text
stringlengths 0
316k
| year
stringclasses 50
values | No
stringclasses 911
values |
|---|---|---|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1179–1188, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics The Influence of Discourse on Syntax A Psycholinguistic Model of Sentence Processing Amit Dubey Abstract Probabilistic models of sentence comprehension are increasingly relevant to questions concerning human language processing. However, such models are often limited to syntactic factors. This paper introduces a novel sentence processing model that consists of a parser augmented with a probabilistic logic-based model of coreference resolution, which allows us to simulate how context interacts with syntax in a reading task. Our simulations show that a Weakly Interactive cognitive architecture can explain data which had been provided as evidence for the Strongly Interactive hypothesis. 1 Introduction Probabilistic grammars have been found to be useful for investigating the architecture of the human sentence processing mechanism (Jurafsky, 1996; Crocker and Brants, 2000; Hale, 2003; Boston et al., 2008; Levy, 2008; Demberg and Keller, 2009). For example, probabilistic models shed light on so-called locality effects: contrast the non-probabilistic hypothesis that dependants which are far away from their head always cause processing difficulty for readers due to the cost of storing the intervening material in memory (Gibson, 1998), compared to the probabilistic prediction that there are cases when faraway dependants facilitate processing, because readers have more time to predict the head (Levy, 2008). Using a computational model to address fundamental questions about sentence comprehension motivates the work in this paper. So far, probabilistic models of sentence processing have been largely limited to syntactic factors. This is unfortunate because many outstanding questions in psycholinguistics concern interactions between different levels of processing. This paper addresses this gap by building a computational model which simulates the influence of discourse on syntax. Going beyond the confines of syntax alone is a sufficiently important problem that it has attracted attention from other authors. In the literature on probabilistic modeling, though, the bulk of this work is focused on lexical semantics (e.g. Pad´o et al., 2006; Narayanan and Jurafsky, 1998) or only considers syntactic decisions in the preceeding text (e.g. Dubey et al., 2009; Levy and Jaeger, 2007). This is the first model we know of which introduces a broad-coverage sentence processing model which takes the effect of coreference and discourse into account. A major question concerning discourse-syntax interactions involves the strength of communication between discourse and syntactic information. The Weakly Interactive (Altmann and Steedman, 1988) hypothesis states that a discourse context can reactively prune syntactic choices that have been proposed by the parser, whereas the Strongly Interactive hypothesis posits that context can proactively suggest choices to the syntactic processor. Support for Weak Interaction comes from experiments in which there are temporary ambiguities, or garden paths, which cause processing difficulty. The general finding is that supportive contexts can reduce the effect of the garden path. However, Grodner et al. (2005) found that supportive contexts even facilitate the processing of unambiguous sentences. As there are no incorrect analyses to prune in unambiguous structures, the authors claimed their results were not consistent with the Weakly Interactive hypothesis, and suggested that their results were best explained by a Strongly Interactive processor. The model we present here implements the Weakly Interactive hypothesis, but we will show that it can nonetheless successfully simulate the results of Grodner et al. (2005). There are three main parts of the model: a syntactic processor, a coreference resolution system, and a simple pragmatics processor which computes certain limited forms of discourse coherence. Following Hale (2001) and Levy (2008), among others, the syntactic processor uses an incremental probabilistic Earley parser to compute a metric which correlates with increased reading difficulty. The coreference resolution system is implemented 1179 in a probabilistic logic known as Markov Logic (Richardson and Domingos, 2006). Finally, the pragmatics processing system contains a small set of probabilistic constraints which convey some intuitive facts about discourse processing. The three components form a pipeline, where each part is probabilistically dependent on the previous one. This allows us to combine all three into a single probability for each reading of an input sentence. The rest of the paper is structured as follows. In Section 2, we discuss the details two experiments showing support of the Weakly and Strongly Interactive hypotheses: we discuss Grodner et al.’s result on unambiguous syntactic structures and we present a new experiment on involving a garden path which was designed to be similar to the Grodner et al. experiment. Section 3 introduces technical details of model, and Section 4 shows the predictions of the model on the experiments discussed in Section 2. Finally, we discuss the theoretical consequences of these predictions in Section 5. 2 Cognitive Experiments 2.1 Discourse and Ambiguity Resolution There is a fairly large literature on garden path experiments involving context (Crain and Steedman, 1985; Mitchell et al., 1992, ibid). The experiments by Altmann and Steedman (1988) involved PP attachment ambiguity. Other authors (e.g. Spivey and Tanenhaus, 1998) have used reduced relative clause attachment ambiguity. In order to be more consistent with the design of the experiment in Section 2.2, however, we performed our own reading-time experiment which partially replicated previous results.1 The experimental items all had a target sentence containing a relative clause, and one of two possible context sentences, one of which supports the relative clause reading and the other which does not. The context sentence was one of: (1) a. There were two postmen, one of whom was injured and carried by paramedics, and another who was unhurt. b. Although there was a medical emergency at the post office earlier today, regular mail delivery was unaffected. 1This experiment was previously reported by Dubey et al. (2010). The target sentences, which were drawn from the experiment of McRae et al. (1998), were either the reduced or unreduced sentences similar to: (2) The postman who was carried by the paramedics was having trouble breathing. The reduced version of the sentence is produced by removing the words who was. We measured reading times in the underlined region, which is the first point at which there is evidence for the relative clause interpretation. The key evidence is given by the word ‘by’, but the previous word is included as readers often do not fixate on short function words, but rather process them while overtly fixating on the previous word (Rayner, 1998). The relative clauses in the target sentence act as restrictive relative clauses, selecting one referent from a larger set. The target sentences are therefore more coherent in a context where a restricted set and a contrast set are easily available, than one in which these sets are absent. This makes the context in Example (1-a) supportive of a reduced relative reading, and the context in Example (1-b) unsupportive of a reduced relative clause. Other experiments, for instance Spivey and Tanenhaus (1998), used an unsupportive context where only one postman was mentioned. Our experiments used a neutral context, where no postmen are mentioned, to be more similar to the Grodner et al. experiment, as described below. Overall, there were 28 items, and 28 participants read these sentences using an EyeLink II eyetracker. Each participant read items one at a time, with fillers between subsequent items so as to obfuscate the nature of the experiment. Results An ANOVA revealed that all conditions with a supportive context were read faster than one with a neutral context (i.e. a main effect of context), and all conditions with unambiguous syntax were read faster than those with a garden path (i.e. a main effect of ambiguity). Finally, there was a statisically significant interaction between syntax and discourse whereby context decreases reading times much more when a garden path is present compared to an unambiguous structure. In other words, a supportive context helped reduce the effect of a garden path. This is the prediction made by both the Weakly Interactive and Strongly Interactive hypothesis. The pattern of results are shown in Figure 2a in Section 4, where they are directly compared to the model results. 1180 2.2 Discourse and Unambiguous Syntax As mentioned in the Introduction, Grodner et al. (2005) proposed an experiment with a supportive or unsupportive discourse followed by an unambiguous target sentence. In their experiment, the target sentence was one of the following: (3) a. The director that the critics praised at a banquet announced that he was retiring to make room for young talent in the industry. b. The director, who the critics praised at a banquet, announced that he was retiring to make room for young talent in the industry. They also manipulated the context, which was either supportive of the target, or a null context. The two supportive contexts are: (4) a. A group of film critics praised a director at a banquet and another director at a film premiere. b. A group of film critics praised a director and a producer for lifetime achievement. The target sentence in (3-a) is a restrictive relative clause, as in the garden path experiments. However, the sentence in (3-b) is a non-restrictive relative clause, which does not assume the presence of a constrast set. Therefore, the context (4-a) is only used with the restrictive relative clause, and the context (4-b), where only one director is mentioned, is used as the context for the non-restrictive relative clause. In the conditions with a null context, the target sentence was not preceded by any contextual sentence. Results Grodner et al. measured residual reading times, i.e. reading times compared to a baseline in the embedded subject NP (‘the critics’). They found that the supportive contexts decreased reading time, and that this effect was stronger for restrictive relatives compared to nonrestricted relatives. As there was no garden path, and hence no incorrect structure for the discourse processor to prune, the authors conclude that this must be evidence for the Strongly Interactive hypothesis. Unlike the garden path experiment above, these results do not appear to be consistent with a Weakly Interactive model. We plot their results in Figure 3a in Section 4, where they are S NP NP The postman VP VBD carried PP IN by NP-LGS The paramedics VP . . . (a) Standard WSJ Tree S NP NPbase The postman VP-LGS VBD1 carried PP:by IN:by by NPbase-LGS The paramedics VP . . . (b) Minimally Modified Tree Figure 1: A schematic representation of the smallest set of grammar transformations which we found were required to accurately parse the experimental items. directly compared to the model results. Because these results are computed as regressions against a baseline, a reading time of 0ms indicates average difficulty, with negative numbers showing some facilitation has occured, and positive number indicating reading difficulty. 3 Model The model comprises three parts: a parser, a coreference resolution system, and a pragmatics subsystem. Let us look at each individually. 3.1 Parser The parser is an incremental unlexicalized probabilistic Earley parser, which is capable of computing prefix probabilities. A PCFG parser outputs the generative probability Pparser(w, t), where w is the text and t is a parse tree. A probabilistic Earley parser can retrieve all possible derivations at word i (Stolcke, 1995), allowing us to compute the probability P(wi . . . w0) = P t Pparser(wi . . . w0, t). Using the prefix probability, we can compute the word-by-word Surprisal (Hale, 2001), by taking the log ratio of the previous word’s prefix probability against this word’s prefix probability: log P(wi−1 . . . w0) P(wi . . . w0) (1) Higher Surprisal scores are interpreted as 1181 being correlated with more reading difficulty, and likewise lower scores with greater reading ease. For most of the remainder of the paper we will simply refer to the prefix probability at word i as P(w). While the prefix probability as presented here is suitable for syntax-based computations, a main technical contribution of our model, detailed in Sections 3.2 and 3.3 below, is that we include non-syntactic probabilities in the computation of Surprisal. As per Hale’s original suggestion, our parser can compute Surprisal using an exhaustive search, which entails summing over each licensed derivation. This can be done efficiently using the packed representation of an Earley chart. However, as the coreference processor takes trees as input, we must therefore unpack parses before resolving referential ambiguity. Given the ambiguity of our grammar, this is not tractable. Therefore, we only consider an n-best list when computing Surprisal. As other authors have found that a relatively small set of analyses can give meaningful predictions (Brants and Crocker, 2000; Boston et al., 2008), we set n = 10. The parser is trained on the Wall Street Journal (WSJ) section of the Penn treebank. Unfortunately, the standard WSJ grammar is not able to give correct incremental parses to our experimental items. We found we could resolve this problem by using four simple transformations, which are shown in Figure 1: (i) adding valency information to verb POS tags (e.g. VBD1 represents a transitive verb); (ii) we lexicalize ‘by’ prepositions; (iii) VPs containing a logical subject (i.e. the agent), get the -LGS label; (iv) non-recursive NPs are renamed NPbase (the coreference system treats each NPbase as a markable). 3.2 Discourse Processor The primary function of the discourse processing module is to perform coreference resolution for each mention in an incrementally processed text. Because each mention in a coreference chains is transitive, we cannot use a simple classifier, as they cannot enforce global transitivity constraints. Therefore, this system is implemented in Markov Logic (Richardson and Domingos, 2006), a probabilistic logic, which does allow us to include such constraints. Markov Logic attempts to combine logic with probabilities by using a Markov random field where logical formulas are features. The Expression Meaning Coref (x, y) x is coreferent with y. First(x) x is a first mention. Order(x, y) x occurs before y. SameHead(x, y) Do x and y share the same syntactic head? ExactMatch(x, y) x and y are same string. SameNumber(x, y) x and y match in number. SameGender(x, y) x and y match in gender. SamePerson(x, y) x and y match in person. Distance(x, y, d) The distance between x and y, in sentences. Pronoun(x) x is a pronoun. EntityType(x, e) x has entity type e (person, organization, etc.) Table 1: Predicates used in the Markov Logic Network Markov Logic Network (MLN) we used for our system uses similar predicates as the MLN-based corference resolution system of Huang et al. (2009).2 Our MLN uses the predicates listed in Table 1. Two of these predicates, Coref and First, are the output of the MLN – they provide a labelling of coreference mentions into entity classes. Note that, unlike Huang et al., we assume an ordering on x and y if Coref (x, y) is true: y must occur earlier in the document than x. The remaining predicates in Table 1 are a subset of features used by other coreference resolution systems (cf. Soon et al., 2001). The predicates we use involve matching strings (checking if two mentions share a head word or if they are exactly the same string), matching argreement features (if the gender, number or person of pairs of NPs are the same; especially important for pronouns), the distance between mentions, and if mentions have the same entity type (i.e. do they refer to a person, organization, etc.) As our main focus is not to produce a state-of-the-art coreference system, we do not include predicates which are irrevelant for our simulations even if they have been shown to be effective for coreference resolution. For example, we do not have predicates if two mentions are in an apposition relationship, or if two mentions are synonyms for each other. Table 2 lists the actual logical formulae which are used as features in the MLN. It should be 2As we are not interested in unsupervised inference, the system of Poon and Domingos (2008) was unsuitable for our needs. 1182 Description Rule Transitivity Coref (x, z) ∧Coref (y, z) ∧Order(x, y) ⇒ Coref (x, y) Coref (x, y) ∧Coref (y, z) ⇒ Coref (x, z) Coref (x, y) ∧Coref (x, z) ∧Order(y, z) ⇒ Coref (y, z) First Mentions Coref (x, y) ⇒¬First(x) First(x) ⇒¬Coref (x, y) String Match ExactMatch(x, y) ⇒ Coref (x, y) SameHead(x, y) ⇒ Coref (x, y) Pronoun Pronoun(x) ∧Pronoun(y) ∧SameGender(x, y) ⇒ Coref (x, y) Pronoun(x) ∧Pronoun(y) ∧SameNumber(x, y) ⇒ Coref (x, y) Pronoun(x) ∧Pronoun(y) ∧SamePerson(x, y) ⇒ Coref (x, y) Other EntityType(x, e) ∧EntityType(y, e) ⇒ Coref (x, y) Distance(x, y, +d) ⇒ Coref (x, y) Table 2: Rules used in the Markov Logic Network noted that, because we are assuming an order on the arguments of Coref (x, y), we need three formulae to capture transivity relationships. To test that the coreference resolution system was producing meaningful results, we evaluated our system on the test section of the ACE-2 dataset. Using b3 scoring (Bagga and Baldwin, 1998), which computes the overlap of a proposed set with the gold set, the system achieves an F-score of 65.4%. While our results are not state-of-the-art, they are reasonable considering the brevity of our feature list. The discourse model is run iteratively at each word. This allows us to find a globally best assignment at each word, which can be reanalyzed at a later point in time. It assumes there is a mention for each base NP outputted by the parser, and for all ordered pairs of mentions x, y, it outputs all the ‘observed’ predicates (i.e. everything but First and Coref ), and feeds them to the Markov Logic system. At each step, we compute both the maximum a posteriori (MAP) assignment of coreference relationships as well as the probability that each individual coreference assignment is true. Taken together, they allow us to calculate, for a coreference assignment c, Pcoref(c|w, t) where w is the text input (of the entire document until this point), and t is the parse of each tree in the document up to and including the current incremental parse. As we have previously calculated Pparser(w, t), it is then possible to compute the joint probability P(c, w, t) at each word, and therefore the prefix probability P(w) due to syntax and coreference. Overall, we have: P(w) = X c X t P(c, w, t) = X c X t Pcoref(c|w, t)Pparser(w, t) Note that we only consider one possible assignment of NPs to coreference entities per parse, as we only retrieve the probabilities of the MAP solution. 3.3 Pragmatics Processor The effect of context in the experiments described in Section 2 cannot be fully explained using a coreference resolution system alone. In the case of restrictive relative clauses, the referential ‘mismatch’ in the unsupported conditions is caused by an expectation elicited by a restrictive relative clause which is inconsistent with the previous discourse when there is no salient restricted subset of a larger set. When the larger set is not found in the discourse, the relative clause becomes incoherent given the context, causing reading difficulty. Modeling this coherence constraint is essentially a pragmatics problem, and is under the purview of the pragmatics processor in our system. The pragmatics processor is quite specialised and, although the information it encapsulates is quite intuitive, it nonetheless relies on hand-coded expert knowledge. The pragmatics processor takes as input an incremental pragmatics configuration p and computes the probability Pprag(p|w, t, c). The pragmatics configuration we consider is quite simple. It is a 3-tuple where one element is true if the current noun phrase being processed is a discourse new definite noun phrase, the second 1183 element is true if the current NP is a discourse new indefinite noun phrase, and the final element is true if we encounter an unsupported restrictive relative clause. We simply conjecture that there is little processing cost (and hence a high probability) if the entire vector is false; there is a small processing cost for discourse new indefinites, a slightly larger processing cost for discourse new definites and a large processing cost for an incoherent reduced relative clause. The first two elements of the 3-tuple depend on the identity of the determiner as recovered by the parser, and on whether the coreference system adduces the predicate First for the current NP. As the coreference system wasn’t designed to find anaphoric contrast sets, these sets were found using a simple post-processing check. This postprocessing approach worked well for our experimental items, but finding such sets is, in general, quite a difficult problem (Modjeska et al., 2003). The distribution Pprag(p|w, t, c) applies a processing penalty for an unsupported restrictive relative clause whenever a restrictive relative clause is in the n best list. Because Surprisal computes a ratio of probabilities, this in effect means we only pay this penality when an unsupported restrictive relative clause first appears in the n best list (otherwise the effect is cancelled out). The penalty for discourse new entities is applied on the first word (ignoring punctuation) following the end of the NP. This spillover processing effect is simply a matter of modeling convenience: without it, we would have to compute Surprisal probabilities over regions rather than individual words. Thus, the overall prefix probability can be computed as: P(w) = P p,c,t Pprag(p|w, t, c)Pcoref(c|w, t)Pparser(w, t), which is then substituted in Equation (1) to get a Surprisal prediction for the current word. 4 Evaluation 4.1 Method When modeling the garden path experiment we presented in Section 2.1, we compute Surprisal values on the word ‘by’, which is the earliest point at which there is evidence for a relative clause interpretation. For the Grodner et al. experiment, we compute Surprisal values on the relativiser ‘who’ or ‘that’. Again, this is the earliest point at which there is evidence for a relative clause, and depending upon the presence or absence of a preceding comma, it will be known to be restrictive or nonrestrictive clause. In addition to the overall Surprisal values, we also compute syntactic Surprisal scores, to test if there is any benefit from the discourse and pragmatics subsystems. As we are outputting n best lists for each parse, it is also straightforward to compute other measures which predict reading difficulty, including pruning (Jurafsky, 1996), whereby processing difficulty is predicted when a parse is removed from the n best list, and attention shift (Crocker and Brants, 2000), which predicts parsing difficulty at words where the most highly ranked parse flips from one interpretation to another. For the garden path experiment, the simulation was run on each of the 28 experimental items in each of the 4 conditions, resulting in a total of 112 runs. For the Grodner et al. experiment, the simulation was run on each of the 20 items in each of the 4 conditions, resulting in a total of 80 runs. For each run, the model was reset, purging all discourse information gained while reading earlier items. As the system is not stochastic, two runs using the exact same items in the same condition will produce the same result. Therefore, we made no attempt to model by-subject variability, but we did perform by-item ANOVAs on the system output. 4.2 Results Garden Path Experiment The simulated results of our experiment are shown in Figure 2. Comparing the full simulated results in Figure 2b to the experimental results in Figure 2a, we find that the simulation, like the actual experiment, finds both main effects and an interaction: there is a main effect of context whereby a supportive context facilitates reading, a main effect of syntax whereby the garden path slows down reading, and an interaction in that the effect of context is strongest in the garden path condition. All these effects were highly significant at p < 0.01. The pattern of results between the full simulation and the experiment differed in two ways. First, the simulated results suggested a much larger reading difficulty due to ambiguity than the experimental results. Also, in the unambiguous case, the model predicted a null cost of an unsupportive context on the word ‘by’, because the model bears the cost of an unsupportive context earlier in the sentence, and assumes no spillover to the word ‘by’. Finally, we note that the syntax-only simulation, shown in Figure 2c, only produced a main effect of ambigu1184 (a) Results from our garden path experiment (b) Simulation of our garden path experiment (c) Syntax-only simulation Figure 2: The simulated results predict the same interaction as the garden path experiment, but show a stronger main effect of ambiguity, and no influence of discourse in the unambiguous condition on the word ‘by’. (a) Results from the Grodner et al. experiment (b) Simulation of the Grodner et al. experiment (c) Syntax-only simulation Figure 3: The simulated results predict the outcome of the Grodner et al. experiment. ity, and was not able to model the effect of context. Grodner et al. Experiment The simulated results of the Grodner et al. experiment are shown in Figure 3. In this experiment, the pattern of simulated results in Figure 3b showed a much closer resemblance to the experimental results in Figure 3a than the garden path experiment. There is a main effect of context, which is much stronger in the restrictive relative case compared to nonrestrictive relatives. As with the garden path experiment, the ANOVA reported that all effects were significant at the p < 0.01 level. Again, as we can see from Figure 3c, there was no effect of context in the syntax-only simulation. The numerical trend did show a slight facilitation in the unrestricted supported condition, with a Surprisal of 4.39 compared to 4.41 in the supported case, but this difference was not significant. 4.3 Discussion We have shown that our incremental sentence processor augmented with discourse processing can successfully simulate syntax-discourse interaction effects which have been shown in the literature. The difference between a Weakly Interactive and Strongly Interactive model can be thought of computationally in terms of a pipeline architecture versus joint inference. In a weaker sense, even a pipeline architecture where the discourse can influence syntactic probabilities could be claimed to be a Strongly Interactive model. However, as our model uses a pipeline where syntactic probabilities are independent of the discourse, we claim that our model is Weakly Interactive. Unlike Altmann and Steedman, who posited that the discourse processor actually removes parsing hypotheses, we were able to simulate this pruning behaviour by simply re-weighting parses in our coreference and pragmatics modules. The fact that a Weakly Interactive system can simulate the result of an experiment proposed in support of the Strongly Interactive hypothesis is initially counter-intuitive. However, this naturally falls out from our decision to use a probabilistic 1185 S NP NPbase The postman VP-LGS VBD1 carried PP:by IN:by by . . . . . . (a) Best parse: p = 9.99 × 10−10 main clause, expecting more dependents S NP NPbase The postman VP-LGS VBD1 carried PP:by IN:by by . . . (b) 2nd parse: p = 9.93 × 10−10 main clause, no more dependents S NP NPbase The postman VP-LGS VBD1 carried PP:by IN:by by . . . . . . (c) 3rd parse: p = 7.69×−10 relative clause Figure 4: The top three parses on the word ‘by’ in the our first experimental item. model: a lower probability, even in an unambiguous structure, is associated with increased reading difficulty. As an aside, we note that when using realistic computational grammars, even the structures used in the Grodner et al. experiment are not unambiguous. In the restrictive relative clause condition, even though there was not any competition between a relative and main clause reading, our n best list was at all times filled with analyses. For example, on the word ‘who’ in the restricted relative clause condition, the parser is already predicting both the subject-relative (‘the postman who was bit by the dog’) and object-relative (‘the postman who the dog bit’) readings. Overall, these results are supportive of the growing importance of probabilistic reasoning as a model of human cognitive behaviour. Therefore, especially with respect to sentence processing, it is necessary to have a proper understanding of how probabilities are linked to real-world behaviours. We note that Surprisal does indeed show processing difficulty on the word ‘by’ in the garden path experiment. However, Figure 4 (which shows the top three parses on the word ‘by’) indicates that not only are there still main clause interpretations present, but in fact, the top two parses are main clause interpretations. This is also true if we limit ourselves to syntactic probabilities (which are the probabilities listed in Figure 4). This suggests that neither Jurafsky (1996)’s notion of pruning as processing difficulty nor Crocker and Brants (2000) notion of attention shifts would correctly predict higher reading times on a region containing the word ‘by’. In fact, the main clause interpretation remains the highestranked interpretation until it is finally pruned at an auxiliary of the main verb of the sentence (‘The postman carried by the paramedics was having’). This result is curious as our experimental items closely match some of those simulated by Crocker and Brants (2000). We conjecture that the difference between our attention shift prediction and theirs is due to differences in the grammar. It is possible that using a more highly tuned grammar would result in attention shift making the correct prediction, but this possibly shows one benefit of using Surprisal as a linking hypothesis. Because Surprisal sums over several derivations, it is not as reliant upon the grammar as the attention shift or pruning linking hypotheses. 5 Conclusions The main result of this paper is that it is possible to produce a Surprisal-based sentence processing model which can simulate the influence of discourse on syntax in both garden path and unambiguous sentences. Computationally, the inclusion of Markov Logic allowed the discourse module to compute well-formed coreference chains, and opens two avenues of future research. First, it ought to be possible to make the probabilistic logic more naturally incremental, rather than re-running from scratch at each word. Second, we would like to make greater use of the logical elements by applying it to problems where inference is necessary, such as resolving bridging anaphora (Haviland and Clark, 1974). Our primary cognitive finding that our model, which assumes the Weakly Interactive hypothesis (whereby discourse is influenced by syntax in a reactive manner), is nonetheless able to simulate the experimental results of Grodner et al. (2005), which were claimed by the authors to be in 1186 support of the Strongly Interactive hypothesis. This suggests that the evidence is in favour of the Strongly Interactive hypothesis may be weaker than thought. Finally, we found that the attention shift (Crocker and Brants, 2000) and pruning (Jurafsky, 1996) linking theories are unable to correctly simulate the results of the garden path experiment. Although our main results above underscore the usefulness of probabilistic modeling, this observation emphasizes the importance of finding a tenable link between probabilities and behaviours. Acknowledgements We would like to thank Frank Keller, Patrick Sturt, Alex Lascarides, Mark Steedman, Mirella Lapata and the anonymous reviewers for their insightful comments. We would also like to thank ESRC for their financial supporting on grant RES-062-23-1450. References Gerry Altmann and Mark Steedman. Interaction with context during human sentence processing. Cognition, 30:191–238, 1988. Amit Bagga and Breck Baldwin. Algorithms for scoring coreference chains. In The First International Conference on Language Resources and Evaluation Workshop on Linguistics Coreference (LREC 98), 1998. Marisa Ferrara Boston, John T. Hale, Reinhold Kliegl, and Shravan Vasisht. Surprising parser actions and reading difficulty. In Proceedings of ACL-08:HLT, Short Papers, pages 5–8, 2008. Thorsten Brants and Matthew Crocker. Probabilistic parsing and psychological plausibility. In Proceedings of 18th International Conference on Computational Linguistics (COLING-2000), pages 111–117, 2000. Stephen Crain and Mark Steedman. On not being led down the garden path: the use of context by the psychological syntax processor. In D. Dowty, L. Karttunen, and A. Zwicky, editors, Natural language parsing: Psychological, computational, and theoretical perspectives. Cambridge University Press, 1985. Matthew Crocker and Thorsten Brants. Wide coverage probabilistic sentence processing. Journal of Psycholinguistic Research, 29(6): 647–669, 2000. Vera Demberg and Frank Keller. A computational model of prediction in human parsing: Unifying locality and surprisal effects. In Proceedings of the 29th meeting of the Cognitive Science Society (CogSci-09), 2009. Amit Dubey, Frank Keller, and Patrick Sturt. A probabilistic corpus-based model of parallelism. Cognition, 109(2):193–210, 2009. Amit Dubey, Patrick Sturt, and Frank Keller. The effect of discourse inferences on syntactic ambiguity resolution. In Proceedings of the 23rd Annual CUNY Conference on Human Sentence Processing (CUNY 2010), page 151, 2010. Ted Gibson. Linguistic complexity: Locality of syntactic dependencies. Cognition, 68:1–76, 1998. Daniel J. Grodner, Edward A. F. Gibson, and Duane Watson. The influence of contextual constrast on syntactic processing: Evidence for strong-interaction in sentence comprehension. Cognition, 95(3):275–296, 2005. John T. Hale. A probabilistic earley parser as a psycholinguistic model. In In Proceedings of the Second Meeting of the North American Chapter of the Asssociation for Computational Linguistics, 2001. John T. Hale. The information conveyed by words in sentences. Journal of Psycholinguistic Research, 32(2):101–123, 2003. Susan E. Haviland and Herbert H. Clark. What’s new? acquiring new information as a process in comprehension. Journal of Verbal Learning and Verbal Behavior, 13:512–521, 1974. Shujian Huang, Yabing Zhang, Junsheng Zhou, and Jiajun Chen. Coreference resolution using markov logic. In Proceedings of the 2009 Conference on Intelligent Text Processing and Computational Linguistics (CICLing 09), 2009. D. Jurafsky. A probabilistic model of lexical and syntactic access and disambiguation. Cognitive Science, 20:137–194, 1996. Roger Levy. Expectation-based syntactic comprehension. Cognition, 106(3):1126–1177, March 2008. Roger Levy and T. Florian Jaeger. Speakers optimize information density through syntactic reduction. In Proceedings of the Twentieth Annual Conference on Neural Information Processing Systems, 2007. 1187 Ken McRae, Michael J. Spivey-Knowlton, and Michael K. Tanenhaus. Modeling the influence of thematic fit (and other constraints) in on-line sentence comprehension. Journal of Memory and Language, 38:283–312, 1998. Don C. Mitchell, Martin M. B. Corley, and Alan Garnham. Effects of context in human sentence parsing: Evidence against a discourse-baed proposal mechanism. Journal of Experimental Psychology: Learning, Memory and Cognition, 18(1):69–88, 1992. Natalia N. Modjeska, Katja Markert, and Malvina Nissim. Using the web in machine learning for other-anaphora resolution. In Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing (EMNLP-2003), pages 176–183, Sapporo, Japan, 2003. Shrini Narayanan and Daniel Jurafsky. Bayesian models of human sentence processing. In Proceedings of the 20th Annual Conference of the Cognitive Science Society (CogSci 98), 1998. Ulrike Pad´o, Matthew Crocker, and Frank Keller. Modelling semantic role plausability in human sentence processing. In Proceedings of the 28th Annual Conference of the Cognitive Science Society (CogSci 2006), pages 657–662, 2006. Hoifung Poon and Pedro Domingos. Joint unsupervised coreference resolution with markov logic. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing (EMNLP-08), 2008. Keith Rayner. Eye movements in reading and information processing: 20 years of research. Psychological Bulletin, 124(3):372–422, 1998. Matthew Richardson and Pedro Domingos. Markov logic networks. Machine Learning, 62 (1-2):107–136, 2006. W. M. Soon, H. T. Ng, and D. C. Y. Lim. A machine learning approach to coreference resolution of noun phrases. Computational Linguistics, 27(4):521–544, 2001. M. J. Spivey and M. K. Tanenhaus. Syntactic ambiguity resolution in discourse: Modeling the effects of referential context and lexical frequency. Journal of Experimental Psychology: Learning, Memory and Cognition, 24(6): 1521–1543, 1998. Andreas Stolcke. An efficient probabilistic context-free parsing algorithm that computes prefix probabilities. Computational Linguistics, 21(2):165–201, 1995. 1188
|
2010
|
120
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1189–1198, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Complexity Metrics in an Incremental Right-corner Parser Stephen Wu Asaf Bachrach† Carlos Cardenas∗ William Schuler◦ Department of Computer Science, University of Minnesota † Unit de Neuroimagerie Cognitive INSERM-CEA ∗Department of Brain & Cognitive Sciences, Massachussetts Institute of Technology ◦University of Minnesota and The Ohio State University [email protected] †[email protected] ∗[email protected] ◦[email protected] Abstract Hierarchical HMM (HHMM) parsers make promising cognitive models: while they use a bounded model of working memory and pursue incremental hypotheses in parallel, they still achieve parsing accuracies competitive with chart-based techniques. This paper aims to validate that a right-corner HHMM parser is also able to produce complexity metrics, which quantify a reader’s incremental difficulty in understanding a sentence. Besides defining standard metrics in the HHMM framework, a new metric, embedding difference, is also proposed, which tests the hypothesis that HHMM store elements represents syntactic working memory. Results show that HHMM surprisal outperforms all other evaluated metrics in predicting reading times, and that embedding difference makes a significant, independent contribution. 1 Introduction Since the introduction of a parser-based calculation for surprisal by Hale (2001), statistical techniques have been become common as models of reading difficulty and linguistic complexity. Surprisal has received a lot of attention in recent literature due to nice mathematical properties (Levy, 2008) and predictive ability on eye-tracking movements (Demberg and Keller, 2008; Boston et al., 2008a). Many other complexity metrics have been suggested as mutually contributing to reading difficulty; for example, entropy reduction (Hale, 2006), bigram probabilities (McDonald and Shillcock, 2003), and split-syntactic/lexical versions of other metrics (Roark et al., 2009). A parser-derived complexity metric such as surprisal can only be as good (empirically) as the model of language from which it derives (Frank, 2009). Ideally, a psychologically-plausible language model would produce a surprisal that would correlate better with linguistic complexity. Therefore, the specification of how to encode a syntactic language model is of utmost importance to the quality of the metric. However, it is difficult to quantify linguistic complexity and reading difficulty. The two commonly-used empirical quantifications of reading difficulty are eye-tracking measurements and word-by-word reading times; this paper uses reading times to find the predictiveness of several parser-derived complexity metrics. Various factors (i.e., from syntax, semantics, discourse) are likely necessary for a full accounting of linguistic complexity, so current computational models (with some exceptions) narrow the scope to syntactic or lexical complexity. Three complexity metrics will be calculated in a Hierarchical Hidden Markov Model (HHMM) parser that recognizes trees in right-corner form (the left-right dual of left-corner form). This type of parser performs competitively on standard parsing tasks (Schuler et al., 2010); also, it reflects plausible accounts of human language processing as incremental (Tanenhaus et al., 1995; Brants and Crocker, 2000), as considering hypotheses probabilistically in parallel (Dahan and Gaskell, 2007), as bounding memory usage to short-term memory limits (Cowan, 2001), and as requiring more memory storage for center-embedding structures than for right- or left-branching ones (Chomsky and Miller, 1963; Gibson, 1998). Also, unlike most other parsers, this parser preserves the arceager/arc-standard ambiguity of Abney and John1189 son (1991). Typical parsing strategies are arcstandard, keeping all right-descendants open for subsequent attachment; but since there can be an unbounded number of such open constituents, this assumption is not compatible with simple models of bounded memory. A consistently arc-eager strategy acknowledges memory bounds, but yields dead-end parses. Both analyses are considered in right-corner HHMM parsing. The purpose of this paper is to determine whether the language model defined by the HHMM parser can also predict reading times — it would be strange if a psychologically plausible model did not also produce viable complexity metrics. In the course of showing that the HHMM parser does, in fact, predict reading times, we will define surprisal and entropy reduction in the HHMM parser, and introduce a third metric called embedding difference. Gibson (1998; 2000) hypothesized two types of syntactic processing costs: integration cost, in which incremental input is combined with existing structures; and memory cost, where unfinished syntactic constructions may incur some short-term memory usage. HHMM surprisal and entropy reduction may be considered forms of integration cost. Though typical PCFG surprisal has been considered a forward-looking metric (Demberg and Keller, 2008), the incremental nature of the right-corner transform causes surprisal and entropy reduction in the HHMM parser to measure the likelihood of grammatical structures that were hypothesized before evidence was observed for them. Therefore, these HHMM metrics resemble an integration cost encompassing both backwardlooking and forward-looking information. On the other hand, embedding difference is designed to model the cost of storing centerembedded structures in working memory. Chen, Gibson, and Wolf (2005) showed that sentences requiring more syntactic memory during sentence processing increased reading times, and it is widely understood that center-embedding incurs significant syntactic processing costs (Miller and Chomsky, 1963; Gibson, 1998). Thus, we would expect for the usage of the center-embedding memory store in an HHMM parser to correlate with reading times (and therefore linguistic complexity). The HHMM parser processes syntactic constructs using a bounded number of store states, defined to represent short-term memory elements; additional states are utilized whenever centerembedded syntactic structures are present. Similar models such as Crocker and Brants (2000) implicitly allow an infinite memory size, but Schuler et al. (2008; 2010) showed that a right-corner HHMM parser can parse most sentences in English with 4 or fewer center-embedded-depth levels. This behavior is similar to the hypothesized size of a human short-term memory store (Cowan, 2001). A positive result in predicting reading times will lend additional validity to the claim that the HHMM parser’s bounded memory corresponds to bounded memory in human sentence processing. The rest of this paper is organized as follows: Section 2 defines the language model of the HHMM parser, including definitions of the three complexity metrics. The methodology for evaluating the complexity metrics is described in Section 3, with actual results in Section 4. Further discussion on results, and comparisons to other work, are in Section 5. 2 Parsing Model This section describes an incremental parser in which surprisal and entropy reduction are simple calculations (Section 2.1). The parser uses a Hierarchical Hidden Markov Model (Section 2.2) and recognizes trees in a right-corner form (Section 2.3 and 2.4). The new complexity metric, embedding difference (Section 2.5), is a natural consequence of this HHMM definition. The model is equivalent to previous HHMM parsers (Schuler, 2009), but reorganized into 5 cases to clarify the right-corner structure of the parsed sentences. 2.1 Surprisal and Entropy in HMMs Hidden Markov Models (HMMs) probabilistically connect sequences of observed states ot and hidden states qt at corresponding time steps t. In parsing, observed states are words; hidden states can be a conglomerate state of linguistic information, here taken to be syntactic. The HMM is an incremental, time-series structure, so one of its by-products is the prefix probability, which will be used to calculate surprisal. This is the probability that that words o1..t have been observed at time t, regardless of which syntactic states q1..t produced them. Bayes’ Law and Markov independence assumptions allow this to 1190 be calculated from two generative probability distributions.1 Pre(o1..t)= X q1..t P(o1..t q1..t) (1) def = X q1..t tY τ=1 PΘA(qτ | qτ–1)·PΘB(oτ | qτ) (2) Here, probabilities arise from a Transition Model (ΘA) between hidden states and an Observation Model (ΘB) that generates an observed state from a hidden state. These models are so termed for historical reasons (Rabiner, 1990). Surprisal (Hale, 2001) is then a straightforward calculation from the prefix probability. Surprisal(t) = log2 Pre(o1..t–1) Pre(o1..t) (3) This framing of prefix probability and surprisal in a time-series model is equivalent to Hale’s (2001; 2006), assuming that q1..t ∈Dt, i.e., that the syntactic states we are considering form derivations Dt, or partial trees, consistent with the observed words. We will see that this is the case for our parser in Sections 2.2–2.4. Entropy is a measure of uncertainty, defined as H(x) = −P(x) log2 P(x). Now, the entropy Ht of a t-word string o1..t in an HMM can be written: Ht = X q1..t P(q1..t o1..t) log2 P(q1..t o1..t) (4) and entropy reduction (Hale, 2003; Hale, 2006) at the tth word is then ER(ot) = max(0, Ht−1 −Ht) (5) Both of these metrics fall out naturally from the time-series representation of the language model. The third complexity metric, embedding difference, will be discussed after additional background in Section 2.5. In the implementation of an HMM, candidate states at a given time qt are kept in a trellis, with step-by-step backpointers to the highestprobability q1..t–1.2 Also, the best qt are often kept in a beam Bt, discarding low-probability states. 1Technically, a prior distribution over hidden states, P(q0), is necessary. This q0 is factored and taken to be a deterministic constant, and is therefore unimportant as a probability model. 2Typical tasks in an HMM include finding the most likely sequence via the Viterbi algorithm, which stores these backpointers to maximum-probability previous states and can uniquely find the most likely sequence. This mitigates the problems of large state spaces (e.g., that of all possible grammatical derivations). Since beams have been shown to perform well (Brants and Crocker, 2000; Roark, 2001; Boston et al., 2008b), complexity metrics in this paper are calculated on a beam rather than over all (unbounded) possible derivations Dt. The equations above, then, will replace the assumption q1..t ∈Dt with qt ∈Bt. 2.2 Hierarchical Hidden Markov Models Hidden states q can have internal structure; in Hierarchical HMMs (Fine et al., 1998; Murphy and Paskin, 2001), this internal structure will be used to represent syntax trees and looks like several HMMs stacked on top of each other. As such, qt is factored into sequences of depth-specific variables — one for each of D levels in the HMM hierarchy. In addition, an intermediate variable ft is introduced to interface between the levels. qt def = ⟨q1 t . . . qD t ⟩ (6) ft def = ⟨f1 t . . . fD t ⟩ (7) Transition probabilities PΘA(qt | qt–1) over complex hidden states qt are calculated in two phases: • Reduce phase. Yields an intermediate state ft, in which component HMMs may terminate. This ft tells “higher” HMMs to hold over their information if “lower” levels are in operation at any time step t, and tells lower HMMs to signal when they’re done. • Shift phase. Yields a modeled hidden state qt, in which unterminated HMMs transition, and terminated HMMs are re-initialized from their parent HMMs. Each phase is factored according to levelspecific reduce and shift models, ΘF and ΘQ: PΘA(qt|qt–1) = X ft P(ft|qt–1)·P(qt|ft qt–1) (8) def = X f1..D t D Y d=1 PΘF(fd t |fd+1 t qd t–1qd–1 t–1 ) · PΘQ(qd t |fd+1 t fd t qd t–1qd–1 t ) (9) with fD+1 t and q0 t defined as constants. Note that only qt is present at the end of the probability calculation. In step t, ft–1 will be unused, so the marginalization of Equation 9 does not lose any information. 1191 . . . . . . . . . . . . f3 t−1 f2 t−1 f1 t−1 q1 t−1 q2 t−1 q3 t−1 ot−1 f3 t f2 t f1 t q1 t q2 t q3 t ot (a) Dependency structure in the HHMM parser. Conditional probabilities at a node are dependent on incoming arcs. d=1 d=2 d=3 word t=1 t=2 t=3 t=4 t=5 t=6 t=7 t=8 the engineers pulled off an engineering trick ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ vbd VBD/PRT ◦ ◦ ◦ dt NP/NN S/VP S/VP S/NP S/NN S/NN S (b) HHMM parser as a store whose elements at each time step are listed vertically, showing a good hypothesis on a sample sentence out of many kept in parallel. Variables corresponding to qd t are shown. S NP DT the NN engineers VP VBD VBD pulled PRT off NP DT an NN NN engineering NN trick (c) A sample sentence in CNF. S S/NN S/NN S/NP S/VP NP NP/NN DT the NN engineers VBD VBD/PRT VBD pulled PRT off DT an NN engineering NN trick (d) The right-corner transformed version of (c). Figure 1: Various graphical representations of HHMM parser operation. (a) shows probabilistic dependencies. (b) considers the qd t store to be incremental syntactic information. (c)–(d) demonstrate the right-corner transform, similar to a left-to-right traversal of (c). In ‘NP/NN’ we say that NP is the active constituent and NN is the awaited. The Observation Model ΘB is comparatively much simpler. It is only dependent on the syntactic state at D (or the deepest active HHMM level). PΘB(ot | qt) def = P(ot | qD t ) (10) Figure 1(a) gives a schematic of the dependency structure of Equations 8–10 for D = 3. Evaluations in this paper are done with D = 4, following the results of Schuler, et al. (2008). 2.3 Parsing right-corner trees In this HHMM formulation, states and dependencies are optimized for parsing right-corner trees (Schuler et al., 2008; Schuler et al., 2010). A sample transformation between CNF and right-corner trees is in Figures 1(c)–1(d). Figure 1(b) shows the corresponding storeelement interpretation3 of the right corner tree in 1(d). These can be used as a case study to see what kind of operations need to occur in an 3This is technically a pushdown automoton (PDA), where the store is limited to D elements. When referring to directions (e.g., up, down), PDAs are typically described opposite of the one in Figure 1(b); here, we push “up” instead of down. HHMM when parsing right-corner trees. There is one unique set of HHMM state values for each tree, so the operations can be seen on either the tree or the store elements. At each time step t, a certain number of elements (maximum D) are kept in memory, i.e., in the store. New words are observed input, and the bottom occupied element (the “frontier” of the store) is the context; together, they determine what the store will look like at t+1. We can characterize the types of store-element changes by when they happen in Figures 1(b) and 1(d): Cross-level Expansion (CLE). Occupies a new store element at a given time step. For example, at t=1, a new store element is occupied which can interact with the observed word, “the.” At t = 3, an expansion occupies the second store element. In-level Reduction (ILR). Completes an active constituent that is a unary child in the rightcorner tree; always accompanied by an inlevel expansion. At t = 2, “engineers” completes the active NP constituent; however, the 1192 level is not yet complete since the NP is along the left-branching trunk of the tree. In-level Expansion (ILE). Starts a new active constituent at an already-occupied store element; always follows an in-level reduction. With the NP complete in t = 2, a new active constituent S is produced at t=3. In-level Transition (ILT). Transitions the store to a new state in the next time step at the same level, where the awaited constituent changes and the active constituent remains the same. This describes each of the steps from t=4 to t=8 at d=1 . Cross-level Reduction (CLR). Vacates a store element on seeing a complete active constituent. This occurs after t = 4; “off” completes the active (at depth 2) VBD constituent, and vacates store element 2. This is accompanied with an in-level transition at depth 1, producing the store at t=5. It should be noted that with some probability, completing the active constituent does not vacate the store element, and the in-level reduction case would have to be invoked. The in-level/cross-level ambiguity occurs in the expansion as well as the reduction, similar to Abney and Johnson’s arc-eager/arc-standard composition strategies (1991). At t=3, another possible hypothesis would be to remain on store element 1 using an ILE instead of a CLE. The HHMM parser, unlike most other parsers, will preserve this in-level/cross-level ambiguity by considering both hypotheses in parallel. 2.4 Reduce and Shift Models With the understanding of what operations need to occur, a formal definition of the language model is in order. Let us begin with the relevant variables. A shift variable qd t at depth d and time step t is a syntactic state that must represent the active and awaited constituents of right-corner form: qd t def = ⟨gA qd t , gW qd t ⟩ (11) e.g., in Figure 1(b), q1 2=⟨NP,NN⟩=NP/NN. Each g is a constituent from the pre-right-corner grammar, G. Reduce variables f are then enlisted to ensure that in-level and cross-level operations are correct. fd t def = ⟨kfd t , gfd t ⟩ (12) First, kfd t is a switching variable that differentiates between ILT, CLE/CLR, and ILE/ILR. This switching is the most important aspect of fd t , so regardless of what gfd t is, we will use: • fd t ∈F0 when kfd t =0, (ILT/no-op) • fd t ∈F1 when kfd t =1, (CLE/CLR) • fd t ∈FG when kfd t ∈G. (ILE/ILR) Then, gfd t is used to keep track of a completelyrecognized constituent whenever a reduction occurs (ILR or CLR). For example, in Figure 1(b), after time step 2, an NP has been completely recognized and precipitates an ILR. The NP gets stored in gf1 3 for use in the ensuing ILE instead of appearing in the store-elements. This leads us to a specification of the reduce and shift probability models. The reduce step happens first at each time step. True to its name, the reduce step handles in-level and cross-level reductions (the second and third case below): PΘF(fd t | fd+1 t qd t−1qd−1 t−1 ) def = (if f d+1 t ̸∈FG : Jf d t = 0K if f d+1 t ∈FG, f d t ∈F1 : ˜PΘF-ILR,d(f d t | qd t−1 qd−1 t−1 ) if f d+1 t ∈FG, f d t ∈FG : ˜PΘF-CLR,d(f d t | qd t−1 qd−1 t−1 ) (13) with edge cases q0 t and fD+1 t defined as appropriate constants. The first case is just store-element maintenance, in which the variable is not on the “frontier” and therefore inactive. Examining ΘF-ILR,d and ΘF-CLR,d, we see that the produced fd t variables are also used in the “if” statement. These models can be thought of as picking out a fd t first, finding the matching case, then applying the probability models that matches. These models are actually two parts of the same model when learned from trees. Probabilities in the shift step are also split into cases based on the reduce variables. More maintenance operations (first case) accompany transitions producing new awaited constituents (second case below) and expansions producing new active constituents (third and fourth case): PΘQ(qd t | fd+1 t fd t qd t−1qd−1 t ) def = if f d+1 t ̸∈FG : Jqd t = qd t−1K if f d+1 t ∈FG, f d t ∈F0 : ˜PΘQ-ILT,d(qd t | f d+1 t qd t−1 qd−1 t ) if f d+1 t ∈FG, f d t ∈F1 : ˜PΘQ-ILE,d(qd t | f d t qd t−1 qd−1 t ) if f d+1 t ∈FG, f d t ∈FG : ˜PΘQ-CLE,d(qd t | qd−1 t ) (14) 1193 FACTOR DESCRIPTION EXPECTED Word order in narrative For each story, words were indexed. Subjects would tend to read faster later in a story. negative slope Reciprocal length Log of the reciprocal of the number of letters in each word. A decrease in the reciprocal (increase in length) might mean longer reading times. positive slope Unigram frequency A log-transformed empirical count of word occurrences in the Brown Corpus section of the Penn Treebank. Higher frequency should indicate shorter reading times. negative slope Bigram probability A log-transformed empirical count of two-successive-word occurrences, with GoodTuring smoothing on words occuring less than 10 times. negative slope Embedding difference Amount of change in HHMM weighted-average embedding depth. Hypothesized to increase with larger working memory requirements, which predict longer reading times. positive slope Entropy reduction Amount of decrease in the HHMM’s uncertainty about the sentence. Larger reductions in uncertainty are hypothesized to take longer. positive slope Surprisal “Surprise value” of a word in the HHMM parser; models were trained on the Wall Street Journal, sections 02–21. More surprising words may take longer to read. positive slope Table 1: A list of factors hypothesized to contribute to reading times. All data was mean-centered. A final note: the notation ˜PΘ(· | ·) has been used to indicate probability models that are empirical, trained directly from frequency counts of rightcorner transformed trees in a large corpus. Alternatively, a standard PCFG could be trained on a corpus (or hand-specified), and then the grammar itself can be right-corner transformed (Schuler, 2009). Taken together, Equations 11–14 define the probabilistic structure of the HHMM for parsing right-corner trees. 2.5 Embedding difference in the HHMM It should be clear from Figure 1 that at any time step while parsing depth-bounded right-corner trees, the candidate hidden state qt will have a “frontier” depth d(qt). At time t, the beam of possible hidden states qt stores the syntactic state (and a backpointer) along with its probability, P(o1..t q1..t). The average embedding depth at a time step is then µEMB(o1..t) = X qt∈Bt d(qt) · P(o1..t q1..t) P q′ t∈Bt P(o1..t q′ 1..t) (15) where we have directly used the beam notation. The embedding difference metric is: EmbDiff(o1..t) = µEMB(o1..t) −µEMB(o1..t−1) (16) There is a strong computational correspondence between this definition of embedding difference and the previous definition of surprisal. To see this, we rewrite Equations 1 and 3: Pre(o1..t)= X qt∈Bt P(o1..t q1..t) (1′) Surprisal(t) = log2 Pre(o1..t–1) −log2 Pre(o1..t) (3′) Both surprisal and embedding difference include summations over the elements of the beam, and are calculated as a difference between previous and current beam states. Most differences between these metrics are relatively inconsequential. For example, the difference in order of subtraction only assures that a positive correlation with reading times is expected. Also, the presence of a logarithm is relatively minor. Embedding difference weighs the probabilities with center-embedding depths and then normalizes the values; since the measure is a weighted average of embedding depths rather than a probability distribution, µEMB is not always less than 1 and the correspondence with KullbackLeibler divergence (Levy, 2008) does not hold, so it does not make sense to take the logs. Therefore, the inclusion of the embedding depth, d(qt), is the only significant difference between the two metrics. The result is a metric that, despite numerical correspondence to surprisal, models the HHMM’s hypotheses about memory cost. 3 Evaluation Surprisal, entropy reduction, and embedding difference from the HHMM parser were evaluated against a full array of factors (Table 1) on a corpus of word-by-word reading times using a linear mixed-effects model. 1194 The corpus of reading times for 23 native English speakers was collected on a set of four narratives (Bachrach et al., 2009), each composed of sentences that were syntactically complex but constructed to appear relatively natural. Using Linger 2.88, words appeared one-by-one on the screen, and required a button-press in order to advance; they were displayed in lines with 11.5 words on average. Following Roark et al.’s (2009) work on the same corpus, reading times above 1500 ms (for diverted attention) or below 150 ms (for button presses planned before the word appeared) were discarded. In addition, the first and last word of each line on the screen were removed; this left 2926 words out of 3540 words in the corpus. For some tests, a division between open- and closed-class words was made, with 1450 and 1476 words, respectively. Closed-class words (e.g., determiners or auxiliary verbs) usually play some kind of syntactic function in a sentence; our evaluations used Roark et al.’s list of stop words. Open class words (e.g., nouns and other verbs) more commonly include new words. Thus, one may expect reading times to differ for these two types of words. Linear mixed-effect regression analysis was used on this data; this entails a set of fixed effects and another of random effects. Reading times y were modeled as a linear combination of factors x, listed in Table 1 (fixed effects); some random variation in the corpus might also be explained by groupings according to subject i, word j, or sentence k (random effects). yijk = β0 + m X ℓ=1 βℓxijkℓ+ bi + bj + bk + ε (17) This equation is solved for each of m fixedeffect coefficients β with a measure of confidence (t-value = ˆβ/SE(ˆβ), where SE is the standard error). β0 is the standard intercept to be estimated along with the rest of the coefficients, to adjust for affine relationships between the dependent and independent variables. We report factors as statistically significant contributors to reading time if the absolute value of the t-value is greater than 2. Two more types of comparisons will be made to see the significance of factors. First, a model of data with the full list of factors can be compared to a model with a subset of those factors. This is done with a likelihood ratio test, producing (for mixed-effects models) a χ2 1 value and corresponding probability that the smaller model could have produced the same estimates as the larger model. A lower probability indicates that the additional factors in the larger model are significant. Second, models with different fixed effects can be compared to each other through various information criteria; these trade off between having a more explanatory model vs. a simpler model, and can be calculated on any model. Here, we use Akaike’s Information Criterion (AIC), where lower values indicate better models. All these statistics were calculated in R, using the lme4 package (Bates et al., 2008). 4 Results Using the full list of factors in Table 1, fixed-effect coefficients were estimated in Table 2. Fitting the best model by AIC would actually prune away some of the factors as relatively insignificant, but these smaller models largely accord with the significance values in the table and are therefore not presented. The first data column shows the regression on all data; the second and third columns divide the data into open and closed classes, because an evaluation (not reported in detail here) showed statistically significant interactions between word class and 3 of the predictors. Additionally, this facilitates comparison with Roark et al. (2009), who make the same division. Out of the non-parser-based metrics, word order and bigram probability are statistically significant regardless of the data subset; though reciprocal length and unigram frequency do not reach significance here, likelihood ratio tests (not shown) confirm that they contribute to the model as a whole. It can be seen that nearly all the slopes have been estimated with signs as expected, with the exception of reciprocal length (which is not statistically significant). Most notably, HHMM surprisal is seen here to be a standout predictive measure for reading times regardless of word class. If the HHMM parser is a good psycholinguistic model, we would expect it to at least produce a viable surprisal metric, and Table 2 attests that this is indeed the case. Though it seems to be less predictive of open classes, a surprisal-only model has the best AIC (-7804) out of any open-class model. Considering the AIC on the full data, the worst model with surprisal 1195 FULL DATA OPEN CLASS CLOSED CLASS Coefficient Std. Err. t-value Coefficient Std. Err. t-value Coefficient Std. Err. t-value (Intcpt) -9.340·10−3 5.347·10−2 -0.175 -1.237·10−2 5.217·10−2 -0.237 -6.295·10−2 7.930·10−2 -0.794 order -3.746·10−5 7.808·10−6 -4.797∗ -3.697·10−5 8.002·10−6 -4.621∗ -3.748·10−5 8.854·10−6 -4.232∗ rlength -2.002·10−2 1.635·10−2 -1.225 9.849·10−3 1.779·10−2 0.554 -2.839·10−2 3.283·10−2 -0.865 unigrm -8.090·10−2 3.690·10−1 -0.219 -1.047·10−1 2.681·10−1 -0.391 -3.847·10+0 5.976·10+0 -0.644 bigrm -2.074·10+0 8.132·10−1 -2.551∗ -2.615·10+0 8.050·10−1 -3.248∗ -5.052·10+1 1.910·10+1 -2.645∗ embdiff 9.390·10−3 3.268·10−3 2.873∗ 2.432·10−3 4.512·10−3 0.539 1.598·10−2 5.185·10−3 3.082∗ etrpyrd 2.753·10−2 6.792·10−3 4.052∗ 6.634·10−4 1.048·10−2 0.063 4.938·10−2 1.017·10−2 4.857∗ srprsl 3.950·10−3 3.452·10−4 11.442∗ 2.892·10−3 4.601·10−4 6.285∗ 5.201·10−3 5.601·10−4 9.286∗ Table 2: Results of linear mixed-effect modeling. Significance (indicated by ∗) is reported at p < 0.05. (Intr) order rlngth ungrm bigrm emdiff entrpy order .000 rlength -.006 -.003 unigrm .049 .000 -.479 bigrm .001 .005 -.006 -.073 emdiff .000 .009 -.049 -.089 .095 etrpyrd .000 .003 .016 -.014 .020 -.010 srprsl .000 -.008 -.033 -.079 .107 .362 .171 Table 3: Correlations in the full model. (AIC=-10589) outperformed the best model without it (AIC=-10478), indicating that the HHMM surprisal is well worth including in the model regardless of the presence of other significant factors. HHMM entropy reduction predicts reading times on the full dataset and on closed-class words. However, its effect on open-class words is insignificant; if we compare the model of column 2 against one without entropy reduction, a likelihood ratio test gives χ2 1 = 0.0022, p = 0.9623 (the smaller model could easily generate the same data). The HHMM’s average embedding difference is also significant except in the case of openclass words — removing embedding difference on open-class data yields χ2 1 = 0.2739, p = 0.6007. But what is remarkable is that there is any significance for this metric at all. Embedding difference and surprisal were relatively correlated compared to other predictors (see Table 3), which is expected because embedding difference is calculated like a weighted version of surprisal. Despite this, it makes an independent contribution to the full-data and closed-class models. Thus, we can conclude that the average embedding depth component affects reading times — i.e., the HHMM’s notion of working memory behaves as we would expect human working memory to behave. 5 Discussion As with previous work on large-scale parserderived complexity metrics, the linear mixedeffect models suggest that sentence-level factors are effective predictors for reading difficulty — in these evaluations, better than commonly-used lexical and near-neighbor predictors (Pollatsek et al., 2006; Engbert et al., 2005). The fact that HHMM surprisal outperforms even n-gram metrics points to the importance of including a notion of sentence structure. This is particularly true when the sentence structure is defined in a language model that is psycholinguistically plausible (here, boundedmemory right-corner form). This accords with an understated result of Boston et al.’s eye-tracking study (2008a): a richer language model predicts eye movements during reading better than an oversimplified one. The comparison there is between phrase structure surprisal (based on Hale’s (2001) calculation from an Earley parser), and dependency grammar surprisal (based on Nivre’s (2007) dependency parser). Frank (2009) similarly reports improvements in the reading-time predictiveness of unlexicalized surprisal when using a language model that is more plausible than PCFGs. The difference in predictivity due to word class is difficult to explain. One theory may be that closed-class words are less susceptible to random effects because there is a finite set of them for any language, making them overall easier to predict via parser-derived metrics. Or, we could note that since closed-class words often serve grammatical functions in addition to their lexical content, they contribute more information to parser-derived measures than open-class words. Previous work with complexity metrics on this corpus (Roark et al., 2009) suggests that these explanations only account for part of the word-class variation in the performance of predictors. 1196 Further comparsion to Roark et al. will show other differences, such as the lesser role of word length and unigram frequency, lower overall correlations between factors, and the greater predictivity of their entropy metric. In addition, their metrics are different from ours in that they are designed to tease apart lexical and syntactic contributions to reading difficulty. Their notion of entropy, in particular, estimates Hale’s definition of entropy on whole derivations (2006) by isolating the predictive entropy; they then proceed to define separate lexical and syntactic predictive entropies. Drawing more directly from Hale, our definition is a whole-derivation metric based on the conditional entropy of the words, given the root. (The root constituent, though unwritten in our definitions, is always included in the HHMM start state, q0.) More generally, the parser used in these evaluations differs from other reported parsers in that it is not lexicalized. One might expect for this to be a weakness, allowing distributions of probabilities at each time step in places not licensed by the observed words, and therefore giving poor probability-based complexity metrics. However, we see that this language model performs well despite its lack of lexicalization. This indicates that lexicalization is not a requisite part of syntactic parser performance with respect to predicting linguistic complexity, corroborating the evidence of Demberg and Keller’s (2008) ‘unlexicalized’ (POS-generating, not word-generating) parser. Another difference is that previous parsers have produced useful complexity metrics without maintaining arc-eager/arc-standard ambiguity. Results show that including this ambiguity in the HHMM at least does not invalidate (and may in fact improve) surprisal or entropy reduction as readingtime predictors. 6 Conclusion The task at hand was to determine whether the HHMM could consistently be considered a plausible psycholinguistic model, producing viable complexity metrics while maintaining other characteristics such as bounded memory usage. The linear mixed-effects models on reading times validate this claim. The HHMM can straightforwardly produce highly-predictive, standard complexity metrics (surprisal and entropy reduction). HHMM surprisal performs very well in predicting reading times regardless of word class. Our formulation of entropy reduction is also significant except in open-class words. The new metric, embedding difference, uses the average center-embedding depth of the HHMM to model syntactic-processing memory cost. This metric can only be calculated on parsers with an explicit representation for short-term memory elements like the right-corner HHMM parser. Results show that embedding difference does predict reading times except in open-class words, yielding a significant contribution independent of surprisal despite the fact that its definition is similar to that of surprisal. Acknowledgments Thanks to Brian Roark for help on the reading times corpus, Tim Miller for the formulation of entropy reduction, Mark Holland for statistical insight, and the anonymous reviewers for their input. This research was supported by National Science Foundation CAREER/PECASE award 0447685. The views expressed are not necessarily endorsed by the sponsors. References Steven P. Abney and Mark Johnson. 1991. Memory requirements and local ambiguities of parsing strategies. J. Psycholinguistic Research, 20(3):233–250. Asaf Bachrach, Brian Roark, Alex Marantz, Susan Whitfield-Gabrieli, Carlos Cardenas, and John D.E. Gabrieli. 2009. Incremental prediction in naturalistic language processing: An fMRI study. Douglas Bates, Martin Maechler, and Bin Dai. 2008. lme4: Linear mixed-effects models using S4 classes. R package version 0.999375-31. Marisa Ferrara Boston, John T. Hale, Reinhold Kliegl, U. Patil, and Shravan Vasishth. 2008a. Parsing costs as predictors of reading difficulty: An evaluation using the Potsdam Sentence Corpus. Journal of Eye Movement Research, 2(1):1–12. Marisa Ferrara Boston, John T. Hale, Reinhold Kliegl, and Shravan Vasishth. 2008b. Surprising parser actions and reading difficulty. In Proceedings of ACL08: HLT, Short Papers, pages 5–8, Columbus, Ohio, June. Association for Computational Linguistics. Thorsten Brants and Matthew Crocker. 2000. Probabilistic parsing and psychological plausibility. In Proceedings of COLING ’00, pages 111–118. 1197 Evan Chen, Edward Gibson, and Florian Wolf. 2005. Online syntactic storage costs in sentence comprehension. Journal of Memory and Language, 52(1):144–169. Noam Chomsky and George A. Miller. 1963. Introduction to the formal analysis of natural languages. In Handbook of Mathematical Psychology, pages 269–321. Wiley. Nelson Cowan. 2001. The magical number 4 in shortterm memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24:87– 185. Matthew Crocker and Thorsten Brants. 2000. Widecoverage probabilistic sentence processing. Journal of Psycholinguistic Research, 29(6):647–669. Delphine Dahan and M. Gareth Gaskell. 2007. The temporal dynamics of ambiguity resolution: Evidence from spoken-word recognition. Journal of Memory and Language, 57(4):483–501. Vera Demberg and Frank Keller. 2008. Data from eyetracking corpora as evidence for theories of syntactic processing complexity. Cognition, 109(2):193–210. Ralf Engbert, Antje Nuthmann, Eike M. Richter, and Reinhold Kliegl. 2005. SWIFT: A dynamical model of saccade generation during reading. Psychological Review, 112:777–813. Shai Fine, Yoram Singer, and Naftali Tishby. 1998. The hierarchical hidden markov model: Analysis and applications. Machine Learning, 32(1):41–62. Stefan L. Frank. 2009. Surprisal-based comparison between a symbolic and a connectionist model of sentence processing. In Proc. Annual Meeting of the Cognitive Science Society, pages 1139–1144. Edward Gibson. 1998. Linguistic complexity: Locality of syntactic dependencies. Cognition, 68(1):1– 76. Edward Gibson. 2000. The dependency locality theory: A distance-based theory of linguistic complexity. In Image, language, brain: Papers from the first mind articulation project symposium, pages 95–126. John Hale. 2001. A probabilistic earley parser as a psycholinguistic model. In Proceedings of the Second Meeting of the North American Chapter of the Association for Computational Linguistics, pages 159–166, Pittsburgh, PA. John Hale. 2003. Grammar, Uncertainty and Sentence Processing. Ph.D. thesis, Cognitive Science, The Johns Hopkins University. John Hale. 2006. Uncertainty about the rest of the sentence. Cognitive Science, 30(4):609–642. Roger Levy. 2008. Expectation-based syntactic comprehension. Cognition, 106(3):1126–1177. Scott A. McDonald and Richard C. Shillcock. 2003. Low-level predictive inference in reading: The influence of transitional probabilities on eye movements. Vision Research, 43(16):1735–1751. George Miller and Noam Chomsky. 1963. Finitary models of language users. In R. Luce, R. Bush, and E. Galanter, editors, Handbook of Mathematical Psychology, volume 2, pages 419–491. John Wiley. Kevin P. Murphy and Mark A. Paskin. 2001. Linear time inference in hierarchical HMMs. In Proc. NIPS, pages 833–840, Vancouver, BC, Canada. Joakim Nivre. 2007. Inductive dependency parsing. Computational Linguistics, 33(2). Alexander Pollatsek, Erik D. Reichle, and Keith Rayner. 2006. Tests of the EZ Reader model: Exploring the interface between cognition and eyemovement control. Cognitive Psychology, 52(1):1– 56. Lawrence R. Rabiner. 1990. A tutorial on hidden Markov models and selected applications in speech recognition. Readings in speech recognition, 53(3):267–296. Brian Roark, Asaf Bachrach, Carlos Cardenas, and Christophe Pallier. 2009. Deriving lexical and syntactic expectation-based measures for psycholinguistic modeling via incremental top-down parsing. Proceedings of the 2009 Conference on Empirical Methods in Natural Langauge Processing, pages 324–333. Brian Roark. 2001. Probabilistic top-down parsing and language modeling. Computational Linguistics, 27(2):249–276. William Schuler, Samir AbdelRahman, Tim Miller, and Lane Schwartz. 2008. Toward a psycholinguistically-motivated model of language. In Proceedings of COLING, pages 785–792, Manchester, UK, August. William Schuler, Samir AbdelRahman, Tim Miller, and Lane Schwartz. 2010. Broad-coverage incremental parsing using human-like memory constraints. Computational Linguistics, 36(1). William Schuler. 2009. Parsing with a bounded stack using a model-based right-corner transform. In Proceedings of the North American Association for Computational Linguistics (NAACL ’09), pages 344–352, Boulder, Colorado. Michael K. Tanenhaus, Michael J. Spivey-Knowlton, Kathy M. Eberhard, and Julie E. Sedivy. 1995. Integration of visual and linguistic information in spoken language comprehension. Science, 268:1632– 1634. 1198
|
2010
|
121
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1199–1208, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics “Ask not what Textual Entailment can do for You...” Mark Sammons V.G.Vinod Vydiswaran Dan Roth University of Illinois at Urbana-Champaign {mssammon|vgvinodv|danr}@illinois.edu Abstract We challenge the NLP community to participate in a large-scale, distributed effort to design and build resources for developing and evaluating solutions to new and existing NLP tasks in the context of Recognizing Textual Entailment. We argue that the single global label with which RTE examples are annotated is insufficient to effectively evaluate RTE system performance; to promote research on smaller, related NLP tasks, we believe more detailed annotation and evaluation are needed, and that this effort will benefit not just RTE researchers, but the NLP community as a whole. We use insights from successful RTE systems to propose a model for identifying and annotating textual inference phenomena in textual entailment examples, and we present the results of a pilot annotation study that show this model is feasible and the results immediately useful. 1 Introduction Much of the work in the field of Natural Language Processing is founded on an assumption of semantic compositionality: that there are identifiable, separable components of an unspecified inference process that will develop as research in NLP progresses. Tasks such as Named Entity and coreference resolution, syntactic and shallow semantic parsing, and information and relation extraction have been identified as worthwhile tasks and pursued by numerous researchers. While many have (nearly) immediate application to real world tasks like search, many are also motivated by their potential contribution to more ambitious Natural Language tasks. It is clear that the components/tasks identified so far do not suffice in themselves to solve tasks requiring more complex reasoning and synthesis of information; many other tasks must be solved to achieve human-like performance on tasks such as Question Answering. But there is no clear process for identifying potential tasks (other than consensus by a sufficient number of researchers), nor for quantifying their potential contribution to existing NLP tasks, let alone to Natural Language Understanding. Recent “grand challenges” such as Learning by Reading, Learning To Read, and Machine Reading are prompting more careful thought about the way these tasks relate, and what tasks must be solved in order to understand text sufficiently well to reliably reason with it. This is an appropriate time to consider a systematic process for identifying semantic analysis tasks relevant to natural language understanding, and for assessing their potential impact on NLU system performance. Research on Recognizing Textual Entailment (RTE), largely motivated by a “grand challenge” now in its sixth year, has already begun to address some of the problems identified above. Techniques developed for RTE have now been successfully applied in the domains of Question Answering (Harabagiu and Hickl, 2006) and Machine Translation (Pado et al., 2009), (Mirkin et al., 2009). The RTE challenge examples are drawn from multiple domains, providing a relatively task-neutral setting in which to evaluate contributions of different component solutions, and RTE researchers have already made incremental progress by identifying sub-problems of entailment, and developing ad-hoc solutions for them. In this paper we challenge the NLP community to contribute to a joint, long-term effort to identify, formalize, and solve textual inference problems motivated by the Recognizing Textual Entailment setting, in the following ways: (a) Making the Recognizing Textual Entailment setting a central component of evaluation for 1199 relevant NLP tasks such as NER, Coreference, parsing, data acquisition and application, and others. While many “component” tasks are considered (almost) solved in terms of expected improvements in performance on task-specific corpora, it is not clear that this translates to strong performance in the RTE domain, due either to problems arising from unrelated, unsolved entailment phenomena that co-occur in the same examples, or to domain change effects. The RTE task offers an application-driven setting for evaluating a broad range of NLP solutions, and will reinforce good practices by NLP researchers. The RTE task has been designed specifically to exercise textual inference capabilities, in a format that would make RTE systems potentially useful components in other “deep” NLP tasks such as Question Answering and Machine Translation. 1 (b) Identifying relevant linguistic phenomena, interactions between phenomena, and their likely impact on RTE/textual inference. Determining the correct label for a single textual entailment example requires human analysts to make many smaller, localized decisions which may depend on each other. A broad, carefully conducted effort to identify and annotate such local phenomena in RTE corpora would allow their distributions in RTE examples to be quantified, and allow evaluation of NLP solutions in the context of RTE. It would also allow assessment of the potential impact of a solution to a specific sub-problem on the RTE task, and of interactions between phenomena. Such phenomena will almost certainly correspond to elements of linguistic theory; but this approach brings a data-driven approach to focus attention on those phenomena that are well-represented in the RTE corpora, and which can be identified with sufficiently close agreement. (c) Developing resources and approaches that allow more detailed assessment of RTE systems. At present, it is hard to know what specific capabilities different RTE systems have, and hence, which aspects of successful systems are worth emulating or reusing. An evaluation framework that could offer insights into the kinds of sub-problems a given system can reliably solve would make it easier to identify significant advances, and thereby promote more rapid advances 1The Parser Training and Evaluation using Textual Entailment track of SemEval 2 takes this idea one step further, by evaluating performance of an isolated NLP task using the RTE methodology. through reuse of successful solutions and focus on unresolved problems. In this paper we demonstrate that Textual Entailment systems are already “interesting”, in that they have made significant progress beyond a “smart” lexical baseline that is surprisingly hard to beat (section 2). We argue that Textual Entailment, as an application that clearly requires sophisticated textual inference to perform well, requires the solution of a range of sub-problems, some familiar and some not yet known. We therefore propose RTE as a promising and worthwhile task for large-scale community involvement, as it motivates the study of many other NLP problems in the context of general textual inference. We outline the limitations of the present model of evaluation of RTE performance, and identify kinds of evaluation that would promote understanding of the way individual components can impact Textual Entailment system performance, and allow better objective evaluation of RTE system behavior without imposing additional burdens on RTE participants. We use this to motivate a large-scale annotation effort to provide data with the mark-up sufficient to support these goals. To stimulate discussion of suitable annotation and evaluation models, we propose a candidate model, and provide results from a pilot annotation effort (section 3). This pilot study establishes the feasibility of an inference-motivated annotation effort, and its results offer a quantitative insight into the difficulty of the TE task, and the distribution of a number of entailment-relevant linguistic phenomena over a representative sample from the NIST TAC RTE 5 challenge corpus. We argue that such an evaluation and annotation effort can identify relevant subproblems whose solution will benefit not only Textual Entailment but a range of other long-standing NLP tasks, and can stimulate development of new ones. We also show how this data can be used to investigate the behavior of some of the highest-scoring RTE systems from the most recent challenge (section 4). 2 NLP Insights from Textual Entailment The task of Recognizing Textual Entailment (RTE), as formulated by (Dagan et al., 2006), requires automated systems to identify when a human reader would judge that given one span of text (the Text) and some unspecified (but restricted) world knowledge, a second span of text (the Hy1200 Text: The purchase of LexCorp by BMI for $2Bn prompted widespread sell-offs by traders as they sought to minimize exposure. Hyp 1: BMI acquired another company. Hyp 2: BMI bought LexCorp for $3.4Bn. Figure 1: Some representative RTE examples. pothesis) is true. The task was extended in (Giampiccolo et al., 2007) to include the additional requirement that systems identify when the Hypothesis contradicts the Text. In the example shown in figure 1, this means recognizing that the Text entails Hypothesis 1, while Hypothesis 2 contradicts the Text. This operational definition of Textual Entailment avoids commitment to any specific knowledge representation, inference method, or learning approach, thus encouraging application of a wide range of techniques to the problem. 2.1 An Illustrative Example The simple RTE examples in figure 1 (most RTE examples have much longer Texts) illustrate some typical inference capabilities demonstrated by human readers in determining whether one span of text contains the meaning of another. To recognize that Hypothesis 1 is entailed by the text, a human reader must recognize that “another company” in the Hypothesis can match “LexCorp”. She must also identify the nominalized relation “purchase”, and determine that “A purchased by B” implies “B acquires A”. To recognize that Hypothesis 2 contradicts the Text, similar steps are required, together with the inference that because the stated purchase price is different in the Text and Hypothesis, but with high probability refers to the same transaction, Hypothesis 2 contradicts the Text. It could be argued that this particular example might be resolved by simple lexical matching; but it should be evident that the Text can be made lexically very dissimilar to Hypothesis 1 while maintaining the Entailment relation, and that conversely, the lexical overlap between the Text and Hypothesis 2 can be made very high, while maintaining the Contradiction relation. This intuition is borne out by the results of the RTE challenges, which show that lexical similarity-based systems are outperformed by systems that use other, more structured analysis, as shown in the next section. Rank System id Accuracy 1 I 0.735 2 E 0.685 3 H 0.670 4 J 0.667 5 G 0.662 6 B 0.638 7 D 0.633 8 F 0.632 9 A 0.615 9 C 0.615 9 K 0.615 Lex 0.612 Table 1: Top performing systems in the RTE 5 2way task. Lex E G H I J Lex 1.000 0.667 0.693 0.678 0.660 0.778 (184,183) (157,132) (168,122) (152,136) (165,137) (165,135) E 1.000 0.667 0.675 0.673 0.702 (224,187) (192,112) (178,131) (201,127) (186,131) G 1.000 0.688 0.713 0.745 (247,150) (186,120) (218,115) (198,125) H 1.000 0.705 0.707 (219,183) (194,139) (178,136) I 1.000 0.705 (260,181) (198,135) J 1.000 (224,178) Table 2: In each cell, top row shows observed agreement and bottom row shows the number of correct (positive, negative) examples on which the pair of systems agree. 2.2 The State of the Art in RTE 5 The outputs for all systems that participated in the RTE 5 challenge were made available to participants. We compared these to each other and to a smart lexical baseline (Do et al., 2010) (lexical match augmented with a WordNet similarity measure, stemming, and a large set of low-semanticcontent stopwords) to assess the diversity of the approaches of different research groups. To get the fullest range of participants, we used results from the two-way RTE task. We have anonymized the system names. Table 1 shows that many participating systems significantly outperform our smart lexical baseline. Table 2 reports the observed agreement between systems and the lexical baseline in terms of the percentage of examples on which a pair of systems gave the same label. The agreement between most systems and the baseline is about 67%, which suggests that systems are not simply augmented versions of the lexical baseline, and are also distinct from each other in their behaviors.2 Common characteristics of RTE systems re2Note that the expected agreement between two random RTE decision-makers is 0.5, so the agreement scores according to Cohen’s Kappa measure (Cohen, 1960) are between 0.3 and 0.4. 1201 ported by their designers were the use of structured representations of shallow semantic content (such as augmented dependency parse trees and semantic role labels); the application of NLP resources such as Named Entity recognizers, syntactic and dependency parsers, and coreference resolvers; and the use of special-purpose ad-hoc modules designed to address specific entailment phenomena the researchers had identified, such as the need for numeric reasoning. However, it is not possible to objectively assess the role these capabilities play in each system’s performance from the system outputs alone. 2.3 The Need for Detailed Evaluation An ablation study that formed part of the official RTE 5 evaluation attempted to evaluate the contribution of publicly available knowledge resources such as WordNet (Fellbaum, 1998), VerbOcean (Chklovski and Pantel, 2004), and DIRT (Lin and Pantel, 2001) used by many of the systems. The observed contribution was in most cases limited or non-existent. It is premature, however, to conclude that these resources have little potential impact on RTE system performance: most RTE researchers agree that the real contribution of individual resources is difficult to assess. As the example in figure 1 illustrates, most RTE examples require a number of phenomena to be correctly resolved in order to reliably determine the correct label (the Interaction problem); a perfect coreference resolver might as a result yield little improvement on the standard RTE evaluation, even though coreference resolution is clearly required by human readers in a significant percentage of RTE examples. Various efforts have been made by individual research teams to address specific capabilities that are intuitively required for good RTE performance, such as (de Marneffe et al., 2008), and the formal treatment of entailment phenomena in (MacCartney and Manning, 2009) depends on and formalizes a divide-and-conquer approach to entailment resolution. But the phenomena-specific capabilities described in these approaches are far from complete, and many are not yet invented. To devote real effort to identify and develop such capabilities, researchers must be confident that the resources (and the will!) exist to create and evaluate their solutions, and that the resource can be shown to be relevant to a sufficiently large subset of the NLP community. While there is widespread belief that there are many relevant entailment phenomena, though each individually may be relevant to relatively few RTE examples (the Sparseness problem), we know of no systematic analysis to determine what those phenomena are, and how sparsely represented they are in existing RTE data. If it were even known what phenomena were relevant to specific entailment examples, it might be possible to more accurately distinguish system capabilities, and promote adoption of successful solutions to sub-problems. An annotation-side solution also maintains the desirable agnosticism of the RTE problem formulation, by not imposing the requirement on system developers of generating an explanation for each answer. Of course, if examples were also annotated with explanations in a consistent format, this could form the basis of a new evaluation of the kind essayed in the pilot study in (Giampiccolo et al., 2007). 3 Annotation Proposal and Pilot Study As part of our challenge to the NLP community, we propose a distributed OntoNotes-style approach (Hovy et al., 2006) to this annotation effort: distributed, because it should be undertaken by a diverse range of researchers with interests in different semantic phenomena; and similar to the OntoNotes annotation effort because it should not presuppose a fixed, closed ontology of entailment phenomena, but rather, iteratively hypothesize and refine such an ontology using interannotator agreement as a guiding principle. Such an effort would require a steady output of RTE examples to form the underpinning of these annotations; and in order to get sufficient data to represent less common, but nonetheless important, phenomena, a large body of data is ultimately needed. A research team interested in annotating a new phenomenon should use examples drawn from the common corpus. Aside from any task-specific gold standard annotation they add to the entailment pairs, they should augment existing explanations by indicating in which examples their phenomenon occurs, and at which point in the existing explanation for each example. In fact, this latter effort – identifying phenomena relevant to textual inference, marking relevant RTE examples, and generating explanations – itself enables other researchers to select from known problems, assess their likely impact, and automatically generate rel1202 evant corpora. To assess the feasibility of annotating RTEoriented local entailment phenomena, we developed an inference model that could be followed by annotators, and conducted a pilot annotation study. We based our initial effort on observations about RTE data we made while participating in RTE challenges, together with intuitive conceptions of the kinds of knowledge that might be available in semi-structured or structured form. In this section, we present our annotation inference model, and the results of our pilot annotation effort. 3.1 Inference Process To identify and annotate RTE sub-phenomena in RTE examples, we need a defensible model for the entailment process that will lead to consistent annotation by different researchers, and to an extensible framework that can accommodate new phenomena as they are identified. We modeled the entailment process as one of manipulating the text and hypothesis to be as similar as possible, by first identifying parts of the text that matched parts of the hypothesis, and then identifying connecting structure. Our inherent assumption was that the meanings of the Text and Hypothesis could be represented as sets of n-ary relations, where relations could be connected to other relations (i.e., could take other relations as arguments). As we followed this procedure for a given example, we marked which entailment phenomena were required for the inference. We illustrate the process using the example in figure 1. First, we would identify the arguments “BMI” and “another company” in the Hypothesis as matching “BMI” and “LexCorp” respectively, requiring 1) Parent-Sibling to recognize that “LexCorp” can match “company”. We would tag the example as requiring 2) Nominalization Resolution to make “purchase” the active relation and 3) Passivization to move “BMI” to the subject position. We would then tag it with 4) Simple Verb Rule to map “A purchase B” to “A acquire B”. These operations make the relevant portion of the Text identical to the Hypothesis, so we are done. For the same Text, but with Hypothesis 2 (a negative example), we follow the same steps 1-3. We would then use 4) Lexical Relation to map “purchase” to “buy”. We would then observe that the only possible match for the hypothesis argument “for $3.4Bn” is the text argument “for $2Bn”. We would label this as a 5) Numerical Quantity Mismatch and 6) Excluding Argument (it can’t be the case that in the same transaction, the same company was sold for two different prices). Note that neither explanation mentions the anaphora resolution connecting “they” to “traders”, because it is not strictly required to determine the entailment label. As our example illustrates, this process makes sense for both positive and negative examples. It also reflects common approaches in RTE systems, many of which have explicit alignment components that map parts of the Hypothesis to parts of the Text prior to a final decision stage. 3.2 Annotation Labels We sought to identify roles for background knowledge in terms of domains and general inference steps, and the types of linguistic phenomena that are involved in representing the same information in different ways, or in detecting key differences in two similar spans of text that indicate a difference in meaning. We annotated examples with domains (such as “Work”) for two reasons: to establish whether some phenomena are correlated with particular domains; and to identify domains that are sufficiently well-represented that a knowledge engineering study might be possible. While we did not generate an explicit representation of our entailment process, i.e. explanations, we tracked which phenomena were strictly required for inference. The annotated corpora and simple CGI scripts for annotation are available at http://cogcomp.cs.illinois.edu/Data/ACL2010 RTE.php. The phenomena that we considered during annotation are presented in Tables 3, 4, 5, and 6. We tried to define each phenomenon so that it would apply to both positive and negative examples, but ran into a problem: often, negative examples can be identified principally by structural differences: the components of the Hypothesis all match components in the Text, but they are not connected by the appropriate structure in the Text. In the case of contradictions, it is often the case that a key relation in the Hypothesis must be matched to an incompatible relation in the Text. We selected names for these structural behaviors, and tagged them when we observed them, but the counterpart for positive examples must always hold: it must necessarily be the case that the structure in the Text linking the arguments that match those in the 1203 Hypothesis must be comparable to the Hypothesis structure. We therefore did not tag this for positive examples. We selected a subset of 210 examples from the NIST TAC RTE 5 (Bentivogli et al., 2009) Test set drawn equally from the three sub-tasks (IE, IR and QA). Each example was tagged by both annotators. Two passes were made over the data: the first covered 50 examples from each RTE sub-task, while the second covered an additional 20 examples from each sub-task. Between the two passes, concepts the annotators identified as difficult to annotate were discussed and more carefully specified, and several new concepts were introduced based on annotator observations. Tables 3, 4, 5, and 6 present information about the distribution of the phenomena we tagged, and the inter-annotator agreement (Cohen’s Kappa (Cohen, 1960)) for each. “Occurrence” lists the average percentage of examples labeled with a phenomenon by the two annotators. Domain Occurrence Agreement work 16.90% 0.918 name 12.38% 0.833 die kill injure 12.14% 0.979 group 9.52% 0.794 be in 8.57% 0.888 kinship 7.14% 1.000 create 6.19% 1.000 cause 6.19% 0.854 come from 5.48% 0.879 win compete 3.10% 0.813 Others 29.52% 0.864 Table 3: Occurrence statistics for domains in the annotated data. Phenomenon Occurrence Agreement Named Entity 91.67% 0.856 locative 17.62% 0.623 Numerical Quantity 14.05% 0.905 temporal 5.48% 0.960 nominalization 4.05% 0.245 implicit relation 1.90% 0.651 Table 4: Occurrence statistics for hypothesis structure features. From the tables it is apparent that good performance on a range of phenomena in our inference model are likely to have a significant effect on RTE results, with coreference being deemed essential to the inference process for 35% of examples, and a number of other phenomena are sufficiently well represented to merit near-future attention (assuming that RTE systems do not already handle these phenomena, a question we address in section 4). It is also clear from the predominance of Simple Rewrite Rule instances, together with Phenomenon Occurrence Agreement coreference 35.00% 0.698 simple rewrite rule 32.62% 0.580 lexical relation 25.00% 0.738 implicit relation 23.33% 0.633 factoid 15.00% 0.412 parent-sibling 11.67% 0.500 genetive relation 9.29% 0.608 nominalization 8.33% 0.514 event chain 6.67% 0.589 coerced relation 6.43% 0.540 passive-active 5.24% 0.583 numeric reasoning 4.05% 0.847 spatial reasoning 3.57% 0.720 Table 5: Occurrence statistics for entailment phenomena and knowledge resources Phenomenon Occurrence Agreement missing argument 16.19% 0.763 missing relation 14.76% 0.708 excluding argument 10.48% 0.952 Named Entity mismatch 9.29% 0.921 excluding relation 5.00% 0.870 disconnected relation 4.52% 0.580 missing modifier 3.81% 0.465 disconnected argument 3.33% 0.764 Numeric Quant. mismatch 3.33% 0.882 Table 6: Occurrences of negative-only phenomena the frequency of most of the domains we selected, that knowledge engineering efforts also have a key role in improving RTE performance. 3.3 Discussion Perhaps surprisingly, given the difficulty of the task, inter-annotator agreement was consistently good to excellent (above 0.6 and 0.8, respectively), with few exceptions, indicating that for most targeted phenomena, the concepts were wellspecified. The results confirmed our initial intuition about some phenomena: for example, that coreference resolution is central to RTE, and that detecting the connecting structure is crucial in discerning negative from positive examples. We also found strong evidence that the difference between contradiction and unknown entailment examples is often due to the behavior of certain relations that either preclude certain other relations holding between the same arguments (for example, winning a contest vs. losing a contest), or which can only hold for a single referent in one argument position (for example, “work” relations such as job title are typically constrained so that a single person holds one position). We found that for some examples, there was more than one way to infer the hypothesis from the text. Typically, for positive examples this involved overlap between phenomena; for example, Coreference might be expected to resolve implicit rela1204 tions induced from appositive structures. In such cases we annotated every way we could find. In future efforts, annotators should record the entailment steps they used to reach their decision. This will make disagreement resolution simpler, and could also form a possible basis for generating gold standard explanations. At a minimum, each inference step must identify the spans of the Text and Hypothesis that are involved and the name of the entailment phenomenon represented; in addition, a partial order over steps must be specified when one inference step requires that another has been completed. Future annotation efforts should also add a category “Other”, to indicate for each example whether the annotator considers the listed entailment phenomena sufficient to identify the label. It might also be useful to assess the difficulty of each example based on the time required by the annotator to determine an explanation, for comparison with RTE system errors. These, together with specifications that minimize the likely disagreements between different groups of annotators, are processes that must be refined as part of the broad community effort we seek to stimulate. 4 Pilot RTE System Analysis In this section, we sketch out ways in which the proposed analysis can be applied to learn something about RTE system behavior, even when those systems do not provide anything beyond the output label. We present the analysis in terms of sample questions we hope to answer with such an analysis. 1. If a system needs to improve its performance, which features should it concentrate on? To answer this question, we looked at the top-5 systems and tried to find which phenomena are active in the mistakes they make. (a) Most systems seem to fail on examples that need numeric reasoning to get the entailment decision right. For example, system H got all 10 examples with numeric reasoning wrong. (b) All top-5 systems make consistent errors in cases where identifying a mismatch in named entities (NE) or numerical quantities (NQ) is important to make the right decision. System G got 69% of cases with NE/NQ mismatches wrong. (c) Most systems make errors in examples that have a disconnected or exclusion component (argument/relation). System J got 81% of cases with a disconnected component wrong. (d) Some phenomena are handled well by certain systems, but not by others. For example, failing to recognize a parent-sibling relation between entities/concepts seems to be one of the top-5 phenomena active in systems E and H. System H also fails to correctly label over 53% of the examples having kinship relation. 2. Which phenomena have strong correlations to the entailment labels among hard examples? We called an example hard if at least 4 of the top 5 systems got the example wrong. In our annotation dataset, there were 41 hard examples. Some of the phenomena that strongly correlate with the TE labels on hard examples are: deeper lexical relation between words (ρ = 0.542), and need for external knowledge (ρ = 0.345). Further, we find that the top-5 systems tend to make mistakes in cases where the lexical approach also makes mistakes (ρ = 0.355). 3. What more can be said about individual systems? In order to better understand the system behavior, we wanted to check if we could predict the system behavior based on the phenomena we identified as important in the examples. We learned SVM classifiers over the identified phenomena and the lexical similarity score to predict both the labels and errors systems make for each of the top-5 systems. We could predict all 10 system behaviors with over 70% accuracy, and could predict labels and mistakes made by two of the top-5 systems with over 77% accuracy. This indicates that although the identified phenomena are indicative of the system performance, it is probably too simplistic to assume that system behavior can be easily reproduced solely as a disjunction of phenomena present in the examples. 4. Does identifying the phenomena correctly help learn a better TE system? We tried to learn an entailment classifier over the phenomenon identified and the top 5 system outputs. The results are summarized in Table 7. All reported numbers are 20-fold cross-validation accuracy from an SVM classifier learned over the features mentioned. The results show that correctly identifying the named-entity and numeric quantity mis1205 No. Feature description No. of Accuracy over which features feats phenomena pheno. + sys. labels (0) Only system labels 5 — 0.714 (1) Domain and hypothesis features (Tables 3, 4) 16 0.510 0.705 (2) (1) + NE + NQ 18 0.619 0.762 (3) (1) + Knowledge resources (subset of Table 5) 22 0.662 0.762 (4) (3) + NE + NQ 24 0.738 0.805 (5) (1) + Entailment and Knowledge resources (Table 5) 29 0.748 0.791 (6) (5) + negative-only phenomena (Table 6) 38 0.971 0.943 Table 7: Accuracy in predicting the label based on the phenomena and top-5 system labels. matches improves the overall accuracy significantly. If we further recognize the need for knowledge resources correctly, we can correctly explain the label for 80% of the examples. Adding the entailment and negation features helps us explain the label for 97% of the examples in the annotated corpus. It must be clarified that the results do not show the textual entailment problem itself is solved with 97% accuracy. However, we believe that if a system could recognize key negation phenomena such as Named Entity mismatch, presence of Excluding arguments, etc. correctly and consistently, it could model them as a Contradiction features in the final inference process to significantly improve its overall accuracy. Similarly, identifying and resolving the key entailment phenomena in the examples, would boost the inference process in positive examples. However, significant effort is still required to obtain near-accurate knowledge and linguistic resources. 5 Discussion NLP researchers in the broader community continually seek new problems to solve, and pose more ambitious tasks to develop NLP and NLU capabilities, yet recognize that even solutions to problems which are considered “solved” may not perform as well on domains different from the resources used to train and develop them. Solutions to such NLP tasks could benefit from evaluation and further development on corpora drawn from a range of domains, like those used in RTE evaluations. It is also worthwhile to consider each task as part of a larger inference process, and therefore motivated not just by performance statistics on special-purpose corpora, but as part of an interconnected web of resources; and the task of Recognizing Textual Entailment has been designed to exercise a wide range of linguistic and reasoning capabilities. The entailment setting introduces a potentially broader context to resource development and assessment, as the hypothesis and text provide context for each other in a way different than local context from, say, the same paragraph in a document: in RTE’s positive examples, the Hypothesis either restates some part of the Text, or makes statements inferable from the statements in the Text. This is not generally true of neighboring sentences in a document. This distinction opens the door to “purposeful”, or goal-directed, inference in a way that may not be relevant to a task studied in isolation. The RTE community seems mainly convinced that incremental advances in local entailment phenomena (including application of world knowledge) are needed to make significant progress. They need ways to identify sub-problems of textual inference, and to evaluate those solutions both in isolation and in the context of RTE. RTE system developers are likely to reward well-engineered solutions by adopting them and citing their authors, because such solutions are easier to incorporate into RTE systems. They are also more likely to adopt solutions with established performance levels. These characteristics promote publication of software developed to solve NLP tasks, attention to its usability, and publication of materials supporting reproduction of results presented in technical papers. For these reasons, we assert that RTE is a natural motivator of new NLP tasks, as researchers look for components capable of improving performance; and that RTE is a natural setting for evaluating solutions to a broad range of NLP problems, though not in its present formulation: we must solve the problem of credit assignment, to recognize component contributions. We have therefore proposed a suitable annotation effort, to provide the resources necessary for more detailed evaluation of RTE systems. We have presented a linguistically-motivated 1206 analysis of entailment data based on a step-wise procedure to resolve entailment decisions, intended to allow independent annotators to reach consistent decisions, and conducted a pilot annotation effort to assess the feasibility of such a task. We do not claim that our set of domains or phenomena are complete: for example, our illustrative example could be tagged with a domain Mergers and Acquisitions, and a different team of researchers might consider Nominalization Resolution to be a subset of Simple Verb Rules. This kind of disagreement in coverage is inevitable, but we believe that in many cases it suffices to introduce a new domain or phenomenon, and indicate its relation (if any) to existing domains or phenomena. In the case of introducing a non-overlapping category, no additional information is needed. In other cases, the annotators can simply indicate the phenomena being merged or split (or even replaced). This information will allow other researchers to integrate different annotation sources and maintain a consistent set of annotations. 6 Conclusions In this paper, we have presented a case for a broad, long-term effort by the NLP community to coordinate annotation efforts around RTE corpora, and to evaluate solutions to NLP tasks relating to textual inference in the context of RTE. We have identified limitations in the existing RTE evaluation scheme, proposed a more detailed evaluation to address these limitations, and sketched a process for generating this annotation. We have proposed an initial annotation scheme to prompt discussion, and through a pilot study, demonstrated that such annotation is both feasible and useful. We ask that researchers not only contribute task specific annotation to the general pool, and indicate how their task relates to those already added to the annotated RTE corpora, but also invest the additional effort required to augment the cross-domain annotation: marking the examples in which their phenomenon occurs, and augmenting the annotator-generated explanations with the relevant inference steps. These efforts will allow a more meaningful evaluation of RTE systems, and of the component NLP technologies they depend on. We see the potential for great synergy between different NLP subfields, and believe that all parties stand to gain from this collaborative effort. We therefore respectfully suggest that you “ask not what RTE can do for you, but what you can do for RTE...” Acknowledgments We thank the anonymous reviewers for their helpful comments and suggestions. This research was partly sponsored by Air Force Research Laboratory (AFRL) under prime contract no. FA875009-C-0181, by a grant from Boeing and by MIAS, the Multimodal Information Access and Synthesis center at UIUC, part of CCICADA, a DHS Center of Excellence. Any opinions, findings, and conclusion or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the view of the sponsors. References Luisa Bentivogli, Ido Dagan, Hoa Trang Dang, Danilo Giampiccolo, and Bernando Magnini. 2009. The fifth pascal recognizing textual entailment challenge. In Notebook papers and Results, Text Analysis Conference (TAC), pages 14–24. Timothy Chklovski and Patrick Pantel. 2004. VerbOcean: Mining the Web for Fine-Grained Semantic Verb Relations. In Proceedings of Conference on Empirical Methods in Natural Language Processing (EMNLP-04), pages 33–40. Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1):37–46. I. Dagan, O. Glickman, and B. Magnini, editors. 2006. The PASCAL Recognising Textual Entailment Challenge., volume 3944. Springer-Verlag, Berlin. Marie-Catherine de Marneffe, Anna N. Rafferty, and Christopher D. Manning. 2008. Finding contradictions in text. In Proceedings of ACL-08: HLT, pages 1039–1047, Columbus, Ohio, June. Association for Computational Linguistics. Quang Do, Dan Roth, Mark Sammons, Yuancheng Tu, and V.G.Vinod Vydiswaran. 2010. Robust, Light-weight Approaches to compute Lexical Similarity. Computer Science Research and Technical Reports, University of Illinois. http://L2R.cs.uiuc.edu/∼danr/Papers/DRSTV10.pdf. C. Fellbaum. 1998. WordNet: An Electronic Lexical Database. MIT Press. Danilo Giampiccolo, Bernardo Magnini, Ido Dagan, and Bill Dolan. 2007. The third pascal recognizing textual entailment challenge. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing, pages 1–9, Prague, June. Association for Computational Linguistics. 1207 Sanda Harabagiu and Andrew Hickl. 2006. Methods for Using Textual Entailment in Open-Domain Question Answering. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 905–912, Sydney, Australia, July. Association for Computational Linguistics. Eduard Hovy, Mitchell Marcus, Martha Palmer, Lance Ramshaw, and Ralph Weischedel. 2006. Ontonotes: The 90% solution. In Proceedings of HLT/NAACL, New York. D. Lin and P. Pantel. 2001. DIRT: discovery of inference rules from text. In Proc. of ACM SIGKDD Conference on Knowledge Discovery and Data Mining 2001, pages 323–328. Bill MacCartney and Christopher D. Manning. 2009. An extended model of natural logic. In The Eighth International Conference on Computational Semantics (IWCS-8), Tilburg, Netherlands. Shachar Mirkin, Lucia Specia, Nicola Cancedda, Ido Dagan, Marc Dymetman, and Idan Szpektor. 2009. Source-language entailment modeling for translating unknown terms. In ACL/AFNLP, pages 791– 799, Suntec, Singapore, August. Association for Computational Linguistics. Sebastian Pado, Michel Galley, Dan Jurafsky, and Christopher D. Manning. 2009. Robust machine translation evaluation with entailment features. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 297–305, Suntec, Singapore, August. Association for Computational Linguistics. 1208
|
2010
|
122
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1209–1219, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Assessing the Role of Discourse References in Entailment Inference Shachar Mirkin, Ido Dagan Bar-Ilan University Ramat-Gan, Israel {mirkins,dagan}@cs.biu.ac.il Sebastian Pad´o University of Stuttgart Stuttgart, Germany [email protected] Abstract Discourse references, notably coreference and bridging, play an important role in many text understanding applications, but their impact on textual entailment is yet to be systematically understood. On the basis of an in-depth analysis of entailment instances, we argue that discourse references have the potential of substantially improving textual entailment recognition, and identify a number of research directions towards this goal. 1 Introduction The detection and resolution of discourse references such as coreference and bridging anaphora play an important role in text understanding applications, like question answering and information extraction. There, reference resolution is used for the purpose of combining knowledge from multiple sentences. Such knowledge is also important for Textual Entailment (TE), a generic framework for modeling semantic inference. TE reduces the inference requirements of many text understanding applications to the problem of determining whether the meaning of a given textual assertion, termed hypothesis (H), can be inferred from the meaning of certain text (T) (Dagan et al., 2006). Consider the following example: (1) T: “Not only had he developed an aversion to the President1 and politics in general, Oswald2 was also a failure with Marina, his wife. [...] Their relationship was supposedly responsible for why he2 killed Kennedy1.” H: “Oswald killed President Kennedy.” The understanding that the second sentence of the text entails the hypothesis draws on two coreference relationships, namely that he is Oswald, and that the Kennedy in question is President Kennedy. However, the utilization of discourse information for such inferences has been so far limited mainly to the substitution of nominal coreferents, while many aspects of the interface between discourse and semantic inference needs remain unexplored. The recently held Fifth Recognizing Textual Entailment (RTE-5) challenge (Bentivogli et al., 2009a) has introduced a Search task, where the text sentences are interpreted in the context of their full discourse, as in Example 1 above. Accordingly, TE constitutes an interesting framework – and the Search task an adequate dataset – to study the interrelation between discourse and inference. The goal of this study is to analyze the roles of discourse references for textual entailment inference, to provide relevant findings and insights to developers of both reference resolvers and entailment systems and to highlight promising directions for the better incorporation of discourse phenomena into inference. Our focus is on a manual, in-depth assessment that results in a classification and quantification of discourse reference phenomena and their utilization for inference. On this basis, we develop an account of formal devices for incorporating discourse references into the inference computation. An additional point of interest is the interrelation between entailment knowledge and coreference. E.g., in Example 1 above, knowing that Kennedy was a president can alleviate the need for coreference resolution. Conversely, coreference resolution can often be used to overcome gaps in entailment knowledge. Structure of the paper. In Section 2, we provide background on the use of discourse references in natural language processing (NLP) in general and specifically in TE. Section 3 describes the goals of this study, followed by our analysis scheme (Section 4) and the required inference 1209 mechanisms (Section 5). Section 6 presents quantitative findings and further observations. Conclusions are discussed in Section 7. 2 Background 2.1 Discourse in NLP Discourse information plays a role in a range of NLP tasks. It is obviously central to discourse processing tasks such as text segmentation (Hearst, 1997). Reference information provided by discourse is also useful for text understanding tasks such as question answering (QA), information extraction (IE) and information retrieval (IR) (Vicedo and Ferrndez, 2006; Zelenko et al., 2004; Na and Ng, 2009), as well as for the acquisition of lexical-semantic “narrative schema” knowledge (Chambers and Jurafsky, 2009). Discourse references have been the subject of attention in both the Message Understanding Conference (Grishman and Sundheim, 1996) and the Automatic Content Extraction program (Strassel et al., 2008). The simplest form of information that discourse provides is coreference, i.e., information that two linguistic expressions refer to the same entity or event. Coreference is particularly important for processing pronouns and other anaphoric expressions, such as he in Example 1. Ability to resolve this reference translates directly into, e.g., a QA system’s ability to answer questions like Who killed Kennedy?. A second, more complex type of information stems from bridging references, such as in the following discourse (Asher and Lascarides, 1998): (2) “I’ve just arrived. The camel is outside.” While coreference indicates equivalence, bridging points to the existence of a salient semantic relation between two distinct entities or events. Here, it is (informally) ‘means of transport’, which would make the discourse (2) relevant for a question like How did I arrive here?. Other types of bridging relations include set-membership, roles in events and consequence (Clark, 1975). Note, however, that text understanding systems are generally limited to the resolution of entity (or even just pronoun) coreference, e.g. (Li et al., 2009; Dali et al., 2009). An important reason is the unavailability of tools to resolve the more complex (and difficult) forms of discourse reference such as event coreference and bridging.1 Another reason is uncertainty about their practical importance. 2.2 Discourse in Textual Entailment Textual Entailment has been introduced in Section 1 as a common-sense notion of inference. It has spawned interest in the computational linguistics community as a common denominator of many NLP tasks including IE, summarization and tutoring (Romano et al., 2006; Harabagiu et al., 2007; Nielsen et al., 2009). Architectures for Textual Entailment. Over the course of recent RTE challenges (Giampiccolo et al., 2007; Giampiccolo et al., 2008), the main benchmark for TE technology, two architectures for modeling TE have emerged as dominant: transformations and alignment. The goal of transformation-based TE models is to determine the entailment relation T ⇒H by finding a “proof”, i.e., a sequence of consequents, (T, T1, . . . , Tn), such that Tn=H (Bar-Haim et al., 2008; Harmeling, 2009), and that in each transformation, Ti →Ti+1, the consequent Ti+1 is entailed by Ti. These transformations commonly include lexical modifications and the generation of syntactic alternatives. The second major approach constructs an alignment between the linguistic entities of the trees (or graphs) of T and H, which can represent syntactic structure, semantic structure, or non-hierarchical phrases (Zanzotto et al., 2009; Burchardt et al., 2009; MacCartney et al., 2008). H is assumed to be entailed by T if its entities are aligned “well” to corresponding entities in T. Alignment quality is generally determined based on features that assess the validity of the local replacement of the T entity by the H entity. While transformation- and alignment-based entailment models look different at first glance, they ultimately have the same goal, namely obtaining a maximal coverage of H by T, i.e. to identify matches of as many elements of H within T as possible.2 To do so, both architectures typically make use of inference rules such as ‘Y was purchased by X →X paid for Y’, either by directly applying them as transformations, or by using them 1Some studies, e.g. (Markert et al., 2003; Poesio et al., 2004), address the resolution of a few specific kinds of bridging relations; yet, wide-scope systems for bridging resolution are unavailable. 2Clearly, the details of how the final entailment decision is made based on the attained coverage differ substantially among models. 1210 to score alignments. Rules are generally drawn from external knowledge resources, such as WordNet (Fellbaum, 1998) or DIRT (Lin and Pantel, 2001), although knowledge gaps remain a key obstacle (Bos, 2005; Balahur et al., 2008; Bar-Haim et al., 2008). Discourse in previous RTE challenges. The first two rounds of the RTE challenge used “selfcontained” texts and hypotheses, where discourse considerations played virtually no role. A first step towards a more comprehensive notion of entailment was taken with RTE-3 (Giampiccolo et al., 2007), when paragraph-length texts were first included and constituted 17% of the texts in the test set. Chambers et al. (2007) report that in a sample of T −H pairs drawn from the development set, 25% involved discourse references. Using the concepts introduced above, the impact of discourse references can be generally described as a coverage problem, independent of the system’s architecture. In Example 1, the hypothesis word Oswald cannot be safely linked to the text pronoun he without further knowledge about he; the same is true for ‘Kennedy →President Kennedy’ which involves a specialization that is only warranted in the specific discourse. A number of systems have tried to address the question of coreference in RTE as a preprocessing step prior to inference proper, with most systems using off-the-shelf coreference resolvers such as JavaRap (Qiu et al., 2004) or OpenNLP3. Generally, anaphoric expressions were textually replaced by their antecedents. Results were inconclusive, however, with several reports about errors introduced by automatic coreference resolution (Agichtein et al., 2008; Adams et al., 2007). Specific evaluations of the contribution of coreference resolution yielded both small negative (Bar-Haim et al., 2008) and insignificant positive (Chambers et al., 2007) results. 3 Motivation and Goals The results of recent studies, as reported in Section 2.2, seem to show that current resolution of discourse references in RTE systems hardly affects performance. However, our intuition is that these results can be attributed to four major limitations shared by these studies: (1) the datasets, where discourse phenomena were not well repre3http://opennlp.sourceforge.net sented; (2) the off-the-shelf coreference resolution systems which may have been not robust enough; (3) the limitation to nominal coreference; and (4) overly simple integration of reference information into the inference engines. The goal of this paper is to assess the impact of discourse references on entailment with an annotation study which removes these limitations. To counteract (1), we use the recent RTE-5 Search dataset (details below). To avoid (2), we perform a manual analysis, assuming discourse references as predicted by an oracle. With regards to (3), our annotation scheme covers coreference and bridging relations of all syntactic categories and classifies them. As for (4), we suggest several operations necessary to integrate the discourse information into an entailment engine. In contrast to the numerous existing datasets annotated for discourse references (Hovy et al., 2006; Strassel et al., 2008), we do not annotate exhaustively. Rather, we are interested specifically in those references instances that impact inference. Furthermore, we analyze each instance from an entailment perspective, characterizing the relevant factors that have an impact on inference. To our knowledge, this is the first such in-depth study.4 The results of our study are of twofold interest. First, they provide guidance for the developers of reference resolvers who might prioritize the scope of their systems to make them more valuable for inference. Second, they point out potential directions for the developers of inference systems by specifying what additional inference mechanisms are needed to utilize discourse information. The RTE-5 Search dataset. We base our annotation on the Search task dataset, a new addition to the recent Fifth RTE challenge (Bentivogli et al., 2009a) that is motivated by the needs of NLP applications and drawn from the TAC summarization track. In the Search task, TE systems are required to find all individual sentences in a given corpus which entail the hypothesis – a setting that is sensible not only for summarization, but also for information access tasks like QA. Sentences are judged individually, but “are to be interpreted in the context of the corpus as they rely on explicit and implicit references to entities, events, dates, places, etc., mentioned elsewhere in the corpus” (Bentivogli et al., 2009b). 4The guidelines and the dataset are available at http://www.cs.biu.ac.il/˜nlp/downloads/ 1211 Text Hypothesis i T ′ Once the reform becomes law, Spain will join the Netherlands and Belgium in allowing homosexual marriages. Massachusetts allows homosexual T Such unions are also legal in six Canadian provinces and the northeastern US state of Massachusetts. marriages T ′ The official name of 2003 UB313 has yet to be determined. ii T Brown said he expected to find a moon orbiting Xena because many Kuiper Belt objects are paired with moons. 2003 UB313 is in the Kuiper Belt iii T ′ a All seven aboard the AS-28 submarine appeared to be in satisfactory condition, naval spokesman said. T ′ b British crews were working with Russian naval authorities to maneuver the unmanned robotic vehicle and untangle the AS-28. The AS-28 mini submarine was trapped underwater T The Russian military was racing against time early Friday to rescue a mini submarine trapped on the seabed. iv T ′ China seeks solutions to its coal mine safety. A mining accident in China has killed several miners T A recent accident has cost more than a dozen miners their lives. v T ′′ A remote-controlled device was lowered to the stricken vessel to cut the cables in which the AS-28 vehicle is caught. T ′ The mini submarine was resting on the seabed at a depth of about 200 meters. The AS-28 mini submarine was trapped underwater T Specialists said it could have become tangled up with a metal cable or in sunken nets from a fishing trawler. vi T . .. dried up lakes in Siberia, because the permafrost beneath them has begun to thaw. The ice is melting in the Arctic Table 1: Examples for discourse-dependent entailment in the RTE-5 dataset, where the inference of H depends on reference information from the discourse sentences T ′ / T ′′. Referring terms (in T) and target terms (in H) are shown in boldface. 4 Analysis Scheme For annotating the RTE-5 data, we operationalize reference relations that are relevant for entailment as those that improve coverage. Recall from Section 2.2 that the concept of coverage is applicable to both transformation and alignment models, all of which aim at maximizing coverage of H by T. We represent T and H as syntactic trees, as common in the RTE literature (Zanzotto et al., 2009; Agichtein et al., 2008). Specifically, we assume MINIPAR-style (Lin, 1993) dependency trees where nodes represent text expressions and edges represent the syntactic relations between them. We use “term” to refer to text expressions, and “components” to refer to nodes, edges, and subtrees. Dependency trees are a popular choice in RTE since they offer a fairly semantics-oriented account of the sentence structure that can still be constructed robustly. In an ideal case of entailment, all nodes and dependency edges of H are covered by T. For each T −H pair, we annotate all relevant discourse references in terms of three items: the target component in H, the focus term in T, and the reference term which stands in a reference relation to the focus term. By resolving this reference, the target component can usually be inferred; sometimes, however, more than one reference term needs to be found. We now define and illustrate these concepts on examples from Table 1.5 The target component is a tree component in H that cannot be covered by the “local” material from T. An example for a tree component is Example (v), where the target component AS-28 mini submarine in H cannot be inferred from the pronoun it in T. Example (vi) demonstrates an edge as target component. In this case, the edge in H connecting melt with the modifier in the Arctic is not found in T. Although each of the hypothesis’ nodes can be covered separately via knowledgebased rules (e.g. ‘Siberia →Arctic’, ‘permafrost →ice’, ‘thaw ↔melt’), the resulting fragments in T are unconnected without the (intra-sentential) coreference between them and lakes in Siberia. For each target component, we identify its focus term as the expression in T that does not cover the target component itself but participates in a reference relation that can help covering it. We follow the focus term’s reference chain to a reference term which can, either separately or in combination with the focus term, help covering the target component. In Example (ii), where the 5In our annotation, we assume throughout that some knowledge about basic admissible transformations is available, such as passive to active or derivational transformations; for brevity, we ignore articles in the examples and treat named entities as single nodes. 1212 target component in H is 2003 UB313, Xena is the focus term in T and the reference term is a mention of 2003 UB313 in a previous sentence, T ′. In this case, the reference term covers the entire target component on its own. An additional attribute that we record for each instance is whether resolving the discourse reference is mandatory for determining entailment, or optional. In Example (v), it is mandatory: the inference cannot be completed without the knowledge provided by the discourse. In contrast, in Example (ii), inferring 2003 UB313 from Xena is optional. It can be done either by identifying their coreference relation, or by using background knowledge in the form of an entailment rule, ‘Xena ↔2003 UB313’, that is applicable in the context of astronomy. Optional discourse references represent instances where discourse information and TE knowledge are interchangeable. As mentioned, knowledge gaps constitute a major obstacle for TE systems, and we cannot rely on the availability of any ceratin piece of knowledge to the inference process. Thus, in our scheme, mandatory references provide a “lower bound” with regards to the necessity to resolve discourse references, even in the presence of complete knowledge; optional references, on the other hand, set an “upper bound” for the contribution of discourse resolution to inference, when no knowledge is available. At the same time, this scheme allows investigating how much TE knowledge can be replaced by (perfect) discourse processing. When choosing a reference term, we search the reference chain of the focus term for the nearest expression that is identical to the target component or a subcomponent of it. If we find such an expression, covering the identical part of the target component requires no entailment knowledge. If no identical reference term exists, we choose the semantically ‘closest’ term from the reference chain, i.e. the term which requires the least knowledge to infer the target component. For instance, we may pick permafrost as the semantically closet term to the target ice if the latter is not found in the focus term’s reference chain. Finally, for each reference relation that we annotate, we record four additional attributes which we assumed to be informative in an evaluation. First, the reference type: Is the relation a coreference or a bridging reference? Second, the syntactic type of the focus and reference terms. Third, the focus/reference terms entailment status – does some kind of entailment relation hold between the two terms? Fourth, the operation that should be performed on the focus and reference terms to obtain coverage of the target component (as specified in Section 5). 5 Integrating Discourse References into Entailment Recognition In initial analysis we found that the standard substitution operation applied by virtually all previous studies for integrating coreference into entailment is insufficient. We identified three distinct cases for the integration of discourse reference knowledge in entailment, which correspond to different relations between the target component, the focus term and the reference term. This section describes the three cases and characterizes them in terms of tree transformations. An initial version of these transformations is described in (Abad et al., 2010). We assume a transformation-based entailment architecture (cf. Section 2.2), although we believe that the key points of our account are also applicable to alignment-based architecture. Transformations create revised trees that cover previously uncovered target components in H. The output of each transformation, T1, is comprised of copies of the components used to construct it, and is appended to the discourse forest, which includes the dependency trees of all sentences and their generated consequents. We assume that we have access to a dependency tree for H, a dependency forest for T and its discourse context, as well as the output of a perfect discourse processor, i.e., a complete set of both coreference and bridging relations, including the type of bridging relation (e.g. part-of, cause). We use the following notation. We use x, y for tree nodes, and Sx to denote a (sub-)tree with root x. lab(x) is the label of the incoming edge of x (i.e., its grammatical function). We write C(x, y) for a coreference relation between Sx and Sy, the corresponding trees of the focus and reference terms, respectively. We write Br(x, y) for a bridging relation, where r is its type. (1) Substitution: This is the most intuitive and widely-used transformation, corresponding to the treatment of discourse information in existing systems. It applies to coreference relations, when an expression found elsewhere in the text (the reference term) can cover all missing information (the 1213 be legal also union such pred mod subj be legal also marriages homosexual pred mod subj mod T T1 marriages homosexual mod T’ pre Figure 1: The Substitution transformation, demonstrated on the relevant subtrees of Example (i). The dashed line denotes a discourse reference. target component) on its own. In such cases, the reference term can replace the entire focus term. Apparently (cf. Section 6), substitution applies also to some types of bridging relations, such as set-membership, when the member is sufficient for representing the entire set for the necessary inference. For example, in “I met two people yesterday. The woman told me a story.” (Clark, 1975), substituting two people with woman results in a text which is entailed from the discourse, and which allows inferring “I met a woman yesterday.” In a parse tree representation, given a coreference relation C(x, y) (or Br(x, y)), the newly generated tree, T1, consists of a copy of T, where the entire tree Sx is replaced by a copy of Sy . In Figure 1, which shows Example (i) from Table 1, such unions is substituted by homosexual marriages. Head-substitution. Occasionally, substituting only the head of the focus term is sufficient. In such cases, only the root nodes x and y are substituted. This is the case, for example, with synonymous verbs with identical subcategorization frames (like melt and thaw). As verbs typically constitute tree roots in dependency parses, substituting or merging (see below) their entire trees might be inappropriate or wasteful. In such cases, the simpler head-substitution may be applied. (2) Merge: In contrast to substitution, where a match for the entire target component is found elsewhere in the text, this transformation is required when parts of the missing information are scattered among multiple locations in the text. We distinguish between two types of merge transformations: (a) dependent-merge, and (b) headmerge, depending on the syntactic roles of the merged components. (a) Dependent-Merge. This operation is applicable when the head of either the focus or reference terms (of both) matches the head node of submarine mini on trapped mod T T1 submarine AS-28 nn T’a pcomp-n pnmod mod seabed submarine mini trapped mod pnmod mod AS-28 nn AS-28 T’b on pcomp-n seabed Figure 2: The dependent-merge (T ′ a) and headmerge (T ′ b) transformations (Example (iii)). the target component, but modifiers from both of them are required to cover the target component’s dependents. The modifiers are therefore merged as dependents of a single head node, to create a tree that covers the entire target component. Dependent-merge is illustrated in Figure 2, using Example (iii). The component we wish to cover in H is the noun phrase AS-28 mini submarine. Unfortunately, the focus term in T, “mini submarine trapped on the seabed”, covers only the modifier mini, but not AS-28. This modifier can however be provided by the coreferent term in T ′ a (left upper corner). Once merged, the inference engine can, e.g., employ the rule ‘on seabed →underwater’ to cover H completely. Formally, assume without loss of generality that y, the reference term’s head, matches the root node of the target component. Given C(x, y), we define T1 as a copy of T, where (i) the subtree Sx is replaced by Sy, and (ii) for all children c of x, a copy of Sc is placed under the copy of y in T1 with its original edge label, lab(c). (b) Head-merge. An alternative way to recover the missing information in Example (iii) is to find a reference term whose head word itself (rather than one of its modifiers) matches the target component’s missing dependent, as with AS-28 in Figure 2 in the bottom left corner (T ′ b). In terms of parse trees, we need to add one tree as a dependent of the other. Formally, given C(x, y), similarly to dependent-merge, T1 is created as a copy of T where the subtree Sx is replaced by either Sx or Sy, depending on whichever of x and y matches the target component’s head. Assume it is x, for example. Then, a copy of Sy is added as a new child to x. In our sample, head-merge operations correspond to internal coreferences within nominal target components (such as between AS-28 and mini submarine in this case). The appropriate label, lab(y), in these cases is nn (nominal modi1214 in T T1 T’ pcomp-n China cost have than more comp1 pcomp-n obj have dozen accident subj recent mod cost have than more comp1 pcomp-n obj have dozen accident subj recent mod mod Solution seek China to mod pcomp-n safety coal mine nn nn its gen obj subj Figure 3: The insertion transformation. Dotted edges mark the newly inserted path (Ex. (iv)). fier). Further analysis is required to specify what other dependencies can hold between such coreferring heads. (3) Insertion: The last transformation, insertion, is used when a relation that is realized in H is missing from T and is only implied via a bridging relation. In Example (iv), the location that is explicitly mentioned in H can only be covered by T by resolving a bridging reference with China in T ′. To connect the bridging referents, a new tree component representing the bridging relation is inserted into the consequent tree T1. In this example, the component connects China and recent accident via the in preposition. Formally, given a bridging relation Br(x, y), we introduce a new subtree Sr z into T1, where z is a child of x and lab(z) = labr. Sr z must contain a variable node that is instantiated with a copy of S(y). This transformation stands out from the others in that it introduces new material. For each bridging relation, it adds a specific subtrees Sr via an edge labeled with labr. These two items form the dependency representation of the bridging relation Br and must be provided by the interface between the discourse and the inference systems. Clearly, their exact form depends on the set of bridging relations provided by the discourse resolver as well as the details of the dependency parses. As shown in Figure 3, the bridging relation located-in (r) is represented by inserting a subtree Sr z headed by in (z) into T1 and connecting it to accident (x) as a modifier (labr). The subtree Sr z consists of a variable node which is connected to in with a pcomp-n dependency (a nominal head of a prepositional phrase), and which is instantiated with the node China (y) when the transformation is applied. Note that the structure of Sr z and the way it is inserted into T1 are predefined by the abovementioned interface; only the node to which it is attached and the contents of the variable node are determined at transformation-time. As another example, consider the following short text from (Clark, 1975): John was murdered yesterday. The knife lay nearby. Here, the bridging relation between the murder event and the instrument, the knife (x), can be addressed by inserting under x a subtree for the clause with which as Sr z, with a variable which is instantiated by the parse-tree (headed by murdered, y) of the entire first sentence John was murdered yesterday. Transformation chaining. Since our transformations are defined to be minimal, some cases require the application of multiple transformations to achieve coverage. Consider Example (v), Table 1. We wish to cover AS-28 mini submarine in H from the coreferring it in T, mini submarine in T ′ and AS-28 vehicle in T ′′. A substitution of it by either coreference does not suffice, since none of the antecedents contains all necessary modifiers. It is therefore necessary to substitute it first by one of the coreferences and then merge it with the other. 6 Results We analyzed 120 sentence-hypothesis pairs of the RTE-5 development set (21 different hypotheses, 111 distinct sentences, 53 different documents). Below, we summarize our findings, focusing on the relation between our findings and the assumptions of previous studies as discussed in Section 3. General statistics. We found that 44% of the pairs contained reference relations whose resolution was mandatory for inference. In another 28%, references could optionally support the inference of the hypothesis. In the remaining 28%, references did not contribute towards inference. The total number of relevant references was 137, and 37 pairs (27%) contained multiple relevant references. These numbers support our assumption that discourse references play an important role in inference. Reference types. 73% of the identified references are coreferences and 27% are bridging relations. The most common bridging relation was the location of events (e.g. Arctic in ice melting events), generally assumed to be known throughout the document. Other bridging relations we encountered include cause (e.g. between injured and attack), event participants and set membership. 1215 (%) Pronoun NE NP VP Focus term 9 19 49 23 Reference term 43 43 14 Table 2: Syntactic types of discourse references (%) Sub. Merge Insertion Coreference 62 38 Bridging 30 70 Total 54 28 18 Table 3: Distribution of transformation types Syntactic types. Table 2 shows that 77% of all focus terms and 86% of the reference terms were nominal phrases, which justifies their prominent position in work on anaphora and coreference resolution. However, almost a quarter of the focus terms were verbal phrases. We found these focus terms to be frequently crucial for entailment since they included the main predicate of the hypothesis.6 This calls for an increased focus on the resolution of event references. Transformations. Table 3 shows the relative frequencies of all transformations. Again, we found that the “default” transformation, substitution, is the most frequent one, and is helpful for both coreference and bridging relations. Substitution is particularly useful for handling pronouns (14% of all substitution instances), the replacement of named entities by synonymous names (32%), the replacement of other NPs (38%), and the substitution of verbal head nodes in event coreference (16%). Yet, in nearly half the cases, a different transformation had to be applied. Insertion accounts for the majority of bridging cases. Head-merge is necessary to integrate proper nouns as modifiers of other head nouns. Dependentmerge, responsible for 85% of the merge transformations, can be used to complete nominal focus terms with missing modifiers (e.g., adjectives), as well as for merging other dependencies between coreferring predicates. This result indicates the importance of incorporating other transformations into inference systems. Distance of reference terms. The distance between the focus and the reference terms varied considerably, ranging from intra-sentential reference relations and up to several dozen sentences. For more than a quarter of the focus terms, we 6The lower proportion of VPs among reference terms stems from bridging relations between VPs and nominal dependents, such as the abovementioned “location” relation. had to go to other documents to find reference terms that, possibly in conjunction with the focus term, could cover the target components. Interestingly, all such cases involved coreference (about equally divided between the merge transformations and substitutions), while bridging was always “document-local”. This result reaffirms the usefulness of cross-document coreference resolution for inference (Huang et al., 2009). Discourse resolution as preprocessing? In existing RTE systems, discourse references are typically resolved as a preprocessing step. While our annotation was manual and cannot yield direct results about processing considerations, we observed that discourse relations often hold between complex, and deeply embedded, expressions, which makes their automatic resolution difficult. Of course, many RTE systems attempt to normalize and simplify H and T, e.g., by splitting conjunctions or removing irrelevant clauses, but these operations are usually considered a part of the inference rather the preprocessing phase (cf. e.g., Bar-Haim et al. (2007)). Since the resolution of discourse references is likely to profit from these steps, it seems desirable to “postpone” it until after simplification. In transformation-based systems, it might be natural to add discourse-based transformations to the set of inference operations, while in alignment-based systems, discourse references can be integrated into the computation of alignment scores. Discourse references vs. entailment knowledge. We have stated before that even if a discourse reference is not strictly necessary for entailment, it may be interesting because it represents an alternative to the use of knowledge rules to cover the hypothesis. Sometimes, these rules are generally applicable (e.g., ‘Alaska →Arctic’). However, often they are context-specific. Consider the following sentence as T for the hypothesis H: “The ice is melting in the Arctic”: (3) T: “The scene at the receding edge of the Exit Glacier was part festive gathering, part nature tour with an apocalyptic edge.” While it is possible to cover melting using a rule ‘melting ↔receding’, this rule is only valid under quite specific conditions (e.g., for the subject ice). Instead of determining the applicability of the rule, a discourse-aware system can take the next sen1216 tence into account, which contains a coreferring event to receding that can cover melting in H: (4) T ′: “. . . people moved closer to the rope line near the glacier as it shied away, practically groaning and melting before their eyes.” Discourse relations can in fact encode arbitrarily complex world knowledge, as in the following pair: (5) H: “The serial killer BTK was accused of at least 7 killings starting in the 1970’s.” T: “Police say BTK may have killed as many as 10 people between 1974 and 1991.” Here, the H modifier serial, which does not occur in T, can be covered either by world knowledge (a person who killed 10 people is a serial killer), or by resolving the coreference of BTK to the term the serial killer BTK which occurs in the discourse around T. Our conclusion is that not only can discourse references often replace world knowledge in principle, in practice it often seems easier to resolve discourse references than to determine whether a rule is applicable in a given context or to formalize complex world knowledge as inference rules. Our annotation provides further empirical support to this claim: An entailment relation exists between the focus and reference terms in 60% of the focus-reference term pairs, and in many of the remainder, entailment holds between the terms’ heads. Thus, discourse provides relations which are many times equivalent to entailment knowledge rules and can therefore be utilized in their stead. 7 Conclusions This work has presented an analysis of the relation between discourse references and textual entailment. We have identified a set of limitations common to the handling of discourse relations in virtually all entailment systems. They include the use of off-the-shelf resolvers that concentrate on nominal coreference, the integration of reference information through substitution, and the RTE evaluation schemes, which played down the role of discourse. Since in practical settings, discourse plays an important role, our goal was to develop an agenda for improving the handling of discourse references in entailment-based inference. Our manual analysis of the RTE-5 dataset shows that while the majority of discourse references that affect inference are nominal coreference relations, another substantial part is made up by verbal terms and bridging relations. Furthermore, we have demonstrated that substitution alone is insufficient to extract all relevant information from the wide range of discourse references that are frequently relevant for inference. We identified three general cases, and suggested matching operations to obtain the relevant inferences, formulated as tree transformations. Furthermore, our evidence suggests that for practical reasons, the resolution of discourse references should be tightly integrated into entailment systems instead of treating it as a preprocessing step. A particularly interesting result concerns the interplay between discourse references and entailment knowledge. While semantic knowledge (e.g., from WordNet or Wikipedia) has been used beneficially for coreference resolution (Soon et al., 2001; Ponzetto and Strube, 2006), reference resolution has, to our knowledge, not yet been employed to validate entailment rules’ applicability. Our analyses suggest that in the context of deciding textual entailment, reference resolution and entailment knowledge can be seen as complementary ways of achieving the same goal, namely enriching T with additional knowledge to allow the inference of H. Given that both of the technologies are still imperfect, we envisage the way forward as a joint strategy, where reference resolution and entailment rules mutually fill each other’s gaps (cf. Example 3). In sum, our study shows that textual entailment can profit substantially from better discourse handling. The next challenge is to translate the theoretical gain into practical benefit. Our analysis demonstrates that improvements are necessary both on the side of discourse reference resolution systems, which need to cover more types of references, as well as a better integration of discourse information in entailment systems, even for those relations which are within the scope of available resolvers. Acknowledgements This work was partially supported by the PASCAL-2 Network of Excellence of the European Community FP7-ICT-2007-1-216886 and the Israel Science Foundation grant 1112/08. 1217 References Azad Abad, Luisa Bentivogli, Ido Dagan, Danilo Giampiccolo, Shachar Mirkin, Emanuele Pianta, and Asher Stern. 2010. A resource for investigating the impact of anaphora and coreference on inference. In Proceedings of LREC. Rod Adams, Gabriel Nicolae, Cristina Nicolae, and Sanda Harabagiu. 2007. Textual entailment through extended lexical overlap and lexico-semantic matching. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing. E. Agichtein, W. Askew, and Y. Liu. 2008. Combining lexical, syntactic, and semantic evidence for textual entailment classification. In Proceedings of TAC. Nicholas Asher and Alex Lascarides. 1998. Bridging. Journal of Semantics, 15(1):83–113. Alexandra Balahur, Elena Lloret, ´Oscar Ferr´andez, Andr´es Montoyo, Manuel Palomar, and Rafael Mu˜noz. 2008. The DLSIUAES team’s participation in the TAC 2008 tracks. In Proceedings of TAC. Roy Bar-Haim, Ido Dagan, Iddo Greental, and Eyal Shnarch. 2007. Semantic inference at the lexicalsyntactic level. In Proceedings of AAAI. Roy Bar-Haim, Jonathan Berant, Ido Dagan, Iddo Greental, Shachar Mirkin, and Eyal Shnarch amd Idan Szpektor. 2008. Efficient semantic deduction and approximate matching over compact parse forests. In Proceedings of TAC. Luisa Bentivogli, Ido Dagan, Hoa Trang Dang, Danilo Giampiccolo, and Bernardo Magnini. 2009a. The fifth pascal recognizing textual entailment challenge. In Proceedings of TAC. Luisa Bentivogli, Ido Dagan, Hoa Trang Dang, Danilo Giampiccolo, Medea Lo Leggio, and Bernardo Magnini. 2009b. Considering discourse references in textual entailment annotation. In Proceedings of the 5th International Conference on Generative Approaches to the Lexicon (GL2009). Johan Bos. 2005. Recognising textual entailment with logical inference. In Proceedings of EMNLP. Aljoscha Burchardt, Marco Pennacchiotti, Stefan Thater, and Manfred Pinkal. 2009. Assessing the impact of frame semantics on textual entailment. Journal of Natural Language Engineering, 15(4):527–550. Nathanael Chambers and Dan Jurafsky. 2009. Unsupervised learning of narrative schemas and their participants. In Proceedings of ACL-IJCNLP. Nathanael Chambers, Daniel Cer, Trond Grenager, David Hall, Chloe Kiddon, Bill MacCartney, MarieCatherine de Marneffe, Daniel Ramage, Eric Yeh, and Christopher D. Manning. 2007. Learning alignments and leveraging natural logic. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing. Herbert H. Clark. 1975. Bridging. In R. C. Schank and B. L. Nash-Webber, editors, Theoretical issues in natural language processing, pages 169–174. Association of Computing Machinery. Ido Dagan, Oren Glickman, and Bernardo Magnini. 2006. The PASCAL recognising textual entailment challenge. In Machine Learning Challenges, volume 3944 of Lecture Notes in Computer Science, pages 177–190. Springer. Lorand Dali, Delia Rusu, Blaz Fortuna, Dunja Mladenic, and Marko Grobelnik. 2009. Question answering based on semantic graphs. In Proceedings of the Workshop on Semantic Search (SemSearch 2009). Christiane Fellbaum, editor. 1998. WordNet: An Electronic Lexical Database (Language, Speech, and Communication). The MIT Press. Danilo Giampiccolo, Bernardo Magnini, Ido Dagan, and Bill Dolan. 2007. The third pascal recognizing textual entailment challenge. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing. Danilo Giampiccolo, Hoa Trang Dang, Bernardo Magnini, Ido Dagan, and Bill Dolan. 2008. The fourth pascal recognizing textual entailment challenge. In Proceedings of TAC. Ralph Grishman and Beth Sundheim. 1996. Message Understanding Conference-6: a brief history. In Proceedings of the 16th conference on Computational Linguistics. Sanda Harabagiu, Andrew Hickl, and Finley Lacatusu. 2007. Satisfying information needs with multidocument summaries. Information Processing & Management, 43:1619–1642. Stefan Harmeling. 2009. Inferring textual entailment with a probabilistically sound calculus. Journal of Natural Language Engineering, pages 459–477. Marti A. Hearst. 1997. Segmenting text into multiparagraph subtopic passages. Computational Linguistics, 23(1):33–64. Eduard Hovy, Mitchell Marcus, Martha Palmer, Lance Ramshaw, and Ralph Weischedel. 2006. Ontonotes: The 90% solution. In Proceedings of HLT-NAACL. Jian Huang, Sarah M. Taylor, Jonathan L. Smith, Konstantinos A. Fotiadis, and C. Lee Giles. 2009. Profile based cross-document coreference using kernelized fuzzy relational clustering. In Proceedings of ACL-IJCNLP. Fangtao Li, Yang Tang, Minlie Huang, and Xiaoyan Zhu. 2009. Answering opinion questions with random walks on graphs. In Proceedings of ACLIJCNLP. 1218 Dekang Lin and Patrick Pantel. 2001. Discovery of inference rules for question answering. Natural Language Engineering, 4(7):343–360. Dekang Lin. 1993. Principle-based parsing without overgeneration. In Proceedings of ACL. Bill MacCartney, Michel Galley, and Christopher D. Manning. 2008. A phrase-based alignment model for natural language inference. In Proceedings of EMNLP. Katja Markert, Malvina Nissim, and Natalia N. Modjeska. 2003. Using the web for nominal anaphora resolution. In Proceedings of EACL Workshop on the Computational Treatment of Anaphora. Seung-Hoon Na and Hwee Tou Ng. 2009. A 2-poisson model for probabilistic coreference of named entities for improved text retrieval. In Proceedings of SIGIR. Rodney D. Nielsen, Wayne Ward, and James H. Martin. 2009. Recognizing entailment in intelligent tutoring systems. Natural Language Engineering, 15(4):479–501. Massimo Poesio, Rahul Mehta, Axel Maroudas, and Janet Hitzeman. 2004. Learning to resolve bridging references. In Proceedings of ACL. Simone Paolo Ponzetto and Michael Strube. 2006. Exploiting semantic role labeling, WordNet and Wikipedia for coreference resolution. In Proceedings of HLT. Long Qiu, Min-Yen Kan, and Tat-Seng Chua. 2004. A public reference implementation of the rap anaphora resolution algorithm. In Proceedings of LREC. Lorenza Romano, Milen Kouylekov, Idan Szpektor, Ido Dagan, and Alberto Lavelli. 2006. Investigating a generic paraphrase-based approach for relation extraction. In Proceedings of EACL. Wee Meng Soon, Hwee Tou Ng, and Daniel Chung Yong Lim. 2001. A machine learning approach to coreference resolution of noun phrases. Computational Linguistics, 27(4):521–544. Stephanie Strassel, Mark Przybocki, Kay Peterson, Zhiyi Song, and Kazuaki Maeda. 2008. Linguistic resources and evaluation techniques for evaluation of cross-document automatic content extraction. In Proceedings of LREC. Jose L. Vicedo and Antonio Ferrndez. 2006. Coreference in Q&A. In Tomek Strzalkowski and Sanda M. Harabagiu, editors, Advances in Open Domain Question Answering, pages 71–96. Springer. Fabio Massimo Zanzotto, Marco Pennacchiotti, and Alessandro Moschitti. 2009. A machine learning approach to textual entailment recognition. Journal of Natural Language Engineering, 15(4):551–582. Dmitry Zelenko, Chinatsu Aone, and Jason Tibbetts. 2004. Coreference resolution for information extraction. In Proceedings of the ACL Workshop on Reference Resolution and its Applications. 1219
|
2010
|
123
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1220–1229, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Global Learning of Focused Entailment Graphs Jonathan Berant Tel-Aviv University Tel-Aviv, Israel [email protected] Ido Dagan Bar-Ilan University Ramat-Gan, Israel [email protected] Jacob Goldberger Bar-Ilan University Ramat-Gan, Israel [email protected] Abstract We propose a global algorithm for learning entailment relations between predicates. We define a graph structure over predicates that represents entailment relations as directed edges, and use a global transitivity constraint on the graph to learn the optimal set of edges, by formulating the optimization problem as an Integer Linear Program. We motivate this graph with an application that provides a hierarchical summary for a set of propositions that focus on a target concept, and show that our global algorithm improves performance by more than 10% over baseline algorithms. 1 Introduction The Textual Entailment (TE) paradigm (Dagan et al., 2009) is a generic framework for applied semantic inference. The objective of TE is to recognize whether a target meaning can be inferred from a given text. For example, a Question Answering system has to recognize that ‘alcohol affects blood pressure’ is inferred from ‘alcohol reduces blood pressure’ to answer the question ‘What affects blood pressure?’ TE systems require extensive knowledge of entailment patterns, often captured as entailment rules: rules that specify a directional inference relation between two text fragments (when the rule is bidirectional this is known as paraphrasing). An important type of entailment rule refers to propositional templates, i.e., propositions comprising a predicate and arguments, possibly replaced by variables. The rule required for the previous example would be ‘X reduce Y →X affect Y’. Because facts and knowledge are mostly expressed by propositions, such entailment rules are central to the TE task. This has led to active research on broad-scale acquisition of entailment rules for predicates, e.g. (Lin and Pantel, 2001; Sekine, 2005; Szpektor and Dagan, 2008). Previous work has focused on learning each entailment rule in isolation. However, it is clear that there are interactions between rules. A prominent example is that entailment is a transitive relation, and thus the rules ‘X →Y ’ and ‘Y →Z’ imply the rule ‘X →Z’. In this paper we take advantage of these global interactions to improve entailment rule learning. First, we describe a structure termed an entailment graph that models entailment relations between propositional templates (Section 3). Next, we show that we can present propositions according to an entailment hierarchy derived from the graph, and suggest a novel hierarchical presentation scheme for corpus propositions referring to a target concept. As in this application each graph focuses on a single concept, we term those focused entailment graphs (Section 4). In the core section of the paper, we present an algorithm that uses a global approach to learn the entailment relations of focused entailment graphs (Section 5). We define a global function and look for the graph that maximizes that function under a transitivity constraint. The optimization problem is formulated as an Integer Linear Program (ILP) and solved with an ILP solver. We show that this leads to an optimal solution with respect to the global function, and demonstrate that the algorithm outperforms methods that utilize only local information by more than 10%, as well as methods that employ a greedy optimization algorithm rather than an ILP solver (Section 6). 2 Background Entailment learning Two information types have primarily been utilized to learn entailment rules between predicates: lexicographic resources and distributional similarity resources. Lexicographic 1220 resources are manually-prepared knowledge bases containing information about semantic relations between lexical items. WordNet (Fellbaum, 1998), by far the most widely used resource, specifies relations such as hyponymy, derivation, and entailment that can be used for semantic inference (Budanitsky and Hirst, 2006). WordNet has also been exploited to automatically generate a training set for a hyponym classifier (Snow et al., 2005), and we make a similar use of WordNet in Section 5.1. Lexicographic resources are accurate but tend to have low coverage. Therefore, distributional similarity is used to learn broad-scale resources. Distributional similarity algorithms predict a semantic relation between two predicates by comparing the arguments with which they occur. Quite a few methods have been suggested (Lin and Pantel, 2001; Bhagat et al., 2007; Yates and Etzioni, 2009), which differ in terms of the specifics of the ways in which predicates are represented, the features that are extracted, and the function used to compute feature vector similarity. Details on such methods are given in Section 5.1. Global learning It is natural to describe entailment relations between predicates by a graph. Nodes represent predicates, and edges represent entailment between nodes. Nevertheless, using a graph for global learning of entailment between predicates has attracted little attention. Recently, Szpektor and Dagan (2009) presented the resource Argument-mapped WordNet, providing entailment relations for predicates in WordNet. Their resource was built on top of WordNet, and makes simple use of WordNet’s global graph structure: new rules are suggested by transitively chaining graph edges, and verified against corpus statistics. The most similar work to ours is Snow et al.’s algorithm for taxonomy induction (2006). Snow et al.’s algorithm learns the hyponymy relation, under the constraint that it is a transitive relation. Their algorithm incrementally adds hyponyms to an existing taxonomy (WordNet), using a greedy search algorithm that adds at each step the set of hyponyms that maximize the probability of the evidence while respecting the transitivity constraint. In this paper we tackle a similar problem of learning a transitive relation, but we use linear programming. A Linear Program (LP) is an optimization problem, where a linear function is minimized (or maximized) under linear constraints. If the variables are integers, the problem is termed an Integer Linear Program (ILP). Linear programming has attracted attention recently in several fields of NLP, such as semantic role labeling, summarization and parsing (Roth and tau Yih, 2005; Clarke and Lapata, 2008; Martins et al., 2009). In this paper we formulate the entailment graph learning problem as an Integer Linear Program, and find that this leads to an optimal solution with respect to the target function in our experiment. 3 Entailment Graph This section presents an entailment graph structure, which resembles the graph in (Szpektor and Dagan, 2009). The nodes of an entailment graph are propositional templates. A propositional template is a path in a dependency tree between two arguments of a common predicate1 (Lin and Pantel, 2001; Szpektor and Dagan, 2008). Note that in a dependency parse, such a path passes through the predicate. We require that a variable appears in at least one of the argument positions, and that each sense of a polysemous predicate corresponds to a separate template (and a separate graph node): X subj ←−−treat#1 obj −−→Y and X subj ←−−treat#1 obj −−→nausea are propositional templates for the first sense of the predicate treat. An edge (u, v) represents the fact that template u entails template v. Note that the entailment relation transcends beyond hyponymy. For example, the template X is diagnosed with asthma entails the template X suffers from asthma, although one is not a hyponoym of the other. An example of an entailment graph is given in Figure 1, left. Since entailment is a transitive relation, an entailment graph is transitive, i.e., if the edges (u, v) and (v, w) are in the graph, so is the edge (u, w). This is why we require that nodes be sensespecified, as otherwise transitivity does not hold: Possibly a →b for one sense of b, b →c for another sense of b, but a ↛c. Because graph nodes represent propositions, which generally have a clear truth value, we can assume that transitivity is indeed maintained along paths of any length in an entailment graph, as entailment between each pair of nodes either occurs or doesn’t occur with very high probability. We support this further in section 4.1, where we show 1We restrict our discussion to templates with two arguments, but generalization is straightforward. 1221 X-related-to-nausea X-associated-with-nausea X-prevent-nausea X-help-with-nausea X-reduce-nausea X-treat-nausea related to nausea headache Oxicontine help with nausea prevent nausea acupuncture ginger reduce nausea relaxation treat nausea drugs Nabilone Lorazepam Figure 1: Left: An entailment graph. For clarity, edges that can be inferred by transitivity are omitted. Right: A hierarchical summary of propositions involving nausea as an argument, such as headache is related to nausea, acupuncture helps with nausea, and Lorazepam treats nausea. that in our experimental setting the length of paths in the entailment graph is relatively small. Transitivity implies that in each strong connectivity component2 of the graph, all nodes are synonymous. Moreover, if we merge every strong connectivity component to a single node, the graph becomes a Directed Acyclic Graph (DAG), and the graph nodes can be sorted and presented hierarchically. Next, we show an application that leverages this property. 4 Motivating Application In this section we propose an application that provides a hierarchical view of propositions extracted from a corpus, based on an entailment graph. Organizing information in large collections has been found to be useful for effective information access (Kaki, 2005; Stoica et al., 2007). It allows for easier data exploration, and provides a compact view of the underlying content. A simple form of structural presentation is by a single hierarchy, e.g. (Hofmann, 1999). A more complex approach is hierarchical faceted metadata, where a number of concept hierarchies are created, corresponding to different facets or dimensions (Stoica et al., 2007). Hierarchical faceted metadata categorizes concepts of a domain in several dimensions, but does not specify the relations between them. For example, in the health-care domain we might have facets for categories such as diseases and symptoms. Thus, when querying about nausea, one might find it is related to vomitting and chicken pox, but not that chicken pox is a cause of nausea, 2A strong connectivity component is a subset of nodes in the graph where there is a path from any node to any other node. while nausea is often accompanied by vomitting. We suggest that the prominent information in a text lies in the propositions it contains, which specify particular relations between the concepts. Propositions have been mostly presented through unstructured textual summaries or manually-constructed ontologies, which are expensive to build. We propose using the entailment graph structure, which describes entailment relations between predicates, to naturally present propositions hierarchically. That is, the entailment hierarchy can be used as an additional facet, which can improve navigation and provide a compact hierarchical summary of the propositions. Figure 1 illustrates a scenario, on which we evaluate later our learning algorithm. Assume a user would like to retrieve information about a target concept such as nausea. We can extract the set of propositions where nausea is an argument automatically from a corprus, and learn an entailment graph over propositional templates derived from the extracted propositions, as illustrated in Figure 1, left. Then, we follow the steps in the process described in Section 3: merge synonymous nodes that are in the same strong connectivity component, and turn the resulting DAG into a predicate hierarchy, which we can then use to present the propositions (Figure 1, right). Note that in all propositional templates one argument is the target concept (nausea), and the other is a variable whose corpus instantiations can be presented according to another hierarchy (e.g. Nabilone and Lorazepam are types of drugs). Moreover, new propositions are inferred from the graph by transitivity. For example, from the proposition ‘relaxation reduces nausea’ we can in1222 fer the proposition ‘relaxation helps with nausea’. 4.1 Focused entailment graphs The application presented above generates entailment graphs of a specific form: (1) Propositional templates have exactly one argument instantiated by the same entity (e.g. nausea). (2) The predicate sense is unspecified, but due to the rather small number of nodes and the instantiating argument, each predicate corresponds to a unique sense. Generalizing this notion, we define a focused entailment graph to be an entailment graph where the number of nodes is relatively small (and consequently paths in the graph are short), and predicates have a single sense (so transitivity is maintained without sense specification). Section 5 presents an algorithm that given the set of nodes of a focused entailment graph learns its edges, i.e., the entailment relations between all pairs of nodes. The algorithm is evaluated in Section 6 using our proposed application. For brevity, from now on the term entailment graph will stand for focused entailment graph. 5 Learning Entailment Graph Edges In this section we present an algorithm for learning the edges of an entailment graph given its set of nodes. The first step is preprocessing: We use a large corpus and WordNet to train an entailment classifier that estimates the likelihood that one propositional template entails another. Next, we can learn on the fly for any input graph: given the graph nodes, we employ a global optimization approach that determines the set of edges that maximizes the probability (or score) of the entire graph, given the edge probabilities (or scores) supplied by the entailment classifier and the graph constraints (transitivity and others). 5.1 Training an entailment classifier We describe a procedure for learning an entailment classifier, given a corpus and a lexicographic resource (WordNet). First, we extract a large set of propositional templates from the corpus. Next, we represent each pair of propositional templates with a feature vector of various distributional similarity scores. Last, we use WordNet to automatically generate a training set and train a classifier. Template extraction We parse the corpus with a dependency parser and extract all propositional templates from every parse tree, employing the procedure used by Lin and Pantel (2001). However, we only consider templates containing a predicate term and arguments3. The arguments are replaced with variables, resulting in propositional templates such as X subj ←−−affect obj −−→Y. Distributional similarity representation We aim to train a classifier that for an input template pair (t1, t2) determines whether t1 entails t2. A template pair is represented by a feature vector where each coordinate is a different distributional similarity score. There are a myriad of distributional similarity algorithms. We briefly describe those used in this paper, obtained through variations along the following dimensions: Predicate representation Most algorithms measure the similarity between templates with two variables (binary templates) such as X subj ←−−affect obj −−→Y (Lin and Pantel, 2001; Bhagat et al., 2007; Yates and Etzioni, 2009). Szpketor and Dagan (2008) suggested learning over templates with one variable (unary templates) such as X subj ←−−affect, and using them to estimate a score for binary templates. Feature representation The features of a template are some representation of the terms that instantiated the argument variables in a corpus. Two representations are used in our experiment (see Section 6). Another variant occurs when using binary templates: a template may be represented by a pair of feature vectors, one for each variable (Lin and Pantel, 2001), or by a single vector, where features represent pairs of instantiations (Szpektor et al., 2004; Yates and Etzioni, 2009). The former variant reduces sparsity problems, while Yates and Etzioni showed the latter is more informative and performs favorably on their data. Similarity function We consider two similarity functions: The Lin (2001) similarity measure, and the Balanced Inclusion (BInc) similarity measure (Szpektor and Dagan, 2008). The former is a symmetric measure and the latter is asymmetric. Therefore, information about the direction of entailment is provided by the BInc measure. We then generate for any (t1, t2) features that are the 12 distributional similarity scores using all combinations of the dimensions. This is reminiscent of Connor and Roth (2007), who used the output of unsupervised classifiers as features for a supervised classifier in a verb disambiguation task. 3Via a simple heuristic, omitted due to space limitations 1223 Training set generation Following the spirit of Snow et al. (2005), WordNet is used to automatically generate a training set of positive (entailing) and negative (non-entailing) template pairs. Let T be the set of propositional templates extracted from the corpus. For each ti ∈T with two variables and a single predicate word w, we extract from WordNet the set H of direct hypernyms and synonyms of w. For every h ∈H, we generate a new template tj from ti by replacing w with h. If tj ∈T, we consider (ti, tj) to be a positive example. Negative examples are generated analogously, by looking at direct co-hyponyms of w instead of hypernyms and synonyms. This follows the notion of “contrastive estimation” (Smith and Eisner, 2005), since we generate negative examples that are semantically similar to positive examples and thus focus the classifier’s attention on identifying the boundary between the classes. Last, we filter training examples for which all features are zero, and sample an equal number of positive and negative examples (for which we compute similarity features), since classifiers tend to perform poorly on the minority class when trained on imbalanced data (Van Hulse et al., 2007; Nikulin, 2008). 5.2 Global learning of edges Once the entailment classifier is trained we learn the graph edges given its nodes. This is equivalent to learning all entailment relations between all propositional template pairs for that graph. To learn edges we consider global constraints, which allow only certain graph topologies. Since we seek a global solution under transitivity and other constraints, linear programming is a natural choice, enabling the use of state of the art optimization packages. We describe two formulations of integer linear programs that learn the edges: one maximizing a global score function, and another maximizing a global probability function. Let Iuv be an indicator denoting the event that node u entails node v. Our goal is to learn the edges E over a set of nodes V . We start by formulating the constraints and then the target functions. The first constraint is that the graph must respect transitivity. Our formulation is equivalent to the one suggested by Finkel and Manning (2008) in a coreference resolution task: ∀u,v,w∈V Iuv + Ivw −Iuw ≤1 In addition, for a few pairs of nodes we have strong evidence that one does not entail the other and so we add the constraint Iuv = 0. Combined with the constraint of transitivity this implies that there must be no path from u to v. This is done in the following two scenarios: (1) When two nodes u and v are identical except for a pair of words wu and wv, and wu is an antonym of wv, or a hypernym of wv at distance ≥2. (2) When two nodes u and v are transitive opposites, that is, if u = X subj ←−−w obj −−→Y and v = X obj ←−−w subj −−→Y , for any word w4. Score-based target function We assume an entailment classifier estimating a positive score Suv if it believes Iuv = 1 and a negative score otherwise (for example, an SVM classifier). We look for a graph G that maximizes the sum of scores over the edges: ˆG = argmax G S(G) = argmax G X u̸=v SuvIuv −λ|E| where λ|E| is a regularization term reflecting the fact that edges are sparse. Note that this constant needs to be optimized on a development set. Probabilistic target function Let Fuv be the features for the pair of nodes (u, v) and F = ∪u̸=vFuv. We assume an entailment classifier estimating the probability of an edge given its features: Puv = P(Iuv = 1|Fuv). We look for the graph G that maximizes the posterior probability P(G|F): ˆG = argmax G P(G|F) Following Snow et al., we make two independence assumptions: First, we assume each set of features Fuv is independent of other sets of features given the graph G, i.e., P(F|G) = Q u̸=v P(Fuv|G). Second, we assume the features for the pair (u, v) are generated by a distribution depending only on whether entailment holds for (u, v). Thus, P(Fuv|G) = P(Fuv|Iuv). Last, for simplicity we assume edges are independent and the prior probability of a graph is a product of the prior probabilities of the edge indicators: 4We note that in some rare cases transitive verbs are indeed reciprocal, as in “X marry Y”, but in the grand majority of cases reciprocal activities are not expressed using a transitive-verb structure. 1224 P(G) = Q u̸=v P(Iuv). Note that although we assume edges are independent, dependency is still expressed using the transitivity constraint. We express P(G|F) using the assumptions above and Bayes rule: P(G|F) ∝P(G)P(F|G) = Y u̸=v [P(Iuv)P(Fuv|Iuv)] = Y u̸=v P(Iuv)P(Iuv|Fuv)P(Fuv) P(Iuv) ∝ Y u̸=v P(Iuv|Fuv) = Y (u,v)∈E Puv · Y (u,v)/∈E (1 −Puv) Note that the prior P(Fuv) is constant with respect to the graph. Now we look for the graph that maximizes log P(G|F): ˆG = argmax G X (u,v)∈E log Puv + X (u,v)/∈E log(1 −Puv) = argmax G X u̸=v [Iuv · log Puv + (1 −Iuv) · log(1 −Puv)] = argmax G X u̸=v log Puv 1 −Puv · Iuv (in the last transition we omit the constant P u̸=v log(1−Puv)). Importantly, while the scorebased formulation contains a parameter λ that requires optimization, this probabilistic formulation is parameter free and does not utilize a development set at all. Since the variables are binary, both formulations are integer linear programs with O(|V |2) variables and O(|V |3) transitivity constraints that can be solved using standard ILP packages. Our work resembles Snow et al.’s in that both try to learn graph edges given a transitivity constraint. However, there are two key differences in the model and in the optimization algorithm. First, Snow et al.’s model attempts to determine the graph that maximizes the likelihood P(F|G) and not the posterior P(G|F). Therefore, their model contains an edge prior P(Iuv) that has to be estimated, whereas in our model it cancels out. Second, they incrementally add hyponyms to a large taxonomy (WordNet) and therefore utilize a greedy algorithm, while we simultaneously learn all edges of a rather small graph and employ integer linear programming, which is more sound theoretically, and as shown in Section 6, leads to an optimal solution. Nevertheless, Snow et al.’s model can also be formulated as a linear program with the following target function: argmax G X u̸=v log Puv · P(Iuv = 0) (1 −Puv) · P(Iuv = 1)Iuv Note that if the prior inverse odds k = P(Iuv=0) P(Iuv=1) = 1, i.e., P(Iuv = 1) = 0.5, then this is equivalent to our probabilistic formulation. We implemented Snow et al’s model and optimization algorithm and in Section 6.3 we compare our model and optimization algorithm to theirs. 6 Experimental Evaluation This section presents our evaluation, which is geared for the application proposed in Section 4. 6.1 Experimental setting A health-care corpus of 632MB was harvested from the web and parsed with the Minipar parser (Lin, 1998). The corpus contains 2,307,585 sentences and almost 50 million word tokens. We used the Unified Medical Language System (UMLS)5 to annotate medical concepts in the corpus. The UMLS is a database that maps natural language phrases to over one million concept identifiers in the health-care domain (termed CUIs). We annotated all nouns and noun phrases that are in the UMLS with their possibly multiple CUIs. We extracted all propositional templates from the corpus, where both argument instantiations are medical concepts, i.e., annotated with a CUI (∼50,000 templates). When computing distributional similarity scores, a template is represented as a feature vector of the CUIs that instantiate its arguments. To evaluate the performance of our algorithm, we constructed 23 gold standard entailment graphs. First, 23 medical concepts, representing typical topics of interest in the medical domain, were manually selected from a list of the most frequent concepts in the corpus. For each concept, nodes were defined by extracting all propositional 5http://www.nlm.nih.gov/research/umls 1225 Using a development set Not using a development set Edges Propositions Edges Propositions R P F1 R P F1 R P F1 R P F1 LP 46.0 50.1 43.8 67.3 69.6 66.2 48.7 41.9 41.2 67.9 62.0 62.3 Greedy 45.7 37.1 36.6 64.2 57.2 56.3 48.2 41.7 41.0 67.8 62.0 62.4 Local-LP 44.5 45.3 38.1 65.2 61.0 58.6 69.3 19.7 26.8 82.7 33.3 42.6 Local1 53.5 34.9 37.5 73.5 50.6 56.1 92.9 11.1 19.7 95.4 18.6 30.6 Local2 52.5 31.6 37.7 69.8 50.0 57.1 63.2 24.9 33.6 77.7 39.3 50.5 Local∗ 1 53.5 38.0 39.8 73.5 54.6 59.1 92.6 11.3 20.0 95.3 18.9 31.1 Local∗ 2 52.5 32.1 38.1 69.8 50.6 57.4 63.1 25.5 34.0 77.7 39.9 50.9 WordNet 10.8 44.1 13.2 39.9 72.4 47.3 Table 1: Results for all experiments templates for which the target concept instantiated an argument at least K(= 3) times (average number of graph nodes=22.04, std=3.66, max=26, min=13). Ten medical students constructed the gold standard of graph edges. Each concept graph was annotated by two students. Following RTE-5 practice (Bentivogli et al., 2009), after initial annotation the two students met for a reconciliation phase. They worked to reach an agreement on differences and corrected their graphs. Inter-annotator agreement was calculated using the Kappa statistic (Siegel and Castellan, 1988) both before (κ = 0.59) and after (κ = 0.9) reconciliation. 882 edges were included in the 23 graphs out of a possible 10,364, providing a sufficiently large data set. The graphs were randomly split into a development set (11 graphs) and a test set (12 graphs)6. The entailment graph fragment in Figure 1 is from the gold standard. The graphs learned by our algorithm were evaluated by two measures, one evaluating the graph directly, and the other motivated by our application: (1) F1 of the learned edges compared to the gold standard edges (2) Our application provides a summary of propositions extracted from the corpus. Note that we infer new propositions by propagating inference transitively through the graph. Thus, we compute F1 for the set of propositions inferred from the learned graph, compared to the set inferred based on the gold standard graph. For example, given the proposition from the corpus ‘relaxation reduces nausea’ and the edge ‘X reduce nausea →X help with nausea’, we evaluate the set {‘relaxation reduces nausea’, ‘relaxation helps with nausea’}. The final score for an algorithm is a macro-average over the 12 graphs of the 6Test set concepts were: asthma, chemotherapy, diarrhea, FDA, headache, HPV, lungs, mouth, salmonella, seizure, smoking and X-ray. test set. 6.2 Evaluated algorithms Local algorithms We described 12 distributional similarity measures computed over our corpus (Section 5.1). For each measure we computed for each template t a list of templates most similar to t (or entailing t for directional measures). In addition, we obtained similarity lists learned by Lin and Pantel (2001), and replicated 3 similarity measures learned by Szpektor and Dagan (2008), over the RCV1 corpus7. For each distributional similarity measure (altogether 16 measures), we learned a graph by inserting any edge (u, v), when u is in the top K templates most similar to v. We also omitted edges for which there was strong evidence that they do not exist, as specified by the constraints in Section 5.2. Another local resource was WordNet where we inserted an edge (u, v) when v was a direct hypernym or synonym of u. For all algorithms, we added all edges inferred by transitivity. Global algorithms We experimented with all 6 combinations of the following two dimensions: (1) Target functions: score-based, probabilistic and Snow et al.’s (2) Optimization algorithms: Snow et al.’s greedy algorithm and a standard ILP solver. A training set of 20,144 examples was automatically generated, each example represented by 16 features using the distributional similarity measures mentioned above. SVMperf (Joachims, 2005) was used to train an SVM classifier yielding Suv, and the SMO classifier from WEKA (Hall et al., 2009) estimated Puv. We used the lpsolve8 package to solve the linear programs. In all results, the relaxation ∀u,v0 ≤Iuv ≤1 was used, which guarantees an optimal output solution. In 7http://trec.nist.gov/data/reuters/reuters.html. The similarity lists were computed using: (1) Unary templates and the Lin function (2) Unary templates and the BInc function (3) Binary templates and the Lin function 8http://lpsolve.sourceforge.net/5.5/ 1226 Global=T/Local=F Global=F/Local=T GS= T 50 143 GS= F 140 1087 Table 2: Comparing disagreements between the best local and global algorithms against the gold standard all experiments the output solution was integer, and therefore it is optimal. Constructing graph nodes and learning its edges given an input concept took 2-3 seconds on a standard desktop. 6.3 Results and analysis Table 1 summarizes the results of the algorithms. The left half depicts methods where the development set was needed to tune parameters, and the right half depicts methods that do not require a (manually created) development set at all. Hence, our score-based LP (tuned-LP), where the parameter λ is tuned, is on the left, and the probabilistic LP (untuned-LP) is on the right. The row Greedy is achieved by using the greedy algorithm instead of lpsolve. The row Local-LP is achieved by omitting global transitivity constraints, making the algorithm completely local. We omit Snow et al.’s formulation, since the optimal prior inverse odds k was almost exactly 1, which conflates with untuned-LP. The rows Local1 and Local2 present the best distributional similarity resources. Local1 is achieved using binary templates, the Lin function, and a single vector with feature pairs. Local2 is identical but employs the BInc function. Local∗ 1 and Local∗ 2 also exploit the local constraints mentioned above. Results on the left were achieved by optimizing the top-K parameter on the development set, and on the right by optimizing on the training set automatically generated from WordNet. The global methods clearly outperform local methods: Tuned-LP outperforms significantly all local methods that require a development set both on the edges F1 measure (p<.05) and on the propositions F1 measure (p<.01)9. The untunedLP algorithm also significantly outperforms all local methods that do not require a development set on the edges F1 measure (p<.05) and on the propositions F1 measure (p<.01). Omitting the global transitivity constraints decreases performance, as shown by Local-LP. Last, local meth9We tested significance using the two-sided Wilcoxon rank test (Wilcoxon, 1945) Global X-treat-headache X-prevent-headache X-reduce-headache X-report-headache X-suffer-from-headache X-experience-headache Figure 2: Subgraph of tuned-LP output for “headache” Global X-treat-headache X-prevent-headache X-reduce-headache X-report-headache X-suffer-from-headache X-experience-headache Figure 3: Subgraph of Local∗ 1 output for“headache” ods are sensitive to parameter tuning and in the absence of a development set their performance dramatically deteriorates. To further establish the merits of global algorithms, we compare (Table 2) tuned-LP, the best global algorithm, with Local∗ 1, the best local algorithm. The table considers all edges where the two algorithms disagree, and counts how many are in the gold standard and how many are not. Clearly, tuned-LP is superior at avoiding wrong edges (false positives). This is because tunedLP refrains from adding edges that subsequently induce many undesirable edges through transitivity. Figures 2 and 3 illustrate this by comparing tuned-LP and Local∗ 1 on a subgraph of the Headache concept, before adding missing edges to satisfy transitivity to Local∗ 1 . Note that Local∗ 1 inserts a single wrong edge X-report-headache → X-prevent-headache, which leads to adding 8 more wrong edges. This is the type of global consideration that is addressed in an ILP formulation, but is ignored in a local approach and often overlooked when employing a greedy algorithm. Figure 2 also illustrates the utility of a local entailment graph for information presentation. Presenting information according to this subgraph distinguishes between propositions dealing with headache treatments and 1227 propositions dealing with headache risk groups. Comparing our use of an ILP algorithm to the greedy one reveals that tuned-LP significantly outperforms its greedy counterpart on both measures (p<.01). However, untuned-LP is practically equivalent to its greedy counterpart. This indicates that in this experiment the greedy algorithm provides a good approximation for the optimal solution achieved by our LP formulation. Last, when comparing WordNet to local distributional similarity methods, we observe low recall and high precision, as expected. However, global methods achieve much higher recall than WordNet while maintaining comparable precision. The results clearly demonstrate that a global approach improves performance on the entailment graph learning task, and the overall advantage of employing an ILP solver rather than a greedy algorithm. 7 Conclusion This paper presented a global optimization algorithm for learning entailment relations between predicates represented as propositional templates. We modeled the problem as a graph learning problem, and searched for the best graph under a global transitivity constraint. We used Integer Linear Programming to solve the optimization problem, which is theoretically sound, and demonstrated empirically that this method outperforms local algorithms as well as a greedy optimization algorithm on the graph learning task. Currently, we are investigating a generalization of our probabilistic formulation that includes a prior on the edges, and the relation of this prior to the regularization term introduced in our scorebased formulation. In future work, we would like to learn general entailment graphs over a large number of nodes. This will introduce a challenge to our current optimization algorithm due to complexity issues, and will require careful handling of predicate ambiguity. Additionally, we will investigate novel features for the entailment classifier. This paper used distributional similarity, but other sources of information are likely to improve performance further. Acknowledgments We would like to thank Roy Bar-Haim, David Carmel and the anonymous reviewers for their useful comments. We also thank Dafna Berant and the nine students who prepared the gold standard data set. This work was developed under the collaboration of FBK-irst/University of Haifa and was partially supported by the Israel Science Foundation grant 1112/08. The first author is grateful to the Azrieli Foundation for the award of an Azrieli Fellowship, and has carried out this research in partial fulllment of the requirements for the Ph.D. degree. References Luisa Bentivogli, Ido Dagan, Hoa Trang Dang, Danilo Giampiccolo, and Bernarde Magnini. 2009. The fifth Pascal recognizing textual entailment challenge. In Proceedings of TAC-09. Rahul Bhagat, Patrick Pantel, and Eduard Hovy. 2007. LEDIR: An unsupervised algorithm for learning directionality of inference rules. In Proceedings of EMNLP-CoNLL. Alexander Budanitsky and Graeme Hirst. 2006. Evaluating wordnet-based measures of lexical semantic relatedness. Computational Linguistics, 32(1):13– 47. James Clarke and Mirella Lapata. 2008. Global inference for sentence compression: An integer linear programming approach. Journal of Artificial Intelligence Research, 31:273–381. Michael Connor and Dan Roth. 2007. Context sensitive paraphrasing with a single unsupervised classifier. In Proceedings of ECML. Ido Dagan, Bill Dolan, Bernardo Magnini, and Dan Roth. 2009. Recognizing textual entailment: Rational, evaluation and approaches. Natural Language Engineering, 15(4):1–17. Christiane Fellbaum, editor. 1998. WordNet: An Electronic Lexical Database (Language, Speech, and Communication). The MIT Press. Jenny Rose Finkel and Christopher D. Manning. 2008. Enforcing transitivity in coreference resolution. In Proceedings of ACL-08: HLT, Short Papers. Mark Hall, Eibe Frank, Geoffrey Holmes, Bernhard Pfahringer, Peter Reutemann, and Ian H. Witten. 2009. The WEKA data mining software: An update. SIGKDD Explorations, 11(1). Thomas Hofmann. 1999. The cluster-abstraction model: Unsupervised learning of topic hierarchies from text data. In Proceedings of IJCAI. Thorsten Joachims. 2005. A support vector method for multivariate performance measures. In Proceedings of ICML. 1228 Mika Kaki. 2005. Findex: Search results categories help users when document ranking fails. In Proceedings of CHI. Dekang Lin and Patrick Pantel. 2001. Discovery of inference rules for question answering. Natural Language Engineering, 7(4):343–360. Dekang Lin. 1998. Dependency-based evaluation of Minipar. In Proceedings of the Workshop on Evaluation of Parsing Systems at LREC. Andre Martins, Noah Smith, and Eric Xing. 2009. Concise integer linear programming formulations for dependency parsing. In Proceedings of ACL. Vladimir Nikulin. 2008. Classification of imbalanced data with random sets and mean-variance filtering. IJDWM, 4(2):63–78. Dan Roth and Wen tau Yih. 2005. Integer linear programming inference for conditional random fields. In Proceedings of ICML, pages 737–744. Satoshi Sekine. 2005. Automatic paraphrase discovery based on context and keywords between ne pairs. In Proceedings of IWP. Sideny Siegel and N. John Castellan. 1988. Nonparametric Statistics for the Behavioral Sciences. McGraw-Hill, New-York. Noah Smith and Jason Eisner. 2005. Contrastive estimation: Training log-linear models on unlabeled data. In Proceedings of ACL. Rion Snow, Daniel Jurafsky, and Andrew Y. Ng. 2005. Learning syntactic patterns for automatic hypernym discovery. In Proceedings of NIPS. Rion Snow, Daniel Jurafsky, and Andrew Y. Ng. 2006. Semantic taxonomy induction from heterogenous evidence. In Proceedings of ACL. Emilia Stoica, Marti Hearst, and Megan Richardson. 2007. Automating creation of hierarchical faceted metadata structures. In Proceedings of NAACLHLT. Idan Szpektor and Ido Dagan. 2008. Learning entailment rules for unary templates. In Proceedings of COLING. Idan Szpektor and Ido Dagan. 2009. Augmenting wordnet-based inference with argument mapping. In Proceedings of TextInfer-2009. Idan Szpektor, Hristo Tanev, Ido Dagan, and Bonaventura Coppola. 2004. Scaling web-based acquisition of entailment relations. In Proceedings of EMNLP. Jason Van Hulse, Taghi Khoshgoftaar, and Amri Napolitano. 2007. Experimental perspectives on learning from imbalanced data. In Proceedings of ICML. Frank Wilcoxon. 1945. Individual comparisons by ranking methods. Biometrics Bulletin, 1:80–83. Alexander Yates and Oren Etzioni. 2009. Unsupervised methods for determining object and relation synonyms on the web. Journal of Artificial Intelligence Research, 34:255–296. 1229
|
2010
|
124
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1230–1238, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Modeling Semantic Relevance for Question-Answer Pairs in Web Social Communities Baoxun Wang, Xiaolong Wang, Chengjie Sun, Bingquan Liu, Lin Sun School of Computer Science and Technology Harbin Institute of Technology Harbin, China {bxwang, wangxl, cjsun, liubq, lsun}@insun.hit.edu.cn Abstract Quantifying the semantic relevance between questions and their candidate answers is essential to answer detection in social media corpora. In this paper, a deep belief network is proposed to model the semantic relevance for question-answer pairs. Observing the textual similarity between the community-driven questionanswering (cQA) dataset and the forum dataset, we present a novel learning strategy to promote the performance of our method on the social community datasets without hand-annotating work. The experimental results show that our method outperforms the traditional approaches on both the cQA and the forum corpora. 1 Introduction In natural language processing (NLP) and information retrieval (IR) fields, question answering (QA) problem has attracted much attention over the past few years. Nevertheless, most of the QA researches mainly focus on locating the exact answer to a given factoid question in the related documents. The most well known international evaluation on the factoid QA task is the Text REtrieval Conference (TREC)1, and the annotated questions and answers released by TREC have become important resources for the researchers. However, when facing a non-factoid question such as why, how, or what about, however, almost no automatic QA systems work very well. The user-generated question-answer pairs are definitely of great importance to solve the nonfactoid questions. Obviously, these natural QA pairs are usually created during people’s communication via Internet social media, among which we are interested in the community-driven 1http://trec.nist.gov question-answering (cQA) sites and online forums. The cQA sites (or systems) provide platforms where users can either ask questions or deliver answers, and best answers are selected manually (e.g., Baidu Zhidao2 and Yahoo! Answers3). Comparing with cQA sites, online forums have more virtual society characteristics, where people hold discussions in certain domains, such as techniques, travel, sports, etc. Online forums contain a huge number of QA pairs, and much noise information is involved. To make use of the QA pairs in cQA sites and online forums, one has to face the challenging problem of distinguishing the questions and their answers from the noise. According to our investigation, the data in the community based sites, especially for the forums, have two obvious characteristics: (a) a post usually includes a very short content, and when a person is initializing or replying a post, an informal tone tends to be used; (b) most of the posts are useless, which makes the community become a noisy environment for question-answer detection. In this paper, a novel approach for modeling the semantic relevance for QA pairs in the social media sites is proposed. We concentrate on the following two problems: 1. How to model the semantic relationship between two short texts using simple textual features? As mentioned above, the user generated questions and their answers via social media are always short texts. The limitation of length leads to the sparsity of the word features. In addition, the word frequency is usually either 0 or 1, that is, the frequency offers little information except the occurrence of a word. Because of this situation, the traditional relevance computing methods based on word co-occurrence, such as Cosine similarity and KL-divergence, are not effective for question2http://zhidao.baidu.com 3http://answers.yahoo.com 1230 answer semantic modeling. Most researchers try to introduce structural features or users’ behavior to improve the models performance, by contrast, the effect of textual features is not obvious. 2. How to train a model so that it has good performance on both cQA and forum datasets? So far, people have been doing QA researches on the cQA and the forum datasets separately (Ding et al., 2008; Surdeanu et al., 2008), and no one has noticed the relationship between the two kinds of data. Since both the cQA systems and the online forums are open platforms for people to communicate, the QA pairs in the cQA systems have similarity with those in the forums. In this case, it is highly valuable and desirable to propose a training strategy to improve the model’s performance on both of the two kinds of datasets. In addition, it is possible to avoid the expensive and arduous hand-annotating work by introducing the method. To solve the first problem, we present a deep belief network (DBN) to model the semantic relevance between questions and their answers. The network establishes the semantic relationship for QA pairs by minimizing the answer-to-question reconstructing error. Using only word features, our model outperforms the traditional methods on question-answer relevance calculating. For the second problem, we make our model to learn the semantic knowledge from the solved question threads in the cQA system. Instead of mining the structure based features from cQA pages and forum threads individually, we consider the textual similarity between the two kinds of data. The semantic information learned from cQA corpus is helpful to detect answers in forums, which makes our model show good performance on social media corpora. Thanks to the labels for the best answers existing in the threads, no manual work is needed in our strategy. The rest of this paper is organized as follows: Section 2 surveys the related work. Section 3 introduces the deep belief network for answer detection. In Section 4, the homogenous data based learning strategy is described. Experimental result is given in Section 5. Finally, conclusions and future directions are drawn in Section 6. 2 Related Work The value of the naturally generated questionanswer pairs has not been recognized until recent years. Early studies mainly focus on extracting QA pairs from frequently asked questions (FAQ) pages (Jijkoun and de Rijke, 2005; Riezler et al., 2007) or service call-center dialogues (Berger et al., 2000). Judging whether a candidate answer is semantically related to the question in the cQA page automatically is a challenging task. A framework for predicting the quality of answers has been presented in (Jeon et al., 2006). Bernhard and Gurevych (2009) have developed a translation based method to find answers. Surdeanu et al. (2008) propose an approach to rank the answers retrieved by Yahoo! Answers. Our work is partly similar to Surdeanu et al. (2008), for we also aim to rank the candidate answers reasonably, but our ranking algorithm needs only word information, instead of the combination of different kinds of features. Because people have considerable freedom to post on forums, there are a great number of irrelevant posts for answering questions, which makes it more difficult to detect answers in the forums. In this field, exploratory studies have been done by Feng et al. (2006) and Huang et al. (2007), who extract input-reply pairs for the discussion-bot. Ding et al.(2008) and Cong et al.(2008) have also presented outstanding research works on forum QA extraction. Ding et al. (2008) detect question contexts and answers using the conditional random fields, and a ranking algorithm based on the authority of forum users is proposed by Cong et al. (2008). Treating answer detection as a binary classification problem is an intuitive idea, thus there are some studies trying to solve it from this view (Hong and Davison, 2009; Wang et al., 2009). Especially Hong and Davison (2009) have achieved a rather high precision on the corpora with less noise, which also shows the importance of “social” features. In order to select the answers for a given question, one has to face the problem of lexical gap. One of the problems with lexical gap embedding is to find similar questions in QA achieves (Jeon et al., 2005). Recently, the statistical machine translation (SMT) strategy has become popular. Lee et al. (2008) use translate models to bridge the lexical gap between queries and questions in QA collections. The SMT based methods are effective on modeling the semantic relationship between questions and answers and expending users’ queries in answer retrieval (Riezler et al., 2007; Berger et al., 1231 2000; Bernhard and Gurevych, 2009). In (Surdeanu et al., 2008), the translation model is used to provide features for answer ranking. The structural features (e.g., authorship, acknowledgement, post position, etc), also called non-textual features, play an important role in answer extraction. Such features are used in (Ding et al., 2008; Cong et al., 2008), and have significantly improved the performance. The studies of Jeon et al. (2006) and Hong et al. (2009) show that the structural features have even more contribution than the textual features. In this case, the mining of textual features tends to be ignored. There are also some other research topics in this field. Cong et al. (2008) and Wang et al. (2009) both propose the strategies to detect questions in the social media corpus, which is proved to be a non-trivial task. The deep research on question detection has been taken by Duan et al. (2008). A graph based algorithm is presented to answer opinion questions (Li et al., 2009). In email summarization field, the QA pairs are also extracted from email contents as the main elements of email summarization (Shrestha and McKeown, 2004). 3 The Deep Belief Network for QA pairs Due to the feature sparsity and the low word frequency of the social media corpus, it is difficult to model the semantic relevance between questions and answers using only co-occurrence features. It is clear that the semantic link exists between the question and its answers, even though they have totally different lexical representations. Thus a specially designed model may learn semantic knowledge by reconstructing a great number of questions using the information in the corresponding answers. In this section, we propose a deep belief network for modeling the semantic relationship between questions and their answers. Our model is able to map the QA data into a low-dimensional semantic-feature space, where a question is close to its answers. 3.1 The Restricted Boltzmann Machine An ensemble of binary vectors can be modeled using a two-layer network called a “restricted Boltzmann machine” (RBM) (Hinton, 2002). The dimension reducing approach based on RBM initially shows good performance on image processing (Hinton and Salakhutdinov, 2006). Salakhutdinov and Hinton (2009) propose a deep graphical model composed of RBMs into the information retrieval field, which shows that this model is able to obtain semantic information hidden in the wordcount vectors. As shown in Figure 1, the RBM is a two-layer network. The bottom layer represents a visible vector v and the top layer represents a latent feature h. The matrix W contains the symmetric interaction terms between the visible units and the hidden units. Given an input vector v, the trained Figure 1: Restricted Boltzmann machine RBM model provides a hidden feature h, which can be used to reconstruct v with a minimum error. The training algorithm for this paper will be described in the next subsection. The ability of the RBM suggests us to build a deep belief network based on RBM so that the semantic relevance between questions and answers can be modeled. 3.2 Pretraining a Deep Belief Network In the social media corpora, the answers are always descriptive, containing one or several sentences. Noticing that an answer has strong semantic association with the question and involves more information than the question, we propose to train a deep belief network by reconstructing the question using its answers. The training object is to minimize the error of reconstruction, and after the pretraining process, a point that lies in a good region of parameter space can be achieved. Firstly, the illustration of the DBN model is given in Figure 2. This model is composed of three layers, and here each layer stands for the RBM or its variant. The bottom layer is a variant form of RBM’s designed for the QA pairs. This layer we design is a little different from the classical RBM’s, so that the bottom layer can generate the hidden features according to the visible answer vector and reconstruct the question vector using the hidden features. The pre-training procedure of this architecture is practically convergent. In the bottom layer, the binary feature vectors based on the statistics of the word occurrence in the answers are used to compute the “hidden features” in the 1232 Figure 2: The Deep Belief Network for QA Pairs hidden units. The model can reconstruct the questions using the hidden features. The processes can be modeled as follows: p(h j = 1|a) = σ(b j + X i wijai) (1) p(qi = 1|h) = σ(bi + X j wijhj) (2) where σ(x) = 1/(1 + e−x), a denotes the visible feature vector of the answer, qi is the ith element of the question vector, and h stands for the hidden feature vector for reconstructing the questions. wij is a symmetric interaction term between word i and hidden feature j, bi stands for the bias of the model for word i, and bj denotes the bias of hidden feature j. Given the training set of answer vectors, the bottom layer generates the corresponding hidden features using Equation 1. Equation 2 is used to reconstruct the Bernoulli rates for each word in the question vectors after stochastically activating the hidden features. Then Equation 1 is taken again to make the hidden features active. We use 1-step Contrastive Divergence (Hinton, 2002) to update the parameters by performing gradient ascent: ∆wij = ϵ(< qihj >qData −< qihj >qRecon) (3) where < qih j >qData denotes the expectation of the frequency with which the word i in a question and the feature j are on together when the hidden features are driven by the question data. < qih j >qRecon defines the corresponding expectation when the hidden features are driven by the reconstructed question data. ϵ is the learning rate. The classical RBM structure is taken to build the middle layer and the top layer of the network. The training method for the higher two layer is similar to that of the bottom one, and we only have to make each RBM to reconstruct the input data using its hidden features. The parameter updates still obeying the rule defined by gradient ascent, which is quite similar to Equation 3. After training one layer, the h vectors are then sent to the higher-level layer as its “training data”. 3.3 Fine-tuning the Weights Notice that a greedy strategy is taken to train each layer individually during the pre-training procedure, it is necessary to fine-tune the weights of the entire network for optimal reconstruction. To finetune the weights, the network is unrolled, taking the answers as the input data to generate the corresponding questions at the output units. Using the cross-entropy error function, we can then tune the network by performing backpropagation through it. The experiment results in section 5.2 will show fine-tuning makes the network performs better for answer detection. 3.4 Best answer detection After pre-training and fine-tuning, a deep belief network for QA pairs is established. To detect the best answer to a given question, we just have to send the vectors of the question and its candidate answers into the input units of the network and perform a level-by-level calculation to obtain the corresponding feature vectors. Then we calculate the distance between the mapped question vector and each candidate answer vector. We consider the candidate answer with the smallest distance as the best one. 4 Learning with Homogenous Data In this section, we propose our strategy to make our DBN model to detect answers in both cQA and forum datasets, while the existing studies focus on one single dataset. 4.1 Homogenous QA Corpora from Different Sources Our motivation of finding the homogenous question-answer corpora from different kind of social media is to guarantee the model’s performance and avoid hand-annotating work. In this paper, we get the “solved question” pages in the computer technology domain from Baidu Zhidao as the cQA corpus, and the threads of 1233 Figure 3: Comparison of the post content lengths in the cQA and the forum datasets ComputerFansClub Forum4 as the online forum corpus. The domains of the corpora are the same. To further explain that the two corpora are homogenous, we will give the detail comparison on text style and word distribution. As shown in Figure 3, we have compared the post content lengths of the cQA and the forum in our corpora. For the comparison, 5,000 posts from the cQA corpus and 5,000 posts from the forum corpus are randomly selected. The left panel shows the statistical result on the Baidu Zhidao data, and the right panel shows the one on the forum data. The number i on the horizontal axis denotes the post contents whose lengths range from 10(i −1) + 1 to 10i bytes, and the vertical axis represents the counts of the post contents. From Figure 3 we observe that the contents of most posts in both the cQA corpus and the forum corpus are short, with the lengths not exceeding 400 bytes. The content length reflects the text style of the posts in cQA systems and online forums. From Figure 3 it can be also seen that the distributions of the content lengths in the two figures are very similar. It shows that the contents in the two corpora are both mainly short texts. Figure 4 shows the percentage of the concurrent words in the top-ranked content words with high frequency. In detail, we firstly rank the words by frequency in the two corpora. The words are chosen based on a professional dictionary to guarantee that they are meaningful in the computer knowledge field. The number k on the horizontal axis in Figure 4 represents the top k content words in the 4http://bbs.cfanclub.net/ corpora, and the vertical axis stands for the percentage of the words shared by the two corpora in the top k words. Figure 4: Distribution of concurrent content words Figure 4 shows that a large number of meaningful words appear in both of the two corpora with high frequencies. The percentage of the concurrent words maintains above 64% in the top 1,400 words. It indicates that the word distributions of the two corpora are quite similar, although they come from different social media sites. Because the cQA corpus and the forum corpus used in this study have homogenous characteristics for answer detecting task, a simple strategy may be used to avoid the hand-annotating work. Apparently, in every “solved question” page of Baidu Zhidao, the best answer is selected by the user who asks this question. We can easily extract the QA pairs from the cQA corpus as the training 1234 set. Because the two corpora are similar, we can apply the deep belief network trained by the cQA corpus to detect answers on both the cQA data and the forum data. 4.2 Features The task of detecting answers in social media corpora suffers from the problem of feature sparsity seriously. High-dimensional feature vectors with only several non-zero dimensions bring large time consumption to our model. Thus it is necessary to reduce the dimension of the feature vectors. In this paper, we adopt two kinds of word features. Firstly, we consider the 1,300 most frequent words in the training set as Salakhutdinov and Hinton (2009) did. According to our statistics, the frequencies of the rest words are all less then 10, which are not statistically significant and may introduce much noise. We take the occurrence of some function words as another kind of features. The function words are quite meaningful for judging whether a short text is an answer or not, especially for the nonfactoid questions. For example, in the answers to the causation questions, the words such as because and so are more likely to appear; and the words such as firstly, then, and should may suggest the answers to the manner questions. We give an example for function word selection in Figure 5. Figure 5: An example for function word selection For this reason, we collect 200 most frequent function words in the answers of the training set. Then for every short text, either a question or an answer, a 1,500-dimensional vector can be generated. Specifically, all the features we have adopted are binary, for they only have to denote whether the corresponding word appears in the text or not. 5 Experiments To evaluate our question-answer semantic relevance computing method, we compare our approach with the popular methods on the answer detecting task. 5.1 Experiment Setup Architecture of the Network: To build the deep belief network, we use a 1500-1500-1000-600 architecture, which means the three layers of the network have individually 1,500×1,500, 1,500×1,000 and 1,000×600 units. Using the network, a 1,500dimensional binary vector is finally mapped to a 600-dimensional real-value vector. During the pretraining stage, the bottom layer is greedily pretrained for 200 passes through the entire training set, and each of the rest two layers is greedily pretrained for 50 passes. For fine-tuning we apply the method of conjugate gradients5, with three line searches performed in each pass. This algorithm is performed for 50 passes to fine-tune the network. Dataset: we have crawled 20,000 pages of “solved question” from the computer and network category of Baidu Zhidao as the cQA corpus. Correspondingly we obtain 90,000 threads from ComputerFansClub, which is an online forum on computer knowledge. We take the forum threads as our forum corpus. From the cQA corpus, we extract 12,600 human generated QA pairs as the training set without any manual work to label the best answers. We get the contents from another 2,000 cQA pages to form a testing set, each content of which includes one question and 4.5 candidate answers on average, with one best answer among them. To get another testing dataset, we randomly select 2,000 threads from the forum corpus. For this training set, human work are necessary to label the best answers in the posts of the threads. There are 7 posts included in each thread on average, among which one question and at least one answer exist. Baseline: To show the performance of our method, three main popular relevance computing methods for ranking candidate answers are considered as our baselines. We will briefly introduce them: Cosine Similarity. Given a question q and its candidate answer a, their cosine similarity can be computed as follows: cos(q, a) = Pn k=1 wqk × wak qPn k=1 w2qk × qPn k=1 w2ak (4) where wqk and wak stand for the weight of the kth word in the question and the answer respectively. 5Code is available at http://www.kyb.tuebingen.mpg.de/bs/people/carl/code/minimize/ 1235 The weights can be get by computing the product of term frequency (tf) and inverse document frequency (idf) HowNet based Similarity. HowNet6 is an electronic world knowledge system, which serves as a powerful tool for meaning computation in human language technology. Normally the similarity between two passages can be calculated by two steps: (1) matching the most semantic-similar words in each passages greedily using the API’s provided by HowNet; (2) computing the weighted average similarities of the word pairs. This strategy is taken as a baseline method for computing the relevance between questions and answers. KL-divergence Language Model. Given a question q and its candidate answer a, we can construct unigram language model Mq and unigram language model Ma. Then we compute KLdivergence between Mq and Ma as below: KL(Ma||Mq) = X w p(w|Ma) log(p(w|Ma)/p(w|Mq)) (5) 5.2 Results and Analysis We evaluate the performance of our approach for answer detection using two metrics: Precision@1 (P@1) and Mean Reciprocal Rank (MRR). Applying the two metrics, we perform the baseline methods and our DBN based methods on the two testing set above. Table 1 lists the results achieved on the forum data using the baseline methods and ours. The additional “Nearest Answer” stands for the method without any ranking strategies, which returns the nearest candidate answer from the question by position. To illustrate the effect of the fine-tuning for our model, we list the results of our method without fine-tuning and the results with fine-tuning. As shown in Table 1, our deep belief network based methods outperform the baseline methods as expected. The main reason for the improvements is that the DBN based approach is able to learn semantic relationship between the words in QA pairs from the training set. Although the training set we offer to the network comes from a different source (the cQA corpus), it still provide enough knowledge to the network to perform better than the baseline methods. This phenomena indicates that the homogenous corpora for training is 6Detail information can be found in: http://www.keenage.com/ effective and meaningful. Method P@1 (%) MRR (%) Nearest Answer 21.25 38.72 Cosine Similarity 23.15 43.50 HowNet 22.55 41.63 KL divergence 25.30 51.40 DBN (without FT) 41.45 59.64 DBN (with FT) 45.00 62.03 Table 1: Results on Forum Dataset We have also investigated the reasons for the unsatisfying performance of the baseline approaches. Basically, the low precision is ascribable to the forum corpus we have obtained. As mentioned in Section 1, the contents of the forum posts are short, which leads to the sparsity of the features. Besides, when users post messages in the online forums, they are accustomed to be casual and use some synonymous words interchangeably in the posts, which is believed to be a significant situation in Chinese forums especially. Because the features for QA pairs are quite sparse and the content words in the questions are usually morphologically different from the ones with the same meaning in the answers, the Cosine Similarity method become less powerful. For HowNet based approaches, there are a large number of words not included by HowNet, thus it fails to compute the similarity between questions and answers. KLdivergence suffers from the same problems with the Cosine Similarity method. Compared with the Cosine Similarity method, this approach has achieved the improvement of 9.3% in P@1, but it performs much better than the other baseline methods in MRR. The baseline results indicate that the online forum is a complex environment with large amount of noise for answer detection. Traditional IR methods using pure textual features can hardly achieve good results. The similar baseline results for forum answer ranking are also achieved by Hong and Davison (2009), which takes some nontextual features to improve the algorithm’s performance. We also notice that, however, the baseline methods have obtained better results on forum corpus (Cong et al., 2008). One possible reason is that the baseline approaches are suitable for their data, since we observe that the “nearest answer” strategy has obtained a 73.5% precision in their work. Our model has achieved the precision of 1236 45.00% in P@1 and 62.03% in MRR for answer detecting on forum data after fine-tuning, while some related works have reported the results with the precision over 90% (Cong et al., 2008; Hong and Davison, 2009). There are mainly two reasons for this phenomena: Firstly, both of the previous works have adopt non-textual features based on the forum structure, such as authorship, position and quotes, etc. The non-textual (or social based) features have played a significant role in improving the algorithms’ performance. Secondly, the quality of corpora influences the results of the ranking strategies significantly, and even the same algorithm may perform differently when the dataset is changed (Hong and Davison, 2009). For the experiments of this paper, large amount of noise is involved in the forum corpus and we have done nothing extra to filter it. Table 2 shows the experimental results on the cQA dataset. In this experiment, each sample is composed of one question and its following several candidate answers. We delete the ones with only one answer to confirm there are at least two candidate answers for each question. The candidate answers are rearranged by post time, so that the real answers do not always appear next to the questions. In this group of experiment, no handannotating work is needed because the real answers have been labeled by cQA users. Method P@1 (%) MRR (%) Nearest Answer 36.05 56.33 Cosine Similarity 44.05 62.84 HowNet 41.10 58.75 KL divergence 43.75 63.10 DBN (without FT) 56.20 70.56 DBN (with FT) 58.15 72.74 Table 2: Results on cQA Dataset From Table 2 we observe that all the approaches perform much better on this dataset. We attribute the improvements to the high quality QA corpus Baidu Zhidao offers: the candidate answers tend to be more formal than the ones in the forums, with less noise information included. In addition, the “Nearest Answer” strategy has reached 36.05% in P@1 on this dataset, which indicates quite a number of askers receive the real answers at the first answer post. This result has supported the idea of introducing position features. What’s more, if the best answer appear immediately, the asker tends to lock down the question thread, which helps to reduce the noise information in the cQA corpus. Despite the baseline methods’ performances have been improved, our approaches still outperform them, with a 32.0% improvement in P@1 and a 15.3% improvement in MRR at least. On the cQA dataset, our model shows better performance than the previous experiment, which is expected because the training set and the testing set come from the same corpus, and the DBN model is more adaptive to the cQA data. We have observed that, from both of the two groups of experiments, fine-tuning is effective for enhancing the performance of our model. On the forum data, the results have been improved by 8.6% in P@1 and 4.0% in MRR, and the improvements are 3.5% and 3.1% individually. 6 Conclusions In this paper, we have proposed a deep belief network based approach to model the semantic relevance for the question answering pairs in social community corpora. The contributions of this paper can be summarized as follows: (1) The deep belief network we present shows good performance on modeling the QA pairs’ semantic relevance using only word features. As a data driven approach, our model learns semantic knowledge from large amount of QA pairs to represent the semantic relevance between questions and their answers. (2) We have studied the textual similarity between the cQA and the forum datasets for QA pair extraction, and introduce a novel learning strategy to make our method show good performance on both cQA and forum datasets. The experimental results show that our method outperforms the traditional approaches on both the cQA and the forum corpora. Our future work will be carried out along two directions. Firstly, we will further improve the performance of our method by adopting the nontextual features. Secondly, more research will be taken to put forward other architectures of the deep networks for QA detection. Acknowledgments The authors are grateful to the anonymous reviewers for their constructive comments. Special thanks to Deyuan Zhang, Bin Liu, Beidong Liu and Ke Sun for insightful suggestions. This work is supported by NSFC (60973076). 1237 References Adam Berger, Rich Caruana, David Cohn, Dayne Freitag, and Vibhu Mittal. 2000. Bridging the lexical chasm: Statistical approaches to answer-finding. In In Proceedings of the 23rd annual international ACM SIGIR conference on Research and development in information retrieval, pages 192–199. Delphine Bernhard and Iryna Gurevych. 2009. Combining lexical semantic resources with question & answer archives for translation-based answer finding. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 728–736, Suntec, Singapore, August. Association for Computational Linguistics. Gao Cong, Long Wang, Chin-Yew Lin, Young-In Song, and Yueheng Sun. 2008. Finding question-answer pairs from online forums. In SIGIR ’08: Proceedings of the 31st annual international ACM SIGIR conference on Research and development in information retrieval, pages 467–474, New York, NY, USA. ACM. Shilin Ding, Gao Cong, Chin-Yew Lin, and Xiaoyan Zhu. 2008. Using conditional random fields to extract contexts and answers of questions from online forums. In Proceedings of ACL-08: HLT, pages 710–718, Columbus, Ohio, June. Association for Computational Linguistics. Huizhong Duan, Yunbo Cao, Chin-Yew Lin, and Yong Yu. 2008. Searching questions by identifying question topic and question focus. In Proceedings of ACL-08: HLT, pages 156–164, Columbus, Ohio, June. Association for Computational Linguistics. Donghui Feng, Erin Shaw, Jihie Kim, and Eduard H. Hovy. 2006. An intelligent discussion-bot for answering student queries in threaded discussions. In Ccile Paris and Candace L. Sidner, editors, IUI, pages 171–177. ACM. G. E. Hinton and R. R. Salakhutdinov. 2006. Reducing the dimensionality of data with neural networks. Science, 313(5786):504–507. Georey E. Hinton. 2002. Training products of experts by minimizing contrastive divergence. Neural Computation, 14. Liangjie Hong and Brian D. Davison. 2009. A classification-based approach to question answering in discussion boards. In SIGIR ’09: Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, pages 171–178, New York, NY, USA. ACM. Jizhou Huang, Ming Zhou, and Dan Yang. 2007. Extracting chatbot knowledge from online discussion forums. In IJCAI’07: Proceedings of the 20th international joint conference on Artifical intelligence, pages 423–428, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc. Jiwoon Jeon, W. Bruce Croft, and Joon Ho Lee. 2005. Finding similar questions in large question and answer archives. In CIKM ’05, pages 84–90, New York, NY, USA. ACM. Jiwoon Jeon, W. Bruce Croft, Joon Ho Lee, and Soyeon Park. 2006. A framework to predict the quality of answers with non-textual features. In SIGIR ’06, pages 228–235, New York, NY, USA. ACM. Valentin Jijkoun and Maarten de Rijke. 2005. Retrieving answers from frequently asked questions pages on the web. In CIKM ’05, pages 76–83, New York, NY, USA. ACM. Jung-Tae Lee, Sang-Bum Kim, Young-In Song, and Hae-Chang Rim. 2008. Bridging lexical gaps between queries and questions on large online q&a collections with compact translation models. In EMNLP ’08: Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 410–418, Morristown, NJ, USA. Association for Computational Linguistics. Fangtao Li, Yang Tang, Minlie Huang, and Xiaoyan Zhu. 2009. Answering opinion questions with random walks on graphs. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 737–745, Suntec, Singapore, August. Association for Computational Linguistics. Stefan Riezler, Alexander Vasserman, Ioannis Tsochantaridis, Vibhu Mittal, and Yi Liu. 2007. Statistical machine translation for query expansion in answer retrieval. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 464–471, Prague, Czech Republic, June. Association for Computational Linguistics. Ruslan Salakhutdinov and Geoffrey Hinton. 2009. Semantic hashing. Int. J. Approx. Reasoning, 50(7):969–978. Lokesh Shrestha and Kathleen McKeown. 2004. Detection of question-answer pairs in email conversations. In Proceedings of Coling 2004, pages 889– 895, Geneva, Switzerland, Aug 23–Aug 27. COLING. Mihai Surdeanu, Massimiliano Ciaramita, and Hugo Zaragoza. 2008. Learning to rank answers on large online QA collections. In Proceedings of ACL-08: HLT, pages 719–727, Columbus, Ohio, June. Association for Computational Linguistics. Baoxun Wang, Bingquan Liu, Chengjie Sun, Xiaolong Wang, and Lin Sun. 2009. Extracting chinese question-answer pairs from online forums. In SMC 2009: Proceedings of the IEEE International Conference on Systems, Man and Cybernetics, 2009., pages 1159–1164. 1238
|
2010
|
125
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1239–1249, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics How Many Words is a Picture Worth? Automatic Caption Generation for News Images Yansong Feng and Mirella Lapata School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB, UK [email protected], [email protected] Abstract In this paper we tackle the problem of automatic caption generation for news images. Our approach leverages the vast resource of pictures available on the web and the fact that many of them are captioned. Inspired by recent work in summarization, we propose extractive and abstractive caption generation models. They both operate over the output of a probabilistic image annotation model that preprocesses the pictures and suggests keywords to describe their content. Experimental results show that an abstractive model defined over phrases is superior to extractive methods. 1 Introduction Recent years have witnessed an unprecedented growth in the amount of digital information available on the Internet. Flickr, one of the best known photo sharing websites, hosts more than three billion images, with approximately 2.5 million images being uploaded every day.1 Many on-line news sites like CNN, Yahoo!, and BBC publish images with their stories and even provide photo feeds related to current events. Browsing and finding pictures in large-scale and heterogeneous collections is an important problem that has attracted much interest within information retrieval. Many of the search engines deployed on the web retrieve images without analyzing their content, simply by matching user queries against collocated textual information. Examples include meta-data (e.g., the image’s file name and format), user-annotated tags, captions, and generally text surrounding the image. As this limits the applicability of search engines (images that 1http://www.techcrunch.com/2008/11/03/ three-billion-photos-at-flickr/ do not coincide with textual data cannot be retrieved), a great deal of work has focused on the development of methods that generate description words for a picture automatically. The literature is littered with various attempts to learn the associations between image features and words using supervised classification (Vailaya et al., 2001; Smeulders et al., 2000), instantiations of the noisychannel model (Duygulu et al., 2002), latent variable models (Blei and Jordan, 2003; Barnard et al., 2002; Wang et al., 2009), and models inspired by information retrieval (Lavrenko et al., 2003; Feng et al., 2004). In this paper we go one step further and generate captions for images rather than individual keywords. Although image indexing techniques based on keywords are popular and the method of choice for image retrieval engines, there are good reasons for using more linguistically meaningful descriptions. A list of keywords is often ambiguous. An image annotated with the words blue, sky, car could depict a blue car or a blue sky, whereas the caption “car running under the blue sky” would make the relations between the words explicit. Automatic caption generation could improve image retrieval by supporting longer and more targeted queries. It could also assist journalists in creating descriptions for the images associated with their articles. Beyond image retrieval, it could increase the accessibility of the web for visually impaired (blind and partially sighted) users who cannot access the content of many sites in the same ways as sighted users can (Ferres et al., 2006). We explore the feasibility of automatic caption generation in the news domain, and create descriptions for images associated with on-line articles. Obtaining training data in this setting does not require expensive manual annotation as many articles are published together with captioned images. Inspired by recent work in summarization, we propose extractive and abstractive caption gen1239 eration models. The backbone for both approaches is a probabilistic image annotation model that suggests keywords for an image. We can then simply identify (and rank) the sentences in the documents that share these keywords or create a new caption that is potentially more concise but also informative and fluent. Our abstractive model operates over image description keywords and document phrases. Their combination gives rise to many caption realizations which we select probabilistically by taking into account dependency and word order constraints. Experimental results show that the model’s output compares favorably to handwritten captions and is often superior to extractive methods. 2 Related Work Although image understanding is a popular topic within computer vision, relatively little work has focused on the interplay between visual and linguistic information. A handful of approaches generate image descriptions automatically following a two-stage architecture. The picture is first analyzed using image processing techniques into an abstract representation, which is then rendered into a natural language description with a text generation engine. A common theme across different models is domain specificity, the use of handlabeled data, and reliance on background ontological information. For example, H´ede et al. (2004) generate descriptions for images of objects shot in uniform background. Their system relies on a manually created database of objects indexed by an image signature (e.g., color and texture) and two keywords (the object’s name and category). Images are first segmented into objects, their signature is retrieved from the database, and a description is generated using templates. Kojima et al. (2002, 2008) create descriptions for human activities in office scenes. They extract features of human motion and interleave them with a concept hierarchy of actions to create a case frame from which a natural language sentence is generated. Yao et al. (2009) present a general framework for generating text descriptions of image and video content based on image parsing. Specifically, images are hierarchically decomposed into their constituent visual patterns which are subsequently converted into a semantic representation using WordNet. The image parser is trained on a corpus, manually annotated with graphs representing image structure. A multi-sentence description is generated using a document planner and a surface realizer. Within natural language processing most previous efforts have focused on generating captions to accompany complex graphical presentations (Mittal et al., 1998; Corio and Lapalme, 1999; Fasciano and Lapalme, 2000; Feiner and McKeown, 1990) or on using the captions accompanying information graphics to infer their intended message, e.g., the author’s goal to convey ostensible increase or decrease of a quantity of interest (Elzer et al., 2005). Little emphasis is placed on image processing; it is assumed that the data used to create the graphics are available, and the goal is to enable users understand the information expressed in them. The task of generating captions for news images is novel to our knowledge. Instead of relying on manual annotation or background ontological information we exploit a multimodal database of news articles, images, and their captions. The latter is admittedly noisy, yet can be easily obtained from on-line sources, and contains rich information about the entities and events depicted in the images and their relations. Similar to previous work, we also follow a two-stage approach. Using an image annotation model, we first describe the picture with keywords which are subsequently realized into a human readable sentence. The caption generation task bears some resemblance to headline generation (Dorr et al., 2003; Banko et al., 2000; Jin and Hauptmann, 2002) where the aim is to create a very short summary for a document. Importantly, we aim to create a caption that not only summarizes the document but is also a faithful to the image’s content (i.e., the caption should also mention some of the objects or individuals depicted in the image). We therefore explore extractive and abstractive models that rely on visual information to drive the generation process. Our approach thus differs from most work in summarization which is solely text-based. 3 Problem Formulation We formulate image caption generation as follows. Given an image I, and a related knowledge database κ, create a natural language description C which captures the main content of the image under κ. Specifically, in the news story scenario, we will generate a caption C for an image I and its accompanying document D. The training data thus consists of document-image-caption tu1240 Thousands of Tongans have attended the funeral of King Taufa’ahau Tupou IV, who died last week at the age of 88. Representatives from 30 foreign countries watched as the king’s coffin was carried by 1,000 men to the official royal burial ground. King Tupou, who was 88, died a week ago. A Nasa satellite has documented startling changes in Arctic sea ice cover between 2004 and 2005. The extent of “perennial” ice declined by 14%, losing an area the size of Pakistan or Turkey. The last few decades have seen ice cover shrink by about 0.7% per year. Satellite instruments can distinguish “old” Arctic ice from “new”. Contaminated Cadbury’s chocolate was the most likely cause of an outbreak of salmonella poisoning, the Health Protection Agency has said. About 36 out of a total of 56 cases of the illness reported between March and July could be linked to the product. Cadbury will increase its contamination testing levels. A third of children in the UK use blogs and social network websites but two thirds of parents do not even know what they are, a survey suggests. The children’s charity NCH said there was “an alarming gap” in technological knowledge between generations. Children were found to be far more internet-wise than parents. Table 1: Each entry in the BBC News database contains a document an image, and its caption. ples like the ones shown in Table 1. During testing, we are given a document and an associated image for which we must generate a caption. Our experiments used the dataset created by Feng and Lapata (2008).2 It contains 3,361 articles downloaded from the BBC News website3 each of which is associated with a captioned news image. The latter is usually 203 pixels wide and 152 pixels high. The average caption length is 9.5 words, the average sentence length is 20.5 words, and the average document length 421.5 words. The caption vocabulary is 6,180 words and the document vocabulary is 26,795. The vocabulary shared between captions and documents is 5,921 words. The captions tend to use half as many words as the document sentences, and more than 50% of the time contain words that are not attested in the document (even though they may be attested in the collection). Generating image captions is a challenging task even for humans, let alone computers. Journalists are given explicit instructions on how to write captions4 and laypersons do not always agree on what a picture depicts (von Ahn and Dabbish, 2004). Along with the title, the lead, and section headings, captions are the most commonly read words 2Available from http://homepages.inf.ed.ac.uk/ s677528/data/ 3http://news.bbc.co.uk/ 4See http://www.theslot.com/captions.html and http://www.thenewsmanual.net/for tips on how to write good captions. in an article. A good caption must be succinct and informative, clearly identify the subject of the picture, establish the picture’s relevance to the article, provide context for the picture, and ultimately draw the reader into the article. It is also worth noting that journalists often write their own captions rather than simply extract sentences from the document. In doing so they rely on general world knowledge but also expertise in current affairs that goes beyond what is described in the article or shown in the picture. 4 Image Annotation As mentioned earlier, our approach relies on an image annotation model to provide description keywords for the picture. Our experiments made use of the probabilistic model presented in Feng and Lapata (2010). The latter is well-suited to our task as it has been developed with noisy, multimodal data sets in mind. The model is based on the assumption that images and their surrounding text are generated by mixtures of latent topics which are inferred from a concatenated representation of words and visual features. Specifically, images are preprocessed so that they are represented by word-like units. Local image descriptors are computed using the Scale Invariant Feature Transform (SIFT) algorithm (Lowe, 1999). The general idea behind the algorithm is to first sample an image with the difference-of-Gaussians point detector at different 1241 scales and locations. Importantly, this detector is, to some extent, invariant to translation, scale, rotation and illumination changes. Each detected region is represented with a SIFT descriptor which is a histogram of edge directions at different locations. Subsequently SIFT descriptors are quantized into a discrete set of visual terms via a clustering algorithm such as K-means. The model thus works with a bag-of-words representation and treats each article-image-caption tuple as a single document dMix consisting of textual and visual words. Latent Dirichlet Allocation (LDA, Blei et al. 2003) is used to infer the latent topics assumed to have generated dMix. The basic idea underlying LDA, and topic models in general, is that each document is composed of a probability distribution over topics, where each topic represents a probability distribution over words. The document-topic and topic-word distributions are learned automatically from the data and provide information about the semantic themes covered in each document and the words associated with each semantic theme. The image annotation model takes the topic distributions into account when finding the most likely keywords for an image and its associated document. More formally, given an image-captiondocument tuple (I,C,D) the model finds the subset of keywords WI (WI ⊆W) which appropriately describe I. Assuming that keywords are conditionally independent, and I, D are represented jointly by dMix, the model estimates: W ∗ I ≈ argmax Wt ∏ wt∈Wt P(wt|dMix) (1) = argmax Wt ∏ wt∈Wt K ∑ k=1 P(wt|zk)P(zk|dMix) Wt denotes a set of description keywords (the subscript t is used to discriminate from the visual words which are not part of the model’s output), K the number of topics, P(wt|zk) the multimodal word distributions over topics, and P(zk|dMix) the estimated posterior of the topic proportions over documents. Given an unseen image-document pair and trained multimodal word distributions over topics, it is possible to infer the posterior of topic proportions over the new data by maximizing the likelihood. The model delivers a ranked list of textual words wt, the n-best of which are used as annotations for image I. It is important to note that the caption generation models we propose are not especially tied to the above annotation model. Any probabilistic model with broadly similar properties could serve our purpose. Examples include PLSA-based approaches to image annotation (e.g., Monay and Gatica-Perez 2007) and correspondence LDA (Blei and Jordan, 2003). 5 Extractive Caption Generation Much work in summarization to date focuses on sentence extraction where a summary is created simply by identifying and subsequently concatenating the most important sentences in a document. Without a great deal of linguistic analysis, it is possible to create summaries for a wide range of documents, independently of style, text type, and subject matter. For our caption generation task, we need only extract a single sentence. And our guiding hypothesis is that this sentence must be maximally similar to the description keywords generated by the annotation model. We discuss below different ways of operationalizing similarity. Word Overlap Perhaps the simplest way of measuring the similarity between image keywords and document sentences is word overlap: Overlap(WI,Sd) = |WI ∩Sd| |WI ∪Sd| (2) where WI is the set of keywords and Sd a sentence in the document. The caption is then the sentence that has the highest overlap with the keywords. Cosine Similarity Word overlap is admittedly a naive measure of similarity, based on lexical identity. We can overcome this by representing keywords and sentences in vector space (Salton and McGill, 1983). The latter is a word-sentence co-occurrence matrix where each row represents a word, each column a sentence, and each entry the frequency with which the word appeared within the sentence. More precisely matrix cells are weighted by their tf-idf values. The similarity of the vectors representing the keywords −→ WI and document sentence −→ Sd can be quantified by measuring the cosine of their angle: sim(−→ WI,−→ Sd) = −→ WI ·−→ Sd | −−−−→ WI||−→ Sd| (3) Probabilistic Similarity Recall that the backbone of our image annotation model is a topic model with images and documents represented as a probability distribution over latent topics. Under this framework, the similarity between an im1242 age and a sentence can be broadly measured by the extent to which they share the same topic distributions (Steyvers and Griffiths, 2007). For example, we may use the KL divergence to measure the difference between the distributions p and q: D(p,q) = K ∑ j=1 pj log2 pj qj (4) where p and q are shorthand for the image topic distribution PdMix and sentence topic distribution PSd, respectively. When doing inference on the document sentence, we also take its neighboring sentences into account to avoid estimating inaccurate topic proportions on short sentences. The KL divergence is asymmetric and in many applications, it is preferable to apply a symmetric measure such as the Jensen Shannon (JS) divergence. The latter measures the “distance” between p and q through (p+q) 2 , the average of p and q: JS(p,q) = 1 2 D(p, (p+q) 2 )+D(q, (p+q) 2 ) (5) 6 Abstractive Caption Generation Although extractive methods yield grammatical captions and require relatively little linguistic analysis, there are a few caveats to consider. Firstly, there is often no single sentence in the document that uniquely describes the image’s content. In most cases the keywords are found in the document but interspersed across multiple sentences. Secondly, the selected sentences make for long captions (sometimes longer than the average document sentence), are not concise and overall not as catchy as human-written captions. For these reasons we turn to abstractive caption generation and present models based on single words but also phrases. Word-based Model Our first abstractive model builds on and extends a well-known probabilistic model of headline generation (Banko et al., 2000). The task is related to caption generation, the aim is to create a short, title-like headline for a given document, without however taking visual information into account. Like captions, headlines have to be catchy to attract the reader’s attention. Banko et al. (2000) propose a bag-of-words model for headline generation. It consists of content selection and surface realization components. Content selection is modeled as the probability of a word appearing in the headline given the same word appearing in the corresponding document and is independent from other words in the headline. The likelihood of different surface realizations is estimated using a bigram model. They also take the distribution of the length of the headlines into account in an attempt to bias the model towards generating concise output: P(w1,w2,...,wn) = n ∏ i=1 P(wi ∈H|wi ∈D) (6) ·P(len(H) = n) · n ∏ i=2 P(wi|wi−1) where wi is a word that may appear in headline H, D the document being summarized, and P(len(H) = n) a headline length distribution model. The above model can be easily adapted to the caption generation task. Content selection is now the probability of a word appearing in the caption given the image and its associated document which we obtain from the output of our image annotation model (see Section 4). In addition we replace the bigram surface realizer with a trigram: P(w1,w2,...,wn) = n ∏ i=1 P(wi ∈C|I,D) (7) ·P(len(C) = n) · n ∏ i=3 P(wi|wi−1,wi−2) where C is the caption, I the image, D the accompanying document, and P(wi ∈C|I,D) the image annotation probability. Despite its simplicity, the caption generation model in (7) has a major drawback. The content selection component will naturally tend to ignore function words, as they are not descriptive of the image’s content. This will seriously impact the grammaticality of the generated captions, as there will be no appropriate function words to glue the content words together. One way to remedy this is to revert to a content selection model that ignores the image and simply estimates the probability of a word appearing in the caption given the same word appearing in the document. At the same time we modify our surface realization component so that it takes note of the image annotation probabilities. Specifically, we use an adaptive language model (Kneser et al., 1997) that modifies an 1243 n-gram model with local unigram probabilities: P(w1,w2,...,wn) = n ∏ i=1 P(wi ∈C|wi ∈D) (8) ·P(len(C) = n) · n ∏ i=3 Padap(wi|wi−1,wi−2) where P(wi ∈C|wi ∈D) is the probability of wi appearing in the caption given that it appears in the document D, and Padap(wi|wi−1,wi−2) the language model adapted with probabilities from our image annotation model: Padap(w|h) = α(w) z(h) Pback(w|h) (9) α(w) ≈(Padap(w) Pback(w) )β (10) z(h) = ∑ w α(w)·Pback(w|h) (11) where Pback(w|h) is the probability of w given the history h of preceding words (i.e., the original trigram model), Padap(w) the probability of w according to the image annotation model, Pback(w) the probability of w according to the original model, and β a scaling parameter. Phrase-based Model The model outlined in equation (8) will generate captions with function words. However, there is no guarantee that these will be compatible with their surrounding context or that the caption will be globally coherent beyond the trigram horizon. To avoid these problems, we turn our attention to phrases which are naturally associated with function words and can potentially capture long-range dependencies. Specifically, we obtain phrases from the output of a dependency parser. A phrase is simply a head and its dependents with the exception of verbs, where we record only the head (otherwise, an entire sentence could be a phrase). For example, from the first sentence in Table 1 (first row, left document) we would extract the phrases: thousands of Tongans, attended, the funeral, King Taufa‘ahau Tupou IV, last week, at the age, died, and so on. We only consider dependencies whose heads are nouns, verbs, and prepositions, as these constitute 80% of all dependencies attested in our caption data. We define a bag-of-phrases model for caption generation by modifying the content selection and caption length components in equation (8) as follows: P(ρ1,ρ2,...,ρm) ≈ m ∏ j=1 P(ρ j ∈C|ρ j ∈D) (12) ·P(len(C) = m ∑ j=1 len(ρ j)) · ∑m j=1 len(ρ j) ∏ i=3 Padap(wi|wi−1,wi−2) Here, P(ρ j ∈C|ρj ∈D) models the probability of phrase ρ j appearing in the caption given that it also appears in the document and is estimated as: P(ρj ∈C|ρj ∈D) = ∏ wj∈ρ j P(w j ∈C|w j ∈D) (13) where wj is a word in the phrase ρ j. One problem with the models discussed thus far is that words or phrases are independent of each other. It is up to the trigram model to enforce coarse ordering constraints. These may be sufficient when considering isolated words, but phrases are longer and their combinations are subject to structural constraints that are not captured by sequence models. We therefore attempt to take phrase attachment constraints into account by estimating the probability of phrase ρ j attaching to the right of phrase ρi as: P(ρj|ρi)= ∑ wi∈ρi ∑ w j∈ρj p(w j|wi) (14) =1 2 ∑ wi∈ρi ∑ wj∈ρ j { f(wi,w j) f(wi,−) + f(wi,w j) f(−,w j) } where p(wj|wi) is the probability of a phrase containing word w j appearing to the right of a phrase containing word wi, f(wi,w j) indicates the number of times wi and w j are adjacent, f(wi,−) is the number of times wi appears on the left of any phrase, and f(−,wi) the number of times it appears on the right.5 After integrating the attachment probabilities into equation (12), the caption generation model becomes: P(ρ1,ρ2,...,ρm) ≈ m ∏ j=1 P(ρ j ∈C|ρ j ∈D) (15) · m ∏ j=2 P(ρj|ρ j−1) ·P(len(C) = ∑m j=1 len(ρ j)) ·∏ m ∑ j=1 len(ρj) i=3 Padap(wi|wi−1,wi−2) 5Equation (14) is smoothed to avoid zero probabilities. 1244 On the one hand, the model in equation (15) takes long distance dependency constraints into account, and has some notion of syntactic structure through the use of attachment probabilities. On the other hand, it has a primitive notion of caption length estimated by P(len(C) = ∑m j=1 len(ρ j)) and will therefore generate captions of the same (phrase) length. Ideally, we would like the model to vary the length of its output depending on the chosen context. However, we leave this to future work. Search To generate a caption it is necessary to find the sequence of words that maximizes P(w1,w2,...,wn) for the word-based model (equation (8)) and P(ρ1,ρ2,...,ρm) for the phrase-based model (equation (15)). We rewrite both probabilities as the weighted sum of their log form components and use beam search to find a near-optimal sequence. Note that we can make search more efficient by reducing the size of the document D. Using one of the models from Section 5, we may rank its sentences in terms of their relevance to the image keywords and consider only the n-best ones. Alternatively, we could consider the single most relevant sentence together with its surrounding context under the assumption that neighboring sentences are about the same or similar topics. 7 Experimental Setup In this section we discuss our experimental design for assessing the performance of the caption generation models presented above. We give details on our training procedure, parameter estimation, and present the baseline methods used for comparison with our models. Data All our experiments were conducted on the corpus created by Feng and Lapata (2008), following their original partition of the data (2,881 image-caption-document tuples for training, 240 tuples for development and 240 for testing). Documents and captions were parsed with the Stanford parser (Klein and Manning, 2003) in order to obtain dependencies for the phrase-based abstractive model. Model Parameters For the image annotation model we extracted 150 (on average) SIFT features which were quantized into 750 visual terms. The underlying topic model was trained with 1,000 topics using only content words (i.e., nouns, verbs, and adjectives) that appeared no less than five times in the corpus. For all models discussed here (extractive and abstractive) we report results with the 15 best annotation keywords. For the abstractive models, we used a trigram model trained with the SRI toolkit on a newswire corpus consisting of BBC and Yahoo! news documents (6.9 M words). The attachment probabilities (see equation (14)) were estimated from the same corpus. We tuned the caption length parameter on the development set using a range of [5,14] tokens for the word-based model and [2,5] phrases for the phrase-based model. Following Banko et al. (2000), we approximated the length distribution with a Gaussian. The scaling parameter β for the adaptive language model was also tuned on the development set using a range of [0.5,0.9]. We report results with β set to 0.5. For the abstractive models the beam size was set to 500 (with at least 50 states for the word-based model). For the phrase-based model, we also experimented with reducing the search scope, either by considering only the n most similar sentences to the keywords (range [2,10]), or simply the single most similar sentence and its neighbors (range [2,5]). The former method delivered better results with 10 sentences (and the KL divergence similarity function). Evaluation We evaluated the performance of our models automatically, and also by eliciting human judgments. Our automatic evaluation was based on Translation Edit Rate (TER, Snover et al. 2006), a measure commonly used to evaluate the quality of machine translation output. TER is defined as the minimum number of edits a human would have to perform to change the system output so that it exactly matches a reference translation. In our case, the original captions written by the BBC journalists were used as reference: TER(E,Er) = Ins+Del+Sub+Shft Nr (16) where E is the hypothetical system output, Er the reference caption, and Nr the reference length. The number of possible edits include insertions (Ins), deletions (Del), substitutions (Sub) and shifts (Shft). TER is similar to word error rate, the only difference being that it allows shifts. A shift moves a contiguous sequence to a different location within the the same system output and is counted as a single edit. The perfect TER score is 0, however note that it can be higher than 1 due to insertions. The minimum translation edit align1245 Model TER AvgLen Lead sentence 2.12† 21.0 Word Overlap 2.46∗† 24.3 Cosine 2.26† 22.0 KL Divergence 1.77∗† 18.4 JS Divergence 1.77∗† 18.6 Abstract Words 1.11∗† 10.0 Abstract Phrases 1.06∗† 10.1 Table 2: TER results for extractive, abstractive models, and lead sentence baseline; ∗: sig. different from lead sentence; †: sig. different from KL and JS divergence. ment is usually found through beam search. We used TER to compare the output of our extractive and abstractive models and also for parameter tuning (see the discussion above). In our human evaluation study participants were presented with a document, an associated image, and its caption, and asked to rate the latter on two dimensions: grammaticality (is the sentence fluent or word salad?) and relevance (does it describe succinctly the content of the image and document?). We used a 1–7 rating scale, participants were encouraged to give high ratings to captions that were grammatical and appropriate descriptions of the image given the accompanying document. We randomly selected 12 document-image pairs from the test set and generated captions for them using the best extractive system, and two abstractive systems (word-based and phrase-based). We also included the original human-authored caption as an upper bound. We collected ratings from 23 unpaid volunteers, all self reported native English speakers. The study was conducted over the Internet. 8 Results Table 2 reports our results on the test set using TER. We compare four extractive models based on word overlap, cosine similarity, and two probabilistic similarity measures, namely KL and JS divergence and two abstractive models based on words (see equation (8)) and phrases (see equation (15)). We also include a simple baseline that selects the first document sentence as a caption and show the average caption length (AvgLen) for each model. We examined whether performance differences among models are statistically significant, using the Wilcoxon test. Model Grammaticality Relevance KL Divergence 6.42∗† 4.10∗† Abstract Words 2.08† 3.20† Abstract Phrases 4.80∗ 4.96∗ Gold Standard 6.39∗† 5.55∗ Table 3: Mean ratings on caption output elicited by humans; ∗: sig. different from wordbased abstractive system; †: sig. different from phrase-based abstractive system. As can be seen the probabilistic models (KL and JS divergence) outperform word overlap and cosine similarity (all differences are statistically significant, p < 0.01).6 They make use of the same topic model as the image annotation model, and are thus able to select sentences that cover common content. They are also significantly better than the lead sentence which is a competitive baseline. It is well known that news articles are written so that the lead contains the most important information in a story.7 This is an encouraging result as it highlights the importance of the visual information for the caption generation task. In general, word overlap is the worst performing model which is not unexpected as it does not take any lexical variation into account. Cosine is slightly better but not significantly different from the lead sentence. The abstractive models obtain the best TER scores overall, however they generate shorter captions in comparison to the other models (closer to the length of the gold standard) and as a result TER treats them favorably, simply because the number of edits is less. For this reason we turn to the results of our judgment elicitation study which assesses in more detail the quality of the generated captions. Recall that participants judge the system output on two dimensions, grammaticality and relevance. Table 3 reports mean ratings for the output of the extractive system (based on the KL divergence), the two abstractive systems, and the human-authored gold standard caption. We performed an Analysis of Variance (ANOVA) to examine the effect of system type on the generation task. Post-hot Tukey tests were carried out on the mean of the ratings shown in Table 3 (for grammaticality and relevance). 6We also note that mean length differences are not significant among these models. 7As a rule of thumb the lead should answer most or all of the five W’s (who, what, when, where, why). 1246 G: King Tupou, who was 88, died a week ago. KL: Last year, thousands of Tongans took part in unprecedented demonstrations to demand greater democracy and public ownership of key national assets. AW : King Toupou IV died at the age of Tongans last week. AP: King Toupou IV died at the age of 88 last week. G: Cadbury will increase its contamination testing levels. KL: Contaminated Cadbury’s chocolate was the most likely cause of an outbreak of salmonella poisoning, the Health Protection Agency has said. AW : Purely dairy milk buttons Easter had agreed to work has caused. AP: The 105g dairy milk buttons Easter egg affected by the recall. G: Satellite instruments can distinguish “old” Arctic ice from “new”. KL: So a planet with less ice warms faster, potentially turning the projected impacts of global warming into reality sooner than anticipated. AW : Dr less winds through ice cover all over long time when. AP: The area of the Arctic covered in Arctic sea ice cover. G: Children were found to be far more internet-wise than parents. KL: That’s where parents come in. AW : The survey found a third of children are about mobile phones. AP: The survey found a third of children in the driving seat. Table 4: Captions written by humans (G) and generated by extractive (KL), word-based abstractive (AW), and phrase-based extractive (AP systems). The word-based system yields the least grammatical output. It is significantly worse than the phrase-based abstractive system (α < 0.01), the extractive system (α < 0.01), and the gold standard (α < 0.01). Unsurprisingly, the phrase-based system is significantly less grammatical than the gold standard and the extractive system, whereas the latter is perceived as equally grammatical as the gold standard (the difference in the means is not significant). With regard to relevance, the word-based system is significantly worse than the phrase-based system, the extractive system, and the gold-standard. Interestingly, the phrase-based system performs on the same level with the human gold standard (the difference in the means is not significant) and significantly better than the extractive system. Overall, the captions generated by the phrase-based system, capture the same content as the human-authored captions, even though they tend to be less grammatical. Examples of system output for the image-document pairs shown in Table 1 are given in Table 4 (the first row corresponds to the left picture (top row) in Table 1, the second row to the right picture, and so on). 9 Conclusions We have presented extractive and abstractive models that generate image captions for news articles. A key aspect of our approach is to allow both the visual and textual modalities to influence the generation task. This is achieved through an image annotation model that characterizes pictures in terms of description keywords that are subsequently used to guide the caption generation process. Our results show that the visual information plays an important role in content selection. Simply extracting a sentence from the document often yields an inferior caption. Our experiments also show that a probabilistic abstractive model defined over phrases yields promising results. It generates captions that are more grammatical than a closely related word-based system and manages to capture the gist of the image (and document) as well as the captions written by journalists. Future extensions are many and varied. Rather than adopting a two-stage approach, where the image processing and caption generation are carried out sequentially, a more general model should integrate the two steps in a unified framework. Indeed, an avenue for future work would be to define a phrase-based model for both image annotation and caption generation. We also believe that our approach would benefit from more detailed linguistic and non-linguistic information. For instance, we could experiment with features related to document structure such as titles, headings, and sections of articles and also exploit syntactic information more directly. The latter is currently used in the phrase-based model by taking attachment probabilities into account. We could, however, improve grammaticality more globally by generating a well-formed tree (or dependency graph). References Banko, Michel, Vibhu O. Mittal, and Micheael J. Witbrock. 2000. Headline generation based on statistical translation. In Proceedings of the 38th Annual Meeting on Association for Computational Linguistics. Hong Kong, pages 318–325. Barnard, Kobus, Pinar Duygulu, David Forsyth, Nando de Freitas, David Blei, and Michael Jordan. 2002. Matching words and pictures. Journal of Machine Learning Research 3:1107– 1135. Blei, David and Michael Jordan. 2003. Modeling annotated data. In Proceedings of the 26th An1247 nual International ACM SIGIR Conference on Research and Development in Information Retrieval. Toronto, ON, pages 127–134. Blei, David, Andrew Ng, and Michael Jordan. 2003. Latent Dirichlet allocation. Journal of Machine Learning Research 3:993–1022. Corio, Marc and Guy Lapalme. 1999. Generation of texts for information graphics. In Proceedings of the 7th European Workshop on Natural Language Generation. Toulouse, France, pages 49–58. Dorr, Bonnie, David Zajic, and Richard Schwartz. 2003. Hedge trimmer: A parse-and-trim approach to headline generation. In Proceedings of the HLT-NAACL 2003 Workshop on Text Summarization. Edmonton, Canada, pages 1–8. Duygulu, Pinar, Kobus Barnard, Nando de Freitas, and David Forsyth. 2002. Object recognition as machine translation: Learning a lexicon for a fixed image vocabulary. In Proceedings of the 7th European Conference on Computer Vision. Copenhagen, Denmark, pages 97–112. Elzer, Stephanie, Sandra Carberry, Ingrid Zukerman, Daniel Chester, Nancy Green, , and Seniz Demir. 2005. A probabilistic framework for recognizing intention in information graphics. In Proceedings of the 19th International Conference on Artificial Intelligence. Edinburgh, Scotland, pages 1042–1047. Fasciano, Massimo and Guy Lapalme. 2000. Intentions in the coordinated generation of graphics and text from tabular data. Knowledge Information Systems 2(3):310–339. Feiner, Steven and Kathleen McKeown. 1990. Coordinating text and graphics in explanation generation. In Proceedings of National Conference on Artificial Intelligence. Boston, MA, pages 442–449. Feng, Shaolei Feng, Victor Lavrenko, and R Manmatha. 2004. Multiple Bernoulli relevance models for image and video annotation. In Proceedings of the International Conference on Computer Vision and Pattern Recognition. Washington, DC, pages 1002–1009. Feng, Yansong and Mirella Lapata. 2008. Automatic image annotation using auxiliary text information. In Proceedings of the 46th Annual Meeting of the Association of Computational Linguistics: Human Language Technologies. Columbus, OH, pages 272–280. Feng, Yansong and Mirella Lapata. 2010. Topic models for image annotation and text illustration. In Proceedings of the 11th Annual Conference of the North American Chapter of the Association for Computational Linguistics. Los Angeles, LA. Ferres, Leo, Avi Parush, Shelley Roberts, and Gitte Lindgaard. 2006. Helping people with visual impairments gain access to graphical information through natural language: The graph system. In Proceedings of 11th International Conference on Computers Helping People with Special Needs. Linz, Austria, pages 1122–1130. H´ede, Patrick, Pierre Allain Mo¨ellic, Jo¨el Bourgeoys, Magali Joint, and Corinne Thomas. 2004. Automatic generation of natural language descriptions for images. In Proceedings of Computer-Assisted Information Retrieval (Recherche d’Information et ses Applications Ordinateur) (RIAO). Avignon, France. Jin, Rong and Alexander G. Hauptmann. 2002. A new probabilistic model for title generation. In Proceedings of the 19th International Conference on Computational linguistics. Taipei, Taiwan, pages 1–7. Klein, Dan and Christopher D. Manning. 2003. Accurate unlexicalized parsing. In Proceedings of the 41st Annual Meeting of the Association of Computational Linguistics. Sapporo, Japan, pages 423–430. Kneser, Reinhard, Jochen Peters, and Dietrich Klakow. 1997. Language model adaptation using dynamic marginals. In Proceedings of 5th European Conference on Speech Communication and Technology. Rhodes, Greece, volume 4, pages 1971–1974. Kojima, Atsuhiro, Mamoru Takaya, Shigeki Aoki, Takao Miyamoto, and Kunio Fukunaga. 2008. Recognition and textual description of human activities by mobile robot. In Proceedings of the 3rd International Conference on Innovative Computing Information and Control. IEEE Computer Society, Washington, DC, pages 53– 56. Kojima, Atsuhiro, Takeshi Tamura, and Kunio Fukunaga. 2002. Natural language description of human activities from video images based on concept hierarchy of actions. International Journal of Computer Vision 50(2):171–184. Lavrenko, Victor, R. Manmatha, and Jiwoon Jeon. 2003. A model for learning the semantics of 1248 pictures. In Proceedings of the 16th Conference on Advances in Neural Information Processing Systems. Vancouver, BC. Lowe, David G. 1999. Object recognition from local scale-invariant features. In Proceedings of International Conference on Computer Vision. IEEE Computer Society, pages 1150–1157. Mittal, Vibhu O., Johanna D. Moore, Giuseppe Carenini, and Steven Roth. 1998. Describing complex charts in natural language: A caption generation system. Computational Linguistics 24:431–468. Monay, Florent and Daniel Gatica-Perez. 2007. Modeling semantic aspects for cross-media image indexing. IEEE Transactions on Pattern Analysis and Machine Intelligence 29(10):1802–1817. Salton, Gerard and M.J. McGill. 1983. Introduction to Modern Information Retrieval. McGraw-Hill, New York. Smeulders, Arnols W.M., Marcel Worring, Simone Santini, Amarnath Gupta, and Ramesh Jain. 2000. Content-based image retrieval at the end of the early years. IEEE Transactions on Pattern Analysis and Machine Intelligence 22(12):1349–1380. Snover, Matthew, Bonnie Dorr, Richard Schwartz, Linnea Micciulla, and John Makhoul. 2006. A study of translation edit rate with targeted human annotation. In Proceedings of the 7th Conference of the Association for Machine Translation in the Americas. Cambridge, pages 223– 231. Steyvers, Mark and Tom Griffiths. 2007. Probabilistic topic models. In T. Landauer, D. McNamara, S Dennis, and W Kintsch, editors, A Handbook of Latent Semantic Analysis, Psychology Press. Vailaya, Aditya, M´ario A. T. Figueiredo, Anil K. Jain, and Hong-Jiang Zhang. 2001. Image classification for content-based indexing. IEEE Transactions on Image Processing 10:117–130. von Ahn, Luis and Laura Dabbish. 2004. Labeling images with a computer game. In ACM Conference on Human Factors in Computing Systems. New York, NY, pages 319–326. Wang, Chong, David Blei, and Li Fei-Fei. 2009. Simultaneous image classification and annotation. In Proceedings of the International Conference on Computer Vision and Pattern Recognition. Miami, FL, pages 1903–1910. Yao, Benjamin, Xiong Yang, Liang Lin, Mun Wai Lee, and Song chun Zhu. 2009. I2t: Image parsing to text description. Proceedings of IEEE (invited for the special issue on Internet Vision) . 1249
|
2010
|
126
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1250–1258, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Generating image descriptions using dependency relational patterns Ahmet Aker University of Sheffield [email protected] Robert Gaizauskas University of Sheffield [email protected] Abstract This paper presents a novel approach to automatic captioning of geo-tagged images by summarizing multiple webdocuments that contain information related to an image’s location. The summarizer is biased by dependency pattern models towards sentences which contain features typically provided for different scene types such as those of churches, bridges, etc. Our results show that summaries biased by dependency pattern models lead to significantly higher ROUGE scores than both n-gram language models reported in previous work and also Wikipedia baseline summaries. Summaries generated using dependency patterns also lead to more readable summaries than those generated without dependency patterns. 1 Introduction The number of images tagged with location information on the web is growing rapidly, facilitated by the availability of GPS (Global Position System) equipped cameras and phones, as well as by the widespread use of online social sites. The majority of these images are indexed with GPS coordinates (latitude and longitude) only and/or have minimal captions. This typically small amount of textual information associated with the image is of limited usefulness for image indexing, organization and search. Therefore methods which could automatically supplement the information available for image indexing and lead to improved image retrieval would be extremely useful. Following the general approach proposed by Aker and Gaizauskas (2009), in this paper we describe a method for automatic image captioning or caption enhancement starting with only a scene or subject type and a set of place names pertaining to an image – for example ⟨church, {St. Paul’s,London}⟩. Scene type and place names can be obtained automatically given GPS coordinates and compass information using techniques such as those described in Xin et al. (2010) – that task is not the focus of this paper. Our method applies only to images of static features of the built or natural landscape, i.e. objects with persistent geo-coordinates, such as buildings and mountains, and not to images of objects which move about in such landscapes, e.g. people, cars, clouds, etc. However, our technique is suitable not only for image captioning but in any application context that requires summary descriptions of instances of object classes, where the instance is to be characterized in terms of the features typically mentioned in describing members of the class. Aker and Gaizauskas (2009) have argued that humans appear to have a conceptual model of what is salient regarding a certain object type (e.g. church, bridge, etc.) and that this model informs their choice of what to say when describing an instance of this type. They also experimented with representing such conceptual models using n-gram language models derived from corpora consisting of collections of descriptions of instances of specific object types (e.g. a corpus of descriptions of churches, a corpus of bridge descriptions, and so on) and reported results showing that incorporating such n-gram language models as a feature in a feature-based extractive summarizer improves the quality of automatically generated summaries. The main weakness of n-gram language models is that they only capture very local information about short term sequences and cannot model long distance dependencies between terms. For example one common and important feature of object descriptions is the simple specification of the object type, e.g. the information that the object London Bridge is a bridge or that the Rhine is a river. If this information is expressed as in the first line of Table 1, n-gram language models are likely to 1250 Table 1: Example of sentences which express the type of an object. London Bridge is a bridge... The Rhine (German: Rhein; Dutch: Rijn; French: Rhin; Romansh: Rain; Italian: Reno; Latin: Rhenus West Frisian Ryn) is one of the longest and most important rivers in Europe... reflect it, since one would expect the tri-gram is a bridge to occur with high frequency in a corpus of bridge descriptions. However, if the type predication occurs with less commonly seen local context, as is the case for the object Rhine in the second row of Table 1 – most important rivers – n-gram language models may well be unable to identify it. Intuitively, what is important in both these cases is that there is a predication whose subject is the object instance of interest and the head of whose complement is the object type: London Bridge ... is ... bridge and Rhine ... is ... river. Sentences matching such patterns are likely to be important ones to include in a summary. This intuition suggests that rather than representing object type conceptual models via corpus-derived language models as do Aker and Gaizauskas (2009), we do so instead using corpus-derived dependency patterns. We pursue this idea in this paper, our hypothesis being that information that is important for describing objects of a given type will frequently be realized linguistically via expressions with the same dependency structure. We explore this hypothesis by developing a method for deriving common dependency patterns from object type corpora (Section 2) and then incorporating these patterns into an extractive summarization system (Section 3). In Section 4 we evaluate the approach both by scoring against model summaries and via a readability assessment. Since our work aims to extend the work of Aker and Gaizauskas (2009) we reproduce their experiments with n-gram language models in the current setting so as to permit accurate comparison. Multi-document summarizers face the problem of avoiding redundancy: often, important information which must be included in the summary is repeated several times across the document set, but must be included in the summary only once. We can use the dependency pattern approach to address this problem in a novel way. The common approach to avoiding redundancy is to use a text similarity measure to block the addition of a further sentence to the summary if it is too similar to one already included. Instead, since specific dependency patterns express specific types of inTable 2: Object types and the number of articles in each object type corpus. Object types which are bold are covered by the evaluation image set. village 39970, school 15794, city 14233, organization 9393, university 7101, area 6934, district 6565, airport 6493, island 6400, railway station 5905, river 5851, company 5734, mountain 5290, park 3754, college 3749, stadium 3665, lake 3649, road 3421, country 3186, church 3005, way 2508, museum 2320, railway 2093, house 2018, arena 1829, field 1731, club 1708, shopping centre 1509, highway 1464, bridge 1383, street 1352, theatre 1330, bank 1310, property 1261, hill 1072, castle 1022, forest 995, court 949, hospital 937, peak 906, bay 899, skyscraper 843, valley 763, hotel 741, garden 739, building 722, market 712, monument 679, port 651, sea 645, temple 625, beach 614, square 605, store 547, campus 525, palace 516, tower 496, cemetery 457, volcano 426, cathedral 402, glacier 392, residence 371, dam 363, waterfall 355, gallery 349, prison 348, cave 341, canal 332, restaurant 329, path 312, observatory 303, zoo 302, coast 298, statue 283, venue 269, parliament 258, shrine 256, desert 248, synagogue 236, bar 229, ski resort 227, arch 223, landscape 220, avenue 202, casino 179, farm 179, seaside 173, waterway 167, tunnel 167, ruin 166, chapel 165, observation wheel 158, basilica 157, woodland 154, wetland 151, cinema 144, gate 142, aquarium 136, entrance 136, opera house 134, spa 125, shop 124, abbey 108, boulevard 108, pub 92, bookstore 76, mosque 56 formation we can group the patterns into groups expressing the same type of information and then, during sentence selection, ensure that sentences matching patterns from different groups are selected in order to guarantee broad, non-redundant coverage of information relevant for inclusion in the summary. We report work experimenting with this idea too. 2 Representing conceptual models 2.1 Object type corpora We derive n-gram language and dependency pattern models using object type corpora made available to us by Aker and Gaizauskas. Aker and Gaizauskas (2009) define an object type corpus as a collection of texts about a specific static object type such as church, bridge, etc. Objects can be named locations such as Eiffel Tower. To refer to such names they use the term toponym. To build such object type corpora the authors categorized Wikipedia articles places by object type. The object type of each article was identified automatically by running Is-A patterns over the first five sentences of the article. The authors report 91% accuracy for their categorization process. The most populated of the categories identified (in total 107 containing articles about places around the world) are shown in Table 2. 2.2 N-gram language models Aker and Gaizauskas (2009) experimented with uni-gram and bi-gram language models to capture the features commonly used when describing an object type and used these to bias the sentence selection of the summarizer towards the sentences that contain these features. As in Song and Croft (1999) they used their language models in a gener1251 ative way, i.e. they calculate the probability that a sentence is generated based on a n-gram language model. They showed that summarizer biased with bi-gram language models produced better results than those biased with uni-gram models. We replicate the experiments of Aker and Gaizauskas and generate a bi-gram language model for each object type corpus. In later sections we use LM to refer to these models. 2.3 Dependency patterns We use the same object type corpora to derive dependency patterns. Our patterns are derived from dependency trees which are obtained using the Stanford parser1. Each article in each object type corpus was pre-processed by sentence splitting and named entity tagging2. Then each sentence was parsed by the Stanford dependency parser to obtain relational patterns. As with the chain model introduced by Sudo et al. (2001) our relational patterns are concentrated on the verbs in the sentences and contain n+1 words (the verb and n words in direct or indirect relation with the verb). The number n is experimentally set to two words. For illustration consider the sentence shown in Table 3 that is taken from an article in the bridge corpus. The first two rows of the table show the original sentence and its form after named entity tagging. The next step in processing is to replace any occurrence of a string denoting the object type by the term “OBJECTTYPE” as shown in the third row of Table 3. The final two rows of the table show the output of the Stanford dependency parser and the relational patterns identified for this example. To obtain the relational patterns from the parser output we first identified the verbs in the output. For each such verb we extracted two further words being in direct or indirect relation to the current verb. Two words are directly related if they occur in the same relational term. The verb built-4, for instance, is directly related to DATE-6 because they both are in the same relational term prepin(built-4, DATE-6). Two words are indirectly related if they occur in two different terms but are linked by a word that occurs in those two terms. The verb was-3 is, for instance, indirectly related to OBJECTTYPE-2 because they are both in different terms but linked with built-4 that occurs in 1http://nlp.stanford.edu/software/lex-parser.shtml 2For performing shallow text analysis the OpenNLP tools (http://opennlp.sourceforge.net/) were used. Table 3: Example sentence for dependency pattern. Original sentence: The bridge was built in 1876 by W. W. After NE tagging: The bridge was built in DATE by W. W. Input to the parser: The OBJECTTYPE was built in DATE by W. W. Output of the parser: det(OBJECTTYPE-2, The-1), nsubjpass(built4, OBJECTTYPE-2), auxpass(built-4, was-3), prep-in(built-4, DATE-6), nn(W-10, W-8), agent(built-4, W-10) Patterns: The OBJECTTYPE built, OBJECTTYPE was built, OBJECTTYPE built DATE, OBJECTTYPE built W, was built DATE, was built W both terms. E.g. for the term nsubjpass(built-4, OBJECTTYPE-2) we use the verb built and extract patterns based on this. OBJECTTYPE is in direct relation to built and The is in indirect relation to built through OBJECTTYPE. So a pattern from these relations is The OBJECTTYPE built. The next pattern extracted from this term is OBJECTTYPE was built. This pattern is based on direct relations. The verb built is in direct relation to OBJECTTYPE and also to was. We continue this until we cover all direct relations with built resulting in two more patterns (OBJECTTYPE built DATE and OBJECTTYPE built W). It should be noted that we consider all direct and indirect relations while generating the patterns. Following these steps we extracted relational patterns for each object type corpus along with the frequency of occurrence of the pattern in the entire corpus. The frequency values are used by the summarizer to score the sentences. In the following sections we will use the term DpM to refer to these dependency pattern models. 2.3.1 Pattern categorization In addition to using dependency patterns as models for biasing sentence selection, we can also use them to control the kind of information to be included in the final summary (see Section 3.2). We may want to ensure that the summary contains a sentence describing the object type of the object, its location and some background information. For example, for the object Eiffel Tower we aim to say that it is a tower, located in Paris, designed by Gustave Eiffel, etc. To be able to do so, we categorize dependency patterns according to the type of information they express. We manually analyzed human written descriptions about instances of different object types and recorded for each sentence in the descriptions the kind of information it contained about the object. We analyzed descriptions of 310 different objects where each object had up to four different human written descriptions (Section 4.1). We categorized the information contained in the descriptions into 1252 the following categories: • type: sentences containing the “type” information of the object such as XXX is a bridge • year: sentences containing information about when the object was built or in case of mountains, for instance, when it was first climbed • location: sentences containing information about where the object is located • background: sentences containing some specific information about the object • surrounding: sentences containing information about what other objects are close to the main object • visiting: sentences containing information about e.g. visiting times, etc. We also manually assigned each dependency pattern in each corpus-derived model to one of the above categories, provided it occurred five or more times in the object type corpora. The patterns extracted for our example sentence shown in Table 3, for instance, are all categorized by year category because all of them contain information about the foundation date of an object. 3 Summarizer We adopted the same overall approach to summarization used by Aker and Gaizauskas (2009) to generate the image descriptions. The summarizer is an extractive, query-based multi-document summarization system. It is given two inputs: a toponym associated with an image and a set of documents to be summarized which have been retrieved from the web using the toponym as a query. The summarizer creates image descriptions in a three step process. First, it applies shallow text analysis, including sentence detection, tokenization, lemmatization and POS-tagging to the given input documents. Then it extracts features from the document sentences. Finally, it combines the features using a linear weighting scheme to compute the final score for each sentence and to create the final summary. We modified the approach to feature extraction and the way the summarizer acquires the weights for feature combination. The following subsections describe how feature extraction/combination is done in more detail. 3.1 Feature Extraction The original summarizer reported in Aker and Gaizauskas (2009) uses the following features to score the sentences: • querySimilarity: Sentence similarity to the query (toponym) (cosine similarity over the vector representation of the sentence and the query). • centroidSimilarity: Sentence similarity to the centroid. The centroid is composed of the 100 most frequently occurring non stop words in the document collection (cosine similarity over the vector representation of the sentence and the centroid). • sentencePosition: Position of the sentence within its document. The first sentence in the document gets the score 1 and the last one gets 1 n where n is the number of sentences in the document. • starterSimilarity: A sentence gets a binary score if it starts with the query term (e.g. Westminster Abbey, The Westminster Abbey, The Westminster or The Abbey) or with the object type, e.g. The church. We also allow gaps (up to four words) between the and the query to capture cases such as The most magnificent Abbey, etc. • LMSim3: The similarity of a sentence S to an n-gram language model LM (the probability that the sentence S is generated by LM). In our experiments we extend this feature set by two dependency pattern related features: DpMSim and DepCat. DpMSim is computed in a similar fashion to LMSim feature. We assign each sentence a dependency similarity score. To compute this score, we first parse the sentence on the fly with the Stanford parser and obtain the dependency patterns for the sentence. We then associate each dependency pattern of the sentence with the occurrence frequency of that pattern in the dependency pattern model (DpM). DpMSim is then computed as given in Equation 1. It is a sum of all occurrence frequencies of the dependency patterns detected in a sentence S that are also contained in the DpM. DpMSim(S, DpM) = X p∈S fDpM(p) (1) The second feature, DepCat, uses dependency patterns to categorize the sentences rather than ranking them. It can be used independently from other features to categorize each sentence by one of the categories described in Section 2.3.1. To do this, we obtain the relational patterns for the current sentence, check whether for each such pattern whether it is included in the DpM, and, if so, we add to the sentence the category the pattern was manually associated with. It should be noted that a sentence can have more than one category. This can occur, for instance, if the sentence contains information about when something was built and at the same time where it is located. It is also important to mention that assigning sentences categories does not change the order in the ranked list. We use DepCat to generate an automated summary by first including sentences containing the category “type”, then “year” and so on until the 3In Aker and Gaizauskas (2009) this feature is called modelSimilarity. 1253 summary length is violated. The sentences are selected according to the order in which they occur in the ranked list. From each of the first three categories (“type”, “year” and “location”) we take a single sentence to avoid redundancy. The same is applied to the final two categories (“surrounding” and “visiting”). Then, if length limit is not violated, we fill the summary with sentences from the “background” category until the word limit of 200 words is reached. Here the number of added sentences is not limited. Finally, we order the sentences by first adding the sentences from the first three categories to the summary, then the “background” related sentences and finally the last two sentences from the “surrounding” and “visiting” categories. However, in cases where we have not reached the summary word limit because of uncovered categories, i.e. there were not, for instance, sentences about “location”, we add to the end of the summary the next top sentence from the ranked list that was not taken. 3.2 Sentence Selection To compute the final score for each sentence Aker and Gaizauskas (2009) use a linear function with weighted features: Sscore = ( n X i=1 featurei ∗weighti) (2) We use the same approach, but whereas the feature weights they use are experimentally set rather than learned, we learn the weights using linear regression instead. We used 2 3 of the 310 images from our image set (see Section 4.1) to train the weights. The image descriptions from this data set are used as model summaries. Our training data contains for each image a set of image descriptions taken from the VirtualTourist travel community web-site 4. From this web-site we took all existing image descriptions about a particular image or object. Note that some of these descriptions about a particular object were used to derive the model summaries for that object (see Section 4.1). Assuming that model summaries contain the most relevant sentences about an object we perform ROUGE comparisons between the sentences in all the image descriptions and the model summaries, i.e. we pair each sentence from all image descriptions about a particular place with every sentence from all the model 4www.virtualtourist.com summaries for that particular object. Sentences which are exactly the same or have common parts will score higher in ROUGE than sentences which do not have anything in common. In this way, we have for each sentence from all existing image descriptions about an object a ROUGE score5 indicating its relevance. We also ran the summarizer for each of these sentences to compute the values for the different features. This gives information about each feature’s value for each sentence. Then the ROUGE scores and feature score values for every sentence were input to the linear regression algorithm to train the weights. Given the weights, Equation 2 is used to compute the final score for each sentence. The final sentence scores are used to sort the sentences in the descending order. This sorted list is then used by the summarizer to generate the final summary as described in Aker and Gaizauskas (2009). 4 Evaluation To evaluate our approach we used two different assessment methods: ROUGE (Lin, 2004) and manual readability. In the following we first describe the data sets used in each of these evaluations, and then we present the results of each assessment. 4.1 Data sets For evaluation we use the image collection described in Aker and Gaizauskas (2010). The image collection contains 310 different images with manually assigned toponyms. The images cover 60 of the 107 object types identified from Wikipedia (see Table 2). For each image there are up to four short descriptions or model summaries. The model summaries were created manually based on image descriptions taken from VirtualTourist and contain a minimum of 190 and a maximum of 210 words. An example model summary about the Eiffel Tower is shown in Table 4. 2 3 of this image collection was used to train the weights and the remaining 1 3 (105 images) for evaluation. To generate automatic captions for the images we automatically retrieved the top 30 related web-documents for each image using the Yahoo! search engine and the toponym associated with the image as a query. The text from these documents was extracted using an HTML parser and passed to the summarizer. The set of documents we used to generate our summaries excluded any VirtualTourist related sites, as these were used to generate 5We used ROUGE 1. 1254 Table 4: Model, Wikipedia baseline and starterSimilarity+LMSim+DepCat summary for Eiffel Tower. Model Summary Wikipedia baseline summary starterSimilarity+LMSim+DepCat summary The Eiffel Tower is the most famous place in Paris. It is made of 15,000 pieces fitted together by 2,500,000 rivets. It’s of 324 m (1070 ft) high structure and weighs about 7,000 tones. This world famous landmark was built in 1889 and was named after its designer, engineer Gustave Alexandre Eiffel. It is now one of the world’s biggest tourist places which is visited by around 6,5 million people yearly. There are three levels to visit: Stages 1 and 2 which can be reached by either taking the steps (680 stairs) or the lift, which also has a restaurant ”Altitude 95” and a Souvenir shop on the first floor. The second floor also has a restaurant ”Jules Verne”. Stage 3, which is at the top of the tower can only be reached by using the lift. But there were times in the history when Tour Eiffel was not at all popular, when the Parisians thought it looked ugly and wanted to pull it down. The Eiffel Tower can be reached by using the Mtro through Trocadro, Ecole Militaire, or Bir-Hakeim stops. The address is: Champ de Mars-Tour Eiffel. The Eiffel Tower (French: Tour Eiffel, [tur efel]) is a 19th century iron lattice tower located on the Champ de Mars in Paris that has become both a global icon of France and one of the most recognizable structures in the world. The Eiffel Tower, which is the tallest building in Paris, is the single most visited paid monument in the world; millions of people ascend it every year. Named after its designer, engineer Gustave Eiffel, the tower was built as the entrance arch for the 1889 World’s Fair. The tower stands at 324 m (1,063 ft) tall, about the same height as an 81-story building. It was the tallest structure in the world from its completion until 1930, when it was eclipsed by the Chrysler Building in New York City. Not including broadcast antennas, it is the second-tallest structure in France, behind the Millau Viaduct, completed in 2004. The tower has three levels for visitors. Tickets can be purchased to ascend either on stairs or lifts to the first and second levels. The Eiffel Tower, which is the tallest building in Paris, is the single most visited paid monument in the world; millions of people ascend it every year. The tower is located on the Left Bank of the Seine River, at the northwestern extreme of the Parc du Champ de Mars, a park in front of the Ecole Militaire that used to be a military parade ground. The tower was met with much criticism from the public when it was built, with many calling it an eyesore. Counting from the ground, there are 347 steps to the first level, 674 steps to the second level, and 1,710 steps to the small platform on the top of the tower. Although it was the world’s tallest structure when completed in 1889, the Eiffel Tower has since lost its standing both as the tallest lattice tower and as the tallest structure in France. The tower has two restaurants: Altitude 95, on the first floor 311ft (95m) above sea level; and the Jules Verne, an expensive gastronomical restaurant on the second floor, with a private lift. Table 5: ROUGE scores for each single feature and Wikipedia baseline. Recall centroidSimilarity sentencePosition querySimilarity starterSimilarity LMSim DpMSim*** Wiki R2 .0734 .066 .0774 .0869 .0895 .093 .097 RSU4 .12 .11 .12 .137 .142 .145 .14 the model summaries. 4.2 ROUGE assessment In the first assessment we compared the automatically generated summaries against model summaries written by humans using ROUGE (Lin, 2004). Following the Document Understanding Conference (DUC) evaluation standards we used ROUGE 2 (R2) and ROUGE SU4 (RSU4) as evaluation metrics (Dang, 2006) . ROUGE 2 gives recall scores for bi-gram overlap between the automatically generated summaries and the reference ones. ROUGE SU4 allows bi-grams to be composed of non-contiguous words, with a maximum of four words between the bi-grams. As baselines for evaluation we used two different summary types. Firstly, we generated summaries for each image using the top-ranked non Wikipedia document retrieved in the Yahoo! search results for the given toponyms. From this document we create a baseline summary by selecting sentences from the beginning until the summary reaches a length of 200 words. As a second baseline we use the Wikipedia article for a given toponym from which we again select sentences from the beginning until the summary length limit is reached. First, we compared the baseline summaries against the VirtualTourist model summaries. The comparison shows that the Wikipedia baseline ROUGE scores (R2 .097***, RSU4 .14***) are significantly higher than the first document ones (R2 0.042, RSU4 .079) 6. Thus, we will focus on the Wikipedia baseline summaries to draw conclusions about our automatic summaries. Table 4 shows the Wikipedia baseline summary about the Eiffel Tower. Secondly, we separately ran the summarizer over the top ten documents for each single feature and compared the automated summaries against the model ones. The results of this comparison are shown in Table 5. Table 5 shows that the dependency model feature (DpMSim) contributes most to the summary quality according to the ROUGE metrics. It is also significantly better than all other feature scores except the LMSim feature. Compared to LMSim ROUGE scores the DpMSim feature offers only a moderate improvement. The same moderate improvement we can see between the DpMSim RSU4 and the Wiki RSU4. The lowest ROUGE scores are obtained if only sentence position (sentecePosition) is used. To see how the ROUGE scores change when features are combined with each other we performed different combinations of the features, ran the summarizer for each combination and compared the automated summaries against the model ones. In the different combinations we 6To assess the statistical significance of ROUGE score differences between multiple summarization results we performed a pairwise Wilcoxon signed-rank test. We use the following conventions for indicating significance level in the tables: *** = p < .0001, ** = p < .001, * = p < .05 and no star indicates non-significance. 1255 Table 6: ROUGE scores of feature combinations which score moderately or significantly higher than dependency pattern model (DpMSim) feature and Wikipedia baseline. Recall starterSimilarity + LMSim starterSimilarity + LMSim + DepCat*** DpmSim Wiki R2 .095 .102 .093 .097 RSU4 .145 .155 .145 .14 also included the dependency pattern categorization (DepCat) feature explained in Section 3.1. Table 6 shows the results of feature combinations which score moderately or significantly higher than the dependency pattern model (DpMSim) feature score shown in Table 5. The results showed that combining DpMSim with other features did not lead to higher ROUGE scores than those produced by that feature alone. The summaries categorized by dependency patterns (starterSimilarity+LMSim+DepCat) achieve significantly higher ROUGE scores than the Wikipedia baseline. For both ROUGE R2 and ROUGE SU4 the significance is at level p < .0001. Table 4 shows a summary about the Eiffel Tower obtained using this starterSimilarity+LMSim+DepCat feature. Table 5 also shows the ROUGE scores of the feature combination starterSimilarity and LMSim used without the dependency categorization (DepCat) feature. It can be seen that this combination without the dependency patterns lead to lower ROUGE scores in ROUGE 2 and only moderate improvement in ROUGE SU4 if compared with Wikipedia baseline ROUGE scores. 4.3 Readability assessment We also evaluated our summaries using a readability assessment as in DUC and TAC. DUC and TAC manually assess the quality of automatically generated summaries by asking human subjects to score each summary using five criteria – grammaticality, redundancy, clarity, focus and structure criteria. Each criterion is scored on a five point scale with high scores indicating a better result (Dang, 2005). For this evaluation we used the same 105 images as in the ROUGE evaluation. As the ROUGE evaluation showed that the dependency pattern categorization (DepCat) renders the best results when used in feature combination starterSimilarity + LMSim + DepCat, we further investigated the contribution of dependency pattern categorization by performing a readability assessment on summaries generated using this feature combination. For comparison we also evaluated summaries which were not structured by dependency patterns (starterSimilarity + LMSim) and also the Wikipedia baseline summaries. We asked four people to assess the summaries. Each person was shown all 315 summaries (105 from each summary type) in a random way and was asked to assess them according to the DUC and TAC manual assessment scheme. The results are shown in Table 7. We see from Table 7 that using dependency patterns to categorize the sentences and produce a structured summary helps to obtain better readable summaries. Looking at the 5 and 4 scores the table shows that the dependency pattern categorized summaries (SLMD) have better clarity (85% of the summaries), are more coherent (74% of the summaries), contain less redundant information (83% of the summaries) and have better grammar (92% of the summaries) than the ones without dependency categorization (80%, 70%, 60%, 84%). The scores of our automated summaries were better than the Wikipedia baseline summaries in the grammar feature. However, in other features the Wikipedia baseline summaries obtained better scores than our automated summaries. This comparison show that there is a gap to fill in order to obtain better readable summaries. 5 Related Work Our approach has an advantage over related work in automatic image captioning in that it requires only GPS information associated with the image in order to generate captions. Other attempts towards automatic generation of image captions generate captions based on the immediate textual context of the image with or without consideration of image related features such as colour, shape or texture (Deschacht and Moens, 2007; Mori et al., 2000; Barnard and Forsyth, 2001; Duygulu et al., 2002; Barnard et al., 2003; Pan et al., 2004; Feng and Lapata, 2008; Satoh et al., 1999; Berg et al., 2005). However, Marsch & White (2003) argue that the content of an image and its immediate text have little semantic agreement and this can, according to Purves et al. (2008), be misleading to image retrieval. Furthermore, these approaches assume that the image has been obtained from a document. In cases where there is no document associated with the image, which is the scenario we are principally concerned with, these techniques are not applicable. 1256 Table 7: Readability evaluation results: Each cell shows the percentage of summaries scoring the ranking score heading the column for each criterion in the row as produced by the summary method indicated by the subcolumn heading – Wikipedia baseline (W), starterSimilarity + LMSim (SLM) and starterSimilarity + LMSim + DepCat (SLMD). The numbers indicate the percentage values averaged over the four people. 5 4 3 2 1 Criterion W SLM SLMD W SLM SLMD W SLM SLMD W SLM SLMD W SLM SLMD clarity 72.6 50.5 53.6 21.7 30.0 31.4 1.2 6.7 5.7 4.0 10.2 6.0 0.5 2.6 3.3 focus 72.1 49.3 51.2 20.5 26.0 25.2 3.8 10.0 10.7 3.3 10.0 10.5 0.2 4.8 2.4 coherence 67.1 39.0 48.3 23.6 31.4 26.9 4.8 12.4 11.9 3.3 10.2 9.8 1.2 6.9 3.1 redundancy 69.8 42.9 55.0 21.7 17.4 28.8 2.4 4.5 4.3 5.0 27.1 8.8 1.2 8.1 3.1 grammar 48.6 55.7 62.9 32.9 29.0 30.0 5.0 3.1 1.9 11.7 12.1 5.2 1.9 0 0 Dependency patterns have been exploited in various language processing applications. In information extraction, for instance, dependency patterns have been used to extract relevant information from text resources (Yangarber et al., 2000; Sudo et al., 2001; Culotta and Sorensen, 2004; Stevenson and Greenwood, 2005; Bunescu and Mooney, 2005; Stevenson and Greenwood, 2009). However, dependency patterns have not been used extensively in summarization tasks. We are aware only of the work described in Nobata et al. (2002) who used dependency patterns in combination with other features to generate extracts in a single document summarization task. The authors found that when learning weights in a simple feature weigthing scheme, the weight assigned to dependency patterns was lower than that assigned to other features. The small contribution of the dependency patterns may have been due to the small number of documents they used to derive their dependency patterns – they gathered dependency patterns from only ten domain specific documents which are unlikely to be sufficient to capture repeated features in a domain. 6 Discussion and Conclusion We have proposed a method by which dependency patterns extracted from corpora of descriptions of instances of particular object types can be used in a multi-document summarizer to automatically generate image descriptions. Our evaluations show that such an approach yields summaries which score more highly than an approach which uses a simpler representation of an object type model in the form of a n-gram language model. When used as the sole feature for sentence ranking, dependency pattern models (DpMSim) produced summaries with higher ROUGE scores than those obtained using the features reported in Aker and Gaizauskas (2009). These dependency pattern models also achieved a modest improvement over Wikipedia baseline ROUGE SU4. Furthermore, we showed that using dependency patterns in combination with features reported in Aker and Gaizauskas to produce a structured summary led to significantly better results than Wikipedia baseline summaries as assessed by ROUGE. However, human assessed readability showed that there is still scope for improvement. These results indicate that dependency patterns are worth investigating for object focused automated summarization tasks. Such investigations should in particular concentrate on how dependency patterns can be used to structure information within the summary, as our best results were achieved when dependency patterns were used for this purpose. There are a number of avenues to pursue in future work. One is to explore how dependency patterns could be used to produce generative summaries and/or perform sentence trimming. Another is to investigate how dependency patterns might be automatically clustered into groups expressing similar or related facts, rather than relying on manual categorization of dependency patterns into categories such as “type”, “year”, etc. as was done here. Evaluation should be extended to investigate the utility of the automatically generated image descriptions for image retrieval. Finally, we also plan to analyze automated ways for learning information structures (e.g. what is the flow of facts to describe a location) from existing image descriptions to produce better summaries. 7 Acknowlegment The research reported was funded by the TRIPOD project supported by the European Commission under the contract No. 045335. We would like to thank Emina Kurtic, Mesude Bicak, Edina Kurtic and Olga Nesic for participating in our manual evaluation. We also would like to thank Trevor Cohn and Mark Hepple for discussions and comments. References A. Aker and R. Gaizauskas. 2009. Summary Generation for Toponym-Referenced Images using Object 1257 Type Language Models. International Conference on Recent Advances in Natural Language Processing (RANLP),2009. A. Aker and R. Gaizauskas. 2010. Model Summaries for Location-related Images. In Proc. of the LREC2010 Conference. K. Barnard and D. Forsyth. 2001. Learning the semantics of words and pictures. In International Conference on Computer Vision, volume 2, pages 408–415. Vancouver: IEEE. K. Barnard, P. Duygulu, D. Forsyth, N. de Freitas, D.M. Blei, and M.I. Jordan. 2003. Matching words and pictures. The Journal of Machine Learning Research, 3:1107–1135. T.L. Berg, A.C. Berg, J. Edwards, and DA Forsyth. 2005. Whos in the Picture? In Advances in Neural Information Processing Systems 17: Proc. Of The 2004 Conference. MIT Press. R.C. Bunescu and R.J. Mooney. 2005. A shortest path dependency kernel for relation extraction. In Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, pages 724–731. Association for Computational Linguistics Morristown, NJ, USA. A. Culotta and J. Sorensen. 2004. Dependency Tree Kernels for Relation Extraction. In Proceedings of the 42nd Meeting of the Association for Computational Linguistics (ACL’04), Main Volume, pages 423–429, Barcelona, Spain, July. H.T. Dang. 2005. Overview of DUC 2005. DUC 05 Workshop at HLT/EMNLP. H.T. Dang. 2006. Overview of DUC 2006. National Institute of Standards and Technology. K. Deschacht and M.F. Moens. 2007. Text Analysis for Automatic Image Annotation. Proc. of the 45th Annual Meeting of the Association for Computational Linguistics. East Stroudsburg: ACL. P. Duygulu, K. Barnard, JFG de Freitas, and D.A. Forsyth. 2002. Object Recognition as Machine Translation: Learning a Lexicon for a Fixed Image Vocabulary. In Seventh European Conference on Computer Vision (ECCV), 4:97–112. X. Fan, A. Aker, M. Tomko, P. Smart, M Sanderson, and R. Gaizauskas. 2010. Automatic Image Captioning From the Web For GPS Photographs. In Proc. of the 11th ACM SIGMM International Conference on Multimedia Information Retrieval, National Constitution Center, Philadelphia, Pennsylvania. Y. Feng and M. Lapata. 2008. Automatic Image Annotation Using Auxiliary Text Information. Proc. of Association for Computational Linguistics (ACL) 2008, Columbus, Ohio, USA. C.Y. Lin. 2004. ROUGE: A Package for Automatic Evaluation of Summaries. Proc. of the Workshop on Text Summarization Branches Out (WAS 2004), pages 25–26. E.E. Marsh and M.D. White. 2003. A taxonomy of relationships between images and text. Journal of Documentation, 59:647–672. Y. Mori, H. Takahashi, and R. Oka. 2000. Automatic word assignment to images based on image division and vector quantization. In Proc. of RIAO 2000: Content-Based Multimedia Information Access. C. Nobata, S. Sekine, H. Isahara, and R. Grishman. 2002. Summarization system integrated with named entity tagging and ie pattern discovery. In Proc. of the LREC-2002 Conference, pages 1742–1745. J.Y. Pan, H.J. Yang, P. Duygulu, and C. Faloutsos. 2004. Automatic image captioning. In Multimedia and Expo, 2004. ICME’04. IEEE International Conference on, volume 3. RS Purves, A. Edwardes, and M. Sanderson. 2008. Describing the where–improving image annotation and search through geography. 1st Intl. Workshop on Metadata Mining for Image Understanding, Funchal, Madeira-Portugal. S. Satoh, Y. Nakamura, and T. Kanade. 1999. Name-It: naming and detecting faces in news videos. Multimedia, IEEE, 6(1):22–35. F. Song and W.B. Croft. 1999. A general language model for information retrieval. In Proc. of the eighth international conference on Information and knowledge management, pages 316–321. ACM New York, NY, USA. M. Stevenson and M.A. Greenwood. 2005. A semantic approach to IE pattern induction. In Proc. of the 43rd Annual Meeting on Association for Computational Linguistics, pages 379–386. Association for Computational Linguistics Morristown, NJ, USA. M. Stevenson and M. Greenwood. 2009. Dependency Pattern Models for Information Extraction. Research on Language and Computation, 7(1):13– 39. K. Sudo, S. Sekine, and R. Grishman. 2001. Automatic pattern acquisition for Japanese information extraction. In Proc. of the first international conference on Human language technology research, page 7. Association for Computational Linguistics. R. Yangarber, R. Grishman, P. Tapanainen, and S. Huttunen. 2000. Automatic acquisition of domain knowledge for information extraction. In Proc. of the 18th International Conference on Computational Linguistics (COLING 2000), pages 940–946. Saarbriicken, Germany, August. 1258
|
2010
|
127
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1259–1267, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Incorporating Extra-linguistic Information into Reference Resolution in Collaborative Task Dialogue Ryu Iida Shumpei Kobayashi Takenobu Tokunaga Tokyo Institute of Technology 2-12-1, ˆOokayama, Meguro, Tokyo 152-8552, Japan {ryu-i,skobayashi,take}@cl.cs.titech.ac.jp Abstract This paper proposes an approach to reference resolution in situated dialogues by exploiting extra-linguistic information. Recently, investigations of referential behaviours involved in situations in the real world have received increasing attention by researchers (Di Eugenio et al., 2000; Byron, 2005; van Deemter, 2007; Spanger et al., 2009). In order to create an accurate reference resolution model, we need to handle extra-linguistic information as well as textual information examined by existing approaches (Soon et al., 2001; Ng and Cardie, 2002, etc.). In this paper, we incorporate extra-linguistic information into an existing corpus-based reference resolution model, and investigate its effects on reference resolution problems within a corpus of Japanese dialogues. The results demonstrate that our proposed model achieves an accuracy of 79.0% for this task. 1 Introduction The task of identifying reference relations including anaphora and coreferences within texts has received a great deal of attention in natural language processing, from both theoretical and empirical perspectives. Recently, research trends for reference resolution have drastically shifted from handcrafted rule-based approaches to corpus-based approaches, due predominately to the growing success of machine learning algorithms (such as Support Vector Machines (Vapnik, 1998)); many researchers have examined ways for introducing various linguistic clues into machine learning-based models (Ge et al., 1998; Soon et al., 2001; Ng and Cardie, 2002; Yang et al., 2003; Iida et al., 2005; Yang et al., 2005; Yang et al., 2008; Poon and Domingos, 2008, etc.). Research has continued to progress each year, focusing on tackling the problem as it is represented in the annotated data sets provided by the Message Understanding Conference (MUC)1 and the Automatic Content Extraction (ACE)2. In these data sets, coreference relations are defined as a limited version of a typical coreference; this generally means that only the relations where expressions refer to the same named entities are addressed, because it makes the coreference resolution task more information extraction-oriented. In other words, the coreference task as defined by MUC and ACE is geared toward only identifying coreference relations anchored to an entity within the text. In contrast to this research trend, investigations of referential behaviour in real world situations have continued to gain interest in the language generation community (Di Eugenio et al., 2000; Byron, 2005; van Deemter, 2007; Foster et al., 2008; Spanger et al., 2009), aiming at applications such as human-robot interaction. Spanger et al. (2009) for example constructed a corpus by recording dialogues of two participants collaboratively solving the Tangram puzzle. The corpus includes extra-lingustic information synchronised with utterances (such as operations on the puzzle pieces). They analysed the relations between referring expressions and the extra-linguistic information, and reported that the pronominal usage of referring expressions is predominant. They also revealed that the multi-modal perspective of reference should be dealt with for more realistic reference understanding. Thus, a challenging issue in reference resolution is to create a model bridging a referring expression in the text and its object in the real world. As a first step, this paper focuses on incorporating extra-linguistic information into an existing corpus-based approach, taking Spanger et al. (2009)’s REX-J corpus3 as the data set. In our 1www-nlpir.nist.gov/related projects/muc/ 2www.itl.nist.gov/iad/mig//tests/ace/ 3The corpus was named REX-J after their publication of 1259 problem setting, a referent needs to be identified by taking into account extra-linguistic information, such as the spatiala relations of puzzle pieces and the participants’ operations on them, as well as any preceding utterances in the dialogue. We particularly focus on the participants’ operation of pieces and so introduce it as several features in a machine learning-based approach. This paper is organised as follows. We first explain the corpus of collaborative work dialogues in Section 2, and then present our approach for identifying a referent given a referring expression in situated dialogues in Section 3. Section 4 shows the results of our empirical evaluation. In Section 5 we compare our work with existing work on reference resolution, and then conclude this paper and discuss future directions in Section 6. 2 REX-J corpus: a corpus of collaborative work dialogue For investigating dialogue from the multi-modal perspective, researchers have developed data sets including extra-linguistic information, bridging objects in the world and their referring expressions. The COCONUT corpus (Di Eugenio et al., 2000) is collected from keyboard-dialogues between two participants, who are collaborating on a simple 2D design task. The setting tends to encourage simple types of expressions by the participants. The COCONUT corpus is also limited to annotations with symbolic information about objects, such as object attributes and location in discrete coordinates. Thus, in addition to the artificial nature of interaction, such as using keyboard input, this corpus only records restricted types of data. On the other hand, though the annotated corpus by Spanger et al. (2009) focuses on a limited domain (i.e. collaborative work dialogues for solving the Tangram puzzle using a puzzle simulator on the computer), the required operations to solve the puzzle, and the situation as it is updated by a series of operations on the pieces are both recorded by the simulator. The relationship between a referring expression in a dialogue and its referent on a computer display is also annotated. For this reason, we selected the REX-J corpus for use in our empirical evaluations on reference resolution. Before explaining the details of our evaluation, we sketch Spanger et al. (2009), which describes its construction. goal shape area working area Figure 1: Screenshot of the Tangram simulator out the REX-J corpus and some of its prominent statistics. 2.1 The REX-J corpus In the process of building the REX-J corpus, Spanger et al. (2009) recruited 12 Japanese graduate students (4 females and 8 males), and split them into 6 pairs. All pairs knew each other previously and were of the same sex and approximately the same age. Each pair was instructed to solve the Tangram puzzle. The goal of the puzzle is to construct a given shape by arranging seven pieces of simple figures as shown in Figure 1. The precise position of every piece and every action that the participants make are recorded by the Tangram simulator in which the pieces on the computer display can be moved, rotated and flipped with simple mouse operations. The piece position and the mouse actions were recorded at intervals of 10 msec. The simulator displays two areas: a goal shape area (the left side of Figure 1) and a working area (the right side of Figure 1) where pieces are shown and can be manipulated. A different role was assigned to each participant of a pair: a solver and an operator. Given a certain goal shape, the solver thinks of the necessary arrangement of the pieces and gives instructions to the operator for how to move them. The operator manipulates the pieces with the mouse according to the solver’s instructions. During this interaction, frequent uttering of referring expressions are needed to distinguish the pieces of the puzzle. This collaboration is achieved by placing a set of participants side by side, each with their own display showing the work area, and a shield screen set between them to prevent the operator from seeing the goal shape, which is visible only on the solver’s screen, and to further restrict their 1260 interaction to only speech. 2.2 Statistics Table 1 lists the syntactic and semantic features of the referring expressions in the corpus with their respective frequencies. Note that multiple features can be used in a single expression. This list demonstrates that ‘pronoun’ and ‘shape’ features are frequently uttered in the corpus. This is because pronominal expressions are often used for pointing to a piece on a computer display. Expressions representing ‘shape’ frequently appear in dialogues even though they may be relatively redundant in the current utterance. From these statistics, capturing these two features can be judged as crucial as a first step toward accurate reference resolution. 3 Reference Resolution using Extra-linguistic Information Before explaining the treatment of extra-linguistic information, let us first describe the task definition, taking the REX-J corpus as target data. In the task of reference resolution, the reference resolution model has to identify a referent (i.e. a piece on a computer display)4. In comparison to conventional problem settings for anaphora resolution, where the model searches for an antecedent out of a set of candidate antecedents from preceding utterances, expressions corresponding to antecedents are sometimes omitted because referring expressions are used as deixis (i.e. physically pointing to a piece on a computer display); they may also refer to a piece that has just been manipulated by an operator due to the temporal salience in a series of operations. For these reasons, even though the model checks all candidates in the preceding utterances, it may not find the antecedent of a given referring expression. However, we do know that each referent exists as a piece on the display. We can therefore establish that when a referring expression is uttered by either a solver or an operator, the model can choose one of seven pieces as a referent of the current referring expression. 3.1 Ranking model to identify referents To investigate the impact of extra-linguistic information on reference resolution, we conduct an em4In the current task on reference resolution, we deal only with referring expressions referring to a single piece to minimise complexity. pirical evaluation in which a reference resolution model chooses a referent (i.e. a piece) for a given referring expression from the set of pieces illustrated on the computer display. As a basis for our reference resolution model, we adopt an existing model for reference resolution. Recently, machine learning-based approaches to reference resolution (Soon et al., 2001; Ng and Cardie, 2002, etc.) have been developed, particularly focussing on identifying anaphoric relations in texts, and have achieved better performance than hand-crafted rule-based approaches. These models for reference resolution take into account linguistic factors, such as relative salience of candidate antecedents, which have been modeled in Centering Theory (Grosz et al., 1995) by ranking candidate antecedents appearing in the preceding discourse (Iida et al., 2003; Yang et al., 2003; Denis and Baldridge, 2008). In order to take advantage of existing models, we adopt the rankingbased approach as a basis for our reference resolution model. In conventional ranking-based models, Yang et al. (2003) and Iida et al. (2003) decompose the ranking process into a set of pairwise comparisons of two candidate antecedents. However, recent work by Denis and Baldridge (2008) reports that appropriately constructing a model for ranking all candidates yields improved performance over those utilising pairwise ranking. Similarly we adopt a ranking-based model, in which all candidate antecedents compete with one another to decide the most likely candidate antecedent. Although the work by Denis and Baldridge (2008) uses Maximum Entropy to create their ranking-based model, we adopt the Ranking SVM algorithm (Joachims, 2002), which learns a weight vector to rank candidates for a given partial ranking of each referent. Each training instance is created from the set of all referents for each referring expression. To define the partial ranking of referents, we simply rank referents referred to by a given referring expression as first place and other referents as second place. 3.2 Use of extra-linguistic information Recent work on multi-modal reference resolution or referring expression generation (Prasov and Chai, 2008; Foster et al., 2008; Carletta et al., 2010) indicates that extra-linguistic information, such as eye-gaze and manipulation of objects, is 1261 Table 1: Referring expressions in REX-J corpus feature tokens example demonstratives 742 adjective 194 “ano migigawa no sankakkei (that triangle at the right side)” pronoun 548 “kore (this)” attribute 795 size 223 “tittyai sankakkei (the small triangle)” shape 566 “ˆokii sankakkei (the large triangle)” direction 6 “ano sita muiteru dekai sankakkei (that large triangle facing to the bottom)” spatial relations 147 projective 143 “hidari no okkii sankakkei (the small triangle on the left)” topological 2 “ˆokii hanareteiru yatu (the big distant one)” overlapping 2 “ sono sita ni aru sankakkei (the triangle underneath it)” action-mentioning 85 “migi ue ni doketa sankakkei (the triangle you put away to the top right)” one of essential clues for distinguishing deictic reference from endophoric reference. For instance, Prasov and Chai (2008) demonstrated that integrating eye-gaze information (especially, relative fixation intensity, the amount of time spent fixating a candidate object) into the conventional dialogue history-based model improved the performance of reference resolution. Foster et al. (2008) investigated the relationship of referring expressions and the manupluation of objects on a collaborative construction task, which is similar to our Tangram task5. They reported about 36% of the initial mentioned referring expressions in their corpus were involved with participant’s operations of objects, such as mouse manipulation. From these background, in addition to the information about the history of the preceding discourse, which has been used in previous machine learning-based approaches, we integrate extralinguistic information into the reference resolution model shown in Section 3.1. More precisely, we introduce the following extra-linguistic information: the information with regards to the history of a piece’s movement and the mouse cursor positions, and the information of the piece currently manipulated by an operator. We next elaborate on these three kinds of features. All the features are summarised in Table 2. 3.2.1 Discourse history features First, ‘type of’ features are acquired from the expressions of a given referring expression and its antecedent in the preceding discourse if the an5Note that the task defined in Foster et al. (2008) makes no distinction between two roles; a operator and a solver. Thus, two partipants both can mamipulate pieces on a computer display, but need to jointly construct to create a predefined goal shape. tecedent explicitly appears. These features have been examined by approaches to anaphora or coreference resolution (Soon et al., 2001; Ng and Cardie, 2002, etc.) to capture the salience of a candidate antecedent. To capture the textual aspect of dialogues for solving Tangram puzzle, we exploit the features such as a binary value indicating whether a referring expression has no antecedent in the preceding discourse and case markers following a candidate antecedent. 3.2.2 Action history features The history of the operations may yield important clues that indicate the salience in terms of the temporal recency of a piece within a series of operations. To introduce this aspect as a set of features, we can use, for example, the time distance of a candidate referent (i.e. a piece in the Tangram puzzle) since the mouse cursor was moved over it. We call this type of feature the action history feature. 3.2.3 Current operation features The recency of operations of a piece is also an important factor on reference resolution because it is directly associated with the focus of attention in terms of the cognition in a series of operations. For example, since a piece which was most recently manipulated is most salient from cognitive perspectives, it might be expected that the piece tends to be referred to by unmarked referring expressions such as pronouns. To incorporate such clues into the reference resolution model, we can use, for example, the time distance of a candidate referent since it was last manipulated in the preceding utterances. We call this type of feature the current operation feature. 1262 Table 2: Feature set (a) Discourse history features DH1 : yes, no a binary value indicating that P is referred to by the most recent referring expression. DH2 : yes, no a binary value indicating that the time distance to the last mention of P is less than or equal to 10 sec. DH3 : yes, no a binary value indicating that the time distance to the last mention of P is more than 10 sec and less than or equal to 20 sec. DH4 : yes, no a binary value indicating that the time distance to the last mention of P is more than 20 sec. DH5 : yes, no a binary value indicating that P has never been referred to by any mentions in the preceding utterances. DH6 : yes, no, N/A a binary value indicating that the attributes of P are compatible with the attributes of R. DH7 : yes, no a binary value indicating that R is followed by the case marker ‘o (accusative)’. DH8 : yes, no a binary value indicating that R is followed by the case marker ‘ni (dative)’. DH9 : yes, no a binary value indicating that R is a pronoun and the most recent reference to P is not a pronoun. DH10 : yes, no a binary value indicating that R is not a pronoun and was most recently referred to by a pronoun. (b) Action history features AH1 : yes, no a binary value indicating that the mouse cursor was over P at the beginning of uttering R. AH2 : yes, no a binary value indicating that P is the last piece that the mouse cursor was over when feature AH1 is ‘no’. AH3 : yes, no a binary value indicating that the time distance is less than or equal to 10 sec after the mouse cursor was over P. AH4 : yes, no a binary value indicating that the time distance is more than 10 sec and less than or equal to 20 sec after the mouse cursor was over P. AH5 : yes, no a binary value indicating that the time distance is more than 20 sec after the mouse cursor was over P. AH6 : yes, no a binary value indicating that the mouse cursor was never over P in the preceding utterances. (c) Current operation features CO1 : yes, no a binary value indicating that P is being manipulated at the beginning of uttering R. CO2 : yes, no a binary value indicating that P is the most recently manipulated piece when feature CO1 is ‘no’. CO3 : yes, no a binary value indicating that the time distance is less than or equal to 10 sec after P was most recently manipulated. CO4 : yes, no a binary value indicating that the time distance is more than 10 sec and less than or equal to 20 sec after P was most recently manipulated. CO5 : yes, no a binary value indicating that the time distance is more than 20 sec after P was most recently manipulated. CO6 : yes, no a binary value indicating that P has never been manipulated. P stands for a piece of the Tangram puzzle (i.e. a candidate referent of a referring expression) and R stands for the target referring expression. 4 Empirical Evaluation In order to investigate the effect of the extralinguistic information introduced in this paper, we conduct an empirical evaluation using the REX-J corpus. 4.1 Models As we see in Section 2.2, the feature testing whether a referring expression is a pronoun or not is crucial because it is directly related to the ‘deictic’ usage of referring expressions, whereas other expressions tend to refer to an expression appearing in the preceding utterances. As described in Denis and Baldridge (2008), when the size of training instances is relatively small, the models induced by learning algorithms (e.g. SVM) should be separately created with regards to distinct features. Therefore, focusing on the difference of the pronominal usage of referring expressions, we separately create the reference resolution models; one is for identifying a referent of a given pronoun, and the other is for all other expressions. We henceforth call the former model the pronoun model and the latter one the non-pronoun model respectively. At the training phase, we use only training instances whose referring expressions are pronouns for creating the pronoun model, and all other training instances are used for the nonpronoun model. The model using one of these models depending on the referring expression to be solved is called the separate model. To verify Denis and Baldridge (2008)’s premise mentioned above, we also create a model using all training instances without dividing pronouns and other. This model is called the combined model hereafter. 4.2 Experimental setting We used 40 dialogues in the REX-J corpus6, containing 2,048 referring expressions. To facilitate the experiments, we conduct 10-fold crossvalidation using 2,035 referring expressions, each of which refers to a single piece in a computer dis6Spanger et al. (2009)’s original corpus contains only 24 dialogues. In addition to this, we obtained anothor 16 dialogues by favour of the authors. 1263 Table 3: Results on reference resolution: accuracy model discourse history +action history* +current operation +action history, (baseline) +current operation* separated model (a+b) 0.664 (1352/2035) 0.790 (1608/2035) 0.685 (1394/2035) 0.780 (1587/2035) a) pronoun model 0.648 (660/1018) 0.886 (902/1018) 0.692 (704/1018) 0.875 (891/1018) b) non-pronoun model 0.680 (692/1017) 0.694 (706/1017) 0.678 (690/1017) 0.684 (696/1017) combined model 0.664 (1352/2035) 0.749 (1524/2035) 0.650 (1322/2035) 0.743 (1513/2035) ‘*’ means the extra-lingustic features (or the combinations of them) significantly contribute to improving performance. For the significant tests, we used McNemar test with Bonferroni’s correction for multiple comparisons, i.e. α/K = 0.05/4 = 0.01. play7. As a baseline model, we adopted a model only using the discourse history features. We utilised SVMrank8 as an implementation of the Ranking SVM algorithm, in which the parameter c was set as 1.0 and the remaining parameters were set to their defaults. 4.3 Results The results of each model are shown in Table 3. First of all, by comparing the models with and without extra-linguistic information (i.e. the model using all features shown in Table 2 and the baseline model), we can see the effectiveness of extra-linguistic information. The results typically show that the former achieved better performance than the latter. In particular, it indicates that exploiting the action history features are significantly useful for reference resolution in this data set. Second, we can also see the impact of extralinguistic information (especially, the action history features) with regards to the pronoun and non-pronoun models. In the former case, the model with extra-linguistic information improved by about 22% compared with the baseline model. On the other hand, in the latter case, the accuracy improved by only 7% over the baseline model. The difference may be caused by the fact that pronouns are more sensitive to the usage of the action history features because pronouns are often uttered as deixis (i.e. a pronoun tends to directly refer to a piece shown in a computer display). The results also show that the model using the discourse history and action history features achieved better performance than the model using all the features. This may be due to the duplicated definitions between the action history and current 7The remaining 13 instances referred to either more than one piece or a class of pieces, thus were excluded in this experiment. 8www.cs.cornell.edu/people/tj/svm light/svm rank.html Table 4: Weights of the features in each model pronoun model non-pronoun model rank feature weight feature weight 1 AH1 0.6371 DH6 0.7060 2 AH3 0.2721 DH2 0.2271 3 DH1 0.2239 AH3 0.2035 4 DH2 0.2191 AH1 0.1839 5 CO1 0.1911 DH1 0.1573 6 DH9 0.1055 DH7 0.0669 7 AH2 0.0988 CO5 0.0433 8 CO3 0.0852 CO3 0.0393 9 DH6 0.0314 CO1 0.0324 10 CO2 0.0249 DH3 0.0177 11 DH10 0 AH4 0.0079 12 DH7 -0.0011 AH2 0.0069 13 DH3 -0.0088 CO4 0.0059 14 CO6 -0.0228 DH10 0.0059 15 CO4 -0.0308 DH9 0 16 CO5 -0.0317 CO2 -0.0167 17 DH8 -0.0371 DH8 -0.0728 18 AH6 -0.0600 CO6 -0.0885 19 AH4 -0.0761 DH4 -0.0924 20 DH5 -0.0910 AH5 -0.1042 21 DH4 -0.1193 AH6 -0.1072 22 AH5 -0.1361 DH5 -0.1524 operation features. As we can see in the feature definitions of CO1 and AH1, some current operation features partially overlap with the action history features, which is effectively used in the ranking process. However, the other current operation features may have bad effects for ranking referents due to their ill-formed definitions. To shed light on this problem, we need additional investigation of the usage of features, and to refine their definitions. Finally, the results show that the performance of the separated model is significantly better than that of the combined model9, which indicates that separately creating models to specialise in distinct factors (i.e. whether a referring expression is a pronoun or not) is important as suggested by Denis and Baldridge (2008). We next investigated the significance of each 9For the significant tests, we used McNemar test (α = 0.05). 1264 Table 5: Frequencies of REs relating to on-mouse pronouns others total # all REs 548 693 1,241 # on-mouse 452 155 607 (82.5%) (22.4%) (48.9%) ‘# all REs’ stands for the frequency of referring expressions uttered in the corpus and ‘# on-mouse’ is the frequency of referring expressions in the situation when a referring expression is uttered and a mouse cursor is over the piece referred to by the expression. feature of the pronoun and non-pronoun models. We calculate the weight of feature f shown in Table 2 according to the following formula. weight(f) = ∑ x∈SV s wxzx(f) (1) where SVs is a set of the support vectors in a ranker induced by SVMrank, wx is the weight of the support vector x, zx(f) is the function that returns 1 if f occurs in x, respectively. The feature weights are shown in Table 4. This demonstrates that in the pronoun model the action history features have the highest weight, while with the non-pronoun model these features are less significant. As we can see in Table 5, pronouns are strongly related to the situation where a mouse cursor is over a piece, directly causing the weights of the features associated with the ‘on-mouse’ situation to become higher than other features. On the other hand, in the non-pronoun model, the discourse history features, such as DH6 and DH2, are the most significant, indicating that the compatibility of the attributes of a piece and a referring expression is more crucial than other action history and current operation features. This is compatible with the previous research concerning textual reference resolution (Mitkov, 2002). Table 4 shows that feature AH3 (aiming at capturing the recency in terms of a series of operations) is also significant. It empirically proves that the recent operation is strongly related to the salience of reference as a kind of ‘focus’ by humans. 5 Related Work There have been increasing concerns about reference resolution in dialogue. Byron and Allen (1998) and Eckert and Strube (2000) reported about 50% of pronouns had no antecedent in TRAINS93 and Switchboard corpora respectively. Strube and M¨uller (2003) attempted to resolve pronominal anaphora in the Switchboard corpus by porting a corpus-based anaphora resolution model focusing on written texts (e.g. Soon et al. (2001) and Ng and Cardie (2002)). They used specialised features for spoken dialogues as well as conventional features. They reported relatively worse results than with written texts. The reason is that the features in their work capture only information derived from transcripts of dialogues, while it is also essential to bridge objects and concepts in the real (or virtual) world and their expressions (especially pronouns) for recognising referential relations intrinsically. To improve performance on reference resolution in dialogue, researchers have focused on anaphoricity determination, which is the task of judging whether an expression explicitly has an antecedent in the text (i.e. in the preceding utterances) (M¨uller, 2006; M¨uller, 2007). Their work presented implementations of pronominal reference resolution in transcribed, multi-party dialogues. M¨uller (2006) focused on the determination of non-referential it, categorising instances of it in the ICSI Meeting Corpus (Janin et al., 2003) into six classes in terms of their grammatical categories. They also took into account each characteristic of these types by using a refined feature set. In the work by M¨uller (2007), they conducted an empirical evaluation including antecedent identification as well as anaphoricity determination. They used the relative frequencies of linguistic patterns as clues to introduce specific patterns for nonreferentials. They reported that their performance for detecting non-referentials was relatively high (80.0% in precision and 60.9% in recall), while the overall performance was still low (18.2% in precision and 19.1% in recall). These results indicate the need for advancing research in reference resolution in dialogue. In contrast to the above mentioned research, our task includes the treatment of entity disambiguation (i.e. selecting a referent out of a set of pieces on a computer display) as well as conventional anaphora resolution. Although our task setting is limited to the problem of solving the Tangram puzzle, we believe it is a good starting point for incorporating real (or virtual) world entities into coventional anaphora resolution. 1265 6 Conclusion This paper presented the task of reference resolution bridging pieces in the real world and their referents in dialogue. We presented an implementation of a reference resolution model exploiting extra-linguistic information, such as action history and current operation features, to capture the salience of operations by a participant and the arrangement of the pieces. Through our empirical evaluation, we demonstrated that the extra-linguistic information introduced in this paper contributed to improving performance. We also analysed the effect of each feature, showing that while action history features were useful for pronominal reference, discourse history features made sense for the other references. In order to enhance this kind of reference resolution, there are several possible future directions. First, in the current problem setting, we exclude zero-anaphora (i.e. omitted expressions refer to either an expression in the previous utterances or an object on a display deictically). However, zero-anaphora is essential for precise modeling and recognition of reference because it is also directly related with the recency of referents, either textually or situationally. Second, representing distractors in a reference resolution model is also a key. Although, this paper presents an implementation of a reference model considering only the relationship between a referring expression and its candidate referents. However, there might be cases when the occurrence of expressions or manipulated pieces intervening between a referring expression and its referent need to be taken into account. Finally, more investigation is needed for considering other extra-linguistic information, such as eye-gaze, for exploring what kinds of information is critical to recognising reference in dialogue. References D. K. Byron and J. F. Allen. 1998. Resolving demonstrative pronouns in the trains93 corpus. In Proceedings of the 2nd Colloquium on Discourse Anaphora and Anaphor Resolution (DAARC2), pages 68–81. D. K. Byron. 2005. Utilizing visual attention for cross-model coreference interpretation. In CONTEXT 2005, pages 83–96. J. Carletta, R. L. Hill, C. Nicol, T. Taylor, J. P. de Ruiter, and E. G. Bard. 2010. Eyetracking for two-person tasks with manipulation of a virtual world. Behavior Research Methods, 42:254–265. P. Denis and J. Baldridge. 2008. Specialized models and ranking for coreference resolution. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 660–669. B. P. W. Di Eugenio, R. H. Thomason, and J. D. Moore. 2000. The agreement process: An empirical investigation of human-human computer-mediated collaborative dialogues. International Journal of HumanComputer Studies, 53(6):1017–1076. M. Eckert and M. Strube. 2000. Dialogue acts, synchronising units and anaphora resolution. Journal of Semantics, 17(1):51–89. M. E. Foster, E. G. Bard, M. Guhe, R. L. Hill, J. Oberlander, and A. Knoll. 2008. The roles of hapticostensive referring expressions in cooperative, taskbased human-robot dialogue. In Proceedings of the 3rd ACM/IEEE international conference on Human robot interaction (HRI ’08), pages 295–302. N. Ge, J. Hale, and E. Charniak. 1998. A statistical approach to anaphora resolution. In Proceedings of the 6th Workshop on Very Large Corpora, pages 161– 170. B. J. Grosz, A. K. Joshi, and S. Weinstein. 1995. Centering: A framework for modeling the local coherence of discourse. Computational Linguistics, 21(2):203–226. R. Iida, K. Inui, H. Takamura, and Y. Matsumoto. 2003. Incorporating contextual cues in trainable models for coreference resolution. In Proceedings of the 10th EACL Workshop on The Computational Treatment of Anaphora, pages 23–30. R. Iida, K. Inui, and Y. Matsumoto. 2005. Anaphora resolution by antecedent identification followed by anaphoricity determination. ACM Transactions on Asian Language Information Processing (TALIP), 4(4):417–434. A. Janin, D. Baron, J. Edwards, D. Ellis, D. Gelbart, N. Morgan, B. Peskin, T. Pfau, E. Shriberg, A. Stolcke, and C. Wooters. 2003. The ICSI meeting corpus. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 364–367. T. Joachims. 2002. Optimizing search engines using clickthrough data. In Proceedings of the ACM Conference on Knowledge Discovery and Data Mining (KDD), pages 133–142. R. Mitkov. 2002. Anaphora Resolution. Studies in Language and Linguistics. Pearson Education. C. M¨uller. 2006. Automatic detection of nonreferential It in spoken multi-party dialog. In Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics, pages 49–56. 1266 C. M¨uller. 2007. Resolving It, This, and That in unrestricted multi-party dialog. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 816–823. V. Ng and C. Cardie. 2002. Improving machine learning approaches to coreference resolution. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), pages 104–111. H. Poon and P. Domingos. 2008. Joint unsupervised coreference resolution with Markov Logic. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 650– 659. Z. Prasov and J. Y. Chai. 2008. What’s in a gaze?: the role of eye-gaze in reference resolution in multimodal conversational interfaces. In Proceedings of the 13th international conference on Intelligent user interfaces (IUI ’08), pages 20–29. W. M. Soon, H. T. Ng, and D. C. Y. Lim. 2001. A machine learning approach to coreference resolution of noun phrases. Computational Linguistics, 27(4):521–544. P. Spanger, Y. Masaaki, R. Iida, and T. Takenobu. 2009. Using extra linguistic information for generating demonstrative pronouns in a situated collaboration task. In Proceedings of Workshop on Production of Referring Expressions: Bridging the gap between computational and empirical approaches to reference. M. Strube and C. M¨uller. 2003. A machine learning approach to pronoun resolution in spoken dialogue. In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, pages 168– 175. K. van Deemter. 2007. TUNA: Towards a unified algorithm for the generation of referring expressions. Technical report, Aberdeen University. V. N. Vapnik. 1998. Statistical Learning Theory. Adaptive and Learning Systems for Signal Processing Communications, and control. John Wiley & Sons. X. Yang, G. Zhou, J. Su, and C. L. Tan. 2003. Coreference resolution using competition learning approach. In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics (ACL), pages 176–183. X. Yang, J. Su, and C. L. Tan. 2005. Improving pronoun resolution using statistics-based semantic compatibility information. In Proceeding of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL), pages 165–172. X. Yang, J. Su, J. Lang, C. L. Tan, T. Liu, and S. Li. 2008. An entity-mention model for coreference resolution with inductive logic programming. In Proceedings of Annual Meeting of the Association for Computational Linguistics (ACL): Human Language Technologies (HLT), pages 843–851. 1267
|
2010
|
128
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1268–1277, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Reading Between the Lines: Learning to Map High-level Instructions to Commands S.R.K. Branavan, Luke S. Zettlemoyer, Regina Barzilay Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology {branavan, lsz, regina}@csail.mit.edu Abstract In this paper, we address the task of mapping high-level instructions to sequences of commands in an external environment. Processing these instructions is challenging—they posit goals to be achieved without specifying the steps required to complete them. We describe a method that fills in missing information using an automatically derived environment model that encodes states, transitions, and commands that cause these transitions to happen. We present an efficient approximate approach for learning this environment model as part of a policygradient reinforcement learning algorithm for text interpretation. This design enables learning for mapping high-level instructions, which previous statistical methods cannot handle.1 1 Introduction In this paper, we introduce a novel method for mapping high-level instructions to commands in an external environment. These instructions specify goals to be achieved without explicitly stating all the required steps. For example, consider the first instruction in Figure 1 — “open control panel.” The three GUI commands required for its successful execution are not explicitly described in the text, and need to be inferred by the user. This dependence on domain knowledge makes the automatic interpretation of high-level instructions particularly challenging. The standard approach to this task is to start with both a manually-developed model of the environment, and rules for interpreting high-level instructions in the context of this model (Agre and 1Code, data, and annotations used in this work are available at http://groups.csail.mit.edu/rbg/code/rl-hli/ Chapman, 1988; Di Eugenio and White, 1992; Di Eugenio, 1992; Webber et al., 1995). Given both the model and the rules, logic-based inference is used to automatically fill in the intermediate steps missing from the original instructions. Our approach, in contrast, operates directly on the textual instructions in the context of the interactive environment, while requiring no additional information. By interacting with the environment and observing the resulting feedback, our method automatically learns both the mapping between the text and the commands, and the underlying model of the environment. One particularly noteworthy aspect of our solution is the interplay between the evolving mapping and the progressively acquired environment model as the system learns how to interpret the text. Recording the state transitions observed during interpretation allows the algorithm to construct a relevant model of the environment. At the same time, the environment model enables the algorithm to consider the consequences of commands before they are executed, thereby improving the accuracy of interpretation. Our method efficiently achieves both of these goals as part of a policy-gradient reinforcement learning algorithm. We apply our method to the task of mapping software troubleshooting guides to GUI actions in the Windows environment (Branavan et al., 2009; Kushman et al., 2009). The key findings of our experiments are threefold. First, the algorithm can accurately interpret 61.5% of high-level instructions, which cannot be handled by previous statistical systems. Second, we demonstrate that explicitly modeling the environment also greatly improves the accuracy of processing low-level instructions, yielding a 14% absolute increase in performance over a competitive baseline (Branavan et al., 2009). Finally, we show the importance of constructing an environment model relevant to the language interpretation task — using textual 1268 "open control panel, double click system, then go to the advanced tab" Document (input): "open control panel" left-click Advanced double-click System left-click Control Panel left-click Settings left-click Start Instructions: high-level instruction low-level instructions Command Sequence (output): : : : :: ::: ::: "double click system" "go to the advanced tab" : : Figure 1: An example mapping of a document containing high-level instructions into a candidate sequence of five commands. The mapping process involves segmenting the document into individual instruction word spans Wa, and translating each instruction into the sequence ⃗c of one or more commands it describes. During learning, the correct output command sequence is not provided to the algorithm. instructions enables us to bias exploration toward transitions relevant for language learning. This approach yields superior performance compared to a policy that relies on an environment model constructed via random exploration. 2 Related Work Interpreting Instructions Our approach is most closely related to the reinforcement learning algorithm for mapping text instructions to commands developed by Branavan et al. (2009) (see Section 4 for more detail). Their method is predicated on the assumption that each command to be executed is explicitly specified in the instruction text. This assumption of a direct correspondence between the text and the environment is not unique to that paper, being inherent in other work on grounded language learning (Siskind, 2001; Oates, 2001; Yu and Ballard, 2004; Fleischman and Roy, 2005; Mooney, 2008; Liang et al., 2009; Matuszek et al., 2010). A notable exception is the approach of Eisenstein et al. (2009), which learns how an environment operates by reading text, rather than learning an explicit mapping from the text to the environment. For example, their method can learn the rules of a card game given instructions for how to play. Many instances of work on instruction interpretation are replete with examples where instructions are formulated as high-level goals, targeted at users with relevant knowledge (Winograd, 1972; Di Eugenio, 1992; Webber et al., 1995; MacMahon et al., 2006). Not surprisingly, automatic approaches for processing such instructions have relied on hand-engineered world knowledge to reason about the preconditions and effects of environment commands. The assumption of a fully specified environment model is also common in work on semantics in the linguistics literature (Lascarides and Asher, 2004). While our approach learns to analyze instructions in a goaldirected manner, it does not require manual specification of relevant environment knowledge. Reinforcement Learning Our work combines ideas of two traditionally disparate approaches to reinforcement learning (Sutton and Barto, 1998). The first approach, model-based learning, constructs a model of the environment in which the learner operates (e.g., modeling location, velocity, and acceleration in robot navigation). It then computes a policy directly from the rich information represented in the induced environment model. In the NLP literature, model-based reinforcement learning techniques are commonly used for dialog management (Singh et al., 2002; Lemon and Konstas, 2009; Schatzmann and Young, 2009). However, if the environment cannot be accurately approximated by a compact representation, these methods perform poorly (Boyan and Moore, 1995; Jong and Stone, 2007). Our instruction interpretation task falls into this latter category,2 rendering standard model-based learning ineffective. The second approach – model-free methods such as policy learning – aims to select the opti2For example, in the Windows GUI domain, clicking on the File menu will result in a different submenu depending on the application. Thus it is impossible to predict the effects of a previously unseen GUI command. 1269 Policy function clicking start word span : LEFT_CLICK( ) start command : Observed text and environment Select run after clicking start. In the open box type "dcomcnfg". State Observed text and environment Select run after clicking start. In the open box type "dcomcnfg". State Action Figure 2: A single step in the instruction mapping process formalized as an MDP. State s is comprised of the state of the external environment E, and the state of the document (d, W), where W is the list of all word spans mapped by previous actions. An action a selects a span Wa of unused words from (d, W), and maps them to an environment command c. As a consequence of a, the environment state changes to E′ ∼p(E′|E, c), and the list of mapped words is updated to W ′ = W ∪Wa. mal action at every step, without explicitly constructing a model of the environment. While policy learners can effectively operate in complex environments, they are not designed to benefit from a learned environment model. We address this limitation by expanding a policy learning algorithm to take advantage of a partial environment model estimated during learning. The approach of conditioning the policy function on future reachable states is similar in concept to the use of postdecision state information in the approximate dynamic programming framework (Powell, 2007). 3 Problem Formulation Our goal is to map instructions expressed in a natural language document d into the corresponding sequence of commands ⃗c = ⟨c1, . . . , cm⟩executable in an environment. As input, we are given a set of raw instruction documents, an environment, and a reward function as described below. The environment is formalized as its states and transition function. An environment state E specifies the objects accessible in the environment at a given time step, along with the objects’ properties. The environment state transition function p(E′|E, c) encodes how the state changes from E to E′ in response to a command c.3 During learning, this function is not known, but samples from it can be collected by executing commands and ob3While in the general case the environment state transitions maybe stochastic, they are deterministic in the software GUI used in this work. serving the resulting environment state. A realvalued reward function measures how well a command sequence⃗c achieves the task described in the document. We posit that a document d is composed of a sequence of instructions, each of which can take one of two forms: • Low-level instructions: these explicitly describe single commands.4 E.g., “double click system” in Figure 1. • High-level instructions: these correspond to a sequence of one or more environment commands, none of which are explicitly described by the instruction. E.g., “open control panel” in Figure 1. 4 Background Our innovation takes place within a previously established general framework for the task of mapping instructions to commands (Branavan et al., 2009). This framework formalizes the mapping process as a Markov Decision Process (MDP) (Sutton and Barto, 1998), with actions encoding individual instruction-to-command mappings, and states representing partial interpretations of the document. In this section, we review the details of this framework. 4Previous work (Branavan et al., 2009) is only able to handle low-level instructions. 1270 starting environment state parts of the environment state space reachable after commands and . state where a control panel icon was observed during previous exploration steps. Figure 3: Using information derived from future states to interpret the high-level instruction “open control panel.” Ed is the starting state, and c1 through c4 are candidate commands. Environment states are shown as circles, with previously visited environment states colored green. Dotted arrows show known state transitions. All else being equal, the information that the control panel icon was observed in state E5 during previous exploration steps can help to correctly select command c3. States and Actions A document is interpreted by incrementally constructing a sequence of actions. Each action selects a word span from the document, and maps it to one environment command. To predict actions sequentially, we track the states of the environment and the document over time as shown in Figure 2. This mapping state s is a tuple (E, d, W) where E is the current environment state, d is the document being interpreted, and W is the list of word spans selected by previous actions. The mapping state s is observed prior to selecting each action. The mapping action a is a tuple (c, Wa) that represents the joint selection of a span of words Wa and an environment command c. Some of the candidate actions would correspond to the correct instruction mappings, e.g., (c = double-click system, Wa = “double click system”). Others such as (c = left-click system, Wa = “double click system”) would be erroneous. The algorithm learns to interpret instructions by learning to construct sequences of actions that assign the correct commands to the words. The interpretation of a document d begins at an initial mapping state s0 = (Ed, d, ∅), Ed being the starting state of the environment for the document. Given a state s = (E, d, W), the space of possible actions a = (c, Wa) is defined by enumerating sub-spans of unused words in d and candidate commands in E.5 The action to execute, a, is selected based on a policy function p(a|s) by finding arg maxa p(a|s). Performing action a in state 5Here, command reordering is possible. At each step, the span of selected words Wa is not required to be adjacent to the previous selections. This reordering is used to interpret sentences such as “Select exit after opening the File menu.” s = (E, d, W) results in a new state s′ according to the distribution p(s′|s, a), where: a = (c, Wa), E′ ∼ p(E′|E, c), W ′ = W ∪Wa, s′ = (E′, d, W ′). The process of selecting and executing actions is repeated until all the words in d have been mapped.6 A Log-Linear Parameterization The policy function used for action selection is defined as a log-linear distribution over actions: p(a|s; θ) = eθ·φ(s,a) X a′ eθ·φ(s,a′) , (1) where θ ∈Rn is a weight vector, and φ(s, a) ∈Rn is an n-dimensional feature function. This representation has the flexibility to incorporate a variety of features computed on the states and actions. Reinforcement Learning Parameters of the policy function p(a|s; θ) are estimated to maximize the expected future reward for analyzing each document d ∈D: θ = arg max θ Ep(h|θ) [r(h)] , (2) where h = (s0, a0, . . . , sm−1, am−1, sm) is a history that records the analysis of document d, p(h|θ) is the probability of selecting this analysis given policy parameters θ, and the reward r(h) is a real valued indication of the quality of h. 6To account for document words that are not part of an instruction, c may be a null command. 1271 5 Algorithm We expand the scope of learning approaches for automatic document interpretation by enabling the analysis of high-level instructions. The main challenge in processing these instructions is that, in contrast to their low-level counterparts, they correspond to sequences of one or more commands. A simple way to enable this one-to-many mapping is to allow actions that do not consume words (i.e., |Wa| = 0). The sequence of actions can then be constructed incrementally using the algorithm described above. However, this change significantly complicates the interpretation problem – we need to be able to predict commands that are not directly described by any words, and allowing action sequences significantly increases the space of possibilities for each instruction. Since we cannot enumerate all possible sequences at decision time, we limit the space of possibilities by learning which sequences are likely to be relevant for the current instruction. To motivate the approach, consider the decision problem in Figure 3, where we need to find a command sequence for the high-level instruction “open control panel.” The algorithm focuses on command sequences leading to environment states where the control panel icon was previously observed. The information about such states is acquired during exploration and is stored in a partial environment model q(E′|E, c). Our goal is to map high-level instructions to command sequences by leveraging knowledge about the long-term effects of commands. We do this by integrating the partial environment model into the policy function. Specifically, we modify the log-linear policy p(a|s; q, θ) by adding lookahead features φ(s, a, q) which complement the local features used in the previous model. These look-ahead features incorporate various measurements that characterize the potential of future states reachable via the selected action. Although primarily designed to analyze high-level instructions, this approach is also useful for mapping low-level instructions. Below, we first describe how we estimate the partial environment transition model and how this model is used to compute the look-ahead features. This is followed by the details of parameter estimation for our algorithm. 5.1 Partial Environment Transition Model To compute the look-ahead features, we first need to collect statistics about the environment transition function p(E′|E, c). An example of an environment transition is the change caused by clicking on the “start” button. We collect this information through observation, and build a partial environment transition model q(E′|E, c). One possible strategy for constructing q is to observe the effects of executing random commands in the environment. In a complex environment, however, such a strategy is unlikely to produce state samples relevant to our text analysis task. Instead, we use the training documents to guide the sampling process. During training, we execute the command sequences predicted by the policy function in the environment, caching the resulting state transitions. Initially, these commands may have little connection to the actual instructions. As learning progresses and the quality of the interpretation improves, more promising parts of the environment will be observed. This process yields samples that are biased toward the content of the documents. 5.2 Look-Ahead Features We wish to select actions that allow for the best follow-up actions, thereby finding the analysis with the highest total reward for a given document. In practice, however, we do not have information about the effects of all possible future actions. Instead, we capitalize on the state transitions observed during the sampling process described above, allowing us to incrementally build an environment model of actions and their effects. Based on this transition information, we can estimate the usefulness of actions by considering the properties of states they can reach. For instance, some states might have very low immediate reward, indicating that they are unlikely to be part of the best analysis for the document. While the usefulness of most states is hard to determine, it correlates with various properties of the state. We encode the following properties as look-ahead features in our policy: • The highest reward achievable by an action sequence passing through this state. This property is computed using the learned environment model, and is therefore an approximation. 1272 • The length of the above action sequence. • The average reward received at the environment state while interpreting any document. This property introduces a bias towards commonly visited states that frequently recur throughout multiple documents’ correct interpretations. Because we can never encounter all states and all actions, our environment model is always incomplete and these properties can only be computed based on partial information. Moreover, the predictive strength of the properties is not known in advance. Therefore we incorporate them as separate features in the model, and allow the learning process to estimate their weights. In particular, we select actions a based on the current state s and the partial environment model q, resulting in the following policy definition: p(a|s; q, θ) = eθ·φ(s,a,q) X a′ eθ·φ(s,a′,q) , (3) where the feature representation φ(s, a, q) has been extended to be a function of q. 5.3 Parameter Estimation The learning algorithm is provided with a set of documents d ∈D, an environment in which to execute command sequences ⃗c, and a reward function r(h). The goal is to estimate two sets of parameters: 1) the parameters θ of the policy function, and 2) the partial environment transition model q(E′|E, c), which is the observed portion of the true model p(E′|E, c). These parameters are mutually dependent: θ is defined over a feature space dependent on q, and q is sampled according to the policy function parameterized by θ. Algorithm 1 shows the procedure for joint learning of these parameters. As in standard policy gradient learning (Sutton et al., 2000), the algorithm iterates over all documents d ∈D (steps 1, 2), selecting and executing actions in the environment (steps 3 to 6). The resulting reward is used to update the parameters θ (steps 8, 9). In the new joint learning setting, this process also yields samples of state transitions which are used to estimate q(E′|E, c) (step 7). This updated q is then used to compute the feature functions φ(s, a, q) during the next iteration of learning (step 4). This process is repeated until the total reward on training documents converges. Input: A document set D, Feature function φ, Reward function r(h), Number of iterations T Initialization: Set θ to small random values. Set q to the empty set. for i = 1 · · · T do 1 foreach d ∈D do 2 Sample history h ∼p(h|θ) where h = (s0, a0, · · · , an−1, sn) as follows: Initialize environment to document specific starting state Ed for t = 0 · · · n −1 do 3 Compute φ(a, st, q) based on latest q 4 Sample action at ∼p(a|st; q, θ) 5 Execute at on state st: st+1 ∼p(s|st, at) 6 Set q = q ∪{(E′, E, c)} where E′, E, c are the 7 environment states and commands from st+1, st, and at end ∆← 8 X t " φ(st, at, q) − X a′ φ(st, a′, q) p(a′|st; q, θ) # θ ←θ + r(h)∆ 9 end end Output: Estimate of parameters θ Algorithm 1: A policy gradient algorithm that also learns a model of the environment. This algorithm capitalizes on the synergy between θ and q. As learning proceeds, the method discovers a more complete state transition function q, which improves the accuracy of the look-ahead features, and ultimately, the quality of the resulting policy. An improved policy function in turn produces state samples that are more relevant to the document interpretation task. 6 Applying the Model We apply our algorithm to the task of interpreting help documents to perform software related tasks (Branavan et al., 2009; Kushman et al., 2009). Specifically, we consider documents from Microsoft’s Help and Support website.7 As in prior work, we use a virtual machine set-up to allow our method to interact with a Windows 2000 environment. Environment States and Actions In this application of our model, the environment state is the set of visible user interface (UI) objects, along 7http://support.microsoft.com/ 1273 with their properties (e.g., the object’s label, parent window, etc). The environment commands consist of the UI commands left-click, right-click, double-click, and type-into. Each of these commands requires a UI object as a parameter, while type-into needs an additional parameter containing the text to be typed. On average, at each step of the interpretation process, the branching factor is 27.14 commands. Reward Function An ideal reward function would be to verify whether the task specified by the help document was correctly completed. Since such verification is a challenging task, we rely on a noisy approximation: we assume that each sentence specifies at least one command, and that the text describing the command has words matching the label of the environment object. If a history h has at least one such command for each sentence, the environment reward function r(h) returns a positive value, otherwise it returns a negative value. This environment reward function is a simplification of the one described in Branavan et al. (2009), and it performs comparably in our experiments. Features In addition to the look-ahead features described in Section 5.2, the policy also includes the set of features used by Branavan et al. (2009). These features are functions of both the text and environment state, modeling local properties that are useful for action selection. 7 Experimental Setup Datasets Our model is trained on the same dataset used by Branavan et al. (2009). For testing we use two datasets: the first one was used in prior work and contains only low-level instructions, while the second dataset is comprised of documents with high-level instructions. This new dataset was collected from the Microsoft Help and Support website, and has on average 1.03 high-level instructions per document. The second dataset contains 60 test documents, while the first is split into 70, 18 and 40 document for training, development and testing respectively. The combined statistics for these datasets is shown below: Total # of documents 188 Total # of words 7448 Vocabulary size 739 Avg. actions per document 10 Reinforcement Learning Parameters Following common practice, we encourage exploration during learning with an ϵ-greedy strategy (Sutton and Barto, 1998), with ϵ set to 0.1. We also identify dead-end states, i.e. states with the lowest possible immediate reward, and use the induced environment model to encourage additional exploration by lowering the likelihood of actions that lead to such dead-end states. During the early stages of learning, experience gathered in the environment model is extremely sparse, causing the look-ahead features to provide poor estimates. To speed convergence, we ignore these estimates by disabling the look-ahead features for a fixed number of initial training iterations. Finally, to guarantee convergence, stochastic gradient ascent algorithms require a learning rate schedule. We use a modified search-thenconverge algorithm (Darken and Moody, 1990), and tie the learning rate to the ratio of training documents that received a positive reward in the current iteration. Baselines As a baseline, we compare our method against the results reported by Branavan et al. (2009), denoted here as BCZB09. As an upper bound for model performance, we also evaluate our method using a reward signal that simulates a fully-supervised training regime. We define a reward function that returns positive one for histories that match the annotations, and zero otherwise. Performing policy-gradient with this function is equivalent to training a fullysupervised, stochastic gradient algorithm that optimizes conditional likelihood (Branavan et al., 2009). Evaluation Metrics We evaluate the accuracy of the generated mapping by comparing it against manual annotations of the correct action sequences. We measure the percentage of correct actions and the percentage of documents where every action is correct. In general, the sequential nature of the interpretation task makes it difficult to achieve high action accuracy. For example, executing an incorrect action early on, often leads to an environment state from which the remaining instructions cannot be completed. When this happens, it is not possible to recover the remaining actions, causing cascading errors that significantly reduce performance. 1274 Low-level instruction dataset High-level instruction dataset action document action high-level action document BCZB09 0.647 0.375 0.021 0.022 0.000 BCZB09 + annotation ∗0.756 0.525 0.035 0.022 0.000 Our model 0.793 0.517 ∗0.419 ∗0.615 ∗0.283 Our model + annotation 0.793 0.650 ∗0.357 0.492 0.333 Table 1: Accuracy of the mapping produced by our model, its variants, and the baseline. Values marked with ∗are statistically significant at p < 0.01 compared to the value immediately above it. 8 Results As shown in Table 1, our model outperforms the baseline on the two datasets, according to all evaluation metrics. In contrast to the baseline, our model can handle high-level instructions, accurately interpreting 62% of them in the second dataset. Every document in this set contains at least one high-level action, which on average, maps to 3.11 environment commands each. The overall action performance on this dataset, however, seems unexpectedly low at 42%. This discrepancy is explained by the fact that in this dataset, high-level instructions are often located towards the beginning of the document. If these initial challenging instructions are not processed correctly, the rest of the actions for the document cannot be interpreted. As the performance on the first dataset indicates, the new algorithm is also beneficial for processing low-level instructions. The model outperforms the baseline by at least 14%, both in terms of the actions and the documents it can process. Not surprisingly, the best performance is achieved when the new algorithm has access to manually annotated data during training. We also performed experiments to validate the intuition that the partial environment model must contain information relevant for the language interpretation task. To test this hypothesis, we replaced the learned environment model with one of the same size gathered by executing random commands. The model with randomly sampled environment transitions performs poorly: it can only process 4.6% of documents and 15% of actions on the dataset with high-level instructions, compared to 28.3% and 41.9% respectively for our algorithm. This result also explains why training with full supervision hurts performance on highlevel instructions (see Table 1). Learning directly from annotations results in a low-quality environment model due to the relative lack of exploration, High-level instruction ∘ open device manager Extracted low-level instruction paraphrase ∘ double click my computer ∘ double click control panel ∘ double click administrative tools ∘ double click computer management ∘ double click device manager High-level instruction ∘ open the network tool in control panel Extracted low-level instruction paraphrase ∘ click start ∘ point to settings ∘ click control panel ∘ double click network and dial-up connections Figure 4: Examples of automatically generated paraphrases for high-level instructions. The model maps the high-level instruction into a sequence of commands, and then translates them into the corresponding low-level instructions. hurting the model’s ability to leverage the lookahead features. Finally, to demonstrate the quality of the learned word–command alignments, we evaluate our method’s ability to paraphrase from high-level instructions to low-level instructions. Here, the goal is to take each high-level instruction and construct a text description of the steps required to achieve it. We did this by finding high-level instructions where each of the commands they are associated with is also described by a low-level instruction in some other document. For example, if the text “open control panel” was mapped to the three commands in Figure 1, and each of those commands was described by a low-level instruction elsewhere, this procedure would create a paraphrase such as “click start, left click setting, and select control panel.” Of the 60 highlevel instructions tagged in the test set, this approach found paraphrases for 33 of them. 29 of 1275 these paraphrases were correct, in the sense that they describe all the necessary commands. Figure 4 shows some examples of the automatically extracted paraphrases. 9 Conclusions and Future Work In this paper, we demonstrate that knowledge about the environment can be learned and used effectively for the task of mapping instructions to actions. A key feature of this approach is the synergy between language analysis and the construction of the environment model: instruction text drives the sampling of the environment transitions, while the acquired environment model facilitates language interpretation. This design enables us to learn to map high-level instructions while also improving accuracy on low-level instructions. To apply the above method to process a broad range of natural language documents, we need to handle several important semantic and pragmatic phenomena, such as reference, quantification, and conditional statements. These linguistic constructions are known to be challenging to learn – existing approaches commonly rely on large amounts of hand annotated data for training. An interesting avenue of future work is to explore an alternative approach which learns these phenomena by combining linguistic information with knowledge gleaned from an automatically induced environment model. Acknowledgments The authors acknowledge the support of the NSF (CAREER grant IIS-0448168, grant IIS0835445, and grant IIS-0835652) and the Microsoft Research New Faculty Fellowship. Thanks to Aria Haghighi, Leslie Pack Kaelbling, Tom Kwiatkowski, Martin Rinard, David Silver, Mark Steedman, Csaba Szepesvari, the MIT NLP group, and the ACL reviewers for their suggestions and comments. Any opinions, findings, conclusions, or recommendations expressed in this paper are those of the authors, and do not necessarily reflect the views of the funding organizations. References Philip E. Agre and David Chapman. 1988. What are plans for? Technical report, Cambridge, MA, USA. J. A. Boyan and A. W. Moore. 1995. Generalization in reinforcement learning: Safely approximating the value function. In Advances in NIPS, pages 369– 376. S.R.K Branavan, Harr Chen, Luke Zettlemoyer, and Regina Barzilay. 2009. Reinforcement learning for mapping instructions to actions. In Proceedings of ACL, pages 82–90. Christian Darken and John Moody. 1990. Note on learning rate schedules for stochastic optimization. In Advances in NIPS, pages 832–838. Barbara Di Eugenio and Michael White. 1992. On the interpretation of natural language instructions. In Proceedings of COLING, pages 1147–1151. Barbara Di Eugenio. 1992. Understanding natural language instructions: the case of purpose clauses. In Proceedings of ACL, pages 120–127. Jacob Eisenstein, James Clarke, Dan Goldwasser, and Dan Roth. 2009. Reading to learn: Constructing features from semantic abstracts. In Proceedings of EMNLP, pages 958–967. Michael Fleischman and Deb Roy. 2005. Intentional context in situated natural language learning. In Proceedings of CoNLL, pages 104–111. Nicholas K. Jong and Peter Stone. 2007. Model-based function approximation in reinforcement learning. In Proceedings of AAMAS, pages 670–677. Nate Kushman, Micah Brodsky, S.R.K. Branavan, Dina Katabi, Regina Barzilay, and Martin Rinard. 2009. Wikido. In Proceedings of HotNets-VIII. Alex Lascarides and Nicholas Asher. 2004. Imperatives in dialogue. In P. Kuehnlein, H. Rieser, and H. Zeevat, editors, The Semantics and Pragmatics of Dialogue for the New Millenium. Benjamins. Oliver Lemon and Ioannis Konstas. 2009. User simulations for context-sensitive speech recognition in spoken dialogue systems. In Proceedings of EACL, pages 505–513. Percy Liang, Michael I. Jordan, and Dan Klein. 2009. Learning semantic correspondences with less supervision. In Proceedings of ACL, pages 91–99. Matt MacMahon, Brian Stankiewicz, and Benjamin Kuipers. 2006. Walk the talk: connecting language, knowledge, and action in route instructions. In Proceedings of AAAI, pages 1475–1482. C. Matuszek, D. Fox, and K. Koscher. 2010. Following directions using statistical machine translation. In Proceedings of Human-Robot Interaction, pages 251–258. Raymond J. Mooney. 2008. Learning to connect language and perception. In Proceedings of AAAI, pages 1598–1601. 1276 James Timothy Oates. 2001. Grounding knowledge in sensors: Unsupervised learning for language and planning. Ph.D. thesis, University of Massachusetts Amherst. Warren B Powell. 2007. Approximate Dynamic Programming. Wiley-Interscience. Jost Schatzmann and Steve Young. 2009. The hidden agenda user simulation model. IEEE Trans. Audio, Speech and Language Processing, 17(4):733–747. Satinder Singh, Diane Litman, Michael Kearns, and Marilyn Walker. 2002. Optimizing dialogue management with reinforcement learning: Experiments with the njfun system. Journal of Artificial Intelligence Research, 16:105–133. Jeffrey Mark Siskind. 2001. Grounding the lexical semantics of verbs in visual perception using force dynamics and event logic. Journal of Artificial Intelligence Research, 15:31–90. Richard S. Sutton and Andrew G. Barto. 1998. Reinforcement Learning: An Introduction. The MIT Press. Richard S. Sutton, David McAllester, Satinder Singh, and Yishay Mansour. 2000. Policy gradient methods for reinforcement learning with function approximation. In Advances in NIPS, pages 1057–1063. Bonnie Webber, Norman Badler, Barbara Di Eugenio, Libby Levison Chris Geib, and Michael Moore. 1995. Instructions, intentions and expectations. Artificial Intelligence, 73(1-2). Terry Winograd. 1972. Understanding Natural Language. Academic Press. Chen Yu and Dana H. Ballard. 2004. On the integration of grounding language and learning objects. In Proceedings of AAAI, pages 488–493. 1277
|
2010
|
129
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 118–127, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Open Information Extraction using Wikipedia Fei Wu University of Washington Seattle, WA, USA [email protected] Daniel S. Weld University of Washington Seattle, WA, USA [email protected] Abstract Information-extraction (IE) systems seek to distill semantic relations from naturallanguage text, but most systems use supervised learning of relation-specific examples and are thus limited by the availability of training data. Open IE systems such as TextRunner, on the other hand, aim to handle the unbounded number of relations found on the Web. But how well can these open systems perform? This paper presents WOE, an open IE system which improves dramatically on TextRunner’s precision and recall. The key to WOE’s performance is a novel form of self-supervised learning for open extractors — using heuristic matches between Wikipedia infobox attribute values and corresponding sentences to construct training data. Like TextRunner, WOE’s extractor eschews lexicalized features and handles an unbounded set of semantic relations. WOE can operate in two modes: when restricted to POS tag features, it runs as quickly as TextRunner, but when set to use dependency-parse features its precision and recall rise even higher. 1 Introduction The problem of information-extraction (IE), generating relational data from natural-language text, has received increasing attention in recent years. A large, high-quality repository of extracted tuples can potentially benefit a wide range of NLP tasks such as question answering, ontology learning, and summarization. The vast majority of IE work uses supervised learning of relationspecific examples. For example, the WebKB project (Craven et al., 1998) used labeled examples of the courses-taught-by relation to induce rules for identifying additional instances of the relation. While these methods can achieve high precision and recall, they are limited by the availability of training data and are unlikely to scale to the thousands of relations found in text on the Web. An alternative paradigm, Open IE, pioneered by the TextRunner system (Banko et al., 2007) and the “preemptive IE” in (Shinyama and Sekine, 2006), aims to handle an unbounded number of relations and run quickly enough to process Webscale corpora. Domain independence is achieved by extracting the relation name as well as its two arguments. Most open IE systems use selfsupervised learning, in which automatic heuristics generate labeled data for training the extractor. For example, TextRunner uses a small set of handwritten rules to heuristically label training examples from sentences in the Penn Treebank. This paper presents WOE (Wikipedia-based Open Extractor), the first system that autonomously transfers knowledge from random editors’ effort of collaboratively editing Wikipedia to train an open information extractor. Specifically, WOE generates relation-specific training examples by matching Infobox1 attribute values to corresponding sentences (as done in Kylin (Wu and Weld, 2007) and Luchs (Hoffmann et al., 2010)), but WOE abstracts these examples to relationindependent training data to learn an unlexicalized extractor, akin to that of TextRunner. WOE can operate in two modes: when restricted to shallow features like part-of-speech (POS) tags, it runs as quickly as Textrunner, but when set to use dependency-parse features its precision and recall rise even higher. We present a thorough experimental evaluation, making the following contributions: • We present WOE, a new approach to open IE that uses Wikipedia for self-supervised learn1An infobox is a set of tuples summarizing the key attributes of the subject in a Wikipedia article. For example, the infobox in the article on “Sweden” contains attributes like Capital, Population and GDP. 118 ing of unlexicalized extractors. Compared with TextRunner (the state of the art) on three corpora, WOE yields between 72% and 91% improved F-measure — generalizing well beyond Wikipedia. • Using the same learning algorithm and features as TextRunner, we compare four different ways to generate positive and negative training data with TextRunner’s method, concluding that our Wikipedia heuristic is responsible for the bulk of WOE’s improved accuracy. • The biggest win arises from using parser features. Previous work (Jiang and Zhai, 2007) concluded that parser-based features are unnecessary for information extraction, but that work assumed the presence of lexical features. We show that abstract dependency paths are a highly informative feature when performing unlexicalized extraction. 2 Problem Definition An open information extractor is a function from a document, d, to a set of triples, {⟨arg1, rel, arg2⟩}, where the args are noun phrases and rel is a textual fragment indicating an implicit, semantic relation between the two noun phrases. The extractor should produce one triple for every relation stated explicitly in the text, but is not required to infer implicit facts. In this paper, we assume that all relational instances are stated within a single sentence. Note the difference between open IE and the traditional approaches (e.g., as in WebKB), where the task is to decide whether some pre-defined relation holds between (two) arguments in the sentence. We wish to learn an open extractor without direct supervision, i.e. without annotated training examples or hand-crafted patterns. Our input is Wikipedia, a collaboratively-constructed encyclopedia2. As output, WOE produces an unlexicalized and relation-independent open extractor. Our objective is an extractor which generalizes beyond Wikipedia, handling other corpora such as the general Web. 3 Wikipedia-based Open IE The key idea underlying WOE is the automatic construction of training examples by heuristically matching Wikipedia infobox values and corresponding text; these examples are used to generate 2We also use DBpedia (Auer and Lehmann, 2007) as a collection of conveniently parsed Wikipedia infoboxes Sentence Splitting NLP Annotating Synonyms Compiling Preprocessor Primary Entity Matching Sentence Matching Matcher Triples Pattern Classifier over Parser Features CRF Extractor over Shallow Features Learner Figure 1: Architecture of WOE. an unlexicalized, relation-independent (open) extractor. As shown in Figure 1, WOE has three main components: preprocessor, matcher, and learner. 3.1 Preprocessor The preprocessor converts the raw Wikipedia text into a sequence of sentences, attaches NLP annotations, and builds synonym sets for key entities. The resulting data is fed to the matcher, described in Section 3.2, which generates the training set. Sentence Splitting: The preprocessor first renders each Wikipedia article into HTML, then splits the article into sentences using OpenNLP. NLP Annotation: As we discuss fully in Section 4 (Experiments), we consider several variations of our system; one version, WOEparse, uses parser-based features, while another, WOEpos, uses shallow features like POS tags, which may be more quickly computed. Depending on which version is being trained, the preprocessor uses OpenNLP to supply POS tags and NP-chunk annotations — or uses the Stanford Parser to create a dependency parse. When parsing, we force the hyperlinked anchor texts to be a single token by connecting the words with an underscore; this transformation improves parsing performance in many cases. Compiling Synonyms: As a final step, the preprocessor builds sets of synonyms to help the matcher find sentences that correspond to infobox relations. This is useful because Wikipedia editors frequently use multiple names for an entity; for example, in the article titled “University of Washington” the token “UW” is widely used to refer the university. Additionally, attribute values are often described differently within the infobox than they are in surrounding text. Without knowledge of these synonyms, it is impossible to construct good matches. Following (Wu and Weld, 2007; Nakayama and Nishio, 2008), the preprocessor uses Wikipedia redirection pages and back119 ward links to automatically construct synonym sets. Redirection pages are a natural choice, because they explicitly encode synonyms; for example, “USA” is redirected to the article on the “United States.” Backward links for a Wikipedia entity such as the “Massachusetts Institute of Technology” are hyperlinks pointing to this entity from other articles; the anchor text of such links (e.g., “MIT”) forms another source of synonyms. 3.2 Matcher The matcher constructs training data for the learner component by heuristically matching attribute-value pairs from Wikipedia articles containing infoboxes with corresponding sentences in the article. Given the article on “Stanford University,” for example, the matcher should associate ⟨established, 1891⟩with the sentence “The university was founded in 1891 by . . . ” Given a Wikipedia page with an infobox, the matcher iterates through all its attributes looking for a unique sentence that contains references to both the subject of the article and the attribute value; these noun phrases will be annotated arg1 and arg2 in the training set. The matcher considers a sentence to contain the attribute value if the value or its synonym is present. Matching the article subject, however, is more involved. Matching Primary Entities: In order to match shorthand terms like “MIT” with more complete names, the matcher uses an ordered set of heuristics like those of (Wu and Weld, 2007; Nguyen et al., 2007): • Full match: strings matching the full name of the entity are selected. • Synonym set match: strings appearing in the entity’s synonym set are selected. • Partial match: strings matching a prefix or suffix of the entity’s name are selected. If the full name contains punctuation, only a prefix is allowed. For example, “Amherst” matches “Amherst, Mass,” but “Mass” does not. • Patterns of “the <type>”: The matcher first identifies the type of the entity (e.g., “city” for “Ithaca”), then instantiates the pattern to create the string “the city.” Since the first sentence of most Wikipedia articles is stylized (e.g. “The city of Ithaca sits . . . ”), a few patterns suffice to extract most entity types. • The most frequent pronoun: The matcher assumes that the article’s most frequent pronoun denotes the primary entity, e.g., “he” for the page on “Albert Einstein.” This heuristic is dropped when “it” is most common, because the word is used in too many other ways. When there are multiple matches to the primary entity in a sentence, the matcher picks the one which is closest to the matched infobox attribute value in the parser dependency graph. Matching Sentences: The matcher seeks a unique sentence to match the attribute value. To produce the best training set, the matcher performs three filterings. First, it skips the attribute completely when multiple sentences mention the value or its synonym. Second, it rejects the sentence if the subject and/or attribute value are not heads of the noun phrases containing them. Third, it discards the sentence if the subject and the attribute value do not appear in the same clause (or in parent/child clauses) in the parse tree. Since Wikipedia’s Wikimarkup language is semantically ambiguous, parsing infoboxes is surprisingly complex. Fortunately, DBpedia (Auer and Lehmann, 2007) provides a cleaned set of infoboxes from 1,027,744 articles. The matcher uses this data for attribute values, generating a training dataset with a total of 301,962 labeled sentences. 3.3 Learning Extractors We learn two kinds of extractors, one (WOEparse) using features from dependency-parse trees and the other (WOEpos) limited to shallow features like POS tags. WOEparse uses a pattern learner to classify whether the shortest dependency path between two noun phrases indicates a semantic relation. In contrast, WOEpos (like TextRunner) trains a conditional random field (CRF) to output certain text between noun phrases when the text denotes such a relation. Neither extractor uses individual words or lexical information for features. 3.3.1 Extraction with Parser Features Despite some evidence that parser-based features have limited utility in IE (Jiang and Zhai, 2007), we hoped dependency paths would improve precision on long sentences. Shortest Dependency Path as Relation: Unless otherwise noted, WOE uses the Stanford Parser to create dependencies in the “collapsedDependency” format. Dependencies involving prepositions, conjuncts as well as information about the referent of relative clauses are collapsed to get direct dependencies between content words. As 120 noted in (de Marneffe and Manning, 2008), this collapsed format often yields simplified patterns which are useful for relation extraction. Consider the sentence: Dan was not born in Berkeley. The Stanford Parser dependencies are: nsubjpass(born-4, Dan-1) auxpass(born-4, was-2) neg(born-4, not-3) prep in(born-4, Berkeley-6) where each atomic formula represents a binary dependence from dependent token to the governor token. These dependencies form a directed graph, ⟨V, E⟩, where each token is a vertex in V , and E is the set of dependencies. For any pair of tokens, such as “Dan” and “Berkeley”, we use the shortest connecting path to represent the possible relation between them: Dan −−−−−−−−−→ nsubjpass born ←−−−−−− prep in Berkeley We call such a path a corePath. While we will see that corePaths are useful for indicating when a relation exists between tokens, they don’t necessarily capture the semantics of that relation. For example, the path shown above doesn’t indicate the existence of negation! In order to capture the meaning of the relation, the learner augments the corePath into a tree by adding all adverbial and adjectival modifiers as well as dependencies like “neg” and “auxpass”. We call the result an expandPath as shown below: WOE traverses the expandPath with respect to the token orders in the original sentence when outputting the final expression of rel. Building a Database of Patterns: For each of the 301,962 sentences selected and annotated by the matcher, the learner generates a corePath between the tokens denoting the subject and the infobox attribute value. Since we are interested in eventually extracting “subject, relation, object” triples, the learner rejects corePaths that don’t start with subject-like dependencies, such as nsubj, nsubjpass, partmod and rcmod. This leads to a collection of 259,046 corePaths. To combat data sparsity and improve learning performance, the learner further generalizes the corePaths in this set to create a smaller set of generalized-corePaths. The idea is to eliminate distinctions which are irrelevant for recognizing (domain-independent) relations. Lexical words in corePaths are replaced with their POS tags. Further, all Noun POS tags and “PRP” are abstracted to “N”, all Verb POS tags to “V”, all Adverb POS tags to “RB” and all Adjective POS tags to “J”. The preposition dependencies such as “prep in” are generalized to “prep”. Take the corePath “Dan −−−−−−−−−→ nsubjpass born ←−−−−−− prep in Berkeley” for example, its generalized-corePath is “N −−−−−−−−−→ nsubjpass V ←−−−− prep N”. We call such a generalized-corePath an extraction pattern. In total, WOE builds a database (named DBp) of 15,333 distinct patterns and each pattern p is associated with a frequency — the number of matching sentences containing p. Specifically, 185 patterns have fp ≥100 and 1929 patterns have fp ≥5. Learning a Pattern Classifier: Given the large number of patterns in DBp, we assume few valid open extraction patterns are left behind. The learner builds a simple pattern classifier, named WOEparse, which checks whether the generalizedcorePath from a test triple is present in DBp, and computes the normalized logarithmic frequency as the probability3: w(p) = max(log(fp) −log(fmin), 0) log(fmax) −log(fmin) where fmax (50,259 in this paper) is the maximal frequency of pattern in DBp, and fmin (set 1 in this work) is the controlling threshold that determines the minimal frequency of a valid pattern. Take the previous sentence “Dan was not born in Berkeley” for example. WOEparse first identifies Dan as arg1 and Berkeley as arg2 based on NP-chunking. It then computes the corePath “Dan −−−−−−−−−→ nsubjpass born ←−−−−−− prep in Berkeley” and abstracts to p=“N −−−−−−−−−→ nsubjpass V ←−−−− prep N”. It then queries DBp to retrieve the frequency fp = 29112 and assigns a probability of 0.95. Finally, WOEparse traverses the triple’s expandPath to output the final expression ⟨Dan, wasNotBornIn, Berkeley⟩. As shown in the experiments on three corpora, WOEparse achieves an F-measure which is between 72% to 91% greater than TextRunner’s. 3.3.2 Extraction with Shallow Features WOEparse has a dramatic performance improvement over TextRunner. However, the improvement comes at the cost of speed — TextRunner 3How to learn a more sophisticated weighting function is left as a future topic. 121 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.0 0.2 0.4 0.6 0.8 1.0 recall precision P/R Curve on WSJ WOEparse WOEpos TextRunner 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.0 0.2 0.4 0.6 0.8 1.0 recall precision P/R Curve on Web WOEparse WOEpos TextRunner 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.0 0.2 0.4 0.6 0.8 1.0 recall precision P/R Curve on Wikipedia WOEparse WOEpos TextRunner Figure 2: WOEposperforms better than TextRunner, especially on precision. WOEparsedramatically improves performance, especially on recall. runs about 30X faster by only using shallow features. Since high speed can be crucial when processing Web-scale corpora, we additionally learn a CRF extractor WOEpos based on shallow features like POS-tags. In both cases, however, we generate training data from Wikipedia by matching sentences with infoboxes, while TextRunner used a small set of hand-written rules to label training examples from the Penn Treebank. We use the same matching sentence set behind DBp to generate positive examples for WOEpos. Specifically, for each matching sentence, we label the subject and infobox attribute value as arg1 and arg2 to serve as the ends of a linear CRF chain. Tokens involved in the expandPath are labeled as rel. Negative examples are generated from random noun-phrase pairs in other sentences when their generalized-CorePaths are not in DBp. WOEpos uses the same learning algorithm and selection of features as TextRunner: a two-order CRF chain model is trained with the Mallet package (McCallum, 2002). WOEpos’s features include POS-tags, regular expressions (e.g., for detecting capitalization, punctuation, etc..), and conjunctions of features occurring in adjacent positions within six words to the left and to the right of the current word. As shown in the experiments, WOEpos achieves an improved F-measure over TextRunner between 18% to 34% on three corpora, and this is mainly due to the increase on precision. 4 Experiments We used three corpora for experiments: WSJ from Penn Treebank, Wikipedia, and the general Web. For each dataset, we randomly selected 300 sentences. Each sentence was examined by two people to label all reasonable triples. These candidate triples are mixed with pseudo-negative ones and submitted to Amazon Mechanical Turk for verification. Each triple was examined by 5 Turkers. We mark a triple’s final label as positive when more than 3 Turkers marked them as positive. 4.1 Overall Performance Analysis In this section, we compare the overall performance of WOEparse, WOEpos and TextRunner (shared by the Turing Center at the University of Washington). In particular, we are going to answer the following questions: 1) How do these systems perform against each other? 2) How does performance vary w.r.t. sentence length? 3) How does extraction speed vary w.r.t. sentence length? Overall Performance Comparison The detailed P/R curves are shown in Figure 2. To have a close look, for each corpus, we randomly divided the 300 sentences into 5 groups and compared the best F-measures of three systems in Figure 3. We can see that: • WOEpos is better than TextRunner, especially on precision. This is due to better training data from Wikipedia via self-supervision. Section 4.2 discusses this in more detail. • WOEparse achieves the best performance, especially on recall. This is because the parser features help to handle complicated and longdistance relations in difficult sentences. In particular, WOEparse outputs 1.42 triples per sentence on average, while WOEpos outputs 1.05 and TextRunner outputs 0.75. Note that we measure TextRunner’s precision & recall differently than (Banko et al., 2007) did. Specifically, we compute the precision & recall based on all extractions, while Banko et al. counted only concrete triples where arg1 is a proper noun, arg2 is a proper noun or date, and 122 Figure 3: WOEposachieves an F-measure, which is between 18% and 34% better than TextRunner’s. WOEparseachieves an improvement between 72% and 91% over TextRunner. The error bar indicates one standard deviation. the frequency of rel is over a threshold. Our experiments show that focussing on concrete triples generally improves precision at the expense of recall.4 Of course, one can apply a concreteness filter to any open extractor in order to trade recall for precision. The extraction errors by WOEparse can be categorized into four classes. We illustrate them with the WSJ corpus. In total, WOEparse got 85 wrong extractions on WSJ, and they are caused by: 1) Incorrect arg1 and/or arg2 from NP-Chunking (18.6%); 2) A erroneous dependency parse from Stanford Parser (11.9%); 3) Inaccurate meaning (27.1%) — for example, ⟨she, isNominatedBy, PresidentBush⟩is wrongly extracted from the sentence “If she is nominated by President Bush ...”5; 4) A pattern inapplicable for the test sentence (42.4%). Note WOEparse is worse than WOEpos in the low recall region. This is mainly due to parsing errors (especially on long-distance dependencies), which misleads WOEparse to extract false highconfidence triples. WOEpos won’t suffer from such parsing errors. Therefore it has better precision on high-confidence extractions. We noticed that TextRunner has a dip point in the low recall region. There are two typical errors responsible for this. A sample error of the first type is ⟨Sources, sold, theCompany⟩ extracted from the sentence “Sources said 4For example, consider the Wikipedia corpus. From our 300 test sentences, TextRunner extracted 257 triples (at 72.0% precision) but only extracted 16 concrete triples (with 87.5% precision). 5These kind of errors might be excluded by monitoring whether sentences contain words such as ‘if,’ ‘suspect,’ ‘doubt,’ etc.. We leave this as a topic for the future. Figure 4: WOEparse’s F-measure decreases more slowly with sentence length than WOEpos and TextRunner, due to its better handling of difficult sentences using parser features. he sold the company”, where “Sources” is wrongly treated as the subject of the object clause. A sample error of the second type is ⟨thisY ear, willStarIn, theMovie⟩ extracted from the sentence “Coming up this year, Long will star in the new movie.”, where “this year” is wrongly treated as part of a compound subject. Taking the WSJ corpus for example, at the dip point with recall=0.002 and precision=0.059, these two types of errors account for 70% of all errors. Extraction Performance vs. Sentence Length We tested how extractors’ performance varies with sentence length; the results are shown in Figure 4. TextRunner and WOEpos have good performance on short sentences, but their performance deteriorates quickly as sentences get longer. This is because long sentences tend to have complicated and long-distance relations which are difficult for shallow features to capture. In contrast, WOEparse’s performance decreases more slowly w.r.t. sentence length. This is mainly because parser features are more useful for handling difficult sentences and they help WOEparse to maintain a good recall with only moderate loss of precision. Extraction Speed vs. Sentence Length We also tested the extraction speed of different extractors. We used Java for implementing the extractors, and tested on a Linux platform with a 2.4GHz CPU and 4G memory. On average, it takes WOEparse 0.679 seconds to process a sentence. For TextRunner and WOEpos, it only takes 0.022 seconds — 30X times faster. The detailed extraction speed vs. sentence length is in Figure 5, showing that TextRunner and WOEpos’s extraction time grows approximately linearly with sentence length, while WOEparse’s extraction time grows 123 Figure 5: Textrnner and WOEpos’s running time seems to grow linearly with sentence length, while WOEparse’s time grows quadratically. quadratically (R2 = 0.935) due to its reliance on parsing. 4.2 Self-supervision with Wikipedia Results in Better Training Data In this section, we consider how the process of matching Wikipedia infobox values to corresponding sentences results in better training data than the hand-written rules used by TextRunner. To compare with TextRunner, we tested four different ways to generate training examples from Wikipedia for learning a CRF extractor. Specifically, positive and/or negative examples are selected by TextRunner’s hand-written rules (tr for short), by WOE’s heuristic of matching sentences with infoboxes (w for short), or randomly (r for short). We use CRF+h1−h2 to denote a particular approach, where “+” means positive samples, “-” means negative samples, and hi ∈{tr, w, r}. In particular, “+w” results in 221,205 positive examples based on the matching sentence set6. All extractors are trained using about the same number of positive and negative examples. In contrast, TextRunner was trained with 91,687 positive examples and 96,795 negative examples generated from the WSJ dataset in Penn Treebank. The CRF extractors are trained using the same learning algorithm and feature selection as TextRunner. The detailed P/R curves are in Figure 6, showing that using WOE heuristics to label positive examples gives the biggest performance boost. CRF+tr−tr (trained using TextRunner’s heuristics) is slightly worse than TextRunner. Most likely, this is because TextRunner’s heuristics rely on parse trees to label training examples, 6This number is smaller than the total number of corePaths (259,046) because we require arg1 to appear before arg2 in a sentence — as specified by TextRunner. and the Stanford parse on Wikipedia is less accurate than the gold parse on WSJ. 4.3 Design Desiderata of WOEparse There are two interesting design choices in WOEparse: 1) whether to require arg1 to appear before arg2 (denoted as 1≺2) in the sentence; 2) whether to allow corePaths to contain prepositional phrase (PP) attachments (denoted as PPa). We tested how they affect the extraction performance; the results are shown in Figure 7. We can see that filtering PP attachments (PPa) gives a large precision boost with a noticeable loss in recall; enforcing a lexical ordering of relation arguments (1≺2) yields a smaller improvement in precision with small loss in recall. Take the WSJ corpus for example: setting 1≺2 and PPa achieves a precision of 0.792 (with recall of 0.558). By changing 1≺2 to 1∼2, the precision decreases to 0.773 (with recall of 0.595). By changing PPa to PPa and keeping 1≺2, the precision decreases to 0.642 (with recall of 0.687) — in particular, if we use gold parse, the precision decreases to 0.672 (with recall of 0.685). We set 1≺2 and PPa as default in WOEparse as a logical consequence of our preference for high precision over high recall. 4.3.1 Different parsing options We also tested how different parsing might effect WOEparse’s performance. We used three parsing options on the WSJ dataset: Stanford parsing, CJ50 parsing (Charniak and Johnson, 2005), and the gold parses from the Penn Treebank. The Stanford Parser is used to derive dependencies from CJ50 and gold parse trees. Figure 8 shows the detailed P/R curves. We can see that although today’s statistical parsers make errors, they have negligible effect on the accuracy of WOE. 5 Related Work Open or Traditional Information Extraction: Most existing work on IE is relation-specific. Occurrence-statistical models (Agichtein and Gravano, 2000; M. Ciaramita, 2005), graphical models (Peng and McCallum, 2004; Poon and Domingos, 2008), and kernel-based methods (Bunescu and R.Mooney, 2005) have been studied. Snow et al. (Snow et al., 2005) utilize WordNet to learn dependency path patterns for extracting the hypernym relation from text. Some seed-based frameworks are proposed for open-domain extraction (Pasca, 2008; Davidov et al., 2007; Davidov and Rappoport, 2008). These works focus 124 0.0 0.1 0.2 0.3 0.4 0.0 0.2 0.4 0.6 0.8 1.0 recall precision P/R Curve on WSJ CRF+w−w=WOEpos CRF+w−tr CRF+w−r CRF+tr−tr TextRunner 0.0 0.1 0.2 0.3 0.4 0.0 0.2 0.4 0.6 0.8 1.0 recall precision P/R Curve on Web CRF+w−w=WOEpos CRF+w−tr CRF+w−r CRF+tr−tr TextRunner 0.0 0.1 0.2 0.3 0.4 0.0 0.2 0.4 0.6 0.8 1.0 recall precision P/R Curve on Wikipedia CRF+w−w=WOEpos CRF+w−tr CRF+w−r CRF+tr−tr TextRunner Figure 6: Matching sentences with Wikipedia infoboxes results in better training data than the handwritten rules used by TextRunner. Figure 7: Filtering prepositional phrase attachments (PPa) shows a strong boost to precision, and we see a smaller boost from enforcing a lexical ordering of relation arguments (1≺2). 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.4 0.6 0.8 1.0 recall precision P/R Curve on WSJ WOEstanford parse =WOEparse WOECJ50 parse WOEgold parse Figure 8: Although today’s statistical parsers make errors, they have negligible effect on the accuracy of WOE compared to operation on gold standard, human-annotated data. on identifying general relations such as class attributes, while open IE aims to extract relation instances from given sentences. Another seedbased system StatSnowball (Zhu et al., 2009) can perform both relation-specific and open IE by iteratively generating weighted extraction patterns. Different from WOE, StatSnowball only employs shallow features and uses L1-normalization to weight patterns. Shinyama and Sekine proposed the “preemptive IE” framework to avoid relation-specificity (Shinyama and Sekine, 2006). They first group documents based on pairwise vector-space clustering, then apply an additional clustering to group entities based on documents clusters. The two clustering steps make it difficult to meet the scalability requirement necessary to process the Web. Mintz et al. (Mintz et al., 2009) uses Freebase to provide distant supervision for relation extraction. They applied a similar heuristic by matching Freebase tuples with unstructured sentences (Wikipedia articles in their experiments) to create features for learning relation extractors. Matching Freebase with arbitrary sentences instead of matching Wikipedia infobox with corresponding Wikipedia articles will potentially increase the size of matched sentences at a cost of accuracy. Also, their learned extractors are relation-specific. Alan Akbik et al. (Akbik and Broß, 2009) annotated 10,000 sentences parsed with LinkGrammar and selected 46 general linkpaths as patterns for relation extraction. In contrast, WOE learns 15,333 general patterns based on an automatically annotated set of 125 301,962 Wikipedia sentences. The KNext system (Durme and Schubert, 2008) performs open knowledge extraction via significant heuristics. Its output is knowledge represented as logical statements instead of information represented as segmented text fragments. Information Extraction with Wikipedia: The YAGO system (Suchanek et al., 2007) extends WordNet using facts extracted from Wikipedia categories. It only targets a limited number of predefined relations. Nakayama et al. (Nakayama and Nishio, 2008) parse selected Wikipedia sentences and perform extraction over the phrase structure trees based on several handcrafted patterns. Wu and Weld proposed the KYLIN system (Wu and Weld, 2007; Wu et al., 2008) which has the same spirit of matching Wikipedia sentences with infoboxes to learn CRF extractors. However, it only works for relations defined in Wikipedia infoboxes. Shallow or Deep Parsing: Shallow features, like POS tags, enable fast extraction over large-scale corpora (Davidov et al., 2007; Banko et al., 2007). Deep features are derived from parse trees with the hope of training better extractors (Zhang et al., 2006; Zhao and Grishman, 2005; Bunescu and Mooney, 2005; Wang, 2008). Jiang and Zhai (Jiang and Zhai, 2007) did a systematic exploration of the feature space for relation extraction on the ACE corpus. Their results showed limited advantage of parser features over shallow features for IE. However, our results imply that abstracted dependency path features are highly informative for open IE. There might be several reasons for the different observations. First, Jiang and Zhai’s results are tested for traditional IE where local lexicalized tokens might contain sufficient information to trigger a correct classification. The situation is different when features are completely unlexicalized in open IE. Second, as they noted, many relations defined in the ACE corpus are short-range relations which are easier for shallow features to capture. In practical corpora like the general Web, many sentences contain complicated long-distance relations. As we have shown experimentally, parser features are more powerful in handling such cases. 6 Conclusion This paper introduces WOE, a new approach to open IE that uses self-supervised learning over unlexicalized features, based on a heuristic match between Wikipedia infoboxes and corresponding text. WOE can run in two modes: a CRF extractor (WOEpos) trained with shallow features like POS tags; a pattern classfier (WOEparse) learned from dependency path patterns. Comparing with TextRunner, WOEpos runs at the same speed, but achieves an F-measure which is between 18% and 34% greater on three corpora; WOEparse achieves an F-measure which is between 72% and 91% higher than that of TextRunner, but runs about 30X times slower due to the time required for parsing. Our experiments uncovered two sources of WOE’s strong performance: 1) the Wikipedia heuristic is responsible for the bulk of WOE’s improved accuracy, but 2) dependency-parse features are highly informative when performing unlexicalized extraction. We note that this second conclusion disagrees with the findings in (Jiang and Zhai, 2007). In the future, we plan to run WOE over the billion document CMU ClueWeb09 corpus to compile a giant knowledge base for distribution to the NLP community. There are several ways to further improve WOE’s performance. Other data sources, such as Freebase, could be used to create an additional training dataset via self-supervision. For example, Mintz et al. consider all sentences containing both the subject and object of a Freebase record as matching sentences (Mintz et al., 2009); while they use this data to learn relation-specific extractors, one could also learn an open extractor. We are also interested in merging lexicalized and open extraction methods; the use of some domain-specific lexical features might help to improve WOE’s practical performance, but the best way to do this is unclear. Finally, we wish to combine WOEparse with WOEpos (e.g., with voting) to produce a system which maximizes precision at low recall. Acknowledgements We thank Oren Etzioni and Michele Banko from Turing Center at the University of Washington for providing the code of their software and useful discussions. We also thank Alan Ritter, Mausam, Peng Dai, Raphael Hoffmann, Xiao Ling, Stefan Schoenmackers, Andrey Kolobov and Daniel Suskin for valuable comments. This material is based upon work supported by the WRF / TJ Cable Professorship, a gift from Google and by the Air Force Research Laboratory (AFRL) under prime contract no. FA8750-09-C-0181. Any opinions, 126 findings, and conclusion or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the view of the Air Force Research Laboratory (AFRL). References E. Agichtein and L. Gravano. 2000. Snowball: Extracting relations from large plain-text collections. In ICDL. Alan Akbik and J¨ugen Broß. 2009. Wanderlust: Extracting semantic relations from natural language text using dependency grammar patterns. In WWW Workshop. S¨oren Auer and Jens Lehmann. 2007. What have innsbruck and leipzig in common? extracting semantics from wiki content. In ESWC. M. Banko, M. Cafarella, S. Soderland, M. Broadhead, and O. Etzioni. 2007. Open information extraction from the Web. In Procs. of IJCAI. Razvan C. Bunescu and Raymond J. Mooney. 2005. Subsequence kernels for relation extraction. In NIPS. R. Bunescu and R.Mooney. 2005. A shortest path dependency kernel for relation extraction. In HLT/EMNLP. Eugene Charniak and Mark Johnson. 2005. Coarseto-fine n-best parsing and maxent discriminative reranking. In ACL. M. Craven, D. DiPasquo, D. Freitag, A. McCallum, T. Mitchell, K. Nigam, and S. Slattery. 1998. Learning to extract symbolic knowledge from the world wide web. In AAAI. Dmitry Davidov and Ari Rappoport. 2008. Unsupervised discovery of generic relationships using pattern clusters and its evaluation by automatically generated sat analogy questions. In ACL. Dmitry Davidov, Ari Rappoport, and Moshe Koppel. 2007. Fully unsupervised discovery of conceptspecific relationships by web mining. In ACL. Marie-Catherine de Marneffe and Christopher D. Manning. 2008. Stanford typed dependencies manual. http://nlp.stanford.edu/downloads/lex-parser.shtml. Benjamin Van Durme and Lenhart K. Schubert. 2008. Open knowledge extraction using compositional language processing. In STEP. R. Hoffmann, C. Zhang, and D. Weld. 2010. Learning 5000 relational extractors. In ACL. Jing Jiang and ChengXiang Zhai. 2007. A systematic exploration of the feature space for relation extraction. In HLT/NAACL. A. Gangemi M. Ciaramita. 2005. Unsupervised learning of semantic relations between concepts of a molecular biology ontology. In IJCAI. Andrew Kachites McCallum. 2002. Mallet: A machine learning for language toolkit. In http://mallet.cs.umass.edu. Mike Mintz, Steven Bills, Rion Snow, and Dan Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In ACL-IJCNLP. T. H. Kotaro Nakayama and S. Nishio. 2008. Wikipedia link structure and text mining for semantic relation extraction. In CEUR Workshop. Dat P.T Nguyen, Yutaka Matsuo, and Mitsuru Ishizuka. 2007. Exploiting syntactic and semantic information for relation extraction from wikipedia. In IJCAI07-TextLinkWS. Marius Pasca. 2008. Turning web text and search queries into factual knowledge: Hierarchical class attribute extraction. In AAAI. Fuchun Peng and Andrew McCallum. 2004. Accurate Information Extraction from Research Papers using Conditional Random Fields. In HLT-NAACL. Hoifung Poon and Pedro Domingos. 2008. Joint Inference in Information Extraction. In AAAI. Y. Shinyama and S. Sekine. 2006. Preemptive information extraction using unristricted relation discovery. In HLT-NAACL. Rion Snow, Daniel Jurafsky, and Andrew Y. Ng. 2005. Learning syntactic patterns for automatic hypernym discovery. In NIPS. Fabian M. Suchanek, Gjergji Kasneci, and Gerhard Weikum. 2007. Yago: A core of semantic knowledge - unifying WordNet and Wikipedia. In WWW. Mengqiu Wang. 2008. A re-examination of dependency path kernels for relation extraction. In IJCNLP. Fei Wu and Daniel Weld. 2007. Autonomouslly Semantifying Wikipedia. In CIKM. Fei Wu, Raphael Hoffmann, and Danel S. Weld. 2008. Information extraction from Wikipedia: Moving down the long tail. In KDD. Min Zhang, Jie Zhang, Jian Su, and Guodong Zhou. 2006. A composite kernel to extract relations between entities with both flat and structured features. In ACL. Shubin Zhao and Ralph Grishman. 2005. Extracting relations with integrated information using kernel methods. In ACL. Jun Zhu, Zaiqing Nie, Xiaojiang Liu, Bo Zhang, and Ji-Rong Wen. 2009. Statsnowball: a statistical approach to extracting entity relationships. In WWW. 127
|
2010
|
13
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1278–1287, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Profiting from Mark-Up: Hyper-Text Annotations for Guided Parsing Valentin I. Spitkovsky Computer Science Department Stanford University and Google Inc. [email protected] Daniel Jurafsky Departments of Linguistics and Computer Science, Stanford University [email protected] Hiyan Alshawi Google Inc. [email protected] Abstract We show how web mark-up can be used to improve unsupervised dependency parsing. Starting from raw bracketings of four common HTML tags (anchors, bold, italics and underlines), we refine approximate partial phrase boundaries to yield accurate parsing constraints. Conversion procedures fall out of our linguistic analysis of a newly available million-word hyper-text corpus. We demonstrate that derived constraints aid grammar induction by training Klein and Manning’s Dependency Model with Valence (DMV) on this data set: parsing accuracy on Section 23 (all sentences) of the Wall Street Journal corpus jumps to 50.4%, beating previous state-of-theart by more than 5%. Web-scale experiments show that the DMV, perhaps because it is unlexicalized, does not benefit from orders of magnitude more annotated but noisier data. Our model, trained on a single blog, generalizes to 53.3% accuracy out-of-domain, against the Brown corpus — nearly 10% higher than the previous published best. The fact that web mark-up strongly correlates with syntactic structure may have broad applicability in NLP. 1 Introduction Unsupervised learning of hierarchical syntactic structure from free-form natural language text is a hard problem whose eventual solution promises to benefit applications ranging from question answering to speech recognition and machine translation. A restricted version of this problem that targets dependencies and assumes partial annotation — sentence boundaries and part-of-speech (POS) tagging — has received much attention. Klein and Manning (2004) were the first to beat a simple parsing heuristic, the right-branching baseline; today’s state-of-the-art systems (Headden et al., 2009; Cohen and Smith, 2009; Spitkovsky et al., 2010a) are rooted in their Dependency Model with Valence (DMV), still trained using variants of EM. Pereira and Schabes (1992) outlined three major problems with classic EM, applied to a related problem, constituent parsing. They extended classic inside-outside re-estimation (Baker, 1979) to respect any bracketing constraints included with a training corpus. This conditioning on partial parses addressed all three problems, leading to: (i) linguistically reasonable constituent boundaries and induced grammars more likely to agree with qualitative judgments of sentence structure, which is underdetermined by unannotated text; (ii) fewer iterations needed to reach a good grammar, countering convergence properties that sharply deteriorate with the number of non-terminal symbols, due to a proliferation of local maxima; and (iii) better (in the best case, linear) time complexity per iteration, versus running time that is ordinarily cubic in both sentence length and the total number of non-terminals, rendering sufficiently large grammars computationally impractical. Their algorithm sometimes found good solutions from bracketed corpora but not from raw text, supporting the view that purely unsupervised, selforganizing inference methods can miss the trees for the forest of distributional regularities. This was a promising break-through, but the problem of whence to get partial bracketings was left open. We suggest mining partial bracketings from a cheap and abundant natural language resource: the hyper-text mark-up that annotates web-pages. For example, consider that anchor text can match linguistic constituents, such as verb phrases, exactly: ..., whereas McCain is secure on the topic, Obama <a>[VP worries about winning the pro-Israel vote]</a>. To validate this idea, we created a new data set, novel in combining a real blog’s raw HTML with tree-bank-like constituent structure parses, gener1278 ated automatically. Our linguistic analysis of the most prevalent tags (anchors, bold, italics and underlines) over its 1M+ words reveals a strong connection between syntax and mark-up (all of our examples draw from this corpus), inspiring several simple techniques for automatically deriving parsing constraints. Experiments with both hard and more flexible constraints, as well as with different styles and quantities of annotated training data — the blog, web news and the web itself, confirm that mark-up-induced constraints consistently improve (otherwise unsupervised) dependency parsing. 2 Intuition and Motivating Examples It is natural to expect hidden structure to seep through when a person annotates a sentence. As it happens, a non-trivial fraction of the world’s population routinely annotates text diligently, if only partially and informally.1 They inject hyper-links, vary font sizes, and toggle colors and styles, using mark-up technologies such as HTML and XML. As noted, web annotations can be indicative of phrase boundaries, e.g., in a complicated sentence: In 1998, however, as I <a>[VP established in <i>[NP The New Republic]</i>]</a> and Bill Clinton just <a>[VP confirmed in his memoirs]</a>, Netanyahu changed his mind and ... In doing so, mark-up sometimes offers useful cues even for low-level tokenization decisions: [NP [NP Libyan ruler] <a>[NP Mu‘ammar al-Qaddafi]</a>] referred to ... (NP (ADJP (NP (JJ Libyan) (NN ruler)) (JJ Mu)) (‘‘ ‘) (NN ammar) (NNS al-Qaddafi)) Above, a backward quote in an Arabic name confuses the Stanford parser.2 Yet mark-up lines up with the broken noun phrase, signals cohesion, and moreover sheds light on the internal structure of a compound. As Vadas and Curran (2007) point out, such details are frequently omitted even from manually compiled tree-banks that err on the side of flat annotations of base-NPs. Admittedly, not all boundaries between HTML tags and syntactic constituents match up nicely: ..., but [S [NP the <a><i>Toronto Star</i>][VP reports [NP this][PP in the softest possible way]</a>,[S stating only that ...]]] Combining parsing with mark-up may not be straight-forward, but there is hope: even above, 1Even when (American) grammar schools lived up to their name, they only taught dependencies. This was back in the days before constituent grammars were invented. 2http://nlp.stanford.edu:8080/parser/ one of each nested tag’s boundaries aligns; and Toronto Star’s neglected determiner could be forgiven, certainly within a dependency formulation. 3 A High-Level Outline of Our Approach Our idea is to implement the DMV (Klein and Manning, 2004) — a standard unsupervised grammar inducer. But instead of learning the unannotated test set, we train with text that contains web mark-up, using various ways of converting HTML into parsing constraints. We still test on WSJ (Marcus et al., 1993), in the standard way, and also check generalization against a hidden data set — the Brown corpus (Francis and Kucera, 1979). Our parsing constraints come from a blog — a new corpus we created, the web and news (see Table 1 for corpora’s sentence and token counts). To facilitate future work, we make the final models and our manually-constructed blog data publicly available.3 Although we are unable to share larger-scale resources, our main results should be reproducible, as both linguistic analysis and our best model rely exclusively on the blog. Corpus Sentences POS Tokens WSJ∞ 49,208 1,028,347 Section 23 2,353 48,201 WSJ45 48,418 986,830 WSJ15 15,922 163,715 Brown100 24,208 391,796 BLOGp 57,809 1,136,659 BLOGt45 56,191 1,048,404 BLOGt15 23,214 212,872 NEWS45 2,263,563,078 32,119,123,561 NEWS15 1,433,779,438 11,786,164,503 WEB45 8,903,458,234 87,269,385,640 WEB15 7,488,669,239 55,014,582,024 Table 1: Sizes of corpora derived from WSJ and Brown, as well as those we collected from the web. 4 Data Sets for Evaluation and Training The appeal of unsupervised parsing lies in its ability to learn from surface text alone; but (intrinsic) evaluation still requires parsed sentences. Following Klein and Manning (2004), we begin with reference constituent parses and compare against deterministically derived dependencies: after pruning out all empty subtrees, punctuation and terminals (tagged # and $) not pronounced where they appear, we drop all sentences with more than a prescribed number of tokens remaining and use automatic “head-percolation” rules (Collins, 1999) to convert the rest, as is standard practice. 3http://cs.stanford.edu/∼valentin/ 1279 Length Marked POS Bracketings Length Marked POS Bracketings Cutoff Sentences Tokens All Multi-Token Cutoff Sentences Tokens All Multi-Token 0 6,047 1,136,659 7,731 6,015 8 485 14,528 710 684 1 of 57,809 149,483 7,731 6,015 9 333 10,484 499 479 2 4,934 124,527 6,482 6,015 10 245 7,887 365 352 3 3,295 85,423 4,476 4,212 15 42 1,519 65 63 4 2,103 56,390 2,952 2,789 20 13 466 20 20 5 1,402 38,265 1,988 1,874 25 6 235 10 10 6 960 27,285 1,365 1,302 30 3 136 6 6 7 692 19,894 992 952 40 0 0 0 0 Table 2: Counts of sentences, tokens and (unique) bracketings for BLOGp, restricted to only those sentences having at least one bracketing no shorter than the length cutoff (but shorter than the sentence). Our primary reference sets are derived from the Penn English Treebank’s Wall Street Journal portion (Marcus et al., 1993): WSJ45 (sentences with fewer than 46 tokens) and Section 23 of WSJ∞(all sentence lengths). We also evaluate on Brown100, similarly derived from the parsed portion of the Brown corpus (Francis and Kucera, 1979). While we use WSJ45 and WSJ15 to train baseline models, the bulk of our experiments is with web data. 4.1 A News-Style Blog: Daniel Pipes Since there was no corpus overlaying syntactic structure with mark-up, we began constructing a new one by downloading articles4 from a newsstyle blog. Although limited to a single genre — political opinion, danielpipes.org is clean, consistently formatted, carefully edited and larger than WSJ (see Table 1). Spanning decades, Pipes’ editorials are mostly in-domain for POS taggers and tree-bank-trained parsers; his recent (internetera) entries are thoroughly cross-referenced, conveniently providing just the mark-up we hoped to study via uncluttered (printer-friendly) HTML.5 After extracting moderately clean text and mark-up locations, we used MxTerminator (Reynar and Ratnaparkhi, 1997) to detect sentence boundaries. This initial automated pass begot multiple rounds of various semi-automated clean-ups that involved fixing sentence breaking, modifying parser-unfriendly tokens, converting HTML entities and non-ASCII text, correcting typos, and so on. After throwing away annotations of fractional words (e.g., <i>basmachi</i>s) and tokens (e.g., <i>Sesame Street</i>-like), we broke up all markup that crossed sentence boundaries (i.e., loosely speaking, replaced constructs like <u>...][S...</u> with <u>...</u> ][S <u>...</u>) and discarded any 4http://danielpipes.org/art/year/all 5http://danielpipes.org/article print.php? id=. . . tags left covering entire sentences. We finalized two versions of the data: BLOGt, tagged with the Stanford tagger (Toutanova and Manning, 2000; Toutanova et al., 2003),6 and BLOGp, parsed with Charniak’s parser (Charniak, 2001; Charniak and Johnson, 2005).7 The reason for this dichotomy was to use state-of-the-art parses to analyze the relationship between syntax and mark-up, yet to prevent jointly tagged (and non-standard AUX[G]) POS sequences from interfering with our (otherwise unsupervised) training.8 4.2 Scaled up Quantity: The (English) Web We built a large (see Table 1) but messy data set, WEB — English-looking web-pages, pre-crawled by a search engine. To avoid machine-generated spam, we excluded low quality sites flagged by the indexing system. We kept only sentence-like runs of words (satisfying punctuation and capitalization constraints), POS-tagged with TnT (Brants, 2000). 4.3 Scaled up Quality: (English) Web News In an effort to trade quantity for quality, we constructed a smaller, potentially cleaner data set, NEWS. We reckoned editorialized content would lead to fewer extracted non-sentences. Perhaps surprisingly, NEWS is less than an order of magnitude smaller than WEB (see Table 1); in part, this is due to less aggressive filtering — we trust sites approved by the human editors at Google News.9 In all other respects, our pre-processing of NEWS pages was identical to our handling of WEB data. 6http://nlp.stanford.edu/software/ stanford-postagger-2008-09-28.tar.gz 7ftp://ftp.cs.brown.edu/pub/nlparser/ parser05Aug16.tar.gz 8However, since many taggers are themselves trained on manually parsed corpora, such as WSJ, no parser that relies on external POS tags could be considered truly unsupervised; for a fully unsupervised example, see Seginer’s (2007) CCL parser, available at http://www.seggu.net/ccl/ 9http://news.google.com/ 1280 5 Linguistic Analysis of Mark-Up Is there a connection between mark-up and syntactic structure? Previous work (Barr et al., 2008) has only examined search engine queries, showing that they consist predominantly of short noun phrases. If web mark-up shared a similar characteristic, it might not provide sufficiently disambiguating cues to syntactic structure: HTML tags could be too short (e.g., singletons like “click <a>here</a>”) or otherwise unhelpful in resolving truly difficult ambiguities (such as PPattachment). We began simply by counting various basic events in BLOGp. Count POS Sequence Frac Sum 1 1,242 NNP NNP 16.1% 2 643 NNP 8.3 24.4 3 419 NNP NNP NNP 5.4 29.8 4 414 NN 5.4 35.2 5 201 JJ NN 2.6 37.8 6 138 DT NNP NNP 1.8 39.5 7 138 NNS 1.8 41.3 8 112 JJ 1.5 42.8 9 102 VBD 1.3 44.1 10 92 DT NNP NNP NNP 1.2 45.3 11 85 JJ NNS 1.1 46.4 12 79 NNP NN 1.0 47.4 13 76 NN NN 1.0 48.4 14 61 VBN 0.8 49.2 15 60 NNP NNP NNP NNP 0.8 50.0 BLOGp +3,869 more with Count ≤49 50.0% Table 3: Top 50% of marked POS tag sequences. Count Non-Terminal Frac Sum 1 5,759 NP 74.5% 2 997 VP 12.9 87.4 3 524 S 6.8 94.2 4 120 PP 1.6 95.7 5 72 ADJP 0.9 96.7 6 61 FRAG 0.8 97.4 7 41 ADVP 0.5 98.0 8 39 SBAR 0.5 98.5 9 19 PRN 0.2 98.7 10 18 NX 0.2 99.0 BLOGp +81 more with Count ≤16 1.0% Table 4: Top 99% of dominating non-terminals. 5.1 Surface Text Statistics Out of 57,809 sentences, 6,047 (10.5%) are annotated (see Table 2); and 4,934 (8.5%) have multitoken bracketings. We do not distinguish HTML tags and track only unique bracketing end-points within a sentence. Of these, 6,015 are multi-token — an average per-sentence yield of 10.4%.10 10A non-trivial fraction of our corpus is older (pre-internet) unannotated articles, so this estimate may be conservative. As expected, many of the annotated words are nouns, but there are adjectives, verbs and other parts of speech too (see Table 3). Mark-up is short, typically under five words, yet (by far) the most frequently marked sequence of POS tags is a pair. 5.2 Common Syntactic Subtrees For three-quarters of all mark-up, the lowest dominating non-terminal is a noun phrase (see Table 4); there are also non-trace quantities of verb phrases (12.9%) and other phrases, clauses and fragments. Of the top fifteen — 35.2% of all — annotated productions, only one is not a noun phrase (see Table 5, left). Four of the fifteen lowest dominating non-terminals do not match the entire bracketing — all four miss the leading determiner, as we saw earlier. In such cases, we recursively split internal nodes until the bracketing aligned, as follows: [S [NP the <a>Toronto Star][VP reports [NP this] [PP in the softest possible way]</a>,[S stating ...]]] S →NP VP →DT NNP NNP VBZ NP PP S We can summarize productions more compactly by using a dependency framework and clipping off any dependents whose subtrees do not cross a bracketing boundary, relative to the parent. Thus, DT NNP NNP VBZ DT IN DT JJS JJ NN becomes DT NNP VBZ, “the <a>Star reports</a>.” Viewed this way, the top fifteen (now collapsed) productions cover 59.4% of all cases and include four verb heads, in addition to a preposition and an adjective (see Table 5, right). This exposes five cases of inexact matches, three of which involve neglected determiners or adjectives to the left of the head. In fact, the only case that cannot be explained by dropped dependents is #8, where the daughters are marked but the parent is left out. Most instances contributing to this pattern are flat NPs that end with a noun, incorrectly assumed to be the head of all other words in the phrase, e.g., ... [NP a 1994 <i>New Yorker</i> article] ... As this example shows, disagreements (as well as agreements) between mark-up and machinegenerated parse trees with automatically percolated heads should be taken with a grain of salt.11 11In a relatively recent study, Ravi et al. (2008) report that Charniak’s re-ranking parser (Charniak and Johnson, 2005) — reranking-parserAug06.tar.gz, also available from ftp://ftp.cs.brown.edu/pub/nlparser/ — attains 86.3% accuracy when trained on WSJ and tested against Brown; its nearly 5% performance loss out-of-domain is consistent with the numbers originally reported by Gildea (2001). 1281 Count Constituent Production Frac Sum 1 746 NP →NNP NNP 9.6% 2 357 NP →NNP 4.6 14.3 3 266 NP →NP PP 3.4 17.7 4 183 NP →NNP NNP NNP 2.4 20.1 5 165 NP →DT NNP NNP 2.1 22.2 6 140 NP →NN 1.8 24.0 7 131 NP →DT NNP NNP NNP 1.7 25.7 8 130 NP →DT NN 1.7 27.4 9 127 NP →DT NNP NNP 1.6 29.0 10 109 S →NP VP 1.4 30.4 11 91 NP →DT NNP NNP NNP 1.2 31.6 12 82 NP →DT JJ NN 1.1 32.7 13 79 NP →NNS 1.0 33.7 14 65 NP →JJ NN 0.8 34.5 15 60 NP →NP NP 0.8 35.3 BLOGp +5,000 more with Count ≤60 64.7% Count Head-Outward Spawn Frac Sum 1 1,889 NNP 24.4% 2 623 NN 8.1 32.5 3 470 DT NNP 6.1 38.6 4 458 DT NN 5.9 44.5 5 345 NNS 4.5 49.0 6 109 NNPS 1.4 50.4 7 98 VBG 1.3 51.6 8 96 NNP NNP NN 1.2 52.9 9 80 VBD 1.0 53.9 10 77 IN 1.0 54.9 11 74 VBN 1.0 55.9 12 73 DT JJ NN 0.9 56.8 13 71 VBZ 0.9 57.7 14 69 POS NNP 0.9 58.6 15 63 JJ 0.8 59.4 BLOGp +3,136 more with Count ≤62 40.6% Table 5: Top 15 marked productions, viewed as constituents (left) and as dependencies (right), after recursively expanding any internal nodes that did not align with the bracketing (underlined). Tabulated dependencies were collapsed, dropping any dependents that fell entirely in the same region as their parent (i.e., both inside the bracketing, both to its left or both to its right), keeping only crossing attachments. 5.3 Proposed Parsing Constraints The straight-forward approach — forcing mark-up to correspond to constituents — agrees with Charniak’s parse trees only 48.0% of the time, e.g., ... in [NP<a>[NP an analysis]</a>[PP of perhaps the most astonishing PC item I have yet stumbled upon]]. This number should be higher, as the vast majority of disagreements are due to tree-bank idiosyncrasies (e.g., bare NPs). Earlier examples of incomplete constituents (e.g., legitimately missing determiners) would also be fine in many linguistic theories (e.g., as N-bars). A dependency formulation is less sensitive to such stylistic differences. We begin with the hardest possible constraint on dependencies, then slowly relax it. Every example used to demonstrate a softer constraint doubles as a counter-example against all previous versions. • strict — seals mark-up into attachments, i.e., inside a bracketing, enforces exactly one external arc — into the overall head. This agrees with head-percolated trees just 35.6% of the time, e.g., As author of <i>The Satanic Verses</i>, I ... • loose — same as strict, but allows the bracketing’s head word to have external dependents. This relaxation already agrees with head-percolated dependencies 87.5% of the time, catching many (though far from all) dropped dependents, e.g., ... the <i>Toronto Star</i> reports ... • sprawl — same as loose, but now allows all words inside a bracketing to attach external dependents.12 This boosts agreement with headpercolated trees to 95.1%, handling new cases, e.g., where “Toronto Star” is embedded in longer mark-up that includes its own parent — a verb: ... the <a>Toronto Star reports ... </a> ... • tear — allows mark-up to fracture after all, requiring only that the external heads attaching the pieces lie to the same side of the bracketing. This propels agreement with percolated dependencies to 98.9%, fixing previously broken PP-attachment ambiguities, e.g., a fused phrase like “Fox News in Canada” that detached a preposition from its verb: ... concession ... has raised eyebrows among those waiting [PP for <a>Fox News][PP in Canada]</a>. Most of the remaining 1.1% of disagreements are due to parser errors. Nevertheless, it is possible for mark-up to be torn apart by external heads from both sides. We leave this section with a (very rare) true negative example. Below, “CSA” modifies “authority” (to its left), appositively, while “AlManar” modifies “television” (to its right):13 The French broadcasting authority, <a>CSA, banned ... Al-Manar</a> satellite television from ... 12This view evokes the trapezoids of the O(n3) recognizer for split head automaton grammars (Eisner and Satta, 1999). 13But this is a stretch, since the comma after “CSA” renders the marked phrase ungrammatical even out of context. 1282 6 Experimental Methods and Metrics We implemented the DMV (Klein and Manning, 2004), consulting the details of (Spitkovsky et al., 2010a). Crucially, we swapped out inside-outside re-estimation in favor of Viterbi training. Not only is it better-suited to the general problem (see §7.1), but it also admits a trivial implementation of (most of) the dependency constraints we proposed.14 5 10 15 20 25 30 35 40 45 4.5 5.0 5.5 WSJk bpt lowest cross-entropy (4.32bpt) attained at WSJ8 x-Entropy h (in bits per token) on WSJ15 Figure 1: Sentence-level cross-entropy on WSJ15 for Ad-Hoc∗initializers of WSJ{1, . . . , 45}. Six settings parameterized each run: • INIT: 0 — default, uniform initialization; or 1 — a high quality initializer, pre-trained using Ad-Hoc∗(Spitkovsky et al., 2010a): we chose the Laplace-smoothed model trained at WSJ15 (the “sweet spot” data gradation) but initialized off WSJ8, since that ad-hoc harmonic initializer has the best cross-entropy on WSJ15 (see Figure 1). • GENRE: 0 — default, baseline training on WSJ; else, uses 1 — BLOGt; 2 — NEWS; or 3 — WEB. • SCOPE: 0 — default, uses all sentences up to length 45; if 1, trains using sentences up to length 15; if 2, re-trains on sentences up to length 45, starting from the solution to sentences up to length 15, as recommended by Spitkovsky et al. (2010a). • CONSTR: if 4, strict; if 3, loose; and if 2, sprawl. We did not implement level 1, tear. Overconstrained sentences are re-attempted at successively lower levels until they become possible to parse, if necessary at the lowest (default) level 0.15 • TRIM: if 1, discards any sentence without a single multi-token mark-up (shorter than its length). • ADAPT: if 1, upon convergence, initializes retraining on WSJ45 using the solution to <GENRE>, attempting domain adaptation (Lee et al., 1991). These make for 294 meaningful combinations. We judged each one by its accuracy on WSJ45, using standard directed scoring — the fraction of correct dependencies over randomized “best” parse trees. 14We analyze the benefits of Viterbi training in a companion paper (Spitkovsky et al., 2010b), which dedicates more space to implementation and to the WSJ baselines used here. 15At level 4, <b> X<u> Y</b> Z</u> is over-constrained. 7 Discussion of Experimental Results Evaluation on Section 23 of WSJ and Brown reveals that blog-training beats all published stateof-the-art numbers in every traditionally-reported length cutoff category, with news-training not far behind. Here is a mini-preview of these results, for Section 23 of WSJ10 and WSJ∞(from Table 8): WSJ10 WSJ∞ (Cohen and Smith, 2009) 62.0 42.2 (Spitkovsky et al., 2010a) 57.1 45.0 NEWS-best 67.3 50.1 BLOGt-best 69.3 50.4 (Headden et al., 2009) 68.8 Table 6: Directed accuracies on Section 23 of WSJ{10,∞} for three recent state-of-the-art systems and our best runs (as judged against WSJ45) for NEWS and BLOGt (more details in Table 8). Since our experimental setup involved testing nearly three hundred models simultaneously, we must take extreme care in analyzing and interpreting these results, to avoid falling prey to any looming “data-snooping” biases.16 In a sufficiently large pool of models, where each is trained using a randomized and/or chaotic procedure (such as ours), the best may look good due to pure chance. We appealed to three separate diagnostics to convince ourselves that our best results are not noise. The most radical approach would be to write off WSJ as a development set and to focus only on the results from the held-out Brown corpus. It was initially intended as a test of out-of-domain generalization, but since Brown was in no way involved in selecting the best models, it also qualifies as a blind evaluation set. We observe that our best models perform even better (and gain more — see Table 8) on Brown than on WSJ — a strong indication that our selection process has not overfitted. Our second diagnostic is a closer look at WSJ. Since we cannot graph the full (six-dimensional) set of results, we begin with a simple linear regression, using accuracy on WSJ45 as the dependent variable. We prefer this full factorial design to the more traditional ablation studies because it allows us to account for and to incorporate every single experimental data point incurred along the 16In the standard statistical hypothesis testing setting, it is reasonable to expect that p% of randomly chosen hypotheses will appear significant at the p% level simply by chance. Consequently, multiple hypothesis testing requires re-evaluating significance levels — adjusting raw p-values, e.g., using the Holm-Bonferroni method (Holm, 1979). 1283 Corpus Marked Sentences All Sentences POS Tokens All Bracketings Multi-Token Bracketings BLOGt45 5,641 56,191 1,048,404 7,021 5,346 BLOG′ t45 4,516 4,516 104,267 5,771 5,346 BLOGt15 1,562 23,214 212,872 1,714 1,240 BLOG′ t15 1,171 1,171 11,954 1,288 1,240 NEWS45 304,129,910 2,263,563,078 32,119,123,561 611,644,606 477,362,150 NEWS′45 205,671,761 205,671,761 2,740,258,972 453,781,081 392,600,070 NEWS15 211,659,549 1,433,779,438 11,786,164,503 365,145,549 274,791,675 NEWS′15 147,848,358 147,848,358 1,397,562,474 272,223,918 231,029,921 WEB45 1,577,208,680 8,903,458,234 87,269,385,640 3,309,897,461 2,459,337,571 WEB′45 933,115,032 933,115,032 11,552,983,379 2,084,359,555 1,793,238,913 WEB15 1,181,696,194 7,488,669,239 55,014,582,024 2,071,743,595 1,494,675,520 WEB′15 681,087,020 681,087,020 5,813,555,341 1,200,980,738 1,072,910,682 Table 7: Counts of sentences, tokens and (unique) bracketings for web-based data sets; trimmed versions, restricted to only those sentences having at least one multi-token bracketing, are indicated by a prime (′). way. Its output is a coarse, high-level summary of our runs, showing which factors significantly contribute to changes in error rate on WSJ45: Parameter (Indicator) Setting ˆβ p-value INIT 1 ad-hoc @WSJ8,15 11.8 *** GENRE 1 BLOGt -3.7 0.06 2 NEWS -5.3 ** 3 WEB -7.7 *** SCOPE 1 @15 -0.5 0.40 2 @15→45 -0.4 0.53 CONSTR 2 sprawl 0.9 0.23 3 loose 1.0 0.15 4 strict 1.8 * TRIM 1 drop unmarked -7.4 *** ADAPT 1 WSJ re-training 1.5 ** Intercept (R2 Adjusted = 73.6%) 39.9 *** We use a standard convention: *** for p < 0.001; ** for p < 0.01 (very signif.); and * for p < 0.05 (signif.). The default training mode (all parameters zero) is estimated to score 39.9%. A good initializer gives the biggest (double-digit) gain; both domain adaptation and constraints also make a positive impact. Throwing away unannotated data hurts, as does training out-of-domain (the blog is least bad; the web is worst). Of course, this overview should not be taken too seriously. Overly simplistic, a first order model ignores interactions between parameters. Furthermore, a least squares fit aims to capture central tendencies, whereas we are more interested in outliers — the best-performing runs. A major imperfection of the simple regression model is that helpful factors that require an interaction to “kick in” may not, on their own, appear statistically significant. Our third diagnostic is to examine parameter settings that give rise to the best-performing models, looking out for combinations that consistently deliver superior results. 7.1 WSJ Baselines Just two parameters apply to learning from WSJ. Five of their six combinations are state-of-the-art, demonstrating the power of Viterbi training; only the default run scores worse than 45.0%, attained by Leapfrog (Spitkovsky et al., 2010a), on WSJ45: Settings SCOPE=0 SCOPE=1 SCOPE=2 INIT=0 41.3 45.0 45.2 1 46.6 47.5 47.6 @45 @15 @15→45 7.2 Blog Simply training on BLOGt instead of WSJ hurts: GENRE=1 SCOPE=0 SCOPE=1 SCOPE=2 INIT=0 39.6 36.9 36.9 1 46.5 46.3 46.4 @45 @15 @15→45 The best runs use a good initializer, discard unannotated sentences, enforce the loose constraint on the rest, follow up with domain adaptation and benefit from re-training — GENRE=TRIM=ADAPT=1: INIT=1 SCOPE=0 SCOPE=1 SCOPE=2 CONSTR=0 45.8 48.3 49.6 (sprawl) 2 46.3 49.2 49.2 (loose) 3 41.3 50.2 50.4 (strict) 4 40.7 49.9 48.7 @45 @15 @15→45 The contrast between unconstrained learning and annotation-guided parsing is higher for the default initializer, still using trimmed data sets (just over a thousand sentences for BLOG′ t15 — see Table 7): INIT=0 SCOPE=0 SCOPE=1 SCOPE=2 CONSTR=0 25.6 19.4 19.3 (sprawl) 2 25.2 22.7 22.5 (loose) 3 32.4 26.3 27.3 (strict) 4 36.2 38.7 40.1 @45 @15 @15→45 Above, we see a clearer benefit to our constraints. 1284 7.3 News Training on WSJ is also better than using NEWS: GENRE=2 SCOPE=0 SCOPE=1 SCOPE=2 INIT=0 40.2 38.8 38.7 1 43.4 44.0 43.8 @45 @15 @15→45 As with the blog, the best runs use the good initializer, discard unannotated sentences, enforce the loose constraint and follow up with domain adaptation — GENRE=2; INIT=TRIM=ADAPT=1: Settings SCOPE=0 SCOPE=1 SCOPE=2 CONSTR=0 46.6 45.4 45.2 (sprawl) 2 46.1 44.9 44.9 (loose) 3 49.5 48.1 48.3 (strict) 4 37.7 36.8 37.6 @45 @15 @15→45 With all the extra training data, the best new score is just 49.5%. On the one hand, we are disappointed by the lack of dividends to orders of magnitude more data. On the other, we are comforted that the system arrives within 1% of its best result — 50.4%, obtained with a manually cleaned up corpus — now using an auto-generated data set. 7.4 Web The WEB-side story is more discouraging: GENRE=3 SCOPE=0 SCOPE=1 SCOPE=2 INIT=0 38.3 35.1 35.2 1 42.8 43.6 43.4 @45 @15 @15→45 Our best run again uses a good initializer, keeps all sentences, still enforces the loose constraint and follows up with domain adaptation, but performs worse than all well-initialized WSJ baselines, scoring only 45.9% (trained at WEB15). We suspect that the web is just too messy for us. On top of the challenges of language identification and sentence-breaking, there is a lot of boiler-plate; furthermore, web text can be difficult for news-trained POS taggers. For example, note that the verb “sign” is twice mistagged as a noun and that “YouTube” is classified as a verb, in the top four POS sequences of web sentences:17 POS Sequence WEB Count Sample web sentence, chosen uniformly at random. 1 DT NNS VBN 82,858,487 All rights reserved. 2 NNP NNP NNP 65,889,181 Yuasa et al. 3 NN IN TO VB RB 31,007,783 Sign in to YouTube now! 4 NN IN IN PRP$ JJ NN 31,007,471 Sign in with your Google Account! 17Further evidence: TnT tags the ubiquitous but ambiguous fragments “click here” and “print post” as noun phrases. 7.5 The State of the Art Our best model gains more than 5% over previous state-of-the-art accuracy across all sentences of WSJ’s Section 23, more than 8% on WSJ20 and rivals the oracle skyline (Spitkovsky et al., 2010a) on WSJ10; these gains generalize to Brown100, where it improves by nearly 10% (see Table 8). We take solace in the fact that our best models agree in using loose constraints. Of these, the models trained with less data perform better, with the best two using trimmed data sets, echoing that “less is more” (Spitkovsky et al., 2010a), pace Halevy et al. (2009). We note that orders of magnitude more data did not improve parsing performance further and suspect a different outcome from lexicalized models: The primary benefit of additional lower-quality data is in improved coverage. But with only 35 unique POS tags, data sparsity is hardly an issue. Extra examples of lexical items help little and hurt when they are mistagged. 8 Related Work The wealth of new annotations produced in many languages every day already fuels a number of NLP applications. Following their early and wide-spread use by search engines, in service of spam-fighting and retrieval, anchor text and link data enhanced a variety of traditional NLP techniques: cross-lingual information retrieval (Nie and Chen, 2002), translation (Lu et al., 2004), both named-entity recognition (Mihalcea and Csomai, 2007) and categorization (Watanabe et al., 2007), query segmentation (Tan and Peng, 2008), plus semantic relatedness and word-sense disambiguation (Gabrilovich and Markovitch, 2007; Yeh et al., 2009). Yet several, seemingly natural, candidate core NLP tasks — tokenization, CJK segmentation, noun-phrase chunking, and (until now) parsing — remained conspicuously uninvolved. Approaches related to ours arise in applications that combine parsing with named-entity recognition (NER). For example, constraining a parser to respect the boundaries of known entities is standard practice not only in joint modeling of (constituent) parsing and NER (Finkel and Manning, 2009), but also in higher-level NLP tasks, such as relation extraction (Mintz et al., 2009), that couple chunking with (dependency) parsing. Although restricted to proper noun phrases, dates, times and quantities, we suspect that constituents identified by trained (supervised) NER systems would also 1285 Model Incarnation WSJ10 WSJ20 WSJ∞ DMV Bilingual Log-Normals (tie-verb-noun) (Cohen and Smith, 2009) 62.0 48.0 42.2 Brown100 Leapfrog (Spitkovsky et al., 2010a) 57.1 48.7 45.0 43.6 default INIT=0,GENRE=0,SCOPE=0,CONSTR=0,TRIM=0,ADAPT=0 55.9 45.8 41.6 40.5 WSJ-best INIT=1,GENRE=0,SCOPE=2,CONSTR=0,TRIM=0,ADAPT=0 65.3 53.8 47.9 50.8 BLOGt-best INIT=1,GENRE=1,SCOPE=2,CONSTR=3,TRIM=1,ADAPT=1 69.3 56.8 50.4 53.3 NEWS-best INIT=1,GENRE=2,SCOPE=0,CONSTR=3,TRIM=1,ADAPT=1 67.3 56.2 50.1 51.6 WEB-best INIT=1,GENRE=3,SCOPE=1,CONSTR=3,TRIM=0,ADAPT=1 64.1 52.7 46.3 46.9 EVG Smoothed (skip-head), Lexicalized (Headden et al., 2009) 68.8 Table 8: Accuracies on Section 23 of WSJ{10, 20,∞} and Brown100 for three recent state-of-the-art systems, our default run, and our best runs (judged by accuracy on WSJ45) for each of four training sets. be helpful in constraining grammar induction. Following Pereira and Schabes’ (1992) success with partial annotations in training a model of (English) constituents generatively, their idea has been extended to discriminative estimation (Riezler et al., 2002) and also proved useful in modeling (Japanese) dependencies (Sassano, 2005). There was demand for partially bracketed corpora. Chen and Lee (1995) constructed one such corpus by learning to partition (English) POS sequences into chunks (Abney, 1991); Inui and Kotani (2001) used n-gram statistics to split (Japanese) clauses. We combine the two intuitions, using the web to build a partially parsed corpus. Our approach could be called lightly-supervised, since it does not require manual annotation of a single complete parse tree. In contrast, traditional semi-supervised methods rely on fully-annotated seed corpora.18 9 Conclusion We explored novel ways of training dependency parsing models, the best of which attains 50.4% accuracy on Section 23 (all sentences) of WSJ, beating all previous unsupervised state-of-the-art by more than 5%. Extra gains stem from guiding Viterbi training with web mark-up, the loose constraint consistently delivering best results. Our linguistic analysis of a blog reveals that web annotations can be converted into accurate parsing constraints (loose: 88%; sprawl: 95%; tear: 99%) that could be helpful to supervised methods, e.g., by boosting an initial parser via self-training (McClosky et al., 2006) on sentences with mark-up. Similar techniques may apply to standard wordprocessing annotations, such as font changes, and to certain (balanced) punctuation (Briscoe, 1994). We make our blog data set, overlaying mark-up and syntax, publicly available. Its annotations are 18A significant effort expended in building a tree-bank comes with the first batch of sentences (Druck et al., 2009). 75% noun phrases, 13% verb phrases, 7% simple declarative clauses and 2% prepositional phrases, with traces of other phrases, clauses and fragments. The type of mark-up, combined with POS tags, could make for valuable features in discriminative models of parsing (Ratnaparkhi, 1999). A logical next step would be to explore the connection between syntax and mark-up for genres other than a news-style blog and for languages other than English. We are excited by the possibilities, as unsupervised parsers are on the cusp of becoming useful in their own right — recently, Davidov et al. (2009) successfully applied Seginer’s (2007) fully unsupervised grammar inducer to the problems of pattern-acquisition and extraction of semantic data. If the strength of the connection between web mark-up and syntactic structure is universal across languages and genres, this fact could have broad implications for NLP, with applications extending well beyond parsing. Acknowledgments Partially funded by NSF award IIS-0811974 and by the Air Force Research Laboratory (AFRL), under prime contract no. FA8750-09-C-0181; first author supported by the Fannie & John Hertz Foundation Fellowship. We thank Angel X. Chang, Spence Green, Christopher D. Manning, Richard Socher, Mihai Surdeanu and the anonymous reviewers for many helpful suggestions, and we are especially grateful to Andy Golding, for pointing us to his sample Map-Reduce over the Google News crawl, and to Daniel Pipes, for allowing us to distribute the data set derived from his blog entries. References S. Abney. 1991. Parsing by chunks. Principle-Based Parsing: Computation and Psycholinguistics. J. K. Baker. 1979. Trainable grammars for speech recognition. In Speech Communication Papers for the 97th Meeting of the Acoustical Society of America. C. Barr, R. Jones, and M. Regelson. 2008. The linguistic structure of English web-search queries. In EMNLP. T. Brants. 2000. TnT — a statistical part-of-speech tagger. In ANLP. 1286 T. Briscoe. 1994. Parsing (with) punctuation, etc. Technical report, Xerox European Research Laboratory. E. Charniak and M. Johnson. 2005. Coarse-to-fine n-best parsing and MaxEnt discriminative reranking. In ACL. E. Charniak. 2001. Immediate-head parsing for language models. In ACL. H.-H. Chen and Y.-S. Lee. 1995. Development of a partially bracketed corpus with part-of-speech information only. In WVLC. S. B. Cohen and N. A. Smith. 2009. Shared logistic normal distributions for soft parameter tying in unsupervised grammar induction. In NAACL-HLT. M. Collins. 1999. Head-Driven Statistical Models for Natural Language Parsing. Ph.D. thesis, University of Pennsylvania. D. Davidov, R. Reichart, and A. Rappoport. 2009. Superior and efficient fully unsupervised pattern-based concept acquisition using an unsupervised parser. In CoNLL. G. Druck, G. Mann, and A. McCallum. 2009. Semisupervised learning of dependency parsers using generalized expectation criteria. In ACL-IJCNLP. J. Eisner and G. Satta. 1999. Efficient parsing for bilexical context-free grammars and head-automaton grammars. In ACL. J. R. Finkel and C. D. Manning. 2009. Joint parsing and named entity recognition. In NAACL-HLT. W. N. Francis and H. Kucera, 1979. Manual of Information to Accompany a Standard Corpus of Present-Day Edited American English, for use with Digital Computers. Department of Linguistic, Brown University. E. Gabrilovich and S. Markovitch. 2007. Computing semantic relatedness using Wikipedia-based Explicit Semantic Analysis. In IJCAI. D. Gildea. 2001. Corpus variation and parser performance. In EMNLP. A. Halevy, P. Norvig, and F. Pereira. 2009. The unreasonable effectiveness of data. IEEE Intelligent Systems, 24. W. P. Headden, III, M. Johnson, and D. McClosky. 2009. Improving unsupervised dependency parsing with richer contexts and smoothing. In NAACL-HLT. S. Holm. 1979. A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics, 6. N. Inui and Y. Kotani. 2001. Robust N-gram based syntactic analysis using segmentation words. In PACLIC. D. Klein and C. D. Manning. 2004. Corpus-based induction of syntactic structure: Models of dependency and constituency. In ACL. C.-H. Lee, C.-H. Lin, and B.-H. Juang. 1991. A study on speaker adaptation of the parameters of continuous density Hidden Markov Models. IEEE Trans. on Signal Processing, 39. W.-H. Lu, L.-F. Chien, and H.-J. Lee. 2004. Anchor text mining for translation of Web queries: A transitive translation approach. ACM Trans. on Information Systems, 22. M. P. Marcus, B. Santorini, and M. A. Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19. D. McClosky, E. Charniak, and M. Johnson. 2006. Effective self-training for parsing. In NAACL-HLT. R. Mihalcea and A. Csomai. 2007. Wikify!: Linking documents to encyclopedic knowledge. In CIKM. M. Mintz, S. Bills, R. Snow, and D. Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In ACL-IJCNLP. J.-Y. Nie and J. Chen. 2002. Exploiting the Web as parallel corpora for cross-language information retrieval. Web Intelligence. F. Pereira and Y. Schabes. 1992. Inside-outside reestimation from partially bracketed corpora. In ACL. A. Ratnaparkhi. 1999. Learning to parse natural language with maximum entropy models. Machine Learning, 34. S. Ravi, K. Knight, and R. Soricut. 2008. Automatic prediction of parser accuracy. In EMNLP. J. C. Reynar and A. Ratnaparkhi. 1997. A maximum entropy approach to identifying sentence boundaries. In ANLP. S. Riezler, T. H. King, R. M. Kaplan, R. Crouch, J. T. Maxwell, III, and M. Johnson. 2002. Parsing the Wall Street Journal using a lexical-functional grammar and discriminative estimation techniques. In ACL. M. Sassano. 2005. Using a partially annotated corpus to build a dependency parser for Japanese. In IJCNLP. Y. Seginer. 2007. Fast unsupervised incremental parsing. In ACL. V. I. Spitkovsky, H. Alshawi, and D. Jurafsky. 2010a. From Baby Steps to Leapfrog: How “Less is More” in unsupervised dependency parsing. In NAACL-HLT. V. I. Spitkovsky, H. Alshawi, D. Jurafsky, and C. D. Manning. 2010b. Viterbi training improves unsupervised dependency parsing. In CoNLL. B. Tan and F. Peng. 2008. Unsupervised query segmentation using generative language models and Wikipedia. In WWW. K. Toutanova and C. D. Manning. 2000. Enriching the knowledge sources used in a maximum entropy part-ofspeech tagger. In EMNLP-VLC. K. Toutanova, D. Klein, C. D. Manning, and Y. Singer. 2003. Feature-rich part-of-speech tagging with a cyclic dependency network. In HLT-NAACL. D. Vadas and J. R. Curran. 2007. Adding noun phrase structure to the Penn Treebank. In ACL. Y. Watanabe, M. Asahara, and Y. Matsumoto. 2007. A graph-based approach to named entity categorization in Wikipedia using conditional random fields. In EMNLPCoNLL. E. Yeh, D. Ramage, C. D. Manning, E. Agirre, and A. Soroa. 2009. WikiWalk: Random walks on Wikipedia for semantic relatedness. In TextGraphs. 1287
|
2010
|
130
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1288–1297, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Phylogenetic Grammar Induction Taylor Berg-Kirkpatrick and Dan Klein Computer Science Division University of California, Berkeley {tberg, klein}@cs.berkeley.edu Abstract We present an approach to multilingual grammar induction that exploits a phylogeny-structured model of parameter drift. Our method does not require any translated texts or token-level alignments. Instead, the phylogenetic prior couples languages at a parameter level. Joint induction in the multilingual model substantially outperforms independent learning, with larger gains both from more articulated phylogenies and as well as from increasing numbers of languages. Across eight languages, the multilingual approach gives error reductions over the standard monolingual DMV averaging 21.1% and reaching as high as 39%. 1 Introduction Learning multiple languages together should be easier than learning them separately. For example, in the domain of syntactic parsing, a range of recent work has exploited the mutual constraint between two languages’ parses of the same bitext (Kuhn, 2004; Burkett and Klein, 2008; Kuzman et al., 2009; Smith and Eisner, 2009; Snyder et al., 2009a). Moreover, Snyder et al. (2009b) in the context of unsupervised part-of-speech induction (and Bouchard-Cˆot´e et al. (2007) in the context of phonology) show that extending beyond two languages can provide increasing benefit. However, multitexts are only available for limited languages and domains. In this work, we consider unsupervised grammar induction without bitexts or multitexts. Without translation examples, multilingual constraints cannot be exploited at the sentence token level. Rather, we capture multilingual constraints at a parameter level, using a phylogeny-structured prior to tie together the various individual languages’ learning problems. Our joint, hierarchical prior couples model parameters for different languages in a way that respects knowledge about how the languages evolved. Aspects of this work are closely related to Cohen and Smith (2009) and Bouchard-Cˆot´e et al. (2007). Cohen and Smith (2009) present a model for jointly learning English and Chinese dependency grammars without bitexts. In their work, structurally constrained covariance in a logistic normal prior is used to couple parameters between the two languages. Our work, though also different in technical approach, differs most centrally in the extension to multiple languages and the use of a phylogeny. Bouchard-Cˆot´e et al. (2007) considers an entirely different problem, phonological reconstruction, but shares with this work both the use of a phylogenetic structure as well as the use of log-linear parameterization of local model components. Our work differs from theirs primarily in the task (syntax vs. phonology) and the variables governed by the phylogeny: in our model it is the grammar parameters that drift (in the prior) rather than individual word forms (in the likelihood model). Specifically, we consider dependency induction in the DMV model of Klein and Manning (2004). Our data is a collection of standard dependency data sets in eight languages: English, Dutch, Danish, Swedish, Spanish, Portuguese, Slovene, and Chinese. Our focus is not the DMV model itself, which is well-studied, but rather the prior which couples the various languages’ parameters. While some choices of prior structure can greatly complicate inference (Cohen and Smith, 2009), we choose a hierarchical Gaussian form for the drift term, which allows the gradient of the observed data likelihood to be easily computed using standard dynamic programming methods. In our experiments, joint multilingual learning substantially outperforms independent monolingual learning. Using a limited phylogeny that 1288 only couples languages within linguistic families reduces error by 5.6% over the monolingual baseline. Using a flat, global phylogeny gives a greater reduction, almost 10%. Finally, a more articulated phylogeny that captures both inter- and intrafamily effects gives an even larger average relative error reduction of 21.1%. 2 Model We define our model over two kinds of random variables: dependency trees and parameters. For each language ℓin a set L, our model will generate a collection tℓof dependency trees ti ℓ. We assume that these dependency trees are generated by the DMV model of Klein and Manning (2004), which we write as ti ℓ∼DMV(θℓ). Here, θℓis a vector of the various model parameters for language ℓ. The prior is what couples the θℓparameter vectors across languages; it is the focus of this work. We first consider the likelihood model before moving on to the prior. 2.1 Dependency Model with Valence A dependency parse is a directed tree t over tokens in a sentence s. Each edge of the tree specifies a directed dependency from a head token to a dependent, or argument token. The DMV is a generative model for trees t, which has been widely used for dependency parse induction. The observed data likelihood, used for parameter estimation, is the marginal probability of generating the observed sentences s, which are simply the leaves of the trees t. Generation in the DMV model involves two types of local conditional probabilities: CONTINUE distributions that capture valence and ATTACH distributions that capture argument selection. First, the Bernoulli CONTINUE probability distributions P CONTINUE(c|h, dir, adj; θℓ) model the fertility of a particular head type h. The outcome c ∈{stop, continue} is conditioned on the head type h, direction dir, and adjacency adj. If a head type’s continue probability is low, tokens of this type will tend to generate few arguments. Second, the ATTACH multinomial probability distributions P ATTACH(a|h, dir; θℓ) capture attachment preferences of heads, where a and h are both token types. We take the same approach as previous work (Klein and Manning, 2004; Cohen and Smith, 2009) and use gold part-of-speech labels as tokens. Thus, the basic observed “word” types are English Dutch Swedish Danish Spanish Portuguese Slovene Chinese Global IndoEuropean Germanic West Germanic North Germanic IberoRomance Italic BaltoSlavic Slavic SinoTibetan Sinitic Figure 1: An example of a linguistically-plausible phylogenetic tree over the languages in our training data. Leaves correspond to (observed) modern languages, while internal nodes represent (unobserved) ancestral languages. actually word classes. 2.1.1 Log-Linear Parameterization The DMV’s local conditional distributions were originally given as simple multinomial distributions with one parameter per outcome. However, they can be re-parameterized to give the following log-linear form (Eisner, 2002; Bouchard-Cˆot´e et al., 2007; Berg-Kirkpatrick et al., 2010): P CONTINUE(c|h, dir, adj; θℓ) = exp ˆ θℓ T f CONTINUE(c, h, dir, adj) ˜ P c′ exp ˆ θℓ T f CONTINUE(c′, h, dir, adj) ˜ P ATTACH(a|h, dir; θℓ) = exp ˆ θℓ T f ATTACH(a, h, dir) ˜ P a′ exp ˆ θℓ T f ATTACH(a′, h, dir) ˜ The parameters are weights θℓwith one weight vector per language. In the case where the vector of feature functions f has an indicator for each possible conjunction of outcome and conditions, the original multinomial distributions are recovered. We refer to these full indicator features as the set of SPECIFIC features. 2.2 Phylogenetic Prior The focus of this work is coupling each of the parameters θℓin a phylogeny-structured prior. Consider a phylogeny like the one shown in Figure 1, where each modern language ℓin L is a leaf. We would like to say that the leaves’ parameter vectors arise from a process which slowly 1289 drifts along each branch. A convenient choice is to posit additional parameter variables θℓ+ at internal nodes ℓ+ ∈L+, a set of ancestral languages, and to assume that the conditional distribution P(θℓ|θpar(ℓ)) at each branch in the phylogeny is a Gaussian centered on θpar(ℓ), where par(ℓ) is the parent of ℓin the phylogeny and ℓranges over L ∪L+. The variance structure of the Gaussian would then determine how much drift (and in what directions) is expected. Concretely, we assume that each drift distribution is an isotropic Gaussian with mean θpar(ℓ) and scalar variance σ2. The root is centered at zero. We have thus defined a joint distribution P(Θ|σ2) where Θ = (θℓ: ℓ∈L∪L+). σ2 is a hyperparameter for this prior which could itself be re-parameterized to depend on branch length or be learned; we simply set it to a plausible constant value. Two primary challenges remain. First, inference under arbitrary priors can become complex. However, in the simple case of our diagonal covariance Gaussians, the gradient of the observed data likelihood can be computed directly using the DMV’s expected counts and maximum-likelihood estimation can be accomplished by applying standard gradient optimization methods. Second, while the choice of diagonal covariance is efficient, it causes components of θ that correspond to features occurring in only one language to be marginally independent of the parameters of all other languages. In other words, only features which fire in more than one language are coupled by the prior. In the next section, we therefore increase the overlap between languages’ features by using coarse projections of parts-of-speech. 2.3 Projected Features With diagonal covariance in the Gaussian drift terms, each parameter evolves independently of the others. Therefore, our prior will be most informative when features activate in multiple languages. In phonology, it is useful to map phonemes to the International Phonetic Alphabet (IPA) in order to have a language-independent parameterization. We introduce a similarly neutral representation here by projecting languagespecific parts-of-speech to a coarse, shared inventory. Indeed, we assume that each language has a distinct tagset, and so the basic configurational features will be language specific. For example, when SPECIFIC: Activate for only one conjunction of outcome and conditions: 1(c = ·, h = ·, dir = ·, adj = ·) SHARED: Activate for heads from multiple languages using cross-lingual POS projection π(·): 1(c = ·, π(h) = ·, dir = ·, adj = ·) CONTINUE distribution feature templates. SPECIFIC: Activate for only one conjunction of outcome and conditions: 1(a = ·, h = ·, dir = ·) SHARED: Activate for heads and arguments from multiple languages using cross-lingual POS projection π(·): 1(π(a) = ·, π(h) = ·, dir = ·) 1(π(a) = ·, h = ·, dir = ·) 1(a = ·, π(h) = ·, dir = ·) ATTACH distribution feature templates. Table 1: Feature templates for CONTINUE and ATTACH conditional distributions. an English VBZ takes a left argument headed by a NNS, a feature will activate specific to VBZ-NNSLEFT. That feature will be used in the log-linear attachment probability for English. However, because that feature does not show up in any other language, it is not usefully controlled by the prior. Therefore, we also include coarser features which activate on more abstract, cross-linguistic configurations. In the same example, a feature will fire indicating a coarse, direction-free NOUN-VERB attachment. This feature will now occur in multiple languages and will contribute to each of those languages’ attachment models. Although such crosslingual features will have different weight parameters in each language, those weights will covary, being correlated by the prior. The coarse features are defined via a projection π from language-specific part-of-speech labels to coarser, cross-lingual word classes, and hence we refer to them as SHARED features. For each corpus used in this paper, we use the tagging annotation guidelines to manually define a fixed mapping from the corpus tagset to the following coarse tagset: noun, verb, adjective, adverb, conjunction, preposition, determiner, interjection, numeral, and pronoun. Parts-of-speech for which this coarse mapping is ambiguous or impossible are not mapped, and do not have corresponding SHARED features. We summarize the feature templates for the CONTINUE and ATTACH conditional distributions in Table 1. Variants of all feature templates that ignore direction and/or adjacency are included. In practice, we found it beneficial for all language1290 independent features to ignore direction. Again, only the coarse features occur in multiple languages, so all phylogenetic influence is through those. Nonetheless, the effect of the phylogeny turns out to be quite strong. 2.4 Learning We now turn to learning with the phylogenetic prior. Since the prior couples parameters across languages, this learning problem requires parameters for all languages be estimated jointly. We seek to find Θ = (θℓ: ℓ∈L ∪L+) which optimizes log P(Θ|s), where s aggregates the observed leaves of all the dependency trees in all the languages. This can be written as log P(Θ) + log P(s|Θ) −log P(s) The third term is a constant and can be ignored. The first term can be written as log P(Θ) = X ℓ∈L∪L+ 1 2σ2 ∥θℓ−θpar(ℓ)∥2 2 + C where C is a constant. The form of log P(Θ) immediately shows how parameters are penalized for being different across languages, more so for languages that are near each other in the phylogeny. The second term log P(s|Θ) = X ℓ∈L log P(sℓ|θℓ) is a sum of observed data likelihoods under the standard DMV models for each language, computable by dynamic programming (Klein and Manning, 2004). Together, this yields the following objective function: l(Θ) = P ℓ∈L∪L+ 1 2σ2 ∥θℓ−θpar(ℓ)∥2 2 + P ℓ∈L log P(sℓ|θℓ) which can be optimized using gradient methods or (MAP) EM. Here we used L-BFGS (Liu et al., 1989). This requires computation of the gradient of the observed data likelihood log P(sℓ|θℓ) which is given by: ∇log P(sℓ|θℓ) = Etℓ|sℓ ∇log P(sℓ, tℓ|θℓ) = P c,h,dir,adj ec,h,dir,adj(sℓ; θℓ) · f CONTINUE(c, h, dir, adj) − P c′ P CONTINUE(c′|h, dir, adj; θℓ)f CONTINUE(c′, h, dir, adj) P a,h,dir ea,h,dir(sℓ; θℓ) · f ATTACH(a, h, dir) − P a′ P ATTACH(a′|h, dir; θℓ)f ATTACH(a′, h, dir) The expected gradient of the log joint likelihood of sentences and parses is equal to the gradient of the log marginal likelihood of just sentences, or the observed data likelihood (Salakhutdinov et al., 2003). ea,h,dir(sℓ; θℓ) is the expected count of the number of times head h is attached to a in direction dir given the observed sentences sℓand DMV parameters θℓ. ec,h,dir,adj(sℓ; θℓ) is defined similarly. Note that these are the same expected counts required to perform EM on the DMV, and are computable by dynamic programming. The computation time is dominated by the computation of each sentence’s posterior expected counts, which are independent given the parameters, so the time required per iteration is essentially the same whether training all languages jointly or independently. In practice, the total number of iterations was also similar. 3 Experimental Setup 3.1 Data We ran experiments with the following languages: English, Dutch, Danish, Swedish, Spanish, Portuguese, Slovene, and Chinese. For all languages but English and Chinese, we used corpora from the 2006 CoNLL-X Shared Task dependency parsing data set (Buchholz and Marsi, 2006). We used the shared task training set to both train and test our models. These corpora provide hand-labeled partof-speech tags (except for Dutch, which is automatically tagged) and provide dependency parses, which are either themselves hand-labeled or have been converted from hand-labeled parses of other kinds. For English and Chinese we use sections 2-21 of the Penn Treebank (PTB) (Marcus et al., 1993) and sections 1-270 of the Chinese Treebank (CTB) (Xue et al., 2002) respectively. Similarly, these sections were used for both training and testing. The English and Chinese data sets have hand-labeled constituency parses and part-ofspeech tags, but no dependency parses. We used the Bikel Chinese head finder (Bikel and Chiang, 2000) and the Collins English head finder (Collins, 1999) to transform the gold constituency parses into gold dependency parses. None of the corpora are bitexts. For all languages, we ran experiments on all sentences of length 10 or less after punctuation has been removed. When constructing phylogenies over the languages we made use of their linguistic classifications. English and Dutch are part of the West Ger1291 English Dutch Swedish Danish Spanish Portuguese Slovene Chinese West Germanic North Germanic IberoRomance Slavic Sinitic Global English Dutch Swedish Danish Spanish Portuguese Slovene Chinese Global (a) (b) (c) English Dutch Swedish Danish Spanish Portuguese Slovene Chinese West Germanic North Germanic IberoRomance Slavic Sinitic Figure 2: (a) Phylogeny for FAMILIES model. (b) Phylogeny for GLOBAL model. (c) Phylogeny for LINGUISTIC model. manic family of languages, whereas Danish and Swedish are part of the North Germanic family. Spanish and Portuguese are both part of the IberoRomance family. Slovene is part of the Slavic family. Finally, Chinese is in the Sinitic family, and is not an Indo-European language like the others. We interchangeably speak of a language family and the ancestral node corresponding to that family’s root language in a phylogeny. 3.2 Models Compared We evaluated three phylogenetic priors, each with a different phylogenetic structure. We compare with two monolingual baselines, as well as an allpairs multilingual model that does not have a phylogenetic interpretation, but which provides very similar capacity for parameter coupling. 3.2.1 Phylogenetic Models The first phylogenetic model uses the shallow phylogeny shown in Figure 2(a), in which only languages within the same family have a shared parent node. We refer to this structure as FAMILIES. Under this prior, the learning task decouples into independent subtasks for each family, but no regularities across families can be captured. The family-level model misses the constraints between distant languages. Figure 2(b) shows another simple configuration, wherein all languages share a common parent node in the prior, meaning that global regularities that are consistent across all languages can be captured. We refer to this structure as GLOBAL. While the global model couples the parameters for all eight languages, it does so without sensitivity to the articulated structure of their descent. Figure 2(c) shows a more nuanced prior structure, LINGUISTIC, which groups languages first by family and then under a global node. This structure allows global regularities as well as regularities within families to be learned. 3.2.2 Parameterization and ALLPAIRS Model Daum´e III (2007) and Finkel and Manning (2009) consider a formally similar Gaussian hierarchy for domain adaptation. As pointed out in Finkel and Manning (2009), there is a simple equivalence between hierarchical regularization as described here and the addition of new tied features in a “flat” model with zero-meaned Gaussian regularization on all parameters. In particular, instead of parameterizing the objective in Section 2.4 in terms of multiple sets of weights, one at each node in the phylogeny (the hierarchical parameterization, described in Section 2.4), it is equivalent to parameterize this same objective in terms of a single set of weights on a larger of group features (the flat parameterization). This larger group of features contains a duplicate set of the features discussed in Section 2.3 for each node in the phylogeny, each of which is active only on the languages that are its descendants. A linear transformation between parameterizations gives equivalence. See Finkel and Manning (2009) for details. In the flat parameterization, it seems equally reasonable to simply tie all pairs of languages by adding duplicate sets of features for each pair. This gives the ALLPAIRS setting, which we also compare to the tree-structured phylogenetic models above. 3.3 Baselines To evaluate the impact of multilingual constraint, we compared against two monolingual baselines. The first baseline is the standard DMV with only SPECIFIC features, which yields the standard multinomial DMV (weak baseline). To facilitate comparison to past work, we used no prior for this monolingual model. The second baseline is the DMV with added SHARED features. This model includes a simple isotropic Gaussian prior on pa1292 Monolingual Multilingual Phylogenetic Corpus Size Baseline Baseline w/ SHARED ALLPAIRS FAMILIES BESTPAIR GLOBAL LINGUISTIC West Germanic English 6008 47.1 51.3 48.5 51.3 51.3 (Ch) 51.2 62.3 Dutch 6678 36.3 36.0 44.0 36.1 36.2 (Sw) 44.0 45.1 North Germanic Danish 1870 33.5 33.6 40.5 31.4 34.2 (Du) 39.6 41.6 Swedish 3571 45.3 44.8 56.3 44.8 44.8 (Ch) 44.5 58.3 Ibero-Romance Spanish 712 28.0 40.5 58.7 63.4 63.8 (Da) 59.4 58.4 Portuguese 2515 38.5 38.5 63.1 37.4 38.4 (Sw) 37.4 63.0 Slavic Slovene 627 38.5 39.7 49.0 – 49.6 (En) 49.4 48.4 Sinitic Chinese 959 36.3 43.3 50.7 – 49.7 (Sw) 50.1 49.6 Macro-Avg. Relative Error Reduction 17.1 5.6 8.5 9.9 21.1 Table 2: Directed dependency accuracy of monolingual and multilingual models, and relative error reduction over the monolingual baseline with SHARED features macro-averaged over languages. Multilingual models outperformed monolingual models in general, with larger gains from increasing numbers of languages. Additionally, more nuanced phylogenetic structures outperformed cruder ones. rameters. This second baseline is the more direct comparison to the multilingual experiments here (strong baseline). 3.4 Evaluation For each setting, we evaluated the directed dependency accuracy of the minimum Bayes risk (MBR) dependency parses produced by our models under maximum (posterior) likelihood parameter estimates. We computed accuracies separately for each language in each condition. In addition, for multilingual models, we computed the relative error reduction over the strong monolingual baseline, macro-averaged over languages. 3.5 Training Our implementation used the flat parameterization described in Section 3.2.2 for both the phylogenetic and ALLPAIRS models. We originally did this in order to facilitate comparison with the non-phylogenetic ALLPAIRS model, which has no equivalent hierarchical parameterization. In practice, optimizing with the hierarchical parameterization also seemed to underperform.1 1We noticed that the weights of features shared across languages had larger magnitude early in the optimization procedure when using the flat parameterization compared to using the hierarchical parameterization, perhaps indicating that cross-lingual influences had a larger effect on learning in its initial stages. All models were trained by directly optimizing the observed data likelihood using L-BFGS (Liu et al., 1989). Berg-Kirkpatrick et al. (2010) suggest that directly optimizing the observed data likelihood may offer improvements over the more standard expectation-maximization (EM) optimization procedure for models such as the DMV, especially when the model is parameterized using features. We stopped training after 200 iterations in all cases. This fixed stopping criterion seemed to be adequate in all experiments, but presumably there is a potential gain to be had in fine tuning. To initialize, we used the harmonic initializer presented in Klein and Manning (2004). This type of initialization is deterministic, and thus we did not perform random restarts. We found that for all models σ2 = 0.2 gave reasonable results, and we used this setting in all experiments. For most models, we found that varying σ2 in a reasonable range did not substantially affect accuracy. For some models, the directed accuracy was less flat with respect to σ2. In these less-stable cases, there seemed to be an interaction between the variance and the choice between head conventions. For example, for some settings of σ2, but not others, the model would learn that determiners head noun phrases. In particular, we observed that even when direct accuracy did fluctuate, undirected accuracy remained more stable. 1293 4 Results Table 2 shows the overall results. In all cases, methods which coupled the languages in some way outperformed the independent baselines that considered each language independently. 4.1 Bilingual Models The weakest of the coupled models was FAMILIES, which had an average relative error reduction of 5.6% over the strong baseline. In this case, most of the average improvement came from a single family: Spanish and Portuguese. The limited improvement of the family-level prior compared to other phylogenies suggests that there are important multilingual interactions that do not happen within families. Table 2 also reports the maximum accuracy achieved for each language when it was paired with another language (same family or otherwise) and trained together with a single common parent. These results appear in the column headed by BESTPAIR, and show the best accuracy for the language on that row over all possible pairings with other languages. When pairs of languages were trained together in isolation, the largest benefit was seen for languages with small training corpora, not necessarily languages with common ancestry. In our setup, Spanish, Slovene, and Chinese have substantially smaller training corpora than the rest of the languages considered. Otherwise, the patterns are not particularly clear; combined with subsequent results, it seems that pairwise constraint is fairly limited. 4.2 Multilingual Models Models that coupled multiple languages performed better in general than models that only considered pairs of languages. The GLOBAL model, which couples all languages, if crudely, yielded an average relative error reduction of 9.9%. This improvement comes as the number of languages able to exert mutual constraint increases. For example, Dutch and Danish had large improvements, over and above any improvements these two languages gained when trained with a single additional language. Beyond the simplistic GLOBAL phylogeny, the more nuanced LINGUISTIC model gave large improvements for English, Swedish, and Portuguese. Indeed, the LINGUISTIC model is the only model we evaluated that gave improvements for all the languages we considered. It is reasonable to worry that the improvements from these multilingual models might be partially due to having more total training data in the multilingual setting. However, we found that halving the amount of data used to train the English, Dutch, and Swedish (the languages with the most training data) monolingual models did not substantially affect their performance, suggesting that for languages with several thousand sentences or more, the increase in statistical support due to additional monolingual data was not an important effect (the DMV is a relatively low-capacity model in any case). 4.3 Comparison of Phylogenies Recall the structures of the three phylogenies presented in Figure 2. These phylogenies differ in the correlations they can represent. The GLOBAL phylogeny captures only “universals,” while FAMILIES captures only correlations between languages that are known to be similar. The LINGUISTIC model captures both of these effects simultaneously by using a two layer hierarchy. Notably, the improvement due to the LINGUISTIC model is more than the sum of the improvements due to the GLOBAL and FAMILIES models. 4.4 Phylogenetic vs. ALLPAIRS The phylogeny is capable of allowing appropriate influence to pass between languages at multiple levels. We compare these results to the ALLPAIRS model in order to see whether limitation to a tree structure is helpful. The ALLPAIRS model achieved an average relative error reduction of 17.1%, certainly outperforming both the simple phylogenetic models. However, the rich phylogeny of the LINGUISTIC model, which incorporates linguistic constraints, outperformed the freer ALLPAIRS model. A large portion of this improvement came from English, a language for which the LINGUISTIC model greatly outperformed all other models evaluated. We found that the improved English analyses produced by the LINGUISTIC model were more consistent with this model’s analyses of other languages. This consistency was not present for the English analyses produced by other models. We explore consistency in more detail in Section 5. 4.5 Comparison to Related Work The likelihood models for both the strong monolingual baseline and the various multilingual mod1294 els are the same, both expanding upon the standard DMV by adding coarse SHARED features. These coarse features, even in a monolingual setting, improved performance slightly over the weak baseline, perhaps by encouraging consistent treatment of the different finer-grained variants of partsof-speech (Berg-Kirkpatrick et al., 2010).2 The only difference between the multilingual systems and the strong baseline is whether or not crosslanguage influence is allowed through the prior. While this progression of model structure is similar to that explored in Cohen and Smith (2009), Cohen and Smith saw their largest improvements from tying together parameters for the varieties of coarse parts-of-speech monolinugally, and then only moderate improvements from allowing cross-linguistic influence on top of monolingual sharing. When Cohen and Smith compared their best shared logistic-normal bilingual models to monolingual counter-parts for the languages they investigate (Chinese and English), they reported a relative error reduction of 5.3%. In comparison, with the LINGUISTIC model, we saw a much larger 16.9% relative error reduction over our strong baseline for these languages. Evaluating our LINGUISTIC model on the same test sets as (Cohen and Smith, 2009), sentences of length 10 or less in section 23 of PTB and sections 271300 of CTB, we achieved an accuracy of 56.6 for Chinese and 60.3 for English. The best models of Cohen and Smith (2009) achieved accuracies of 52.0 and 62.0 respectively on these same test sets. Our results indicate that the majority of our model’s power beyond that of the standard DMV is derived from multilingual, and in particular, more-than-bilingual, interaction. These are, to the best of our knowledge, the first results of this kind for grammar induction without bitext. 5 Analysis By examining the proposed parses we found that the LINGUISTIC and ALLPAIRS models produced analyses that were more consistent across languages than those of the other models. We also observed that the most common errors can be summarized succinctly by looking at attachment counts between coarse parts-of-speech. Figure 3 shows matrix representations of dependency 2Coarse features that only tie nouns and verbs are explored in Berg-Kirkpatrick et al. (2010). We found that these were very effective for English and Chinese, but gave worse performance for other languages. counts. The area of a square is proportional to the number of order-collapsed dependencies where the column label is the head and the row label is the argument in the parses from each system. For ease of comprehension, we use the cross-lingual projections and only show counts for selected interesting classes. Comparing Figure 3(c), which shows dependency counts proposed by the LINGUISTIC model, to Figure 3(a), which shows the same for the strong monolingual baseline, suggests that the analyses proposed by the LINGUISTIC model are more consistent across languages than are the analyses proposed by the monolingual model. For example, the monolingual learners are divided as to whether determiners or nouns head noun phrases. There is also confusion about which labels head whole sentences. Dutch has the problem that verbs modify pronouns more often than pronouns modify verbs, and pronouns are predicted to head sentences as often as verbs are. Spanish has some confusion about conjunctions, hypothesizing that verbs often attach to conjunctions, and conjunctions frequently head sentences. More subtly, the monolingual analyses are inconsistent in the way they head prepositional phrases. In the monolingual Portuguese hypotheses, prepositions modify nouns more often than nouns modify prepositions. In English, nouns modify prepositions, and prepositions modify verbs. Both the Dutch and Spanish models are ambivalent about the attachment of prepositions. As has often been observed in other contexts (Liang et al., 2008), promoting agreement can improve accuracy in unsupervised learning. Not only are the analyses proposed by the LINGUISTIC model more consistent, they are also more in accordance with the gold analyses. Under the LINGUISTIC model, Dutch now attaches pronouns to verbs, and thus looks more like English, its sister in the phylogenetic tree. The LINGUISTIC model has also chosen consistent analyses for prepositional phrases and noun phrases, calling prepositions and nouns the heads of each, respectively. The problem of conjunctions heading Spanish sentences has also been corrected. Figure 3(b) shows dependency counts for the GLOBAL multilingual model. Unsurprisingly, the analyses proposed under global constraint appear somewhat more consistent than those proposed under no multi-lingual constraint (now three lan1295 Figure 3: Dependency counts in proposed parses. Row label modifies column label. (a) Monolingual baseline with SHARED features. (b) GLOBAL model. (c) LINGUISTIC model. (d) Dependency counts in hand-labeled parses. Analyses proposed by monolingual baseline show significant inconsistencies across languages. Analyses proposed by LINGUISTIC model are more consistent across languages than those proposed by either the monolingual baseline or the GLOBAL model. guages agree that prepositional phrases are headed by prepositions), but not as consistent as those proposed by the LINGUISTIC model. Finally, Figure 3(d) shows dependency counts in the hand-labeled dependency parses. It appears that even the very consistent LINGUISTIC parses do not capture the non-determinism of prepositional phrase attachment to both nouns and verbs. 6 Conclusion Even without translated texts, multilingual constraints expressed in the form of a phylogenetic prior on parameters can give substantial gains in grammar induction accuracy over treating languages in isolation. Additionally, articulated phylogenies that are sensitive to evolutionary structure can outperform not only limited flatter priors but also unconstrained all-pairs interactions. 7 Acknowledgements This project is funded in part by the NSF under grant 0915265 and DARPA under grant N10AP20007. 1296 References T. Berg-Kirkpatrick, A. Bouchard-Cˆot´e, J. DeNero, and D. Klein. 2010. Painless unsupervised learning with features. In North American Chapter of the Association for Computational Linguistics. D. M. Bikel and D. Chiang. 2000. Two statistical parsing models applied to the Chinese treebank. In Second Chinese Language Processing Workshop. A. Bouchard-Cˆot´e, P. Liang, D. Klein, and T. L. Griffiths. 2007. A probabilistic approach to diachronic phonology. In Empirical Methods in Natural Language Processing. S. Buchholz and E. Marsi. 2006. Computational Natural Language Learning-X shared task on multilingual dependency parsing. In Conference on Computational Natural Language Learning. D. Burkett and D. Klein. 2008. Two languages are better than one (for syntactic parsing). In Empirical Methods in Natural Language Processing. S. B. Cohen and N. A. Smith. 2009. Shared logistic normal distributions for soft parameter tying in unsupervised grammar induction. In North American Chapter of the Association for Computational Linguistics. M. Collins. 1999. Head-driven statistical models for natural language parsing. In Ph.D. thesis, University of Pennsylvania, Philadelphia. H. Daum´e III. 2007. Frustratingly easy domain adaptation. In Association for Computational Linguistics. J. Eisner. 2002. Parameter estimation for probabilistic finite-state transducers. In Association for Computational Linguistics. J. R. Finkel and C. D. Manning. 2009. Hierarchical bayesian domain adaptation. In North American Chapter of the Association for Computational Linguistics. D. Klein and C. D. Manning. 2004. Corpus-based induction of syntactic structure: Models of dependency and constituency. In Association for Computational Linguistics. J. Kuhn. 2004. Experiments in parallel-text based grammar induction. In Association for Computational Linguistics. G. Kuzman, J. Gillenwater, and B. Taskar. 2009. Dependency grammar induction via bitext projection constraints. In Association for Computational Linguistics/International Joint Conference on Natural Language Processing. P. Liang, D. Klein, and M. I. Jordan. 2008. Agreement-based learning. In Advances in Neural Information Processing Systems. D. C. Liu, J. Nocedal, and C. Dong. 1989. On the limited memory BFGS method for large scale optimization. Mathematical Programming. M. P. Marcus, M. A. Marcinkiewicz, and B. Santorini. 1993. Building a large annotated corpus of English: the penn treebank. Computational Linguistics. R. Salakhutdinov, S. Roweis, and Z. Ghahramani. 2003. Optimization with EM and expectationconjugate-gradient. In International Conference on Machine Learning. D. A. Smith and J. Eisner. 2009. Parser adaptation and projection with quasi-synchronous grammar features. In Empirical Methods in Natural Language Processing. B. Snyder, T. Naseem, and R. Barzilay. 2009a. Unsupervised multilingual grammar induction. In Association for Computational Linguistics/International Joint Conference on Natural Language Processing. B. Snyder, T. Naseem, J. Eisenstein, and R. Barzilay. 2009b. Adding more languages improves unsupervised multilingual part-of-speech tagging: A Bayesian non-parametric approach. In North American Chapter of the Association for Computational Linguistics. N. Xue, F-D Chiou, and M. Palmer. 2002. Building a large-scale annotated Chinese corpus. In International Conference on Computational Linguistics. 1297
|
2010
|
131
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1298–1307, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Improved Unsupervised POS Induction through Prototype Discovery Omri Abend1∗Roi Reichart2 Ari Rappoport1 1Institute of Computer Science, 2ICNC Hebrew University of Jerusalem {omria01|roiri|arir}@cs.huji.ac.il Abstract We present a novel fully unsupervised algorithm for POS induction from plain text, motivated by the cognitive notion of prototypes. The algorithm first identifies landmark clusters of words, serving as the cores of the induced POS categories. The rest of the words are subsequently mapped to these clusters. We utilize morphological and distributional representations computed in a fully unsupervised manner. We evaluate our algorithm on English and German, achieving the best reported results for this task. 1 Introduction Part-of-speech (POS) tagging is a fundamental NLP task, used by a wide variety of applications. However, there is no single standard POS tagging scheme, even for English. Schemes vary significantly across corpora and even more so across languages, creating difficulties in using POS tags across domains and for multi-lingual systems (Jiang et al., 2009). Automatic induction of POS tags from plain text can greatly alleviate this problem, as well as eliminate the efforts incurred by manual annotations. It is also a problem of great theoretical interest. Consequently, POS induction is a vibrant research area (see Section 2). In this paper we present an algorithm based on the theory of prototypes (Taylor, 2003), which posits that some members in cognitive categories are more central than others. These practically define the category, while the membership of other elements is based on their association with the ∗Omri Abend is grateful to the Azrieli Foundation for the award of an Azrieli Fellowship. central members. Our algorithm first clusters words based on a fine morphological representation. It then clusters the most frequent words, defining landmark clusters which constitute the cores of the categories. Finally, it maps the rest of the words to these categories. The last two stages utilize a distributional representation that has been shown to be effective for unsupervised parsing (Seginer, 2007). We evaluated the algorithm in both English and German, using four different mapping-based and information theoretic clustering evaluation measures. The results obtained are generally better than all existing POS induction algorithms. Section 2 reviews related work. Sections 3 and 4 detail the algorithm. Sections 5, 6 and 7 describe the evaluation, experimental setup and results. 2 Related Work Unsupervised and semi-supervised POS tagging have been tackled using a variety of methods. Sch¨utze (1995) applied latent semantic analysis. The best reported results (when taking into account all evaluation measures, see Section 5) are given by (Clark, 2003), which combines distributional and morphological information with the likelihood function of the Brown algorithm (Brown et al., 1992). Clark’s tagger is very sensitive to its initialization. Reichart et al. (2010b) propose a method to identify the high quality runs of this algorithm. In this paper, we show that our algorithm outperforms not only Clark’s mean performance, but often its best among 100 runs. Most research views the task as a sequential labeling problem, using HMMs (Merialdo, 1994; Banko and Moore, 2004; Wang and Schuurmans, 2005) and discriminative models (Smith and Eisner, 2005; Haghighi and Klein, 2006). Several 1298 techniques were proposed to improve the HMM model. A Bayesian approach was employed by (Goldwater and Griffiths, 2007; Johnson, 2007; Gao and Johnson, 2008). Van Gael et al. (2009) used the infinite HMM with non-parametric priors. Grac¸a et al. (2009) biased the model to induce a small number of possible tags for each word. The idea of utilizing seeds and expanding them to less reliable data has been used in several papers. Haghighi and Klein (2006) use POS ‘prototypes’ that are manually provided and tailored to a particular POS tag set of a corpus. Freitag (2004) and Biemann (2006) induce an initial clustering and use it to train an HMM model. Dasgupta and Ng (2007) generate morphological clusters and use them to bootstrap a distributional model. Goldberg et al. (2008) use linguistic considerations for choosing a good starting point for the EM algorithm. Zhao and Marcus (2009) expand a partial dictionary and use it to learn disambiguation rules. Their evaluation is only at the type level and only for half of the words. Ravi and Knight (2009) use a dictionary and an MDLinspired modification to the EM algorithm. Many of these works use a dictionary providing allowable tags for each or some of the words. While this scenario might reduce human annotation efforts, it does not induce a tagging scheme but remains tied to an existing one. It is further criticized in (Goldwater and Griffiths, 2007). Morphological representation. Many POS induction models utilize morphology to some extent. Some use simplistic representations of terminal letter sequences (e.g., (Smith and Eisner, 2005; Haghighi and Klein, 2006)). Clark (2003) models the entire letter sequence as an HMM and uses it to define a morphological prior. Dasgupta and Ng (2007) use the output of the Morfessor segmentation algorithm for their morphological representation. Morfessor (Creutz and Lagus, 2005), which we use here as well, is an unsupervised algorithm that segments words and classifies each segment as being a stem or an affix. It has been tested on several languages with strong results. Our work has several unique aspects. First, our clustering method discovers prototypes in a fully unsupervised manner, mapping the rest of the words according to their association with the prototypes. Second, we use a distributional representation which has been shown to be effective for unsupervised parsing (Seginer, 2007). Third, we use a morphological representation based on signatures, which are sets of affixes that represent a family of words sharing an inflectional or derivational morphology (Goldsmith, 2001). 3 Distributional Algorithm Our algorithm is given a plain text corpus and optionally a desired number of clusters k. Its output is a partitioning of words into clusters. The algorithm utilizes two representations, distributional and morphological. Although eventually the latter is used before the former, for clarity of presentation we begin by detailing the base distributional algorithm. In the next section we describe the morphological representation and its integration into the base algorithm. Overview. The algorithm consists of two main stages: landmark clusters discovery, and word mapping. For the former, we first compute a distributional representation for each word. We then cluster the coordinates corresponding to high frequency words. Finally, we define landmark clusters. In the word mapping stage we map each word to the most similar landmark cluster. The rationale behind using only the high frequency words in the first stage is twofold. First, prototypical members of a category are frequent (Taylor, 2003), and therefore we can expect the salient POS tags to be represented in this small subset. Second, higher frequency implies more reliable statistics. Since this stage determines the cores of all resulting clusters, it should be as accurate as possible. Distributional representation. We use a simplified form of the elegant representation of lexical entries used by the Seginer unsupervised parser (Seginer, 2007). Since a POS tag reflects the grammatical role of the word and since this representation is effective to parsing, we were motivated to apply it to the present task. Let W be the set of word types in the corpus. The right context entry of a word x ∈W is a pair of mappings r intx : W →[0, 1] and r adjx : W →[0, 1]. For each w ∈W, r adjx(w) is an adjacency score of w to x, reflecting w’s tendency to appear on the right hand side of x. For each w ∈W, r intx(w) is an interchangeability score of x with w, reflecting the tendency of w to appear to the left of words that tend to appear to the right of x. This can be viewed as a 1299 similarity measure between words with respect to their right context. The higher the scores the more the words tend to be adjacent/interchangeable. Left context parameters l intx and l adjx are defined analogously. There are important subtleties in these definitions. First, for two words x, w ∈W, r adjx(w) is generally different from l adjw(x). For example, if w is a high frequency word and x is a low frequency word, it is likely that w appears many times to the right of x, yielding a high r adjx(w), but that x appears only a few times to the left of w yielding a low l adjw(x). Second, from the definition of r intx(w) and r intw(x), it is clear that they need not be equal. These functions are computed incrementally by a bootstrapping process. We initialize all mappings to be identically 0. We iterate over the words in the training corpus. For every word instance x, we take the word immediately to its right y and update x’s right context using y’s left context: ∀w ∈W : r intx(w) += l adjy(w) N(y) ∀w ∈W : r adjx(w) += ( 1 w = y l inty(w) N(y) w ̸= y The division by N(y) (the number of times y appears in the corpus before the update) is done in order not to give a disproportional weight to high frequency words. Also, r intx(w) and r adjx(w) might become larger than 1. We therefore normalize them after all updates are performed by the number of occurrences of x in the corpus. We update l intx and l adjx analogously using the word z immediately to the left of x. The updates of the left and right functions are done in parallel. We define the distributional representation of a word type x to be a 4|W| + 2 dimensional vector vx. Each word w yields four coordinates, one for each direction (left/right) and one for each mapping type (int/adj). Two additional coordinates represent the frequency in which the word appears to the left and to the right of a stopping punctuation. Of the 4|W| coordinates corresponding to words, we allow only 2n to be non-zero: the n top scoring among the right side coordinates (those of r intx and r adjx), and the n top scoring among the left side coordinates (those of l intx and l adjx). We used n = 50. The distance between two words is defined to be one minus the cosine of the angle between their representation vectors. Coordinate clustering. Each of our landmark clusters will correspond to a set of high frequency words (HFWs). The number of HFWs is much larger than the number of expected POS tags. Hence we should cluster HFWs. Our algorithm does that by unifying some of the non-zero coordinates corresponding to HFWs in the distributional representation defined above. We extract the words that appear more than N times per million1 and apply the following procedure I times (5 in our experiments). We run average link clustering with a threshold α (AVGLINKα, (Jain et al., 1999)) on these words, in each iteration initializing every HFW to have its own cluster. AVGLINKα means running the average link algorithm until the two closest clusters have a distance larger than α. We then use the induced clustering to update the distributional representation, by collapsing all coordinates corresponding to words appearing in the same cluster into a single coordinate whose value is the sum of the collapsed coordinates’ values. In order to produce a conservative (fine) clustering, we used a relatively low α value of 0.25. Note that the AVGLINKα initialization in each of the I iterations assigns each HFW to a separate cluster. The iterations differ in the distributional representation of the HFWs, resulting from the previous iterations. In our English experiments, this process reduced the dimension of the HFWs set (the number of coordinates that are non-zero in at least one of the HFWs) from 14365 to 10722. The average number of non-zero coordinates per word decreased from 102 to 55. Since all eventual POS categories correspond to clusters produced at this stage, to reduce noise we delete clusters of less than five elements. Landmark detection. We define landmark clusters using the clustering obtained in the final iteration of the coordinate clustering stage. However, the number of clusters might be greater than the desired number k, which is an optional parameter of the algorithm. In this case we select a subset of k clusters that best covers the HFW space. We use the following heuristic. We start from the most frequent cluster, and greedily select the clus1We used N = 100, yielding 1242 words for English and 613 words for German. 1300 ter farthest from the clusters already selected. The distance between two clusters is defined to be the average distance between their members. A cluster’s distance from a set of clusters is defined to be its minimal distance from the clusters in the set. The final set of clusters {L1, ..., Lk} and their members are referred to as landmark clusters and prototypes, respectively. Mapping all words. Each word w ∈W is assigned the cluster Li that contains its nearest prototype: d(w, Li) = minx∈Li{1 −cos(vw, vx)} Map(w) = argminLi{d(w, Li)} Words that appear less than 5 times are considered as unknown words. We consider two schemes for handling unknown words. One randomly maps each such word to a cluster, using a probability proportional to the number of unique known words already assigned to that cluster. However, when the number k of landmark clusters is relatively large, it is beneficial to assign all unknown words to a separate new cluster (after running the algorithm with k −1). In our experiments, we use the first option when k is below some threshold (we used 15), otherwise we use the second. 4 Morphological Model The morphological model generates another word clustering, based on the notion of a signature. This clustering is integrated with the distributional model as described below. 4.1 Morphological Representation We use the Morfessor (Creutz and Lagus, 2005) word segmentation algorithm. First, all words in the corpus are segmented. Then, for each stem, the set of all affixes with which it appears (its signature, (Goldsmith, 2001)) is collected. The morphological representation of a word type is then defined to be its stem’s signature in conjunction with its specific affixes2 (See Figure 1). We now collect all words having the same representation. For instance, if the words joined and painted are found to have the same signature, they would share the same cluster since both have the affix ‘ ed’. The word joins does not share the same cluster with them since it has a different affix, ‘ s’. This results in coarse-grained clusters exclusively defined according to morphology. 2A word may contain more than a single affix. Types join joins joined joining Stem join join join join Affixes φ s ed ing Signature {φ, ed, s, ing} Figure 1: An example for a morphological representation, defined to be the conjunction of its affix(es) with the stem’s signature. In addition, we incorporate capitalization information into the model, by constraining all words that appear capitalized in more than half of their instances to belong to a separate cluster, regardless of their morphological representation. The motivation for doing so is practical: capitalization is used in many languages to mark grammatical categories. For instance, in English capitalization marks the category of proper names and in German it marks the noun category . We report English results both with and without this modification. Words that contain non-alphanumeric characters are represented as the sequence of the nonalphanumeric characters they include, e.g., ‘vis-`avis’ is represented as (“-”, “-”). We do not assign a morphological representation to words including more than one stem (like weatherman), to words that have a null affix (i.e., where the word is identical to its stem) and to words whose stem is not shared by any other word (signature of size 1). Words that were not assigned a morphological representation are included as singletons in the morphological clustering. 4.2 Distributional-Morphological Algorithm We detail the modifications made to our base distributional algorithm given the morphological clustering defined above. Coordinate clustering and landmarks. We constrain AVGLINKα to begin by forming links between words appearing in the same morphological cluster. Only when the distance between the two closest clusters gets above α we remove this constraint and proceed as before. This is equivalent to performing AVGLINKα separately within each morphological cluster and then using the result as an initial condition for an AVGLINKα coordinate clustering. The modified algorithm in this stage is otherwise identical to the distributional algorithm. Word mapping. In this stage words that are not prototypes are mapped to one of the landmark 1301 clusters. A reasonable strategy would be to map all words sharing a morphological cluster as a single unit. However, these clusters are too coarsegrained. We therefore begin by partitioning the morphological clusters into sub-clusters according to their distributional behavior. We do so by applying AVGLINKβ (the same as AVGLINKα but with a different parameter) to each morphological cluster. Since our goal is cluster refinement, we use a β that is considerably higher than α (0.9). We then find the closest prototype to each such sub-cluster (averaging the distance across all of the latter’s members) and map it as a single unit to the cluster containing that prototype. 5 Clustering Evaluation We evaluate the clustering produced by our algorithm using an external quality measure: we take a corpus tagged by gold standard tags, tag it using the induced tags, and compare the two taggings. There is no single accepted measure quantifying the similarity between two taggings. In order to be as thorough as possible, we report results using four known measures, two mapping-based measures and two information theoretic ones. Mapping-based measures. The induced clusters have arbitrary names. We define two mapping schemes between them and the gold clusters. After the induced clusters are mapped, we can compute a derived accuracy. The Many-to-1 measure finds the mapping between the gold standard clusters and the induced clusters which maximizes accuracy, allowing several induced clusters to be mapped to the same gold standard cluster. The 1-to-1 measure finds the mapping between the induced and gold standard clusters which maximizes accuracy such that no two induced clusters are mapped to the same gold cluster. Computing this mapping is equivalent to finding the maximal weighted matching in a bipartite graph, whose weights are given by the intersection sizes between matched classes/clusters. As in (Reichart and Rappoport, 2008), we use the Kuhn-Munkres algorithm (Kuhn, 1955; Munkres, 1957) to solve this problem. Information theoretic measures. These are based on the observation that a good clustering reduces the uncertainty of the gold tag given the induced cluster, and vice-versa. Several such measures exist; we use V (Rosenberg and Hirschberg, 2007) and NVI (Reichart and Rappoport, 2009), VI’s (Meila, 2007) normalized version. 6 Experimental Setup Since a goal of unsupervised POS tagging is inducing an annotation scheme, comparison to an existing scheme is problematic. To address this problem we compare to three different schemes in two languages. In addition, the two English schemes we compare with were designed to tag corpora contained in our training set, and have been widely and successfully used with these corpora by a large number of applications. Our algorithm was run with the exact same parameters on both languages: N = 100 (high frequency threshold), n = 50 (the parameter that determines the effective number of coordinates), α = 0.25 (cluster separation during landmark cluster generation), β = 0.9 (cluster separation during refinement of morphological clusters). The algorithm we compare with in most detail is (Clark, 2003), which reports the best current results for this problem (see Section 7). Since Clark’s algorithm is sensitive to its initialization, we ran it a 100 times and report its average and standard deviation in each of the four measures. In addition, we report the percentile in which our result falls with respect to these 100 runs. Punctuation marks are very frequent in corpora and are easy to cluster. As a result, including them in the evaluation greatly inflates the scores. For this reason we do not assign a cluster to punctuation marks and we report results using this policy, which we recommend for future work. However, to be able to directly compare with previous work, we also report results for the full POS tag set. We do so by assigning a singleton cluster to each punctuation mark (in addition to the k required clusters). This simple heuristic yields very high performance on punctuation, scoring (when all other words are assumed perfect tagging) 99.6% (99.1%) 1-to-1 accuracy when evaluated against the English fine (coarse) POS tag sets, and 97.2% when evaluated against the German POS tag set. For English, we trained our model on the 39832 sentences which constitute sections 2-21 of the PTB-WSJ and on the 500K sentences from the NYT section of the NANC newswire corpus (Graff, 1995). We report results on the WSJ part of our data, which includes 950028 words tokens in 44389 types. Of the tokens, 832629 (87.6%) 1302 English Fine k=13 Coarse k=13 Fine k=34 Prototype Clark Prototype Clark Prototype Clark Tagger µ σ % Tagger µ σ % Tagger µ σ % Many–to–1 61.0 55.1 1.6 100 70.0 66.9 2.1 94 71.6 69.8 1.5 90 55.5 48.8 1.8 100 66.1 62.6 2.3 94 67.5 65.5 1.7 90 1–to–1 60.0 52.2 1.9 100 58.1 49.4 2.9 100 63.5 54.5 1.6 100 54.9 46.0 2.2 100 53.7 43.8 3.3 100 58.8 48.5 1.8 100 NVI 0.652 0.773 0.027 100 0.841 0.972 0.036 100 0.663 0.725 0.018 100 0.795 0.943 0.033 100 1.052 1.221 0.046 100 0.809 0.885 0.022 100 V 0.636 0.581 0.015 100 0.590 0.543 0.018 100 0.677 0.659 0.008 100 0.542 0.478 0.019 100 0.484 0.429 0.023 100 0.608 0.588 0.010 98 German k=17 k=26 Prototype Clark Prototype Clark Tagger µ σ % Tagger µ σ % Many–to-1 64.6 64.7 1.2 41 68.2 67.8 1.0 60 58.9 59.1 1.4 40 63.2 62.8 1.2 60 1–to–1 53.7 52.0 1.8 77 56.0 52.0 2.1 99 48.0 46.0 2.3 78 50.7 45.9 2.6 99 NVI 0.667 0.675 0.019 66 0.640 0.682 0.019 100 0.819 0.829 0.025 66 0.785 0.839 0.025 100 V 0.646 0.645 0.010 50 0.675 0.657 0.008 100 0.552 0.553 0.013 48 0.596 0.574 0.010 100 Table 1: Top: English. Bottom: German. Results are reported for our model (Prototype Tagger), Clark’s average score (µ), Clark’s standard deviation (σ) and the fraction of Clark’s results that scored worse than our model (%). For the mapping based measures, results are accuracy percentage. For V ∈[0, 1], higher is better. For high quality output, NV I ∈[0, 1] as well, and lower is better. In each entry, the top number indicates the score when including punctuation and the bottom number the score when excluding it. In English, our results are always better than Clark’s. In German, they are almost always better. are not punctuation. The percentage of unknown words (those appearing less than five times) is 1.6%. There are 45 clusters in this annotation scheme, 34 of which are not punctuation. We ran each algorithm both with k=13 and k=34 (the number of desired clusters). We compare the output to two annotation schemes: the fine grained PTB WSJ scheme, and the coarse grained tags defined in (Smith and Eisner, 2005). The output of the k=13 run is evaluated both against the coarse POS tag annotation (the ‘Coarse k=13’ scenario) and against the full PTB-WSJ annotation scheme (the ‘Fine k=13’ scenario). The k=34 run is evaluated against the full PTB-WSJ annotation scheme (the ‘Fine k=34’ scenario). The POS cluster frequency distribution tends to be skewed: each of the 13 most frequent clusters in the PTB-WSJ cover more than 2.5% of the tokens (excluding punctuation) and together 86.3% of them. We therefore chose k=13, since it is both the number of coarse POS tags (excluding punctuation) as well as the number of frequent POS tags in the PTB-WSJ annotation scheme. We chose k=34 in order to evaluate against the full 34 tags PTB-WSJ annotation scheme (excluding punctuation) using the same number of clusters. For German, we trained our model on the 20296 sentences of the NEGRA corpus (Brants, 1997) and on the first 450K sentences of the DeWAC corpus (Baroni et al., 2009). DeWAC is a corpus extracted by web crawling and is therefore out of domain. We report results on the NEGRA part, which includes 346320 word tokens of 49402 types. Of the tokens, 289268 (83.5%) are not punctuation. The percentage of unknown words (those appearing less than five times) is 8.1%. There are 62 clusters in this annotation scheme, 51 of which are not punctuation. We ran the algorithms with k=17 and k=26. k=26 was chosen since it is the number of clusters that cover each more than 0.5% of the NEGRA tokens, and in total cover 96% of the (nonpunctuation) tokens. In order to test our algorithm in another scenario, we conducted experiments with k=17 as well, which covers 89.9% of the tokens. All outputs are compared against NEGRA’s gold standard scheme. We do not report results for k=51 (where the number of gold clusters is the same as the number of induced clusters), since our algorithm produced only 42 clusters in the landmark detection stage. We could of course have modified the parameters to allow our algorithm to produce 51 clusters. However, we wanted to use the exact same parameters as those used for the English experiments to minimize the issue of parameter tuning. In addition to the comparisons described above, we present results of experiments (in the ‘Fine 1303 B B+M B+C F(I=1) F M-to-1 53.3 54.8 58.2 57.3 61.0 1-to-1 50.2 51.7 55.1 54.8 60.0 NVI 0.782 0.720 0.710 0.742 0.652 V 0.569 0.598 0.615 0.597 0.636 Table 2: A comparison of partial versions of the model in the ‘Fine k=13’ WSJ scenario. M-to-1 and 1-to-1 results are reported in accuracy percentage. Lower NVI is better. B is the strictly distributional algorithm, B+M adds the morphological model, B+C adds capitalization to B, F(I=1) consists of all components, where only one iteration of coordinate clustering is performed, and F is the full model. M-to-1 1-to-1 V VI Prototype 71.6 63.5 0.677 2.00 Clark 69.8 54.5 0.659 2.18 HK – 41.3 – – J 43–62 37–47 – 4.23–5.74 GG – – – 2.8 GJ – 40–49.9 – 4.03–4.47 VG – – 0.54-0.59 2.5–2.9 GGTP-45 65.4 44.5 – – GGTP-17 70.2 49.5 – – Table 4: Comparison of our algorithms with the recent fully unsupervised POS taggers for which results are reported. The models differ in the annotation scheme, the corpus size and the number of induced clusters (k) that they used. HK: (Haghighi and Klein, 2006), 193K tokens, fine tags, k=45. GG: (Goldwater and Griffiths, 2007), 24K tokens, coarse tags, k=17. J : (Johnson, 2007), 1.17M tokens, fine tags, k=25–50. GJ: (Gao and Johnson, 2008), 1.17M tokens, fine tags, k=50. VG: (Van Gael et al., 2009), 1.17M tokens, fine tags, k=47–192. GGTP-45: (Grac¸a et al., 2009), 1.17M tokens, fine tags, k=45. GGTP-17: (Grac¸a et al., 2009), 1.17M tokens, coarse tags, k=17. Lower VI values indicate better clustering. VI is computed using e as the base of the logarithm. Our algorithm gives the best results. k=13’ scenario) that quantify the contribution of each component of the algorithm. We ran the base distributional algorithm, a variant which uses only capitalization information (i.e., has only one nonsingleton morphological class, that of words appearing capitalized in most of their instances) and a variant which uses no capitalization information, defining the morphological clusters according to the morphological representation alone. 7 Results Table 1 presents results for the English and German experiments. For English, our algorithm obtains better results than Clark’s in all measures and scenarios. It is without exception better than the average score of Clark’s and in most cases better than the maximal Clark score obtained in 100 runs. A significant difference between our algorithm and Clark’s is that the latter, like most algorithms which addressed the task, induces the clustering 0 5 10 15 20 25 30 35 40 45 0 0.2 0.4 0.6 0.8 1 Gold Standard Induced Figure 2: POS class frequency distribution for our model and the gold standard, in the ‘Fine k=34’ scenario. The distributions are similar. by maximizing a non-convex function. These functions have many local maxima and the specific solution to which algorithms that maximize them converge strongly depends on their (random) initialization. Therefore, their output’s quality often significantly diverges from the average. This issue is discussed in depth in (Reichart et al., 2010b). Our algorithm is deterministic3. For German, in the k=26 scenario our algorithm outperforms Clark’s, often outperforming even its maximum in 100 runs. In the k=17 scenario, our algorithm obtains a higher score than Clark with probability 0.4 to 0.78, depending on the measure and scenario. Clark’s average score is slightly better in the Many-to-1 measure, while our algorithm performs somewhat better than Clark’s average in the 1-to-1 and NVI measures. The DeWAC corpus from which we extracted statistics for the German experiments is out of domain with respect to NEGRA. The corresponding corpus in English, NANC, is a newswire corpus and therefore clearly in-domain with respect to WSJ. This is reflected by the percentage of unknown words, which was much higher in German than in English (8.1% and 1.6%), lowering results. Table 2 shows the effect of each of our algorithm’s components. Each component provides an improvement over the base distributional algorithm. The full coordinate clustering stage (several iterations, F) considerably improves the score over a single iteration (F(I=1)). Capitalization information increases the score more than the morphological information, which might stem from the granularity of the POS tag set with respect to names. This analysis is supported by similar experiments we made in the ‘Coarse k=13’ scenario (not shown in tables here). There, the decrease in performance was only of 1%–2% in the mapping 3The fluctuations inflicted on our algorithm by the random mapping of unknown words are of less than 0.1% . 1304 Excluding Punctuation Including Punctuation Perfect Punctuation M-to-1 1-to-1 NVI V M-to-1 1-to-1 NVI V M-to-1 1-to-1 NVI V Van Gael 59.1 48.4 0.999 0.530 62.3 51.3 0.861 0.591 64.0 54.6 0.820 0.610 Prototype 67.5 58.8 0.809 0.608 71.6 63.5 0.663 0.677 71.6 63.9 0.659 0.679 Table 3: Comparison between the iHMM: PY-fixed model (Van Gael et al., 2009) and ours with various punctuation assignment schemes. Left section: punctuation tokens are excluded. Middle section: punctuation tokens are included. Right section: perfect assignment of punctuation is assumed. based measures and 3.5% in the V measure. Finally, Table 4 presents reported results for all recent algorithms we are aware of that tackled the task of unsupervised POS induction from plain text. Results for our algorithm’s and Clark’s are reported for the ‘Fine, k=34’ scenario. The settings of the various experiments vary in terms of the exact annotation scheme used (coarse or fine grained) and the size of the test set. However, the score differences are sufficiently large to justify the claim that our algorithm is currently the best performing algorithm on the PTB-WSJ corpus for POS induction from plain text4. Since previous works provided results only for the scenario in which punctuation is included, the reported results are not directly comparable. In order to quantify the effect various punctuation schemes have on the results, we evaluated the ‘iHMM: PY-fixed’ model (Van Gael et al., 2009) and ours when punctuation is excluded, included or perfectly tagged5. The results (Table 3) indicate that most probably even after an appropriate correction for punctuation, our model remains the best performing one. 8 Discussion In this work we presented a novel unsupervised algorithm for POS induction from plain text. The algorithm first generates relatively accurate clusters of high frequency words, which are subsequently used to bootstrap the entire clustering. The distributional and morphological representations that we use are novel for this task. We experimented on two languages with mapping and information theoretic clustering evaluation measures. Our algorithm obtains the best reported results on the English PTB-WSJ corpus. In addition, our results are almost always better than Clark’s on the German NEGRA corpus. 4Grac¸a et al. (2009) report very good results for 17 tags in the M-1 measure. However, their 1-1 results are quite poor, and results for the common IT measures were not reported. Their results for 45 tags are considerably lower. 5We thank the authors for sending us their data. We have also performed a manual error analysis, which showed that our algorithm performs much better on closed classes than on open classes. In order to asses this quantitatively, let us define a random variable for each of the gold clusters, which receives a value corresponding to each induced cluster with probability proportional to their intersection size. For each gold cluster, we compute the entropy of this variable. In addition, we greedily map each induced cluster to a gold cluster and compute the ratio between their intersection size and the size of the gold cluster (mapping accuracy). We experimented in the ‘Fine k=34’ scenario. The clusters that obtained the best scores were (brackets indicate mapping accuracy and entropy for each of these clusters) coordinating conjunctions (95%, 0.32), prepositions (94%, 0.32), determiners (94%, 0.44) and modals (93%, 0.45). These are all closed classes. The classes on which our algorithm performed worst consist of open classes, mostly verb types: past tense verbs (47%, 2.2), past participle verbs (44%, 2.32) and the morphologically unmarked non-3rd person singular present verbs (32%, 2.86). Another class with low performance is the proper nouns (37%, 2.9). The errors there are mostly of three types: confusions between common and proper nouns (sometimes due to ambiguity), unknown words which were put in the unknown words cluster, and abbreviations which were given a separate class by our algorithm. Finally, the algorithm’s performance on the heterogeneous adverbs class (19%, 3.73) is the lowest. Clark’s algorithm exhibits6 a similar pattern with respect to open and closed classes. While his algorithm performs considerably better on adverbs (15% mapping accuracy difference and 0.71 entropy difference), our algorithm scores considerably better on prepositions (17%, 0.77), superlative adjectives (38%, 1.37) and plural proper names (45%, 1.26). 6Using average mapping accuracy and entropy over the 100 runs. 1305 Naturally, this analysis might reflect the arbitrary nature of a manually design POS tag set rather than deficiencies in automatic POS induction algorithms. In future work we intend to analyze the output of such algorithms in order to improve POS tag sets. Our algorithm and Clark’s are monosemous (i.e., they assign each word exactly one tag), while most other algorithms are polysemous. In order to assess the performance loss caused by the monosemous nature of our algorithm, we took the M-1 greedy mapping computed for the entire dataset and used it to compute accuracy over the monosemous and polysemous words separately. Results are reported for the English ‘Fine k=34’ scenario (without punctuation). We define a word to be monosemous if more than 95% of its tokens are assigned the same gold standard tag. For English, there are approximately 255K polysemous tokens and 578K monosemous ones. As expected, our algorithm is much more accurate on the monosemous tokens, achieving 76.6% accuracy, compared to 47.1% on the polysemous tokens. The evaluation in this paper is done at the token level. Type level evaluation, reflecting the algorithm’s ability to detect the set of possible POS tags for each word type, is important as well. It could be expected that a monosemous algorithm such as ours would perform poorly in a type level evaluation. In (Reichart et al., 2010a) we discuss type level evaluation at depth and propose type level evaluation measures applicable to the POS induction problem. In that paper we compare the performance of our Prototype Tagger with leading unsupervised POS tagging algorithms (Clark, 2003; Goldwater and Griffiths, 2007; Gao and Johnson, 2008; Van Gael et al., 2009). Our algorithm obtained the best results in 4 of the 6 measures in a margin of 4–6%, and was second best in the other two measures. Our results were better than Clark’s (the only other monosemous algorithm evaluated there) on all measures in a margin of 5–21%. The fact that our monosemous algorithm was better than good polysemous algorithms in a type level evaluation can be explained by the prototypical nature of the POS phenomenon (a longer discussion is given in (Reichart et al., 2010a)). However, the quality upper bound for monosemous algorithms is obviously much lower than that for polysemous algorithms, and we expect polysemous algorithms to outperform monosemous algorithms in the future in both type level and token level evaluations. The skewed (Zipfian) distribution of POS class frequencies in corpora is a problem for many POS induction algorithms, which by default tend to induce a clustering having a balanced distribution. Explicit modifications to these algorithms were introduced in order to bias their model to produce such a distribution (see (Clark, 2003; Johnson, 2007; Reichart et al., 2010b)). An appealing property of our model is its ability to induce a skewed distribution without being explicitly tuned to do so, as seen in Figure 2. Acknowledgements. We would like to thank Yoav Seginer for his help with his parser. References Michele Banko and Robert C. Moore, 2004. Part of Speech Tagging in Context. COLING ’04. Marco Baroni, Silvia Bernardini, Adriano Ferraresi and Eros Zanchetta, 2009. The WaCky Wide Web: A Collection of Very Large Linguistically Processed Web-Crawled Corpora. Language Resources and Evaluation. Chris Biemann, 2006. Unsupervised Part-ofSpeech Tagging Employing Efficient Graph Clustering. COLING-ACL ’06 Student Research Workshop. Thorsten Brants, 1997. The NEGRA Export Format. CLAUS Report, Saarland University. Peter F. Brown, Vincent J. Della Pietra, Peter V. de Souze, Jenifer C. Lai and Robert Mercer, 1992. Class-Based N-Gram Models of Natural Language. Computational Linguistics, 18(4):467–479. Alexander Clark, 2003. Combining Distributional and Morphological Information for Part of Speech Induction. EACL ’03. Mathias Creutz and Krista Lagus, 2005. Inducing the Morphological Lexicon of a Natural Language from Unannotated Text. AKRR ’05. Sajib Dasgupta and Vincent Ng, 2007. Unsupervised Part-of-Speech Acquisition for ResourceScarce Languages. EMNLP-CoNLL ’07. Dayne Freitag, 2004. Toward Unsupervised WholeCorpus Tagging. COLING ’04. Jianfeng Gao and Mark Johnson, 2008. A Comparison of Bayesian Estimators for Unsupervised Hidden Markov Model POS Taggers. EMNLP ’08. Yoav Goldberg, Meni Adler and Michael Elhadad, 2008. EM Can Find Pretty Good HMM POSTaggers (When Given a Good Start). ACL ’08. 1306 John Goldsmith, 2001. Unsupervised Learning of the Morphology of a Natural Language. Computational Linguistics, 27(2):153–198. Sharon Goldwater and Tom Griffiths, 2007. Fully Bayesian Approach to Unsupervised Part-of-Speech Tagging. ACL ’07. Jo˜ao Grac¸a, Kuzman Ganchev, Ben Taskar and Frenando Pereira, 2009. Posterior vs. Parameter Sparsity in Latent Variable Models. NIPS ’09. David Graff, 1995. North American News Text Corpus. Linguistic Data Consortium. LDC95T21. Aria Haghighi and Dan Klein, 2006. Prototype-driven Learning for Sequence Labeling. HLT–NAACL ’06. Anil K. Jain, Narasimha M. Murty and Patrick J. Flynn, 1999. Data Clustering: A Review. ACM Computing Surveys 31(3):264–323. Wenbin Jiang, Liang Huang and Qun Liu, 2009. Automatic Adaptation of Annotation Standards: Chinese Word Segmentation and POS Tagging – A Case Study. ACL ’09. Mark Johnson, 2007. Why Doesnt EM Find Good HMM POS-Taggers? EMNLP-CoNLL ’07. Harold W. Kuhn, 1955. The Hungarian method for the Assignment Problem. Naval Research Logistics Quarterly, 2:83-97. Marina Meila, 2007. Comparing Clustering – an Information Based Distance. Journal of Multivariate Analysis, 98:873–895. Bernard Merialdo, 1994. Tagging English Text with a Probabilistic Model. Computational Linguistics, 20(2):155–172. James Munkres, 1957. Algorithms for the Assignment and Transportation Problems. Journal of the SIAM, 5(1):32–38. Sujith Ravi and Kevin Knight, 2009. Minimized Models for Unsupervised Part-of-Speech Tagging. ACL ’09. Roi Reichart and Ari Rappoport, 2008. Unsupervised Induction of Labeled Parse Trees by Clustering with Syntactic Features. COLING ’08. Roi Reichart and Ari Rappoport, 2009. The NVI Clustering Evaluation Measure. CoNLL ’09. Roi Reichart, Omri Abend and Ari Rappoport, 2010a. Type Level Clustering Evaluation: New Measures and a POS Induction Case Study. CoNLL ’10. Roi Reichart, Raanan Fattal and Ari Rappoport, 2010b. Improved Unsupervised POS Induction Using Intrinsic Clustering Quality and a Zipfian Constraint. CoNLL ’10. Andrew Rosenberg and Julia Hirschberg, 2007. VMeasure: A Conditional Entropy-Based External Cluster Evaluation Measure. EMNLP ’07. Hinrich Sch¨utze, 1995. Distributional part-of-speech tagging. EACL ’95. Yoav Seginer, 2007. Fast Unsupervised Incremental Parsing. ACL ’07. Noah A. Smith and Jason Eisner, 2005. Contrastive Estimation: Training Log-Linear Models on Unlabeled Data. ACL ’05. John R. Taylor, 2003. Linguistic Categorization: Prototypes in Linguistic Theory, Third Edition. Oxford University Press. Jurgen Van Gael, Andreas Vlachos and Zoubin Ghahramani, 2009. The Infinite HMM for Unsupervised POS Tagging. EMNLP ’09. Qin Iris Wang and Dale Schuurmans, 2005. Improved Estimation for Unsupervised Part-of-Speech Tagging. IEEE NLP–KE ’05. Qiuye Zhao and Mitch Marcus, 2009. A Simple Unsupervised Learner for POS Disambiguation Rules Given Only a Minimal Lexicon. EMNLP ’09. 1307
|
2010
|
132
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1308–1317, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Extraction and Approximation of Numerical Attributes from the Web Dmitry Davidov ICNC The Hebrew University Jerusalem, Israel [email protected] Ari Rappoport Institute of Computer Science The Hebrew University Jerusalem, Israel [email protected] Abstract We present a novel framework for automated extraction and approximation of numerical object attributes such as height and weight from the Web. Given an object-attribute pair, we discover and analyze attribute information for a set of comparable objects in order to infer the desired value. This allows us to approximate the desired numerical values even when no exact values can be found in the text. Our framework makes use of relation defining patterns and WordNet similarity information. First, we obtain from the Web and WordNet a list of terms similar to the given object. Then we retrieve attribute values for each term in this list, and information that allows us to compare different objects in the list and to infer the attribute value range. Finally, we combine the retrieved data for all terms from the list to select or approximate the requested value. We evaluate our method using automated question answering, WordNet enrichment, and comparison with answers given in Wikipedia and by leading search engines. In all of these, our framework provides a significant improvement. 1 Introduction Information on various numerical properties of physical objects, such as length, width and weight is fundamental in question answering frameworks and for answering search engine queries. While in some cases manual annotation of objects with numerical properties is possible, it is a hard and labor intensive task, and is impractical for dealing with the vast amount of objects of interest. Hence, there is a need for automated semantic acquisition algorithms targeting such properties. In addition to answering direct questions, the ability to make a crude comparison or estimation of object attributes is important as well. For example, it allows to disambiguate relationships between objects such as X part-of Y or X inside Y. Thus, a coarse approximation of the height of a house and a window is sufficient to decide that in the ‘house window’ nominal compound, ‘window’ is very likely to be a part of house and not vice versa. Such relationship information can, in turn, help summarization, machine translation or textual entailment tasks. Due to the importance of relationship and attribute acquisition in NLP, numerous methods were proposed for extraction of various lexical relationships and attributes from text. Some of these methods can be successfully used for extracting numerical attributes. However, numerical attribute extraction is substantially different in two aspects, verification and approximation. First, unlike most general lexical attributes, numerical attribute values are comparable. It usually makes no sense to compare the names of two actors, but it is meaningful to compare their ages. The ability to compare values of different objects allows to improve attribute extraction precision by verifying consistency with attributes of other similar objects. For example, suppose that for Toyota Corolla width we found two different values, 1.695m and 27cm. The second value can be either an extraction error or a length of a toy car. Extracting and looking at width values for different car brands and for ‘cars’ in general we find: • Boundaries: Maximal car width is 2.195m, minimal is 88cm. • Average: Estimated avg. car width is 1.7m. • Direct/indirect comparisons: Toyota Corolla is wider than Toyota Corona. • Distribution: Car width is distributed normally around the average. 1308 Usage of all this knowledge allows us to select the correct value of 1.695m and reject other values. Thus we can increase the precision of value extraction by finding and analyzing an entire group of comparable objects. Second, while it is usually meaningless and impossible to approximate general lexical attribute values like an actor’s name, numerical attributes can be estimated even if they are not explicitly mentioned in the text. In general, attribute extraction frameworks usually attempt to discover a single correct value (e.g., capital city of a country) or a set of distinct correct values (e.g., actors of a movie). So there is essentially nothing to do when there is no explicit information present in the text for a given object and an attribute. In contrast, in numerical attribute extraction it is possible to provide an approximation even when no explicit information is present in the text, by using values of comparable objects for which information is provided. In this paper we present a pattern-based framework that takes advantage of the properties of similar objects to improve extraction precision and allow approximation of requested numerical object properties. Our framework comprises three main stages. First, given an object name we utilize WordNet and pattern-based extraction to find a list of similar objects and their category labels. Second, we utilize a predefined set of lexical patterns in order to extract attribute values of these objects and available comparison/boundary information. Finally, we analyze the obtained information and select or approximate the attribute value for the given (object, attribute) pair. We performed a thorough evaluation using three different applications: Question Answering (QA), WordNet (WN) enrichment, and comparison with Wikipedia and answers provided by leading search engines. QA evaluation was based on a designed dataset of 1250 questions on size, height, width, weight, and depth, for which we created a gold standard and compared against it automatically1. For WN enrichment evaluation, our framework discovered size and weight values for 300 WN physical objects, and the quality of results was evaluated by human judges. For interactive search, we compared our results to information obtained through Wikipedia, Google and Wolfram Alpha. 1This dataset is available in the authors’ websites for the research community. Utilization of information about comparable objects provided a significant boost to numerical attribute extraction quality, and allowed a meaningful approximation of missing attribute values. Section 2 discusses related work, Section 3 details the algorithmic framework, Section 4 describes the experimental setup, and Section 5 presents our results. 2 Related work Numerous methods have been developed for extraction of diverse semantic relationships from text. While several studies propose relationship identification methods using distributional analysis of feature vectors (Turney, 2005), the majority of the proposed open-domain relations extraction frameworks utilize lexical patterns connecting a pair of related terms. (Hearst, 1992) manually designed lexico-syntactic patterns for extracting hypernymy relations. (Berland and Charniak, 1999; Girju et al, 2006) proposed a set of patterns for meronymy relations. Davidov and Rappoport (2008a) used pattern clusters to disambiguate nominal compound relations. Extensive frameworks were proposed for iterative discovery of any pre-specified (e.g., (Riloff and Jones, 1999; Chklovski and Pantel, 2004)) and unspecified (e.g., (Banko et al., 2007; Rosenfeld and Feldman, 2007; Davidov and Rappoport, 2008b)) relation types. The majority of the above methods utilize the following basic strategy. Given (or discovering automatically) a set of patterns or relationshiprepresenting term pairs, these methods mine the web for these patterns and pairs, iteratively obtaining more instances. The proposed strategies generally include some weighting/frequency/contextbased algorithms (e.g. (Pantel and Pennacchiotti, 2006)) to reduce noise. Some of the methods are suitable for retrieval of numerical attributes. However, most of them do not exploit the numerical nature of the attribute data. Our research is related to a sub-domain of question answering (Prager, 2006), since one of the applications of our framework is answering questions on numerical values. The majority of the proposed QA frameworks rely on pattern-based relationship acquisition (Ravichandran and Hovy, 2009). However, most QA studies focus on different types of problems than our paper, including question classification, paraphrasing, etc. 1309 Several recent studies directly target the acquisition of numerical attributes from the Web and attempt to deal with ambiguity and noise of the retrieved attribute values. (Aramaki et al., 2007) utilize a small set of patterns to extract physical object sizes and use the averages of the obtained values for a noun compound classification task. (Banerjee et al, 2009) developed a method for dealing with quantity consensus queries (QCQs) where there is uncertainty about the answer quantity (e.g. “driving time from Paris to Nice”). They utilize a textual snippet feature and snippet quantity in order to select and rank intervals of the requested values. This approach is particularly useful when it is possible to obtain a substantial amount of a desired attribute values for the requested query. (Moriceau, 2006) proposed a rulebased system which analyzes the variation of the extracted numerical attribute values using information in the textual context of these values. A significant body of recent research deals with extraction of various data from web tables and lists (e.g., (Cafarella et al., 2008; Crestan and Pantel, 2010)). While in the current research we do not utilize this type of information, incorporation of the numerical data extracted from semistructured web pages can be extremely beneficial for our framework. All of the above numerical attribute extraction systems utilize only direct information available in the discovered object-attribute co-occurrences and their contexts. However, as we show, indirect information available for comparable objects can contribute significantly to the selection of the obtained values. Using such indirect information is particularly important when only a modest amount of values can be obtained for the desired object. Also, since the above studies utilize only explicitly available information they were unable to approximate object values in cases where no explicit information was found. 3 The Attribute Mining Framework Our algorithm is given an object and an attribute. In the WN enrichment scenario, it is also given the object’s synset. The algorithm comprises three main stages: (1) mining for similar objects and determination of a class label; (2) mining for attribute values and comparison statements; (3) processing the results. 3.1 Similar objects and class label To verify and estimate attribute values for the given object we utilize similar objects (cohyponyms) and the object’s class label (hypernym). In the WN enrichment scenario we can easily obtain these, since we get the object’s synset as input. However, in Question Answering (QA) scenarios we do not have such information. To obtain it we employ a strategy which uses WordNet along with pattern-based web mining. Our web mining part follows common patternbased retrieval practice (Davidov et al., 2007). We utilize Yahoo! Boss API to perform search engine queries. For an object name Obj we query the Web using a small set of pre-defined co-hyponymy patterns like “as * and/or [Obj]”2. In the WN enrichment scenario, we can add the WN class label to each query in order to restrict results to the desired word sense. In the QA scenario, if we are given the full question and not just the (object, attribute) pair we can add terms appearing in the question and having a strong PMI with the object (this can be estimated using any fixed corpus). However, this is not essential. We then extract new terms from the retrieved web snippets and use these terms iteratively to retrieve more terms from the Web. For example, when searching for an object ‘Toyota’, we execute a search engine query [ “as * and Toyota”] and we might retrieve a text snippet containing “...as Honda and Toyota ...”. We then extract from this snippet the additional word ‘Honda’ and use it for iterative retrieval of additional similar terms. We attempt to avoid runaway feedback loop by requiring each newly detected term to co-appear with the original term in at least a single co-hyponymy pattern. WN class labels are used later for the retrieval of boundary values, and here for expansion of the similar object set. In the WN enrichment scenario, we already have the class label of the object. In the QA scenario, we automatically find class labels as follows. We compute for each WN subtree a coverage value, the number of retrieved terms found in the subtree divided by the number of subtree terms, and select the subtree having the highest coverage. In all scenarios, we add all terms found in this subtree to the retrieved term list. If no WN subtree with significant (> 0.1) coverage is found, 2“*” means a search engine wildcard. Square brackets indicate filled slots and are not part of the query. 1310 we retrieve a set of category labels from the Web using hypernymy detection patterns like “* such as [Obj]” (Hearst, 1992). If several label candidates were found, we select the most frequent. Note that we perform this stage only once for each object and do not need to repeat it for different attribute types. 3.2 Querying for values, bounds and comparison data Now we would like to extract the attribute values for the given object and its similar objects. We will also extract bounds and comparison information in order to verify the extracted values and to approximate the missing ones. To allow us to extract attribute-specific information, we provided the system with a seed set of extraction patterns for each attribute type. There are three kinds of patterns: value extraction, bounds and comparison patterns. We used up to 10 patterns of each kind. These patterns are the only attribute-specific resource in our framework. Value extraction. The first pattern group, Pvalues, allows extraction of the attribute values from the Web. All seed patterns of this group contain a measurement unit name, attribute name, and some additional anchoring words, e.g., ‘Obj is * [height unit] tall’ or ‘Obj width is * [width unit]’. As in Section 3.1, we execute search engine queries and collect a set of numerical values for each pattern. We extend this group iteratively from the given seed as commonly done in pattern-based acquisition methods. To do this we re-query the Web with the obtained (object, attribute value, attribute name) triplets (e.g., ‘[Toyota width 1.695m]’). We then extract new patterns from the retrieved search engine snippets and re-query the Web with the new patterns to obtain more attribute values. We provided the framework with unit names and with an appropriate conversion table which allows to convert between different measurement systems and scales. The provided names include common abbreviations like cm/centimeter. All value acquisition patterns include unit names, so we know the units of each extracted value. At the end of the value extraction stage, we convert all values to a single unit format for comparison. Boundary extraction. The second group, Pboundary, consists of boundary-detection patterns like ‘the widest [label] is * [width unit]’. These patterns incorporate the class labels discovered in the previous stage. They allow us to find maximal and minimal values for the object category defined by labels. If we get several lower bounds and several upper bounds, we select the highest upper bound and the lowest lower bound. Extraction of comparison information. The third group, Pcompare, consists of comparison patterns. They allow to compare objects directly even when no attribute values are mentioned. This group includes attribute equality patterns such as ‘[Object1] has the same width as [Object2]’, and attribute inequality ones such as ‘[Object1] is wider than [Object2]’. We execute search queries for each of these patterns, and extract a set of ordered term pairs, keeping track of the relationships encoded by the pairs. We use these pairs to build a directed graph (Widdows and Dorow, 2002; Davidov and Rappoport, 2006) in which nodes are objects (not necessarily with assigned values) and edges correspond to extracted co-appearances of objects inside the comparison patterns. The directions of edges are determined by the comparison sign. If two objects co-appear inside an equality pattern we put a bidirectional edge between them. 3.3 Processing the collected data As a result of the information collection stage, for each object and attribute type we get: • A set of attribute values for the requested object. • A set of objects similar or comparable to the requested object, some of them annotated with one or many attribute values. • Upper and lowed bounds on attribute values for the given object category. • A comparison graph connecting some of the retrieved objects by comparison edges. Obviously, some of these components may be missing or noisy. Now we combine these information sources to select a single attribute value for the requested object or to approximate this value. First we apply bounds, removing out-of-range values, then we use comparisons to remove inconsistent comparisons. Finally we examine the remaining values and the comparison graph. Processing bounds. First we verify that indeed most (≥50%) of the retrieved values fit the retrieved bounds. If the lower and/or upper bound 1311 contradicts more than half of the data, we reject the bound. Otherwise we remove all values which do not satisfy one or both of the accepted bounds. If no bounds are found or if we disable the bound retrieval (see Section 4.1), we assign the maximal and minimal observed values as bounds. Since our goal is to obtain a value for the single requested object, if at the end of this stage we remain with a single value, no further processing is needed. However, if we obtain a set of values or no values at all, we have to utilize comparison data to select one of the retrieved values or to approximate the value in case we do not have an exact answer. Processing comparisons. First we simplify the comparison graph. We drop all graph components that are not connected (when viewing the graph as undirected) to the desired object. Now we refine the graph. Note that each graph node may have a single value, many assigned values, or no assigned values. We define assigned nodes as nodes that have at least one value. For each directed edge E(A →B), if both A and B are assigned nodes, we check if Avg(A) ≤ Avg(B)3. If the average values violate the equation, we gradually remove up to half of the highest values for A and up to half of the lowest values for B till the equation is satisfied. If this cannot be done, we drop the edge. We repeat this process until every edge that connects two assigned nodes satisfies the inequality. Selecting an exact attribute value. The goal now is to select an attribute value for the given object. During the first stage it is possible that we directly extract from the text a set of values for the requested object. The bounds processing step rejects some of these values, and the comparisons step may reject some more. If we still have several values remaining, we choose the most frequent value based on the number of web snippets retrieved during the value acquisition stage. If there are several values with the same frequency we select the median of these values. Approximating the attribute value. In the case when we do not have any values remaining after the bounds processing step, the object node will remain unassigned after construction of the comparison graph, and we would like to estimate its value. Here we present an algorithm which allows 3Avg. is of values of an object, without similar objects. us to set the values of all unassigned nodes, including the node of the requested object. In the algorithm below we treat all node groups connected by bidirectional (equality) edges as a same-value group, i.e., if a value is assigned to one node in the group, the same value is immediately assigned to the rest of the nodes in the same group. We start with some preprocessing. We create dummy lower and upper bound nodes L and U with corresponding upper/lower bound values obtained during the previous stage. These dummy nodes will be used when we encounter a graph which ends with one or more nodes with no available numerical information. We then connect them to the graph as follows: (1) if A has no incoming edges, we add an edge L →A; (2) if A has no outgoing edges, we add an edge A →U. We define a legal unassigned path as a directed path A0 →A1 →. . . →An →An+1 where A0 and An+1 are assigned satisfying Avg(A0) ≤ Avg(An+1) and A1 . . . An are unassigned. We would like to use dummy bound nodes only in cases when no other information is available. Hence we consider paths L →. . . →U connecting both bounds are illegal. First we assign values for all unassigned nodes that belong to a single legal unassigned path, using a simple linear combination: V al(Ai)i∈(1...n) = n + 1 −i n + 1 Avg(A0) + i n + 1Avg(An+1) Then, for all unassigned nodes that belong to multiple legal unassigned paths, we compute node value as above for each path separately and assign to the node the average of the computed values. Finally we assign the average of all extracted values within bounds to all the remaining unassigned nodes. Note that if we have no comparison information and no value information for the requested object, the requested object will receive the average of the extracted values of the whole set of the retrieved comparable objects and the comparison step will be essentially empty. 4 Experimental Setup We performed automated question answering (QA) evaluation, human-based WN enrichment evaluation, and human-based comparison of our results to data available through Wikipedia and to the top results of leading search engines. 1312 4.1 Experimental conditions In order to test the main system components, we ran our framework under five different conditions: • FULL: All system components were used. • DIRECT: Only direct pattern-based acquisition of attribute values (Section 3.2, value extraction) for the given object was used, as done in most general-purpose attribute acquisition systems. If several values were extracted, the most common value was used as an answer. • NOCB: No boundary and no comparison data were collected and processed (Pcompare and Pbounds were empty). We only collected and processed a set of values for the similar objects. • NOB: As in FULL but no boundary data was collected and processed (Pbounds was empty). • NOC: As in FULL but no comparison data was collected and processed (Pcompare was empty). 4.2 Automated QA Evaluation We created two QA datasets, Web and TREC based. Web-based QA dataset. We created QA datasets for size, height, width, weight, and depth attributes. For each attribute we extracted from the Web 250 questions in the following way. First, we collected several thousand questions, querying for the following patterns: “How long/tall/wide/heavy/deep/high is”,“What is the size/width/height/depth/weight of”. Then we manually filtered out non-questions and heavily context-specific questions, e.g., “what is the width of the triangle”. Next, we retained only a single question for each entity by removing duplicates. For each of the extracted questions we manually assigned a gold standard answer using trusted resources including books and reliable Web data. For some questions, the exact answer is the only possible one (e.g., the height of a person), while for others it is only the center of a distribution (e.g., the weight of a coffee cup). Questions with no trusted and exact answers were eliminated. From the remaining questions we randomly selected 250 questions for each attribute. TREC-based QA dataset. As a small complementary dataset we used relevant questions from the TREC Question Answering Track 1999-2007. From 4355 questions found in this set we collected 55 (17 size, 2 weight, 3 width, 3 depth and 30 height) questions. Examples. Some example questions from our datasets are (correct answers are in parentheses): How tall is Michelle Obama? (180cm); How tall is the tallest penguin? (122cm); What is the height of a tennis net? (92cm); What is the depth of the Nile river? (1000cm = 10 meters); How heavy is a cup of coffee? (360gr); How heavy is a giraffe? (1360000gr = 1360kg); What is the width of a DNA molecule? (2e-7cm); What is the width of a cow? (65cm). Evaluation protocol. Evaluation against the datasets was done automatically. For each question and each condition our framework returned a numerical value marked as either an exact answer or as an approximation. In cases where no data was found for an approximation (no similar objects with values were found), our framework returned no answer. We computed precision4, comparing results to the gold standard. Approximate answers are considered to be correct if the approximation is within 10% of the gold standard value. While a choice of 10% may be too strict for some applications and too generous for others, it still allows to estimate the quality of our framework. 4.3 WN enrichment evaluation We manually selected 300 WN entities from about 1000 randomly selected objects below the object tree in WN, by filtering out entities that clearly do not possess any of the addressed numerical attributes. Evaluation was done using human subjects. It is difficult to do an automated evaluation, since the nature of the data is different from that of the QA dataset. Most of the questions asked over the Web target named entities like specific car brands, places and actors. There is usually little or no variability in attribute values of such objects, and the major source of extraction errors is name ambiguity of the requested objects. WordNet physical objects, in contrast, are much less specific and their attributes such as size and 4Due to the nature of the task recall/f-score measures are redundant here 1313 weight rarely have a single correct value, but usually possess an acceptable numerical range. For example, the majority of the selected objects like ‘apple’ are too general to assign an exact size. Also, it is unclear how to define acceptable values and an approximation range. Crudeness of desired approximation depends both on potential applications and on object type. Some objects show much greater variability in size (and hence a greater range of acceptable approximations) than others. This property of the dataset makes it difficult to provide a meaningful gold standard for the evaluation. Hence in order to estimate the quality of our results we turn to an evaluation based on human judges. In this evaluation we use only approximate retrieved values, keeping out the small amount of returned exact values5. We have mixed (Object, Attribute name, Attribute value) triplets obtained through each of the conditions, and asked human subjects to assign these to one of the following categories: • The attribute value is reasonable for the given object. • The value is a very crude approximation of the given object attribute. • The value is incorrect or clearly misleading. • The object is not familiar enough to me so I cannot answer the question. Each evaluator was provided with a random sample of 40 triplets. In addition we mixed in 5 manually created clearly correct triplets and 5 clearly incorrect ones. We used five subjects, and the agreement (inter-annotator Kappa) on shared evaluated triplets was 0.72. 4.4 Comparisons to search engine output Recently there has been a significant improvement both in the quality of search engine results and in the creation of manual well-organized and annotated databases such as Wikipedia. Google and Yahoo! queries frequently provide attribute values in the top snippets or in search result web pages. Many Wikipedia articles include infoboxes with well-organized attribute values. Recently, the Wolfram Alpha computational knowledge engine presented the computation of attribute values from a given query text. 5So our results are in fact higher than shown. Hence it is important to test how well our framework can complement the manual extraction of attributes from resources such as Wikipedia and top Google snippets. In order to test this, we randomly selected 100 object-attribute pairs from our Web QA and WordNet datasets and used human subjects to test the following: 1. Go1: Querying Google for [object-name attribute-name] gives in some of the first three snippets a correct value or a good approximation value6 for this pair. 2. Go2: Querying Google for [object-name attribute-name] and following the first three links gives a correct value or a good approximation value. 3. Wi: There is a Wikipedia page for the given object and it contains an appropriate attribute value or an approximation in an infobox. 4. Wf: A Wolfram Alpha query for [objectname attribute-name] retrieves a correct value or a good approximation value 5 Results 5.1 QA results We applied our framework to the above QA datasets. Table 1 shows the precision and the percentage of approximations and exact answers. Looking at %Exact+%Approx, we can see that for all datasets only 1-9% of the questions remain unanswered, while correct exact answers are found for 65%/87% of the questions for Web/TREC (% Exact and Prec(Exact) in the table). Thus approximation allows us to answer 1324% of the requested values which are either simply missing from the retrieved text or cannot be detected using the current pattern-based framework. Comparing performance of FULL to DIRECT, we see that our framework not only allows an approximation when no exact answer can be found, but also significantly increases the precision of exact answers using the comparison and the boundary information. It is also apparent that both boundary and comparison features are needed to achieve good performance and that using both of them achieves substantially better results than each of them separately. 6As defined in the human subject questionnaire. 1314 FULL DIRECT NOCB NOB NOC Web QA Size %Exact 80 82 82 82 80 Prec(Exact) 76 40 40 54 65 %Approx 16 14 14 16 Prec(Appr) 64 34 53 46 Height %Exact 79 84 84 84 79 Prec(Exact) 86 56 56 69 70 %Approx 16 11 11 16 Prec(Appr) 72 25 65 53 Width %Exact 74 76 76 76 74 Prec(Exact) 86 45 45 60 72 %Approx 17 15 15 17 Prec(Appr) 75 26 63 55 Weight %Exact 71 73 73 73 71 Prec(Exact) 82 57 57 64 70 Prec(Appr) 24 22 22 24 %Approx 61 39 51 46 Depth %Exact 82 82 82 82 82 Prec(Exact) 89 60 60 71 78 %Approx 19 19 19 19 Prec(Appr) 92 58 76 63 Total average %Exact 77 79 79 79 77 Prec(Exact) 84 52 52 64 71 %Approx 18 16 16 19 Prec(Appr) 72 36 62 53 TREC QA %Exact 87 90 90 90 87 Prec(Exact) 100 62 62 84 76 %Approx 13 9 9 13 Prec(Appr) 57 20 40 57 Table 1: Precision and amount of exact and approximate answers for QA datasets. Comparing results for different question types we can see substantial performance differences between the attribute types. Thus depth shows much better overall results than width. This is likely due to a lesser difficulty of depth questions or to a more exact nature of available depth information compared to width or size. 5.2 WN enrichment As shown in Table 2, for the majority of examined WN objects, the algorithm returned an approximate value, and only for 13-15% of the objects (vs. 70-80% in QA data) the algorithm could retrieve exact answers. Note that the common pattern-based acquisition framework, presented as the DIRECT condition, could only extract attribute values for 15% of the objects since it does not allow approximations and FULL DIRECT NOCB NOB NOC Size %Exact 15.3 18.0 18.0 18.0 15.3 %Approx 80.3 38.2 20.0 23.6 Weight %Exact 11.8 12.5 12.5 12.5 11.8 %Approx 71.7 38.2 20.0 23.6 Table 2: Percentage of exact and approximate values for the WordNet enrichment dataset. FULL NOCB NOB NOC Size %Correct 73 21 49 28 %Crude 15 54 31 49 %Incorrect 8 21 16 19 Weight %Correct 64 24 46 38 %Crude 24 45 30 41 %Incorrect 6 25 18 15 Table 3: Human evaluation of approximations for the WN enrichment dataset (the percentages are averaged over the human subjects). may only extract values from the text where they explicitly appear. Table 3 shows human evaluation results. We see that the majority of approximate values were clearly accepted by human subjects, and only 68% were found to be incorrect. We also observe that both boundary and comparison data significantly improve the approximation results. Note that DIRECT is missing from this table since no approximations are possible in this condition. Some examples for WN objects and approximate values discovered by the algorithm are: Sandfish, 15gr; skull, 1100gr; pilot, 80.25kg. The latter value is amusing due to the high variability of the value. However, even this value is valuable, as a sanity check measure for automated inference systems and for various NLP tasks (e.g., ‘pilot jacket’ likely refers to a jacket used by pilots and not vice versa). 5.3 Comparison with search engines and Wikipedia Table 4 shows results for the above datasets in comparison to the proportion of correct results and the approximations returned by our framework under the FULL condition (correct exact values and approximations are taken together). We can see that our framework, due to its approximation capability, currently shows significantly greater coverage than manual extraction of data from Wikipedia infoboxes or from the first 1315 FULL Go1 Go2 Wi Wf Web QA 83 32 40 15 21 WordNet 87 24 27 18 5 Table 4: Comparison of our attribute extraction framework to manual extraction using Wikipedia and search engines. search engine results. 6 Conclusion We presented a novel framework which allows an automated extraction and approximation of numerical attributes from the Web, even when no explicit attribute values can be found in the text for the given object. Our framework retrieves similarity, boundary and comparison information for objects similar to the desired object, and combines this information to approximate the desired attribute. While in this study we explored only several specific numerical attributes like size and weight, our framework can be easily augmented to work with any other consistent and comparable attribute type. The only change required for incorporation of a new attribute type is the development of attribute-specific Pboundary, Pvalues, and Pcompare pattern groups; the rest of the system remains unchanged. In our evaluation we showed that our framework achieves good results and significantly outperforms the baseline commonly used for general lexical attribute retrieval7. While there is a growing justification to rely on extensive manually created resources such as Wikipedia, we have shown that in our case automated numerical attribute acquisition could be a preferable option and provides excellent coverage in comparison to handcrafted resources or manual examination of the leading search engine results. Hence a promising direction would be to use our approach in combination with Wikipedia data and with additional manually created attribute rich sources such as Web tables, to achieve the best possible performance and coverage. We would also like to explore the incorporation of approximate discovered numerical attribute data into existing NLP tasks such as noun compound classification and textual entailment. 7It should be noted, however, that in our DIRECT baseline we used a basic pattern-based retrieval strategy; more sophisticated strategies for value selection might bring better results. References Eiji Aramaki, Takeshi Imai, Kengo Miyo and Kazuhiko Ohe. 2007 UTH: SVM-based Semantic Relation Classification using Physical Sizes. Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007). Somnath Banerjee, Soumen Chakrabarti and Ganesh Ramakrishnan. 2009. Learning to Rank for Quantity Consensus Queries. SIGIR ’09. Michele Banko, Michael J Cafarella , Stephen Soderland, Matt Broadhead and Oren Etzioni. 2007. Open information extraction from the Web. IJCAI ’07. Matthew Berland, Eugene Charniak, 1999. Finding parts in very large corpora. ACL ’99. Michael Cafarella, Alon Halevy, Yang Zhang, Daisy Zhe Wang and Eugene Wu. 2008. WebTables: Exploring the Power of Tables on the Web. VLDB ’08. Timothy Chklovski and Patrick Pantel. 2004. VerbOcean: mining the Web for fine-grained semantic verb relations. EMNLP ’04. Eric Crestan and Patrick Pantel. 2010. Web-Scale Knowledge Extraction from Semi-Structured Tables. WWW ’10. Dmitry Davidov and Ari Rappoport. 2006. Efficient Unsupervised Discovery of Word Categories Using Symmetric Patterns and High Frequency Words. ACL-Coling ’06. Dmitry Davidov, Ari Rappoport and Moshe Koppel. 2007. Fully unsupervised discovery of conceptspecific relationships by web mining. ACL ’07. Dmitry Davidov and Ari Rappoport. 2008a. Classification of Semantic Relationships between Nominals Using Pattern Clusters. ACL ’08. Dmitry Davidov and Ari Rappoport. 2008b. Unsupervised Discovery of Generic Relationships Using Pattern Clusters and its Evaluation by Automatically Generated SAT Analogy Questions. ACL ’08. Roxana Girju, Adriana Badulescu, and Dan Moldovan. 2006. Automatic discovery of part-whole relations. Computational Linguistics, 32(1). Marty Hearst, 1992. Automatic acquisition of hyponyms from large text corpora. COLING ’92. Veronique Moriceau, 2006. Numerical Data Integration for Cooperative Question-Answering. EACL KRAQ06 ’06. John Prager, 2006. Open-domain question-answering. In Foundations and Trends in Information Retrieval,vol. 1, pp 91-231. 1316 Patrick Pantel and Marco Pennacchiotti. 2006. Espresso: leveraging generic patterns for automatically harvesting semantic relations. COLING-ACL ’06. Deepak Ravichandran and Eduard Hovy. 2002 Learning Surface Text Patterns for a Question Answering System. ACL ’02. Ellen Riloff and Rosie Jones. 1999. Learning Dictionaries for Information Extraction by Multi-Level Bootstrapping. AAAI ’99. Benjamin Rosenfeld and Ronen Feldman. 2007. Clustering for unsupervised relation identification. CIKM ’07. Peter Turney, 2005. Measuring semantic similarity by latent relational analysis, IJCAI ’05. Dominic Widdows and Beate Dorow. 2002. A graph model for unsupervised Lexical acquisition. COLING ’02. 1317
|
2010
|
133
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1318–1327, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Learning Word-Class Lattices for Definition and Hypernym Extraction Roberto Navigli and Paola Velardi Dipartimento di Informatica Sapienza Universit`a di Roma {navigli,velardi}@di.uniroma1.it Abstract Definition extraction is the task of automatically identifying definitional sentences within texts. The task has proven useful in many research areas including ontology learning, relation extraction and question answering. However, current approaches – mostly focused on lexicosyntactic patterns – suffer from both low recall and precision, as definitional sentences occur in highly variable syntactic structures. In this paper, we propose WordClass Lattices (WCLs), a generalization of word lattices that we use to model textual definitions. Lattices are learned from a dataset of definitions from Wikipedia. Our method is applied to the task of definition and hypernym extraction and compares favorably to other pattern generalization methods proposed in the literature. 1 Introduction Textual definitions constitute a fundamental source to look up when the meaning of a term is sought. Definitions are usually collected in dictionaries and domain glossaries for consultation purposes. However, manually constructing and updating glossaries requires the cooperative effort of a team of domain experts. Further, in the presence of new words or usages, and – even worse – new domains, such resources are of no help. Nonetheless, terms are attested in texts and some (usually few) of the sentences in which a term occurs are typically definitional, that is they provide a formal explanation for the term of interest. While it is not feasible to manually search texts for definitions, this task can be automatized by means of Machine Learning (ML) and Natural Language Processing (NLP) techniques. Automatic definition extraction is useful not only in the construction of glossaries, but also in many other NLP tasks. In ontology learning, definitions are used to create and enrich concepts with textual information (Gangemi et al., 2003), and extract taxonomic and non-taxonomic relations (Snow et al., 2004; Navigli and Velardi, 2006; Navigli, 2009a). Definitions are also harvested in Question Answering to deal with “what is” questions (Cui et al., 2007; Saggion, 2004). In eLearning, they are used to help students assimilate knowledge (Westerhout and Monachesi, 2007), etc. Much of the current literature focuses on the use of lexico-syntactic patterns, inspired by Hearst’s (1992) seminal work. However, these methods suffer both from low recall and precision, as definitional sentences occur in highly variable syntactic structures, and because the most frequent definitional pattern – X is a Y – is inherently very noisy. In this paper we propose a generalized form of word lattices, called Word-Class Lattices (WCLs), as an alternative to lexico-syntactic pattern learning. A lattice is a directed acyclic graph (DAG), a subclass of non-deterministic finite state automata (NFA). The lattice structure has the purpose of preserving the salient differences among distinct sequences, while eliminating redundant information. In computational linguistics, lattices have been used to model in a compact way many sequences of symbols, each representing an alternative hypothesis. Lattice-based methods differ in the types of nodes (words, phonemes, concepts), the interpretation of links (representing either a sequential or hierarchical ordering between nodes), their means of creation, and the scoring method used to extract the best consensus output from the lattice (Schroeder et al., 2009). In speech processing, phoneme or word lattices (Campbell et al., 2007; Mathias and Byrne, 2006; Collins et al., 2004) are used as an interface between speech recognition and understanding. Lat1318 tices are adopted also in Chinese word segmentation (Jiang et al., 2008), decompounding in German (Dyer, 2009), and to represent classes of translation models in machine translation (Dyer et al., 2008; Schroeder et al., 2009). In more complex text processing tasks, such as information retrieval, information extraction and summarization, the use of word lattices has been postulated but is considered unrealistic because of the dimension of the hypothesis space. To reduce this problem, concept lattices have been proposed (Carpineto and Romano, 2005; Klein, 2008; Zhong et al., 2008). Here links represent hierarchical relations, rather than the sequential order of symbols like in word/phoneme lattices, and nodes are clusters of salient words aggregated using synonymy, similarity, or subtrees of a thesaurus. However, salient word selection and aggregation is non-obvious and furthermore it falls into word sense disambiguation, a notoriously AI-hard problem (Navigli, 2009b). In definition extraction, the variability of patterns is higher than for “traditional” applications of lattices, such as translation and speech, however not as high as in unconstrained sentences. The methodology that we propose to align patterns is based on the use of star (wildcard *) characters to facilitate sentence clustering. Each cluster of sentences is then generalized to a lattice of word classes (each class being either a frequent word or a part of speech). A key feature of our approach is its inherent ability to both identify definitions and extract hypernyms. The method is tested on an annotated corpus of Wikipedia sentences and a large Web corpus, in order to demonstrate the independence of the method from the annotated dataset. WCLs are shown to generalize over lexico-syntactic patterns, and outperform well-known approaches to definition and hypernym extraction. The paper is organized as follows: Section 2 discusses related work, WCLs are introduced in Section 3 and illustrated by means of an example in Section 4, experiments are presented in Section 5. We conclude the paper in Section 6. 2 Related Work Definition Extraction. A great deal of work is concerned with definition extraction in several languages (Klavans and Muresan, 2001; Storrer and Wellinghoff, 2006; Gaudio and Branco, 2007; Iftene et al., 2007; Westerhout and Monachesi, 2007; Przepi´orkowski et al., 2007; Deg´orski et al., 2008). The majority of these approaches use symbolic methods that depend on lexico-syntactic patterns or features, which are manually crafted or semi-automatically learned (Zhang and Jiang, 2009; Hovy et al., 2003; Fahmi and Bouma, 2006; Westerhout, 2009). Patterns are either very simple sequences of words (e.g. “refers to”, “is defined as”, “is a”) or more complex sequences of words, parts of speech and chunks. A fully automated method is instead proposed by Borg et al. (2009): they use genetic programming to learn simple features to distinguish between definitions and non-definitions, and then they apply a genetic algorithm to learn individual weights of features. However, rules are learned for only one category of patterns, namely “is” patterns. As we already remarked, most methods suffer from both low recall and precision, because definitional sentences occur in highly variable and potentially noisy syntactic structures. Higher performance (around 6070% F1-measure) is obtained only for specific domains (e.g., an ICT corpus) and patterns (Borg et al., 2009). Only few papers try to cope with the generality of patterns and domains in real-world corpora (like the Web). In the GlossExtractor web-based system (Velardi et al., 2008), to improve precision while keeping pattern generality, candidates are pruned using more refined stylistic patterns and lexical filters. Cui et al. (2007) propose the use of probabilistic lexico-semantic patterns, called soft patterns, for definitional question answering in the TREC contest1. The authors describe two soft matching models: one is based on an n-gram language model (with the Expectation Maximization algorithm used to estimate the model parameter), the other on Profile Hidden Markov Models (PHMM). Soft patterns generalize over lexicosyntactic “hard” patterns in that they allow a partial matching by calculating a generative degree of match probability between the test instance and the set of training instances. Thanks to its generalization power, this method is the most closely related to our work, however the task of definitional question answering to which it is applied is slightly different from that of definition extraction, so a direct performance comparison is not possi1Text REtrieval Conferences: http://trec.nist. gov 1319 ble2. In fact, the TREC evaluation datasets cannot be considered true definitions, but rather text fragments providing some relevant fact about a target term. For example, sentences like: “Bollywood is a Bombay-based film industry” and “700 or more films produced by India with 200 or more from Bollywood” are both “vital” answers for the question “Bollywood”, according to TREC classification, but the second sentence is not a definition. Hypernym Extraction. The literature on hypernym extraction offers a higher variability of methods, from simple lexical patterns (Hearst, 1992; Oakes, 2005) to statistical and machine learning techniques (Agirre et al., 2000; Caraballo, 1999; Dolan et al., 1993; Sanfilippo and Pozna´nski, 1992; Ritter et al., 2009). One of the highest-coverage methods is proposed by Snow et al. (2004). They first search sentences that contain two terms which are known to be in a taxonomic relation (term pairs are taken from WordNet (Miller et al., 1990)); then they parse the sentences, and automatically extract patterns from the parse trees. Finally, they train a hypernym classifer based on these features. Lexico-syntactic patterns are generated for each sentence relating a term to its hypernym, and a dependency parser is used to represent them. 3 Word-Class Lattices 3.1 Preliminaries Notion of definition. In our work, we rely on a formal notion of textual definition. Specifically, given a definition, e.g.: “In computer science, a closure is a first-class function with free variables that are bound in the lexical environment”, we assume that it contains the following fields (Storrer and Wellinghoff, 2006): • The DEFINIENDUM field (DF): this part of the definition includes the definiendum (that is, the word being defined) and its modifiers (e.g., “In computer science, a closure”); • The DEFINITOR field (VF): it includes the verb phrase used to introduce the definition (e.g., “is”); 2In the paper, a 55% recall and 34% precision is achieved with the best experiment on TREC-13 data. Furthermore, the classifier of Cui et al. (2007) is based on soft patterns but also on a bag-of-word relevance heuristic. However, the relative influence of the two methods on the final performance is not discussed. • The DEFINIENS field (GF): it includes the genus phrase (usually including the hypernym, e.g., “a first-class function”); • The REST field (RF): it includes additional clauses that further specify the differentia of the definiendum with respect to its genus (e.g., “with free variables that are bound in the lexical environment”). Further examples of definitional sentences annotated with the above fields are shown in Table 1. For each sentence, the definiendum (that is, the word being defined) and its hypernym are marked in bold and italic, respectively. Given the lexicosyntactic nature of the definition extraction models we experiment with, training and test sentences are part-of-speech tagged with the TreeTagger system, a part-of-speech tagger available for many languages (Schmid, 1995). Word Classes and Generalized Sentences. We now introduce our notion of word class, on which our learning model is based. Let T be the set of training sentences, manually bracketed with the DF, VF, GF and RF fields. We first determine the set F of words in T whose frequency is above a threshold θ (e.g., the, a, is, of, refer, etc.). In our training sentences, we replace the term being defined with ⟨TARGET⟩, thus this frequent token is also included in F. We use the set of frequent words F to generalize words to “word classes”. We define a word class as either a word itself or its part of speech. Given a sentence s = w1, w2, . . . , w|s|, where wi is the i-th word of s, we generalize its words wi to word classes ωi as follows: ωi = ( wi if wi ∈F POS(wi) otherwise that is, a word wi is left unchanged if it occurs frequently in the training corpus (i.e., wi ∈F) or is transformed to its part of speech (POS(wi)) otherwise. As a result, we obtain a generalized sentence s′ = ω1, ω2, . . . , ω|s|. For instance, given the first sentence in Table 1, we obtain the corresponding generalized sentence: “In NN, a ⟨TARGET⟩is a JJ NN”, where NN and JJ indicate the noun and adjective classes, respectively. 3.2 Algorithm We now describe our learning algorithm based on Word-Class Lattices. The algorithm consists of three steps: 1320 [In arts, a chiaroscuro]DF [is]VF [a monochrome picture]GF. [In mathematics, a graph]DF [is]VF [a data structure]GF [that consists of ...]REST. [In computer science, a pixel]DF [is]VF [a dot]GF [that is part of a computer image]REST. Table 1: Example definitions (defined terms are marked in bold face, their hypernyms in italic). • Star patterns: each sentence in the training set is pre-processed and generalized to a star pattern. For instance, “In arts, a chiaroscuro is a monochrome picture” is transformed to “In *, a ⟨TARGET⟩is a *” (Section 3.2.1); • Sentence clustering: the training sentences are then clustered based on the star patterns to which they belong (Section 3.2.2); • Word-Class Lattice construction: for each sentence cluster, a WCL is created by means of a greedy alignment algorithm (Section 3.2.3). We present two variants of our WCL model, dealing either globally with the entire sentence or separately with its definition fields (Section 3.2.4). The WCL models can then be used to classify any input sentence of interest (Section 3.2.5). 3.2.1 Star Patterns Let T be the set of training sentences. In this step, we associate a star pattern σ(s) with each sentence s ∈T . To do so, let s ∈T be a sentence such that s = w1, w2, . . . , w|s|, where wi is its i-th word. Given the set F of most frequent words in T (cf. Section 3.1), the star pattern σ(s) associated with s is obtained by replacing with * all the words wi ̸∈F, that is all the tokens that are non-frequent words. For instance, given the sentence “In arts, a chiaroscuro is a monochrome picture”, the corresponding star pattern is “In *, a ⟨TARGET⟩is a *”, where ⟨TARGET⟩is the defined term. Note that, here and in what follows, we discard the sentence fragments tagged with the REST field, which is used only to delimit the core part of definitional sentences. 3.2.2 Sentence Clustering In the second step, we cluster the sentences in our training set T based on their star patterns. Formally, let Σ = (σ1, . . . , σm) be the set of star patterns associated with the sentences in T . We create a clustering C = (C1, . . . , Cm) such that Ci = {s ∈T : σ(s) = σi}, that is Ci contains all the sentences whose star pattern is σi. As an example, assume σ3 = “In *, a ⟨TARGET⟩is a *”. The sentences reported in Table 1 are all grouped into cluster C3. We note that each cluster Ci contains sentences whose degree of variability is generally much lower than for any pair of sentences in T belonging to two different clusters. 3.2.3 Word-Class Lattice Construction Finally, the third step consists of the construction of a Word-Class Lattice for each sentence cluster. Given such a cluster Ci ∈C, we apply a greedy algorithm that iteratively constructs the WCL. Let Ci = {s1, s2, . . . , s|Ci|} and consider its first sentence s1 = w1 1, w1 2, . . . , w1 |s1| (wj i denotes the i-th token of the j-th sentence). We first produce the corresponding generalized sentence s′ 1 = ω1 1, ω1 2, . . . , ω1 |s1| (cf. Section 3.1). We then create a directed graph G = (V, E) such that V = {ω1 1, . . . , ω1 |s1|} and E = {(ω1 1, ω1 2), (ω1 2, ω1 3), . . . , (ω1 |s1|−1, ω1 |s1|)}. Next, for the subsequent sentences in Ci, that is, for each j = 2, . . . , |Ci|, we determine the alignment between the sentence sj and each sentence sk ∈Ci such that k < j based on the following dynamic programming formulation (Cormen et al., 1990, pp. 314–319): Ma,b = max {Ma−1,b−1 +Sa,b, Ma,b−1, Ma−1,b} where a ∈{1, . . . , |sk|} and b ∈{1, . . . , |sj|}, Sa,b is a score of the matching between the a-th token of sk and the b-th token of sj, and M0,0, M0,b and Ma,0 are initially set to 0 for all a and b. The matching score Sa,b is calculated on the generalized sentences s′ k of sk and s′ j of sj as follows: Sa,b = ( 1 if ωk a = ωj b 0 otherwise where ωk a and ωj b are the a-th and b-th word classes of s′ k and s′ j, respectively. In other words, the matching score equals 1 if the a-th and the b-th tokens of the two original sentences have the same word class. Finally, the alignment score between sk and sj is given by M|sk|,|sj|, which calculates the mini1321 In arts science mathematics NN1 NN4 computer , a ⟨TARGET⟩ pixel graph chiaroscuro is a monochrome JJ NN2 structure picture dot NN3 data Figure 1: The Word-Class Lattice for the sentences in Table 1. The support of each word class is reported beside the corresponding node. mal number of misalignments between the two token sequences. We repeat this calculation for each sentence sk (k = 1, . . . , j −1) and choose the one that maximizes its alignment score with sj. We then use the best alignment to add sj to the graph G. Such alignment is obtained by means of backtracking from M|sk|,|sj| to M0,0. We add to the set of vertices V the tokens of the generalized sentence s′ j for which there is no alignment to s′ k and we add to E the edges (ωj 1, ωj 2), . . . , (ωj |sj|−1, ωj |sj|). Furthermore, in the final lattice, nodes associated with the hypernym words in the learning sentences are marked as hypernyms in order to be able to determine the hypernym of a test sentence at classification time. 3.2.4 Variants of the WCL Model So far, we have assumed that our WCL model learns lattices from the training sentences in their entirety (we call this model WCL-1). We now propose a second model that learns separate WCLs for each field of the definition, namely: the DEFINIENDUM (DF), DEFINITOR (VF) and DEFINIENS (GF) fields (see Section 3.1). We refer to this latter model as WCL-3. Rather than applying the WCL algorithm to the entire sentence, the very same method is applied to the sentence fragments tagged with one of the three definition fields. The reason for introducing the WCL-3 model is that, while definitional patterns are highly variable, DF, VF and GF individually exhibit a lower variability, thus WCL-3 should improve the generalization power. 3.2.5 Classification Once the learning process is over, a set of WCLs is produced. Given a test sentence s, the classification phase for the WCL-1 model consists of determining whether it exists a lattice that matches s. In the case of WCL-3, we consider any combination of DEFINIENDUM, DEFINITOR and DEFINIENS lattices. While WCL-1 is applied as a yes-no classifier as there is a single WCL that can possibly match the input sentence, WCL-3 selects, if any, the combination of the three WCLs that best fits the sentence. In fact, choosing the most appropriate combination of lattices impacts the performance of hypernym extraction. The best combination of WCLs is selected by maximizing the following confidence score: score(s, lDF, lVF, lGF) = coverage · log(support) where s is the candidate sentence, lDF, lVF and lGF are three lattices one for each definition field, coverage is the fraction of words of the input sentence covered by the three lattices, and support is the sum of the number of sentences in the star patterns corresponding to the three lattices. Finally, when a sentence is classified as a definition, its hypernym is extracted by selecting the words in the input sentence that are marked as “hypernyms” in the WCL-1 lattice (or in the WCL-3 GF lattice). 4 Example As an example, consider the definitions in Table 1. As illustrated in Section 3.2.2, their star pattern is “In *, a ⟨TARGET⟩is a *”. The corresponding WCL is built as follows: the first partof-speech tagged sentence, “In/IN arts/NN , a/DT ⟨TARGET⟩/NN is/VBZ a/DT monochrome/JJ picture/NN”, is considered. The corresponding generalized sentence is “In NN , a ⟨TARGET⟩is a JJ NN”. The initially empty graph is thus populated with one node for each word class and one edge for each pair of consecutive tokens, as shown in Figure 1 (the central sequence of nodes in the graph). Note that we draw the hypernym token NN2 with a rectangle shape. We also add to the 1322 graph a start node • and an end node •⃝, and connect them to the corresponding initial and final sentence tokens. Next, the second sentence, “In mathematics, a graph is a data structure that consists of...”, is aligned to the first sentence. The alignment of the generalized sentence is perfect, apart from the NN3 node corresponding to “data”. The node is added to the graph together with the edges a→NN3 and NN3 →NN2 . Finally, the third sentence in Table 1, “In computer science, a pixel is a dot that is part of a computer image”, is generalized as “In NN NN , a ⟨TARGET⟩is a NN”. Thus, a new node NN4 is added, corresponding to “computer” and new edges are added: In→NN4 and NN4→NN1. Figure 1 shows the resulting WCL-1 lattice. 5 Experiments 5.1 Experimental Setup Datasets. We conducted experiments on two different datasets: • A corpus of 4,619 Wikipedia sentences, that contains 1,908 definitional and 2,711 nondefinitional sentences. The former were obtained from a random selection of the first sentences of Wikipedia articles3. The defined terms belong to different Wikipedia domain categories4, so as to capture a representative and cross-domain sample of lexical and syntactic patterns for definitions. These sentences were manually annotated with DEFINIENDUM, DEFINITOR, DEFINIENS and REST fields by an expert annotator, who also marked the hypernyms. The associated set of negative examples (“syntactically plausible” false definitions) was obtained by extracting from the same Wikipedia articles sentences in which the page title occurs. • A subset of the ukWaC Web corpus (Ferraresi et al., 2008), a large corpus of the English language constructed by crawling the .uk domain of the Web. The subset includes over 300,000 sentences in which occur any of 239 terms selected from the terminology of four different domains (COMPUTER SCI3The first sentence of Wikipedia entries is, in the large majority of cases, a definition of the page title. 4en.wikipedia.org/wiki/Wikipedia:Categories ENCE, ASTRONOMY, CARDIOLOGY, AVIATION). The reason for using the ukWaC corpus is that, unlike the “clean” Wikipedia dataset, in which relatively simple patterns can achieve good results, ukWaC represents a real-world test, with many complex cases. For example, there are sentences that should be classified as definitional according to Section 3.1 but are rather uninformative, like “dynamic programming was the brainchild of an american mathematician”, as well as informative sentences that are not definitional (e.g., they do not have a hypernym), like “cubism was characterised by muted colours and fragmented images”. Even more frequently, the dataset includes sentences which are not definitions but have a definitional pattern (“A Pacific Northwest tribe’s saga refers to a young woman who [..]”), or sentences with very complex definitional patterns (“white body cells are the body’s clean up squad” and “joule is also an expression of electric energy”). These cases can be correctly handled only with fine-grained patterns. Additional details on the corpus and a more thorough linguistic analysis of complex cases can be found in Navigli et al. (2010). Systems. For definition extraction, we experiment with the following systems: • WCL-1 and WCL-3: these two classifiers are based on our Word-Class Lattice model. WCL-1 learns from the training set a lattice for each cluster of sentences, whereas WCL3 identifies clusters (and lattices) separately for each sentence field (DEFINIENDUM, DEFINITOR and DEFINIENS) and classifies a sentence as a definition if any combination from the three sets of lattices matches (cf. Section 3.2.4, the best combination is selected). • Star patterns: a simple classifier based on the patterns learned as a result of step 1 of our WCL learning algorithm (cf. Section 3.2.1): a sentence is classified as a definition if it matches any of the star patterns in the model. • Bigrams: an implementation of the bigram classifier for soft pattern matching proposed by Cui et al. (2007). The classifier selects as definitions all the sentences whose probability is above a specific threshold. The probability is calculated as a mixture of bigram and 1323 Algorithm P R F1 A WCL-1 99.88 42.09 59.22 76.06 WCL-3 98.81 60.74 75.23 83.48 Star patterns 86.74 66.14 75.05 81.84 Bigrams 66.70 82.70 73.84 75.80 Random BL 50.00 50.00 50.00 50.00 Table 2: Performance on the Wikipedia dataset. unigram probabilities, with Laplace smoothing on the latter. We use the very same settings of Cui et al. (2007), including threshold values. While the authors propose a second soft-pattern approach based on Profile HMM (cf. Section 2), their results do not show significant improvements over the bigram language model. For hypernym extraction, we compared WCL1 and WCL-3 with Hearst’s patterns, a system that extracts hypernyms from sentences based on the lexico-syntactic patterns specified in Hearst’s seminal work (1992). These include (hypernym in italic): “such NP as {NP ,} {(or | and)} NP”, “NP {, NP} {,} or other NP”, “NP {,} including { NP ,} {or | and} NP”, “NP {,} especially { NP ,} {or | and} NP”, and variants thereof. However, it should be noted that hypernym extraction methods in the literature do not extract hypernyms from definitional sentences, like we do, but rather from specific patterns like “X such as Y”. Therefore a direct comparison with these methods is not possible. Nonetheless, we decided to implement Hearst’s patterns for the sake of completeness. We could not replicate the more refined approach by Snow et al. (2004) because it requires the annotation of a possibly very large dataset of sentence fragments. In any case Snow et al. (2004) reported the following performance figures on a corpus of dimension and complexity comparable with ukWaC: the recall-precision graph indicates precision 85% at recall 10% and precision 25% at recall of 30% for the hypernym classifier. A variant of the classifier that includes evidence from coordinate terms (terms with a common ancestor in a taxonomy) obtains an increased precision of 35% at recall 30%. We see no reasons why these figures should vary dramatically on the ukWaC. Finally, we compare all systems with the random baseline, that classifies a sentence as a definition with probability 1 2. Algorithm P R† WCL-1 98.33 39.39 WCL-3 94.87 56.57 Star patterns 44.01 63.63 Bigrams 46.60 45.45 Random BL 50.00 50.00 Table 3: Performance on the ukWaC dataset († Recall is estimated). Measures. To assess the performance of our systems, we calculated the following measures: • precision – the number of definitional sentences correctly retrieved by the system over the number of sentences marked by the system as definitional. • recall – the number of definitional sentences correctly retrieved by the system over the number of definitional sentences in the dataset. • the F1-measure – a harmonic mean of precision (P) and recall (R) given by 2PR P+R. • accuracy – the number of correctly classified sentences (either as definitional or nondefinitional) over the total number of sentences in the dataset. 5.2 Results and Discussion Definition Extraction. In Table 2 we report the results of definition extraction systems on the Wikipedia dataset. Given this dataset is also used for training, experiments are performed with 10fold cross validation. The results show very high precision for WCL-1, WCL-3 (around 99%) and star patterns (86%). As expected, bigrams and star patterns exhibit a higher recall (82% and 66%, respectively). The lower recall of WCL-1 is due to its limited ability to generalize compared to WCL3 and the other methods. In terms of F1-measure, star patterns and WCL-3 achieve 75%, and are thus the best systems. Similar performance is observed when we also account for negative sentences – that is we calculate accuracy (with WCL3 performing better). All the systems perform significantly better than the random baseline. From our Wikipedia corpus, we learned over 1,000 lattices (and star patterns). Using WCL3, we learned 381 DF, 252 VF and 395 GF lattices, that then we used to extract definitions from 1324 Algorithm Full Substring WCL-1 42.75 77.00 WCL-3 40.73 78.58 Table 4: Precision in hypernym extraction on the Wikipedia dataset the ukWaC dataset. To calculate precision on this dataset, we manually validated the definitions output by each system. However, given the large size of the test set, recall could only be estimated. To this end, we manually analyzed 50,000 sentences and identified 99 definitions, against which recall was calculated. The results are shown in Table 3. On the ukWaC dataset, WCL-3 performs best, obtaining 94.87% precision and 56.57% recall (we did not calculate F1, as recall is estimated). Interestingly, star patterns obtain only 44% precision and around 63% recall. Bigrams achieve even lower performance, namely 46.60% precision, 45.45% recall. The reason for such bad performance on ukWaC is due to the very different nature of the two datasets: for example, in Wikipedia most “is a” sentences are definitional, whereas this property is not verified in the real world (that is, on the Web, of which ukWaC is a sample). Also, while WCL does not need any parameter tuning5, the same does not hold for bigrams6, whose probability threshold and mixture weights need to be best tuned on the task at hand. Hypernym Extraction. For hypernym extraction, we tested WCL-1, WCL-3 and Hearst’s patterns. Precision results are reported in Tables 4 and 5 for the two datasets, respectively. The Substring column refers to the case in which the captured hypernym is a substring of what the annotator considered to be the correct hypernym. Notice that this is a complex matter, because often the selection of a hypernym depends on semantic and contextual issues. For example, “Fluoroscopy is an imaging method” and “the Mosaic was an interesting project” have precisely the same genus pattern, but (probably depending on the vagueness of the noun in the first sentence, and of the adjective in the second) the annotator selected respec5WCL has only one threshold value θ to be set for determining frequent words (cf. Section 3.1). However, no tuning was made for choosing the best value of θ. 6We had to re-tune the system parameters on ukWaC, since with the original settings of Cui et al. (2007) performance was much lower. Algorithm Full Substring WCL-1 86.19 (206) 96.23 (230) WCL-3 89.27 (383) 96.27 (413) Hearst 65.26 (62) 88.42 (84) Table 5: Precision in hypernym extraction on the ukWaC dataset (number of hypernyms in parentheses). tively imaging method and project as hypernyms. For the above reasons it is difficult to achieve high performance in capturing the correct hypernym (e.g. 40.73% with WCL-3 on Wikipedia). However, our performance of identifying a substring of the correct hypernym is much higher (around 78.58%). In Table 4 we do not report the precision of Hearst’s patterns, as only one hypernym was found, due to the inherently low coverage of the method. On the ukWaC dataset, the hypernyms returned by the three systems were manually validated and precision was calculated. Both WCL-1 and WCL3 obtained a very high precision (86-89% and 96% in identifying the exact hypernym and a substring of it, respectively). Both WCL models are thus equally robust in identifying hypernyms, whereas WCL-1 suffers from a lack of generalization in definition extraction (cf. Tables 2 and 3). Also, given that the ukWaC dataset contains sentences in which any of 239 domain terms occur, WCL-3 extracts on average 1.6 and 1.7 full and substring hypernyms per term, respectively. Hearst’s patterns also obtain high precision, especially when substrings are taken into account. However, the number of hypernyms returned by this method is much lower, due to the specificity of the patterns (62 vs. 383 hypernyms returned by WCL-3). 6 Conclusions In this paper, we have presented a lattice-based approach to definition and hypernym extraction. The novelty of our approach is: 1. The use of a lattice structure to generalize over lexico-syntactic definitional patterns; 2. The ability of the system to jointly identify definitions and extract hypernyms; 3. The generality of the method, which applies to generic Web documents in any domain and style, and needs no parameter tuning; 1325 4. The high performance as compared with the best-known methods for both definition and hypernym extraction. Our approach outperforms the other systems particularly where the task is more complex, as in real-world documents (i.e., the ukWaC corpus). Even though definitional patterns are learned from a manually annotated dataset, the dimension and heterogeneity of the training dataset ensures that training needs not to be repeated for specific domains7, as demonstrated by the cross-domain evaluation on the ukWaC corpus. The datasets used in our experiments are available from http://lcl.uniroma1.it/wcl. We also plan to release our system to the research community. In the near future, we aim to apply the output of our classifiers to the task of automated taxonomy building, and to test the WCL approach on other information extraction tasks, like hypernym extraction from generic sentence fragments, as in Snow et al. (2004). References Eneko Agirre, Ansa Olatz, Xabier Arregi, Xabier Artola, Arantza Daz de Ilarraza Snchez, Mikel Lersundi, David Martnez, Kepa Sarasola, and Ruben Urizar. 2000. Extraction of semantic relations from a basque monolingual dictionary using constraint grammar. In Proceedings of Euralex. Claudia Borg, Mike Rosner, and Gordon Pace. 2009. Evolutionary algorithms for definition extraction. In Proceedings of the 1st Workshop on Definition Extraction 2009 (wDE’09). William M. Campbell, M. F. Richardson, and D. A. Reynolds. 2007. Language recognition with word lattices and support vector machines. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2007), pages 989–992, Honolulu, HI. Sharon A. Caraballo. 1999. Automatic construction of a hypernym-labeled noun hierarchy from text. In Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics (ACL), pages 120–126, Maryland, USA. Claudio Carpineto and Giovanni Romano. 2005. Using concept lattices for text retrieval and mining. In B. Ganter, G. Stumme, and R. Wille, editors, Formal Concept Analysis, pages 161–179. Christopher Collins, Bob Carpenter, and Gerald Penn. 2004. Head-driven parsing for word lattices. In Proceedings of the 42nd Meeting of the Association for 7Of course, it would need some additional work if applied to languages other than English. However, the approach does not need to be adapted to the language of interest. Computational Linguistics (ACL’04), Main Volume, pages 231–238, Barcelona, Spain, July. Thomas H. Cormen, Charles E. Leiserson, and Ronald L. Rivest. 1990. Introduction to algorithms. the MIT Electrical Engineering and Computer Science Series. MIT Press, Cambridge, MA. Hang Cui, Min-Yen Kan, and Tat-Seng Chua. 2007. Soft pattern matching models for definitional question answering. ACM Transactions on Information Systems, 25(2):8. Łukasz Deg´orski, Michał Marcinczuk, and Adam Przepi´orkowski. 2008. Definition extraction using a sequential combination of baseline grammars and machine learning classifiers. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC 2008), Marrakech, Morocco. William Dolan, Lucy Vanderwende, and Stephen D. Richardson. 1993. Automatically deriving structured knowledge bases from on-line dictionaries. In Proceedings of the First Conference of the Pacific Association for Computational Linguistics, pages 5– 14. Christopher Dyer, Smaranda Muresan, and Philip Resnik. 2008. Generalizing word lattice translation. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL 2008), pages 1012–1020, Columbus, Ohio, USA. Christopher Dyer. 2009. Using a maximum entropy model to build segmentation lattices for mt. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics (HLT-NAACL 2009), pages 406–414, Boulder, Colorado, USA. Ismail Fahmi and Gosse Bouma. 2006. Learning to identify definitions using syntactic features. In Proceedings of the EACL 2006 workshop on Learning Structured Information in Natural Language Applications, pages 64–71, Trento, Italy. Adriano Ferraresi, Eros Zanchetta, Marco Baroni, and Silvia Bernardini. 2008. Introducing and evaluating ukwac, a very large Web-derived corpus of english. In Proceedings of the 4th Web as Corpus Workshop (WAC-4), Marrakech, Morocco. Aldo Gangemi, Roberto Navigli, and Paola Velardi. 2003. The OntoWordNet project: Extension and axiomatization of conceptual relations in WordNet. In Proceedings of the International Conference on Ontologies, Databases and Applications of SEmantics (ODBASE 2003), pages 820–838, Catania, Italy. Rosa Del Gaudio and Ant´onio Branco. 2007. Automatic extraction of definitions in portuguese: A rulebased approach. In Proceedings of the TeMa Workshop. Marti Hearst. 1992. Automatic acquisition of hyponyms from large text corpora. In Proceedings of the 14th International Conference on Computational Linguistics (COLING), pages 539–545, Nantes, France. 1326 Eduard Hovy, Andrew Philpot, Judith Klavans, Ulrich Germann, and Peter T. Davis. 2003. Extending metadata definitions by automatically extracting and organizing glossary definitions. In Proceedings of the 2003 Annual National Conference on Digital Government Research, pages 1–6. Digital Government Society of North America. Adrian Iftene, Diana Trandab˘a, and Ionut Pistol. 2007. Natural language processing and knowledge representation for elearning environments. In Proc. of Applications for Romanian. Proceedings of RANLP workshop, pages 19–25. Wenbin Jiang, Haitao Mi, and Qun Liu. 2008. Word lattice reranking for chineseword segmentation and part-of-speech tagging. In Proceedings of the 22nd International Conference on Computational Linguistics (COLING 2008), pages 385–392, Manchester, UK. Judith Klavans and Smaranda Muresan. 2001. Evaluation of the DEFINDER system for fully automatic glossary construction. In Proc. of the American Medical Informatics Association (AMIA) Symposium. Michael Tully Klein. 2008. Understanding English with Lattice-Learning, Master thesis. MIT, Cambridge, MA, USA. Lambert Mathias and William Byrne. 2006. Statistical phrase-based speech translation. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2006), Toulouse, France. George A. Miller, R.T. Beckwith, Christiane D. Fellbaum, D. Gross, and K. Miller. 1990. WordNet: an online lexical database. International Journal of Lexicography, 3(4):235–244. Roberto Navigli and Paola Velardi. 2006. Ontology enrichment through automatic semantic annotation of on-line glossaries. In Proceedings of the 15th International Conference on Knowledge Engineering and Knowledge Management (EKAW 2006), pages 126–140, Podebrady, Czech Republic. Roberto Navigli, Paola Velardi, and Juana Mar´ıa RuizMart´ınez. 2010. An annotated dataset for extracting definitions and hypernyms from the Web. In Proceedings of the 7th International Conference on Language Resources and Evaluation (LREC 2010), Valletta, Malta. Roberto Navigli. 2009a. Using cycles and quasi-cycles to disambiguate dictionary glosses. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2009), pages 594–602, Athens, Greece. Roberto Navigli. 2009b. Word Sense Disambiguation: A survey. ACM Computing Surveys, 41(2):1–69. Michael P. Oakes. 2005. Using hearst’s rules for the automatic acquisition of hyponyms for mining a pharmaceutical corpus. In Proceedings of the Workshop Text Mining Research. Adam Przepi´orkowski, Lukasz Deg´orski, Beata W´ojtowicz, Miroslav Spousta, Vladislav Kuboˇn, Kiril Simov, Petya Osenova, and Lothar Lemnitzer. 2007. Towards the automatic extraction of definitions in slavic. In Proceedings of the Workshop on Balto-Slavonic Natural Language Processing (in ACL ’07), pages 43–50, Prague, Czech Republic. Association for Computational Linguistics. Alan Ritter, Stephen Soderland, and Oren Etzioni. 2009. What is this, anyway: Automatic hypernym discovery. In Proceedings of the 2009 AAAI Spring Symposium on Learning by Reading and Learning to Read, pages 88–93. Horacio Saggion. 2004. Identifying denitions in text collections for question answering. In Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC 2004), Lisbon, Portugal. Antonio Sanfilippo and Victor Pozna´nski. 1992. The acquisition of lexical knowledge from combined machine-readable dictionary sources. In Proceedings of the third Conference on Applied Natural Language Processing, pages 80–87. Helmut Schmid. 1995. Improvements in part-ofspeech tagging with an application to german. In Proceedings of the ACL SIGDAT-Workshop, pages 47–50. Josh Schroeder, Trevor Cohn, and Philipp Koehn. 2009. Word lattices for multi-source translation. In Proceedings of the European Chapter of the Association for Computation Linguistics (EACL 2009), pages 719–727, Athens, Greece. Rion Snow, Dan Jurafsky, and Andrew Y. Ng. 2004. Learning syntactic patterns for automatic hypernym discovery. In Proceedings of Advances in Neural Information Processing Systems, pages 1297–1304. Angelika Storrer and Sandra Wellinghoff. 2006. Automated detection and annotation of term definitions in german text corpora. In Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC 2006), Genova, Italy. Paola Velardi, Roberto Navigli, and Pierluigi D’Amadio. 2008. Mining the Web to create specialized glossaries. IEEE Intelligent Systems, 23(5):18–25. Eline Westerhout and Paola Monachesi. 2007. Extraction of dutch definitory contexts for eLearning purposes. In Proceedings of CLIN. Eline Westerhout. 2009. Definition extraction using linguistic and structural features. In Proceedings of the RANLP 2009 Workshop on Definition Extraction, pages 61–67. Chunxia Zhang and Peng Jiang. 2009. Automatic extraction of definitions. In Proceedings of 2nd IEEE International Conference on Computer Science and Information Technology, pages 364–368. Zhao-man Zhong, Zong-tian Liu, and Yan Guan. 2008. Precise information extraction from text based on two-level concept lattice. In Proceedings of the 2008 International Symposiums on Information Processing (ISIP ’08), pages 275–279, Washington, DC, USA. 1327
|
2010
|
134
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1328–1336, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics On Learning Subtypes of the Part-Whole Relation: Do Not Mix your Seeds Ashwin Ittoo University of Groningen Groningen, The Netherlands [email protected] Gosse Bouma University of Groningen Groningen, The Netherlands [email protected] Abstract An important relation in information extraction is the part-whole relation. Ontological studies mention several types of this relation. In this paper, we show that the traditional practice of initializing minimally-supervised algorithms with a single set that mixes seeds of different types fails to capture the wide variety of part-whole patterns and tuples. The results obtained with mixed seeds ultimately converge to one of the part-whole relation types. We also demonstrate that all the different types of part-whole relations can still be discovered, regardless of the type characterized by the initializing seeds. We performed our experiments with a state-ofthe-art information extraction algorithm. 1 Introduction A fundamental semantic relation in many disciplines such as linguistics, cognitive science, and conceptual modelling is the part-whole relation, which exists between parts and the wholes they compise (Winston et al., 1987; Gerstl and Pribbenow, 1995). Different types of part-whole relations, classified in various taxonomies, are mentioned in literature (Winston et al., 1987; Odell, 1994; Gerstl and Pribbenow, 1995; Keet and Artale, 2008). The taxonomy of Keet and Artale (2008), for instance, distinguishes part-whole relations based on their transitivity, and on the semantic classes of entities they sub-categorize. Part-whole relations are also crucial for many information extraction (IE) tasks (Girju et al., 2006). Annotated corpora and semantic dictionaries used in IE, such as the ACE corpus1 and WordNet (Fellbaum, 1998), include examples of part-whole relations. Also, previous relation extraction work, 1http://projects.ldc.upenn.edu/ace/ such as Berland and Charniak (1999) and Girju et al. (2006), have specifically targeted the discovery of part-whole relations from text. Furthermore, part-whole relations are de-facto benchmarks for evaluating the performance of general relation extraction systems (Pantel and Pennacchiotti, 2006; Beamer et al., 2008; Pyysalo et al., 2009). However, these relation extraction efforts have overlooked the ontological distinctions between the different types of part-whole relations. They assume the existence of a single relation, subsuming the different part-whole relation types. In this paper, we show that enforcing the ontological distinctions between the different types of part-whole relations enable information extraction systems to capture a wider variety of both generic and specialised part-whole lexico-syntactic patterns and tuples. Specifically, we address 3 major questions. 1. Is information extraction (IE) harder when learning the individual types of part-whole relations? That is, we determine whether the performance of state-of-the-art IE systems in learning the individual part-whole relation types increases (due to more coherency in the relations’ linguistic realizations) or drops (due to fewer examples), compared to the traditional practice of considering a single partwhole relation. 2. Are the patterns and tuples discovered when focusing on a specific part-whole relation type confined to that particular type? That is, we investgate whether IE systems discover examples representative of the different types by targetting one particular part-whole relation type. 3. Are more distinct examples discovered when IE systems learn the individual part-whole relation types? That is, we determine whether 1328 a wider variety of unique patterns and tuples are extracted when IE systems target the different types of part-whole relations instead of considering a single part-whole relation that subsumes all the different types. To answer these questions, we bootstrapped a minimally-supervised relation extraction algorithm, based on Espresso (Pantel and Pennacchiotti, 2006), with different seed-sets for the various types of part-whole relations, and analyzed the harvested tuples and patterns. 2 Previous Work Investigations on the part-whole relations span across many disciplines, such as conceptual modeling (Artale et al., 1996; Keet, 2006; Keet and Artale, 2008), which focus on the ontological aspects, and linguistics and cognitive sciences, which focus on natural language semantics. Several linguistically-motivated taxonomies (Odell, 1994; Gerstl and Pribbenow, 1995), based on the work of Winston et al. (1987), have been proposed to clarify the semantics of the different part-whole relations types across these various disciplines. Keet and Artale (2008) developed a formal taxonomy, distinguishing transitive mereological partwhole relations from intransitive meronymic ones. Meronymic relations identified are: 1) memberof, between a physical object (or role) and an aggregation, e.g. player-team, 2) constituted-of, between a physical object and an amount of matter e.g. clay-statue, 3) sub-quantity-of, between amounts of matter or units, e.g. oxygen-water or m-km, and 4)participates-in, between an entity and a process e.g. enzyme-reaction. Mereological relations are: 1)involved-in, between a phase and a process, e.g. chewing-eating, 2) locatedin, between an entity and its 2-dimensional region, e.g. city-region, 3)contained-in, between an entity and its 3-dimensional region, e.g.tooltrunk, and 4)structural part-of, between integrals and their (functional) components, e.g. enginecar. This taxonomy further discriminates between part-whole relation types by enforcing semantical selectional restrictions, in the form of DOLCE ontology (Gangemi et al., 2002) classes, on their entities. In NLP, information extraction (IE) techniques, for discovering part-whole relations from text have also been developed. Berland and Charniak (1999) use manually-crafted patterns, similar to Hearst (1992), and on initial “seeds” denoting “whole” objects (e.g. building) to harvest possible “part” objects (e.g. room) from the North Americal News Corpus (NANC) of 1 million words. They rank their results with measures like log-likelihood (Dunning, 1993), and report a maximum accuracy of 70% over their top-20 results. In the supervised approaches in Girju et al. (2003) and Girju et al. (2006), lexical patterns expressing partwhole relations between WordNet concept pairs are manually extracted from 20,000 sentences of the L.A Times and SemCor corpora (Miller et al., 1993), and used to generate a training corpus, with manually-annotated positive and negative examples of part-whole relations. Classification rules, induced over the training data, achieve a precision of 80.95% and recall of 75.91% in predicting whether an unseen pattern encode a partwhole relation. Van Hage et al. (2006) acquire 503 part-whole pairs from dedicated thesauri (e.g. AGROVOC2) to learn 91 reliable part-whole patterns. They substituted the patterns’ “part” arguments with known entities to formulate websearch queries. Corresponding “whole” entities were then discovered from documents in the query results with a precision of 74%. The part-whole relation is also a benchmark to evaluate the performance of general information extraction systems. The Espresso algorithm (Pantel and Pennacchiotti, 2006) achieves a precision of 80% in learning partwhole relations from the Acquaint (TREC-9) corpus of nearly 6M words. Despite the reasonable performance of the above IE systems in discovering part-whole relations, they overlook the ontological distinctions between the different relation types. For example, Girju et al. (2003) and Girju et al. (2006) assume a single part-whole relation, encompassing all the different types mentioned in the taxonomy of Winston et al. (1987). Similarly, the minimally-supervised Espresso algorithm (Pantel and Pennacchiotti, 2006) is initialized with a single set that mixes seeds of heterogeneous types, such as leader-panel and oxygen-water, which respectively correspond to the member-of and sub-quantity-of relations in the taxonomy of Keet and Artale (2008). 2http://aims.fao.org/website/ AGROVOC-Thesaurus/sub 1329 3 Methodology Our aim is to compare the relations harvested when a minimally-supervised IE algorithm is initialized with separate sets of seeds for each type of part-whole relation, and when it is initialized following the traditional practice of a single set that mixes seeds of the different types. To distinguish between types of part-whole relations, we commit to the taxonomy of Keet and Artale (2008) (Keet’s taxonomy), which uses sound ontological formalisms to unambiguously discrimate the relation types. Also, this taxonomy classifies the various part-whole relations introduced in literature, including ontologically-motivated mereological relations and linguistically-motivated meronymic ones. We adopt a 3-step approach to address our questions from section 1. 1. Define prototypical seeds (part-whole tuples) as follows: • (Separate) sets of seeds for each type of part-whole relation in Keet’s taxonomy. • A single set that mixes seeds denoting all the different part-whole relations types. 2. Part-whole relations extraction from a corpus by initializing a minimally-supervised IE algorithm with the seed-sets 3. Evaluation of the harvested relations to determine performance gain/loss, types of partwhole relations extracted, and distinct and unique patterns and tuples discovered. The corpora and IE algorithm we used, and the seed-sets construction are described below. Results are presented in the next section. 3.1 Corpora We used the English and Dutch Wikipedia texts since their broad-coverage and size ensures that they include sufficient lexical realizations of the different types of part-whole relations. Wikipedia has also been targeted by recent IE efforts (Nguyen et al., 2007; Wu and Weld, 2007). However, while they exploited the structured features (e.g. infoboxes), we only consider the unstructured texts. The English corpus size is approximately 470M words (∼80% of the August 2007 dump), while for Dutch, we use the full text collection (February 2008 dump) of approximately 110M words. We parsed the English and Dutch corpora respectively with the Stanford3 (Klein and Manning, 2003) and the Alpino4 (van Noord, 2006) parsers, and formalized the relations between terms (entities) as dependency paths. A dependency path is the shortest path of lexico-syntactic elements, i.e. shortest lexico-syntactic pattern, connecting entities (proper and common nouns) in their parsetrees. Such a formalization has been successfully employed in previous IE tasks (see Stevenson and Greenwood (2009) for an overview). Compared to traditional surface-pattern representations, used by Pantel and Pennacchiotti (2006), dependency paths abstract from surface texts to capture long range dependencies between terms. They also alleviate the manual authoring of large numbers of surface patterns. In our formalization, we substitute entities in the dependency paths with generic placeholders PART and WHOLE. Below, we show two dependency paths (1-b) and (2-b), respectively derived from English and Dutch Wikipedia sentences (1-a) and (2-a), and denoting the relations between sample-song, and alkalo¨ıde-plant. (1) a. The song “Mao Tse Tung Said” by Alabama 3 contains samples of a speech by Jim Jones b. WHOLE+nsubj ←contains →dobj+PART (2) a. Alle delen van de planten bevatten alkalo¨ıden en zijn daarmee giftig (All parts of the plants contain alkaloids and therefore are poisonous) b. WHOLE+obj1+van+mod+deel+su ← bevat→obj1+PART In our experiments, we only consider those entity pairs (tuples), patterns, and co-occuring pairspatterns with a minimum frequency of 10 in the English corpus, and 5 in the Dutch corpus. Statistics on the number of tuples and patterns preserved after applying the frequency cut-off are given in Table 1. 3.2 Information Extraction Algorithm As IE algorithm for extracting part-whole relations from our texts, we relied on Espresso, a minimally-supervised algorithm, as described by Pantel and Pennacchiotti (2006). They show 3http://nlp.stanford.edu/software/ lex-parser.shtml 4http://www.let.rug.nl/˜vannoord/alp/ Alpino 1330 English Dutch words 470.0 110.0 pairs 328.0 28.8 unique pairs 6.7 1.4 patterns 238.0 54.0 unique patterns 2.0 0.9 Table 1: Corpus Statistics in millions that the algorithm achieves state-of-the-art performance when initialized with relatively small seedsets over the Acquaint corpus (∼6M words). Recall is improved with web search queries as additional source of information. Espresso extracts surface patterns connecting the seeds (tuples) in a corpus. The reliability of a pattern p, r(p), given a set of input tuples I, is computed using (3), as its average strength of association with each tuple,i, weighted by each tuple’s reliability, rι(i). (3) rπ(p) = P i∈I pmi(i,p) maxpmi ×rι(i) |I| In this equation, pmi(i, p) is the pointwise mutual information score (Church and Hanks, 1990) between a pattern, p (e.g. consist-of), and a tuple, i (e.g. engine-car), and maxpmi is the maximum PMI score between all patterns and tuples. The reliability of the initializing seeds is set to 1. The top-k most reliable patterns are selected to find new tuples. The reliability of each tuple i, rι(i) is computed according to (4), where P is the set of harvested patterns. The top-m most reliable tuples are used to infer new patterns. (4) rι(i) = P i∈I pmi(i,p) maxpmi ×rπ(p) |P| The recursive discovery of patterns from tuples and vice-versa is repeated until a threshold number of patterns and/or tuples have been extracted. In our implementation, we maintain the core of the original Espresso algorithm, which pertains to estimating the reliability of patterns and tuples. Pantel and Pennacchiotti (2006) mention that their method is independent of the way patterns are formulated. Thus, instead of relying on surface patterns, we use dependency paths (as described above). Another difference is that while Pantel and Pennacchiotti (2006) complement their small corpus with documents retrieved from the web, we only rely on patterns extracted from our (much larger) corpora. Finally, we did not apply the discounting factor suggested in Pantel and Pennacchiotti (2006) to correct for the fact that PMI overestimates the importance of low-frequency events. Instead, as explained above, we applied a general frequency cut-off.5 3.3 Seed Selection Initially,we selected seeds from WordNet (Fellbaum, 1998) (for English) and EuroWordNet (Vossen, 1998) (for Dutch) to initialize the IE algorithm. However, we found that these pairs, such as acinos-mother of thyme or radarschermradarapparatuur (radar screen - radar equipment, hardly co-occured with reasonable frequency in Wikipedia sentences, hindering pattern extraction. We therefore adopted the following strategy. We searched our corpora for archetypal patterns, e.g. contain , which characterize all the different types of part-whole relations. The tuples sub-categorized by these patterns in the English texts were automatically6 typed to appropriate DOLCE ontology7 classes, corresponding to those employed by Keet and Artale for constraining the entity pairs participating in different types of partwhole relations. The types of part-whole relations instantiated by the tuples could then be determined based on their ontological classes. Separate sets of 20 tuples, with each set corresponding to a specific relation type in the taxonomy of Keet and Artale (Keet’s taxonomy), were then created. For example, the English Wikipedia tuple t1 =actor-cast was used as a seed to discover member-of partwhole relations since both its elements were typed to the SOCIAL OBJECT class of the DOLCE ontology, and according to Keet’s taxonomy, they instantiate a member-of relation. Seeds for extracting relations from the Dutch corpus were defined in a similar way, except that we manually determined their ontological classes based on the class glossary of DOLCE. Below, we only report on the member-of and sub-quantity-of meronymic relations, and on the located-in, contained-in and structural part-of mereological relations. We were unable to find sufficient seeds for the constituted-of meronymic 5We experimented with the suggested discounting factor for PMI, but were not able to improve over the accuracy scores reported later. 6Using the Java-OWL API, from http://protege. stanford.edu/plugins/owl/api/ 7OWL Version 0.72, downloaded from http://www. loa-cnr.it/DOLCE.html/ 1331 Lg Part Whole # Type EN grave church 155 contain NL beeld kerk 120 contain (statue) (church) EN city region 3735 located NL abdij gemeente 36 located (abbey) (community) EN actor cast 432 member NL club voetbal bond 178 member (club) (soccer union) EN engine car 3509 structural NL geheugen computer 14 structural (memory) (computer) EN alcohol wine 260 subquant NL alcohol bier 28 subquant (alcohol) (beer) Table 2: Seeds used for learning part-whole relations (contained-in, located-in, member-of, structural part-of, sub-quantity-of). relations (e.g. clay-statue). Also, we did not experiment with the participates-in and involved-in relations since their lexical realizations in our corpora are sparse, and they contain at least one verbal argument, whereas we only targeted patterns connecting nomimals. Sample seeds, their corpus frequency, and the part-whole relation type they instantiate from the English (EN) and Dutch (NL) corpora are illustrated in Table 2. Besides the five specialized seed-sets of 20 prototypical tuples for the aforementioned relations, we also defined a general set of mixed seeds, which combines four seeds from each of the specialized sets. 4 Experiments and Evaluation We initialized our IE algorithm with the seed-sets to extract part-whole relations from our corpora. The same parameters as Pantel and Pennacchiotti (2006) were used. That is, the 10 most reliable patterns inferred from the initial seeds are bootstrapped to induce 100 part-whole tuples. In each subsequent iteration, we learn one additional pattern and 100 additional tuples. We evaluated our results after 5 iterations since the performance in later iterations was almost constant. The results are discussed next. meronomic mereological memb subq cont struc locat gen EN 0.67 0.74 0.70 0.82 0.75 0.80 NL 0.68 0.60 0.60 0.60 0.70 0.71 Table 3: Precision for seed-sets representing specific types of part-whole relations (member-of, sub-quantity-of, contained-in, structural part-of and located-in), and for the general set composed of all types. 4.1 Precision of Extracted Relations Two human judges manually evaluated the tuples extracted from the English and Dutch corpora per seed-set in each iteration of our algorithm. Tuples that unambiguously instantiated part-whole relations were considered true positives. Those that did not were considered false positives. Ambiguous tuples were discarded. The precision of the tuples discovered by the different seed-sets in the last iteration of our algorithm are in Table 3. These results reveal that the precision of harvested tuples varies depending on the part-whole relation type that the initializing seeds denote. Mereological seeds (cont, struct, locat sets) outperformed their meronymic counterparts (memb, subq) in extracting relations with higher precision from the English texts. This could be attributed to their formal ontological grounding, making them less ambiguous than the linguistically-motivated meronymic relations (Keet, 2006; Keet and Artale, 2008). The precision variations were less discernible for tuples extracted from the Dutch corpus, although the best precision was still achieved with mereological located-in seeds. We also noticed that the precision of tuples extracted from both the English and Dutch corpora by the general set of mixed seeds was as high as the maximum precision obtained by the individual sets of specialized seeds over these two corpora, i.e. 0.80 (general seeds) vs. 0.82 (structural partof seeds) for English, and 0.71 (general seeds) vs. 0.70 (located-in seeds) for Dutch. Based on these findings, we address our first question, and conclude that 1) the type of relation instantiated by the initializing seeds affects the performance of IE algorithms, with mereological seeds being in general more fertile than their meronymic counterparts, and generating higher-precision tuples; 2) the precision achieved when initializing IE algorithms with a general set, which mixes 1332 seeds of heterogeneous part-whole relation types, is comparable to the best results obtained with individual sets of specialized seeds, denoting specific part-whole relations. An evaluation of the patterns and tuples extracted indicated considerable precision drop between successive iterations of our algorithm. This appears to be due to semantic drift (McIntosh and Curran, 2009), where highly-ambiguous patterns promote incorrect tuples , which in turn, compound the precision loss. 4.2 Types of Extracted Relations Initializing our algorithm with seeds of a particular type always led to the discovery of tuples characterizing other types of part-whole relations in the English corpus. This can be explained by prototypical patterns, e.g. “include”, generated regardless of the seeds’ types, and which are highy correlated with, and hence, trigger tuples denoting other part-whole relation types. An almost similar observation was made for the Dutch corpus, except that tuples instantiating the member-of relation could only be learnt using initial seeds of that particular type (i.e. member-of). Upon inspecting our results, it was found that this phenomenon was due to the distinct and specific patterns, such as “treedt toe tot” (“become member of”), which linguistically realize the member-of relations in the Dutch corpus. Thus, initializing our IE algorithm with seeds that instantiate relations other than member-of fails to detect these unique patterns, and fails to subsequently discover partwhole tuples describing the member-of relations. Our findings are illustrated in Table 4, where each cell lists a tuple of a particular type (column), which was harvested from seeds of a given type (row). These results answer our second question. 4.3 Distinct Patterns and Tuples We address our third question by comparing the output of our algorithm to determine whether the results obtained by initializing with the individual specialized seeds were (dis)similar and/or distinct. Each result set consisted of maximally 520 tuples (including 20 initializing seeds) and 15 lexicosyntactic patterns, obtained after five iterations. Tuples extracted from the English corpus using the member-of and contained-in seed-sets exhibited a high degree of similarity, with 465 common tuples discovered by both sets. These identical tuples were also assigned the same ranks (reliability) in the results generated by the memberof and contained-in seeds, with a Spearman rank correlation of 0.82 between their respective outputs. This convergence was also reflected in the fact that the member-of and contained-in seeds generated around 80% of common patterns. These patterns were mostly prototypical ones indicative of part-whole relations, such as WHOLE+nsubj ←include →dobj+PART (“include”) and their cognates involving passive forms and relative clauses. However, the specialized seeds also generated distinct patterns, like “joined as” and “released with” for the member-of and contained-in seeds respectively. The most distinct tuples and patterns were harvested with the sub-quantity-of, structural part-of, and located-in seeds. Negative Spearman correlation scores were obtained when comparing the results of these three sets among themselves, and with the results of the member-of and containedin seeds, indicating insignificant similarity and overlap. Examining the patterns harvested by the sub-quantity-of, structural part-of, and located-in seeds revealed a high prominence of specialized and unique patterns, which specifically characterize these relations. Examples of such patterns include “made with”, “released with” and “found in”, which lexically realize the sub-quantity-of, structural part-of, and located-in relations respectively. For the Dutch corpus, the seeds that generated the most similar tuples were those corresponding to the sub-quantity-of, contained-in, and structural part-of relations, with 490 common tuples discovered, and a Spearman rank correlation in the range of 0.89-0.93 between their respective outputs. As expected, these seeds also led to the discovery of a substantial number of common and prototypical part-whole patterns. Examples include “bevat” (“contain”), “omvat” (“comprise”), and their variants. The most distinct results were harvested by the located-in and member-of seeds, with negative Spearman correlation scores between the output tuples indicating hardly any overlap. We also found out that the patterns harvested by the located-in and member-of seeds characteristically pertained to these relations. Example of such patterns include “ligt in” (“lie in”), “is gelegen in” (“is located in”), and “treedt toe tot” (“become member of”), respectively describing the located-in and member-of relations. Thus, we observed that 1) tuples harvested from 1333 meronomic mereological Tuples→ member subquant contained struct located Seeds↓ EN member ship-convoy alcohol-wine card-deck proton-nucleus lake-park subquant aircraft-fleet moisture-soil building-complex engine-car commune-canton contained aircraft-fleet alcohol-wine relic-church base-spacecraft campus-city structural brother-family mineral-bone library-building inlay-fingerboard hamlet-town located performer-cast alcohol-blood artifact-museum chassis-car city-shore NL member sporter-ploeg helium-atmosfeer stalagmieten-grot shirt-tenue boerderij-dorp (athlete-team) (helium-atmosphere) (stalagnites-cave) (shirt-outfit) (farm-village) subquant — vet-kaas pijp orgel-kerk kam-gitaar paleis-stad (fat-cheese) (pipe-organ-church) (bridge-guitar) (palace-city) contained — tannine-wijn kamer-toren atoom-molecule paleis-stad (tannine-wine) (room-tower) (atom-molecule) (palace-city) structural — kinine-tonic beeld-kerk wervel-ruggengraat paleis-stad (quinine-tonic) statue-church) (vertebra-backbone) (palace-city) located — — kunst werk-kathedraal poort-muur metro station-wijk (work of art-cathedral) (gate-wall) (metro station-quarter) Table 4: Sample tuples found per relation type. both the English and Dutch corpora by seeds instantiating a single particular type of part-whole relation highly correlated with tuples discovered by at least one other type of seeds (member-of and contained-in for English, and sub-quantityof, contained-in and structural part-of for Dutch); 2) some part-whole relations are manifested by a wide variety of specialized patterns (sub-quantityof, structural part-of, and located-in for English, and located-in and member-of for Dutch). Finally, instead of a single set that mixes seeds of different types, we created five such general sets by picking four different seeds from each of the specialized sets, and used them to initialize our algorithm. When examining the results of each of the five general sets, we found out that they were unstable, and always correlated with the output of a different specialized set. Based on these findings, we believe that the traditional practice of initializing IE algorithms with general sets that mix seeds denoting different partwhole relation types leads to inherently unstable results. As we have shown, the relations extracted by combining seeds of heterogeneous types almost always converge to one specific part-whole relation type, which cannot be conclusively predicted. Furthermore, general seeds are unable to capture the specific and distinct patterns that lexically realize the individual types of part-whole relations. 5 Conclusions In this paper, we have investigated the effect of ontologically-motivated distinctions in part-whole relations on IE systems that learn instances of these relations from text. We have shown that learning from specialized seeds-sets, denoting specific types of the partwhole relations, results in precision that is as high as or higher than the precision achieved with a general set that mixes seeds of different types. By comparing the outputs generated by different seed-sets, we observed that the tuples learnt with seeds denoting a specific part-whole relation type are not confined to that particular type. In most case, we are still able to discover tuples across all the different types of part-whole relations, regardless of the type instantiated by the initializing seeds. Most importantly, we demonstrated that IE algorithms initialized with general sets of mixed seeds harvest results that tend to converge towards a specific type of part-whole relation. Conversely, when starting with seeds representing a specific type, it is likely to discover tuples and patterns that are completely distinct from those found by a mixed seed-set. Our results also illustrate that the outputs of IE algorithms are heavily influenced by the initializing seeds, concurring with the findings of McIntosh and Curran (2009). We believe that our results show a drastic form of this phenomenon: given a set of mixed seeds, denoting heterogeneous relations, the harvested tuples may converge towards any of the relations instantiated by the seeds. Predicting the convergent relation is in usual cases impossible, and may depend on factors pertaining to corpus characteristics. This instability strongly suggests that seeds instantiating different types of relations should not be mixed, partic1334 ularly when learning part-whole relations, which are characterized by many subtypes. Seeds should be defined such that they represent an ontologically well-defined class, for which one may hope to find a coherent set of extraction patterns. Acknowledgement Ashwin Ittoo is part of the project “Merging of Incoherent Field Feedback Data into Prioritized Design Information (DataFusion)” (http://www. iopdatafusion.org//), sponsored by the Dutch Ministry of Economic Affairs under the IOP-IPCR program. Gosse Bouma acknowledges support from the Stevin LASSY project (www.let.rug.nl/ ˜vannoord/Lassy/). References A. Artale, E. Franconi, N. Guarino, and L. Pazzi. 1996. Part-whole relations in object-centered systems: An overview. Data & Knowledge Engineering, 20(3):347–383. B. Beamer, A. Rozovskaya, and R. Girju. 2008. Automatic semantic relation extraction with multiple boundary generation. In Proceedings of the 23rd national conference on Artificial intelligence-Volume 2, pages 824–829. AAAI Press. Matthew Berland and Eugene Charniak. 1999. Finding parts in very large corpora. In Proceedings of the 37th annual meeting of the Association for Computational Linguistics on Computational Linguistics, pages 57–64, Morristown, NJ, USA. Association for Computational Linguistics. K.W. Church and P. Hanks. 1990. Word association norms, mutual information, and lexicography. Computational linguistics, 16(1):22–29. T. Dunning. 1993. Accurate methods for the statistics of surprise and coincidence. Computational linguistics, 19(1):74. Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database. MIT, Cambridge. A. Gangemi, N. Guarino, C. Masolo, A. Oltramari, and L. Schneider. 2002. Sweetening ontologies with DOLCE. Knowledge Engineering and Knowledge Management: Ontologies and the Semantic Web, Lecture Notes in Computer Science, pages 223–233. P. Gerstl and S. Pribbenow. 1995. Midwinters, end games, and body parts: a classification of part-whole relations. International Journal of Human Computer Studies, 43:865–890. R. Girju, A. Badulescu, and D. Moldovan. 2003. Learning semantic constraints for the automatic discovery of part-whole relations. In Proceedings of HLT/NAACL, volume 3, pages 80–87. R. Girju, A. Badulescu, and D. Moldovan. 2006. Automatic discovery of part-whole relations. Computational Linguistics, 32(1):83–135. M.A. Hearst. 1992. Automatic acquisition of hyponyms from large text corpora. In Proceedings of the 14th conference on Computational linguisticsVolume 2, pages 539–545. Association for Computational Linguistics Morristown, NJ, USA. C.M. Keet and A. Artale. 2008. Representing and reasoning over a taxonomy of part–whole relations. Applied Ontology, 3(1):91–110. C.M. Keet. 2006. Part-whole relations in objectrole models. On the Move to Meaningful Internet Systems 2006, Lecture Notes in Computer Science, 4278:1118–1127. D. Klein and C.D. Manning. 2003. Accurate unlexicalized parsing. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics-Volume 1, pages 423–430. Association for Computational Linguistics Morristown, NJ, USA. T. McIntosh and J.R. Curran. 2009. Reducing semantic drift with bagging and distributional similarity. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Confe rence on Natural Language Processing of the AFNLP, pages 396–404. G.A. Miller, C. Leacock, R. Tengi, and R.T. Bunker. 1993. A semantic concordance. In Proceedings of the 3rd DARPA workshop on Human Language Technology, pages 303–308. New Jersey. D.P.T. Nguyen, Y. Matsuo, and M. Ishizuka. 2007. Relation extraction from wikipedia using subtree mining. In Proceedings of the National Conference on Artificial Intelligence, volume 22, page 1414. Menlo Park, CA; Cambridge, MA; London; AAAI Press; MIT Press; 1999. J. Odell. 1994. Six different kinds of composition. Journal of Object-Oriented Programming, 5(8):10– 15. Patrick Pantel and Marco Pennacchiotti. 2006. Espresso: Leveraging generic patterns for automatically harvesting semantic relations. In Proceedings of Conference on Computational Linguistics / Association for Computational Linguistics (COLING/ACL-06), pages 113–120, Sydney, Australia. S. Pyysalo, T. Ohta, J.D. Kim, and J. Tsujii. 2009. Static relations: a piece in the biomedical information extraction puzzle. In Proceedings of the Workshop on BioNLP, pages 1–9. Association for Computational Linguistics. 1335 Mark Stevenson and Mark Greenwood. 2009. Dependency pattern models for information extraction. Research on Language and Computation, 3:13–39. W.R. Van Hage, H. Kolb, and G. Schreiber. 2006. A method for learning part-whole relations. The Semantic Web - ISWC 2006, Lecture Notes in Computer Science, 4273:723–735. Gertjan van Noord. 2006. At last parsing is now operational. In Piet Mertens, Cedrick Fairon, Anne Dister, and Patrick Watrin, editors, TALN06. Verbum Ex Machina. Actes de la 13e conference sur le traitement automatique des langues naturelles, pages 20– 42. Presses univ. de Louvain. P. Vossen, editor. 1998. EuroWordNet A Multilingual Database with Lexical Semantic Networks. Kluwer Academic publishers. M.E. Winston, R. Chaffin, and D. Herrmann. 1987. A taxonomy of part-whole relations. Cognitive science, 11(4):417–444. F. Wu and D.S. Weld. 2007. Autonomously semantifying wikipedia. In Proceedings of the sixteenth ACM conference on Conference on information and knowledge management, pages 41–50. ACM. 1336
|
2010
|
135
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1337–1345, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Understanding the Semantic Structure of Noun Phrase Queries Xiao Li Microsoft Research One Microsoft Way Redmond, WA 98052 USA [email protected] Abstract Determining the semantic intent of web queries not only involves identifying their semantic class, which is a primary focus of previous works, but also understanding their semantic structure. In this work, we formally define the semantic structure of noun phrase queries as comprised of intent heads and intent modifiers. We present methods that automatically identify these constituents as well as their semantic roles based on Markov and semi-Markov conditional random fields. We show that the use of semantic features and syntactic features significantly contribute to improving the understanding performance. 1 Introduction Web queries can be considered as implicit questions or commands, in that they are performed either to find information on the web or to initiate interaction with web services. Web users, however, rarely express their intent in full language. For example, to find out “what are the movies of 2010 in which johnny depp stars”, a user may simply query “johnny depp movies 2010”. Today’s search engines, generally speaking, are based on matching such keywords against web documents and ranking relevant results using sophisticated features and algorithms. As search engine technologies evolve, it is increasingly believed that search will be shifting away from “ten blue links” toward understanding intent and serving objects. This trend has been largely driven by an increasing amount of structured and semi-structured data made available to search engines, such as relational databases and semantically annotated web documents. Searching over such data sources, in many cases, can offer more relevant and essential results compared with merely returning web pages that contain query keywords. Table 1 shows a simplified view of a structured data source, where each row represents a movie object. Consider the query “johnny depp movies 2010”. It is possible to retrieve a set of movie objects from Table 1 that satisfy the constraints Year = 2010 and Cast ∋ Johnny Depp. This would deliver direct answers to the query rather than having the user sort through list of keyword results. In no small part, the success of such an approach relies on robust understanding of query intent. Most previous works in this area focus on query intent classification (Shen et al., 2006; Li et al., 2008b; Arguello et al., 2009). Indeed, the intent class information is crucial in determining if a query can be answered by any structured data sources and, if so, by which one. In this work, we go one step further and study the semantic structure of a query, i.e., individual constituents of a query and their semantic roles. In particular, we focus on noun phrase queries. A key contribution of this work is that we formally define query semantic structure as comprised of intent heads (IH) and intent modifiers (IM), e.g., [IM:Title alice in wonderland] [IM:Year 2010] [IH cast] It is determined that “cast” is an IH of the above query, representing the essential information the user intends to obtain. Furthermore, there are two IMs, “alice in wonderland” and “2010”, serving as filters of the information the user receives. Identifying the semantic structure of queries can be beneficial to information retrieval. Knowing the semantic role of each query constituent, we 1337 Title Year Genre Director Cast Review Precious 2009 Drama Lee Daniels Gabby Sidibe, Mo’Nique,. . . 2012 2009 Action, Sci Fi Roland Emmerich John Cusack, Chiwetel Ejiofor,. . . Avatar 2009 Action, Sci Fi James Cameron Sam Worthington, Zoe Saldana,. . . The Rum Diary 2010 Adventure, Drama Bruce Robinson Johnny Depp,Giovanni Ribisi,. . . Alice in Wonderland 2010 Adventure, Family Tim Burton Mia Wasikowska, Johnny Depp,. . . Table 1: A simplified view of a structured data source for the Movie domain. can reformulate the query into a structured form or reweight different query constituents for structured data retrieval (Robertson et al., 2004; Kim et al., 2009; Paparizos et al., 2009). Alternatively, the knowledge of IHs, IMs and semantic labels of IMs may be used as additional evidence in a learning to rank framework (Burges et al., 2005). A second contribution of this work is to present methods that automatically extract the semantic structure of noun phrase queries, i.e., IHs, IMs and the semantic labels of IMs. In particular, we investigate the use of transition, lexical, semantic and syntactic features. The semantic features can be constructed from structured data sources or by mining query logs, while the syntactic features can be obtained by readily-available syntactic analysis tools. We compare the roles of these features in two discriminative models, Markov and semiMarkov conditional random fields. The second model is especially interesting to us since in our task it is beneficial to use features that measure segment-level characteristics. Finally, we evaluate our proposed models and features on manuallyannotated query sets from three domains, while our techniques are general enough to be applied to many other domains. 2 Related Works 2.1 Query intent understanding As mentioned in the introduction, previous works on query intent understanding have largely focused on classification, i.e., automatically mapping queries into semantic classes (Shen et al., 2006; Li et al., 2008b; Arguello et al., 2009). There are relatively few published works on understanding the semantic structure of web queries. The most relevant ones are on the problem of query tagging, i.e., assigning semantic labels to query terms (Li et al., 2009; Manshadi and Li, 2009). For example, in “canon powershot sd850 camera silver”, the word “canon” should be tagged as Brand. In particular, Li et al. leveraged clickthrough data and a database to automatically derive training data for learning a CRF-based tagger. Manshadi and Li developed a hybrid, generative grammar model for a similar task. Both works are closely related to one aspect of our work, which is to assign semantic labels to IMs. A key difference is that they do not conceptually distinguish between IHs and IMs. On the other hand, there have been a series of research studies related to IH identification (Pasca and Durme, 2007; Pasca and Durme, 2008). Their methods aim at extracting attribute names, such as cost and side effect for the concept Drug, from documents and query logs in a weakly-supervised learning framework. When used in the context of web queries, attribute names usually serve as IHs. In fact, one immediate application of their research is to understand web queries that request factual information of some concepts, e.g. “asiprin cost” and “aspirin side effect”. Their framework, however, does not consider the identification and categorization of IMs (attribute values). 2.2 Question answering Query intent understanding is analogous to question understanding for question answering (QA) systems. Many web queries can be viewed as the keyword-based counterparts of natural language questions. For example, the query “california national” and “national parks califorina” both imply the question “What are the national parks in California?”. In particular, a number of works investigated the importance of head noun extraction in understanding what-type questions (Metzler and Croft, 2005; Li et al., 2008a). To extract head nouns, they applied syntax-based rules using the information obtained from part-of-speech (POS) tagging and deep parsing. As questions posed in natural language tend to have strong syntactic structures, such an approach was demonstrated to be accurate in identifying head nouns. In identifying IHs in noun phrase queries, however, direct syntactic analysis is unlikely to be as effective. This is because syntactic structures are in general less pronounced in web queries. In this 1338 work, we propose to use POS tagging and parsing outputs as features, in addition to other features, in extracting the semantic structure of web queries. 2.3 Information extraction Finally, there exist large bodies of work on information extraction using models based on Markov and semi-Markov CRFs (Lafferty et al., 2001; Sarawagi and Cohen, 2004), and in particular for the task of named entity recognition (McCallum and Li, 2003). The problem studied in this work is concerned with identifying more generic “semantic roles” of the constituents in noun phrase queries. While some IM categories belong to named entities such as IM:Director for the intent class Movie, there can be semantic labels that are not named entities such as IH and IM:Genre (again for Movie). 3 Query Semantic Structure Unlike database query languages such as SQL, web queries are usually formulated as sequences of words without explicit structures. This makes web queries difficult to interpret by computers. For example, should the query “aspirin side effect” be interpreted as “the side effect of aspirin” or “the aspirin of side effect”? Before trying to build models that can automatically makes such decisions, we first need to understand what constitute the semantic structure of a noun phrase query. 3.1 Definition We let C denote a set of query intent classes that represent semantic concepts such as Movie, Product and Drug. The query constituents introduced below are all defined w.r.t. the intent class of a query, c ∈C, which is assumed to be known. Intent head An intent head (IH) is a query segment that corresponds to an attribute name of an intent class. For example, the IH of the query “alice in wonderland 2010 cast” is “cast”, which is an attribute name of Movie. By issuing the query, the user intends to find out the values of the IH (i.e., cast). A query can have multiple IHs, e.g., “movie avatar director and cast”. More importantly, there can be queries without an explicit IH. For example, “movie avatar” does not contain any segment that corresponds to an attribute name of Movie. Such a query, however, does have an implicit intent which is to obtain general information about the movie. Intent modifier In contrast, an intent modifier (IM) is a query segment that corresponds to an attribute value (of some attribute name). The role of IMs is to imposing constraints on the attributes of an intent class. For example, there are two constraints implied in the query “alice in wonderland 2010 cast”: (1) the Title of the movie is “alice in wonderland”; and (2) the Year of the movie is “2010”. Interestingly, the user does not explicitly specify the attribute names, i.e., Title and Year, in this query. Such information, however, can be inferred given domain knowledge. In fact, one important goal of this work is to identify the semantic labels of IMs, i.e., the attribute names they implicitly refer to. We use Ac to denote the set of IM semantic labels for the intent class c. Other Additionally, there can be query segments that do not play any semantic roles, which we refer to as Other. 3.2 Syntactic analysis The notion of IHs and IMs in this work is closely related to that of linguistic head nouns and modifiers for noun phrases. In many cases, the IHs of noun phrase queries are exactly the head nouns in the linguistic sense. Exceptions mostly occur in queries without explicit IHs, e.g., “movie avatar” in which the head noun “avatar” serves as an IM instead. Due to the strong resemblance, it is interesting to see if IHs can be identified by extracting linguistic head nouns from queries based on syntactic analysis. To this end, we apply the following heuristics for head noun extraction. We first run a POS-tagger and a chunker jointly on each query, where the POS-tagger/chunker is based on an HMM system trained on English Penn Treebank (Gao et al., 2001). We then mark the right most NP chunk before any prepositional phrase or adjective clause, and apply the NP head rules (Collins, 1999) to the marked NP chunk. The main problem with this approach, however, is that a readily-available POS tagger or chunker is usually trained on natural language sentences and thus is unlikely to produce accurate results on web queries. As shown in (Barr et al., 2008), the lexical category distribution of web queries is dramatically different from that of natural languages. For example, prepositions and subordinating conjunctions, which are strong indicators of the syntactic 1339 structure in natural languages, are often missing in web queries. Moreover, unlike most natural languages that follow the linear-order principle, web queries can have relatively free word orders (although some orders may occur more often than others statistically). These factors make it difficult to produce reliable syntactic analysis outputs. Consequently, the head nouns and hence the IHs extracted therefrom are likely to be error-prone, as will be shown by our experiments in Section 6.3. Although a POS tagger and a chunker may not work well on queries, their output can be used as features for learning statistical models for semantic structure extraction, which we introduce next. 4 Models This section presents two statistical models for semantic understanding of noun phrase queries. Assuming that the intent class c ∈C of a query is known, we cast the problem of extracting the semantic structure of the query into a joint segmentation/classification problem. At a high level, we would like to identify query segments that correspond to IHs, IMs and Others. Furthermore, for each IM segment, we would like to assign a semantic label, denoted by IM:a, a ∈Ac, indicating which attribute name it refers to. In other words, our label set consists of Y = {IH, {IM:a}a∈Ac, Other}. Formally, we let x = (x1, x2, . . . , xM) denote an input query of length M. To avoid confusion, we use i to represent the index of a word token and j to represent the index of a segment in the following text. Our goal is to obtain s∗= argmax s p(s|c, x) (1) where s = (s1, s2, . . . , sN) denotes a query segmentation as well as a classification of all segments. Each segment sj is represented by a tuple (uj, vj, yj). Here uj and vj are the indices of the starting and ending word tokens respectively; yj ∈Y is a label indicating the semantic role of s. We further augment the segment sequence with two special segments: Start and End, represented by s0 and sN+1 respectively. For notional simplicity, we assume that the intent class is given and use p(s|x) as a shorthand for p(s|c, x), but keep in mind that the label space and hence the parameter space is class-dependent. Now we introduce two methods of modeling p(s|x). 4.1 CRFs One natural approach to extracting the semantic structure of queries is to use linear-chain CRFs (Lafferty et al., 2001). They model the conditional probability of a label sequence given the input, where the labels, denoted as y = (y1, y2, . . . , yM), yi ∈Y, have a one-to-one correspondence with the word tokens in the input. Using linear-chain CRFs, we aim to find the label sequence that maximizes pλ(y|x) = 1 Zλ(x) exp (M+1 X i=1 λ · f(yi−1, yi, x, i) ) . (2) The partition function Zλ(x) is a normalization factor. λ is a weight vector and f(yi−1, yi, x) is a vector of feature functions referred to as a feature vector. The features used in CRFs will be described in Section 5. Given manually-labeled queries, we estimate λ that maximizes the conditional likelihood of training data while regularizing model parameters. The learned model is then used to predict the label sequence y for future input sequences x. To obtain s in Equation (1), we simply concatenate the maximum number of consecutive word tokens that have the same label and treat the resulting sequence as a segment. By doing this, we implicitly assume that there are no two adjacent segments with the same label in the true segment sequence. Although this assumption is not always correct in practice, we consider it a reasonable approximation given what we empirically observed in our training data. 4.2 Semi-Markov CRFs In contrast to standard CRFs, semi-Markov CRFs directly model the segmentation of an input sequence as well as a classification of the segments (Sarawagi and Cohen, 2004), i.e., p(s|x) = 1 Zλ(x) exp N+1 X j=1 λ · f(sj−1, sj, x) (3) In this case, the features f(sj−1, sj, x) are defined on segments instead of on word tokens. More precisely, they are of the function form f(yj−1, yj, x, uj, vj). It is easy to see that by imposing a constraint ui = vi, the model is reduced to standard linear-chain CRFs. SemiMarkov CRFs make Markov assumptions at the segment level, thereby naturally offering means to 1340 CRF features A1: Transition δ(yi−1 = a)δ(yi = b) transiting from state a to b A2: Lexical δ(xi = w)δ(yi = b) current word is w A3: Semantic δ(xi ∈WL)δ(yi = b) current word occurs in lexicon L A4: Semantic δ(xi−1:i ∈WL)δ(yi = b) current bigram occurs in lexicon L A5: Syntactic δ(POS(xi) = z)δ(yi = b) POS tag of the current word is z Semi-Markov CRF features B1: Transition δ(yj−1 = a)δ(yj = b) Transiting from state a to b B2: Lexical δ(xuj:vj = w)δ(yj = b) Current segment is w B3: Lexical δ(xuj:vj ∋w)δ(yj = b) Current segment contains word w B4: Semantic δ(xuj:vj ∈L)δ(yj = b) Current segment is an element in lexicon L B5: Semantic max l∈L s(xuj:vj, l)δ(yj = b) The max similarity between the segment and elements in L B6: Syntactic δ(POS(xuj:vj) = z)δ(yj = b) Current segment’s POS sequence is z B7: Syntactic δ(Chunk(xuj:vj) = c)δ(yj = b) Current segment is a chunk with phrase type c Table 2: A summary of feature types in CRFs and segmental CRFs for query understanding. We assume that the state label is b in all features and omit this in the feature descriptions. incorporate segment-level features, as will be presented in Section 5. 5 Features In this work, we explore the use of transition, lexical, semantic and syntactic features in Markov and semi-Markov CRFs. The mathematical expression of these features are summarized in Table 2 with details described as follows. 5.1 Transition features Transition features, i.e., A1 and B1 in Table 2, capture state transition patterns between adjacent word tokens in CRFs, and between adjacent segments in semi-Markov CRFs. We only use firstorder transition features in this work. 5.2 Lexical features In CRFs, a lexical feature (A2) is implemented as a binary function that indicates whether a specific word co-occurs with a state label. The set of words to be considered in this work are those observed in the training data. We can also generalize this type of features from words to n-grams. In other words, instead of inspecting the word identity at the current position, we inspect the n-gram identity by applying a window of length n centered at the current position. Since feature functions are defined on segments in semi-Markov CRFs, we create B2 that indicates whether the phrase in a hypothesized query segment co-occurs with a state label. Here the set of phrase identities are extracted from the query segments in the training data. Furthermore, we create another type of lexical feature, B3, which is activated when a specific word occurs in a hypothesized query segment. The use of B3 would favor unseen words being included in adjacent segments rather than to be isolated as separate segments. 5.3 Semantic features Models relying on lexical features may require very large amounts of training data to produce accurate prediction performance, as the feature space is in general large and sparse. To make our model generalize better, we create semantic features based on what we call lexicons. A lexicon, denoted as L, is a cluster of semantically-related words/phrases. For example, a cluster of movie titles or director names can be such a lexicon. Before describing how such lexicons are generated for our task, we first introduce the forms of the semantic features assuming the availability of the lexicons. We let L denote a lexicon, and WL denote the set of n-grams extracted from L. For CRFs, we create a binary function that indicates whether any n-gram in WL co-occurs with a state label, with n = 1, 2 for A3, A4 respectively. For both A3 and A4, the number of such semantic features is equal to the number of lexicons multiplied by the number of state labels. The same source of semantic knowledge can be conveniently incorporated in semi-Markov CRFs. One set of semantic features (B4) inspect whether the phrase of a hypothesized query segment matches any element in a given lexicon. A second set of semantic features (B5) relax the exact match constraints made by B4, and take as the feature value the maximum “similarity” between the query segment and all lexicon elements. The fol1341 lowing similarity function is used in this work , s(xuj:vj, l) = 1 −Lev(xuj:vj, l)/|l| (4) where Lev represents the Levenshtein distance. Notice that we normalize the Levenshtein distance by the length of the lexicon element, as we empirically found it performing better compared with normalizing by the length of the segment. In computing the maximum similarity, we first retrieve a set of lexicon elements with a positive tf-idf cosine distance with the segment; we then evaluate Equation (4) for each retrieved element and find the one with the maximum similarity score. Lexicon generation To create the semantic features described above, we generate two types of lexicons leveraging databases and query logs for each intent class. The first type of lexicon is an IH lexicon comprised of a list of attribute names for the intent class, e.g., “box office” and “review” for the intent class Movie. One easy way of composing such a list is by aggregating the column names in the corresponding database such as Table 1. However, this approach may result in low coverage on IHs for some domains. Moreover, many database column names, such as Title, are unlikely to appear as IHs in queries. Inspired by Pasca and Van Durme (2007), we apply a bootstrapping algorithm that automatically learns attribute names for an intent class from query logs. The key difference from their work is that we create templates that consist of semantic labels at the segment level from training data. For example, “alice in wonderland 2010 cast” is labeled as “IM:Title IM:Year IH”, and thus “IM:Title + IM:Year + #” is used as a template. We select the most frequent templates (top 2 in this work) from training data and use them to discover new IH phrases from the query log. Secondly, we have a set IM lexicons, each comprised of a list of attribute values of an attribute name in Ac. We exploit internal resources to generate such lexicons. For example, the lexicon for IM:Title (in Movie) is a list of movie titles generated by aggregating the values in the Title column of a movie database. Similarly, the lexicon for IM:Employee (in Job) is a list of employee names extracted from a job listing database. Note that a substantial amount of research effort has been dedicated to automatic lexicon acquisition from the Web (Pantel and Pennacchiotti, 2006; Pennacchiotti and Pantel, 2009). These techniques can be used in expanding the semantic lexicons for IMs when database resources are not available. But we do not use such techniques in our work since the lexicons extracted from databases in general have good precision and coverage. 5.4 Syntactic features As mentioned in Section 3.2, web queries often lack syntactic cues and do not necessarily follow the linear order principle. Consequently, applying syntactic analysis such as POS tagging or chunking using models trained on natural language corpora is unlikely to give accurate results on web queries, as supported by our experimental evidence in Section 6.3. It may be beneficial, however, to use syntactic analysis results as additional evidence in learning. To this end, we generate a sequence of POS tags for a given query, and use the co-occurrence of POS tag identities and state labels as syntactic features (A5) for CRFs. For semi-Markov CRFs, we instead examine the POS tag sequence of the corresponding phrase in a query segment. Again their identities are combined with state labels to create syntactic features B6. Furthermore, since it is natural to incorporate segment-level features in semi-Markov CRFs, we can directly use the output of a syntactic chunker. To be precise, if a query segment is determined by the chunker to be a chunk, we use the indicator of the phrase type of the chunk (e.g., NP, PP) combined with a state label as the feature, denoted by B7 in the Table. Such features are not activated if a query segment is determined not to be a chunk. 6 Evaluation 6.1 Data To evaluate our proposed models and features, we collected queries from three domains, Movie, Job and National Park, and had them manually annotated. The annotation was given on both segmentation of the queries and classification of the segments according to the label sets defined in Table 3. There are 1000/496 samples in the training/test set for the Movie domain, 600/366 for the Job domain and 491/185 for the National Park domain. In evaluation, we report the test-set performance in each domain as well as the average performance (weighted by their respectively test-set size) over all domains. 1342 Movie Job National Park IH trailer, box office IH listing, salary IH lodging, calendar IM:Award oscar best picture IM:Category engineering IM:Category national forest IM:Cast johnny depp IM:City las vegas IM:City page IM:Character michael corleone IM:County orange IM:Country us IM:Category tv series IM:Employer walmart IM:Name yosemite IM:Country american IM:Level entry level IM:POI volcano IM:Director steven spielberg IM:Salary high-paying IM:Rating best IM:Genre action IM:State florida IM:State flordia IM:Rating best IM:Type full time IM:Title the godfather Other the, in, that Other the, in, that Other the, in, that Table 3: Label sets and their respective query segment examples for the intent class Movie, Job and National Park. 6.2 Metrics There are two evaluation metrics used in our work: segment F1 and sentence accuracy (Acc). The first metric is computed based on precision and recall at the segment level. Specifically, let us assume that the true segment sequence of a query is s = (s1, s2, . . . , sN), and the decoded segment sequence is s′ = (s′ 1, s′ 2, . . . , s′ K). We say that s′ k is a true positive if s′ k ∈s. The precision and recall, then, are measured as the total number of true positives divided by the total number of decoded and true segments respectively. We report the F1-measure which is computed as 2 · prec · recall/(prec + recall). Secondly, a sentence is correct if all decoded segments are true positives. Sentence accuracy is measured by the total number of correct sentences divided by the total number of sentences. 6.3 Results We start with models that incorporate first-order transition features which are standard for both Markov and semi-Markov CRFs. We then experiment with lexical features, semantic features and syntactic features for both models. Table 4 and Table 5 give a summarization of all experimental results. Lexical features The first experiment we did is to evaluate the performance of lexical features (combined with transition features). This involves the use of A2 in Table 2 for CRFs, and B2 and B3 for semi-Markov CRFs. Note that adding B3, i.e., indicators of whether a query segment contains a word identity, gave an absolute 7.0%/3.2% gain in sentence accuracy and segment F1 on average, as shown in the row B1-B3 in Table 5. For both A2 and B3, we also tried extending the features based on word IDs to those based on n-gram IDs, where n = 1, 2, 3. This greatly increased the number of lexical features but did not improve learning performance, most likely due to the limited amounts of training data coupled with the sparsity of such features. In general, lexical features do not generalize well to the test data, which accounts for the relatively poor performance of both models. Semantic features We created IM lexicons from three in-house databases on Movie, Job and National Parks. Some lexicons, e.g., IM:State, are shared across domains. Regarding IH lexicons, we applied the bootstrapping algorithm described in Section 5.3 to a 1-month query log of Bing. We selected the most frequent 57 and 131 phrases to form the IH lexicons for Movie and National Park respectively. We do not have an IH lexicon for Job as the attribute names in that domain are much fewer and are well covered by training set examples. We implemented A3 and A4 for CRFs, which are based on the n-gram sets created from lexicons; and B4 and B5 for semi-Markov CRFs, which are based on exact and fuzzy match with lexicon items. As shown in Table 4 and 5, drastic increases in sentence accuracies and F1-measures were observed for both models. Syntactic features As shown in the row A1-A5 in Table 4, combined with all other features, the syntactic features (A5) built upon POS tags boosted the CRF model performance. Table 6 listed the most dominant positive and negative features based on POS tags for Movie (features for the other two domains are not reported due to space limit). We can see that many of these features make intuitive sense. For 1343 Movie Job National Park Average Features Acc F1 Acc F1 Acc F1 Acc F1 A1,A2: Tran + Lex 59.9 75.8 65.6 84.7 61.6 75.6 62.1 78.9 A1-A3: Tran + Lex + Sem 67.9 80.2 70.8 87.4 70.5 80.8 69.4 82.8 A1-A4: Tran + Lex + Sem 72.4 83.5 72.4 89.7 71.1 82.3 72.2 85.0 A1-A5: Tran + Lex + Sem + Syn 74.4 84.8 75.1 89.4 75.1 85.4 74.8 86.5 A2-A5: Lex + Sem + Syn 64.9 78.8 68.1 81.1 64.8 83.7 65.4 81.0 Table 4: Sentence accuracy (Acc) and segment F1 (F1) using CRFs with different features. Movie Job National Park Average Features Acc F1 Acc F1 Acc F1 Acc F1 B1,B2: Tran + Lex 53.4 71.6 59.6 83.8 60.0 77.3 56.7 76.9 B1-B3: Tran + Lex 61.3 77.7 65.9 85.9 66.0 80.7 63.7 80.1 B1-B4: Tran + Lex + Sem 73.8 83.6 76.0 89.7 74.6 85.3 74.7 86.1 B1-B5: Tran + Lex + Sem 75.0 84.3 76.5 89.7 76.8 86.8 75.8 86.6 B1-B6: Tran + Lex + Sem + Syn 75.8 84.3 76.2 89.7 76.8 87.2 76.1 86.7 B1-B5,B7: Tran + Lex + Sem + Syn 75.6 84.1 76.0 89.3 76.8 86.8 75.9 86.4 B2-B6:Lex + Sem + Syn 72.0 82.0 73.2 87.9 76.5 89.3 73.8 85.6 Table 5: Sentence accuracy (Acc) and segment F1 (F1) using semi-Markov CRFs with different features. example, IN (preposition or subordinating conjunction) is a strong indicator of Other, while TO and IM:Date usually do not co-occur. Some features, however, may appear less “correct”. This is largely due to the inaccurate output of the POS tagger. For example, a large number of actor names were mis-tagged as RB, resulting in a high positive weight of the feature (RB, IM:Cast). Positive Negative (IN, Other), (TO, IM:Date) (VBD, Other) (IN, IM:Cast) (CD, IM:Date) (CD, IH) (RB, IM:Cast) (IN, IM:Character) Table 6: Syntactic features with the largest positive/negative weights in the CRF model for Movie Similarly, we added segment-level POS tag features (B6) to semi-Markov CRFs, which lead to the best overall results as shown by the highlighted numbers in Table 5. Again many of the dominant features are consistent with our intuition. For example, the most positive feature for Movie is (CD JJS, IM:Rating) (e.g. 100 best). When syntactic features based on chunking results (B7) are used instead of B6, the performance is not as good. Transition features In addition, it is interesting to see the importance of transition features in both models. Since web queries do not generally follow the linear order principle, is it helpful to incorporate transition features in learning? To answer this question, we dropped the transition features from the best systems, corresponding to the last rows in Table 4 and 5. This resulted in substantial degradations in performance. One intuitive explanation is that although web queries are relatively “order-free”, statistically speaking, some orders are much more likely to occur than others. This makes it beneficial to use transition features. Comparison to syntactic analysis Finally, we conduct a simple experiment by using the heuristics described in Section 3.2 in extracting IHs from queries. The precision and recall of IHs averaged over all 3 domains are 50.4% and 32.8% respectively. The precision and recall numbers from our best model-based system, i.e., B1B6 in Table 5, are 89.9% and 84.6% respectively, which are significantly better than those based on pure syntactic analysis. 7 Conclusions In this work, we make the first attempt to define the semantic structure of noun phrase queries. We propose statistical methods to automatically extract IHs, IMs and the semantic labels of IMs using a variety of features. Experiments show the effectiveness of semantic features and syntactic features in both Markov and semi-Markov CRF models. In the future, it would be useful to explore other approaches to automatic lexicon discovery to improve the quality or to increase the coverage of both IH and IM lexicons, and to systematically evaluate their impact on query understanding performance. The author would like to thank Hisami Suzuki and Jianfeng Gao for useful discussions. 1344 References Jaime Arguello, Fernando Diaz, Jamie Callan, and Jean-Francois Crespo. 2009. Sources of evidence for vertical selection. In SIGIR’09: Proceedings of the 32st Annual International ACM SIGIR conference on Research and Development in Information Retrieval. Cory Barr, Rosie Jones, and Moira Regelson. 2008. The linguistic structure of English web-search queries. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 1021–1030. Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. 2005. Learning to rank using gradient descent. In ICML’05: Proceedings of the 22nd international conference on Machine learning, pages 89–96. Michael Collins. 1999. Head-Driven Statistical Models for Natural Language Parsing. Ph.D. thesis, University of Pennsylvania. Jianfeng Gao, Jian-Yun Nie, Jian Zhang, Endong Xun, Ming Zhou, and Chang-Ning Huang. 2001. Improving query translation for CLIR using statistical models. In SIGIR’01: Proceedings of the 24th Annual International ACM SIGIR conference on Research and Development in Information Retrieval. Jinyoung Kim, Xiaobing Xue, and Bruce Croft. 2009. A probabilistic retrieval model for semistructured data. In ECIR’09: Proceedings of the 31st European Conference on Information Retrieval, pages 228–239. John Lafferty, Andrew McCallum, and Ferdando Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the International Conference on Machine Learning, pages 282–289. Fangtao Li, Xian Zhang, Jinhui Yuan, and Xiaoyan Zhu. 2008a. Classifying what-type questions by head noun tagging. In COLING’08: Proceedings of the 22nd International Conference on Computational Linguistics, pages 481–488. Xiao Li, Ye-Yi Wang, and Alex Acero. 2008b. Learning query intent from regularized click graph. In SIGIR’08: Proceedings of the 31st Annual International ACM SIGIR conference on Research and Development in Information Retrieval, July. Xiao Li, Ye-Yi Wang, and Alex Acero. 2009. Extracting structured information from user queries with semi-supervised conditional random fields. In SIGIR’09: Proceedings of the 32st Annual International ACM SIGIR conference on Research and Development in Information Retrieva. Mehdi Manshadi and Xiao Li. 2009. Semantic tagging of web search queries. In Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP. Andrew McCallum and Wei Li. 2003. Early results for named entity recognition with conditional random fields, feature induction and web-enhanced lexicons. In Proceedings of the seventh conference on Natural language learning at HLT-NAACL 2003, pages 188– 191. Donald Metzler and Bruce Croft. 2005. Analysis of statistical question classification for fact-based questions. Jounral of Information Retrieval, 8(3). Patrick Pantel and Marco Pennacchiotti. 2006. Espresso: Leveraging generic patterns for automatically har-vesting semantic relations. In Proceedings of the 21st International Conference on Computational Linguis-tics and the 44th annual meeting of the ACL, pages 113–120. Stelios Paparizos, Alexandros Ntoulas, John Shafer, and Rakesh Agrawal. 2009. Answering web queries using structured data sources. In Proceedings of the 35th SIGMOD international conference on Management of data. Marius Pasca and Benjamin Van Durme. 2007. What you seek is what you get: Extraction of class attributes from query logs. In IJCAI’07: Proceedings of the 20th International Joint Conference on Artificial Intelligence. Marius Pasca and Benjamin Van Durme. 2008. Weakly-supervised acquisition of open-domain classes and class attributes from web documents and query logs. In Proceedings of ACL-08: HLT. Marco Pennacchiotti and Patrick Pantel. 2009. Entity extraction via ensemble semantics. In EMNLP’09: Proceedings of Conference on Empirical Methods in Natural Language Processing, pages 238–247. Stephen Robertson, Hugo Zaragoza, and Michael Taylor. 2004. Simple BM25 extension to multiple weighted fields. In CIKM’04: Proceedings of the thirteenth ACM international conference on Information and knowledge management, pages 42–49. Sunita Sarawagi and William W. Cohen. 2004. SemiMarkov conditional random fields for information extraction. In Advances in Neural Information Processing Systems (NIPS’04). Dou Shen, Jian-Tao Sun, Qiang Yang, and Zheng Chen. 2006. Building bridges for web query classification. In SIGIR’06: Proceedings of the 29th Annual International ACM SIGIR conference on research and development in information retrieval, pages 131–138. 1345
|
2010
|
136
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1346–1356, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Multilingual Pseudo-Relevance Feedback: Performance Study of Assisting Languages Manoj K. Chinnakotla Karthik Raman Pushpak Bhattacharyya Department of Computer Science and Engineering Indian Institute of Technology, Bombay, Mumbai, India {manoj,karthikr,pb}@cse.iitb.ac.in Abstract In a previous work of ours Chinnakotla et al. (2010) we introduced a novel framework for Pseudo-Relevance Feedback (PRF) called MultiPRF. Given a query in one language called Source, we used English as the Assisting Language to improve the performance of PRF for the source language. MulitiPRF showed remarkable improvement over plain Model Based Feedback (MBF) uniformly for 4 languages, viz., French, German, Hungarian and Finnish with English as the assisting language. This fact inspired us to study the effect of any source-assistant pair on MultiPRF performance from out of a set of languages with widely different characteristics, viz., Dutch, English, Finnish, French, German and Spanish. Carrying this further, we looked into the effect of using two assisting languages together on PRF. The present paper is a report of these investigations, their results and conclusions drawn therefrom. While performance improvement on MultiPRF is observed whatever the assisting language and whatever the source, observations are mixed when two assisting languages are used simultaneously. Interestingly, the performance improvement is more pronounced when the source and assisting languages are closely related, e.g., French and Spanish. 1 Introduction The central problem of Information Retrieval (IR) is to satisfy the user’s information need, which is typically expressed through a short (typically 2-3 words) and often ambiguous query. The problem of matching the user’s query to the documents is rendered difficult by natural language phenomena like morphological variations, polysemy and synonymy. Relevance Feedback (RF) tries to overcome these problems by eliciting user feedback on the relevance of documents obtained from the initial ranking and then uses it to automatically refine the query. Since user input is hard to obtain, Pseudo-Relevance Feedback (PRF) (Buckley et al., 1994; Xu and Croft, 2000; Mitra et al., 1998) is used as an alternative, wherein RF is performed by assuming the top k documents from the initial retrieval as being relevant to the query. Based on the above assumption, the terms in the feedback document set are analyzed to choose the most distinguishing set of terms that characterize the feedback documents and as a result the relevance of a document. Query refinement is done by adding the terms obtained through PRF, along with their weights, to the actual query. Although PRF has been shown to improve retrieval, it suffers from the following drawbacks: (a) the type of term associations obtained for query expansion is restricted to co-occurrence based relationships in the feedback documents, and thus other types of term associations such as lexical and semantic relations (morphological variants, synonyms) are not explicitly captured, and (b) due to the inherent assumption in PRF, i.e., relevance of top k documents, performance is sensitive to that of the initial retrieval algorithm and as a result is not robust. Multilingual Pseudo-Relevance Feedback (MultiPRF) (Chinnakotla et al., 2010) is a novel framework for PRF to overcome both the above limitations of PRF. It does so by taking the help of a different language called the assisting language. In MultiPRF, given a query in source language L1, the query is automatically translated into the assisting language L2 and PRF performed in the assisting language. The resultant terms are translated back into L1 using a probabilistic bi-lingual dictionary. The translated feedback 1346 model, is then combined with the original feedback model of L1 to obtain the final model which is used to re-rank the corpus. MulitiPRF showed remarkable improvement on standard CLEF collections over plain Model Based Feedback (MBF) uniformly for 4 languages, viz., French, German, Hungarian and Finnish with English as the assisting language. This fact inspired us to study the effect of any source-assistant pair on PRF performance from out of a set of languages with widely different characteristics, viz., Dutch, English, Finnish, French, German and Spanish. Carrying this further, we looked into the effect of using two assisting languages together on PRF. The present paper is a report of these investigations, their results and conclusions drawn therefrom. While performance improvement on PRF is observed whatever the assisting language and whatever the source, observations are mixed when two assisting languages are used simultaneously. Interestingly, the performance improvement is more pronounced when the source and assisting languages are closely related, e.g., French and Spanish. The paper is organized as follows: Section 2, discusses the related work. Section 3, explains the Language Modeling (LM) based PRF approach. Section 4, describes the MultiPRF approach. Section 5 discusses the experimental set up. Section 6 presents the results, and studies the effect of varying the assisting language and incorporates multiple assisting languages. Finally, Section 7 concludes the paper by summarizing and outlining future work. 2 Related Work PRF has been successfully applied in various IR frameworks like vector space models, probabilistic IR and language modeling (Buckley et al., 1994; Jones et al., 2000; Lavrenko and Croft, 2001; Zhai and Lafferty, 2001). Several approaches have been proposed to improve the performance and robustness of PRF. Some of the representative techniques are (i) Refining the feedback document set (Mitra et al., 1998; Sakai et al., 2005), (ii) Refining the terms obtained through PRF by selecting good expansion terms (Cao et al., 2008) and (iii) Using selective query expansion (Amati et al., 2004; Cronen-Townsend et al., 2004) and (iv) Varying the importance of documents in the feedback set (Tao and Zhai, 2006). Another direction of work, often reported in the TREC Robust Track, is to use a large external collection like Wikipedia or the Web as a source of expansion terms (Xu et al., 2009; Voorhees, 2006). The intuition behind the above approach is that if the query does not have many relevant documents in the collection then any improvements in the modeling of PRF is bound to perform poorly due to query drift. Several approaches have been proposed for including different types of lexically and semantically related terms during query expansion. Voorhees (1994) use Wordnet for query expansion and report negative results. Recently, random walk models (Lafferty and Zhai, 2001; CollinsThompson and Callan, 2005) have been used to learn a rich set of term level associations by combining evidence from various kinds of information sources like WordNet, Web etc. Metzler and Croft (2007) propose a feature based approach called latent concept expansion to model term dependencies. All the above mentioned approaches use the resources available within the language to improve the performance of PRF. However, we make use of a second language to improve the performance of PRF. Our proposed approach is especially attractive in the case of resource-constrained languages where the original retrieval is bad due to poor coverage of the collection and/or inherent complexity of query processing (for example term conflation) in those languages. Jourlin et al. (1999) use parallel blind relevance feedback, i.e. they use blind relevance feedback on a larger, more reliable parallel corpus, to improve retrieval performance on imperfect transcriptions of speech. Another related idea is by Xu et al. (2002), where a statistical thesaurus is learned using the probabilistic bilingual dictionaries of Arabic to English and English to Arabic. Meij et al. (2009) tries to expand a query in a different language using language models for domainspecific retrieval, but in a very different setting. Since our method uses a corpus in the assisting language from a similar time period, it can be likened to the work by Talvensaari et al. (2007) who used comparable corpora for Cross-Lingual Information Retrieval (CLIR). Other work pertaining to document alignment in comparable corpora, such as Braschler and Sch¨auble (1998), Lavrenko et al. (2002), also share certain common themes with our approach. Recent work by Gao et al. 1347 (2008) uses English to improve the performance over a subset of Chinese queries whose translations in English are unambiguous. They use interdocument similarities across languages to improve the ranking performance. However, cross language document similarity measurement is in itself known to be an hard problem and the scale of their experimentation is quite small. 3 PRF in the LM Framework The Language Modeling (LM) Framework allows PRF to be modelled in a principled manner. In the LM approach, documents and queries are modeled using multinomial distribution over words called document language model P(w|D) and query language model P(w|ΘQ) respectively. For a given query, the document language models are ranked based on their proximity to the query language model, measured using KL-Divergence. KL(ΘQ||D) = X w P (w|ΘQ) · log P (w|ΘQ) P (w|D) Since the query length is short, it is difficult to estimate ΘQ accurately using the query alone. In PRF, the top k documents obtained through the initial ranking algorithm are assumed to be relevant and used as feedback for improving the estimation of ΘQ. The feedback documents contain both relevant and noisy terms from which the feedback language model is inferred based on a Generative Mixture Model (Zhai and Lafferty, 2001). Let DF = {d1, d2, . . . , dk} be the top k documents retrieved using the initial ranking algorithm. Zhai and Lafferty (Zhai and Lafferty, 2001) model the feedback document set DF as a mixture of two distributions: (a) the feedback language model and (b) the collection model P(w|C). The feedback language model is inferred using the EM Algorithm (Dempster et al., 1977), which iteratively accumulates probability mass on the most distinguishing terms, i.e. terms which are more frequent in the feedback document set than in the entire collection. To maintain query focus the final converged feedback model, ΘF is interpolated with the initial query model ΘQ to obtain the final query model ΘFinal. ΘF inal = (1 −α) · ΘQ + α · ΘF ΘFinal is used to re-rank the corpus using the KL-Divergence ranking function to obtain the final ranked list of documents. Henceforth, we refer Initial Retrieval Algorithm (LM Based Query Likelihood) Initial Retrieval Algorithm (LM Based Query Likelihood) Top ‘k’ Results Top ‘k’ Results PRF (Model Based Feedback) PRF (Model Based Feedback) L1 Index L2 Index Final Ranked List Of Documents in L1 Feedback Model Interpolation Relevance Model Translation KL-Divergence Ranking Function Feedback Model θL2 Feedback Model θL1 Query in L1 Translated Query to L2 Probabilistic Dictionary L2 → L1 Translated Feedback Model Query Model θQ Figure 1: Schematic of the Multilingual PRF Approach Symbol Description ΘQ Query Language Model ΘF L1 Feedback Language Model obtained from PRF in L1 ΘF L2 Feedback Language Model obtained from PRF in L2 ΘT rans L1 Feedback Model Translated from L2 to L1 t(f|e) Probabilistic Bi-Lingual Dictionary from L2 to L1 β, γ Interpolation coefficients coefficients used in MultiPRF Table 2: Glossary of Symbols used in explaining MultiPRF to the above technique as Model Based Feedback (MBF). 4 Multilingual PRF (MultiPRF) The schematic of the MultiPRF approach is shown in Figure 1. Given a query Q in the source language L1, we automatically translate the query into the assisting language L2. We then rank the documents in the L2 collection using the query likelihood ranking function (John Lafferty and Chengxiang Zhai, 2003). Using the top k documents, we estimate the feedback model using MBF as described in the previous section. Similarly, we also estimate a feedback model using the original query and the top k documents retrieved from the initial ranking in L1. Let the resultant feedback models be ΘF L2 and ΘF L1 respectively. The feedback model estimated in the assisting language ΘF L2 is translated back into language L1 using a probabilistic bi-lingual dictionary t(f|e) from L2 →L1 as follows: P (f|ΘT rans L1 ) = X ∀e in L2 t(f|e) · P (e|ΘF L2) (1) The probabilistic bi-lingual dictionary t(f|e) is 1348 Language CLEF Collection Identifier Description No. of Documents No. of Unique Terms CLEF Topics (No. of Topics) English EN-00+01+02 LA Times 94 113005 174669 EN-03+05+06 LA Times 94, Glasgow Herald 95 169477 234083 EN-02+03 LA Times 94, Glasgow Herald 95 169477 234083 91-200 (67) French FR-00 Le Monde 94 44013 127065 1-40 (29) FR-01+02 Le Monde 94, French SDA 94 87191 159809 41-140 (88) FR-02+03 Le Monde 94, French SDA 94-95 129806 182214 91-200 (67) FR-03+05 Le Monde 94, French SDA 94-95 129806 182214 141-200,251-300 (99) FR-06 Le Monde 94-95, French SDA 94-95 177452 231429 301-350 (48) German DE-00 Frankfurter Rundschau 94, Der Spiegel 94-95 153694 791093 1-40 (33) DE-01+02 Frankfurter Rundschau 94, Der Spiegel 94-95, German SDA 94 225371 782304 41-140 (85) DE-02+03 Frankfurter Rundschau 94, Der Spiegel 94-95, German SDA 94-95 294809 867072 91-200 (67) DE-03 Frankfurter Rundschau 94, Der Spiegel 94-95, German SDA 94-95 294809 867072 141-200 (51) Finnish FI-02+03+04 Aamulehti 94-95 55344 531160 91-250 (119) FI-02+03 Aamulehti 94-95 55344 531160 91-200 (67) Dutch NL-02+03 NRC Handelsblad 94-95, Algemeen Dagblad 9495 190604 575582 91-200 (67) Spanish ES-02+03 EFE 94, EFE 95 454045 340250 91-200 (67) Table 1: Details of the CLEF Datasets used for Evaluating the MultiPRF approach. The number shown in brackets of the final column CLEF Topics indicate the actual number of topics used during evaluation. Source Term Top Aligned Terms in Target French English am´ericain american, us, united, state, america nation nation, un, united, state, country e´tude study, research, assess, investigate, survey German English flugzeug aircraft, plane, aeroplane, air, flight spiele play, game, stake, role, player verh¨altnis relationship, relate, balance, proportion Table 3: Top Translation Alternatives for some sample words in Probabilistic Bi-Lingual Dictionary learned from a parallel sentence-aligned corpora in L1−L2 based on word level alignments. Tiedemann (Tiedemann, 2001) has shown that the translation alternatives found using word alignments could be used to infer various morphological and semantic relations between terms. In Table 3, we show the top translation alternatives for some sample words. For example, the French word am´ericain (american) brings different variants of the translation like american, america, us, united, state, america which are lexically and semantically related. Hence, the probabilistic bi-lingual dictionary acts as a rich source of morphologically and semantically related feedback terms. Thus, during this step, of translating the feedback model as given in Equation 1, the translation model adds related terms in L1 which have their source as the term from feedback model ΘF L2. The final MultiPRF model is obtained by interpolating the above translated feedback model with the original query model and the feedback model of language L1 as given below: ΘMulti L1 = (1 −β −γ) · ΘQ + β · ΘF L1 + γ · ΘT rans L1 (2) Since we want to retain the query focus during back translation the feedback model in L2 is interpolated with the translated query before translation of the L2 feedback model. The parameters β and γ control the relative importance of the original query model, feedback model of L1 and the translated feedback model obtained from L1 and are tuned based on the choice of L1 and L2. 5 Experimental Setup We evaluate the performance of our system using the standard CLEF evaluation data in six languages, widely varying in their familial relationships - Dutch, German, English, French, Spanish and Finnish using more than 600 topics. The details of the collections and their corresponding topics used for MultiPRF are given in Table 1. Note that, in each experiment, we choose assisting collections such that the topics in the source language are covered in the assisting collection so as to get meaningful feedback terms. In all the topics, we only use the title field. We ignore the topics which have no relevant documents as the true performance on those topics cannot be evaluated. We demonstrate the performance of MultiPRF approach with French, German and Finnish as source languages and Dutch, English and Spanish as the assisting language. We later vary the assisting language, for each source language and study the effects. We use the Terrier IR platform (Ounis et al., 2005) for indexing the documents. We perform standard tokenization, stop word removal and stemming. We use the Porter Stemmer for English and the stemmers available through the Snowball package for other languages. Other than these, we do not perform any language-specific processing on the languages. In case of French, 1349 Collection Assist. Lang P@5 P@10 MAP GMAP MBF MultiPRF % Impr. MBF MultiPRF % Impr. MBF MultiPRF % Impr. MBF MultiPRF % Impr. FR-00 EN 0.4690 0.5241 11.76‡ 0.4000 0.4000 0.00 0.4220 0.4393 4.10 0.2961 0.3413 15.27 ES 0.5034 7.35‡ 0.4103 2.59 0.4418 4.69 0.3382 14.22 NL 0.5034 7.35 0.4103 2.59 0.4451 5.47 0.3445 16.34 FR-01+02 EN 0.4636 0.4818 3.92 0.4068 0.4386 7.82‡ 0.4342 0.4535 4.43‡ 0.2395 0.2721 13.61 ES 0.4977 7.35‡ 0.4363 7.26‡ 0.4416 1.70 0.2349 -1.92 NL 0.4818 3.92 0.4409 8.38‡ 0.4375 0.76 0.2534 5.80 FR-03+05 EN 0.4545 0.4768 4.89‡ 0.4040 0.4202 4‡ 0.3529 0.3694 4.67‡ 0.1324 0.1411 6.57 ES 0.4727 4.00 0.4080 1.00 0.3582 1.50 0.1325 0.07 NL 0.4525 -0.44 0.4010 -0.75 0.3513 0.45 0.1319 -0.38 FR-06 EN 0.4917 0.5083 3.39 0.4625 0.4729 2.25 0.3837 0.4104 6.97 0.2174 0.2810 29.25 ES 0.5083 3.39 0.4687 1.35 0.3918 2.12 0.2617 20.38 NL 0.5083 3.39 0.4646 0.45 0.3864 0.71 0.2266 4.23 DE-00 EN 0.2303 0.3212 39.47‡ 0.2394 0.2939 22.78‡ 0.2158 0.2273 5.31 0.0023 0.0191 730.43 ES 0.3212 39.47‡ 0.2818 17.71‡ 0.2376 10.09 0.0123 434.78 NL 0.3151 36.82‡ 0.2818 17.71‡ 0.2331 8.00 0.0122 430.43 DE-01+02 EN 0.5341 0.6000 12.34‡ 0.4864 0.5318 9.35‡ 0.4229 0.4576 8.2‡ 0.1765 0.2721 9.19 ES 0.5682 6.39‡ 0.5091 4.67‡ 0.4459 5.43 0.2309 30.82 NL 0.5773 8.09‡ 0.5114 5.15‡ 0.4498 6.35‡ 0.2355 33.43 DE-03 EN 0.5098 0.5412 6.15 0.4784 0.4980 4.10 0.4274 0.4355 1.91 0.1243 0.1771 42.48 ES 0.5647 10.77‡ 0.4980 4.10 0.4568 6.89‡ 0.1645 32.34 NL 0.5529 8.45‡ 0.4941 3.27 0.4347 1.72 0.1490 19.87 FI-02+03+04 EN 0.3782 0.4034 6.67‡ 0.3059 0.3319 8.52‡ 0.3966 0.4246 7.06‡ 0.1344 0.2272 69.05 ES 0.3879 2.58 0.3267 6.81 0.3881 -2.15 0.1755 30.58 NL 0.3948 4.40 0.3301 7.92 0.4077 2.79 0.1839 36.83 Table 4: Results comparing the performance of MultiPRF over baseline MBF on CLEF collections with English (EN), Spanish (ES) and Dutch (NL) as assisting languages. Results marked as ‡ indicate that the improvement was found to be statistically significant over the baseline at 90% confidence level (α = 0.01) when tested using a paired two-tailed t-test. since some function words like l’, d’ etc., occur as prefixes to a word, we strip them off during indexing and query processing, since it significantly improves the baseline performance. We use standard evaluation measures like MAP, P@5 and P@10 for evaluation. Additionally, for assessing robustness, we use the Geometric Mean Average Precision (GMAP) metric (Robertson, 2006) which is also used in the TREC Robust Track (Voorhees, 2006). The probabilistic bi-lingual dictionary used in MultiPRF was learnt automatically by running GIZA++: a word alignment tool (Och and Ney, 2003) on a parallel sentence aligned corpora. For all the above language pairs we used the Europarl Corpus (Philipp, 2005). We use Google Translate as the query translation system as it has been shown to perform well for the task (Wu et al., 2008). We use the MBF approach explained in Section 3 as a baseline for comparison. We use two-stage Dirichlet smoothing with the optimal parameters tuned based on the collection (Zhai and Lafferty, 2004). We tune the parameters of MBF, specifically λ and α, and choose the values which give the optimal performance on a given collection. We uniformly choose the top ten documents for feedback. Table 4 gives the overall results. 6 Results and Discussion In Table 4, we see the performance of the MultiPRF approach for three assisting languages, and how it compares with the baseline MBF methods. We find MultiPRF to consistently outperform the baseline value on all metrics, namely MAP (where significant improvements range from 4.4% to 7.1%); P@5 (significant improvements range from 4.9% to 39.5% and P@10 (where MultiPRF has significant gains varying from 4% to 22.8%). Additionally we also find MultiPRF to be more robust than the baseline, as indicated by the GMAP score, where improvements vary from 4.2% to 730%. Furthermore we notice these trends hold across different assisting languages, with Spanish and Dutch outperforming English as the assisting language on some of the French and German collections. On performing a more detailed study of the results we identify the main reason for improvements in our approach is the ability to obtain good feedback terms in the assisting language coupled with the introduction of lexically and semantically related terms during the backtranslation step. In Table 5, we see some examples, which illustrates the feedback terms brought by the MultiPRF method. As can be seen by these example, the gains achieved by MultiPRF are primarily due to one of three reasons: (a) Good Feedback in Assisting Language: If the feedback model in the assisting language contains good terms, then the back-translation process will introduce the corresponding feedback terms in the source language, thus leading to improved performance. As an example of this phenomena, consider the French Query “Maladie de Creutzfeldt-Jakob”. In this case the original feedback model also performs 1350 TOPIC NO ASSIST LANG. SOURCE LANGUAGE QUERY TRANSLATED QUERY QUERY MEANING MBF MAP MPRF MAP MBF- Top Representative Terms (With Meaning) Excl. Query Terms MultiPRF- Top Representative Terms (With Meaning) Excl. Query Terms GERMAN '01: TOPIC 61 EN Ölkatastrophe in Sibirien Oil Spill in Siberia Siberian Oil Catastrophe 0.618 0.812 exxon, million, ol (oil), tonn, russisch (russian), olp (oil), moskau (moscow), us olverschmutz (oil pollution), ol, russisch, erdol (petroleum), russland (russia), olunfall(oil spill), olp GERMAN '02: TOPIC 105 ES Bronchialasthma El asma bronquial Bronchial Asthma 0.062 0.636 chronisch (chronic), pet, athlet (athlete), ekrank (ill), gesund (healthy), tuberkulos (tuberculosis), patient, reis (rice), person asthma, allergi, krankheit (disease), allerg (allergenic), chronisch, hauterkrank (illness of skin), arzt (doctor), erkrank (ill) FRENCH '02: TOPIC 107 NL Ingénierie génétique Genetische Manipulatie Genetic Engineering 0.145 0.357 développ (developed), évolu (evolved), product, produit (product), moléculair (molecular) genetic, gen, engineering, développ, product FRENCH '06: TOPIC 256 EN Maladie de Creutzfeldt-Jakob Creutzfeldt-Jakob CreutzfeldtJakob Disease 0.507 0.688 malad (illness), produit (product), animal (animal), hormon (hormone) malad, humain (human), bovin (bovine), encéphalopath (suffering from encephalitis), scientif, recherch (research) GERMAN '03: TOPIC 157 EN Siegerinnen von Wimbledon Champions of Wimbledon Wimbledon Lady Winners 0.074 0.146 telefonbuch (phone book), sieg (victory), titelseit (front page), telekom (telecommunication), graf gross (large), verfecht (champion), sampra (sampras), 6, champion, steffi, verteidigt (defendending), martina, jovotna, navratilova GERMAN '01: TOPIC 91 ES AI in Lateinamerika La gripe aviar en América Latina AI in Latin America 0.456 0.098 international, amnesty, strassenkind (street child), kolumbi (Columbian), land, brasili (Brazil), menschenrecht (human rights), polizei (police) karib (Caribbean), land, brasili, schuld (blame), amerika, kalt (cold), welt (world), forschung (research) GERMAN '03: TOPIC 196 EN Fusion japanischer Banken Fusion of Japanese banks Merger of Japanese Banks 0.572 0.264 daiwa, tokyo, filial (branch), zusammenschluss (merger) kernfusion (nuclear fusion), zentralbank (central bank), daiwa, weltbank (world bank), investitionsbank (investment bank) FRENCH '03: TOPIC 152 NL Les droits de l'enfant De rechten van het kind Child Rights 0.479 0.284 convent (convention), franc, international, onun (united nations), réserv (reserve) per (father), convent, franc, jurid (legal), homm (man), cour (court), biolog Table 5: Qualitative comparison of feedback terms given by MultiPRF and MBF on representative queries where positive and negative results were observed in French and German collections. quite strongly with a MAP score of 0.507. Although there is no significant topic drift in this case, there are not many relevant terms apart from the query terms. However the same query performs very well in English with all the documents in the feedback set of the English corpus being relevant, thus resulting in informative feedback terms such as {bovin, scientif, recherch}. (b) Finding Synonyms/Morphological Variations: Another situation in which MultiPRF leads to large improvements is when it finds semantically/lexically related terms to the query terms which the original feedback model was unable to. For example, consider the French query “Ing´enierie g´n´tique”. While the feedback model was unable to find any of the synonyms of the query terms, due to their lack of co-occurence with the query terms, the MultiPRF model was able to get these terms, which are introduced primarily during the backtranslation process. Thus terms like {genetic, gen, engineering}, which are synonyms of the query words, are found thus resulting in improved performance. (c) Combination of Above Factors: Sometimes a combination of the above two factors causes improvements in the performance as in the German query “ ¨Olkatastrophein Sibirien”. For this query, MultiPRF finds good feedback terms such as {russisch, russland} while also obtaining semantically related terms such as {olverschmutz, erdol, olunfall}. Although all of the previously described examples had good quality translations of the query in the assisting language, as mentioned in (Chinnakotla et al., 2010), the MultiPRF approach is robust to suboptimal translation quality as well. To see how MultiPRF leads to improvements even with errors in query translation consider the German Query “Siegerinnen von Wimbledon”. When this is translated to English, the term “Lady” is dropped, this causes only “Wimbledon Champions” to remain. As can be observed, this causes terms like sampras to come up in the MultiPRF model. However, while the MultiPRF model has some terms pertaining to Men’s Winners of Wimbledon as well, the original feedback model suffers from severe topic drift, with irrelevant terms such as {telefonbuch, telekom} also amongst the top terms. Thus we notice that despite the error in query translation MultiPRF still manages to correct the drift of the original feedback model, while also introducing relevant terms such as {verfecht, steffi, martina, novotna, navratilova} as well. Thus as shown in (Chinnakotla et al., 2010), having a better query translation system can only lead to better performance. We also perform a detailed error analysis and found three main reasons for MultiPRF failing: (i) Inaccuracies in query translation (including the presence of out-of-vocabulary terms). This is seen in the German Query AI in Lateinamerika, which wrongly translates to Avian Flu in Latin America in Spanish thus affecting performance. (ii) Poor retrieval in Assisting Language. Consider the French query Les droits de l’enfant, for which due to topic drift in English, MultiPRF performance reduces. (iii) In a few rare cases inaccuracy in the back transla1351 (a) Source:French (FR-01+02) Assist:Spanish (b) Source:German (DE-01+02) Assist:Dutch (c) Source:Finnish (FI-02+03+04) Assist:English Figure 2: Results showing the sensitivity of MultiPRF performance to parameters β and γ for French, German and Finnish. tion affects performance as well. 6.1 Parameter Sensitivity Analysis The MultiPRF parameters β and γ in Equation 2 control the relative importance assigned to the original feedback model in source language L1, the translated feedback model obtained from assisting language L2 and the original query terms. We varied the β and γ parameters for French, German and Finnish collections with English, Dutch and Spanish as assisting languages and studied its effect on MAP of MultiPRF. The results are shown in Figure 2. The results show that, in all the three collections, the optimal value of the parameters almost remains the same and lies in the range of 0.4-0.48. Due to the above reason, we arbitrarily choose the parameters in the above range and do not use any technique to learn these parameters. 6.2 Effect of Assisting Language Choice In this section, we discuss the effect of varying the assisting language. Besides, we also study the inter and intra familial behaviour of sourceassisting language pairs. In order to ensure that the results are comparable across languages, we indexed the collections from the years 2002, 2003 and use common topics from the topic range 91200 that have relevant documents across all the six languages. The number of such common topics were 67. For each source language, we use the other languages as assisting collections and study the performance of MultiPRF. Since query translation quality varies across language pairs, we analyze the behaviour of MultiPRF in the following two scenarios: (a) Using ideal query translation (b) Using Google Translate for query translation. In ideal query translation setup, in order to eliminate its effect, we skip the query translation step and use the corresponding original topics for each target language instead. The results for both the above scenarios are given in Tables 6 and 7. From the results, we firstly observe that besides English, other languages such as French, Spanish, German and Dutch act as good assisting languages and help in improving performance over monolingual MBF. We also observe that the best assisting language varies with the source language. However, the crucial factors of the assisting language which influence the performance of MultiPRF are: (a) Monolingual PRF Performance: The main motivation for using a different language was to get good feedback terms, especially in case of queries which fail in the source language. Hence, an assisting language in which the monolingual feedback performance itself is poor, is unlikely to give any performance gains. This observation is evident in case of Finnish, which has the lowest Monolingual MBF performance. The results show that Finnish is the least helpful of assisting languages, with performance similar to those of the baselines. We also observe that the three best performing assistant languages, i.e. English, French and Spanish, have the highest monolingual performances as well, thus further validating the claim. One possible reason for this is the relative 1352 Source Lang. Assisting Language Source Lang.MBF English German Dutch Spanish French Finnish English MAP 0.4464 (-0.7%) 0.4471 (-0.5%) 0.4566 (+1.6%) 0.4563 (+1.5%) 0.4545 (+1.1%) 0.4495 P@5 0.4925 (-0.6%) 0.5045 (+1.8%) 0.5164 (+4.2%) 0.5075 (+2.4%) 0.5194 (+4.8%) 0.4955 P@10 0.4343 (+0.4%) 0.4373 (+1.0%) 0.4537 (+4.8%) 0.4343 (+0.4%) 0.4373 (+1.0%) 0.4328 German MAP 0.4229 (+4.9%) 0.4346 (+7.8%) 0.4314 (+7.0%) 0.411 (+1.9%) 0.3863 (-4.2%) 0.4033 P@5 0.5851 (+14%) 0.5851 (+14%) 0.5791 (+12.8%) 0.594 (+15.7%) 0.5522 (+7.6%) 0.5134 P@10 0.5284 (+11.3%) 0.5209 (+9.8%) 0.5179 (+9.1%) 0.5149 (+8.5%) 0.5075 (+6.9%) 0.4746 Dutch MAP 0.4317 (+4%) 0.4453 (+7.2%) 0.4275 (+2.9%) 0.4241 (+2.1%) 0.3971 (-4.4%) 0.4153 P@5 0.5642 (+11.8%) 0.5731 (+13.6%) 0.5343 (+5.9%) 0.5582 (+10.6%) 0.5045 (0%) 0.5045 P@10 0.5075 (+9%) 0.4925 (+5.8%) 0.4896 (+5.1%) 0.5015 (+7.7%) 0.4806 (+3.2%) 0.4657 Spanish MAP 0.4667 (-2.9%) 0.4749 (-1.2%) 0.4744 (-1.3%) 0.4609 (-4.1%) 0.4311 (10.3%) 0.4805 P@5 0.62 (-2.9%) 0.6418 (+0.5%) 0.6299 (-1.4%) 0.6269 (-1.6%) 0.6149 (-3.7%) 0.6388 P@10 0.5625 (-1.8%) 0.5806 (+1.3%) 0.5851 (+2.1%) 0.5627 (-1.8%) 0.5478 (-4.4%) 0.5731 French MAP 0.4658 (+6.9%) 0.4526 (+3.9%) 0.4374 (+0.4%) 0.4634 (+6.4%) 0.4451 (+2.2%) 0.4356 P@5 0.4925 (+3.1%) 0.4806 (+0.6%) 0.4567 (-4.4%) 0.4925 (+3.1%) 0.4836 (+1.3%) 0.4776 P@10 0.4358 (+3.9%) 0.4239 (+1%) 0.4224 (+0.7%) 0.4388 (+4.6%) 0.4209 (+0.4%) 0.4194 Finnish MAP 0.3411 (-4.7%) 0.3796 (+6.1%) 0.3722 (+4%) 0.369 (+3.1%) 0.3553 (-0.7%) 0.3578 P@5 0.394 (+3.1%) 0.403 (+5.5%) 0.406 (+6.3%) 0.4119 (+7.8%) 0.397 (+3.9%) 0.3821 P@10 0.3463 (+11.5%) 0.3582 (+15.4%) 0.3478 (+12%) 0.3448 (+11%) 0.3433 (+10.6%) 0.3105 Table 6: Results showing the performance of MultiPRF with different source and assisting languages using Google Translate for query translation step. The intra-familial affinity could be observed from the elements close to the diagonal. ease of processing in these languages. (b) Familial Similarity Between Languages: We observe that the performance of MultiPRF is good if the assisting language is from the same language family. Birch et al. (2008) show that the language family is a strong predictor of machine translation performance. Hence, the query translation and back translation quality improves if the source and assisting languages belong to the same family. For example, in the Germanic family, the sourceassisting language pairs German-English, DutchEnglish, Dutch-German and German-Dutch show good performance. Similarly, in Romance family, the performance of French-Spanish confirms this behaviour. In some cases, we observe that MultiPRF scores decent improvements even when the assisting language does not belong to the same language family as witnessed in French-English and English-French. This is primarily due to their strong monolingual MBF performance. 6.3 Effect of Language Family on Back Translation Performance As already mentioned, the performance of MultiPRF is good if the source and assisting languages belong to the same family. In this section, we verify the above intuition by studying the impact of language family on back translation performance. The experiment designed is as follows: Given a query in source language L1, the ideal translation in assisting language L2 is used to compute the query model in L2 using only the query terms. Then, without performing PRF the query model Source Lang. Assisting Language MBF MPRF FR ES DE NL EN FI French 0.3686 0.3113 0.3366 0.4338 0.3011 0.4342 0.4535 Spanish 0.3647 0.3440 0.3476 0.3954 0.3036 0.5000 0.4892 German 0.2729 0.2736 0.2951 0.2107 0.2266 0.4229 0.4576 Dutch 0.2663 0.2836 0.2902 0.2757 0.2372 0.3968 0.3989 Table 8: Effect of Language Family on Back Translation Performance measured through MultiPRF MAP. 100 Topics from years 2001 and 2002 were used for all languages. is directly back translated from L2 into L1 and finally documents are re-ranked using this translated feedback model. Since the automatic query translation and PRF steps have been eliminated, the only factor which influences the MultiPRF performance is the back-translation step. This means that the source-assisting language pairs for which the back-translation is good will score a higher performance. The results of the above experiment is shown in Table 8. For each source language, the best performing assisting languages have been highlighted. The results show that the performance of closely related languages like French-Spanish and German-Dutch is more when compared to other source-assistant language pairs. This shows that in case of closely related languages, the backtranslation step succeeds in adding good terms which are relevant like morphological variants, synonyms and other semantically related terms. Hence, familial closeness of the assisting language helps in boosting the MultiPRF performance. An exception to this trend is English as assisting lan1353 Source Lang. Assisting Language Source Lang.MBF English German Dutch Spanish French Finnish English MAP 0.4513 (+0.4%) 0.4475 (-0.4%) 0.4695 (+4.5%) 0.4665 (+3.8%) 0.4416 (-1.7%) 0.4495 P@5 0.5104 (+3.0%) 0.5104 (+3.0%) 0.5343 (+7.8%) 0.5403 (+9.0%) 0.4806 (-3.0%) 0.4955 P@10 0.4373 (+1.0%) 0.4358 (+0.7%) 0.4597 (+6.2%) 0.4582 (+5.9%) 0.4164 (-3.8%) 0.4328 German MAP 0.4427 (+9.8%) 0.4306 (+6.8%) 0.4404 (+9.2%) 0.4104 (+1.8%) 0.3993 (-1.0%) 0.4033 P@5 0.606 (+18%) 0.5672 (+10.5%) 0.594 (+15.7%) 0.5761 (+12.2%) 0.5552 (+8.1%) 0.5134 P@10 0.5373 (+13.2%) 0.503 (+6.0%) 0.5299 (+11.7%) 0.494 (+4.1%) 0.5 (+5.4%) 0.4746 Dutch MAP 0.4361 (+5.0%) 0.4344 (+4.6%) 0.4227 (+1.8%) 0.4304 (+3.6%) 0.4134 (-0.5%) 0.4153 P@5 0.5761 (+14.2%) 0.5552 (+10%) 0.5403 (+7.1%) 0.5463 (+8.3%) 0.5433 (+7.7%) 0.5045 P@10 0.5254 (+12.8%) 0.497 (+6.7%) 0.4776 (+2.6%) 0.5134 (+10.2%) 0.4925 (+5.8%) 0.4657 Spanish MAP 0.4665 (-2.9%) 0.4773 (-0.7%) 0.4733 (-1.5%) 0.4839 (+0.7%) 0.4412 (-8.2%) 0.4805 P@5 0.6507 (+1.8%) 0.6448 (+0.9%) 0.6507 (+1.8%) 0.6478 (+1.4%) 0.597 (-6.5%) 0.6388 P@10 0.5791 (+1.0%) 0.5791 (+1.0%) 0.5761 (+0.5%) 0.5866 (+2.4%) 0.5567 (-2.9%) 0.5731 French MAP 0.4591 (+5.4%) 0.4514 (+3.6%) 0.4409 (+1.2%) 0.4712 (+8.2%) 0.4354 (0%) 0.4356 P@5 0.4925 (+3.1%) 0.4776 (0%) 0.4776 (0%) 0.4995 (+4.6%) 0.4955 (+3.8%) 0.4776 P@10 0.4463 (+6.4%) 0.4313 (+2.8%) 0.4373 (+4.3%) 0.4448 (+6.1%) 0.4209 (+0.3%) 0.4194 Finnish MAP 0.3733 (+4.3%) 0.3559 (-0.5%) 0.3676 (+2.7%) 0.3594 (+0.4%) 0.371 (+3.7%) 0.3578 P@5 0.4149 (+8.6%) 0.385 (+0.7%) 0.388 (+1.6%) 0.388 (+1.6%) 0.3911 (+2.4%) 0.3821 P@10 0.3567 (+14.9%) 0.31 (-0.2%) 0.3253 (+4.8%) 0.32 (+3.1%) 0.3239 (+4.3%) 0.3105 Table 7: Results showing the performance of MultiPRF without using automatic query translation i.e. by using corresponding original queries in assisting collection. The results show the potential of MultiPRF by establishing a performance upper bound. guage which shows good performance across both families. 6.4 Multiple Assisting Languages So far, we have only considered a single assisting language. However, a natural extension to the method which comes to mind, is using multiple assisting languages. In other words, combining the evidence from all the feedback models of more than one assisting language, to get a feedback model which is better than that obtained using a single assisting language. To check how this simple extension works, we performed experiments using a pair of assisting languages. In these experiments for a given source language (from amongst the 6 previously mentioned languages) we tried using all pairs of assisting languages (for each source language, we have 10 pairs possible). To obtain the final model, we simply interpolate all the feedback models with the initial query model, in a similar manner as done in MultiPRF. The results for these experiments are given in Table 9. As we see, out of the 60 possible combinations of source language and assisting language pairs, we obtain improvements of greater than 3% in 16 cases. Here the improvements are with respect to the best model amongst the two MultiPRF models corresponding to each of the two assisting languages, with the same source language. Thus we observe that a simple linear interpolation of models is not the best way of combining evidence from multiple assisting languages. We also observe than when German or Spanish are used as one of the two assisting languages, they are most likely to Source Language Assisting Language Pairs with Improvement >3% English FR-DE (4.5%), FR-ES (4.8%), DE-NL (+3.1%) French EN-DE (4.1%), DE-ES (3.4%), NL-FI (4.8%) German None Spanish None Dutch EN-DE (3.9%), DE-FR (4.1%), FR-ES (3.8%), DE-ES (3.9%) Finnish EN-ES (3.2%), FR-DE (4.6%), FR-ES (6.4%), DE-ES (11.2%), DE-NL (4.4%), ES-NL (5.9%) Total - 16 EN – 3 Pairs; FR – 6 Pairs; DE – 10 Pairs; ES - 8 Pairs; NL – 4 Pairs; FI – 1 Pair Table 9: Summary of MultiPRF Results with Two Assisting Languages. The improvements described above are with respect to maximum MultiPRF MAP obtained using either L1 or L2 alone as assisting language. lead to improvements. A more detailed study of this observation needs to be done to explain this. 7 Conclusion and Future Work We studied the effect of different source-assistant pairs and multiple assisting languages on the performance of MultiPRF. Experiments across a wide range of language pairs with varied degree of familial relationships show that MultiPRF improves performance in most cases with the performance improvement being more pronounced when the source and assisting languages are closely related. We also notice that the results are mixed when two assisting languages are used simultaneously. As part of future work, we plan to vary the model interpolation parameters dynamically to improve the performance in case of multiple assisting languages. Acknowledgements The first author was supported by a fellowship award from Infosys Technologies Ltd., India. We would like to thank Mr. Vishal Vachhani for his help in running the experiments. 1354 References Giambattista Amati, Claudio Carpineto, and Giovanni Romano. 2004. Query Difficulty, Robustness, and Selective Application of Query Expansion. In ECIR ’04, pages 127–137. Alexandra Birch, Miles Osborne and Philipp Koehn. 2008. Predicting Success in Machine Translation. In EMNLP ’08, pages 745-754, ACL. Martin Braschler and Carol Peters. 2004. Cross-Language Evaluation Forum: Objectives, Results, Achievements. Inf. Retr., 7(1-2):7–31. Martin Braschler and Peter Sch¨auble. 1998. Multilingual Information Retrieval based on Document Alignment Techniques. In ECDL ’98, pages 183–197, Springer-Verlag. Chris Buckley, Gerald Salton, James Allan, and Amit Singhal. 1994. Automatic Query Expansion using SMART : TREC 3. In TREC-3, pages 69–80. Guihong Cao, Jian-Yun Nie, Jianfeng Gao, and Stephen Robertson. 2008. Selecting Good Expansion Terms for Pseudo-Relevance Feedback. In SIGIR ’08, pages 243– 250. ACM. Manoj K. Chinnakotla, Karthik Raman, and Pushpak Bhattacharyya. 2010. Multilingual PRF: English Lends a Helping Hand. In SIGIR ’10, ACM. Kevyn Collins-Thompson and Jamie Callan. 2005. Query Expansion Using Random Walk Models. In CIKM ’05, pages 704–711. ACM. Steve Cronen-Townsend, Yun Zhou, and W. Bruce Croft. 2004. A Framework for Selective Query Expansion. In CIKM ’04, pages 236–237. ACM. Ido Dagan, Alon Itai, and Ulrike Schwall. 1991. Two Languages Are More Informative Than One. In ACL ’91, pages 130–137. ACL. A. Dempster, N. Laird, and D. Rubin. 1977. Maximum Likelihood from Incomplete Data via the EM Algorithm. Journal of the Royal Statistical Society, 39:1–38. T. Susan Dumais, A. Todd Letsche, L. Michael Littman, and K. Thomas Landauer. 1997. Automatic Cross-Language Retrieval Using Latent Semantic Indexing. In AAAI ’97, pages 18–24. Wei Gao, John Blitzer, and Ming Zhou. 2008. Using English Information in Non-English Web Search. In iNEWS ’08, pages 17–24. ACM. David Hawking, Paul Thistlewaite, and Donna Harman. 1999. Scaling Up the TREC Collection. Inf. Retr., 1(12):115–137. Hieu Hoang, Alexandra Birch, Chris Callison-burch, Richard Zens, Rwth Aachen, Alexandra Constantin, Marcello Federico, Nicola Bertoldi, Chris Dyer, Brooke Cowan, Wade Shen, Christine Moran, and Ondej Bojar. 2007. Moses: Open Source Toolkit for Statistical Machine Translation. In ACL ’07, pages 177–180. P. Jourlin, S. E. Johnson, K. Sp¨arck Jones and P. C. Woodland. 1999. Improving Retrieval on Imperfect Speech Transcriptions (Poster Abstract). In SIGIR ’99, pages 283–284. ACM. John Lafferty and Chengxiang Zhai. 2003. Probabilistic Relevance Models Based on Document and Query Generation. Language Modeling for Information Retrieval, pages 1–10. Kluwer International Series on IR. K. Sparck Jones, S. Walker, and S. E. Robertson. 2000. A Probabilistic Model of Information Retrieval: Development and Comparative Experiments. Inf. Process. Manage., 36(6):779–808. John Lafferty and Chengxiang Zhai. 2001. Document Language Models, Query Models, and Risk Minimization for Information Retrieval. In SIGIR ’01, pages 111–119. ACM. Victor Lavrenko and W. Bruce Croft. 2001. Relevance Based Language Models. In SIGIR ’01, pages 120–127. ACM. Victor Lavrenko, Martin Choquette, and W. Bruce Croft. 2002. Cross-Lingual Relevance Models. In SIGIR ’02, pages 175–182, ACM. Edgar Meij, Dolf Trieschnigg, Maarten Rijke de, and Wessel Kraaij. 2009. Conceptual Language Models for Domainspecific Retrieval. Information Processing & Management, 2009. Donald Metzler and W. Bruce Croft. 2007. Latent Concept Expansion Using Markov Random Fields. In SIGIR ’07, pages 311–318. ACM. Mandar Mitra, Amit Singhal, and Chris Buckley. 1998. Improving Automatic Query Expansion. In SIGIR ’98, pages 206–214. ACM. Franz Josef Och and Hermann Ney. 2003. A Systematic Comparison of Various Statistical Alignment Models. Computational Linguistics, 29(1):19–51. I. Ounis, G. Amati, Plachouras V., B. He, C. Macdonald, and Johnson. 2005. Terrier Information Retrieval Platform. In ECIR ’05, volume 3408 of Lecture Notes in Computer Science, pages 517–519. Springer. Koehn Philipp. 2005. Europarl: A Parallel Corpus for Statistical Machine Translation. In MT Summit ’05. Stephen Robertson. 2006. On GMAP: and Other Transformations. In CIKM ’06, pages 78–83. ACM. Tetsuya Sakai, Toshihiko Manabe, and Makoto Koyama. 2005. Flexible Pseudo-Relevance Feedback Via Selective Sampling. ACM TALIP, 4(2):111–135. Tao Tao and ChengXiang Zhai. 2006. Regularized Estimation of Mixture Models for Robust Pseudo-Relevance Feedback. In SIGIR ’06, pages 162–169. ACM. Tuomas Talvensaari, Jorma Laurikkala, Kalervo J¨arvelin, Martti Juhola, and Heikki Keskustalo. 2007. Creating and Exploiting a Comparable Corpus in Cross-language Information Retrieval. ACM Trans. Inf. Syst., 25(1):4, 2007. Jrg Tiedemann. 2001. The Use of Parallel Corpora in Monolingual Lexicography - How word alignment can identify morphological and semantic relations. In COMPLEX ’01, pages 143–151. Ellen M. Voorhees. 1994. Query Expansion Using LexicalSemantic Relations. In SIGIR ’94, pages 61–69. SpringerVerlag. 1355 Ellen Voorhees. 2006. Overview of the TREC 2005 Robust Retrieval Track. In TREC 2005, Gaithersburg, MD. NIST. Dan Wu, Daqing He, Heng Ji, and Ralph Grishman. 2008. A Study of Using an Out-of-Box Commercial MT System for Query Translation in CLIR. In iNEWS ’08, pages 71– 76. ACM. Jinxi Xu and W. Bruce Croft. 2000. Improving the Effectiveness of Information Retrieval with Local Context Analysis. ACM Trans. Inf. Syst., 18(1):79–112. Jinxi Xu, Alexander Fraser, and Ralph Weischedel. 2002. Empirical Studies in Strategies for Arabic Retrieval. In SIGIR ’02, pages 269–274. ACM. Yang Xu, Gareth J.F. Jones, and Bin Wang. 2009. Query Dependent Pseudo-Relevance Feedback Based on Wikipedia. In SIGIR ’09, pages 59–66. ACM. Chengxiang Zhai and John Lafferty. 2001. Model-based Feedback in the Language Modeling approach to Information Retrieval. In CIKM ’01, pages 403–410. ACM. Chengxiang Zhai and John Lafferty. 2004. A Study of Smoothing Methods for Language Models applied to Information Retrieval. ACM Transactions on Information Systems, 22(2):179–214. 1356
|
2010
|
137
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1357–1366, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Wikipedia as Sense Inventory to Improve Diversity in Web Search Results Celina Santamar´ıa, Julio Gonzalo and Javier Artiles nlp.uned.es UNED, c/Juan del Rosal, 16, 28040 Madrid, Spain [email protected] [email protected] [email protected] Abstract Is it possible to use sense inventories to improve Web search results diversity for one word queries? To answer this question, we focus on two broad-coverage lexical resources of a different nature: WordNet, as a de-facto standard used in Word Sense Disambiguation experiments; and Wikipedia, as a large coverage, updated encyclopaedic resource which may have a better coverage of relevant senses in Web pages. Our results indicate that (i) Wikipedia has a much better coverage of search results, (ii) the distribution of senses in search results can be estimated using the internal graph structure of the Wikipedia and the relative number of visits received by each sense in Wikipedia, and (iii) associating Web pages to Wikipedia senses with simple and efficient algorithms, we can produce modified rankings that cover 70% more Wikipedia senses than the original search engine rankings. 1 Motivation The application of Word Sense Disambiguation (WSD) to Information Retrieval (IR) has been subject of a significant research effort in the recent past. The essential idea is that, by indexing and matching word senses (or even meanings) , the retrieval process could better handle polysemy and synonymy problems (Sanderson, 2000). In practice, however, there are two main difficulties: (i) for long queries, IR models implicitly perform disambiguation, and thus there is little room for improvement. This is the case with most standard IR benchmarks, such as TREC (trec.nist.gov) or CLEF (www.clef-campaign.org) ad-hoc collections; (ii) for very short queries, disambiguation may not be possible or even desirable. This is often the case with one word and even two word queries in Web search engines. In Web search, there are at least three ways of coping with ambiguity: • Promoting diversity in the search results (Clarke et al., 2008): given the query ”oasis”, the search engine may try to include representatives for different senses of the word (such as the Oasis band, the Organization for the Advancement of Structured Information Standards, the online fashion store, etc.) among the top results. Search engines are supposed to handle diversity as one of the multiple factors that influence the ranking. • Presenting the results as a set of (labelled) clusters rather than as a ranked list (Carpineto et al., 2009). • Complementing search results with search suggestions (e.g. ”oasis band”, ”oasis fashion store”) that serve to refine the query in the intended way (Anick, 2003). All of them rely on the ability of the search engine to cluster search results, detecting topic similarities. In all of them, disambiguation is implicit, a side effect of the process but not its explicit target. Clustering may detect that documents about the Oasis band and the Oasis fashion store deal with unrelated topics, but it may as well detect a group of documents discussing why one of the Oasis band members is leaving the band, and another group of documents about Oasis band lyrics; both are different aspects of the broad topic Oasis band. A perfect hierarchical clustering should distinguish between the different Oasis senses at a first level, and then discover different topics within each of the senses. Is it possible to use sense inventories to improve search results for one word queries? To answer 1357 this question, we will focus on two broad-coverage lexical resources of a different nature: WordNet (Miller et al., 1990), as a de-facto standard used in Word Sense Disambiguation experiments and many other Natural Language Processing research fields; and Wikipedia (www.wikipedia.org), as a large coverage and updated encyclopedic resource which may have a better coverage of relevant senses in Web pages. Our hypothesis is that, under appropriate conditions, any of the above mechanisms (clustering, search suggestions, diversity) might benefit from an explicit disambiguation (classification of pages in the top search results) using a wide-coverage sense inventory. Our research is focused on four relevant aspects of the problem: 1. Coverage: Are Wikipedia/Wordnet senses representative of search results? Otherwise, trying to make a disambiguation in terms of a fixed sense inventory would be meaningless. 2. If the answer to (1) is positive, the reverse question is also interesting: can we estimate search results diversity using our sense inventories? 3. Sense frequencies: knowing sense frequencies in (search results) Web pages is crucial to have a usable sense inventory. Is it possible to estimate Web sense frequencies from currently available information? 4. Classification: The association of Web pages to word senses must be done with some unsupervised algorithm, because it is not possible to hand-tag training material for every possible query word. Can this classification be done accurately? Can it be effective to promote diversity in search results? In order to provide an initial answer to these questions, we have built a corpus consisting of 40 nouns and 100 Google search results per noun, manually annotated with the most appropriate Wordnet and Wikipedia senses. Section 2 describes how this corpus has been created, and in Section 3 we discuss WordNet and Wikipedia coverage of search results according to our testbed. As this initial results clearly discard Wordnet as a sense inventory for the task, the rest of the paper mainly focuses on Wikipedia. In Section 4 we estimate search results diversity from our testbed, finding that the use of Wikipedia could substantially improve diversity in the top results. In Section 5 we use the Wikipedia internal link structure and the number of visits per page to estimate relative frequencies for Wikipedia senses, obtaining an estimation which is highly correlated with actual data in our testbed. Finally, in Section 6 we discuss a few strategies to classify Web pages into word senses, and apply the best classifier to enhance diversity in search results. The paper concludes with a discussion of related work (Section 7) and an overall discussion of our results in Section 8. 2 Test Set 2.1 Set of Words The most crucial step in building our test set is choosing the set of words to be considered. We are looking for words which are susceptible to form a one-word query for a Web search engine, and therefore we should focus on nouns which are used to denote one or more named entities. At the same time we want to have some degree of comparability with previous research on Word Sense Disambiguation, which points to noun sets used in Senseval/SemEval evaluation campaigns1. Our budget for corpus annotation was enough for two persons-month, which limited us to handle 40 nouns (usually enough to establish statistically significant differences between WSD algorithms, although obviously limited to reach solid figures about the general behaviour of words in the Web). With these arguments in mind, we decided to choose: (i) 15 nouns from the Senseval-3 lexical sample dataset, which have been previously employed by (Mihalcea, 2007) in a related experiment (see Section 7); (ii) 25 additional words which satisfy two conditions: they are all ambiguous, and they are all names for music bands in one of their senses (not necessarily the most salient). The Senseval set is: {argument, arm, atmosphere, bank, degree, difference, disc, image, paper, party, performance, plan, shelter, sort, source}. The bands set is {amazon, apple, camel, cell, columbia, cream, foreigner, fox, genesis, jaguar, oasis, pioneer, police, puma, rainbow, shell, skin, sun, tesla, thunder, total, traffic, trapeze, triumph, yes}. For each noun, we looked up all its possible senses in WordNet 3.0 and in Wikipedia (using 1http://senseval.org 1358 Table 1: Coverage of Search Results: Wikipedia vs. WordNet Wikipedia WordNet # senses # documents # senses # documents available/used assigned to some sense available/used assigned to some sense Senseval set 242/100 877 (59%) 92/52 696 (46%) Bands set 640/174 1358 (54%) 78/39 599 (24%) Total 882/274 2235 (56%) 170/91 1295 (32%) Wikipedia disambiguation pages). Wikipedia has an average of 22 senses per noun (25.2 in the Bands set and 16.1 in the Senseval set), and Wordnet a much smaller figure, 4.5 (3.12 for the Bands set and 6.13 for the Senseval set). For a conventional dictionary, a higher ambiguity might indicate an excess of granularity; for an encyclopaedic resource such as Wikipedia, however, it is just an indication of larger coverage. Wikipedia entries for camel which are not in WordNet, for instance, include the Apache Camel routing and mediation engine, the British rock band, the brand of cigarettes, the river in Cornwall, and the World World War I fighter biplane. 2.2 Set of Documents We retrieved the 150 first ranked documents for each noun, by submitting the nouns as queries to a Web search engine (Google). Then, for each document, we stored both the snippet (small description of the contents of retrieved document) and the whole HTML document. This collection of documents contain an implicit new inventory of senses, based on Web search, as documents retrieved by a noun query are associated with some sense of the noun. Given that every document in the top Web search results is supposed to be highly relevant for the query word, we assume a ”one sense per document” scenario, although we allow annotators to assign more than one sense per document. In general this assumption turned out to be correct except in a few exceptional cases (such as Wikipedia disambiguation pages): only nine documents received more than one WordNet sense, and 44 (1.1% of all annotated pages) received more than one Wikipedia sense. 2.3 Manual Annotation We implemented an annotation interface which stored all documents and a short description for every Wordnet and Wikipedia sense. The annotators had to decide, for every document, whether there was one or more appropriate senses in each of the dictionaries. They were instructed to provide annotations for 100 documents per name; if an URL in the list was corrupt or not available, it had to be discarded. We provided 150 documents per name to ensure that the figure of 100 usable documents per name could be reached without problems. Each judge provided annotations for the 4,000 documents in the final data set. In a second round, they met and discussed their independent annotations together, reaching a consensus judgement for every document. 3 Coverage of Web Search Results: Wikipedia vs Wordnet Table 1 shows how Wikipedia and Wordnet cover the senses in search results. We report each noun subset separately (Senseval and bands subsets) as well as aggregated figures. The most relevant fact is that, unsurprisingly, Wikipedia senses cover much more search results (56%) than Wordnet (32%). If we focus on the top ten results, in the bands subset (which should be more representative of plausible web queries) Wikipedia covers 68% of the top ten documents. This is an indication that it can indeed be useful for promoting diversity or help clustering search results: even if 32% of the top ten documents are not covered by Wikipedia, it is still a representative source of senses in the top search results. We have manually examined all documents in the top ten results that are not covered by Wikipedia: a majority of the missing senses consists of names of (generally not well-known) companies (45%) and products or services (26%); the other frequent type (12%) of non annotated document is disambiguation pages (from Wikipedia and also from other dictionaries). It is also interesting to examine the degree of overlap between Wikipedia and Wordnet senses. Being two different types of lexical resource, they might have some degree of complementarity. Table 2 shows, however, that this is not the case: most of the (annotated) documents either fit Wikipedia senses (26%) or both Wikipedia and Wordnet (29%), and just 3% fit Wordnet only. 1359 Table 2: Overlap between Wikipedia and Wordnet in Search Results # documents annotated with Wikipedia & Wordnet Wikipedia only Wordnet only none Senseval set 607 (40%) 270 (18%) 89 (6%) 534 (36%) Bands set 572 (23%) 786 (31%) 27 (1%) 1115 (45%) Total 1179 (29%) 1056 (26%) 116 (3%) 1649 (41%) Therefore, Wikipedia seems to extend the coverage of Wordnet rather than providing complementary sense information. If we wanted to extend the coverage of Wikipedia, the best strategy seems to be to consider lists of companies, products and services, rather than complementing Wikipedia with additional sense inventories. 4 Diversity in Google Search Results Once we know that Wikipedia senses are a representative subset of actual Web senses (covering more than half of the documents retrieved by the search engine), we can test how well search results respect diversity in terms of this subset of senses. Table 3 displays the number of different senses found at different depths in the search results rank, and the average proportion of total senses that they represent. These results suggest that diversity is not a major priority for ranking results: the top ten results only cover, in average, 3 Wikipedia senses (while the average number of senses listed in Wikipedia is 22). When considering the first 100 documents, this number grows up to 6.85 senses per noun. Another relevant figure is the frequency of the most frequent sense for each word: in average, 63% of the pages in search results belong to the most frequent sense of the query word. This is roughly comparable with most frequent sense figures in standard annotated corpora such as Semcor (Miller et al., 1993) and the Senseval/Semeval data sets, which suggests that diversity may not play a major role in the current Google ranking algorithm. Of course this result must be taken with care, because variability between words is high and unpredictable, and we are using only 40 nouns for our experiment. But what we have is a positive indication that Wikipedia could be used to improve diversity or cluster search results: potentially the first top ten results could cover 6.15 different senses in average (see Section 6.5), which would be a substantial growth. 5 Sense Frequency Estimators for Wikipedia Wikipedia disambiguation pages contain no systematic information about the relative importance of senses for a given word. Such information, however, is crucial in a lexicon, because sense distributions tend to be skewed, and knowing them can help disambiguation algorithms. We have attempted to use two estimators of expected sense distribution: • Internal relevance of a word sense, measured as incoming links for the URL of a given sense in Wikipedia. • External relevance of a word sense, measured as the number of visits for the URL of a given sense (as reported in http://stats.grok.se). The number of internal incoming links is expected to be relatively stable for Wikipedia articles. As for the number of visits, we performed a comparison of the number of visits received by the bands noun subset in May, June and July 2009, finding a stable-enough scenario with one notorious exception: the number of visits to the noun Tesla raised dramatically in July, because July 10 was the anniversary of the birth of Nicola Tesla, and a special Google logo directed users to the Wikipedia page for the scientist. We have measured correlation between the relative frequencies derived from these two indicators and the actual relative frequencies in our testbed. Therefore, for each noun w and for each sense wi, we consider three values: (i) proportion of documents retrieved for w which are manually assigned to each sense wi; (ii) inlinks(wi): relative amount of incoming links to each sense wi; and (iii) visits(wi): relative number of visits to the URL for each sense wi. We have measured the correlation between these three values using a linear regression correlation coefficient, which gives a correlation value of .54 for the number of visits and of .71 for the number of incoming links. Both estimators seem 1360 Table 3: Diversity in Search Results according to Wikipedia average # senses in search results average coverage of Wikipedia senses Bands set Senseval set Total Bands set Senseval set Total First 10 docs 2.88 3.2 3.00 .21 .21 .21 First 25 4.44 4.8 4.58 .28 .33 .30 First 50 5.56 5.47 5.53 .33 .36 .34 First 75 6.56 6.33 6.48 .37 .43 .39 First 100 6.96 6.67 6.85 .38 .45 .41 to be positively correlated with real relative frequencies in our testbed, with a strong preference for the number of links. We have experimented with weighted combinations of both indicators, using weights of the form (k, 1 −k), k ∈{0, 0.1, 0.2 . . . 1}, reaching a maximal correlation of .73 for the following weights: freq(wi) = 0.9∗inlinks(wi)+0.1∗visits(wi) (1) This weighted estimator provides a slight advantage over the use of incoming links only (.73 vs .71). Overall, we have an estimator which has a strong correlation with the distribution of senses in our testbed. In the next section we will test its utility for disambiguation purposes. 6 Association of Wikipedia Senses to Web Pages We want to test whether the information provided by Wikipedia can be used to classify search results accurately. Note that we do not want to consider approaches that involve a manual creation of training material, because they can’t be used in practice. Given a Web page p returned by the search engine for the query w, and the set of senses w1 . . . wn listed in Wikipedia, the task is to assign the best candidate sense to p. We consider two different techniques: • A basic Information Retrieval approach, where the documents and the Wikipedia pages are represented using a Vector Space Model (VSM) and compared with a standard cosine measure. This is a basic approach which, if successful, can be used efficiently to classify search results. • An approach based on a state-of-the-art supervised WSD system, extracting training examples automatically from Wikipedia content. We also compute two baselines: • A random assignment of senses (precision is computed as the inverse of the number of senses, for every test case). • A most frequent sense heuristic which uses our estimation of sense frequencies and assigns the same sense (the most frequent) to all documents. Both are naive baselines, but it must be noted that the most frequent sense heuristic is usually hard to beat for unsupervised WSD algorithms in most standard data sets. We now describe each of the two main approaches in detail. 6.1 VSM Approach For each word sense, we represent its Wikipedia page in a (unigram) vector space model, assigning standard tf*idf weights to the words in the document. idf weights are computed in two different ways: 1. Experiment VSM computes inverse document frequencies in the collection of retrieved documents (for the word being considered). 2. Experiment VSM-GT uses the statistics provided by the Google Terabyte collection (Brants and Franz, 2006), i.e. it replaces the collection of documents with statistics from a representative snapshot of the Web. 3. Experiment VSM-mixed combines statistics from the collection and from the Google Terabyte collection, following (Chen et al., 2009). The document p is represented in the same vector space as the Wikipedia senses, and it is compared with each of the candidate senses wi via the cosine similarity metric (we have experimented 1361 with other similarity metrics such as χ2, but differences are irrelevant). The sense with the highest similarity to p is assigned to the document. In case of ties (which are rare), we pick the first sense in the Wikipedia disambiguation page (which in practice is like a random decision, because senses in disambiguation pages do not seem to be ordered according to any clear criteria). We have also tested a variant of this approach which uses the estimation of sense frequencies presented above: once the similarities are computed, we consider those cases where two or more senses have a similar score (in particular, all senses with a score greater or equal than 80% of the highest score). In that cases, instead of using the small similarity differences to select a sense, we pick up the one which has the largest frequency according to our estimator. We have applied this strategy to the best performing system, VSM-GT, resulting in experiment VSM-GT+freq. 6.2 WSD Approach We have used TiMBL (Daelemans et al., 2001), a state-of-the-art supervised WSD system which uses Memory-Based Learning. The key, in this case, is how to extract learning examples from the Wikipedia automatically. For each word sense, we basically have three sources of examples: (i) occurrences of the word in the Wikipedia page for the word sense; (ii) occurrences of the word in Wikipedia pages pointing to the page for the word sense; (iii) occurrences of the word in external pages linked in the Wikipedia page for the word sense. After an initial manual inspection, we decided to discard external pages for being too noisy, and we focused on the first two options. We tried three alternatives: • TiMBL-core uses only the examples found in the page for the sense being trained. • TiMBL-inlinks uses the examples found in Wikipedia pages pointing to the sense being trained. • TiMBL-all uses both sources of examples. In order to classify a page p with respect to the senses for a word w, we first disambiguate all occurrences of w in the page p. Then we choose the sense which appears most frequently in the page according to TiMBL results. In case of ties we pick up the first sense listed in the Wikipedia disambiguation page. We have also experimented with a variant of the approach that uses our estimation of sense frequencies, similarly to what we did with the VSM approach. In this case, (i) when there is a tie between two or more senses (which is much more likely than in the VSM approach), we pick up the sense with the highest frequency according to our estimator; and (ii) when no sense reaches 30% of the cases in the page to be disambiguated, we also resort to the most frequent sense heuristic (among the candidates for the page). This experiment is called TiMBL-core+freq (we discarded ”inlinks” and ”all” versions because they were clearly worse than ”core”). 6.3 Classification Results Table 4 shows classification results. The accuracy of systems is reported as precision, i.e. the number of pages correctly classified divided by the total number of predictions. This is approximately the same as recall (correctly classified pages divided by total number of pages) for our systems, because the algorithms provide an answer for every page containing text (actual coverage is 94% because some pages only contain text as part of an image file such as photographs and logotypes). Table 4: Classification Results Experiment Precision random .19 most frequent sense (estimation) .46 TiMBL-core .60 TiMBL-inlinks .50 TiMBL-all .58 TiMBL-core+freq .67 VSM .67 VSM-GT .68 VSM-mixed .67 VSM-GT+freq .69 All systems are significantly better than the random and most frequent sense baselines (using p < 0.05 for a standard t-test). Overall, both approaches (using TiMBL WSD machinery and using VSM) lead to similar results (.67 vs. .69), which would make VSM preferable because it is a simpler and more efficient approach. Taking a 1362 Figure 1: Precision/Coverage curves for VSM-GT+freq classification algorithm closer look at the results with TiMBL, there are a couple of interesting facts: • There is a substantial difference between using only examples taken from the Wikipedia Web page for the sense being trained (TiMBL-core, .60) and using examples from the Wikipedia pages pointing to that page (TiMBL-inlinks, .50). Examples taken from related pages (even if the relationship is close as in this case) seem to be too noisy for the task. This result is compatible with findings in (Santamar´ıa et al., 2003) using the Open Directory Project to extract examples automatically. • Our estimation of sense frequencies turns out to be very helpful for cases where our TiMBL-based algorithm cannot provide an answer: precision rises from .60 (TiMBLcore) to .67 (TiMBL-core+freq). The difference is statistically significant (p < 0.05) according to the t-test. As for the experiments with VSM, the variations tested do not provide substantial improvements to the baseline (which is .67). Using idf frequencies obtained from the Google Terabyte corpus (instead of frequencies obtained from the set of retrieved documents) provides only a small improvement (VSM-GT, .68), and adding the estimation of sense frequencies gives another small improvement (.69). Comparing the baseline VSM with the optimal setting (VSM-GT+freq), the difference is small (.67 vs .69) but relatively robust (p = 0.066 according to the t-test). Remarkably, the use of frequency estimations is very helpful for the WSD approach but not for the SVM one, and they both end up with similar performance figures; this might indicate that using frequency estimations is only helpful up to certain precision ceiling. 6.4 Precision/Coverage Trade-off All the above experiments are done at maximal coverage, i.e., all systems assign a sense for every document in the test collection (at least for every document with textual content). But it is possible to enhance search results diversity without annotating every document (in fact, not every document can be assigned to a Wikipedia sense, as we have discussed in Section 3). Thus, it is useful to investigate which is the precision/coverage trade-off in our dataset. We have experimented with the best performing system (VSM-GT+freq), introducing a similarity threshold: assignment of a document to a sense is only done if the similarity of the document to the Wikipedia page for the sense exceeds the similarity threshold. We have computed precision and coverage for every threshold in the range [0.00 −0.90] (beyond 0.90 coverage was null) and represented the results in Figure 1 (solid line). The graph shows that we 1363 can classify around 20% of the documents with a precision above .90, and around 60% of the documents with a precision of .80. Note that we are reporting disambiguation results using a conventional WSD test set, i.e., one in which every test case (every document) has been manually assigned to some Wikipedia sense. But in our Web Search scenario, 44% of the documents were not assigned to any Wikipedia sense: in practice, our classification algorithm would have to cope with all this noise as well. Figure 1 (dotted line) shows how the precision/coverage curve is affected when the algorithm attempts to disambiguate all documents retrieved by Google, whether they can in fact be assigned to a Wikipedia sense or not. At a coverage of 20%, precision drops approximately from .90 to .70, and at a coverage of 60% it drops from .80 to .50. We now address the question of whether this performance is good enough to improve search results diversity in practice. 6.5 Using Classification to Promote Diversity We now want to estimate how the reported classification accuracy may perform in practice to enhance diversity in search results. In order to provide an initial answer to this question, we have re-ranked the documents for the 40 nouns in our testbed, using our best classifier (VSM-GT+freq) and making a list of the top-ten documents with the primary criterion of maximising the number of senses represented in the set, and the secondary criterion of maximising the similarity scores of the documents to their assigned senses. The algorithm proceeds as follows: we fill each position in the rank (starting at rank 1), with the document which has the highest similarity to some of the senses which are not yet represented in the rank; once all senses are represented, we start choosing a second representative for each sense, following the same criterion. The process goes on until the first ten documents are selected. We have also produced a number of alternative rankings for comparison purposes: • clustering (centroids): this method applies Hierarchical Agglomerative Clustering – which proved to be the most competitive clustering algorithm in a similar task (Artiles et al., 2009) – to the set of search results, forcing the algorithm to create ten clusters. The centroid of each cluster is then selected Table 5: Enhancement of Search Results Diversity rank@10 # senses coverage Original rank 2.80 49% Wikipedia 4.75 77% clustering (centroids) 2.50 42% clustering (top ranked) 2.80 46% random 2.45 43% upper bound 6.15 97% as one of the top ten documents in the new rank. • clustering (top ranked): Applies the same clustering algorithm, but this time the top ranked document (in the original Google rank) of each cluster is selected. • random: Randomly selects ten documents from the set of retrieved results. • upper bound: This is the maximal diversity that can be obtained in our testbed. Note that coverage is not 100%, because some words have more than ten meanings in Wikipedia and we are only considering the top ten documents. All experiments have been applied on the full set of documents in the testbed, including those which could not be annotated with any Wikipedia sense. Coverage is computed as the ratio of senses that appear in the top ten results compared to the number of senses that appear in all search results. Results are presented in Table 5. Note that diversity in the top ten documents increases from an average of 2.80 Wikipedia senses represented in the original search engine rank, to 4.75 in the modified rank (being 6.15 the upper bound), with the coverage of senses going from 49% to 77%. With a simple VSM algorithm, the coverage of Wikipedia senses in the top ten results is 70% larger than in the original ranking. Using Wikipedia to enhance diversity seems to work much better than clustering: both strategies to select a representative from each cluster are unable to improve the diversity of the original ranking. Note, however, that our evaluation has a bias towards using Wikipedia, because only Wikipedia senses are considered to estimate diversity. Of course our results do not imply that the Wikipedia modified rank is better than the original 1364 Google rank: there are many other factors that influence the final ranking provided by a search engine. What our results indicate is that, with simple and efficient algorithms, Wikipedia can be used as a reference to improve search results diversity for one-word queries. 7 Related Work Web search results clustering and diversity in search results are topics that receive an increasing attention from the research community. Diversity is used both to represent sub-themes in a broad topic, or to consider alternative interpretations for ambiguous queries (Agrawal et al., 2009), which is our interest here. Standard IR test collections do not usually consider ambiguous queries, and are thus inappropriate to test systems that promote diversity (Sanderson, 2008); it is only recently that appropriate test collections are being built, such as (Paramita et al., 2009) for image search and (Artiles et al., 2009) for person name search. We see our testbed as complementary to these ones, and expect that it can contribute to foster research on search results diversity. To our knowledge, Wikipedia has not explicitly been used before to promote diversity in search results; but in (Gollapudi and Sharma, 2009), it is used as a gold standard to evaluate diversification algorithms: given a query with a Wikipedia disambiguation page, an algorithm is evaluated as promoting diversity when different documents in the search results are semantically similar to different Wikipedia pages (describing the alternative senses of the query). Although semantic similarity is measured automatically in this work, our results confirm that this evaluation strategy is sound, because Wikipedia senses are indeed representative of search results. (Clough et al., 2009) analyses query diversity in a Microsoft Live Search, using click entropy and query reformulation as diversity indicators. It was found that at least 9.5% - 16.2% of queries could benefit from diversification, although no correlation was found between the number of senses of a word in Wikipedia and the indicators used to discover diverse queries. This result does not discard, however, that queries where applying diversity is useful cannot benefit from Wikipedia as a sense inventory. In the context of clustering, (Carmel et al., 2009) successfully employ Wikipedia to enhance automatic cluster labeling, finding that Wikipedia labels agree with manual labels associated by humans to a cluster, much more than with significant terms that are extracted directly from the text. In a similar line, both (Gabrilovich and Markovitch, 2007) and (Syed et al., 2008) provide evidence suggesting that categories of Wikipedia articles can successfully describe common concepts in documents. In the field of Natural Language Processing, there has been successful attempts to connect Wikipedia entries to Wordnet senses: (RuizCasado et al., 2005) reports an algorithm that provides an accuracy of 84%. (Mihalcea, 2007) uses internal Wikipedia hyperlinks to derive sensetagged examples. But instead of using Wikipedia directly as sense inventory, Mihalcea then manually maps Wikipedia senses into Wordnet senses (claiming that, at the time of writing the paper, Wikipedia did not consistently report ambiguity in disambiguation pages) and shows that a WSD system based on acquired sense-tagged examples reaches an accuracy well beyond an (informed) most frequent sense heuristic. 8 Conclusions We have investigated whether generic lexical resources can be used to promote diversity in Web search results for one-word, ambiguous queries. We have compared WordNet and Wikipedia and arrived to a number of conclusions: (i) unsurprisingly, Wikipedia has a much better coverage of senses in search results, and is therefore more appropriate for the task; (ii) the distribution of senses in search results can be estimated using the internal graph structure of the Wikipedia and the relative number of visits received by each sense in Wikipedia, and (iii) associating Web pages to Wikipedia senses with simple and efficient algorithms, we can produce modified rankings that cover 70% more Wikipedia senses than the original search engine rankings. We expect that the testbed created for this research will complement the - currently short - set of benchmarking test sets to explore search results diversity and query ambiguity. Our testbed is publicly available for research purposes at http://nlp.uned.es. Our results endorse further investigation on the use of Wikipedia to organize search results. Some limitations of our research, however, must be 1365 noted: (i) the nature of our testbed (with every search result manually annotated in terms of two sense inventories) makes it too small to extract solid conclusions on Web searches (ii) our work does not involve any study of diversity from the point of view of Web users (i.e. when a Web query addresses many different use needs in practice); research in (Clough et al., 2009) suggests that word ambiguity in Wikipedia might not be related with diversity of search needs; (iii) we have tested our classifiers with a simple re-ordering of search results to test how much diversity can be improved, but a search results ranking depends on many other factors, some of them more crucial than diversity; it remains to be tested how can we use document/Wikipedia associations to improve search results clustering (for instance, providing seeds for the clustering process) and to provide search suggestions. Acknowledgments This work has been partially funded by the Spanish Government (project INES/Text-Mess) and the Xunta de Galicia. References R. Agrawal, S. Gollapudi, A. Halverson, and S. Leong. 2009. Diversifying Search Results. In Proc. of WSDM’09. ACM. P. Anick. 2003. Using Terminological Feedback for Web Search Refinement : a Log-based Study. In Proc. ACM SIGIR 2003, pages 88–95. ACM New York, NY, USA. J. Artiles, J. Gonzalo, and S. Sekine. 2009. WePS 2 Evaluation Campaign: overview of the Web People Search Clustering Task. In 2nd Web People Search Evaluation Workshop (WePS 2009), 18th WWW Conference. 2009. T. Brants and A. Franz. 2006. Web 1T 5-gram, version 1. Philadelphia: Linguistic Data Consortium. D. Carmel, H. Roitman, and N. Zwerdling. 2009. Enhancing Cluster Labeling using Wikipedia. In Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, pages 139–146. ACM. C. Carpineto, S. Osinski, G. Romano, and Dawid Weiss. 2009. A Survey of Web Clustering Engines. ACM Computing Surveys, 41(3). Y. Chen, S. Yat Mei Lee, and C. Huang. 2009. PolyUHK: A Robust Information Extraction System for Web Personal Names. In Proc. WWW’09 (WePS2 Workshop). ACM. C. Clarke, M. Kolla, G. Cormack, O. Vechtomova, A. Ashkan, S. B¨uttcher, and I. MacKinnon. 2008. Novelty and Diversity in Information Retrieval Evaluation. In Proc. SIGIR’08, pages 659–666. ACM. P. Clough, M. Sanderson, M. Abouammoh, S. Navarro, and M. Paramita. 2009. Multiple Approaches to Analysing Query Diversity. In Proc. of SIGIR 2009. ACM. W. Daelemans, J. Zavrel, K. van der Sloot, and A. van den Bosch. 2001. TiMBL: Tilburg Memory Based Learner, version 4.0, Reference Guide. Technical report, University of Antwerp. E. Gabrilovich and S. Markovitch. 2007. Computing Semantic Relatedness using Wikipedia-based Explicit Semantic Analysis. In Proceedings of The 20th International Joint Conference on Artificial Intelligence (IJCAI), Hyderabad, India. S. Gollapudi and A. Sharma. 2009. An Axiomatic Approach for Result Diversification. In Proc. WWW 2009, pages 381–390. ACM New York, NY, USA. R. Mihalcea. 2007. Using Wikipedia for Automatic Word Sense Disambiguation. In Proceedings of NAACL HLT, volume 2007. G. Miller, C. R. Beckwith, D. Fellbaum, Gross, and K. Miller. 1990. Wordnet: An on-line lexical database. International Journal of Lexicograph, 3(4). G.A Miller, C. Leacock, R. Tengi, and Bunker R. T. 1993. A Semantic Concordance. In Proceedings of the ARPA WorkShop on Human Language Technology. San Francisco, Morgan Kaufman. M. Paramita, M. Sanderson, and P. Clough. 2009. Diversity in Photo Retrieval: Overview of the ImageCLEFPhoto task 2009. CLEF working notes, 2009. M. Ruiz-Casado, E. Alfonseca, and P. Castells. 2005. Automatic Assignment of Wikipedia Encyclopaedic Entries to Wordnet Synsets. Advances in Web Intelligence, 3528:380–386. M. Sanderson. 2000. Retrieving with Good Sense. Information Retrieval, 2(1):49–69. M. Sanderson. 2008. Ambiguous Queries: Test Collections Need More Sense. In Proceedings of the 31st annual international ACM SIGIR conference on Research and development in information retrieval, pages 499–506. ACM New York, NY, USA. C. Santamar´ıa, J. Gonzalo, and F. Verdejo. 2003. Automatic Association of Web Directories to Word Senses. Computational Linguistics, 29(3):485–502. Z. S. Syed, T. Finin, and Joshi. A. 2008. Wikipedia as an Ontology for Describing Documents. In Proc. ICWSM’08. 1366
|
2010
|
138
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1367–1375, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics A Unified Graph Model for Sentence-based Opinion Retrieval Binyang Li, Lanjun Zhou, Shi Feng, Kam-Fai Wong Department of Systems Engineering and Engineering Management The Chinese University of Hong Kong {byli, ljzhou, sfeng, kfwong}@se.cuhk.edu.hk Abstract There is a growing research interest in opinion retrieval as on-line users’ opinions are becoming more and more popular in business, social networks, etc. Practically speaking, the goal of opinion retrieval is to retrieve documents, which entail opinions or comments, relevant to a target subject specified by the user’s query. A fundamental challenge in opinion retrieval is information representation. Existing research focuses on document-based approaches and documents are represented by bag-of-word. However, due to loss of contextual information, this representation fails to capture the associative information between an opinion and its corresponding target. It cannot distinguish different degrees of a sentiment word when associated with different targets. This in turn seriously affects opinion retrieval performance. In this paper, we propose a sentence-based approach based on a new information representation, namely topic-sentiment word pair, to capture intra-sentence contextual information between an opinion and its target. Additionally, we consider inter-sentence information to capture the relationships among the opinions on the same topic. Finally, the two types of information are combined in a unified graph-based model, which can effectively rank the documents. Compared with existing approaches, experimental results on the COAE08 dataset showed that our graph-based model achieved significant improvement. 1 Introduction In recent years, there is a growing interest in sharing personal opinions on the Web, such as product reviews, economic analysis, political polls, etc. These opinions cannot only help independent users make decisions, but also obtain valuable feedbacks (Pang et al., 2008). Opinion oriented research, including sentiment classification, opinion extraction, opinion question answering, and opinion summarization, etc. are receiving growing attention (Wilson, et al., 2005; Liu et al., 2005; Oard et al., 2006). However, most existing works concentrate on analyzing opinions expressed in the documents, and none on how to represent the information needs required to retrieve opinionated documents. In this paper, we focus on opinion retrieval, whose goal is to find a set of documents containing not only the query keyword(s) but also the relevant opinions. This requirement brings about the challenge on how to represent information needs for effective opinion retrieval. In order to solve the above problem, previous work adopts a 2-stage approach. In the first stage, relevant documents are determined and ranked by a score, i.e. tf-idf value. In the second stage, an opinion score is generated for each relevant document (Macdonald and Ounis, 2007; Oard et al., 2006). The opinion score can be acquired by either machine learning-based sentiment classifiers, such as SVM (Zhang and Yu, 2007), or a sentiment lexicons with weighted scores from training documents (Amati et al., 2007; Hannah et al., 2007; Na et al., 2009). Finally, an overall score combining the two is computed by using a score function, e.g. linear combination, to re-rank the retrieved documents. Retrieval in the 2-stage approach is based on document and document is represented by bag-of-word. This representation, however, can only ensure that there is at least one opinion in each relevant document, but it cannot determine the relevance pairing of individual opinion to its target. In general, by simply representing a document in bag-of-word, contextual information i.e. the corresponding target of an opinion, is neglected. This may result in possible mismatch between an opinion and a target and in turn affects opinion retrieval performance. By the same token, the effect to documents consisting of mul1367 tiple topics, which is common in blogs and on-line reviews, is also significant. In this setting, even if a document is regarded opinionated, it cannot ensure that all opinions in the document are indeed relevant to the target concerned. Therefore, we argue that existing information representation i.e. bag-of-word, cannot satisfy the information needs for opinion retrieval. In this paper, we propose to handle opinion retrieval in the granularity of sentence. It is observed that a complete opinion is always expressed in one sentence, and the relevant target of the opinion is mostly the one found in it. Therefore, it is crucial to maintain the associative information between an opinion and its target within a sentence. We define the notion of a topic-sentiment word pair, which is composed of a topic term (i.e. the target) and a sentiment word (i.e. opinion) of a sentence. Word pairs can maintain intra-sentence contextual information to express the potential relevant opinions. In addition, inter-sentence contextual information is also captured by word pairs to represent the relationship among opinions on the same topic. In practice, the inter-sentence information reflects the degree of a word pair. Finally, we combine both intra-sentence and inter-sentence contextual information to construct a unified undirected graph to achieve effective opinion retrieval. The rest of the paper is organized as follows. In Section 2, we describe the motivation of our approach. Section 3 presents a novel unified graph-based model for opinion retrieval. We evaluated our model and the results are presented in Section 4. We review related works on opinion retrieval in Section 5. Finally, in Section 6, the paper is concluded and future work is suggested. 2 Motivation In this section, we start from briefly describing the objective of opinion retrieval. We then illustrate the limitations of current opinion retrieval approaches, and analyze the motivation of our method. 2.1 Formal Description of Problem Opinion retrieval was first presented in the TREC 2006 Blog track, and the objective is to retrieve documents that express an opinion about a given target. The opinion target can be a “traditional” named entity (e.g. a name of person, location, or organization, etc.), a concept (e.g. a type of technology), or an event (e.g. presidential election). The topic of the document is not required to be the same as the target, but an opinion about the target has to be presented in the document or one of the comments to the document (Macdonald and Ounis, 2006). Therefore, in this paper we regard the information needs for opinion retrieval as relevant opinion. 2.2 Motivation of Our Approach In traditional information retrieval (IR) bag-of-word representation is the most common way to express information needs. However, in opinion retrieval, information need target at relevant opinion, and this renders bag-of-word representation ineffective. Consider the example in Figure 1. There are three sentences A, B, and C in a document di. Now given an opinion-oriented query Q related to ‘Avatar’. According to the conventional 2-stage opinion retrieval approach, di is represented by a bag-of-word. Among the words, there is a topic term Avatar (t1) occurring twice, i.e. Avatar in A and Avatar in C, and two sentiment words comfortable (o1) and favorite (o2) (refer to Figure 2 (a)). In order to rank this document, an overall score of the document di is computed by a simple combination of the relevant score ( ܵܿݎ݁୰ୣ୪) and the opinion score (ܵܿݎ݁୭୮), e.g. equal weighted linear combination, as follows. ܵܿݎ݁ୢ୭ୡൌܵܿݎ݁୰ୣ୪ ܵܿݎ݁୭୮ For simplicity, we let ܵܿݎ݁୰ୣ୪ൌݐ݂୕ൈ݂݅݀୕, and ܵܿݎ݁୭୮ be computed by using lexicon-based method: ܵܿݎ݁୭୮ൌݓ݁݅݃ℎݐୡ୭୫୭୰୲ୟୠ୪ୣݓ݁݅݃ℎݐୟ୴୭୰୧୲ୣ. Figure 1: A retrieved document di on the target ‘Avatar’. Although bag-of-word representation achieves good performance in retrieving relevant documents, our study shows that it cannot satisfy the information needs for retrieval of relevant opinion. It suffers from the following limitations: (1) It cannot maintain contextual information; thus, an opinion may not be related to the target of the retrieved document is neglected. In this example, only the opinion favorite (o2) on Avatar in C is the relevant opinion. But due to loss of contextual information between the opinion and its corresponding target, Avatar in A and comA. 阿凡达明日将在中国上映。 Tomorrow, Avatar will be shown in China. B. 我预订到了IMAX 影院中最舒服的位子。 I’ve reserved a comfortable seat in IMAX. C. 阿凡达是我最喜欢的一部3D 电影。 Avatar is my favorite 3D movie. 1368 fortable (o1) are also regarded as relevant opinion mistakenly, creating a false positive. In reality comfortable (o1) describes “the seats in IMAX”, which is an irrelevant opinion, and sentence A is a factual statement rather than an opinion statement. (a) (b) Figure 2: Two kinds of information representation of opinion retrieval. (t1=‘Avatar’ o1= ‘comfortable’, o2=‘favorite’) (1) Current approaches cannot capture the relationship among opinions about the same topic. Suppose there is another document including sentence C which expresses the same opinion on Avatar. Existing information representation simply does not cater for the two identical opinions from different documents. In addition, if many documents contain opinions on Avatar, the relationship among them is not clearly represented by existing approaches. In this paper, we process opinion retrieval in the granularity of sentence as we observe that a complete opinion always exists within a sentence (refer to Figure 2 (b)). To represent a relevant opinion, we define the notion of topic-sentiment word pair, which consists of a topic term and a sentiment word. A word pair maintains the associative information between the two words, and enables systems to draw up the relationship among all the sentences with the same opinion on an identical target. This relationship information can identify all documents with sentences including the sentiment words and to determine the contributions of such words to the target (topic term). Furthermore, based on word pairs, we designed a unified graph-based method for opinion retrieval (see later in Section 3). 3 Graph-based model 3.1 Basic Idea Different from existing approaches which simply make use of document relevance to reflect the relevance of opinions embedded in them, our approach concerns more on identifying the relevance of individual opinions. Intuitively, we believed that the more relevant opinions appear in a document, the more relevant is that document for subsequent opinion analysis operations. Further, since the lexical scope of an opinion does not usually go beyond a sentence, we propose to handle opinion retrieval in the granularity of sentence. Without loss of generality, we assume that there is a document set ࣞൌሼ݀ଵ, ݀ଶ, ݀ଷ, ڮ , ݀ሽ, and a specific query ࣫ൌሼݍଵ, ݍଶ, ݍଷ, ڮ , ݍ௭ሽ, where ݍଵ, ݍଶ, ݍଷ, ڮ , ݍ௭ are query keywords. Opinion retrieval aims at retrieving documents from ࣞ with relevant opinion about the query ࣫. In addition, we construct a sentiment word lexicon ܸ and a topic term lexicon ܸ௧ (see Section 4). To maintain the associative information between the target and the opinion, we consider the document set as a bag of sentences, and define a sentence set as ࣭ൌሼݏଵ, ݏଶ, ݏଷ, ڮ , ݏேሽ. For each sentence, we capture the intra-sentence information through the topic-sentiment word pair. Definition 1. topic-sentiment word pair consists of two elements, one is from ܸ௧, and the other one is from ܸ. ൌ൛൏ݐ, |ݐא ܸ௧ , א ܸሻൟ. The topic term from ܸ௧ determines relevance by the query term matching, and the sentiment word from ܸ is used to express an opinion. We use the word pair to maintain the associative information between the topic term and the opinion word (also referred to as sentiment word). The word pair is used to identify a relevant opinion in a sentence. In Figure 2 (b), t1, i.e. Avatar in C, is a topic term relevant to the query, and o2 (‘favorite’) is supposed to be an opinion; and the word pair < t1, o2> indicates sentence C contains a relevant opinion. Similarly, we map each sentence in word pairs by the following rule, and express the intra-sentence information using word pairs. For each sentiment word of a sentence, we choose the topic term with minimum distance as the other element of the word pair: ݏ՜ ൛൏ݐ, |ݐൌminܦ݅ݏݐ൫ݐ, ൯ for each ൟ According to the mapping rule, although a sentence may give rise to a number of word pairs, only the pair with the minimum word distance is selected. We do not take into consideration of the other words in a sentence as relevant opinions are generally formed in close proximity. A sentence is regarded non-opinionated unless it contains at least one word pair. In practice, not all word pairs carry equal weights to express a relevant opinion as the contribution of an opinion word differs from different target topics, and vice versa. For example, the word pair < t1, o2> should be more probable as a relevant opinion than < t1, o1>. To consider 1369 that, inter-sentence contextual information is explored. This is achieved by assigning a weight to each word pair to measure their associative degrees to different queries. We believe that the more a word pair appears the higher should be the weight between the opinion and the target in the context. We will describe how to utilize intra-sentence contextual information to express relevant opinion, and inter-sentence information to measure the degree of each word pair through a graph-based model in the following section. 3.2 HITS Model We propose an opinion retrieval model based on HITS, a popular graph ranking algorithm (Kleinberg, 1999). By considering both intra-sentence information and inter-sentence information, we can determine the weight of a word pair and rank the documents. HITS algorithm distinguishes hubs and authorities in objects. A hub object has links to many authorities. An authority object, which has high-quality content, would have many hubs linking to it. The hub scores and authority scores are computed in an iterative way. Our proposed opinion retrieval model contains two layers. The upper level contains all the topic-sentiment word pairs ൌ൛൏ݐ, |ݐא ܸ௧, א ܸሻൟ. The lower level contains all the documents to be retrieved. Figure 3 gives the bipartite graph representation of the HITS model. Figure 3: Bipartite link graph. For our purpose, the word pairs layer is considered as hubs and the documents layer authorities. If a word pair occurs in one sentence of a document, there will be an edge between them. In Figure 3, we can see that the word pair that has links to many documents can be assigned a high weight to denote a strong associative degree between the topic term and a sentiment word, and it likely expresses a relevant opinion. On the other hand, if a document has links to many word pairs, the document is with many relevant opinions, and it will result in high ranking. Formally, the representation for the bipartite graph is denoted as ܩൌ൏ܪ, ܣௗ, ܧௗ, where ܪൌሼሽ is the set of all pairs of topic words and sentiment words, which appear in one sentence. ܣௗൌሼ݀ ሽ is the set of documents. ܧௗൌሼ݁ |א ܪ, ݀א ܣௗሽ corresponds to the connection between documents and topic-sentiment word pairs. Each edge ݁ is associated with a weight ݓ א ሾ0,1ሿ denoting the contribution of to the document ݀. The weight ݓ is computed by the contribution of word pair in all sentences of ݀ as follows: ݓ ൌ ଵ |ௗೖ| ∑ ൣߣ· ݎ݈݁ሺݐ, ݏሻሺ1 െߣሻ݊൫, ݏ൯൧ ೕא௦אௗೖ ሺ1ሻ |݀| is the number of sentences in ݀; ߣ is introduced as the trade-off parameter to balance the ݎ݈݁ሺݐ, ݏሻ and ݊൫, ݏ൯; ݎ݈݁ሺݐ, ݏሻ is computed to judge the relevance of ݐ in ݏ which belongs to ݀; ݎ݈݁ሺݐ, ݏሻൌݐ݂௧,௦ൈ݅ݏ݂௧ (2) where ݐ݂௧,௦ౢ is the number of ݐ appears in ݏ, and ݅ݏ݂௧ൌlogሺ ேାଵ .ହା௦ ሻ (3) where ݏ݂௧ is the number of sentences that the word ݐ appears in. ݊൫, ݏ൯ is the contribution of in ݏ which belongs to ݀. ݊൫, ݏ൯ൌ ݐ݂݆,ݏ݈ ݐ݂݆,ݏ݈0.5ሺ1.5ൈ ݈݁݊ሺݏ݈ሻ ܽݏ݈ሻ (4) where ܽݏ݈ is the average number of sentences in ݀; ݐ݂௧,ೕ is the number of appears in ݏ (Allan et al., 2003; Otterbacher et al., 2005). It is found that the contribution of a sentiment word will not decrease even if it appears in all the sentences. Therefore in Equation 4, we just use the length of a sentence instead of ݅ݏ݂ೕ to normalize long sentences which would likely contain more sentiment words. The authority score ܣݑݐ݄ܵܿݎ݁ሺ்ାଵሻሺ݀ሻ of document ݀ and a hub score ܪݑܾܵܿݎ݁ሺ்ାଵሻሺሻ of at the ሺܶ1ሻ୲୦ iteration are computed based on the hub scores and authority scores in the ܶ୲୦ iteration as follows. ܣݑݐ݄ܵܿݎ݁ሺ்ାଵሻሺ݀ሻൌ∑ ݓ ൈܪݑܾܵܿݎ்݁ሺሻ ౠאு౦ (5) ܪݑܾܵܿݎ݁ሺ்ାଵሻ൫൯ൌ∑ ݓ ൈܣݑݐ݄ܵܿݎ்݁ሺ݀ሻ ௗౡאౚ (6) We let ܮൌ൫ܮ,൯|ு|ൈ|| denote the adjacency matrix. ܽറሺ்ାଵሻൌܮ݄ሬറሺ்ሻ (7) ݄ሬറሺ்ାଵሻൌܮ்ܽറሺ்ሻ (8) where ܽറሺ்ሻൌሾܣݑݐ݄ܵܿݎ்݁ሺ݀ሻሿ||ൈଵ is the vector of authority scores for documents at the ܶ୲୦ iteration and ݄ሬറሺ்ሻൌሾܪݑܾܵܿݎ்݁ሺሻሿ|ு|ൈଵ is the vector of hub scores for the word pairs at ܶ୲୦ iteration. In order to ensure convergence of the iterative form, ܽറ and ݄ሬറ are normalized in each iteration cycle. 1370 For computation of the final scores, the initial scores of all documents are set to ଵ √, and topic-sentiment word pairs are set to ଵ √ൈெ. The above iterative steps are then used to compute the new scores until convergence. Usually the convergence of the iteration algorithm is achieved when the difference between the scores computed at two successive iterations for any nodes falls below a given threshold (Wan et al., 2008; Li et al., 2009; Erkan and Radev, 2004). In our model, we use the hub scores to denote the associative degree of each word pair and the authority scores as the total scores. The documents are then ranked based on the total scores. 4 Experiment We performed the experiments on the Chinese benchmark dataset to verify our proposed approach for opinion retrieval. We first tested the effect of the parameter ߣ of our model. To demonstrate the effectiveness of our opinion retrieval model, we compared its performance with the same of other approaches. In addition, we studied each individual query to investigate the influence of query to our model. Furthermore, we showed the top-5 highest weight word pairs of 5 queries to further demonstrate the effect of word pair. 4.1 Experiment Setup 4.1.1 Benchmark Datasets Our experiments are based on the Chinese benchmark dataset, COAE08 (Zhao et al., 2008). COAE dataset is the benchmark data set for the opinion retrieval track in the Chinese Opinion Analysis Evaluation (COAE) workshop, consisting of blogs and reviews. 20 queries are provided in COAE08. In our experiment, we created relevance judgments through pooling method, where documents are ranked at different levels: irrelevant, relevant but without opinion, and relevant with opinion. Since polarity is not considered, all relevant documents with opinion are classified into the same level. 4.1.2 Sentiment Lexicon In our experiment, the sentiment lexicon is composed by the following resources (Xu et al., 2007): (1) The Lexicon of Chinese Positive Words, which consists of 5,054 positive words and the Lexicon of Chinese Negative Words, which consists of 3,493 negative words; (2) The opinion word lexicon provided by National Taiwan University which consists of 2,812 positive words and 8,276 negative words; (3) Sentiment word lexicon and comment word lexicon from Hownet. It contains 1836 positive sentiment words, 3,730 positive comments, 1,254 negative sentiment words and 3,116 negative comment words. The different graphemes corresponding to Traditional Chinese and Simplified Chinese are both considered so that the sentiment lexicons from different sources are applicable to process Simplified Chinese text. The lexicon was manually verified. 4.1.3 Topic Term Collection In order to acquire the collection of topic terms, we adopt two expansion methods, dictionary-based method and pseudo relevance feedback method. The dictionary-based method utilizes Wikipedia (Popescu and Etzioni, 2005) to find an entry page for a phrase or a single term in a query. If such an entry exists, all titles of the entry page are extracted as synonyms of the query concept. For example, if we search “绿坝” (Green Tsunami, a firewall) in Wikipedia, it is re-directed to an entry page titled “花季护航” (Youth Escort). This term is then added as a synonym of “绿坝” (Green Tsunami) in the query. Synonyms are treated the same as the original query terms in a retrieval process. The content words in the entry page are ranked by their frequencies in the page. The top-k terms are returned as potential expanded topic terms. The second query expansion method is a web-based method. It is similar to the pseudo relevance feedback expansion but using web documents as the document collection. The query is submitted to a web search engine, such as Google, which returns a ranked list of documents. In the top-n documents, the top-m topic terms which are highly correlated to the query terms are returned. 4.2 Performance Evaluation 4.2.1 Parameter Tuning We first studied how the parameter ߣ (see Equation 1) influenced the mean average precision (MAP) in our model. The result is given in Figure 4. 1371 Figure 4: Performance of MAP with varying ߣ. Best MAP performance was achieved in COAE08 evaluation, when ߣ was set between 0.4 and 0.6. Therefore, in the following experiments, we set ߣൌ0.4. 4.2.2 Opinion Retrieval Model Comparison To demonstrate the effectiveness of our proposed model, we compared it with the following models using different evaluation metrics: (1) IR: We adopted a classical information retrieval model, and further assumed that all retrieved documents contained relevant opinions. (2) Doc: The 2-stage document-based opinion retrieval model was adopted. The model used sentiment lexicon-based method for opinion identification and a conventional information retrieval method for relevance detection. (3) ROSC: This was the model which achieved the best run in TREC Blog 07. It employed machine learning method to identify opinions for each sentence, and to determine the target topic by a NEAR operator. (4) ROCC: This model was similar to ROSC, but it considered the factor of sentence and regarded the count of relevant opinionated sentence to be the opinion score (Zhang and Yu, 2007). In our experiment, we treated this model as the evaluation baseline. (5) GORM: our proposed graph-based opinion retrieval model. Approach COAE08 Evaluation metrics Run id MAP R-pre bPref P@10 IR 0.2797 0.3545 0.2474 0.4868 Doc 0.3316 0.3690 0.3030 0.6696 ROSC 0.3762 0.4321 0.4162 0.7089 Baseline 0.3774 0.4411 0.4198 0.6931 GORM 0.3978 0.4835 0.4265 0.7309 Table 1: Comparison of different approaches on COAE08 dataset, and the best is highlighted. Most of the above models were originally designed for opinion retrieval in English, and re-designed them to handle Chinese opinionated documents. We incorporated our own Chinese sentiment lexicon for this purpose. In our experiments, in addition to MAP, other metrics such as R-precision (R-prec), binary Preference (bPref) and Precision at 10 documents (P@10) were also used. The evaluation results based on these metrics are shown in Table 1. Table 1 summarized the results obtained. We found that GORM achieved the best performance in all the evaluation metrics. Our baseline, ROSC and GORM which were sentence-based approaches achieved better performance than the document-based approaches by 20% in average. Moreover, our GORM approach did not use machine learning techniques, but it could still achieve outstanding performance. To study GORM influenced by different queries, the MAP from median average precision on individual topic was shown in Figure 5. Figure 5: Difference of MAP from Median on COAE08 dataset. (MAP of Median is 0.3724) As shown in Figure 5, the MAP performance was very low on topic 8 and topic 11. Topic 8, i.e. ‘成龙’ (Jackie Chan), it was influenced by topic 7, i.e. ‘李连杰’ (Jet Lee) as there were a number of similar relevant targets for the two topics, and therefore many word pairs ended up the same. As a result, documents belonging to topic 7 and topic 8 could not be differentiated, and they both performed badly. In order to solve this problem, we extracted the topic term with highest relevant weight in the sentence to form word pairs so that it reduce the impact on the topic terms in common. 24% and 30% improvement were achieved, respectively. As to topic 11, i.e. ‘指环王’ (Lord of King), there were only 8 relevant documents without any opinion and 14 documents with relevant opinions. As a result, the graph constructed by insufficient documents worked ineffectively. Except for the above queries, GORM performed well in most of the others. To further investigate the effect of word pair, we summarized the top-5 word pairs with highest weight of 5 queries in Table 2. 0.2 0.25 0.3 0.35 0.4 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 MAP λ COAE08 ‐0.4 ‐0.3 ‐0.2 ‐0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 1 2 3 4 5 6 7 8 9 1011121314151617181920 Difference Topic Difference from Median Average Precision per Topic 1372 Table 2: Top-5 highest weight word pairs for 5 queries in COAE08 dataset. Table 2 showed that most word pairs could represent the relevant opinions about the corresponding queries. This showed that inter-sentence information was very helpful to identify the associative degree of a word pair. Furthermore, since word pairs can indicate relevant opinions effectively, it is worth further study on how they could be applied to other opinion oriented applications, e.g. opinion summarization, opinion prediction, etc. 5 Related Work Our research focuses on relevant opinion rather than on relevant document retrieval. We, therefore, review related works in opinion identification research. Furthermore, we do not support the conventional 2-stage opinion retrieval approach. We conducted literature review on unified opinion retrieval models and related work in this area is presented in the section. 5.1 Lexicon-based Opinion Identification Different from traditional IR, opinion retrieval focuses on the opinion nature of documents. During the last three years, NTICR and TREC evaluations have shown that sentiment lexicon-based methods led to good performance in opinion identification. A lightweight lexicon-based statistical approach was proposed by Hannah et al. (2007). In this method, the distribution of terms in relevant opinionated documents was compared to their distribution in relevant fact-based documents to calculate an opinion weight. These weights were used to compute opinion scores for each retrieved document. A weighted dictionary was generated from previous TREC relevance data (Amati et al., 2007). This dictionary was submitted as a query to a search engine to get an initial query-independent opinion score of all retrieved documents. Similarly, a pseudo opinionated word composed of all opinion words was first created, and then used to estimate the opinion score of a document (Na et al., 2009). This method was shown to be very effective in TREC evaluations (Lee et al., 2008). More recently, Huang and Croft (2009) proposed an effective relevance model, which integrated both query-independent and query-dependent sentiment words into a mixture model. In our approach, we also adopt sentiment lexicon-based method for opinion identification. Unlike the above methods, we generate a weight to a sentiment word for each target (associated topic term) rather than assign a unified weight or an equal weight to the sentiment word for the whole topics. Besides, in our model no training data is required. We just utilize the structure of our graph to generate a weight to reflect the associative degree between the two elements of a word pair in different context. 5.2 Unified Opinion Retrieval Model In addition to conventional 2-stage approach, there has been some research on unified opinion retrieval models. Eguchi and Lavrenko proposed an opinion retrieval model in the framework of generative language modeling (Eguchi and Lavrenko, 2006). They modeled a collection of natural language documents or statements, each of which consisted of some topic-bearing and some sentiment-bearing words. The sentiment was either represented by a group of predefined seed words, or extracted from a training sentiment corpus. This model was shown to be effective on the MPQA corpus. Mei et al. tried to build a fine-grained opinion retrieval system for consumer products (Mei et al., 2007). The opinion score for a product was a mixture of several facets. Due to the difficulty in Top-5 MAP 陈凯歌 Chen Kaige 国六条 Six States 宏观调控 Macro-regulation 周星驰 Stephen Chow Vista Vista <陈凯歌 支持> Chen Kaige Support <陈凯歌 最佳> Chen Kaige Best <《无极》 骂> Limitless Revile <影片 优秀> Movie Excellent <阵容 强大的> Cast Strong <房价 上涨> Room rate Rise <调控 加强> Regulate Strengthen <中央 加强> CCP Strengthen <房价 平稳> Room rate Steady <住房 保障> Housing Security <经济 平稳> Economics Steady <价格 上涨> Price Rise <发展 平稳> Development Steady <消费 上涨> Consume Rise <社会 保障> Social Security <电影 喜欢> Movie Like <周星驰 喜欢> Stephen Chow Like <主角 最佳> Protagonist Best <喜剧 好> Comedy Good <作品 精彩> Works Splendid <价格 贵> Price Expensive <微软 喜欢> Microsoft Like <Vista 推荐> Vista Recommend <问题 重要> Problem Vital <性能 不> Performance No 1373 associating sentiment with products and facets, the system was only tested using small scale text collections. Zhang and Ye proposed a generative model to unify topic relevance and opinion generation (Zhang and Ye, 2008). This model led to satisfactory performance, but an intensive computation load was inevitable during retrieval, since for each possible candidate document, an opinion score was summed up from the generative probability of thousands of sentiment words. Huang and Croft proposed a unified opinion retrieval model according to the Kullback-Leib- ler divergence between the two probability distributions of opinion relevance model and document model (Huang and Croft, 2009). They divided the sentiment words into query-dependent and query-independent by utilizing several sentiment expansion techniques, and integrated them into a mixed model. However, in this model, the contribution of a sentiment word was its corresponding incremental mean average precision value. This method required that large amount of training data and manual labeling. Different from the above opinion retrieval approaches, our proposed graph-based model processes opinion retrieval in the granularity of sentence. Instead of bag-of-word, the sentence is split into word pairs which can maintain the contextual information. On the one hand, word pair can identify the relevant opinion according to intra-sentence contextual information. On the other hand, it can measure the degree of a relevant opinion by considering the inter-sentence contextual information. 6 Conclusion and Future Work In this work we focus on the problem of opinion retrieval. Different from existing approaches, which regard document relevance as the key indicator of opinion relevance, we propose to explore the relevance of individual opinion. To do that, opinion retrieval is performed in the granularity of sentence. We define the notion of word pair, which can not only maintain the association between the opinion and the corresponding target in the sentence, but it can also build up the relationship among sentences through the same word pair. Furthermore, we convert the relationships between word pairs and sentences into a unified graph, and use the HITS algorithm to achieve document ranking for opinion retrieval. Finally, we compare our approach with existing methods. Experimental results show that our proposed model performs well on COAE08 dataset. The novelty of our work lies in using word pairs to represent the information needs for opinion retrieval. On the one hand, word pairs can identify the relevant opinion according to intra-sentence contextual information. On the other hand, word pairs can measure the degree of a relevant opinion by taking inter-sentence contextual information into consideration. With the help of word pairs, the information needs for opinion retrieval can be represented appropriately. In the future, more research is required in the following directions: (1) Since word pairs can indicate relevant opinions effectively, it is worth further study on how they could be applied to other opinion oriented applications, e.g. opinion summarization, opinion prediction, etc. (2) The characteristics of blogs will be taken into consideration, i.e., the post time, which could be helpful to create a more time sensitivity graph to filter out fake opinions. (3) Opinion holder is another important role of an opinion, and the identification of opinion holder is a main task in NTCIR. It would be interesting to study opinion holders, e.g. its seniority, for opinion retrieval. Acknowledgements: This work is partially supported by the Innovation and Technology Fund of Hong Kong SAR (No. ITS/182/08) and National 863 program (No. 2009AA01Z150). Special thanks to Xu Hongbo for providing the Chinese sentiment resources. We also thank Bo Chen, Wei Gao, Xu Han and anonymous reviewers for their helpful comments. References James Allan, Courtney Wade, and Alvaro Bolivar. 2003. Retrieval and novelty detection at the sentence level. In SIGIR ’03: Proceedings of the 26th annual international ACM SIGIR conference on Research and development in information retrieval, pages 314-321. ACM. Giambattista Amati, Edgardo Ambrosi, Marco Bianchi, Carlo Gaibisso, and Giorgio Gambosi. 2007. FUB, IASI-CNR and University of Tor Vergata at TREC 2007 Blog Track. In Proceedings of the 15th Text Retrieval Conference. Koji Eguchi and Victor Lavrenko. Sentiment retrieval using generative models. 2006. In EMNLP ’06, Proceedings of 2006 Conference on Empirical Methods in Natural Language Processing, page 345-354. 1374 Gunes Erkan and Dragomir R. Radev. 2004. Lexpagerank: Prestige in multi-document text summarization. In EMNLP ’04, Proceedings of 2004 Conference on Empirical Methods in Natural Language Processing. David Hannah, Craig Macdonald, Jie Peng, Ben He, and Iadh Ounis. 2007. University of Glasgow at TREC 2007: Experiments in Blog and Enterprise Tracks with Terrier. In Proceedings of the 15th Text Retrieval Conference. Xuanjing Huang, William Bruce Croft. 2009. A Unified Relevance Model for Opinion Retrieval. In Proceedings of CIKM. Jon M. Kleinberg. 1999. Authoritative sources in a hyperlinked environment. J. ACM, 46(5): 604-632. Yeha Lee, Seung-Hoon Na, Jungi Kim, Sang-Hyob Nam, Hun-young Jung, Jong-Hyeok Lee. 2008. KLE at TREC 2008 Blog Track: Blog Post and Feed Retrieval. In Proceedings of the 15th Text Retrieval Conference. Fangtao Li, Yang Tang, Minlie Huang, and Xiaoyan Zhu. 2009. Answering Opinion Questions with Random Walks on Graphs. In ACL ’09, Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics. Bing Liu, Minqing Hu, and Junsheng Cheng. 2005. Opinion observer: Analyzing and comparing opinion s on the web. In WWW ’05: Proceedings of the 14th International Conference on World Wide Web. Craig Macdonald and Iadh Ounis. 2007. Overview of the TREC-2007 Blog Track. In Proceedings of the 15th Text Retrieval Conference. Craig Macdonald and Iadh Ounis. 2006. Overview of the TREC-2006 Blog Track. In Proceedings of the 14th Text Retrieval Conference. Qiaozhu Mei, Xu Ling, Matthew Wondra, Hang Su, and Chengxiang Zhai. 2007. Topic sentiment mixture: Modeling facets and opinions in weblogs. In WWW ’07: Proceedings of the 16 International Conference on World Wide Web. Seung-Hoon Na, Yeha Lee, Sang-Hyob Nam, and Jong-Hyeok Lee. 2009. Improving opinion retrieval based on query-specific sentiment lexicon. In ECIR ’09: Proceedings of the 31st annual European Conference on Information Retrieval, pages 734-738. Douglas Oard, Tamer Elsayed, Jianqiang Wang, Yejun Wu, Pengyi Zhang, Eileen Abels, Jimmy Lin, and Dagbert Soergel. 2006. TREC-2006 at Maryland: Blog, Enterprise, Legal and QA Tracks. In Proceedings of the 15th Text Retrieval Conference. Jahna Otterbacher, Gunes Erkan, and Dragomir R. Radev. 2005. Using random walks for question-focused sentence retrieval. In EMNLP ’05, Proceedings of 2005 Conference on Empirical Methods in Natural Language Processing. Larry Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. 1998. The pagerank citation ranking: Bringing order to the web. Technical report, Stanford University. Bo Pang and Lillian Lee. 2008. Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval, 2(1-2): 1-135. Ana-Maria Popescu and Oren Etzioni. 2005. Extracting product features and opinion s from reviews. In EMNLP ’05, Proceedings of 2005 Conference on Empirical Methods in Natural Language Processing. Xiaojun Wan and Jianwu Yang. 2008. Multi-document summarization using cluster-based link analysis. In SIGIR ’08: Proceedings of the 31th annual international ACM SIGIR conference on Research and development in information retrieval, pages 299-306. ACM. Theresa Wilson, Janyce Wiebe, and Paul Hoffmann. 2005. Recognizing contextual polarity in phrase-level sentiment analysis. In EMNLP ’05, Proceedings of 2005 Conference on Empirical Methods in Natural Language Processing. Ruifeng Xu, Kam-Fai Wong and Yunqing Xia. 2007. Opinmine - Opinion Analysis System by CUHK for NTCIR-6 Pilot Task. In Proceedings of NTCIR-6. Min Zhang and Xingyao Ye. 2008. A generation model to unify topic relevance and lexicon-based sentiment for opinion retrieval. In SIGIR ’08: Proceedings of the 31st Annual International ACM SIGIR conference on Research and Development in Information Retrieval, pages 411-418. ACM. Wei Zhang and Clement Yu. 2007. UIC at TREC 2007 Blog Track. In Proceedings of the 15th Text Retrieval Conference. Jun Zhao, Hongbo Xu, Xuanjing Huang, Songbo Tan, Kang Liu, and Qi Zhang. 2008. Overview of Chinese Opinion Analysis Evaluation 2008. In Proceedings of the First Chinese Opinion Analysis Evaluation. 1375
|
2010
|
139
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 128–137, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics SystemT: An Algebraic Approach to Declarative Information Extraction Laura Chiticariu Rajasekar Krishnamurthy Yunyao Li Sriram Raghavan Frederick R. Reiss Shivakumar Vaithyanathan IBM Research – Almaden San Jose, CA, USA {chiti,sekar,yunyaoli,rsriram,frreiss,vaithyan}@us.ibm.com Abstract As information extraction (IE) becomes more central to enterprise applications, rule-based IE engines have become increasingly important. In this paper, we describe SystemT, a rule-based IE system whose basic design removes the expressivity and performance limitations of current systems based on cascading grammars. SystemT uses a declarative rule language, AQL, and an optimizer that generates high-performance algebraic execution plans for AQL rules. We compare SystemT’s approach against cascading grammars, both theoretically and with a thorough experimental evaluation. Our results show that SystemT can deliver result quality comparable to the state-of-theart and an order of magnitude higher annotation throughput. 1 Introduction In recent years, enterprises have seen the emergence of important text analytics applications like compliance and data redaction. This increase, combined with the inclusion of text into traditional applications like Business Intelligence, has dramatically increased the use of information extraction (IE) within the enterprise. While the traditional requirement of extraction quality remains critical, enterprise applications also demand efficiency, transparency, customizability and maintainability. In recent years, these systemic requirements have led to renewed interest in rule-based IE systems (Doan et al., 2008; SAP, 2010; IBM, 2010; SAS, 2010). Until recently, rule-based IE systems (Cunningham et al., 2000; Boguraev, 2003; Drozdzynski et al., 2004) were predominantly based on the cascading grammar formalism exemplified by the Common Pattern Specification Language (CPSL) specification (Appelt and Onyshkevych, 1998). In CPSL, the input text is viewed as a sequence of annotations, and extraction rules are written as pattern/action rules over the lexical features of these annotations. In a single phase of the grammar, a set of rules are evaluated in a left-to-right fashion over the input annotations. Multiple grammar phases are cascaded together, with the evaluation proceeding in a bottom-up fashion. As demonstrated by prior work (Grishman and Sundheim, 1996), grammar-based IE systems can be effective in many scenarios. However, these systems suffer from two severe drawbacks. First, the expressivity of CPSL falls short when used for complex IE tasks over increasingly pervasive informal text (emails, blogs, discussion forums etc.). To address this limitation, grammar-based IE systems resort to significant amounts of userdefined code in the rules, combined with preand post-processing stages beyond the scope of CPSL (Cunningham et al., 2010). Second, the rigid evaluation order imposed in these systems has significant performance implications. Three decades ago, the database community faced similar expressivity and efficiency challenges in accessing structured information. The community addressed these problems by introducing a relational algebra formalism and an associated declarative query language SQL. The groundbreaking work on System R (Chamberlin et al., 1981) demonstrated how the expressivity of SQL can be efficiently realized in practice by means of a query optimizer that translates an SQL query into an optimized query execution plan. Borrowing ideas from the database community, we have developed SystemT, a declarative IE system based on an algebraic framework, to address both expressivity and performance issues. In SystemT, extraction rules are expressed in a declarative language called AQL. At compilation time, 128 ({First} {Last} ) :full :full.Person ({Caps} {Last} ) :full :full.Person ({Last} {Token.orth = comma} {Caps | First}) : reverse :reverse.Person ({First}) : fn :fn.Person ({Last}) : ln :ln.Person ({Lookup.majorType = FirstGaz}) : fn :fn.First ({Lookup.majorType = LastGaz}) : ln :ln.Last ({Token.orth = upperInitial} | {Token.orth = mixedCaps } ) :cw :cw.Caps Rule Patterns 50 20 10 10 10 50 50 10 Priority P2R1 P2R2 P2R3 P2R4 P2R5 P1R1 P1R2 P1R3 RuleId Input First Last Caps Token Output Person Input Lookup Token Output First Last Caps Types Phase P2 P1 P2R3 ({Last} {Token.orth = comma} {Caps | First}) : reverse :reverse.Person Last followed by Token whose orth attribute has value comma followed by Caps or First Rule part Action part Create Person annotation Bind match to variables Syntax: Gazetteers containing first names and last names Figure 1: Cascading grammar for identifying Person names SystemT translates AQL statements into an algebraic expression called an operator graph that implements the semantics of the statements. The SystemT optimizer then picks a fast execution plan from many logically equivalent plans. SystemT is currently deployed in a multitude of realworld applications and commercial products1. We formally demonstrate the superiority of AQL and SystemT in terms of both expressivity and efficiency (Section 4). Specifically, we show that 1) the expressivity of AQL is a strict superset of CPSL grammars not using external functions and 2) the search space explored by the SystemT optimizer includes operator graphs corresponding to efficient finite state transducer implementations. Finally, we present an extensive experimental evaluation that validates that high-quality annotators can be developed with SystemT, and that their runtime performance is an order of magnitude better when compared to annotators developed with a state-of-the-art grammar-based IE system (Section 5). 2 Grammar-based Systems and CPSL A cascading grammar consists of a sequence of phases, each of which consists of one or more rules. Each phase applies its rules from left to right over an input sequence of annotations and generates an output sequence of annotations that the next phase consumes. Most cascading grammar systems today adhere to the CPSL standard. Fig. 1 shows a sample CPSL grammar that identifies person names from text in two phases. The first phase, P1, operates over the results of the tok1A trial version is available at http://www.alphaworks.ibm.com/tech/systemt Rule skipped due to priority semantics CPSL Phase P1 Last(P1R2) Last(P1R2) … Mark Scott , Howard Smith … First(P1R1) First(P1R1) First(P1R1) Last(P1R2) CPSL Phase P2 … Mark Scott , Howard Smith … Person(P2R1) Person (P2R4) Person(P2R4) Person (P2R5) Person(P2R4) … Mark Scott , Howard Smith … First(P1R1) First(P1R1) First(P1R1) Last(P1R2) JAPE Phase P1 (Brill) Caps(P1R3) Last(P1R2) Last(P1R2) Caps(P1R3) Caps(P1R3) Caps(P1R3) … Mark Scott , Howard Smith … Person(P2R1) Person (P2R4, P2R5) JAPE Phase P2 (Appelt) Person(P2R1) Person (P2R2) Some discarded matches omitted for clarity … Tomorrow, we will meet Mark Scott, Howard Smith and … Document d1 Rule fired Legend 3 persons identified 2 persons identified (a) (b) Figure 2: Sample output of CPSL and JAPE enizer and gazetteer (input types Token and Lookup, respectively) to identify words that may be part of a person name. The second phase, P2, identifies complete names using the results of phase P1. Applying the above grammar to document d1 (Fig. 2), one would expect that to match “Mark Scott” and “Howard Smith” as Person. However, as shown in Fig. 2(a), the grammar actually finds three Person annotations, instead of two. CPSL has several limitations that lead to such discrepancies: L1. Lossy sequencing. In a CPSL grammar, each phase operates on a sequence of annotations from left to right. If the input annotations to a phase may overlap with each other, the CPSL engine must drop some of them to create a nonoverlapping sequence. For instance, in phase P1 (Fig. 2(a)), “Scott” has both a Lookup and a Token annotation. The system has made an arbitrary choice to retain the Lookup annotation and discard the Token annotation. Consequently, no Caps annotations are output by phase P1. L2. Rigid matching priority. CPSL specifies that, for each input annotation, only one rule can actually match. When multiple rules match at the same start position, the following tie-breaker conditions are applied (in order): (a) the rule matching the most annotations in the input stream; (b) the rule with highest priority; and (c) the rule declared earlier in the grammar. This rigid matching priority can lead to mistakes. For instance, as illustrated in Fig. 2(a), phase P1 only identifies “Scott” as a First. Matching priority causes the grammar to skip the corresponding match for “Scott” as a Last. Consequently, phase P2 fails to identify “Mark Scott” as one single Person. L3. Limited expressivity in rule patterns. It is not possible to express rules that compare annotations overlapping with each other. E.g., “Identify 129 [A-Z]{\w|-}+ Document Input Tuple … we will meet Mark Scott, … Output Tuple 2 Span 2 Document Span 1 Output Tuple 1 Document Regex Caps Figure 3: Regular Expression Extraction Operator words that are both capitalized and present in the FirstGaz gazetteer” or “Identify Person annotations that occur within an EmailAddress”. Extensions to CPSL In order to address the above limitations, several extensions to CPSL have been proposed in JAPE, AFst and XTDL (Cunningham et al., 2000; Boguraev, 2003; Drozdzynski et al., 2004). The extensions are summarized as below, where each solution Si corresponds to limitation Li. • S1. Grammar rules are allowed to operate on graphs of input annotations in JAPE and AFst. • S2. JAPE introduces more matching regimes besides the CPSL’s matching priority and thus allows more flexibility when multiple rules match at the same starting position. • S3. The rule part of a pattern has been expanded to allow more expressivity in JAPE, AFst and XTDL. Fig. 2(b) illustrates how the above extensions help in identifying the correct matches ‘Mark Scott’ and ‘Howard Smith’ in JAPE. Phase P1 uses a matching regime (denoted by Brill) that allows multiple rules to match at the same starting position, and phase P2 uses CPSL’s matching priority, Appelt. 3 SystemT SystemT is a declarative IE system based on an algebraic framework. In SystemT, developers write rules in a language called AQL. The system then generates a graph of operators that implement the semantics of the AQL rules. This decoupling allows for greater rule expressivity, because the rule language is not constrained by the need to compile to a finite state transducer. Likewise, the decoupled approach leads to greater flexibility in choosing an efficient execution strategy, because many possible operator graphs may exist for the same AQL annotator. In the rest of the section, we describe the parts of SystemT, starting with the algebraic formalism behind SystemT’s operators. 3.1 Algebraic Foundation of SystemT SystemT executes IE rules using graphs of operators. The formal definition of these operators takes the form of an algebra that is similar to the relational algebra, but with extensions for text processing. The algebra operates over a simple relational data model with three data types: span, tuple, and relation. In this data model, a span is a region of text within a document identified by its “begin” and “end” positions; a tuple is a fixed-size list of spans. A relation is a multiset of tuples, where every tuple in the relation must be of the same size. Each operator in our algebra implements a single basic atomic IE operation, producing and consuming sets of tuples. Fig. 3 illustrates the regular expression extraction operator in the algebra, which performs character-level regular expression matching. Overall, the algebra contains 12 different operators, a full description of which can be found in (Reiss et al., 2008). The following four operators are necessary to understand the examples in this paper: • The Extract operator (E) performs characterlevel operations such as regular expression and dictionary matching over text, creating a tuple for each match. • The Select operator (σ) takes as input a set of tuples and a predicate to apply to the tuples. It outputs all tuples that satisfy the predicate. • The Join operator (⊲⊳) takes as input two sets of tuples and a predicate to apply to pairs of tuples from the input sets. It outputs all pairs of input tuples that satisfy the predicate. • The consolidate operator (Ω) takes as input a set of tuples and the index of a particular column in those tuples. It removes selected overlapping spans from the indicated column, according to the specified policy. 3.2 AQL Extraction rules in SystemT are written in AQL, a declarative relational language similar in syntax to the database language SQL. We chose SQL as a basis for our language due to its expressivity and its familiarity. The expressivity of SQL, which consists of first-order logic predicates 130 Figure 4: Person annotator as AQL query over sets of tuples, is well-documented and wellunderstood (Codd, 1990). As SQL is the primary interface to most relational database systems, the language’s syntax and semantics are common knowledge among enterprise application programmers. Similar to SQL terminology, we call a collection of AQL rules an AQL query. Fig. 4 shows portions of an AQL query. As can be seen, the basic building block of AQL is a view: A logical description of a set of tuples in terms of either the document text (denoted by a special view called Document) or the contents of other views. Every SystemT annotator consists of at least one view. The output view statement indicates that the tuples in a view are part of the final results of the annotator. Fig. 4 also illustrates three of the basic constructs that can be used to define a view. • The extract statement specifies basic character-level extraction primitives to be applied directly to a tuple. • The select statement is similar to the SQL select statement but it contains an additional consolidate on clause, along with an extensive collection of text-specific predicates. • The union all statement merges the outputs of one or more select or extract statements. To keep rules compact, AQL also provides a shorthand sequence pattern notation similar to the syntax of CPSL. For example, the CapsLast view in Figure 4 could have been written as: create view CapsLast as extract pattern <C.name> <L.name> from Caps C, Last L; Internally, SystemT translates each of these extract pattern statements into one or more select and extract statements. AQL SystemT Optimizer SystemT Runtime Compiled Operator Graph Figure 5: The compilation process in SystemT Figure 6: Execution strategies for the CapsLast rule in Fig. 4 SystemT has built-in multilingual support including tokenization, part of speech and gazetteer matching for over 20 languages using LanguageWare (IBM, 2010). Rule developers can utilize the multilingual support via AQL without having to configure or manage any additional resources. In addition, AQL allows user-defined functions to be used in a restricted context in order to support operations such as validation (e.g. for extracted credit card numbers), or normalization (e.g., compute abbreviations of multi-token organization candidates that are useful in generating additional candidates). More details on AQL can be found in the AQL manual (SystemT, 2010). 3.3 Optimizer and Operator Graph Grammar-based IE engines place rigid restrictions on the order in which rules can be executed. Due to the semantics of the CPSL standard, systems that implement the standard must use a finite state transducer that evaluates each level of the cascade with one or more left to right passes over the entire token stream. In contrast, SystemT places no explicit constraints on the order of rule evaluation, nor does it require that intermediate results of an annotator collapse to a fixed-size sequence. As shown in Fig. 5, the SystemT engine does not execute AQL directly; instead, the SystemT optimizer compiles AQL into a graph of operators. By tying a collection of operators together by their inputs and outputs, the system can implement a wide variety of different execution strategies. Different execution strategies are associated with different evaluation costs. The optimizer chooses the execution strategy with the lowest estimated evaluation cost. 131 Fig. 6 presents three possible execution strategies for the CapsLast rule in Fig. 4. If the optimizer estimates that the evaluation cost of Last is much lower than that of Caps, then it can determine that Plan C has the lowest evaluation cost among the three, because Plan C only evaluates Caps in the “left” neighborhood for each instance of Last. More details of our algorithms for enumerating plans can be found in (Reiss et al., 2008). The optimizer in SystemT chooses the best execution plan from a large number of different algebra graphs available to it. Many of these graphs implement strategies that a transducer could not express: such as evaluating rules from right to left, sharing work across different rules, or selectively skipping rule evaluations. Within this large search space, there generally exists an execution strategy that implements the rule semantics far more efficiently than the fastest transducer could. We refer the reader to (Reiss et al., 2008) for a detailed description of the types of plan the optimizer considers, as well as an experimental analysis of the performance benefits of different parts of this search space. Several parallel efforts have been made recently to improve the efficiency of IE tasks by optimizing low-level feature extraction (Ramakrishnan et al., 2006; Ramakrishnan et al., 2008; Chandel et al., 2006) or by reordering operations at a macroscopic level (Ipeirotis et al., 2006; Shen et al., 2007; Jain et al., 2009). However, to the best of our knowledge, SystemT is the only IE system in which the optimizer generates a full end-to-end plan, beginning with low-level extraction primitives and ending with the final output tuples. 3.4 Deployment Scenarios SystemT is designed to be usable in various deployment scenarios. It can be used as a standalone system with its own development and runtime environment. Furthermore, SystemT exposes a generic Java API that enables the integration of its runtime environment with other applications. For example, a specific instantiation of this API allows SystemT annotators to be seamlessly embedded in applications using the UIMA analytics framework (UIMA, 2010). 4 Grammar vs. Algebra Having described both the traditional cascading grammar approach and the declarative approach Figure 7: Supporting Complex Rule Interactions used in SystemT, we now compare the two in terms of expressivity and performance. 4.1 Expressivity In Section 2, we described three expressivity limitations of CPSL grammars: Lossy sequencing, rigid matching priority, and limited expressivity in rule patterns. As we noted, cascading grammar systems extend the CPSL specification in various ways to provide workarounds for these limitations. In SystemT, the basic design of the AQL language eliminates these three problems without the need for any special workaround. The key design difference is that AQL views operate over sets of tuples, not sequences of tokens. The input or output tuples of a view can contain spans that overlap in arbitrary ways, so the lossy sequencing problem never occurs. The annotator will retain these overlapping spans across any number of views until a view definition explicitly removes the overlap. Likewise, the tuples that a given view produces are in no way constrained by the outputs of other, unrelated views, so the rigid matching priority problem never occurs. Finally, the select statement in AQL allows arbitrary predicates over the cross-product of its input tuple sets, eliminating the limited expressivity in rule patterns problem. Beyond eliminating the major limitations of CPSL grammars, AQL provides a number of other information extraction operations that even extended CPSL cannot express without custom code. Complex rule interactions. Consider an example document from the Enron corpus (Minkov et al., 2005), shown in Fig. 7, which contains a list of person names. Because the first person in the list (‘Skilling’) is referred to by only a last name, rule P2R3 in Fig. 1 incorrectly identifies ‘Skilling, Cindy’ as a person. Consequently, the output of phase P2 of the cascading grammar contains several mistakes as shown in the figure. This problem 132 went to the Switchfoot concert at the Roxy. It was pretty fun,… The lead singer/guitarist was really good, and even though there was another guitarist (an Asian guy), he ended up playing most of the guitar parts, which was really impressive. The biggest surprise though is that I actually liked the opening bands. …I especially liked the first band Consecutive review snippets are within 25 tokens At least 4 occurrences of MusicReviewSnippet or GenericReviewSnippet At least 3 of them should be MusicReviewSnippets Review ends with one of these. Start with ConcertMention Complete review is within 200 tokens ConcertMention MusicReviewSnippet GenericReviewSnippet Example Rule Informal Band Review Figure 8: Extracting informal band reviews from web logs occurs because CPSL only evaluates rules over the input sequence in a strict left-to-right fashion. On the other hand, the AQL query Q1 shown in the figure applies the following condition: “Always discard matches to Rule P2R3 if they overlap with matches to rules P2R1 or P2R2” (even if the match to Rule P2R3 starts earlier). Applying this rule ensures that the person names in the list are identified correctly. Obtaining the same effect in grammar-based systems would require the use of custom code (as recommended by (Cunningham et al., 2010)). Counting and Aggregation. Complex extraction tasks sometimes require operations such as counting and aggregation that go beyond the expressivity of regular languages, and thus can be expressed in CPSL only using external functions. One such task is that of identifying informal concert reviews embedded within blog entries. Fig. 8 describes, by example, how these reviews consist of reference to a live concert followed by several review snippets, some specific to musical performances and others that are more general review expressions. An example rule to identify informal reviews is also shown in the figure. Notice how implementing this rule requires counting the number of MusicReviewSnippet and GenericReviewSnippet annotations within a region of text and aggregating this occurrence count across the two review types. While this rule can be written in AQL, it can only be approximated in CPSL grammars. Character-Level Regular Expression CPSL cannot specify character-level regular expressions that span multiple tokens. In contrast, the extract regex statement in AQL fully supports these expressions. We have described above several cases where AQL can express concepts that can only be expressed through external functions in a cascading grammar. These examples naturally raise the question of whether similar cases exist where a cascading grammar can express patterns that cannot be expressed in AQL. It turns out that we can make a strong statement that such examples do not exist. In the absence of an escape to arbitrary procedural code, AQL is strictly more expressive than a CPSL grammar. To state this relationship formally, we first introduce the following definitions. We refer to a grammar conforming to the CPSL specification as a CPSL grammar. When a CPSL grammar contains no external functions, we refer to it as a Code-free CPSL grammar. Finally, we refer to a grammar that conforms to one of the CPSL, JAPE, AFst and XTDL specifications as an expanded CPSL grammar. Ambiguous Grammar Specification An expanded CPSL grammar may be under-specified in some cases. For example, a single rule containing the disjunction operator (|) may match a given region of text in multiple ways. Consider the evaluation of Rule P2R3 over the text fragment “Scott, Howard” from document d1 (Fig. 1). If “Howard” is identified both as Caps and First, then there are two evaluations for Rule P2R3 over this text fragment. Since the system has to arbitrarily choose one evaluation, the results of the grammar can be non-deterministic (as pointed out in (Cunningham et al., 2010)). We refer to a grammar G as an ambiguous grammar specification for a document collection D if the system makes an arbitrary choice while evaluating G over D. Definition 1 (UnambigEquiv) A query Q is UnambigEquiv to a cascading grammar G if and only if for every document collection D, where G is not an ambiguous grammar specification for D, the results of the grammar invocation and the query evaluation are identical. We now formally compare the expressivity of AQL and expanded CPSL grammars. The detailed proof is omitted due to space limitations. Theorem 1 The class of extraction tasks expressible as AQL queries is a strict superset of that expressible through expanded code-free CPSL grammars. Specifically, (a) Every expanded code-free CPSL grammar can be expressed as an UnambigEquiv AQL query. (b) AQL supports information extraction operations that cannot be expressed in expanded codefree CPSL grammars. 133 Proof Outline: (a) A single CPSL grammar can be expressed in AQL as follows. First, each rule r in the grammar is translated into a set of AQL statements. If r does not contain the disjunct (|) operator, then it is translated into a single AQL select statement. Otherwise, a set of AQL statements are generated, one for each disjunct operator in rule r, and the results merged using union all statements. Then, a union all statement is used to combine the results of individual rules in the grammar phase. Finally, the AQL statements for multiple phases are combined in the same order as the cascading grammar specification. The main extensions to CPSL supported by expanded CPSL grammars (listed in Sec. 2) are handled as follows. AQL queries operate on graphs on annotations just like expanded CPSL grammars. In addition, AQL supports different matching regimes through consolidation operators, span predicates through selection predicates and coreferences through join operators. (b) Example operations supported in AQL that cannot be expressed in expanded code-free CPSL grammars include (i) character-level regular expressions spanning multiple tokens, (ii) counting the number of annotations occurring within a given bounded window and (iii) deleting annotations if they overlap with other annotations starting later in the document. 2 4.2 Performance For the annotators we test in our experiments (See Section 5), the SystemT optimizer is able to choose algebraic plans that are faster than a comparable transducer-based implementation. The question arises as to whether there are other annotators for which the traditional transducer approach is superior. That is, for a given annotator, might there exist a finite state transducer that is combinatorially faster than any possible algebra graph? It turns out that this scenario is not possible, as the theorem below shows. Definition 2 (Token-Based FST) A token-based finite state transducer (FST) is a nondeterministic finite state machine in which state transitions are triggered by predicates on tokens. A token-based FST is acyclic if its state graph does not contain any cycles and has exactly one “accept” state. Definition 3 (Thompson’s Algorithm) Thompson’s algorithm is a common strategy for evaluating a token-based FST (based on (Thompson, 1968)). This algorithm processes the input tokens from left to right, keeping track of the set of states that are currently active. Theorem 2 For any acyclic token-based finite state transducer T, there exists an UnambigEquiv operator graph G, such that evaluating G has the same computational complexity as evaluating T with Thompson’s algorithm starting from each token position in the input document. Proof Outline: The proof constructs G by structural induction over the transducer T. The base case converts transitions out of the start state into Extract operators. The inductive case adds a Select operator to G for each of the remaining state transitions, with each selection predicate being the same as the predicate that drives the corresponding state transition. For each state transition predicate that T would evaluate when processing a given document, G performs a constant amount of work on a single tuple. 2 5 Experimental Evaluation In this section we present an extensive comparison study between SystemT and implementations of expanded CPSL grammar in terms of quality, runtime performance and resource requirements. Tasks We chose two tasks for our evaluation: • NER : named-entity recognition for Person, Organization, Location, Address, PhoneNumber, EmailAddress, URL and DateTime. • BandReview : identify informal reviews in blogs (Fig. 8). We chose NER primarily because named-entity recognition is a well-studied problem and standard datasets are available for evaluation. For this task we use GATE and ANNIE for comparison3. We chose BandReview to conduct performance evaluation for a more complex extraction task. Datasets. For quality evaluation, we use: • EnronMeetings (Minkov et al., 2005): collection of emails with meeting information from the Enron corpus4 with Person labeled data; • ACE (NIST, 2005): collection of newswire reports and broadcast news/conversations with Person, Organization, Location labeled data5. 3To the best of our knowledge, ANNIE (Cunningham et al., 2002) is the only publicly available NER library implemented in a grammar-based system (JAPE in GATE). 4http://www.cs.cmu.edu/ enron/ 5Only entities of type NAM have been considered. 134 Table 1: Datasets for performance evaluation. Dataset Description of the Content Number of Document size documents range average Enronx Emails randomly sampled from the Enron corpus of average size xKB (0.5 < x < 100)2 1000 xKB +/ −10% xKB WebCrawl Small to medium size web pages representing company news, with HTML tags removed 1931 68b - 388.6KB 8.8KB FinanceM Medium size financial regulatory filings 100 240KB - 0.9MB 401KB FinanceL Large size financial regulatory filings 30 1MB - 3.4MB 1.54MB Table 2: Quality of Person on test datasets. Precision (%) Recall (%) F1 measure (%) (Exact/Partial) (Exact/Partial) (Exact/Partial) EnronMeetings ANNIE 57.05/76.84 48.59/65.46 52.48/70.69 T-NE 88.41/92.99 82.39/86.65 85.29/89.71 Minkov 81.1/NA 74.9/NA 77.9/NA ACE ANNIE 39.41/78.15 30.39/60.27 34.32/68.06 T-NE 93.90/95.82 90.90/92.76 92.38/94.27 Table 1 lists the datasets used for performance evaluation. The size of FinanceLis purposely small because GATE takes a significant amount of time processing large documents (see Sec. 5.2). Set Up. The experiments were run on a server with two 2.4 GHz 4-core Intel Xeon CPUs and 64GB of memory. We use GATE 5.1 (build 3431) and two configurations for ANNIE: 1) the default configuration, and 2) an optimized configuration where the Ontotext Japec Transducer6 replaces the default NE transducer for optimized performance. We refer to these configurations as ANNIE and ANNIE-Optimized, respectively. 5.1 Quality Evaluation The goal of our quality evaluation is two-fold: to validate that annotators can be built in SystemT with quality comparable to those built in a grammar-based system; and to ensure a fair performance comparison between SystemT and GATE by verifying that the annotators used in the study are comparable. Table 2 shows results of our comparison study for Person annotators. We report the classical (exact) precision, recall, and F1 measures that credit only exact matches, and corresponding partial measures that credit partial matches in a fashion similar to (NIST, 2005). As can be seen, TNE produced results of significantly higher quality than ANNIE on both datasets, for the same Person extraction task. In fact, on EnronMeetings, the F1 measure of T-NE is 7.4% higher than the best published result (Minkov et al., 2005). Similar results 6http://www.ontotext.com/gate/japec.html a) Throughput on Enron 0 100 200 300 400 500 600 700 0 20 40 60 80 100 Average document size (KB) Throughput (KB/sec) ANNIE ANNIE-Optimized T-NE x b) Memory Utilization on Enron 0 200 400 600 0 20 40 60 80 100 Average document size (KB) Avg Heap size (MB) ANNIE ANNIE-Optimized T-NE Error bars show 25th and 75th percentile x Figure 9: Throughput (a) and memory consumption (b) comparisons on Enronx datasets. can be observed for Organization and Location on ACE (exact numbers omitted in interest of space). Clearly, considering the large gap between ANNIE’s F1 and partial F1 measures on both datasets, ANNIE’s quality can be improved via dataset-specific tuning as demonstrated in (Maynard et al., 2003). However, dataset-specific tuning for ANNIE is beyond the scope of this paper. Based on the experimental results above and our previous formal comparison in Sec. 4, we believe it is reasonable to conclude that annotators can be built in SystemT of quality at least comparable to those built in a grammar-based system. 5.2 Performance Evaluation We now focus our attention on the throughput and memory behavior of SystemT, and draw a comparison with GATE. For this purpose, we have configured both ANNIE and T-NE to identify only the same eight types of entities listed for NER task. Throughput. Fig. 9(a) plots the throughput of the two systems on multiple Enronx datasets with average document sizes of between 0.5KB and 100KB. For this experiment, both systems ran with a maximum Java heap size of 1GB. 135 Table 3: Throughput and mean heap size. ANNIE ANNIE-Optimized T-NE Dataset ThroughputMemoryThroughput Memory ThroughputMemory (KB/s) (MB) (KB/s) (MB) (KB/s) (MB) WebCrawl 23.9 212.6 42.8 201.8 498.9 77.2 FinanceM 18.82 715.1 26.3 601.8 703.5 143.7 FinanceL 19.2 2586.2 21.1 2683.5 954.5 189.6 As shown in Fig. 9(a), even though the throughput of ANNIE-Optimized (using the optimized transducer) increases two-fold compared to ANNIE under default configuration, T-NE is between 8 and 24 times faster compared to ANNIE-Optimized. For both systems, throughput varied with document size. For T-NE, the relatively low throughput on very small document sizes (less than 1KB) is due to fixed overhead in setting up operators to process a document. As document size increases, the overhead becomes less noticeable. We have observed similar trends on the rest of the test collections. Table 3 shows that TNE is at least an order of magnitude faster than ANNIE-Optimized across all datasets. In particular, on FinanceL T-NE’s throughput remains high, whereas the performance of both ANNIE and ANNIE-Optimized degraded significantly. To ascertain whether the difference in performance in the two systems is due to low-level components such as dictionary evaluation, we performed detailed profiling of the systems. The profiling revealed that 8.2%, 16.2% and respectively 14.2% of the execution time was spent on average on low-level components in the case of ANNIE, ANNIE-Optimized and T-NE, respectively, thus leading us to conclude that the observed differences are due to SystemT’s efficient use of resources at a macroscopic level. Memory utilization. In theory, grammar based systems can stream tuples through each stage for minimal memory consumption, whereas SystemT operator graphs may need to materialize intermediate results for the full document at certain points to evaluate the constraints in the original AQL. The goal of this study is to evaluate whether this potential problem does occur in practice. In this experiment we ran both systems with a maximum heap size of 2GB, and used the Java garbage collector’s built-in telemetry to measure the total quantity of live objects in the heap over time while annotating the different test corpora. Fig. 9(b) plots the minimum, maximum, and mean heap sizes with the Enronx datasets. On small documents of size up to 15KB, memory consumption is dominated by the fixed size of the data structures used (e.g., dictionaries, FST/operator graph), and is comparable for both systems. As documents get larger, memory consumption increases for both systems. However, the increase is much smaller for T-NE compared to that for both ANNIE and ANNIE-Optimized. A similar trend can be observed on the other datasets as shown in Table 3. In particular, for FinanceL, both ANNIE and ANNIE-Optimized required 8GB of Java heap size to achieve reasonable throughput7, in contrast to TNE which utilized at most 300MB out of the 2GB of maximum Java heap size allocation. SystemT requires much less memory than GATE in general due to its runtime, which monitors data dependencies between operators and clears out low-level results when they are no longer needed. Although a streaming CPSL implementation is theoretically possible, in practice mechanisms that allow an escape to custom code make it difficult to decide when an intermediate result will no longer be used, hence GATE keeps most intermediate data in memory until it is done analyzing the current document. The BandReview Task. We conclude by briefly discussing our experience with the BandReview task from Fig. 8. We built two versions of this annotator, one in AQL, and the other using expanded CPSL grammar. The grammar implementation processed a 4.5GB collection of 1.05 million blogs in 5.6 hours and output 280 reviews. In contrast, the SystemT version (85 AQL statements) extracted 323 reviews in only 10 minutes! 6 Conclusion In this paper, we described SystemT, a declarative IE system based on an algebraic framework. We presented both formal and empirical arguments for the benefits of our approach to IE. Our extensive experimental results show that highquality annotators can be built using SystemT, with an order of magnitude throughput improvement compared to state-of-the-art grammar-based systems. Going forward, SystemT opens up several new areas of research, including implementing better optimization strategies and augmenting the algebra with additional operators to support advanced features such as coreference resolution. 7GATE ran out of memory when using less than 5GB of Java heap size, and thrashed when run with 5GB to 7GB 136 References Douglas E. Appelt and Boyan Onyshkevych. 1998. The common pattern specification language. In TIPSTER workshop. Branimir Boguraev. 2003. Annotation-based finite state processing in a large-scale nlp arhitecture. In RANLP, pages 61–80. D. D. Chamberlin, A. M. Gilbert, and Robert A. Yost. 1981. A history of System R and SQL/data system. In vldb. Amit Chandel, P. C. Nagesh, and Sunita Sarawagi. 2006. Efficient batch top-k search for dictionarybased entity recognition. In ICDE. E. F. Codd. 1990. The relational model for database management: version 2. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA. H. Cunningham, D. Maynard, and V. Tablan. 2000. JAPE: a Java Annotation Patterns Engine (Second Edition). Research Memorandum CS–00–10, Department of Computer Science, University of Sheffield, November. H. Cunningham, D. Maynard, K. Bontcheva, and V. Tablan. 2002. GATE: A framework and graphical development environment for robust NLP tools and applications. In Proceedings of the 40th Anniversary Meeting of the Association for Computational Linguistics, pages 168 – 175. Hamish Cunningham, Diana Maynard, Kalina Bontcheva, Valentin Tablan, Marin Dimitrov, Mike Dowman, Niraj Aswani, Ian Roberts, Yaoyong Li, and Adam Funk. 2010. Developing language processing components with gate version 5 (a user guide). AnHai Doan, Luis Gravano, Raghu Ramakrishnan, and Shivakumar Vaithyanathan. 2008. Special issue on managing information extraction. SIGMOD Record, 37(4). Witold Drozdzynski, Hans-Ulrich Krieger, Jakub Piskorski, Ulrich Sch¨afer, and Feiyu Xu. 2004. Shallow processing with unification and typed feature structures — foundations and applications. K¨unstliche Intelligenz, 1:17–23. Ralph Grishman and Beth Sundheim. 1996. Message understanding conference - 6: A brief history. In COLING, pages 466–471. IBM. 2010. IBM LanguageWare. P. G. Ipeirotis, E. Agichtein, P. Jain, and L. Gravano. 2006. To search or to crawl?: towards a query optimizer for text-centric tasks. In SIGMOD. Alpa Jain, Panagiotis G. Ipeirotis, AnHai Doan, and Luis Gravano. 2009. Join optimization of information extraction output: Quality matters! In ICDE. Diana Maynard, Kalina Bontcheva, and Hamish Cunningham. 2003. Towards a semantic extraction of named entities. In Recent Advances in Natural Language Processing. Einat Minkov, Richard C. Wang, and William W. Cohen. 2005. Extracting personal names from emails: Applying named entity recognition to informal text. In HLT/EMNLP. NIST. 2005. The ACE evaluation plan. Ganesh Ramakrishnan, Sreeram Balakrishnan, and Sachindra Joshi. 2006. Entity annotation based on inverse index operations. In EMNLP. Ganesh Ramakrishnan, Sachindra Joshi, Sanjeet Khaitan, and Sreeram Balakrishnan. 2008. Optimization issues in inverted index-based entity annotation. In InfoScale. Frederick Reiss, Sriram Raghavan, Rajasekar Krishnamurthy, Huaiyu Zhu, and Shivakumar Vaithyanathan. 2008. An algebraic approach to rule-based information extraction. In ICDE, pages 933–942. SAP. 2010. Inxight ThingFinder. SAS. 2010. Text Mining with SAS Text Miner. Warren Shen, AnHai Doan, Jeffrey F. Naughton, and Raghu Ramakrishnan. 2007. Declarative information extraction using datalog with embedded extraction predicates. In vldb. SystemT. 2010. AQL Manual. http://www.alphaworks.ibm.com/tech/systemt. Ken Thompson. 1968. Regular expression search algorithm. pages 419–422. UIMA. 2010. Unstructured Information Management Architecture. http://uima.apache.org. 137
|
2010
|
14
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1376–1385, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Generating Fine-Grained Reviews of Songs From Album Reviews Swati Tata and Barbara Di Eugenio Computer Science Department University of Illinois, Chicago, IL, USA {stata2 | bdieugen}@uic.edu Abstract Music Recommendation Systems often recommend individual songs, as opposed to entire albums. The challenge is to generate reviews for each song, since only full album reviews are available on-line. We developed a summarizer that combines information extraction and generation techniques to produce summaries of reviews of individual songs. We present an intrinsic evaluation of the extraction components, and of the informativeness of the summaries; and a user study of the impact of the song review summaries on users’ decision making processes. Users were able to make quicker and more informed decisions when presented with the summary as compared to the full album review. 1 Introduction In recent years, the personal music collection of many individuals has significantly grown due to the availability of portable devices like MP3 players and of internet services. Music listeners are now looking for techniques to help them manage their music collections and explore songs they may not even know they have (Clema, 2006). Currently, most of those electronic devices follow a Universal Plug and Play (UPNP) protocol (UPN, 2008), and can be used in a simple network, on which the songs listened to can be monitored. Our interest is in developing a Music Recommendation System (Music RS) for such a network. Commercial web-sites such as Amazon (www. amazon.com) and Barnes and Nobles (www. bnn.com) have deployed Product Recommendation Systems (Product RS) to help customers choose from large catalogues of products. Most Product RSs include reviews from customers who bought or tried the product. As the number of reviews available for each individual product increases, RSs may overwhelm the user if they make all those reviews available. Additionally, in some reviews only few sentences actually describe the recommended product, hence, the interest in opinion mining and in summarizing those reviews. A Music RS could be developed along the lines of Product RSs. However, Music RSs recommend individual tracks, not full albums, e.g. see www.itunes.com. Summarizing reviews becomes more complex: available data consists of album reviews, not individual song reviews (www. amazon.com, www.epinions.com). Comments about a given song are fragmented all over an album review. Though some web-sites like www.last.fm allow users to comment on individual songs, the comments are too short (a few words such as “awesome song”) to be counted as a full review. In this paper, after presenting related work and contrasting it to our goals in Section 2, we discuss our prototype Music RS in Section 3. We devote Section 4 to our summarizer, that extracts comments on individual tracks from album reviews and produces a summary of those comments for each individual track recommended to the user. In Section 5, we report two types of evaluation: an intrinsic evaluation of the extraction components, and of the coverage of the summary; an extrinsic evaluation via a between-subject study. We found that users make quicker and more informed decisions when presented with the song review summaries as opposed to the full album review. 2 Related Work Over the last decade, summarization has become a hot topic for research. Quite a few systems were developed for different tasks, including multidocument summarization (Barzilay and McKeown, 2005; Soubbotin and Soubbotin, 2005; Nastase, 2008). 1376 What’s not to get? Yes, Maxwell, and Octopus are a bit silly! ... ..... “Something” and “Here Comes The Sun” are two of George’s best songs ever (and “Something” may be the single greatest love song ever). “Oh Darling” is a bluesy masterpiece with Paul screaming..... ....... “Come Together” contains a great riff, but he ended up getting sued over the lyrics by Chuck Berry...... Figure 1: A sample review for the album “Abbey Road” Whereas summarizing customer reviews can be seen as multi-document summarization, an added necessary step is to first extract the most important features customers focus on. Hence, summarizing customer reviews has mostly been studied as a combination of machine learning and NLP techniques (Hu and Liu, 2004; Gamon et al., 2005). For example, (Hu and Liu, 2004) use associative mining techniques to identify features that frequently occur in reviews taken from www.epinions.com and www. amazon.com. Then, features are paired to the nearest words that express some opinion on that feature. Most work on product reviews focuses on identifying sentences and polarity of opinion terms, not on generating a coherent summary from the extracted features, which is the main goal of our research. Exceptions are (Carenini et al., 2006; Higashinaka et al., 2006), whose focus was on extracting domain specific ontologies in order to structure summarization of customer reviews. Summarizing reviews on objects different from products, such as restaurants (Nguyen et al., 2007), or movies (Zhuang et al., 2006), has also been tackled, although not as extensively. We are aware of only one piece of work that focuses on music reviews (Downie and Hu, 2006). This study is mainly concerned with identifying descriptive patterns in positive or negative reviews but not on summarizing the reviews. 2.1 Summarizing song reviews is different As mentioned earlier, using album reviews for song summarization poses new challenges: a) Comments on features of a song are embedded and fragmented within the album reviews, as shown in Figure 1. It is necessary to correctly map features to songs. b) Each song needs to be identified each time it is referred to in the review. Titles are often abbreviated, and in different ways, even in the same review – e.g. see Octopus for Octopus’s Garden in Figure 1. Additionally, song titles need not be noun phrases and hence NP extraction algorithms miss many occurrences, as was shown by preliminary experiments we ran. c) Reviewers focus on both inherent features such as lyrics, genre and instruments, but also on people (artist, lyricist, producer etc.), unlike in product reviews where manufacturer/designer are rarely mentioned. This variety of features makes it harder to generate a coherent summary. 3 SongRecommend: Prototype Music RS Figure 2 shows the interface of our prototype Music RS. It is a simple interface dictated by our focus on the summarization process (but it was informed by a small pilot study). Moving from window to window and from top to bottom: a) The top leftmost window shows different devices on which the user listens to songs. These devices are monitored with a UPNP control point. Based on the messages received by the control point, the user activities, including the metadata of the song, are logged. b) Once the user chooses a certain song on one of the devices (see second window on top), we display more information about the song (third top window); we also identify related songs from the internet, including: other songs from the same album, popular songs of the artist and popular songs of related artists, as obtained from Yahoo Music. c) The top 25 recommendations are shown in the fourth top window. We use the SimpleKMeans Clustering (Mitchell, 1997) to identify and rank the top twenty-five songs which belong to the same cluster and are closest to the given song. Closeness between two songs in a cluster is measured as the number of attributes (album, artist etc) of the songs that match. d) When the user clicks on More Info for one of the recommended songs, the pop-up, bottom window is displayed, which contains the summary of the reviews for the specific song. 4 Extraction and Summarization Our summarization framework consists of the five tasks illustrated in Figure 3. The first two tasks pertain to information extraction, the last three to repackaging the information and generating a co1377 Figure 2: SongRecommend Interface Figure 3: Summarization Pipeline herent summary. Whereas the techniques we use for each individual step are state-of-the-art, our approach is innovative in that it integrates them into an effective end-to-end system. Its effectiveness is shown by the promising results obtained both via the intrinsic evaluation, and the user study. Our framework can be applied to any domain where reviews of individual components need to be summarized from reviews of collections, such as reviews of different hotels and restaurants in a city. Our corpus was opportunistically collected from www.amazon.com and www.epinions.com. It consists of 1350 album reviews across 27 albums (50 reviews per album). 50 randomly chosen reviews were used for development. Reviews have noise, since the writing is informal. We did not clean it, for example we did not correct spelling mistakes. This corpus was annotated for song titles and song features. Feature annotation consists of marking a phrase as a feature and matching it with the song to which the feature is attributed. Note that we have no a priori inventory of features; what counts as features of songs emerged from the annotation, since annotators were asked to annotate for noun phrases which contain “any song related term or terms spoken in the context of a song”. Further, they were given about 5 positive and 5 negative 1378 What’s not to get? Yes, <song id=3>Maxwell</song>, and <song id=5>Octopus</song> are a bit silly! ... ......... ......... <song id=2>“Something”</song> and <song id=7>“Here Comes The Sun”</song> are two of <feature id=(2,7)>George’s</feature> best songs ever (and <song id=2>“Something”</song> may be ...... <song id=4>“Oh Darling”</song> is a <feature id=4>bluesy masterpiece</feature> with <feature id=4>Paul</feature> screaming...... ..... <song id=1>“Come Together”</song> contains a great <feature id=1>riff</feature>, but ... Figure 4: A sample annotated review examples of features. Figure 4 shows annotations for the excerpt in Figure 1. For example in Figure 4, George, Paul, bluesy masterpiece and riff have been marked as features. Ten randomly chosen reviews were doubly annotated for song titles and features. The Kappa co-efficient of agreement on both was excellent (0.9), hence the rest of the corpus was annotated by one annotator only. The two annotators were considered to be in agreement on a feature if they marked the same head of phrase and attributed it to the same song. We will now turn to describing the component tasks. The algorithms are described in full in (Tata, 2010). 4.1 Title Extraction Song identification is the first step towards summarization of reviews. We identify a string of words as the title of a song to be extracted from an album review if it (1) includes some or all the words in the title of a track of that album, and (2) this string occurs in the right context. Constraint (2) is necessary because the string of words corresponding to the title may appear in the lyrics of the song or anywhere else in the review. The string Maxwell’s Silver Hammer counts as a title only in sentence (a) below; the second sentence is a verse in the lyrics: a. Then, the wild and weird “Maxwell’s Silver Hammer.” b. Bang, Bang, maxwell’s silver hammer cam down on her head. Similar to Named Entity Recognition (Schedl et al., 2007), our approach to song title extraction is based on n-grams. We proceed album by album. Given the reviews for an album and the list of songs in that album, first, we build a lexicon of all the words in the song titles. We also segment the reviews into sentences via sentence boundary detection. All 1,2,3,4-grams for each sentence (the upper-bound 4 was determined experimentally) in the review are generated. First, n-grams that contain at least one word with an edit distance greater than one from a word in the lexicon are filtered out. Second, if higher and lower order n-grams overlap at the same position in the same sentence, lower order n-grams are filtered out. Third, the n-grams are merged if they occur sequentially in a sentence. Fourth, the n-grams are further filtered to include only those where (i) the n-gram is within quotation marks; and/or (ii) the first character of each word in the n-gram is upper case. This filters n-grams such as those shown in sentence (b) above. All the n-grams remaining at this point are potential song titles. Finally, for each n-gram, we retrieve the set of IDs for each of its words and intersect those sets. This intersection always resulted in one single song ID, since song titles in each album differ by at least one content word. Recall that the algorithm is run on reviews for each album separately. 4.2 Feature Extraction Once the song titles are identified in the album review, sentences with song titles are used as anchors to (1) identify segments of texts that talk about a specific song, and then (2) extract the feature(s) that the pertinent text segment discusses. The first step roughly corresponds to identifying the flow of topics in a review. The second step corresponds to identifying the properties of each song. Both steps would greatly benefit from reference resolution, but current algorithms still have a low accuracy. We devised an approach that combines text tiling (Hearst, 1994) and domain heuristics. The text tiling algorithm divides the text into coherent discourse units, to describe the sub-topic structure of the given text. We found the relatively coarse segments the text tiling algorithm provides sufficient to identify different topics. An album review is first divided into segments using the text tiling algorithm. Let [seg1, seg2, ..., segk] be the segments obtained. The segments that contain potential features of a song are identified using the following heuristics: Step 1: Include segi if it contains a song title. 1379 These segments are more likely to contain features of songs as they are composed of the sentences surrounding the song title. Step 2: Include segi+1 if segi is included and segi+1 contains one or more feature terms. Since we have no a priori inventory of features (the feature annotation will be used for evaluation, not for development), we use WordNet (Fellbaum, 1998) to identify feature terms: i.e., those nouns whose synonyms, direct hypernym or direct hyponym, or the definitions of any of those, contain the terms “music” or “song”, or any form of these words like “musical”, “songs” etc, for at least one sense of the noun. Feature terms exclude the words “music”, “song”, the artist/band/album name as they are likely to occur across album reviews. All feature terms in the final set of segments selected by the heuristics are taken to be features of the song described by that segment. 4.3 Sentence Partitioning and Regeneration After extracting the sentences containing the features, the next step is to divide the sentences into two or more “sub-sentences”, if necessary. For example, “McCartney’s bouncy bass-line is especially wonderful, and George comes in with an excellent, minimal guitar solo.” discusses both features bass and guitar. Only a portion of the sentence describes the guitar. This sentence can thus be divided into two individual sentences. Removing parts of sentences that describe another feature, will have no effect on the summary as a whole as the portions that are removed will be present in the group of sentences that describe the other feature. To derive n sentences, each concerning a single feature f, from the original sentence that covered n features, we need to: 1. Identify portions of sentences relevant to each feature f (partitioning) 2. Regenerate each portion as an independent sentence, which we call f-sentence. To identify portions of the sentence relevant to the single feature f, we use the Stanford Typed Dependency Parser (Klein and Manning, 2002; de Marnee and Manning, 2008). Typed Dependencies describe grammatical relationships between pairs of words in a sentence. Starting from the feature term f in question, we collect all the nouns, adjectives and verbs that are directly related to it in the sentence. These nouns, adjectives and verbs 1. “Maxwell” is a bit silly. 2. “Octopus” is a bit silly. 3. “Something” is George’s best song. 4. “Here Comes The Sun” is George’s best song. 5. “Something” may be the single greatest love song. 6. “Oh! Darling” is a bluesy masterpiece. 7. “Come Together” contains a great riff. Figure 5: f-sentences corresponding to Figure 1 become the components of the new f-sentence. Next, we need to adjust their number and forms. This is a natural language generation task, specifically, sentence realization. We use YAG (McRoy et al., 2003), a template based sentence realizer. clause is the main template used to generate a sentence. Slots in a template can in turn be templates. The grammatical relationships obtained from the Typed Dependency Parser such as subject and object identify the slots and the template the slots follows; the words in the relationship fill the slot. We use a morphological tool (Minnen et al., 2000) to obtain the base form from the original verb or noun, so that YAG can generate grammatical sentences. Figure 5 shows the regenerated review from Figure 1. YAG regenerates as many f-sentences from the original sentence, as many features were contained in it. By the end of this step, for each feature f of a certain song si, we have generated a set of f-sentences. This set also contains every original sentence that only covered the single feature f. 4.4 Grouping f-sentences are further grouped, by sub-feature and by polarity. As concerns sub-feature grouping, consider the following f-sentences for the feature guitar: a. George comes in with an excellent, minimal guitar solo. b. McCartney laid down the guitar lead for this track. c. Identical lead guitar provide the rhythmic basis for this song. The first sentence talks about the guitar solo, the second and the third about the lead guitar. This step will create two subgroups, with sentence a in one group and sentences b and c in another. We 1380 Let [fx-s1, fx-s2, ...fx-sn] be the set of sentences for feature fx and song Sy Step 1: Find the longest common n-gram (LCN) between fx-si and fx-sj for all i ̸= j: LCN(fx-si, fx-sj) Step 2: If LCN(fx-si, fx-sj) contains the feature term and is not the feature term alone, fx-si and fx-sj are in the same group. Step 3: For any fx-si, if LCN(fx-si, fx-sj) for all i and j, is the feature term, then fx-si belongs to the default group for the feature. Figure 6: Grouping sentences by sub-features identify subgroups via common n-grams between f-sentences, and make sure that only n-grams that are related to feature f are identified at this stage, as detailed in Figure 6. When the procedure described in Figure 6 is applied to the three sentences above, it identifies guitar as the longest pertinent LCN between a and b, and between a and c; and guitar lead between b and c (we do not take into account linear order within n-grams, hence guitar lead and lead guitar are considered identical). Step 2 in Figure 6 will group b and c together since guitar lead properly contains the feature term guitar. In Step 3, sentence a is sentence fx-si such that its LCN with all other sentences (b and c) contains only the feature term; hence, sentence a is left on its own. Note that Steps 2 and 3 ensure that, among all the possible LNCs between pair of sentences, we only consider the ones containing the feature in question. As concerns polarity grouping, different reviews may express different opinions regarding a particular feature. To generate a coherent summary that mentions conflicting opinions, we need to subdivide f-sentences according to polarity. We use SentiWordNet (Esuli and Sebastiani, 2006), an extension of WordNet where each sense of a word is augmented with the probability of that sense being positive, negative or neutral. The overall sentence score is based on the scores of the adjectives contained in the sentence. Since there are a number of senses for each word, an adjective ai in a sentence is scored as the normalized weighted scores of each sense of the adjective. For each ai, we compute three scores, positive, as shown in Formula 1, negative and obExample: The lyrics are the best Adjectives in the sentence: best Senti-wordnet Scores of best: Sense 1 (frequency=2): positive = 0.625, negative =0 , objective = 0.375 Sense 2 (frequency=1): positive = 0.75, negative = 0, objective = 0.25 Polarity Scores Calculation: positive(best) = 2∗0.625+1∗0.75 (2+1) = 0.67 negative(best) = 2∗0+1∗0 (2+1) = 0 objective(best) = 2∗0.375+1∗0.25 (2+1) = 0.33 Since the sentence contains only the adjective best, its polarity is positive, from: Max (positive(best), negative(best), objective(best)) Figure 7: Polarity Calculation jective, which are computed analogously: pos(ai) = freq1 ∗pos1 + ... + freqn ∗posn (freq1 + .... + freqn) (1) ai is the ith adjective, freqj is the frequency of the jth sense of ai as given by Wordnet, and posj is the positive score of the jth sense of ai, as given by SentiWordnet. Figure 7 shows an example of calculating the polarity of a sentence. For an f-sentence, three scores will be computed, as the sum of the corresponding scores (positive, negative, objective) of all the adjectives in the sentence. The polarity of the sentence is determined by the maximum of these three scores. 4.5 Selection and Ordering Finally, the generation of a coherent summary involves selection of the sentences to be included, and ordering them in a coherent fashion. This step has in input groups of f-sentences, where each group pertains to the feature f, one of its subfeatures, and one polarity type (positive, negative, objective). We need to select one sentence from each subgroup to make sure that all essential concepts are included in the summary. Note that if there are contrasting opinions on one feature or subfeatures, one sentence per polarity will be extracted, resulting in potentially inconsistent opinions on that feature to be included in the review (we did not observe this happening frequently, and even if it did, it did not appear to confuse our users). Recall that at this point, most f-sentences have been regenerated from portions of original sen1381 tences (see Section 4.3). Each f-sentence in a subgroup is assigned a score which is equivalent to the number of features in the original sentence from which the f-sentence was obtained. The sentence which has the lowest score in each subgroup is chosen as the representative for that subgroup. If multiple sentences have the lowest score, one sentence is selected randomly. Our assumption is that among the original sentences, a sentence that talks about one feature only is likely to express a stronger opinion about that feature than a sentence in which other features are present. We order the sentences by exploiting a music ontology (Giasson and Raimond, 2007). We have extended this ontology to include few additional concepts that correspond to features identified in our corpus. Also, we extended each of the classes by adding the domain to which it belongs. We identified a total of 20 different domains for all the features. For example, [saxophone,drums] belongs to the domain Instrument, and [tone, vocals] belong to the domain Sound. We also identified the priority order in which each of these domains should appear in the final summary. The ordering of the domains is such that first we present the general features of the song (e.g. Song) domain, then present more specific domains (e.g. Sound, Instrument). f−sentences of a single domain form one paragraph in the final summary. However, features domains that are considered as sub-domains of another domain are included in the same paragraph, but are ordered next to the features of the parent domain. The complete list of domains is described in (Tata, 2010). f-sentences are grouped and ordered according to the domain of the features. Figure 8 shows a sample summary when the extracted sentences are ordered via this method. “The Song That Jane Likes” is cute. The song has some nice riffs by Leroi Moore. “The Song That Jane Likes” is also amazing funk number. The lyrics are sweet and loving. The song carries a light-hearted tone. It has a catchy tune. The song features some nice accents. “The Song That Jane Likes” is beautiful song with great rhythm. The funky beat will surely make a move. It is a heavily acoustic guitar-based song. Figure 8: Sample summary 5 Evaluation In this section we report three evaluations, two intrinsic and one extrinsic: evaluation of the song title and feature extraction steps; evaluation of the informativeness of summaries; and a user study to judge how summaries affect decision making. 5.1 Song Title and Feature Extraction The song title extraction and feature extraction algorithms (Sections 4.1 and 4.2) were manually evaluated on 100 reviews randomly taken from the corpus (2 or 3 from each album). This relatively small number is due to the need to conduct the evaluation manually. The 100 reviews contained 1304 occurrences of song titles and 898 occurrences of song features, as previously annotated. 1294 occurrences of song titles were correctly identified; additionally, 123 spurious occurrences were also identified. This results in a precision of 91.3%, and recall of 98%. The 10 occurrences that were not identified contained either abbreviations like Dr. for Doctor or spelling mistakes (recall that we don’t clean up mistakes). Of the 898 occurrences of song features, 853 were correctly identified by our feature extraction algorithm, with an additional 41 spurious occurrences. This results in a precision of 95.4% and a recall of 94.9%. Note that a feature (NP) is considered as correctly identified, if its head noun is annotated in a review for the song with correct ID. As a baseline comparison, we implemented the feature extraction algorithm from (Hu and Liu, 2004). We compared their algorithm to ours on 10 randomly chosen reviews from our corpus, for a total of about 500 sentences. Its accuracy (40.8% precision, and 64.5% recall) is much lower than ours, and than their original results on product reviews (72% precision, and 80% recall). 5.2 Informativeness of the summaries To evaluate the information captured in the summary, we randomly selected 5 or 6 songs from 10 albums, and generated the corresponding 52 summaries, one per song – this corresponds to a test set of about 500 album reviews (each album has about 50 reviews). Most summary evaluation schemes, for example the Pyramid method (Harnly et al., 2005), make use of reference summaries written by humans. We approximate those goldstandard reference summaries with 2 or 3 critic reviews per album taken from www.pitchfork. 1382 com, www.rollingstone.com and www. allmusic.com. First, we manually annotated both critic reviews and the automatically generated summaries for song titles and song features. 302, i.e., 91.2% of the features identified in the critic reviews are also identified in the summaries (recall that a feature is considered as identified, if the head-noun of the NP is identified by both the critic review and the summary, and attributed to the same song). 64 additional features were identified, for a recall of 82%. It is not surprising that additional features may appear in the summaries: even if only one of the 50 album reviews talks about that feature, it is included in the summary. Potentially, a threshold on frequency of feature mention could increase recall, but we found out that even a threshold of two significantly affects precision. In a second evaluation, we used our Feature Extraction algorithm to extract features from the critic reviews, for each song whose summary needs to be evaluated. This is an indirect evaluation of that algorithm, in that it shows it is not affected by somewhat different data, since the critic reviews are more formally written. 375, or 95% of the features identified in the critic reviews are also identified in the summaries. 55 additional features were additionally identified, for a recall of 87.5%. These values are comparable, even if slightly higher, to the precision and recall of the manual annotation described above. 5.3 Between-Subject User Study Our intrinsic evaluation gives satisfactory results. However, we believe the ultimate measure of such a summarization algorithm is an end-to-end evaluation to ascertain whether it affects user behavior, and how. We conducted a between-subject user study, where users were presented with two different versions of our Music RS. For each of the recommended songs, the baseline version provides only whole album reviews, the experimental version provides the automatically generated song feature summary, as shown in Figure 2. The interface for the baseline version is similar, but the summary in the bottom window is replaced by the corresponding album review. The presented review is the one among the 50 reviews for that album whose length is closest to the average length of album reviews in the corpus (478 words). Each user was presented with 5 songs in succession, with 3 recommendations each (only the top 3 recommendations were presented among the available 25, see Section 3). Users were asked to select at least one recommendation for each song, namely, to click on the url where they can listen to the song. They were also asked to base their selection on the information provided by the interface. The first song was a test song for users to get acquainted with the system. We collected comprehensive timed logs of the user actions, including clicks, when windows are open and closed, etc. After using the system, users were administered a brief questionnaire which included questions on a 5-point Likert Scale. 18 users interacted with the baseline version and 21 users with the experimental version (five additional subjects were run but their log data was not properly saved). All users were students at our University, and most of them, graduate students (no differences were found due to gender, previous knowledge of music, or education level). Our main measure is time on task, the total time taken to select the recommendations from song 2 to song 5 – this excludes the time spent listening to the songs. A t-test showed that users in the experimental version take less time to make their decision when compared to baseline subjects (p = 0.019, t = 2.510). This is a positive result, because decreasing time to selection is important, given that music collections can include millions of songs. However, time-on-task basically represents the time it takes users to peruse the review or summary, and the number of words in the summaries is significantly lower than the number of words in the reviews (p < 0.001, t = 16.517). Hence, we also analyzed the influence of summaries on decision making, to see if they have any effects beyond cutting down on the number of words to read. Our assumption is that the default choice is to choose the first recommendation. Users in the baseline condition picked the first recommendation as often as the other two recommendations combined; users in the experimental condition picked the second and third recommendations more often than the first, and the difference between the two conditions is significant (χ2 = 8.74, df = 1, p = 0.003). If we examine behavior song by song, this holds true especially for song 3 (χ2 = 12.3, df = 1, p < 0.001) and song 4 (χ2 = 5.08, df = 1, p = 0.024). We speculate that users in the experimental condition 1383 are more discriminatory in their choices, because important features of the recommended songs are evident in the summaries, but are buried in the album reviews. For example, for Song 3, only one of the 20 sentences in the album review is about the first recommended song, and is not very positive. Negative opinions are much more evident in the review summaries. The questionnaires included three common questions between the two conditions. The experimental subjects gave a more positive assessment of the length of the summary than the baseline subjects (p = 0.003, t = −3.248, df = 31.928). There were no significant differences on the other two questions, feeling overwhelmed by the information provided; and whether the review/summary helped them to quickly make their selection. A multiple Linear Regression with, as predictors, the number of words the user read before making the selection and the questions, and time on task as dependent variable, revealed only one, not surprising, correlation: the number of words the user read correlates with time on task (R2 = 0.277, β = 0.509, p = 0.004). Users in the experimental version were also asked to rate the grammaticality and coherence of the summary. The average rating was 3.33 for grammaticality, and 3.14 for coherence. Whereas these numbers in isolation are not too telling, they are at least suggestive that users did not find these summaries badly written. We found no significant correlations between grammaticality and coherence of summaries, and time on task. 6 Discussion and Conclusions Most summarization research on customer reviews focuses on obtaining features of the products, but not much work has been done on presenting them as a coherent summary. In this paper, we described a system that uses information extraction and summarization techniques in order to generate summaries of individual songs from multiple album reviews. Whereas the techniques we have used are state-of-the-art, the contribution of our work is integrating them in an effective end-to-end system. We first evaluated it intrinsically as concerns information extraction, and the informativeness of the summaries. Perhaps more importantly, we also ran an extrinsic evaluation in the context of our prototype Music RS. Users made quicker decisions and their choice of recommendations was more varied when presented with song review summaries than with album reviews. Our framework can be applied to any domain where reviews of individual components need to be summarized from reviews of collections, such as travel reviews that cover many cities in a country, or different restaurants in a city. References Regina Barzilay and Kathleen McKeown. 2005. Sentence fusion for multidocument news summarization. Computational Linguistics, 31(3):297–328. Giuseppe Carenini, Raymond Ng, and Adam Pauls. 2006. Multi-document summarization of evaluative text. In Proceedings of EACL. Oscar Clema. 2006. Interaction Design for Recommender Systems. Ph.D. thesis, Universitat Pompeu Fabra, Barcelona, July. Marie-Catherine de Marnee and Christopher D. Manning. 2008. Stanford Typed Dependencies Manual. http://nlp.stanford.edu/software/dependencies manual.pdf. J. Stephen Downie and Xiao Hu. 2006. Review mining for music digital libraries: Phase ii. In Proceedings of the 6th ACM/IEEE-CS Joint Conference on Digital Libraries, pages 196–197, Chapel Hill, NC, USA. Andrea Esuli and Fabrizio Sebastiani. 2006. SentiWordNet: A publicly available lexical resource for opinion mining. In Proceedings of LREC-06, the 5th Conference on Language Resources and Evaluation, Genova, IT. Christiane Fellbaum, editor. 1998. WordNet: an electronic lexical database. MIT Press. Michael Gamon, Anthony Aue, Simon Corston-Oliver, and Eric Ringger. 2005. Pulse: Mining customer opinions from free text. In Advances in Intelligent Data Analysis VI, volume 3646/2005 of Lecture Notes in Computer Science, pages 121–132. Springer Berlin / Heidelberg. Frederick Giasson and Yves Raimond. 2007. Music ontology specification. Working draft, February. http://pingthesemanticweb.com/ontology/mo/. Aaron Harnly, Ani Nenkova, Rebecca Passonneau, and Owen Rambow. 2005. Automation of summary evaluation by the Pyramid method. In Proceedings of the Conference on Recent Advances in Natural Language Processing. Marti A. Hearst. 1994. Multi-paragraph segmentation of expository text. In Proceedings of the 32nd Meeting of the Association for Computational Linguistics, Las Cruces, NM, June. 1384 Ryuichiro Higashinaka, Rashmi Prasad, and Marilyn Walker. 2006. Learning to Generate Naturalistic Utterances Using Reviews in Spoken Dialogue Systems. In COLING-ACL06, Sidney, Australia. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In Proceedings of KDD, Seattle, Washington, USA, August. Dan Klein and Christopher D. Manning. 2002. Fast exact inference with a factored model for natural language parsing. In Advances in Neural Information Processing Systems 15, pages 3–10. Susan McRoy, Songsak Ukul, and Syed Ali. 2003. An augmented template-based approach to text realization. In Natural Language Engineering, pages 381– 420. Cambridge Press. Guido Minnen, John Carroll, and Darren Pearce. 2000. Robust, applied morphological generation. In Proceedings of the 1st International Natural Language Generation Conference. Tom Mitchell. 1997. Machine Learning. McGraw Hill. Vivi Nastase. 2008. Topic-driven multi-document summarization with encyclopedic knowledge and spreading activation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Patrick Nguyen, Milind Mahajan, and Geoffrey Zweig. 2007. Summarization of multiple user reviews in the restaurant domain. Technical Report MSR-TR2007-126, Microsoft, September. Markus Schedl, Gerhard Widmer, Tim Pohle, and Klaus Seyerlehner. 2007. Web-based detection of music band members and line-up. In Proceedings of the Australian Computer Society. M. Soubbotin and S. Soubbotin. 2005. Trade-Off Between Factors Influencing Quality of the Summary. In Document Understanding Workshop (DUC), Vancouver, BC, Canada. Swati Tata. 2010. SongRecommend: a Music Recommendation System with Fine-Grained Song Reviews. Ph.D. thesis, University of Illinois, Chicago, IL. 2008. UPnP Device Architecture Version 1.0. (www.upnp.org). Li Zhuang, Feng Jing, and Xiaoyan Zhu. 2006. Movie review mining and summarization. In Conference on Information and Knowledge Management, Arlington, Virginia, USA. 1385
|
2010
|
140
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1386–1395, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics A study of Information Retrieval weighting schemes for sentiment analysis Georgios Paltoglou University of Wolverhampton Wolverhampton, United Kingdom [email protected] Mike Thelwall University of Wolverhampton Wolverhampton, United Kingdom [email protected] Abstract Most sentiment analysis approaches use as baseline a support vector machines (SVM) classifier with binary unigram weights. In this paper, we explore whether more sophisticated feature weighting schemes from Information Retrieval can enhance classification accuracy. We show that variants of the classic tf.idf scheme adapted to sentiment analysis provide significant increases in accuracy, especially when using a sublinear function for term frequency weights and document frequency smoothing. The techniques are tested on a wide selection of data sets and produce the best accuracy to our knowledge. 1 Introduction The increase of user-generated content on the web in the form of reviews, blogs, social networks, tweets, fora, etc. has resulted in an environment where everyone can publicly express their opinion about events, products or people. This wealth of information is potentially of vital importance to institutions and companies, providing them with ways to research their consumers, manage their reputations and identify new opportunities. Wright (2009) claims that “for many businesses, online opinion has turned into a kind of virtual currency that can make or break a product in the marketplace”. Sentiment analysis, also known as opinion mining, provides mechanisms and techniques through which this vast amount of information can be processed and harnessed. Research in the field has mainly, but not exclusively, focused in two subproblems: detecting whether a segment of text, either a whole document or a sentence, is subjective or objective, i.e. contains an expression of opinion, and detecting the overall polarity of the text, i.e. positive or negative. Most of the work in sentiment analysis has focused on supervised learning techniques (Sebastiani, 2002), although there are some notable exceptions (Turney, 2002; Lin and He, 2009). Previous research has shown that in general the performance of the former tend to be superior to that of the latter (Mullen and Collier, 2004; Lin and He, 2009). One of the main issues for supervised approaches has been the representation of documents. Usually a bag of words representation is adopted, according to which a document is modeled as an unordered collection of the words that it contains. Early research by Pang et al. (2002) in sentiment analysis showed that a binary unigrambased representation of documents, according to which a document is modeled only by the presence or absence of words, provides the best baseline classification accuracy in sentiment analysis in comparison to other more intricate representations using bigrams, adjectives, etc. Later research has focused on extending the document representation with more complex features such as structural or syntactic information (Wilson et al., 2005), favorability measures from diverse sources (Mullen and Collier, 2004), implicit syntactic indicators (Greene and Resnik, 2009), stylistic and syntactic feature selection (Abbasi et al., 2008), “annotator rationales” (Zaidan et al., 2007) and others, but no systematic study has been presented exploring the benefits of employing more sophisticated models for assigning weights to word features. In this paper, we examine whether term weighting functions adopted from Information Retrieval (IR) based on the standard tf.idf formula and adapted to the particular setting of sentiment analysis can help classification accuracy. We demonstrate that variants of the original tf.idf weighting scheme provide significant increases in classification performance. The advantages of the approach are that it is intuitive, computationally efficient 1386 and doesn’t require additional human annotation or external sources. Experiments conducted on a number of publicly available data sets improve on the previous state-of-the art. The next section provides an overview of relevant work in sentiment analysis. In section 3 we provide a brief overview of the original tf.idf weighting scheme along with a number of variants and show how they can be applied to a classification scenario. Section 4 describes the corpora that were used to test the proposed weighting schemes and section 5 discusses the results. Finally, we conclude and propose future work in section 6. 2 Prior Work Sentiment analysis has been a popular research topic in recent years. Most of the work has focused on analyzing the content of movie or general product reviews, but there are also applications to other domains such as debates (Thomas et al., 2006; Lin et al., 2006), news (Devitt and Ahmad, 2007) and blogs (Ounis et al., 2008; Mishne, 2005). The book of Pang and Lee (2008) presents a thorough overview of the research in the field. This section presents the most relevant work. Pang et al. (2002) conducted early polarity classification of reviews using supervised approaches. They employed Support Vector Machines (SVMs), Naive Bayes and Maximum Entropy classifiers using a diverse set of features, such as unigrams, bigrams, binary and term frequency feature weights and others. They concluded that sentiment classification is more difficult that standard topic-based classification and that using a SVM classifier with binary unigrambased features produces the best results. A subsequent innovation was the detection and removal of the objective parts of documents and the application of a polarity classifier on the rest (Pang and Lee, 2004). This exploited text coherence with adjacent text spans which were assumed to belong to the same subjectivity or objectivity class. Documents were represented as graphs with sentences as nodes and association scores between them as edges. Two additional nodes represented the subjective and objective poles. The weights between the nodes were calculated using three different, heuristic decaying functions. Finding a partition that minimized a cost function separated the objective from the subjective sentences. They reported a statistically significant improvement over a Naive Bayes baseline using the whole text but only slight increase compared to using a SVM classifier on the entire document. Mullen and Collier (2004) used SVMs and expanded the feature set for representing documents with favorability measures from a variety of diverse sources. They introduced features based on Osgood’s Theory of Semantic Differentiation (Osgood, 1967) using WordNet to derive the values of potency, activity and evaluative of adjectives and Turney’s semantic orientation (Turney, 2002). Their results showed that using a hybrid SVM classifier, that uses as features the distance of documents from the separating hyperplane, with all the above features produces the best results. Whitelaw et al. (2005) added fine-grained semantic distinctions in the feature set. Their approach was based on a lexicon created in a semisupervised fashion and then manually refined It consists of 1329 adjectives and their modifiers categorized under several taxonomies of appraisal attributes based on Martin and White’s Appraisal Theory (2005). They combined the produced appraisal groups with unigram-based document representations as features to a Support Vector Machine classifier (Witten and Frank, 1999), resulting in significant increases in accuracy. Zaidan et al. (2007) introduced “annotator rationales”, i.e. words or phrases that explain the polarity of the document according to human annotators. By deleting rationale text spans from the original documents they created several contrast documents and constrained the SVM classifier to classify them less confidently than the originals. Using the largest training set size, their approach significantly increased the accuracy on a standard data set (see section 4). Prabowo and Thelwall (2009) proposed a hybrid classification process by combining in sequence several ruled-based classifiers with a SVM classifier. The former were based on the General Inquirer lexicon (Wilson et al., 2005), the MontyLingua part-of-speech tagger (Liu, 2004) and co-occurrence statistics of words with a set of predefined reference words. Their experiments showed that combining multiple classifiers can result in better effectiveness than any individual classifier, especially when sufficient training data isn’t available. In contrast to machine learning approaches that require labeled corpora for training, Lin and 1387 He (2009) proposed an unsupervised probabilistic modeling framework, based on Latent Dirichlet Allocation (LDA). The approach assumes that documents are a mixture of topics, i.e. probability distribution of words, according to which each document is generated through an hierarchical process and adds an extra sentiment layer to accommodate the opinionated nature (positive or negative) of the document. Their best attained performance, using a filtered subjectivity lexicon and removing objective sentences in a manner similar to Pang and Lee (2004), is only slightly lower than that of a fully-supervised approach. 3 A study of non-binary weights We use the terms “features”, “words” and “terms” interchangeably in this paper, since we mainly focus on unigrams. The approach nonetheless can easily be extended to higher order n-grams. Each document D therefore is represented as a bag-ofwords feature vector: D = w1, w2, ..., w|V | where |V | is the size of the vocabulary (i.e. the number of unique words) and wi, i = 1, . . . , |V | is the weight of term i in document D. Despite the significant attention that sentiment analysis has received in recent years, the best accuracy without using complex features (Mullen and Collier, 2004; Whitelaw et al., 2005) or additional human annotations (Zaidan et al., 2007) is achieved by employing a binary weighting scheme (Pang et al., 2002), where wi = 1, if tfi > 0 and wi = 0, if tfi = 0, where tfi is the number of times that term i appears in document D (henceforth raw term frequency) and utilizing a SVM classifier. It is of particular interest that using tfi in the document representation usually results in decreased accuracy, a result that appears to be in contrast with topic classification (Mccallum and Nigam, 1998; Pang et al., 2002). In this paper, we also utilize SVMs but our study is centered on whether more sophisticated than binary or raw term frequency weighting functions can improve classification accuracy. We base our approach on the classic tf.idf weighting scheme from Information Retrieval (IR) and adapt it to the domain of sentiment classification. 3.1 The classic tf.idf weighting schemes The classic tf.idf formula assigns weight wi to term i in document D as: wi = tfi · idfi = tfi · log N dfi (1) where tfi is the number of times term i occurs in D, idfi is the inverse document frequency of term i, N is the total number of documents and dfi is the number of documents that contain term i. The utilization of tfi in classification is rather straightforward and intuitive but, as previously discussed, usually results in decreased accuracy in sentiment analysis. On the other hand, using idf to assign weights to features is less intuitive, since it only provides information about the general distribution of term i amongst documents of all classes, without providing any additional evidence of class preference. The utilization of idf in information retrieval is based on its ability to distinguish between content-bearing words (words with some semantical meaning) and simple function words, but this behavior is at least ambiguous in classification. Table 1: SMART notation for term frequency variants. maxt(tf) is the maximum frequency of any term in the document and avg dl is the average number of terms in all the documents. For ease of reference, we also include the BM25 tf scheme. The k1 and b parameters of BM25 are set to their default values of 1.2 and 0.95 respectively (Jones et al., 2000). Notation Term frequency n (natural) tf l (logarithm) 1 + log(tf) a (augmented) 0.5 + 0.5·tf maxt(tf) b (boolean) 1, tf > 0 0, otherwise L (log ave) 1+log(tf) 1+log(avg dl) o (BM25) (k1+1)·tf k1 (1−b)+b· dl avg dl +tf 3.2 Delta tf.idf Martineau and Finin (2009) provide a solution to the above issue of idf utilization in a classification scenario by localizing the estimation of idf to the documents of one or the other class and subtracting the two values. Therefore, the weight of term 1388 Table 2: SMART notation for inverse document frequency variants. For ease of reference we also include the BM25 idf factor and also present the extensions of the original formulations with their ∆variants. Notation Inverse Document Frequency n (no) 1 t (idf) log N df p (prob idf) log N−df df k (BM25 idf) log N−df+0.5 df+0.5 ∆(t) (Delta idf) log N1·df2 N2·df1 ∆(t′) (Delta smoothed idf) log N1·df2+0.5 N2·df1+0.5 ∆(p) (Delta prob idf) log (N1−df1)·df2 df1·(N2−df2) ∆(p′) (Delta smoothed prob idf) log (N1−df1)·df2+0.5 (N2−df2)·df1+0.5 ∆(k) (Delta BM25 idf) log (N1−df1+0.5)·df2+0.5 (N2−df2+0.5)·df1+0.5 i in document D is estimated as: wi = tfi · log2( N1 dfi,1 ) −tfi · log2( N2 dfi,2 ) = tfi · log2(N1 · dfi,2 dfi,1 · N2 ) (2) where Nj is the total number of training documents in class cj and dfi,j is the number of training documents in class cj that contain term i. The above weighting scheme was appropriately named Delta tf.idf. The produced results (Martineau and Finin, 2009) show that the approach produces better results than the simple tf or binary weighting scheme. Nonetheless, the approach doesn’t take into consideration a number of tested notions from IR, such as the non-linearity of term frequency to document relevancy (e.g. Robertson et al. (2004)) according to which, the probability of a document being relevant to a query term is typically sublinear in relation to the number of times a query term appears in the document. Additionally, their approach doesn’t provide any sort of smoothing for the dfi,j factor and is therefore susceptible to errors in corpora where a term occurs in documents of only one or the other class and therefore dfi,j = 0 . 3.3 SMART and BM25 tf.idf variants The SMART retrieval system by Salton (1971) is a retrieval system based on the vector space model (Salton and McGill, 1986). Salton and Buckley (1987) provide a number of variants of the tf.idf weighting approach and present the SMART notation scheme, according to which each weighting function is defined by triples of letters; the first one denotes the term frequency factor, the second one corresponds to the inverse document frequency function and the last one declares the normalization that is being applied. The upper rows of tables 1, 2 and 3 present the three most commonly used weighting functions for each factor respectively. For example, a binary document representation would be equivalent to SMART.bnn1 or more simply bnn, while a simple raw term frequency based would be notated as nnn or nnc with cosine normalization. Table 3: SMART normalization. Notation Normalization n (none) 1 c (cosine) 1 √ w2 1+w2 2+...+w2n Significant research has been done in IR on diverse weighting functions and not all versions of SMART notations are consistent (Manning et al., 2008). Zobel and Moffat (1998) provide an exhaustive study but in this paper, due to space constraints, we will follow the concise notation presented by Singhal et al. (1995). The BM25 weighting scheme (Robertson et al., 1994; Robertson et al., 1996) is a probabilistic model for information retrieval and is one of the most popular and effective algorithms used in information retrieval. For ease of reference, we incorporate the BM25 tf and idf factors into the SMART annotation scheme (last row of table 1 and 4th row of table 2), therefore the weight wi of term i in document D according to the BM25 scheme is notated as SMART.okn or okn. Most of the tf weighting functions in SMART and the BM25 model take into consideration the non-linearity of document relevance to term fre1Typically, a weighting function in the SMART system is defined as a pair of triples, i.e. ddd.qqq where the first triple corresponds to the document representation and the second to the query representation. In the context that the SMART annotation is used here, we will use the prefix SMART for the first part and a triple for the document representation in the second part, i.e. SMART.ddd, or more simply ddd. 1389 quency and thus employ tf factors that scale sublinearly in relation to term frequency. Additionally, the BM25 tf variant also incorporates a scaling for the length of the document, taking into consideration that longer documents will by definition have more term occurences2. Effective weighting functions is a very active research area in information retrieval and it is outside the scope of this paper to provide an in-depth analysis but significant research can be found in Salton and McGill (1986), Robertson et al. (2004), Manning et al. (2008) or Armstrong et al. (2009) for a more recent study. 3.4 Introducing SMART and BM25 Delta tf.idf variants We apply the idea of localizing the estimation of idf values to documents of one class but employ more sophisticated term weighting functions adapted from the SMART retrieval system and the BM25 probabilistic model. The resulting idf weighting functions are presented in the lower part of table 2. We extend the original SMART annotation scheme by adding Delta (∆) variants of the original idf functions and additionally introduce smoothed Delta variants of the idf and the prob idf factors for completeness and comparative reasons, noted by their accented counterparts. For example, the weight of term i in document D according to the o∆(k)n weighting scheme where we employ the BM25 tf weighting function and utilize the difference of class-based BM25 idf values would be calculated as: wi = (k1 + 1) · tfi K + tfi · log(N1 −dfi,1 + 0.5 dfi,1 + 0.5 ) − (k1 + 1) · tfi K + tfi · log(N2 −dfi,2 + 0.5 dfi,2 + 0.5 ) = (k1 + 1) · tfi K + tfi · log (N1 −dfi,1 + 0.5) · (dfi,2 + 0.5) (N2 −dfi,2 + 0.5) · (dfi,1 + 0.5) where K is defined as k1 (1 −b) + b · dl avg dl . However, we used a minor variation of the above formulation for all the final accented weighting functions in which the smoothing factor is added to the product of dfi with Ni (or its variation for ∆(p′) and ∆(k)), rather than to the dfi alone as the 2We deliberately didn’t extract the normalization component from the BM25 tf variant, as that would unnecessarily complicate the notation. above formulation would imply (see table 2). The above variation was made for two reasons: firstly, when the dfi’s are larger than 1 then the smoothing factor influences the final idf value only in a minor way in the revised formulation, since it is added only after the multiplication of the dfi with Ni (or its variation). Secondly, when dfi = 0, then the smoothing factor correctly adds only a small mass, avoiding a potential division by zero, where otherwise it would add a much greater mass, because it would be multiplied by Ni. According to this annotation scheme therefore, the original approach by Martineau and Finin (2009) can be represented as n∆(t)n. We hypothesize that the utilization of sophisticated term weighting functions that have proved effective in information retrieval, thus providing an indication that they appropriately model the distinctive power of terms to documents and the smoothed, localized estimation of idf values will prove beneficial in sentiment classification. Table 4: Reported accuracies on the Movie Review data set. Only the best reported accuracy for each approach is presented, measured by 10-fold cross validation. The list is not exhaustive and because of differences in training/testing data splits the results are not directly comparable. It is produced here only for reference. Approach Acc. SVM with unigrams & binary weights (Pang et al., 2002), reported at (Pang and Lee, 2004) 87.15% Hybrid SVM with Turney/Osgood Lemmas (Mullen and Collier, 2004) 86% SVM with min-cuts (Pang and Lee, 2004) 87.2% SVM with appraisal groups 90.2% (Whitelaw et al., 2005) SVM with log likehood ratio feature selection (Aue and Gamon, 2005) 90.45% SVM with annotator rationales 92.2% (Zaidan et al., 2007) LDA with filtered lexicon, subjectivity detection (Lin and He, 2009) 84.6% The approach is straightforward, intuitive, computationally efficient, doesn’t require additional human effort and takes into consideration standardized and tested notions from IR. The results presented in section 5 show that a number 1390 of weighting functions solidly outperform other state-of-the-art approaches. In the next section, we present the corpora that were used to study the effectiveness of different weighting schemes. 4 Experimental setup We have experimented with a number of publicly available data sets. The movie review dataset by Pang et al. (2002) has been used extensively in the past by a number of researchers (see Table 4), presenting the opportunity to compare the produced results with previous approaches. The dataset comprises 2,000 movie reviews, equally divided between positive and negative, extracted from the Internet Movie Database3 archive of the rec.arts.movies.reviews newsgroup. In order to avoid reviewer bias, only 20 reviews per author were kept, resulting in a total of 312 reviewers4. The best attained accuracies by previous research on the specific data are presented in table 4. We do not claim that those results are directly comparable to ours, because of potential subtle differences in tokenization, classifier implementations etc, but we present them here for reference. The Multi-Domain Sentiment data set (MDSD) by Blitzer et al. (2007) contains Amazon reviews for four different product types: books, electronics, DVDs and kitchen appliances. Reviews with ratings of 3 or higher, on a 5-scale system, were labeled as positive and reviews with a rating less than 3 as negative. The data set contains 1,000 positive and 1,000 negative reviews for each product category for a total of 8,000 reviews. Typically, the data set is used for domain adaptation applications but in our setting we only split the reviews between positive and negative5. Lastly, we present results from the BLOGS06 (Macdonald and Ounis, 2006) collection that is comprised of an uncompressed 148GB crawl of approximately 100,000 blogs and their respective RSS feeds. The collection has been used for 3 consecutive years by the Text REtrieval Conferences (TREC)6. Participants of the conference are provided with the task of finding documents (i.e. web pages) expressing an opinion about specific enti3http://www.imdb.com 4The dataset can be found at: http://www.cs.cornell.edu/ People/pabo/movie-review-data/review polarity.tar.gz. 5The data set can be found at http://www.cs.jhu.edu/ mdredze/datasets/sentiment/ 6http://www.trec.nist.gov ties X, which may be people, companies, films etc. The results are given to human assessors who then judge the content of the webpages (i.e. blog post and comments) and assign each webpage a score: “1” if the document contains relevant, factual information about the entity but no expression of opinion, “2” if the document contains an explicit negative opinion towards the entity and “4” is the document contains an explicit positive opinion towards the entity. We used the produced assessments from all 3 years of the conference in our data set, resulting in 150 different entity searches and, after duplicate removal, 7,930 negative documents (i.e. having an assessment of “2”) and 9,968 positive documents (i.e. having an assessment of “4”), which were used as the “gold standard” 7. Documents are annotated at the document-level, rather than at the post level, making this data set somewhat noisy. Additionally, the data set is particularly large compared to the other ones, making classification especially challenging and interesting. More information about all data sets can be found at table 5. We have kept the pre-processing of the documents to a minimum. Thus, we have lower-cased all words and removed all punctuation but we have not removed stop words or applied stemming. We have also refrained from removing words with low or high occurrence. Additionally, for the BLOGS06 data set, we have removed all html formatting. We utilize the implementation of a support vector classifier from the LIBLINEAR library (Fan et al., 2008). We use a linear kernel and default parameters. All results are based on leave-one out cross validation accuracy. The reason for this choice of cross-validation setting, instead of the most standard ten-fold, is that all of the proposed approaches that use some form of idf utilize the training documents for extracting document frequency statistics, therefore more information is available to them in this experimental setting. Because of the high number of possible combinations between tf and idf variants (6·9·2 = 108) and due to space constraints we only present results from a subset of the most representative combinations. Generally, we’ll use the cosine normalized variants of unsmoothed delta weighting schemes, since they perform better than their un7More information about the data set, as well as information on how it can be obtained can be found at: http://ir.dcs.gla.ac.uk/test collections/blogs06info.html 1391 Table 5: Statistics about the data sets used. Data set #Documents #Terms #Unique Terms Average #Terms per Document Movie Reviews 2,000 1,336,883 39,399 668 Multi-Domain Sentiment Dataset (MDSD) 8,000 1,741,085 455,943 217 BLOGS06 17,898 51,252,850 367,899 2,832 Figure 1: Reported accuracy on the Movie Review data set. normalized counterparts. We’ll avoid using normalization for the smoothed versions, in order to focus our attention on the results of smoothing, rather than normalization. 5 Results Results for the Movie Reviews, Multi-Domain Sentiment Dataset and BLOGS06 corpora are reported in figures 1, 2 and 3 respectively. On the Movie Review data set, the results reconfirm that using binary features (bnc) is better than raw term frequency (nnc) (83.40%) features. For reference, in this setting the unnormalized vector using the raw tf approach (nnn) performs similar to the normalized (nnc) (83.40% vs. 83.60%), the former not present in the graph. Nonetheless, using any scaled tf weighting function (anc or onc) performs as well as the binary approach (87.90% and 87.50% respectively). Of interest is the fact that although the BM25 tf algorithm has proved much more successful in IR, the same doesn’t apply in this setting and its accuracy is similar to the simpler augmented tf approach. Incorporating un-localized variants of idf (middle graph section) produces only small increases in accuracy. Smoothing also doesn’t provide any particular advantage, e.g. btc (88.20%) vs. bt′c (88.45%), since no zero idf values are present. Again, using more sophisticated tf functions provides an advantage over raw tf, e.g. nt′c attains an accuracy of 86.6% in comparison to at′c’s 88.25%, although the simpler at′c is again as effective than the BM25 tf (ot′c), which performs at 88%. The actual idf weighting function is of some importance, e.g. ot′c (88%) vs. okc (87.65%) and akc (88%) vs. at′c (88.25%), with simpler idf factors performing similarly, although slightly better than BM25. Introducing smoothed, localized variants of idf and scaled or binary tf weighting schemes produces significant advantages. In this setting, smoothing plays a role, e.g. n∆(t)c8 (91.60%) vs. n∆(t′)n (95.80%) and a∆(p)c (92.80%) vs. a∆(p′)n (96.55%), since we can expect zero class-based estimations of idf values, supporting our initial hypothesis on its importance. Additionally, using augmented, BM25 or binary tf weights is always better than raw term frequency, providing further support on the advantages of using sublinear tf weighting functions9. In this setting, the best accuracy of 96.90% is attained using BM25 tf weights with the BM25 delta idf variant, although binary or augmented tf weights using 8The original Delta tf.idf by Martineau and Finin (2009) has a limitation of utilizing features with df > 2. In our experiments it performed similarly to n∆(t)n (90.60%) but still lower than the cosine normalized variant n∆(t)c included in the graph (91.60%). 9Although not present in the graph, for completeness reasons it should be noted that l∆(s)n and L∆(s)n also perform very well, both reaching accuracies of approx. 96%. 1392 Figure 2: Reported accuracy on the Multi-Domain Sentiment data set. delta idf perform similarly (96.50% and 96.60% respectively). The results indicate that the tf and the idf factor themselves aren’t of significant importance, as long as the former are scaled and the latter smoothed in some manner. For example, a∆(p′)n vs. a∆(t′)n perform quite similarly. The results from the Multi-Domain Sentiment data set (figure 2) largely agree with the findings on the Movie Review data set, providing a strong indication that the approach isn’t limited to a specific domain. Binary weights outperform raw term frequency weights and perform similarly with scaled tf’s. Non-localized variants of idf weights do provide a small advantage in this data set although the actual idf variant isn’t important, e.g. btc, bt′c, and okc all perform similarly. The utilized tf variant also isn’t important, e.g. at′c (88.39%) vs. bt′c (88.25%). We focus our attention on the delta idf variants which provide the more interesting results. The importance of smoothing becomes apparent when comparing the accuracy of a∆(p)c and its smoothed variant a∆(p′)n (92.56% vs. 95.6%). Apart from that, all smoothed delta idf variants perform very well in this data set, including somewhat surprisingly, n∆(t′)n which uses raw tf (94.54%). Considering that the average tf per document is approx. 1.9 in the Movie Review data set and 1.1 in the MDSD, the results can be attributed to the fact that words tend to typically appear only once per document in the latter, therefore minimizing the difference of the weights attributed by different tf functions10. The best attained accuracy is 96.40% but as the MDSD has mainly been used for domain adaptation applications, there is no clear baseline to compare it with. 10For reference, the average tf per document in the BLOGS06 data set is 2.4. Lastly, we present results on the BLOGS06 dataset in figure 3. As previously noted, this data set is particularly noisy, because it has been annotated at the document-level rather than the postlevel and as a result, the differences aren’t as profound as in the previous corpora, although they do follow the same patterns. Focusing on the delta idf variants, the importance of smoothing becomes apparent, e.g. a∆(p)c vs. a∆(p′)n and n∆(t)c vs. n∆(t′)n. Additionally, because of the fact that documents tend to be more verbose in this data set, the scaled tf variants also perform better than the simple raw tf ones, n∆(t′)n vs. a∆(t′)n. Lastly, as previously, the smoothed localized idf variants perform better than their unsmoothed counterparts, e.g. n∆(t)n vs. n∆(t′)n and a∆(p)c vs. a∆(p′)n. 6 Conclusions In this paper, we presented a study of document representations for sentiment analysis using term weighting functions adopted from information retrieval and adapted to classification. The proposed weighting schemes were tested on a number of publicly available datasets and a number of them repeatedly demonstrated significant increases in accuracy compared to other state-of-theart approaches. We demonstrated that for accurate classification it is important to use term weighting functions that scale sublinearly in relation to the number of times a term occurs in a document and that document frequency smoothing is a significant factor. In the future we plan to test the proposed weighting functions in other domains such as topic classification and additionally extend the approach to accommodate multi-class classification. 1393 Figure 3: Reported accuracy on the BLOGS06 data set. Acknowledgments This work was supported by a European Union grant by the 7th Framework Programme, Theme 3: Science of complex systems for socially intelligent ICT. It is part of the CyberEmotions Project (Contract 231323). References Ahmed Abbasi, Hsinchun Chen, and Arab Salem. 2008. Sentiment analysis in multiple languages: Feature selection for opinion classification in web forums. ACM Trans. Inf. Syst., 26(3):1–34. Timothy G. Armstrong, Alistair Moffat, William Webber, and Justin Zobel. 2009. Improvements that don’t add up: ad-hoc retrieval results since 1998. In David Wai Lok Cheung, Il Y. Song, Wesley W. Chu, Xiaohua Hu, Jimmy J. Lin, David Wai Lok Cheung, Il Y. Song, Wesley W. Chu, Xiaohua Hu, and Jimmy J. Lin, editors, CIKM, pages 601–610, New York, NY, USA. ACM. Anthony Aue and Michael Gamon. 2005. Customizing sentiment classifiers to new domains: A case study. In Proceedings of Recent Advances in Natural Language Processing (RANLP). John Blitzer, Mark Dredze, and Fernando Pereira. 2007. Biographies, bollywood, boom-boxes and blenders: Domain adaptation for sentiment classification. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 440–447, Prague, Czech Republic, June. Association for Computational Linguistics. Ann Devitt and Khurshid Ahmad. 2007. Sentiment polarity identification in financial news: A cohesionbased approach. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 984–991, Prague, Czech Republic, June. Association for Computational Linguistics. Rong-En Fan, Kai-Wei Chang, Cho-Jui Hsieh, XiangRui Wang, and Chih-Jen Lin. 2008. LIBLINEAR: A library for large linear classification. Journal of Machine Learning Research, 9:1871–1874. Stephan Greene and Philip Resnik. 2009. More than words: Syntactic packaging and implicit sentiment. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 503–511, Boulder, Colorado, June. Association for Computational Linguistics. K. Sparck Jones, S. Walker, and S. E. Robertson. 2000. A probabilistic model of information retrieval: development and comparative experiments. Inf. Process. Manage., 36(6):779–808. Chenghua Lin and Yulan He. 2009. Joint sentiment/topic model for sentiment analysis. In CIKM ’09: Proceeding of the 18th ACM conference on Information and knowledge management, pages 375– 384, New York, NY, USA. ACM. Wei-Hao Lin, Theresa Wilson, Janyce Wiebe, and Alexander Hauptmann. 2006. Which side are you on? identifying perspectives at the document and sentence levels. In Proceedings of the Conference on Natural Language Learning (CoNLL). Hugo Liu. 2004. MontyLingua: An end-to-end natural language processor with common sense. Technical report, MIT. C. Macdonald and I. Ounis. 2006. The trec blogs06 collection : Creating and analysing a blog test collection. DCS Technical Report Series. Christopher D. Manning, Prabhakar Raghavan, and Hinrich Sch¨utze. 2008. Introduction to Information Retrieval. Cambridge University Press, 1 edition, July. J. R. Martin and P. R. R. White. 2005. The language of evaluation : appraisal in English / J.R. Martin and P.R.R. White. Palgrave Macmillan, Basingstoke :. Justin Martineau and Tim Finin. 2009. Delta TFIDF: An Improved Feature Space for Sentiment Analysis. In Proceedings of the Third AAAI Internatonal Conference on Weblogs and Social Media, San Jose, CA, May. AAAI Press. (poster paper). A. Mccallum and K. Nigam. 1998. A comparison of event models for naive bayes text classification. 1394 G. Mishne. 2005. Experiments with mood classification in blog posts. In 1st Workshop on Stylistic Analysis Of Text For Information Access. Tony Mullen and Nigel Collier. 2004. Sentiment analysis using support vector machines with diverse information sources. In Dekang Lin and Dekai Wu, editors, Proceedings of EMNLP 2004, pages 412– 418, Barcelona, Spain, July. Association for Computational Linguistics. Charles E. Osgood. 1967. The measurement of meaning / [by] [Charles E. Osgood, George J. Suci [and] Percy H. Tannenbaum]. University of Illinois Press, Urbana :, 2nd ed. edition. Iadh Ounis, Craig Macdonald, and Ian Soboroff. 2008. Overview of the trec-2008 blog trac. In The Seventeenth Text REtrieval Conference (TREC 2008) Proceedings. NIST. Bo Pang and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In In Proceedings of the ACL, pages 271–278. B. Pang and L. Lee. 2008. Opinion Mining and Sentiment Analysis. Now Publishers Inc. Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. 2002. Thumbs up? sentiment classification using machine learning techniques. In Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing (EMNLP). Rudy Prabowo and Mike Thelwall. 2009. Sentiment analysis: A combined approach. Journal of Informetrics, 3(2):143–157, April. Stephen E. Robertson, Steve Walker, Susan Jones, Micheline Hancock-Beaulieu, and Mike Gatford. 1994. Okapi at trec-3. In TREC, pages 0–. S E Robertson, S Walker, S Jones, M M HancockBeaulieu, and M Gatford. 1996. Okapi at trec-2. In In The Second Text REtrieval Conference (TREC2), NIST Special Special Publication 500-215, pages 21–34. Stephen Robertson, Hugo Zaragoza, and Michael Taylor. 2004. Simple bm25 extension to multiple weighted fields. In CIKM ’04: Proceedings of the thirteenth ACM international conference on Information and knowledge management, pages 42–49, New York, NY, USA. ACM. Gerard Salton and Chris Buckley. 1987. Term weighting approaches in automatic text retrieval. Technical report, Ithaca, NY, USA. Gerard Salton and Michael J. McGill. 1986. Introduction to Modern Information Retrieval. McGrawHill, Inc., New York, NY, USA. G. Salton. 1971. The SMART Retrieval System— Experiments in Automatic Document Processing. Prentice-Hall, Inc., Upper Saddle River, NJ, USA. Fabrizio Sebastiani. 2002. Machine learning in automated text categorization. ACM Computing Surveys, 34(1):1˜n47. Amit Singhal, Gerard Salton, and Chris Buckley. 1995. Length normalization in degraded text collections. Technical report, Ithaca, NY, USA. Matt Thomas, Bo Pang, and Lillian Lee. 2006. Get out the vote: Determining support or opposition from congressional floor-debate transcripts. CoRR, abs/cs/0607062. Peter D. Turney. 2002. Thumbs up or thumbs down? semantic orientation applied to unsupervised classification of reviews. In ACL, pages 417–424. Casey Whitelaw, Navendu Garg, and Shlomo Argamon. 2005. Using appraisal groups for sentiment analysis. In CIKM ’05: Proceedings of the 14th ACM international conference on Information and knowledge management, pages 625–631, New York, NY, USA. ACM. Theresa Wilson, Janyce Wiebe, and Paul Hoffmann. 2005. Recognizing contextual polarity in phraselevel sentiment analysis. In Proceedings of Human Language Technologies Conference/Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP 2005), Vancouver, CA. Ian H. Witten and Eibe Frank. 1999. Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations (The Morgan Kaufmann Series in Data Management Systems). Morgan Kaufmann, 1st edition, October. Alex Wright. 2009. Mining the web for feelings, not facts. August 23, NY Times, last accessed October 2, 2009, http://http://www.nytimes.com/2009/08/24/ technology/internet/ 24emotion.html? r=1. O.F. Zaidan, J. Eisner, and C.D. Piatko. 2007. Using Annotator Rationales to Improve Machine Learning for Text Categorization. Proceedings of NAACL HLT, pages 260–267. Justin Zobel and Alistair Moffat. 1998. Exploring the similarity space. SIGIR Forum, 32(1):18–34. 1395
|
2010
|
141
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1396–1411, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Supervised Noun Phrase Coreference Research: The First Fifteen Years Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 [email protected] Abstract The research focus of computational coreference resolution has exhibited a shift from heuristic approaches to machine learning approaches in the past decade. This paper surveys the major milestones in supervised coreference research since its inception fifteen years ago. 1 Introduction Noun phrase (NP) coreference resolution, the task of determining which NPs in a text or dialogue refer to the same real-world entity, has been at the core of natural language processing (NLP) since the 1960s. NP coreference is related to the task of anaphora resolution, whose goal is to identify an antecedent for an anaphoric NP (i.e., an NP that depends on another NP, specifically its antecedent, for its interpretation) [see van Deemter and Kibble (2000) for a detailed discussion of the difference between the two tasks]. Despite its simple task definition, coreference is generally considered a difficult NLP task, typically involving the use of sophisticated knowledge sources and inference procedures (Charniak, 1972). Computational theories of discourse, in particular focusing (see Grosz (1977) and Sidner (1979)) and centering (Grosz et al. (1983; 1995)), have heavily influenced coreference research in the 1970s and 1980s, leading to the development of numerous centering algorithms (see Walker et al. (1998)). The focus of coreference research underwent a gradual shift from heuristic approaches to machine learning approaches in the 1990s. This shift can be attributed in part to the advent of the statistical NLP era, and in part to the public availability of annotated coreference corpora produced as part of the MUC-6 (1995) and MUC-7 (1998) conferences. Learning-based coreference research has remained vibrant since then, with results regularly published not only in general NLP conferences, but also in specialized conferences (e.g., the biennial Discourse Anaphora and Anaphor Resolution Colloquium (DAARC)) and workshops (e.g., the series of Bergen Workshop on Anaphora Resolution (WAR)). Being inherently a clustering task, coreference has also received a lot of attention in the machine learning community. Fifteen years have passed since the first paper on learning-based coreference resolution was published (Connolly et al., 1994). Our goal in this paper is to provide NLP researchers with a survey of the major milestones in supervised coreference research, focusing on the computational models, the linguistic features, the annotated corpora, and the evaluation metrics that were developed in the past fifteen years. Note that several leading coreference researchers have published books (e.g., Mitkov (2002)), written survey articles (e.g., Mitkov (1999), Strube (2009)), and delivered tutorials (e.g., Strube (2002), Ponzetto and Poesio (2009)) that provide a broad overview of coreference research. This survey paper aims to complement, rather than supersede, these previously published materials. In particular, while existing survey papers discuss learning-based coreference research primarily in the context of the influential mention-pair model, we additionally survey recently proposed learning-based coreference models, which attempt to address the weaknesses of the mention-pair model. Due to space limitations, however, we will restrict our discussion to the most commonly investigated kind of coreference relation: the identity relation for NPs, excluding coreference among clauses and bridging references (e.g., part/whole and set/subset relations). 2 Annotated Corpora The widespread popularity of machine learning approaches to coreference resolution can be attributed in part to the public availability of an1396 notated coreference corpora. The MUC-6 and MUC-7 corpora, though relatively small (60 documents each) and homogeneous w.r.t. document type (newswire articles only), have been extensively used for training and evaluating coreference models. Equally popular are the corpora produced by the Automatic Content Extraction (ACE1) evaluations in the past decade: while the earlier ACE corpora (e.g., ACE-2) consist of solely English newswire and broadcast news articles, the later ones (e.g., ACE 2005) have also included Chinese and Arabic documents taken from additional sources such as broadcast conversations, webblog, usenet, and conversational telephone speech. Coreference annotations are also publicly available in treebanks. These include (1) the English Penn Treebank (Marcus et al., 1993), which is labeled with coreference links as part of the OntoNotes project (Hovy et al., 2006); (2) the T¨ubingen Treebank (Telljohann et al., 2004), which is a collection of German news articles consisting of 27,125 sentences; (3) the Prague Dependency Treebank (Haji˘c et al., 2006), which consists of 3168 news articles taken from the Czech National Corpus; (4) the NAIST Text Corpus (Iida et al., 2007b), which consists of 287 Japanese news articles; (5) the AnCora Corpus (Recasens and Mart´ı, 2009), which consists of Spanish and Catalan journalist texts; and (6) the GENIA corpus (Ohta et al., 2002), which contains 2000 MEDLINE abstracts. Other publicly available coreference corpora of interest include two annotated by Ruslan Mitkov’s research group: (1) a 55,000-word corpus in the domain of security/terrorism (Hasler et al., 2006); and (2) training data released as part of the 2007 Anaphora Resolution Exercise (Or˘asan et al., 2008), a coreference resolution shared task. There are also two that consist of spoken dialogues: the TRAINS93 corpus (Heeman and Allen, 1995) and the Switchboard data set (Calhoun et al., in press). Additional coreference data will be available in the near future. For instance, the SemEval-2010 shared task on Coreference Resolution in Multiple Languages (Recasens et al., 2009) has promised to release coreference data in six languages. In addition, Massimo Poesio and his colleagues are leading an annotation project that aims to collect large amounts of coreference data for English via a Web Collaboration game called Phrase Detectives2. 1http://www.itl.nist.gov/iad/mig/tests/ace/ 2http://www.phrasedetectives.org 3 Learning-Based Coreference Models In this section, we examine three important classes of coreference models that were developed in the past fifteen years, namely, the mention-pair model, the entity-mention model, and ranking models. 3.1 Mention-Pair Model The mention-pair model is a classifier that determines whether two NPs are coreferent. It was first proposed by Aone and Bennett (1995) and McCarthy and Lehnert (1995), and is one of the most influential learning-based coreference models. Despite its popularity, this binary classification approach to coreference is somewhat undesirable: the transitivity property inherent in the coreference relation cannot be enforced, as it is possible for the model to determine that A and B are coreferent, B and C are coreferent, but A and C are not coreferent. Hence, a separate clustering mechanism is needed to coordinate the pairwise classification decisions made by the model and construct a coreference partition. Another issue that surrounds the acquisition of the mention-pair model concerns the way training instances are created. Specifically, to determine whether a pair of NPs is coreferent or not, the mention-pair model needs to be trained on a data set where each instance represents two NPs and possesses a class value that indicates whether the two NPs are coreferent. Hence, a natural way to assemble a training set is to create one instance from each pair of NPs appearing in a training document. However, this instance creation method is rarely employed: as most NP pairs in a text are not coreferent, this method yields a training set with a skewed class distribution, where the negative instances significantly outnumber the positives. As a result, in practical implementations of the mention-pair model, one needs to specify not only the learning algorithm for training the model and the linguistic features for representing an instance, but also the training instance creation method for reducing class skewness and the clustering algorithm for constructing a coreference partition. 3.1.1 Creating Training Instances As noted above, the primary purpose of training instance creation is to reduce class skewness. Many heuristic instance creation methods have been proposed, among which Soon et al.’s (1999; 2001) is arguably the most popular choice. Given 1397 an anaphoric noun phrase3, NPk, Soon et al.’s method creates a positive instance between NPk and its closest preceding antecedent, NPj, and a negative instance by pairing NPk with each of the intervening NPs, NPj+1, . . ., NPk−1. With an eye towards improving the precision of a coreference resolver, Ng and Cardie (2002c) propose an instance creation method that involves a single modification to Soon et al.’s method: if NPk is non-pronominal, a positive instance should be formed between NPk and its closest preceding nonpronominal antecedent instead. This modification is motivated by the observation that it is not easy for a human, let alone a machine learner, to learn from a positive instance where the antecedent of a non-pronominal NP is a pronoun. To further reduce class skewness, some researchers employ a filtering mechanism on top of an instance creation method, thereby disallowing the creation of training instances from NP pairs that are unlikely to be coreferent, such as NP pairs that violate gender and number agreement (e.g., Strube et al. (2002), Yang et al. (2003)). While many instance creation methods are heuristic in nature (see Uryupina (2004) and Hoste and Daelemans (2005)), some are learning-based. For example, motivated by the fact that some coreference relations are harder to identify than the others (see Harabagiu et al. (2001)), Ng and Cardie (2002a) present a method for mining easy positive instances, in an attempt to avoid the inclusion of hard training instances that may complicate the acquisition of an accurate coreference model. 3.1.2 Training a Coreference Classifier Once a training set is created, we can train a coreference model using an off-the-shelf learning algorithm. Decision tree induction systems (e.g., C5 (Quinlan, 1993)) are the first and one of the most widely used learning algorithms by coreference researchers, although rule learners (e.g., RIPPER (Cohen, 1995)) and memory-based learners (e.g., TiMBL (Daelemans and Van den Bosch, 2005)) are also popular choices, especially in early applications of machine learning to coreference resolution. In recent years, statistical learners such as maximum entropy models (Berger et al., 1996), voted perceptrons (Freund and Schapire, 1999), 3In this paper, we use the term anaphoric to describe any NP that is part of a coreference chain but is not the head of the chain. Hence, proper names can be anaphoric under this overloaded definition, but linguistically, they are not. and support vector machines (Joachims, 1999) have been increasingly used, in part due to their ability to provide a confidence value (e.g., in the form of a probability) associated with a classification, and in part due to the fact that they can be easily adapted to train recently proposed rankingbased coreference models (see Section 3.3). 3.1.3 Generating an NP Partition After training, we can apply the resulting model to a test text, using a clustering algorithm to coordinate the pairwise classification decisions and impose an NP partition. Below we describe some commonly used coreference clustering algorithms. Despite their simplicity, closest-first clustering (Soon et al., 2001) and best-first clustering (Ng and Cardie, 2002c) are arguably the most widely used coreference clustering algorithms. The closest-first clustering algorithm selects as the antecedent for an NP, NPk, the closest preceding noun phrase that is classified as coreferent with it.4 However, if no such preceding noun phrase exists, no antecedent is selected for NPk. The best-first clustering algorithm aims to improve the precision of closest-first clustering, specifically by selecting as the antecedent of NPk the most probable preceding NP that is classified as coreferent with it. One criticism of the closest-first and best-first clustering algorithms is that they are too greedy. In particular, clusters are formed based on a small subset of the pairwise decisions made by the model. Moreover, positive pairwise decisions are unjustifiably favored over their negative counterparts. For example, three NPs are likely to end up in the same cluster in the resulting partition even if there is strong evidence that A and C are not coreferent, as long as the other two pairs (i.e., (A,B) and (B,C)) are classified as positive. Several algorithms that address one or both of these problems have been used for coreference clustering. Correlation clustering (Bansal et al., 2002), which produces a partition that respects as many pairwise decisions as possible, is used by McCallum and Wellner (2004), Zelenko et al. (2004), and Finley and Joachims (2005). Graph partitioning algorithms are applied on a weighted, undirected graph where a vertex corresponds to an NP and an edge is weighted by the pairwise coreference scores between two NPs (e.g., McCallum and Wellner (2004), Nicolae and Nico4If a probabilistic model is used, we can define a threshold above which a pair of NPs is considered coreferent. 1398 lae (2006)). The Dempster-Shafer rule (Dempster, 1968), which combines the positive and negative pairwise decisions to score a partition, is used by Kehler (1997) and Bean and Riloff (2004) to identify the most probable NP partition. Some clustering algorithms bear a closer resemblance to the way a human creates coreference clusters. In these algorithms, not only are the NPs in a text processed in a left-to-right manner, the later coreference decisions are dependent on the earlier ones (Cardie and Wagstaff, 1999; Klenner and Ailloud, 2008).5 For example, to resolve an NP, NPk, Cardie and Wagstaff’s algorithm considers each preceding NP, NPj, as a candidate antecedent in a right-to-left order. If NPk and NPj are likely to be coreferent, the algorithm imposes an additional check that NPk does not violate any constraint on coreference (e.g., gender agreement) with any NP in the cluster containing NPj before positing that the two NPs are coreferent. Luo et al.’s (2004) Bell-tree-based algorithm is another clustering algorithm where the later coreference decisions are dependent on the earlier ones. A Bell tree provides an elegant way of organizing the space of NP partitions. Informally, a node in the ith level of a Bell tree corresponds to an ithorder partial partition (i.e., a partition of the first i NPs of the given document), and the ith level of the tree contains all possible ith-order partial partitions. Hence, a leaf node contains a complete partition of the NPs, and the goal is to search for the leaf node that contains the most probable partition. The search starts at the root, and a partitioning of the NPs is incrementally constructed as we move down the tree. Specifically, based on the coreference decisions it has made in the first i−1 levels of the tree, the algorithm determines at the ith level whether the ith NP should start a new cluster, or to which preceding cluster it should be assigned. While many coreference clustering algorithms have been developed, there have only been a few attempts to compare their effectiveness. For example, Ng and Cardie (2002c) report that bestfirst clustering is better than closest-first clustering. Nicolae and Nicolae (2006) show that bestfirst clustering performs similarly to Bell-treebased clustering, but neither of these algorithms 5When applying closest-first and best-first clustering, Soon et al. (2001) and Ng and Cardie (2002c) also process the NPs in a sequential manner, but since the later decisions are not dependent on the earlier ones, the order in which the NPs are processed does not affect their clustering results. performs as well as their proposed minimum-cutbased graph partitioning algorithm. 3.1.4 Determining NP Anaphoricity While coreference clustering algorithms attempt to resolve each NP encountered in a document, only a subset of the NPs are anaphoric and therefore need to be resolved. Hence, knowledge of the anaphoricity of an NP can potentially improve the precision of a coreference resolver. Traditionally, the task of anaphoricity determination has been tackled independently of coreference resolution using a variety of techniques. For example, pleonastic it has been identified using heuristic approaches (e.g., Paice and Husk (1987), Lappin and Leass (1994), Kennedy and Boguraev (1996)), supervised approaches (e.g., Evans (2001), M¨uller (2006), Versley et al. (2008a)), and distributional methods (e.g., Bergsma et al. (2008)); and non-anaphoric definite descriptions have been identified using rule-based techniques (e.g., Vieira and Poesio (2000)) and unsupervised techniques (e.g., Bean and Riloff (1999)). Recently, anaphoricity determination has been evaluated in the context of coreference resolution, with results showing that training an anaphoricity classifier to identify and filter non-anaphoric NPs prior to coreference resolution can improve a learning-based resolver (e.g., Ng and Cardie (2002b), Uryupina (2003), Poesio et al. (2004b)). Compared to earlier work on anaphoricity determination, recently proposed approaches are more “global” in nature, taking into account the pairwise decisions made by the mention-pair model when making anaphoricity decisions. Examples of such approaches have exploited techniques including integer linear programming (ILP) (Denis and Baldridge, 2007a), label propagation (Zhou and Kong, 2009), and minimum cuts (Ng, 2009). 3.1.5 Combining Classification & Clustering From a learning perspective, a two-step approach to coreference — classification and clustering — is undesirable. Since the classification model is trained independently of the clustering algorithm, improvements in classification accuracy do not guarantee corresponding improvements in clustering-level accuracy. That is, overall performance on the coreference task might not improve. To address this problem, McCallum and Wellner (2004) and Finley and Joachims (2005) eliminate the classification step entirely, treating coref1399 erence as a supervised clustering task where a similarity metric is learned to directly maximize clustering accuracy. Klenner (2007) and Finkel and Manning (2008) use ILP to ensure that the pairwise classification decisions satisfy transitivity.6 3.1.6 Weaknesses of the Mention-Pair Model While many of the aforementioned algorithms for clustering and anaphoricity determination have been shown to improve coreference performance, the underlying model with which they are used in combination — the mention-pair model — remains fundamentally weak. The model has two commonly-cited weaknesses. First, since each candidate antecedent for an anaphoric NP to be resolved is considered independently of the others, the model only determines how good a candidate antecedent is relative to the anaphoric NP, but not how good a candidate antecedent is relative to other candidates. In other words, it fails to answer the question of which candidate antecedent is most probable. Second, it has limitations in its expressiveness: the information extracted from the two NPs alone may not be sufficient for making an informed coreference decision, especially if the candidate antecedent is a pronoun (which is semantically empty) or a mention that lacks descriptive information such as gender (e.g., “Clinton”). Below we discuss how these weaknesses are addressed by the entity-mention model and ranking models. 3.2 Entity-Mention Model The entity-mention model addresses the expressiveness problem with the mention-pair model. To motivate the entity-mention model, consider an example taken from McCallum and Wellner (2003), where a document consists of three NPs: “Mr. Clinton,” “Clinton,” and “she.” The mentionpair model may determine that “Mr. Clinton” and “Clinton” are coreferent using string-matching features, and that “Clinton” and “she” are coreferent based on proximity and lack of evidence for gender and number disagreement. However, these two pairwise decisions together with transitivity imply that “Mr. Clinton” and “she” will end up in the same cluster, which is incorrect due to gender mismatch. This kind of error arises in part because the later coreference decisions are not dependent on the earlier ones. In particular, had the model taken into consideration that “Mr. Clinton” 6Recently, however, Klenner and Ailloud (2009) have become less optimistic about ILP approaches to coreference. and “Clinton” were in the same cluster, it probably would not have posited that “she” and “Clinton” are coreferent. The aforementioned Cardie and Wagstaff algorithm attempts to address this problem in a heuristic manner. It would be desirable to learn a model that can classify whether an NP to be resolved is coreferent with a preceding, possibly partially-formed, cluster. This model is commonly known as the entity-mention model. Since the entity-mention model aims to classify whether an NP is coreferent with a preceding cluster, each of its training instances (1) corresponds to an NP, NPk, and a preceding cluster, Cj, and (2) is labeled with either POSITIVE or NEGATIVE, depending on whether NPk should be assigned to Cj. Consequently, we can represent each instance by a set of cluster-level features (i.e., features that are defined over an arbitrary subset of the NPs in Cj). A cluster-level feature can be computed from a feature employed by the mention-pair model by applying a logical predicate. For example, given the NUMBER AGREEMENT feature, which determines whether two NPs agree in number, we can apply the ALL predicate to create a cluster-level feature, which has the value YES if NPk agrees in number with all of the NPs in Cj and NO otherwise. Other commonly-used logical predicates for creating cluster-level features include relaxed versions of the ALL predicate, such as MOST, which is true if NPk agrees in number with more than half of the NPs in Cj, and ANY, which is true as long as NPk agrees in number with just one of the NPs in Cj. The ability of the entity-mention model to employ cluster-level features makes it more expressive than its mention-pair counterpart. Despite its improved expressiveness, the entitymention model has not yielded particularly encouraging results. For example, Luo et al. (2004) apply the ANY predicate to generate cluster-level features for their entity-mention model, which does not perform as well as the mention-pair model. Yang et al. (2004b; 2008a) also investigate the entity-mention model, which produces results that are only marginally better than those of the mention-pair model. However, it appears that they are not fully exploiting the expressiveness of the entity-mention model, as cluster-level features only comprise a small fraction of their features. Variants of the entity-mention model have been investigated. For example, Culotta et al. (2007) present a first-order logic model that determines 1400 the probability that an arbitrary set of NPs are all co-referring. Their model resembles the entitymention model in that it enables the use of clusterlevel features. Daum´e III and Marcu (2005) propose an online learning model for constructing coreference chains in an incremental fashion, allowing later coreference decisions to be made by exploiting cluster-level features that are computed over the coreference chains created thus far. 3.3 Ranking Models While the entity-mention model addresses the expressiveness problem with the mention-pair model, it does not address the other problem: failure to identify the most probable candidate antecedent. Ranking models, on the other hand, allow us to determine which candidate antecedent is most probable given an NP to be resolved. Ranking is arguably a more natural reformulation of coreference resolution than classification, as a ranker allows all candidate antecedents to be considered simultaneously and therefore directly captures the competition among them. Another desirable consequence is that there exists a natural resolution strategy for a ranking approach: an anaphoric NP is resolved to the candidate antecedent that has the highest rank. This contrasts with classification-based approaches, where many clustering algorithms have been employed to coordinate the pairwise classification decisions, and it is still not clear which of them is the best. The notion of ranking candidate antecedents can be traced back to centering algorithms, many of which use grammatical roles to rank forwardlooking centers (see Walker et al. (1998)). Ranking is first applied to learning-based coreference resolution by Connolly et al. (1994; 1997), where a model is trained to rank two candidate antecedents. Each training instance corresponds to the NP to be resolved, NPk, as well as two candidate antecedents, NPi and NPj, one of which is an antecedent of NPk and the other is not. Its class value indicates which of the two candidates is better. This model is referred to as the tournament model by Iida et al. (2003) and the twin-candidate model by Yang et al. (2003; 2008b). To resolve an NP during testing, one way is to apply the model to each pair of its candidate antecedents, and the candidate that is classified as better the largest number of times is selected as its antecedent. Advances in machine learning have made it possible to train a mention ranker that ranks all of the candidate antecedents simultaneously. While mention rankers have consistently outperformed the mention-pair model (Versley, 2006; Denis and Baldridge, 2007b), they are not more expressive than the mention-pair model, as they are unable to exploit cluster-level features, unlike the entitymention model. To enable rankers to employ cluster-level features, Rahman and Ng (2009) propose the cluster-ranking model, which ranks preceding clusters, rather than candidate antecedents, for an NP to be resolved. Cluster rankers therefore address both weaknesses of the mention-pair model, and have been shown to improve mention rankers. Cluster rankers are conceptually similar to Lappin and Leass’s (1994) heuristic pronoun resolver, which resolves an anaphoric pronoun to the most salient preceding cluster. An important issue with ranking models that we have eluded so far concerns the identification of non-anaphoric NPs. As a ranker simply imposes a ranking on candidate antecedents or preceding clusters, it cannot determine whether an NP is anaphoric (and hence should be resolved). To address this problem, Denis and Baldridge (2008) apply an independently trained anaphoricity classifier to identify non-anaphoric NPs prior to ranking, and Rahman and Ng (2009) propose a model that jointly learns coreference and anaphoricity. 4 Knowledge Sources Another thread of supervised coreference research concerns the development of linguistic features. Below we give an overview of these features. String-matching features can be computed robustly and typically contribute a lot to the performance of a coreference system. Besides simple string-matching operations such as exact string match, substring match, and head noun match for different kinds of NPs (see Daum´e III and Marcu (2005)), slightly more sophisticated stringmatching facilities have been attempted, including minimum edit distance (Strube et al., 2002) and longest common subsequence (Casta˜no et al., 2002). Yang et al. (2004a) treat the two NPs involved as two bags of words, and compute their similarity using metrics commonly-used in information retrieval, such as the dot product, with each word weighted by their TF-IDF value. Syntactic features are computed based on a syntactic parse tree. Ge et al. (1998) implement 1401 a Hobbs distance feature, which encodes the rank assigned to a candidate antecedent for a pronoun by Hobbs’s (1978) seminal syntax-based pronoun resolution algorithm. Luo and Zitouni (2005) extract features from a parse tree for implementing Binding Constraints (Chomsky, 1988). Given an automatically parsed corpus, Bergsma and Lin (2006) extract from each parse tree a dependency path, which is represented as a sequence of nodes and dependency labels connecting a pronoun and a candidate antecedent, and collect statistical information from these paths to determine the likelihood that a pronoun and a candidate antecedent connected by a given path are coreferent. Rather than deriving features from parse trees, Iida et al. (2006) and Yang et al. (2006) employ these trees directly as structured features for pronoun resolution. Specifically, Yang et al. define tree kernels for efficiently computing the similarity between two parse trees, and Iida et al. use a boosting-based algorithm to compute the usefulness of a subtree. Grammatical features encode the grammatical properties of one or both NPs involved in an instance. For example, Ng and Cardie’s (2002c) resolver employs 34 grammatical features. Some features determine NP type (e.g., are both NPs definite or pronouns?). Some determine the grammatical role of one or both of the NPs. Some encode traditional linguistic (hard) constraints on coreference. For example, coreferent NPs have to agree in number and gender and cannot span one another (e.g., “Google” and “Google employees”). There are also features that encode general linguistic preferences either for or against coreference. For example, an indefinite NP (that is not in apposition to an anaphoric NP) is not likely to be coreferent with any NP that precedes it. There has been an increasing amount of work on investigating semantic features for coreference resolution. One of the earliest kinds of semantic knowledge employed for coreference resolution is perhaps selectional preference (Dagan and Itai, 1990; Kehler et al., 2004b; Yang et al., 2005; Haghighi and Klein, 2009): given a pronoun to be resolved, its governing verb, and its grammatical role, we prefer a candidate antecedent that can be governed by the same verb and be in the same role. Semantic knowledge has also been extracted from WordNet and unannotated corpora for computing the semantic compatibility/similarity between two common nouns (Harabagiu et al., 2001; Versley, 2007) as well as the semantic class of a noun (Ng, 2007a; Huang et al., 2009). One difficulty with deriving knowledge from WordNet is that one has to determine which sense of a given word to use. Some researchers simply use the first sense (Soon et al., 2001) or all possible senses (Ponzetto and Strube, 2006a), while others overcome this problem with word sense disambiguation (Nicolae and Nicolae, 2006). Knowledge has also been mined from Wikipedia for measuring the semantic relatedness of two NPs, NPj and NPk (Ponzetto and Strube (2006a; 2007)), such as: whether NPj/k appears in the first paragraph of the Wiki page that has NPk/j as the title or in the list of categories to which this page belongs, and the degree of overlap between the two pages that have the two NPs as their titles (see Poesio et al. (2007) for other uses of encyclopedic knowledge for coreference resolution). Contextual roles (Bean and Riloff, 2004), semantic relations (Ji et al., 2005), semantic roles (Ponzetto and Strube, 2006b; Kong et al., 2009), and animacy (Or˘asan and Evans, 2007) have also been exploited to improve coreference resolution. Lexico-syntactic patterns have been used to capture the semantic relatedness between two NPs and hence the likelihood that they are coreferent. For instance, given the pattern X is a Y (which is highly indicative that X and Y are coreferent), we can instantiate it with a pair of NPs and search for the instantiated pattern in a large corpus or the Web (Daum´e III and Marcu, 2005; Haghighi and Klein, 2009). The more frequently the pattern occurs, the more likely they are coreferent. This technique has been applied to resolve different kinds of anaphoric references, including other-anaphora (Modjeska et al., 2003; Markert and Nissim, 2005) and bridging references (Poesio et al., 2004a). While these patterns are typically hand-crafted (e.g., Garera and Yarowsky (2006)), they can also be learned from an annotated corpus (Yang and Su, 2007) or bootstrapped from an unannotated corpus (Bean and Riloff, 2004). Despite the large amount of work on discoursebased anaphora resolution in the 1970s and 1980s (see Hirst (1981)), learning-based resolvers have only exploited shallow discourse-based features, which primarily involve characterizing the salience of a candidate antecedent by measuring its distance from the anaphoric NP to be resolved or determining whether it is in a prominent grammatical role (e.g., subject). A notable exception 1402 is Iida et al. (2009), who train a ranker to rank the candidate antecedents for an anaphoric pronoun by their salience. It is worth noting that Tetreault (2005) has employed Grosz and Sidner’s (1986) discourse theory and Veins Theory (Ide and Cristea, 2000) to identify and remove candidate antecedents that are not referentially accessible to an anaphoric pronoun in his heuristic pronoun resolvers. It would be interesting to incorporate this idea into a learning-based resolver. There are also features that do not fall into any of the preceding categories. For example, a memorization feature is a word pair composed of the head nouns of the two NPs involved in an instance (Bengtson and Roth, 2008). Memorization features have been used as binary-valued features indicating the presence or absence of their words (Luo et al., 2004) or as probabilistic features indicating the probability that the two heads are coreferent according to the training data (Ng, 2007b). An anaphoricity feature indicates whether an NP to be resolved is anaphoric, and is typically computed using an anaphoricity classifier (Ng, 2004), hand-crafted patterns (Daum´e III and Marcu, 2005), and automatically acquired patterns (Bean and Riloff, 1999). Finally, the outputs of rule-based pronoun and coreference resolvers have also been used as features for learning-based coreference resolution (Ng and Cardie, 2002c). For an empirical evaluation of the contribution of a subset of these features to the mention-pair model, see Bengtson and Roth (2008). 5 Evaluation Issues Two important issues surround the evaluation of a coreference resolver. First, how do we obtain the set of NPs that a resolver will partition? Second, how do we score the partition it produces? 5.1 Extracting Candidate Noun Phrases To obtain the set of NPs to be partitioned by a resolver, three methods are typically used. In the first method, the NPs are extracted automatically from a syntactic parser. The second method involves extracting the NPs directly from the gold standard. In the third method, a mention detector is first trained on the gold-standard NPs in the training texts, and is then applied to automatically extract system mentions in a test text.7 Note that 7An exception is Daum´e III and Marcu (2005), whose model jointly learns to extract NPs and perform coreference. these three extraction methods typically produce different numbers of NPs: the NPs extracted from a parser tend to significantly outnumber the system mentions, which in turn outnumber the gold NPs. The reasons are two-fold. First, in some coreference corpora (e.g., MUC-6 and MUC-7), the NPs that are not part of any coreference chain are not annotated. Second, in corpora such as those produced by the ACE evaluations, only the NPs that belong to one of the ACE entity types (e.g., PERSON, ORGANIZATION, LOCATION) are annotated. Owing in large part to the difference in the number of NPs extracted by these three methods, a coreference resolver can produce substantially different results when applied to the resulting three sets of NPs, with gold NPs yielding the best results and NPs extracted from a parser yielding the worst (Nicolae and Nicolae, 2006). While researchers who evaluate their resolvers on gold NPs point out that the results can more accurately reflect the performance of their coreference algorithm, Stoyanov et al. (2009) argue that such evaluations are unrealistic, as NP extraction is an integral part of an end-to-end fully-automatic resolver. Whichever NP extraction method is employed, it is clear that the use of gold NPs can considerably simplify the coreference task, and hence resolvers employing different extraction methods should not be compared against each other. 5.2 Scoring a Coreference Partition The MUC scorer (Vilain et al., 1995) is the first program developed for scoring coreference partitions. It has two often-cited weaknesses. As a linkbased measure, it does not reward correctly identified singleton clusters since there is no coreference link in these clusters. Also, it tends to underpenalize partitions with overly large clusters. To address these problems, two coreference scoring programs have been developed: B3 (Bagga and Baldwin, 1998) and CEAF (Luo, 2005). Note that both scorers have only been defined for the case where the key partition has the same set of NPs as the response partition. To apply these scorers to automatically extracted NPs, different methods have been proposed (see Rahman and Ng (2009) and Stoyanov et al. (2009)). Since coreference is a clustering task, any general-purpose method for evaluating a response partition against a key partition (e.g., Kappa (Carletta, 1996)) can be used for coreference scor1403 ing (see Popescu-Belis et al. (2004)). In practice, these general-purpose methods are typically used to provide scores that complement those obtained via the three coreference scorers discussed above. It is worth mentioning that there is a trend towards evaluating a resolver against multiple scorers, which can indirectly help to counteract the bias inherent in a particular scorer. For further discussion on evaluation issues, see Byron (2001). 6 Concluding Remarks While we have focused our discussion on supervised approaches, coreference researchers have also attempted to reduce a resolver’s reliance on annotated data by combining a small amount of labeled data and a large amount of unlabeled data using general-purpose semi-supervised learning algorithms such as co-training (M¨uller et al., 2002), self-training (Kehler et al., 2004a), and EM (Cherry and Bergsma, 2005; Ng, 2008). Interestingly, recent results indicate that unsupervised approaches to coreference resolution (e.g., Haghighi and Klein (2007; 2010), Poon and Domingos (2008)) rival their supervised counterparts, casting doubts on whether supervised resolvers are making effective use of the available labeled data. Another issue that we have not focused on but which is becoming increasingly important is multilinguality. While many of the techniques discussed in this paper were originally developed for English, they have been applied to learn coreference models for other languages, such as Chinese (e.g., Converse (2006)), Japanese (e.g., Iida (2007)), Arabic (e.g., Luo and Zitouni (2005)), Dutch (e.g., Hoste (2005)), German (e.g., Wunsch (2010)), Swedish (e.g., Nilsson (2010)), and Czech (e.g., Ngu.y et al. (2009)). In addition, researchers have developed approaches that are targeted at handling certain kinds of anaphora present in non-English languages, such as zero anaphora (e.g., Iida et al. (2007a), Zhao and Ng (2007)). As Mitkov (2001) puts it, coreference resolution is a “difficult, but not intractable problem,” and we have been making “slow, but steady progress” on improving machine learning approaches to the problem in the past fifteen years. To ensure further progress, researchers should compare their results against a baseline that is stronger than the commonly-used Soon et al. (2001) system, which relies on a weak model (i.e., the mention-pair model) and a small set of linguistic features. As recent systems are becoming more sophisticated, we suggest that researchers make their systems publicly available in order to facilitate performance comparisons. Publicly available coreference systems currently include JavaRAP (Qiu et al., 2004), GuiTaR (Poesio and Kabadjov, 2004), BART (Versley et al., 2008b), CoRTex (Denis and Baldridge, 2008), the Illinois Coreference Package (Bengtson and Roth, 2008), CherryPicker (Rahman and Ng, 2009), Reconcile (Stoyanov et al., 2010), and Charniak and Elsner’s (2009) pronoun resolver. We conclude with a discussion of two questions regarding supervised coreference research. First, what is the state of the art? This is not an easy question, as researchers have been evaluating their resolvers on different corpora using different evaluation metrics and preprocessing tools. In particular, preprocessing tools can have a large impact on the performance of a resolver (Barbu and Mitkov, 2001). Worse still, assumptions about whether gold or automatically extracted NPs are used are sometimes not explicitly stated, potentially causing results to be interpreted incorrectly. To our knowledge, however, the best results on the MUC-6 and MUC-7 data sets using automatically extracted NPs are reported by Yang et al. (2003) (71.3 MUC F-score) and Ng and Cardie (2002c) (63.4 MUC F-score), respectively;8 and the best results on the ACE data sets using gold NPs can be found in Luo (2007) (88.4 ACE-value). Second, what lessons can we learn from fifteen years of learning-based coreference research? The mention-pair model is weak because it makes coreference decisions based on local information (i.e., information extracted from two NPs). Expressive models (e.g., those that can exploit cluster-level features) generally offer better performance, and so are models that are “global” in nature. Global coreference models may refer to any kind of models that can exploit non-local information, including models that can consider multiple candidate antecedents simultaneously (e.g., ranking models), models that allow joint learning for coreference resolution and related tasks (e.g., anaphoricity determination), models that can directly optimize clustering-level (rather than classification) accuracy, and models that can coordinate with other components of a resolver, such as training instance creation and clustering. 8These results by no means suggest that no progress has been made since 2003: most of the recently proposed coreference models were evaluated on the ACE data sets. 1404 Acknowledgments We thank the three anonymous reviewers for their invaluable comments on an earlier draft of the paper. This work was supported in part by NSF Grant IIS-0812261. Any opinions, findings, and conclusions or recommendations expressed are those of the author and do not necessarily reflect the views or official policies, either expressed or implied, of the NSF. References Chinatsu Aone and Scott William Bennett. 1995. Evaluating automated and manual acquisition of anaphora resolution strategies. In Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics, pages 122–129. Amit Bagga and Breck Baldwin. 1998. Algorithms for scoring coreference chains. In Proceedings of the LREC Workshop on Linguistic Coreference, pages 563–566. Nikhil Bansal, Avrim Blum, and Shuchi Chawla. 2002. Correlation clustering. In Proceedings of the 43rd Annual IEEE Symposium on Foundations of Computer Science, pages 238–247. Catalina Barbu and Ruslan Mitkov. 2001. Evaluation tool for rule-based anaphora resolution methods. In Proceedings of the 39th Annual Meeting of the Association for Computational Linguistics, pages 34–41. David Bean and Ellen Riloff. 1999. Corpus-based identification of non-anaphoric noun phrases. In Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics, pages 373– 380. David Bean and Ellen Riloff. 2004. Unsupervised learning of contextual role knowledge for coreference resolution. In Human Language Technologies 2004: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 297– 304. Eric Bengtson and Dan Roth. 2008. Understanding the values of features for coreference resolution. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 294– 303. Adam L. Berger, Stephen A. Della Pietra, and Vincent J. Della Pietra. 1996. A maximum entropy approach to natural language processing. Computational Linguistics, 22(1):39–71. Shane Bergsma and Dekang Lin. 2006. Bootstrapping path-based pronoun resolution. In Proceedings of the 21st International Conference on Computational Linguistics and the 44th Annual Meeting of the Association for Computational Linguistics, pages 33–40. Shane Bergsma, Dekang Lin, and Randy Goebel. 2008. Distributional identification of non-referential pronouns. In Proceedings of ACL-08: HLT, pages 10–18. Donna Byron. 2001. The uncommon denominator: A proposal for consistent reporting of pronoun resolution results. Computational Linguistics, 27(4):569– 578. Sasha Calhoun, Jean Carletta, Jason Brenier, Neil Mayo, Dan Jurafsky, Mark Steedman, and David Beaver. (in press). The NXT-format Switchboard corpus: A rich resource for investigating the syntax, semantics, pragmatics and prosody of dialogue. Language Resources and Evaluation. Claire Cardie and Kiri Wagstaff. 1999. Noun phrase coreference as clustering. In Proceedings of the 1999 Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora, pages 82–89. Jean Carletta. 1996. Assessing agreement on classification tasks: the kappa statistic. Computational Linguistics, 22(2):249–254. Jos´e Casta˜no, Jason Zhang, and James Pustejovsky. 2002. Anaphora resolution in biomedical literature. In Proceedings of the 2002 International Symposium on Reference Resolution. Eugene Charniak and Micha Elsner. 2009. EM works for pronoun anaphora resolution. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics, pages 148–156. Eugene Charniak. 1972. Towards a Model of Children’s Story Comphrension. AI-TR 266, Artificial Intelligence Laboratory, Massachusetts Institute of Technology, USA. Colin Cherry and Shane Bergsma. 2005. An expectation maximization approach to pronoun resolution. In Proceedings of the Ninth Conference on Computational Natural Language Learning, pages 88–95. Noam Chomsky. 1988. Language and Problems of Knowledge. The Managua Lectures. MIT Press, Cambridge, Massachusetts. William Cohen. 1995. Fast effective rule induction. In Proceedings of the 12th International Conference on Machine Learning, pages 115–123. Dennis Connolly, John D. Burger, and David S. Day. 1994. A machine learning approach to anaphoric reference. In Proceedings of International Conference on New Methods in Language Processing, pages 255–261. Dennis Connolly, John D. Burger, and David S. Day. 1997. A machine learning approach to anaphoric reference. In D. Jones and H. Somers, editors, New Methods in Language Processing, pages 133–144. UCL Press. 1405 Susan Converse. 2006. Pronominal Anaphora Resolution in Chinese. Ph.D. thesis, University of Pennsylvania, USA. Aron Culotta, Michael Wick, and Andrew McCallum. 2007. First-order probabilistic models for coreference resolution. In Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 81–88. Walter Daelemans and Antal Van den Bosch. 2005. Memory-Based Language Processing. Cambridge University Press, Cambridge, UK. Ido Dagan and Alon Itai. 1990. Automatic processing of large corpora for the resolution of anaphora references. In Proceedings of the 13th International Conference on Computational Linguistics, pages 330–332. Hal Daum´e III and Daniel Marcu. 2005. A largescale exploration of effective global features for a joint entity detection and tracking model. In Proceedings of the Human Language Technology Conference and the Conference on Empirical Methods in Natural Language Processing, pages 97–104. Arthur Dempster. 1968. A generalization of Bayesian inference. Journal of the Royal Statistical Society, 30:205–247. Pascal Denis and Jason Baldridge. 2007a. Global, joint determination of anaphoricity and coreference resolution using integer programming. In Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 236–243. Pascal Denis and Jason Baldridge. 2007b. A ranking approach to pronoun resolution. In Proceedings of the Twentieth International Conference on Artificial Intelligence, pages 1588–1593. Pascal Denis and Jason Baldridge. 2008. Specialized models and ranking for coreference resolution. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 660–669. Richard Evans. 2001. Applying machine learning toward an automatic classification of it. Literary and Linguistic Computing, 16(1):45–57. Jenny Rose Finkel and Christopher Manning. 2008. Enforcing transitivity in coreference resolution. In Proceedings of ACL-08: HLT, Short Papers, pages 45–48. Thomas Finley and Thorsten Joachims. 2005. Supervised clustering with support vector machines. In Proceedings of the 22nd International Conference on Machine Learning, pages 217–224. Yoav Freund and Robert E. Schapire. 1999. Large margin classification using the perceptron algorithm. Machine Learning, 37(3):277–296. Nikesh Garera and David Yarowsky. 2006. Resolving and generating definite anaphora by modeling hypernymy using unlabeled corpora. In Proceedings of the Tenth Conference on Computational Natural Language Learning, pages 37–44. Niyu Ge, John Hale, and Eugene Charniak. 1998. A statistical approach to anaphora resolution. In Proceedings of the Sixth Workshop on Very Large Corpora, pages 161–170. Barbara J. Grosz and Candace L. Sidner. 1986. Attention, intentions, and the structure of discourse. Computational Linguistics, 12(3):175–204. Barbara J. Grosz, Aravind K. Joshi, and Scott Weinstein. 1983. Providing a unified account of definite noun phrases in discourse. In Proceedings of the 21st Annual Meeting of the Association for Computational Linguistics, pages 44–50. Barbara J. Grosz, Aravind K. Joshi, and Scott Weinstein. 1995. Centering: A framework for modeling the local coherence of discourse. Computational Linguistics, 21(2):203–226. Barbara J. Grosz. 1977. The representation and use of focus in a system for understanding dialogs. In Proceedings of the Fifth International Joint Conference on Artificial Intelligence, pages 67–76. Aria Haghighi and Dan Klein. 2007. Unsupervised coreference resolution in a nonparametric bayesian model. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 848–855. Aria Haghighi and Dan Klein. 2009. Simple coreference resolution with rich syntactic and semantic features. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1152–1161. Aria Haghighi and Dan Klein. 2010. Coreference resolution in a modular, entity-centered model. In Proceedings of Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics. Jan Haji˘c, Jarmila Panevov´a, Eva Haji˘cov´a, Jarmila Panevov´a, Petr Sgall, Petr Pajas, Jan St˘ep´anek, Ji˘r´ı Havelka, and Marie Mikulov´a. 2006. The Prague Dependency Treebank 2.0. In Linguistic Data Consortium. Sanda Harabagiu, R˘azvan Bunescu, and Steven Maiorano. 2001. Text and knowledge mining for coreference resolution. In Proceedings of the 2nd Meeting of the North American Chapter of the Association for Computational Linguistics, pages 55–62. 1406 Laura Hasler, Constantin Orasan, and Karin Naumann. 2006. NPs for events: Experiments in coreference annotation. In Proceedings of the 5th International Conference on Language Resources and Evaluation, pages 1167–1172. Peter Heeman and James Allen. 1995. The TRAINS spoken dialog corpus. CD-ROM, Linguistic Data Consortium. Graeme Hirst. 1981. Discourse-oriented anaphora resolution in natural language understanding: A review. American Journal of Computational Linguistics, 7(2):85–98. Jerry Hobbs. 1978. Resolving pronoun references. Lingua, 44:311–338. V´eronique Hoste and Walter Daelemans. 2005. Comparing learning approaches to coreference resolution. There is more to it than bias. In Proceedings of the ICML Workshop on Meta-Learning. V´eronique Hoste. 2005. Optimization Issues in Machine Learning of Coreference Resolution. Ph.D. thesis, University of Antewerp, Belgium. Eduard Hovy, Mitchell Marcus, Martha Palmer, Lance Ramshaw, and Ralph Weischedel. 2006. Ontonotes: The 90% solution. In Proceedings of the Human Language Technology Conference of the NAACL, Companion Volume: Short Papers, pages 57–60. Zhiheng Huang, Guangping Zeng, Weiqun Xu, and Asli Celikyilmaz. 2009. Accurate semantic class classifier for coreference resolution. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1232–1240. Nancy Ide and Dan Cristea. 2000. A hierarchical account of referential accessibility. In Proceedings of the 38th Annual Meeting of the Association for Computational Linguistics, pages 416–424. Ryu Iida, Kentaro Inui, Hiroya Takamura, and Yuji Matsumoto. 2003. Incorporating contextual cues in trainable models for coreference resolution. In Proceedings of the EACL Workshop on The Computational Treatment of Anaphora. Ryu Iida, Kentaro Inui, and Yuji Matsumoto. 2006. Exploting syntactic patterns as clues in zeroanaphora resolution. In Proceedings of the 21st International Conference on Computational Linguistics and the 44th Annual Meeting of the Association for Computational Linguistics, pages 625–632. Ryu Iida, Kentaro Inui, and Yuji Matsumoto. 2007a. Zero-anaphora resolution by learning rich syntactic pattern features. ACM Transactions on Asian Language Information Processing, 6(4). Ryu Iida, Mamoru Komachi, Kentaro Inui, and Yuji Matsumoto. 2007b. Annotating a Japanese text corpus with predicate-argument and coreference relations. In Proceedings of the ACL Workshop ’Linguistic Annotation Workshop’, pages 132–139. Ryu Iida, Kentaro Inui, and Yuji Matsumoto. 2009. Capturing salience with a trainable cache model for zero-anaphora resolution. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 647–655. Ryu Iida. 2007. Combining Linguistic Knowledge and Machine Learning for Anaphora Resolution. Ph.D. thesis, Nara Institute of Science and Technology, Japan. Heng Ji, David Westbrook, and Ralph Grishman. 2005. Using semantic relations to refine coreference decisions. In Proceedings of the Human Language Technology Conference and the Conference on Empirical Methods in Natural Language Processing, pages 17–24. Thorsten Joachims. 1999. Making large-scale SVM learning practical. In Bernhard Scholkopf and Alexander Smola, editors, Advances in Kernel Methods - Support Vector Learning, pages 44–56. MIT Press. Andrew Kehler, Douglas Appelt, Lara Taylor, and Aleksandr Simma. 2004a. Competitive self-trained pronoun interpretation. In Proceedings of HLTNAACL 2004: Short Papers, pages 33–36. Andrew Kehler, Douglas Appelt, Lara Taylor, and Aleksandr Simma. 2004b. The (non)utility of predicate-argument frequencies for pronoun interpretation. In Human Language Technologies 2004: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 289–296. Andrew Kehler. 1997. Probabilistic coreference in information extraction. In Proceedings of the Second Conference on Empirical Methods in Natural Language Processing, pages 163–173. Christopher Kennedy and Branimir Boguraev. 1996. Anaphor for everyone: Pronominal anaphora resolution without a parser. In Proceedings of the 16th International Conference on Computational Linguistics, pages 113–118. Manfred Klenner and ´Etienne Ailloud. 2008. Enhancing coreference clustering. In Proceedings of the Second Workshop on Anaphora Resolution, pages 31–40. Manfred Klenner and ´Etienne Ailloud. 2009. Optimization in coreference resolution is not needed: A nearly-optimal algorithm with intensional constraints. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics, pages 442–450. Manfred Klenner. 2007. Enforcing consistency on coreference sets. In Proceedings of Recent Advances in Natural Language Processing. 1407 Fang Kong, GuoDong Zhou, and Qiaoming Zhu. 2009. Employing the centering theory in pronoun resolution from the semantic perspective. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 987–996. Shalom Lappin and Herbert Leass. 1994. An algorithm for pronominal anaphora resolution. Computational Linguistics, 20(4):535–562. Xiaoqiang Luo and Imed Zitouni. 2005. Multi-lingual coreference resolution with syntactic features. In Proceedings of the Human Language Technology Conference and the Conference on Empirical Methods in Natural Language Processing, pages 660– 667. Xiaoqiang Luo, Abe Ittycheriah, Hongyan Jing, Nanda Kambhatla, and Salim Roukos. 2004. A mentionsynchronous coreference resolution algorithm based on the Bell tree. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics, pages 135–142. Xiaoqiang Luo. 2005. On coreference resolution performance metrics. In Proceedings of the Human Language Technology Conference and the Conference on Empirical Methods in Natural Language Processing, pages 25–32. Xiaoqiang Luo. 2007. Coreference or not: A twin model for coreference resolution. In Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 73–80. Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19(2):313–330. Katja Markert and Malvina Nissim. 2005. Comparing knowledge sources for nominal anaphora resolution. Computational Linguistics, 31(3):367–402. Andrew McCallum and Ben Wellner. 2003. Toward conditional models of identity uncertainty with application to proper noun coreference. In Proceedings of the IJCAI Workshop on Information Integration on the Web. Andrew McCallum and Ben Wellner. 2004. Conditional models of identity uncertainty with application to noun coreference. In Advances in Neural Information Proceesing Systems. Joseph McCarthy and Wendy Lehnert. 1995. Using decision trees for coreference resolution. In Proceedings of the Fourteenth International Conference on Artificial Intelligence, pages 1050–1055. Ruslan Mitkov. 1999. Anaphora resolution: The state of the art. Technical Report (Based on the COLING/ACL-98 tutorial on anaphora resolution), University of Wolverhampton, Wolverhampton. Ruslan Mitkov. 2001. Outstanding issues in anaphora resolution. In Al. Gelbukh, editor, Computational Linguistics and Intelligent Text Processing, pages 110–125. Springer. Ruslan Mitkov. 2002. Anaphora Resolution. Longman. Natalia N. Modjeska, Katja Markert, and Malvina Nissim. 2003. Using the web in machine learning for other-anaphora resolution. In Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing, pages 176–183. MUC-6. 1995. Proceedings of the Sixth Message Understanding Conference. MUC-7. 1998. Proceedings of the Seventh Message Understanding Conference. Christoph M¨uller, Stefan Rapp, and Michael Strube. 2002. Applying co-training to reference resolution. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 352– 359. Christoph M¨uller. 2006. Automatic detection of nonreferential it in spoken multi-party dialog. In Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics, pages 49–56. Vincent Ng and Claire Cardie. 2002a. Combining sample selection and error-driven pruning for machine learning of coreference rules. In Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing, pages 55–62. Vincent Ng and Claire Cardie. 2002b. Identifying anaphoric and non-anaphoric noun phrases to improve coreference resolution. In Proceedings of the 19th International Conference on Computational Linguistics, pages 730–736. Vincent Ng and Claire Cardie. 2002c. Improving machine learning approaches to coreference resolution. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 104– 111. Vincent Ng. 2004. Learning noun phrase anaphoricity to improve conference resolution: Issues in representation and optimization. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics, pages 151–158. Vincent Ng. 2007a. Semantic class induction and coreference resolution. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 536–543. Vincent Ng. 2007b. Shallow semantics for coreference resolution. In Proceedings of the Twentieth International Joint Conference on Artificial Intelligence, pages 1689–1694. 1408 Vincent Ng. 2008. Unsupervised models for coreference resolution. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 640–649. Vincent Ng. 2009. Graph-cut-based anaphoricity determination for coreference resolution. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 575–583. Giang Linh Ngu.y, V´aclav Nov´ak, and Zdenˇek ˇZabokrtsk´y. 2009. Comparison of classification and ranking approaches to pronominal anaphora resolution in Czech. In Proceedings of the SIGDIAL 2009 Conference, pages 276–285. Cristina Nicolae and Gabriel Nicolae. 2006. BestCut: A graph algorithm for coreference resolution. In Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing, pages 275–283. Kristina Nilsson. 2010. Hybrid Methods for Coreference Resolution in Swedish. Ph.D. thesis, Stockholm University, Sweden. Tomoko Ohta, Yuka Tateisi, and Jin-Dong Kim. 2002. The GENIA corpus: An annotated research abstract corpus in molecular biology domain. In Proceedings of the Second International Conference on Human Language Technology Research, pages 82–86. Constantin Or˘asan and Richard Evans. 2007. NP animacy identification for anaphora resolution. Journal of Artificial Intelligence Research, 29:79 – 103. Constantin Or˘asan, Dan Cristea, Ruslan Mitkov, and Ant´onio H. Branco. 2008. Anaphora Resolution Exercise: An overview. In Proceedings of the 6th Language Resources and Evaluation Conference, pages 2801–2805. Chris Paice and Gareth Husk. 1987. Towards the automatic recognition of anaphoric features in English text: the impersonal pronoun ’it’. Computer Speech and Language, 2:109–132. Massimo Poesio and Mijail A. Kabadjov. 2004. A general-purpose, off-the-shelf anaphora resolution module: Implementation and preliminary evaluation. In Proceedings of the 4th International Conference on Language Resources and Evaluation, pages 663–668. Massimo Poesio, Rahul Mehta, Axel Maroudas, and Janet Hitzeman. 2004a. Learning to resolve bridging references. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics, pages 143–150. Massimo Poesio, Olga Uryupina, Renata Vieira, Mijail Alexandrov-Kabadjov,and Rodrigo Goulart. 2004b. Discourse-new detectors for definite description resolution: A survey and a preliminary proposal. In Proeedings of the ACL Workshop on Reference Resolution. Massimo Poesio, David Day, Ron Artstein, Jason Duncan, Vladimir Eidelman, Claudio Giuliano, Rob Hall, Janet Hitzeman, Alan Jern, Mijail Kabadjov, Stanley Yong Wai Keong, Gideon Mann, Alessandro Moschitti, Simone Ponzetto, Jason Smith, Josef Steinberger, Michael Strube, Jian Su, Yannick Versley, Xiaofeng Yang, and Michael Wick. 2007. ELERFED: Final report of the research group on Exploiting Lexical and Encyclopedic Resources For Entity Disambiguation. Technical report, Summer Workshop on Language Engineering, Center for Language and Speech Processing, Johns Hopkins University, Baltimore, MD. Simone Paolo Ponzetto and Massimo Poesio. 2009. State-of-the-art NLP approaches to coreference resolution: Theory and practical recipes. In Tutorial Abstracts of ACL-IJCNLP 2009, page 6. Simone Paolo Ponzetto and Michael Strube. 2006a. Exploiting semantic role labeling, WordNet and Wikipedia for coreference resolution. In Human Language Technologies 2006: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 192–199. Simone Paolo Ponzetto and Michael Strube. 2006b. Semantic role labeling for coreference resolution. In Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics, pages 143–146. Simone Paolo Ponzetto and Michael Strube. 2007. Knowledge derived from Wikipedia for computing semantic relatedness. Journal of Artificial Intelligence Research, 30:181–212. Hoifung Poon and Pedro Domingos. 2008. Joint unsupervised coreference resolution with Markov Logic. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 650–659. Andrei Popescu-Belis, Lo¨ıs Rigouste, Susanne Salmon-Alt, and Laurent Romary. 2004. Online evaluation of coreference resolution. In Proceedings of the 4th International Conference on Language Resources and Evaluation, pages 1507–1510. Long Qiu, Min-Yen Kan, and Tat-Seng Chua. 2004. A public reference implementation of the RAP anaphora resolution algorithm. In Proceedings of the 4th International Conference on Language Resources and Evaluation, pages 291–294. John Ross Quinlan. 1993. C4.5: Programs for Machine Learning. Morgan Kaufmann, San Mateo, CA. Altaf Rahman and Vincent Ng. 2009. Supervised models for coreference resolution. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 968–977. 1409 Marta Recasens and M. Ant´onia Mart´ı. 2009. AnCoraCO: Coreferentially annotated corpora for Spanish and Catalan. Language Resources and Evaluation, 43(4). Marta Recasens, Toni Mart´ı, Mariona Taul´e, Llu´ıs M`arquez, and Emili Sapena. 2009. SemEval2010 Task 1: Coreference resolution in multiple languages. In Proceedings of the Workshop on Semantic Evaluations: Recent Achievements and Future Directions (SEW-2009), pages 70–75. Candace Sidner. 1979. Towards a Computational Theory of Definite Anaphora Comprehension in English Discourse. Ph.D. thesis, Massachusetts Institute of Technology, USA. Wee Meng Soon, Hwee Tou Ng, and Chung Yong Lim. 1999. Corpus-based learning for noun phrase coreference resolution. In Proceedings of the 1999 Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora, pages 285–291. Wee Meng Soon, Hwee Tou Ng, and Daniel Chung Yong Lim. 2001. A machine learning approach to coreference resolution of noun phrases. Computational Linguistics, 27(4):521–544. Veselin Stoyanov, Nathan Gilbert, Claire Cardie, and Ellen Riloff. 2009. Conundrums in noun phrase coreference resolution: Making sense of the stateof-the-art. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 656–664. Veselin Stoyanov, Claire Cardie, Nathan Gilbert, Ellen Riloff, David Buttler, and David Hysom. 2010. Coreference resolution with Reconcile. In Proceedings of the ACL 2010 Conference Short Papers. Michael Strube, Stefan Rapp, and Christoph M¨uller. 2002. The influence of minimum edit distance on reference resolution. In Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing, pages 312–319. Michael Strube. 2002. NLP approaches to reference resolution. In Tutorial Abstracts of ACL 2002, page 124. Michael Strube. 2009. Anaphernresolution. In Computerlinguistik und Sprachtechnologie. Eine Einfuhrung. Springer, Heidelberg, Germany, 3rd edition. Heike Telljohann, Erhard Hinrichs, and Sandra K¨ubler. 2004. The t¨uba-d/z treebank: Annotating German with a context-free backbone. In Proceedings of the 4th International Conference on Language Resources and Evaluation, pages 2229–2235. Joel Tetreault. 2005. Empirical Evaluations of Pronoun Resolution. Ph.D. thesis, University of Rochester, USA. Olga Uryupina. 2003. High-precision identification of discourse new and unique noun phrases. In Proceedings of the ACL Student Research Workshop, pages 80–86. Olga Uryupina. 2004. Linguistically motivated sample selection for coreference resolution. In Proceedings of the 5th Discourse Anaphora and Anaphor Resolution Colloquium. Kees van Deemter and Rodger Kibble. 2000. On coreferring: Coreference in MUC and related annotation schemes. Computational Linguistics, 26(4):629– 637. Yannick Versley, Alessandro Moschitti, Massimo Poesio, and Xiaofeng Yang. 2008a. Coreference systems based on kernels methods. In Proceedings of the 22nd International Conference on Computational Linguistics, pages 961–968. Yannick Versley, Simone Paolo Ponzetto, Massimo Poesio, Vladimir Eidelman, Alan Jern, Jason Smith, Xiaofeng Yang, and Alessandro Moschitti. 2008b. BART: A modular toolkit for coreference resolution. In Proceedings of the ACL-08: HLT Demo Session, pages 9–12. Yannick Versley. 2006. A constraint-based approach to noun phrase coreference resolution in German newspaper text. In Konferenz zur Verarbeitung Nat¨urlicher Sprache. Yannick Versley. 2007. Antecedent selection techniques for high-recall coreference resolution. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 496–505. Renata Vieira and Massimo Poesio. 2000. Processing definite descriptions in corpora. In S. Botley and A. McEnery, editors, Corpus-based and Computational Approaches to Discourse Anaphora, pages 189–212. UCL Press. Marc Vilain, John Burger, John Aberdeen, Dennis Connolly, and Lynette Hirschman. 1995. A modeltheoretic coreference scoring scheme. In Proceedings of the Sixth Message Understanding Conference, pages 45–52. Marilyn Walker, Aravind Joshi, and Ellen Prince, editors. 1998. Centering Theory in Discourse. Oxford University Press. Holger Wunsch. 2010. Rule-based and Memory-based Pronoun Resolution for German: A Comparison and Assessment of Data Sources. Ph.D. thesis, University of T¨ubingen, Germany. Xiaofeng Yang and Jian Su. 2007. Coreference resolution using semantic relatedness information from automatically discovered patterns. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics, pages 528–535. 1410 Xiaofeng Yang, Guodong Zhou, Jian Su, and Chew Lim Tan. 2003. Coreference resolution using competitive learning approach. In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, pages 176–183. Xiaofeng Yang, Jian Su, and Chew Lim Tan. 2004a. Improving noun phrase coreference resolution by matching strings. In Proceedings of the First International Joint Conference on Natural Language Processing, pages 22–31. Xiaofeng Yang, Jian Su, GuoDong Zhou, and Chew Lim Tan. 2004b. An NP-cluster based approach to coreference resolution. In Proceedings of the 20th International Conference on Computational Linguistics, pages 226–232. Xiaofeng Yang, Jian Su, and Chew Lim Tan. 2005. Improving pronoun resolution using statistics-based semantic compatibility information. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, pages 165–172. Xiaofeng Yang, Jian Su, and Chew Lim Tan. 2006. Kernel based pronoun resolution with structured syntactic knowledge. In Proceedings of the 21st International Conference on Computational Linguistics and the 44th Annual Meeting of the Association for Computational Linguistics, pages 41–48. Xiaofeng Yang, Jian Su, Jun Lang, Chew Lim Tan, and Sheng Li. 2008a. An entity-mention model for coreference resolution with inductive logic programming. In Proceedings of ACL-08: HLT, pages 843–851. Xiaofeng Yang, Jian Su, and Chew Lim Tan. 2008b. A twin-candidate model for learning-based anaphora resolution. Computational Linguistics, 34(3):327– 356. Dmitry Zelenko, Chinatsu Aone, and Jason Tibbetts. 2004. Coreference resolution for information extraction. In Proceedings of the ACL Workshop on Reference Resolution and its Applications, pages 9– 16. Shanheng Zhao and Hwee Tou Ng. 2007. Identification and resolution of Chinese zero pronouns: A machine learning approach. In Proceedings of the 2007 Joint Conference on Empirical Methods on Natural Language Processing and Computational Natural Language Learning, pages 541–550. GuoDong Zhou and Fang Kong. 2009. Global learning of noun phrase anaphoricity in coreference resolution via label propagation. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 978–986. 1411
|
2010
|
142
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1412–1422, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Unsupervised Event Coreference Resolution with Rich Linguistic Features Cosmin Adrian Bejan Institute for Creative Technologies University of Southern California Marina del Rey, CA 90292, USA Sanda Harabagiu Human Language Technology Institute University of Texas at Dallas Richardson, TX 75083, USA Abstract This paper examines how a new class of nonparametric Bayesian models can be effectively applied to an open-domain event coreference task. Designed with the purpose of clustering complex linguistic objects, these models consider a potentially infinite number of features and categorical outcomes. The evaluation performed for solving both within- and cross-document event coreference shows significant improvements of the models when compared against two baselines for this task. 1 Introduction The event coreference task consists of finding clusters of event mentions that refer to the same event. Although it has not been extensively studied in comparison with the related problem of entity coreference resolution, solving event coreference has already proved its usefulness in various applications such as topic detection and tracking (Allan et al., 1998), information extraction (Humphreys et al., 1997), question answering (Narayanan and Harabagiu, 2004), textual entailment (Haghighi et al., 2005), and contradiction detection (de Marneffe et al., 2008). Previous approaches for solving event coreference relied on supervised learning methods that explore various linguistic properties in order to decide if a pair of event mentions is coreferential or not (Humphreys et al., 1997; Bagga and Baldwin, 1999; Ahn, 2006; Chen and Ji, 2009). In spite of being successful for a particular labeled corpus, these pairwise models are dependent on the domain or language that they are trained on. Moreover, since event coreference resolution is a complex task that involves exploring a rich set of linguistic features, annotating a large corpus with event coreference information for a new language or domain of interest requires a substantial amount of manual effort. Also, since these models are dependent on local pairwise decisions, they are unable to capture a global event distribution at topic or document collection level. To address these limitations and to provide a more flexible representation for modeling observable data with rich properties, we present two novel, fully generative, nonparametric Bayesian models for unsupervised withinand crossdocument event coreference resolution. The first model extends the hierarchical Dirichlet process (Teh et al., 2006) to take into account additional properties associated with observable objects (i.e., event mentions). The second model overcomes some of the limitations of the first model. It uses the infinite factorial hidden Markov model (Van Gael et al., 2008b) coupled to the infinite hidden Markov model (Beal et al., 2002) in order to (1) consider a potentially infinite number of features associated with observable objects, (2) perform an automatic selection of the most salient features, and (3) capture the structural dependencies of observable objects at the discourse level. Furthermore, both models are designed to account for a potentially infinite number of categorical outcomes (i.e., events). These models provide additional details and experimental results to our preliminary work on unsupervised event coreference resolution (Bejan et al., 2009). 2 Event Coreference The problem of determining if two events are identical was originally studied in philosophy. One relevant theory on event identity was proposed by Davidson (1969) who argued that two events are identical if they have the same causes and effects. Later on, a different theory was proposed by Quine (1985) who considered that each event refers to a physical object (which is well defined in space and time), and therefore, two events are identical 1412 if they have the same spatiotemporal location. In (Davidson, 1985), Davidson abandoned his suggestion to embrace the Quinean theory on event identity (Malpas, 2009). 2.1 An Example In accordance with the Quinean theory, we consider that two event mentions are coreferential if they have the same event properties and share the same event participants. For instance, the sentences from Example 1 encode event mentions that refer to several individuated events. These sentences are extracted from a newly annotated corpus with event coreference information (see Section 4). In this corpus, we organize documents that describe the same seminal event into topics. In particular, the topics shown in this example describe the seminal event of buying ATI by AMD (topic 43) and the seminal event of buying EDS by HP (topic 44). Although all the event mentions of interest emphasized in boldface in Example 1 evoke the same generic event buy, they refer to three individuated events: e1 = {em1, em2}, e2 = {em3−6, em8}, and e3 = {em7}. For example, em1(buy) and em3(buy) correspond to different individuated events since they have a different AGENT ([BUYER(em1)=AMD] ̸= [BUYER(em3)=HP]). This organization of event mentions leads to the idea of creating an event hierarchy which has on the first level, event mentions, on the second level, individuated events, and on the third level, generic events. In particular, the event hierarchy corresponding to the event mentions annotated in our example is illustrated in Figure 1. Solving the event coreference problem poses many interesting challenges. For instance, in order to solve the coreference chain of event mentions that refer to the event e2, we need to take into account the following issues: (i) a coreference chain can encode both within- and cross-document coreference information; (ii) two mentions from the same chain can have different word classes (e.g., em3(buy)–verb, em4(purchase)–noun); (iii) not all the mentions from the same chain are synonymous (e.g., em3(buy) and em8(acquire)), although a semantic relation might exist between them (e.g., in WordNet (Fellbaum, 1998), the genus of buy is acquire); (iv) partial (or all) properties and participants of an event mention can be omitted in text (e.g., em4(purchase)). In Section Topic 43 Document 3 s4: AMD agreed to [buy]em1 Markham, Ontario-based ATI for around $5.4 billion in cash and stock, the companies announced Monday. s5: The [acquisition]em2 would turn AMD into one of the world’s largest providers of graphics chips. Topic 44 Document 2 s1: Hewlett-Packard is negotiating to [buy]em3 technology services provider Electronic Data Systems. s8: With a market value of about $115 billion, HP could easily use its own stock to finance the [purchase]em4. s9: If the [deal]em5 is completed, it would be HP’s biggest [acquisition]em6 since it [bought]em7 Compaq Computer Corp. for $19 billion in 2002. Document 5 s2: Industry sources have confirmed to eWEEK that Hewlett-Packard will [acquire]em8 Electronic Data Systems for about $13 billion. Example 1: Examples of event mention annotations. buy em7 e2 e3 e1 em5 em6 em3 em2 em1 em4 em8 Figure 1: Fragment from the event hierarchy. 5, we discuss additional aspects of the event coreference problem that are not revealed in Example 1. 2.2 Linguistic Features The events representing coreference clusters of event mentions are characterized by a large set of linguistic features. To compute an accurate event distribution for event coreference resolution, we associate the following categories of linguistic features with each annotated event mention. Lexical Features (LF) We capture the lexical context of an event mention by extracting the following features: the head word (HW), the lemmatized head word (HL), the lemmatized left and right words surrounding the mention (LHL,RHL), and the HL features corresponding to the left and right mentions (LHE,RHE). For instance, the lexical features extracted for the event mention em7(bought) from our example are HW:bought, HL:buy, LHL:it, RHL:Compaq, LHE:acquisition, and RHE:acquire. Class Features (CF) These features aim to group mentions into several types of classes: the partof-speech of the HW feature (POS), the word class of the HW feature (HWC), and the event class of the mention (EC). The HWC feature can take one of the following values: VERB, NOUN, ADJEC1413 TIVE, and OTHER. As values for the EC feature, we consider the seven event classes defined in the TimeML specification language (Pustejovsky et al., 2003a): OCCURRENCE, PERCEPTION, REPORTING, ASPECTUAL, STATE, I ACTION, and I STATE. In order to extract the event classes corresponding to the event mentions from a given dataset, we employed the event extractor described in (Bejan, 2007). This extractor is trained on the TimeBank corpus (Pustejovsky et al., 2003b), which is a TimeML resource encoding temporal elements such as events, time expressions, and temporal relations. WordNet Features (WF) In our efforts to create clusters of event mention attributes as close as possible to the true attribute clusters of the individuated events, we build two sets of word clusters using the entire lexical information from the WordNet database. After creating these sets of clusters, we then associate each event mention with only one cluster from each set. The first set uses the transitive closure of the WordNet SYNONYMOUS relation to form clusters with all the words from WordNet (WNS). For instance, the verbs buy and purchase correspond to the same cluster ID because there exist a chain of SYNONYMOUS relations between them in WordNet. The second set considers as grouping criteria the categorization of words from the WordNet lexicographer’s files (WNL). In addition, for each word that is not covered in WordNet, we create a new cluster ID in each set of clusters. Semantic Features (SF) To extract features that characterize participants and properties of event mentions, we use the semantic parser described in (Bejan and Hathaway, 2007). One category of semantic features that we identify for event mentions is the predicate argument structures encoded in PropBank annotations (Palmer et al., 2005). In PropBank, the predicate argument structures are represented by events expressed as verbs in text and by the semantic roles, or predicate arguments, associated with these events. For example, ARG0 annotates a specific type of semantic role which represents the AGENT, DOER, or ACTOR of a specific event. Another argument is ARG1, which plays the role of the PATIENT, THEME, or EXPERIENCER of an event. In particular, the predicate arguments associated to the event mention em8(bought) from Example 1 are ARG0:[it], ARG1:[Compaq Computer Corp.], ARG3:[for $19 billion], and ARG-TMP:[in 2002]. Event mentions are not only expressed as verbs in text, but also as nouns and adjectives. Therefore, for a better coverage of semantic features, we also employ the semantic annotations encoded in the FrameNet corpus (Baker et al., 1998). FrameNet annotates word expressions capable of evoking conceptual structures, or semantic frames, which describe specific situations, objects, or events (Fillmore, 1982). The semantic roles associated with a word in FrameNet, or frame elements, are locally defined for the semantic frame evoked by the word. In general, the words annotated in FrameNet are expressed as verbs, nouns, and adjectives. To preserve the consistency of semantic role features, we align frame elements to predicate arguments by running the PropBank semantic parser on the manual annotations from FrameNet; conversely, we also run the FrameNet parser on the manual annotations from PropBank. Moreover, to obtain a better alignment of semantic roles, we run both parsers on a large amount of unlabeled text. The result of this process is a map with all frame elements statistically aligned to all predicate arguments. For instance, in 99.7% of the cases the frame element BUYER of the semantic frame COMMERCE BUY is mapped to ARG0, and in the remaining 0.3% of the cases to ARG1. Additionally, we use this map to create a more general semantic feature which assigns to each predicate argument a frame element label. In particular, the features for em8(acquire) are FEA0:BUYER, FEA1:GOODS, FEA3:MONEY, and FEATMP:TIME. Two additional semantic features used in our experiments are: (1) the semantic frame (FR) evoked by every mention;1 and (2) the WNS feature applied to the head word of every semantic role (e.g., WSA0, WSA1). Feature Combinations (FC) We also explore various combinations of the features presented above. Examples include HW+HWC, HL+FR, FR+ARG1, LHL+RHL, etc. It is worth noting that there exist event mentions for which not all the features can be extracted. For example, the LHE and RHE features are missing for the first and last event mentions in a document, respectively. Also, many semantic roles can be absent for an event mention in a given context. 1 The reason for extracting this feature is given by the fact that, in general, frames are able to capture properties of generic events (Lowe et al., 1997). 1414 3 Nonparametric Bayesian Models As input for our models, we consider a collection of I documents, where each document i has Ji event mentions. For features, we make the distinction between feature types and feature values (e.g., POS is a feature type and has values such as NN and VB). Each event mention is characterized by L feature types, FT, and each feature type is represented by a finite vocabulary of feature values, fv. Thus, we can represent the observable properties of an event mention as a vector of L feature type – feature value pairs ⟨(FT1 : fv1i), . . . , (FTL : fvLi)⟩, where each feature value index i ranges in the feature value space associated with a feature type. 3.1 A Finite Feature Model We present an extension of the hierarchical Dirichlet process (HDP) model which is able to represent each observable object (i.e., event mention) by a finite number of feature types L. Our HDP extension is also inspired from the Bayesian model proposed by Haghighi and Klein (2007). However, their model is strictly customized for entity coreference resolution, and therefore, extending it to include additional features for each observable object is a challenging task (Ng, 2008; Poon and Domingos, 2008). In the HDP model, a Dirichlet process (DP) (Ferguson, 1973) is associated with each document, and each mixture component (i.e., event) is shared across documents. To describe its extension, we consider Z the set of indicator random variables for indices of events, φz the set of parameters associated with an event z, φ a notation for all model parameters, and X a notation for all random variables that represent observable features.2 Given a document collection annotated with event mentions, the goal is to find the best assignment of event indices Z∗, which maximize the posterior probability P(Z|X). In a Bayesian approach, this probability is computed by integrating out all model parameters: P(Z|X)= Z P(Z, φ|X)dφ= Z P(Z|X, φ)P(φ|X)dφ Our HDP extension is depicted graphically in Figure 2(a). Similar to the HDP model, the distribution over events associated with each document, β, is generated by a Dirichlet process with a 2 In this subsection, the feature term is used in context of a feature type. concentration parameter α > 0. Since this setting enables a clustering of event mentions at the document level, it is desirable that events be shared across documents and the number of events K be inferred from data. To ensure this flexibility, a global nonparametric DP prior with a hyperparameter γ and a global base measure H can be considered for β (Teh et al., 2006). The global distribution drawn from this DP prior, denoted as β0 in Figure 2(a), encodes the event mixing weights. Thus, same global events are used for each document, but each event has a document specific distribution βi that is drawn from a DP prior centered on the global weights β0. To infer the true posterior probability of P(Z|X), we follow (Teh et al., 2006) and use the Gibbs sampling algorithm (Geman and Geman, 1984) based on the direct assignment sampling scheme. In this sampling scheme, the parameters β and φ are integrated out analytically. Moreover, to reduce the complexity of computing P(Z|X), we make the na¨ıve Bayes assumption that the feature variables X are conditionally independent given Z. This allows us to factorize the joint distribution of feature variables X conditioned on Z into product of marginals. Thus, by Bayes rule, the formula for sampling an event index for mention j from document i, Zi,j, is:3 P(Zi,j | Z−i,j, X) ∝P(Zi,j | Z−i,j) Y X∈X P(Xi,j |Z, X−i,j) where Xi,j represents the feature value of a feature type corresponding to the event mention j from the document i. In the process of generating an event mention, an event index z is first sampled by using a mechanism that facilitates sampling from a prior for infinite mixture models called the Chinese restaurant franchise (CRF) representation, as reported in (Teh et al., 2006): P(Zi,j = z | Z−i,j, β0) ∝ ( αβu 0 , if z = znew nz + αβz 0, otherwise Here, nz is the number of event mentions with event index z, znew is a new event index not used already in Z−i,j, βz 0 are the global mixing proportions associated with the K events, and βu 0 is the weight for the unknown mixture component. Next, to generate a feature value x (with the feature type X) of the event mention, the event z is 3 Z−i,j represents a notation for Z −{Zi,j}. 1415 H Zi ∞ β α γ φ ∞ Xi (a) β0 ∞ Ji I L φ ∞ HLi FRi POSi α γ H θ F2 0 F2 1 F2 2 F2 T ∞ β β0 ∞ I Ji Zi (b) F1 0 Y1 F1 1 Y2 F1 2 YT F1 T S0 FM 0 FM 1 FM 2 FM T S1 S2 ST Phase 1 Phase 2 (c) Figure 2: Graphical representation of our models: nodes correspond to random variables; shaded nodes denote observable variables; a rectangle captures the replication of the structure it contains, where the number of replications is indicated in the bottom-right corner. The model in (a) illustrates a flat representation of a limited number of features in a generalized framework (henceforth, HDPflat). The model in (b) captures a simple example of structured network topology of three feature variables (henceforth, HDPstruct). The dependencies involving parameters φ and θ in these models are omitted for clarity. The model from (c) shows the representation of the iFHMM-iHMM model as well as the main phases of its generative process. associated with a multinomial emission distribution over the feature values of X having the parameters φ = ⟨φx Z⟩. We assume that this emission distribution is drawn from a symmetric Dirichlet distribution with concentration λX: P(Xi,j = x | Z, X−i,j) ∝nx,z + λX where Xi,j is the feature type of the mention j from the document i, and nx,z is the number of times the feature value x has been associated with the event index z in (Z, X−i,j). We also apply the Lidstone’s smoothing method to this distribution. In cases when only a feature type is considered (e.g., X = ⟨HL⟩), the HDPflat model is identical with the original HDP model. We denote this one feature model by HDP1f. When dependencies between feature variables exist (e.g., in our case, frame elements are dependent on the semantic frames that define them, and frames are dependent on the words that evoke them), various global distributions are involved for computing P(Z|X). For the model depicted in Figure 2(b), for instance, the posterior probability is given by: P(Zi,j)P(FRi,j |HLi,j, θ) Y X∈X P(Xi,j |Z) In this formula, P(FRi,j|HLi,j, θ) is a global distribution parameterized by θ, and X is a feature variable from the set X = ⟨HL, POS, FR⟩. For the sake of clarity, we omit the conditioning components of Z, HL, FR, and POS. 3.2 An Infinite Feature Model To relax some of the restrictions of the first model, we devise an approach that combines the infinite factorial hidden Markov model (iFHMM) with the infinite hidden Markov model (iHMM) to form the iFHMM-iHMM model. The iFHMM framework uses the Markov Indian buffet process (mIBP) (Van Gael et al., 2008b) in order to represent each object as a sparse subset of a potentially unbounded set of latent features (Griffiths and Ghahramani, 2006; Ghahramani et al., 2007; Van Gael et al., 2008a).4 Specifically, the mIBP defines a distribution over an unbounded set of binary Markov chains, where each chain can be associated with a binary latent feature that evolves over time according to Markov dynamics. Therefore, if we denote by M the total number of feature chains and by T the number of observable components, the mIBP defines a probability distribution over a binary matrix F with T rows, which correspond to observations, and an unbounded number of columns M, which correspond to features. An observation yt contains a subset from the unbounded set of features {f 1, f 2, . . . , f M} that is represented in the matrix by a binary vector Ft =⟨F 1 t , F 2 t , . . . , F M t ⟩, where F i t = 1 indicates that f i is associated with yt. In other words, F decomposes the observations and represents them as feature factors, which can then be associated with hidden variables in an iFHMM model as depicted in Figure 2(c). 4 In this subsection, a feature will be represented by a (feature type:feature value) pair. 1416 Although the iFHMM allows a more flexible representation of the latent structure by letting the number of parallel Markov chains M be learned from data, it cannot be used as a framework where the number of clustering components K is infinite. On the other hand, the iHMM represents a nonparametric extension of the hidden Markov model (HMM) (Rabiner, 1989) that allows performing inference on an infinite number of states K. To further increase the representational power for modeling discrete time series data, we propose a nonparametric extension that combines the best of the two models, and lets the parameters M and K be learned from data. As shown in Figure 2(c), each step in the new iHMM-iFHMM generative process is performed in two phases: (i) the latent feature variables from the iFHMM framework are sampled using the mIBP mechanism; and (ii) the features sampled so far, which become observable during this second phase, are used in an adapted version of the beam sampling algorithm (Van Gael et al., 2008a) to infer the clustering components (i.e., latent events). In the first phase, the stochastic process for sampling features in F is defined as follows. The first component samples a number of Poisson(α′) features. In general, depending on the value that was sampled in the previous step (t −1), a feature f m is sampled for the tth component according to the P(F m t = 1 | F m t−1 = 1) and P(F m t = 1 | F m t−1 = 0) probabilities.5 After all features are sampled for the tth component, a number of Poisson(α′/t) new features are assigned for this component, and M gets incremented accordingly. To describe the adapted beam sampler, which is employed in the second phase of the generative process, we introduce additional notations. We denote by (s1, . . . , sT ) the sequence of hidden states corresponding to the sequence of event mentions (y1, . . . , yT ), where each state st belongs to one of the K events, st ∈{1, . . . , K}, and each mention yt is represented by a sequence of latent features ⟨F 1 t , F 2 t , . . . , F M t ⟩. One element of the transition probability π is defined as πij = P(st = j | st−1 =i), and a mention yt is generated according to a likelihood model F that is parameterized by a state-dependent parameter φst (yt | st ∼F(φst)). The observation parameters φ are drawn independently from an identical prior base distribution H. The beam sampling algorithm combines the 5 Technical details for computing these probabilities are described in (Van Gael et al., 2008b). ideas of slice sampling and dynamic programming for an efficient sampling of state trajectories. Since in time series models the transition probabilities have independent priors (Beal et al., 2002), Van Gael and colleagues (2008a) also used the HDP mechanism to allow couplings across transitions. For sampling the whole hidden state trajectory s, this algorithm employs a forward filteringbackward sampling technique. In the forward step of our adapted beam sampler, for each mention yt, we sample features using the mIBP mechanism and the auxiliary variable ut ∼Uniform(0, πst−1st). As explained in (Van Gael et al., 2008a), the auxiliary variables u are used to filter only those trajectories s for which πst−1st ≥ut for all t. Also, in this step, we compute the probabilities P(st |y1:t, u1:t) for all t: P(st|y1:t,u1:t)∝P(yt|st) X st−1:ut<πst−1st P(st−1|y1:t−1,u1:t−1) Here, the dependencies involving parameters π and φ are omitted for clarity. In the backward step, we first sample the event for the last state sT directly from P(sT | y1:T, u1:T ) and then, for all t : T −1 . . . 1, we sample each state st given st+1 by using the formula P(st | st+1, y1:T , u1:T) ∝P(st | y1:t, u1:t)P(st+1 | st, ut+1). To sample the emission distribution φ efficiently, and to ensure that each mention is characterized by a finite set of representative features, we set the base distribution H to be conjugate with the data distribution F in a Dirichletmultinomial model with the multinomial parameters (o1, . . . , oK) defined as: ok = T X t=1 X fm∈Bt nmk In this formula, nmk counts how many times the feature f m was sampled for the event k, and Bt stores a finite set of features for yt. The mechanism for building a finite set of representative features for the mention yt is based on slice sampling (Neal, 2003). Letting qm be the number of times the feature f m was sampled in the mIBP, and vt an auxiliary variable for yt such that vt ∼Uniform(1, max{qm : F m t = 1}), we define the finite feature set Bt for the observation yt as Bt = {f m : F m t = 1∧qm ≥vt}. The finiteness of this feature set is based on the observation that, in the generative process of the mIBP, only a finite set 1417 of features are sampled for a component. We denote this model as iFHMM-iHMMuniform. Also, it is worth mentioning that, by using this type of sampling, only the most representative features of yt get selected in Bt. Furthermore, we explore the mechanism for selecting a finite set of features associated with an observation by: (1) considering all the observation’s features whose corresponding feature counter qm ≥1 (unfiltered); (2) selecting only the higher half of the feature distribution consisting of the observation’s features that were sampled at least once in the mIBP model (median); and (3) sampling vt from a discrete distribution of the observation’s features that were sampled at least once in the mIBP (discrete). 4 Experiments Datasets One dataset we employed is the automatic content extraction (ACE) (ACE-Event, 2005). However, the utilization of the ACE corpus for the task of solving event coreference is limited because this resource provides only withindocument event coreference annotations using a restricted set of event types such as LIFE, BUSINESS, CONFLICT, and JUSTICE. Therefore, as a second dataset, we created the EventCorefBank (ECB) corpus6 to increase the diversity of event types and to be able to evaluate our models for both within- and cross-document event coreference resolution. One important step in the creation process of this corpus consists in finding sets of related documents that describe the same seminal event such that the annotation of coreferential event mentions across documents is possible. For this purpose, we selected from the GoogleNews archive7 various topics whose description contains keywords such as commercial transaction, attack, death, sports, terrorist act, election, arrest, natural disaster, etc. The entire annotation process for creating the ECB resource is described in (Bejan and Harabagiu, 2008). Table 1 lists several basic statistics extracted from these two corpora. Evaluation For a more realistic approach, we not only trained the models on the manually annotated event mentions (i.e., true mentions), but also on all the possible mentions encoded in the two datasets. To extract all event mentions, we ran the event identifier described in (Bejan, 2007). The mentions extracted by this system (i.e., system men6 ECB is available at http://www.hlt.utdallas.edu/∼ady. 7 http://news.google.com/ ACE ECB Number of topics – 43 Number of documents 745 482 Number of within-topic events – 339 Number of cross-document events – 208 Number of within-document events 4946 1302 Number of true mentions 6553 1744 Number of system mentions 45289 21175 Number of distinct feature values 391798 237197 Table 1: Statistics of the ACE and ECB corpora. tions) were able to cover all the true mentions from both datasets. As shown in Table 1, we extracted from ACE and ECB corpora 45289 and 21175 system mentions, respectively. We report results in terms of recall (R), precision (P), and F-score (F) by employing the mention-based B3 metric (Bagga and Baldwin, 1998), the entity-based CEAF metric (Luo, 2005), and the pairwise F1 (PW) metric. All the results are averaged over 5 runs of the generative models. In the evaluation process, we considered only the true mentions of the ACE test dataset, and the event mentions of the test sets derived from a 5fold cross validation scheme on the ECB dataset. For evaluating the cross-document coreference annotations, we adopted the same approach as described in (Bagga and Baldwin, 1999) by merging all the documents from the same topic into a meta-document and then scoring this document as performed for within-document evaluation. For both corpora, we considered a set of 132 feature types, where each feature type consists on average of 3900 distinct feature values. Baselines We consider two baselines for event coreference resolution (rows 1&2 in Tables 2&3). One baseline groups each event mention by its event class (BLeclass). Therefore, for this baseline, we cluster mentions according to their corresponding EC feature value. Similarly, the second baseline uses as grouping criteria for event mentions their corresponding WNS feature value (BLsyn). HDP Extensions Due to memory limitations, we evaluated the HDP models on a restricted set of manually selected feature types. In general, the HDP1f model with the feature type HL, which plays the role of a baseline for the HDPflat and HDPstruct models, outperforms both baselines on the ACE and ECB datasets. For the HDPflat models (rows 4–7 in Tables 2&3), we classified the experiments according to the set of feature types described in Section 2. Our experiments reveal that the best configuration of features for this model 1418 Model configuration B3 CEAF PW B3 CEAF PW R P F R P F R P F R P F R P F R P F ECB | WD ECB | CD 1 BLeclass 97.7 55.8 71.0 44.5 80.1 57.2 93.7 25.4 39.8 93.8 49.6 64.9 36.6 72.7 48.7 90.7 28.6 43.3 2 BLsyn 91.5 57.4 70.5 45.7 75.9 57.0 65.3 21.9 32.6 84.6 48.1 61.3 32.8 63.6 43.3 66.2 26.0 37.3 3 HDP1f (HL) 84.3 89.0 86.5 83.4 79.6 81.4 36.6 53.4 42.6 67.0 86.2 75.3 76.2 57.1 65.2 34.9 58.9 43.5 4 HDPflat (LF) 81.4 98.2 89.0 92.7 77.2 84.2 24.7 82.8 37.7 63.8 97.3 77.0 84.9 54.3 66.1 27.2 88.5 41.5 5 (LF+CF) 81.5 98.0 89.0 92.8 77.9 84.7 24.6 80.7 37.4 64.6 97.3 77.6 85.3 55.6 67.2 27.6 88.7 42.0 6 (LF+CF+WF) 82.0 98.9 89.6 93.7 78.4 85.3 26.8 89.9 41.0 65.8 98.0 78.7 86.7 57.1 68.8 29.6 93.0 44.8 7 (LF+CF+WF+SF) 82.1 99.2 89.8 93.9 78.2 85.3 27.0 92.4 41.3 65.0 98.7 78.3 86.9 56.0 68.0 29.2 95.1 44.4 8 HDPstruct (HL→FR→FEA) 84.3 97.1 90.2 92.7 81.1 86.5 34.4 83.0 48.6 69.3 95.8 80.4 86.2 60.1 70.8 37.5 85.6 52.1 9 iFHMM-iHMMunfiltered 82.6 97.7 89.5 92.7 78.5 85.0 28.5 82.4 41.8 67.2 96.4 79.1 85.6 58.0 69.1 32.5 87.7 47.2 10 iFHMM-iHMMdiscrete 82.6 98.1 89.7 93.2 79.0 85.5 29.7 85.4 44.0 66.2 96.2 78.4 84.8 57.2 68.3 32.2 88.1 47.1 11 iFHMM-iHMMmedian 82.6 97.8 89.5 92.9 78.8 85.3 29.3 83.7 43.0 67.0 96.5 79.0 86.1 58.3 69.5 33.1 88.1 47.9 12 iFHMM-iHMMuniform 82.5 98.1 89.6 93.1 78.8 85.3 29.4 86.6 43.7 67.0 96.4 79.0 85.5 58.0 69.1 33.3 88.3 48.2 Table 2: Results for within-document (WD) and cross-document (WD) coreference resolution on the ECB dataset. B3 CEAF PW R P F R P F R P F ACE | WD 1 97.9 25.0 39.9 14.7 64.4 24.0 93.5 8.2 15.2 2 89.3 36.7 52.1 25.1 64.8 36.2 63.8 10.5 18.1 3 86.0 70.6 77.5 62.3 76.4 68.6 50.5 27.7 35.8 4 82.9 82.6 82.7 74.9 75.8 75.3 42.4 41.9 42.1 5 82.0 84.9 83.4 77.8 75.3 76.6 37.9 45.1 41.2 6 83.3 83.6 83.4 76.3 76.2 76.3 42.2 43.9 43.0 7 83.4 84.2 83.8 76.9 76.5 76.7 43.3 47.1 45.1 8 86.2 76.9 81.3 69.0 77.5 73.0 53.2 38.1 44.4 9 82.8 83.6 83.2 75.8 75.0 75.4 41.4 42.6 42.0 10 83.1 81.5 82.3 73.7 75.1 74.4 41.9 40.1 41.0 11 83.0 81.3 82.1 73.2 75.2 74.2 40.7 39.0 39.8 12 81.9 82.2 82.1 74.6 74.5 74.5 37.2 39.0 38.1 Table 3: Results for WD coreference resolution on ACE. consists of a combination of feature types from all the categories of features (row 7). For the HDPstruct experiments, we considered the set of features of the best HDPflat experiment as well as the dependencies between HL, FR, and FEA. Overall, we can assert that HDPflat achieved the best performance results on the ACE test dataset (Table 3), whereas HDPstruct proved to be more effective on the ECB dataset (Table 2). Moreover, the results of the HDPflat and HDPstruct models show an F-score increase by 4-10% over HDP1f, and therefore, the results prove that the HDP extension provides a more flexible representation for clustering objects with rich properties. We also plot the evolution of our generative processes. For instance, Figure 3(a) shows that the HDPflat model corresponding to row 7 in Table 3 converges in 350 iteration steps to a posterior distribution over event mentions from ACE with around 2000 latent events. Additionally, our experiments with different values of the λ parameter for the Lidstone’s smoothing method indicate that this smoothing method is useful for improving the performance of the HDP models. However, we could not find a λ value in our experiments that brings a major improvement over the non-smoothed HDP models. Figure3(b) shows the performances of HDPstruct on ECB with various λ values.8 The HDP results from Tables 2&3 correspond to a λ value of 10−4 and 10−2 for HDPflat and HDPstruct, respectively. iFHMM-iHMM In spite of the fact that the iFHMM-iHMM model employs automatic feature selection, its results remain competitive against the results of the HDP models, where the feature types were manually tuned. When comparing the strategies for filtering feature values in this framework, we could not find a distinct separation between the results obtained by the unfiltered, discrete, median, and uniform models. As observed from Tables 2&3, most of the iFHMMiHMM results fall in between the HDPflat and HDPstruct results. The results were obtained by automatically selecting only up to 1.5% of distinct feature values. Figure 3(c) shows the percents of features employed by this model for various values of the parameter α′ that controls the number of sampled features. The best results (also listed in Tables 2&3) were obtained for α′ = 10 (0.05%) on ACE and α′ = 150 (0.91%) on ECB. To show the usefulness of the sampling schemes considered for this model, we also compare in Table 4 the results obtained by an iFHMMiHMM model that considers all the feature values associated with an observable object (iFHMMiHMMall) against the iFHMM-iHMM models that employ the mIBP sampling scheme together with the unfiltered, discrete, median, and uniform filtering schemes. Because of the memory limitation constraints, we performed the experiments listed in Table 4 by selecting only a subset from 8 A configuration λ = 0 in the Lidstone’s smoothing method is equivalent with a non-smoothed version of the model on which it is applied. 1419 1000 1500 2000 2500 HDPflat | ACE | WD Number of categories 0 50 100 150 200 250 300 350 −4.5 −4 −3.5 −3 −2.5 x 10 5 Number of iterations Log−likelihood (a) 30 40 50 60 70 80 90 100 90.27 86.53 48.62 0 10−7 10−6 10−4 10−3 10−2 101 102 λ F1−measure HDPstruct | ECB | WD B3 CEAF PW (b) 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 10 50 100 150 200 250 0.07 0.32 0.63 0.91 1.20 1.47 α’ Number of feature values (%) iFHMM−iHMM | ECB | WD&CD (c) Figure 3: (a) Evolution of K and log-likelihood in the HDPflat model. (b) Evaluation of the Lidstone’s smoothing method in the HDPstruct model. (c) Counts of features employed by the iFHMM-iHMM model for various α′ values. Model B3 CEAF PW R P F R P F R P F ACE | WD all 89.3 39.8 55.0 30.2 68.8 42.0 62.7 9.1 15.9 unfiltered 83.3 77.7 80.4 70.6 75.9 73.2 42.1 34.6 38.0 discrete 83.8 80.7 82.2 73.0 75.8 74.4 43.9 39.1 41.4 median 83.5 80.2 81.8 72.2 75.3 73.7 42.7 38.2 40.3 uniform 82.8 80.7 81.7 72.8 75.2 73.9 41.4 39.3 40.3 ECB | WD all 89.5 62.5 73.6 53.3 76.5 62.8 60.7 22.9 33.2 unfiltered 82.6 96.6 89.0 92.0 79.1 85.1 28.4 75.6 41.0 discrete 83.1 96.7 89.4 91.6 79.2 84.9 30.5 79.0 43.9 median 82.5 97.3 89.3 92.8 78.9 85.3 29.2 78.8 42.0 uniform 82.7 96.0 88.9 91.1 79.0 84.6 29.3 74.9 41.6 ECB | CD all 79.3 54.4 64.5 43.3 61.3 50.7 59.6 26.2 36.4 unfiltered 67.2 94.5 78.5 84.7 59.2 69.6 32.8 82.5 46.8 discrete 67.6 94.8 78.9 83.8 58.3 68.8 34.3 85.3 48.9 median 66.7 95.2 78.4 84.5 57.7 68.5 32.2 83.7 46.3 uniform 67.7 93.6 78.4 83.6 59.2 69.2 33.6 79.5 46.9 Table 4: Feature non-sampling vs. feature sampling in the iFHMM-iHMM model. the feature types which proved to be salient in the HDP experiments. As listed in Table 4, all the iFHMM-iHMM models that used a feature sampling scheme significantly outperform the iFHMM-iHMMall model; this proves that all the sampling schemes considered in the iFHMMiHMM framework are able to successfully filter out noisy and redundant feature values. The closest comparison to prior work is the supervised approach described in (Chen and Ji, 2009) that achieved a 92.2% B3 F-measure on the ACE corpus. However, for this result, ground truth event mentions as well as a manually tuned coreference threshold were employed. 5 Error Analysis One frequent error occurs when a more complex form of semantic inference is needed to find a correspondence between two event mentions of the same individuated event. For instance, since all properties and participants of em3(deal) are omitted in our example and no common features exist between em3(buy) and em1(buy) to indicate a similarity between these mentions, they will most probably be assigned to different clusters. This example also suggests the need for a better modeling of the discourse salience for event mentions. Another common error is made when matching the semantic roles corresponding to coreferential event mentions. Although we simulated entity coreference by using various semantic features, the task of matching participants of coreferential event mentions is not completely solved. This is because, in many coreferential cases, partonomic relations between semantic roles need to be inferred.9 Examples of such relations extracted from ECB are Israeli forces PART OF −−−−→Israel, an Indian warship PART OF −−−−→the Indian navy, his cell PART OF −−−−→Sicilian jail. Similarly for event properties, many coreferential examples do not specify a clear location and time interval (e.g., Jabaliya refugee camp PART OF −−−−→Gaza, Tuesday PART OF −−−−→this week). In future work, we plan to build relevant clusters using partonomies and taxonomies such as the WordNet hierarchies built from MERONYMY/HOLONYMY and HYPERNYMY/HYPONYMY relations, respectively.10 6 Conclusion We have presented two novel, nonparametric Bayesian models that are designed to solve complex problems that require clustering objects characterized by a rich set of properties. Our experiments for event coreference resolution proved that these models are able to solve real data applications in which the feature and cluster numbers are treated as free parameters, and the selection of feature values is performed automatically. 9 This observation was also reported in (Hasler and Orasan, 2009). 10 This task is not trivial since, if applying the transitive closure on these relations, all words will end up being part from the same cluster with entity for instance. 1420 References ACE-Event. 2005. ACE (Automatic Content Extraction) English Annotation Guidelines for Events, version 5.4.3 2005.07.01. David Ahn. 2006. The stages of event extraction. In Proceedings of the Workshop on Annotating and Reasoning about Time and Events, pages 1–8. James Allan, Jaime Carbonell, George Doddington, Jonathan Yamron, and Yiming Yang. 1998. Topic Detection and Tracking Pilot Study: Final Report. In Proceedings of the Broadcast News Understanding and Transcription Workshop, pages 194–218. Amit Bagga and Breck Baldwin. 1998. Algorithms for Scoring Coreference Chains. In Proceedings of the 1st International Conference on Language Resources and Evaluation (LREC-1998). Amit Bagga and Breck Baldwin. 1999. CrossDocument Event Coreference: Annotations, Experiments, and Observations. In Proceedings of the ACL Workshop on Coreference and its Applications, pages 1–8. Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998. The Berkeley FrameNet project. In Proceedings of the 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics (COLING-ACL). Matthew J. Beal, Zoubin Ghahramani, and Carl Edward Rasmussen. 2002. The Infinite Hidden Markov Model. In Advances in Neural Information Processing Systems 14 (NIPS). Cosmin Adrian Bejan and Sanda Harabagiu. 2008. A Linguistic Resource for Discovering Event Structures and Resolving Event Coreference. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC). Cosmin Adrian Bejan and Chris Hathaway. 2007. UTD-SRL: A Pipeline Architecture for Extracting Frame Semantic Structures. In Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval), pages 460–463. Cosmin Adrian Bejan, Matthew Titsworth, Andrew Hickl, and Sanda Harabagiu. 2009. Nonparametric Bayesian Models for Unsupervised Event Coreference Resolution. In Advances in Neural Information Processing Systems 23 (NIPS). Cosmin Adrian Bejan. 2007. Deriving Chronological Information from Texts through a Graph-based Algorithm. In Proceedings of the 20th Florida Artificial Intelligence Research Society International Conference (FLAIRS), Applied Natural Language Processing track. Zheng Chen and Heng Ji. 2009. Graph-based Event Coreference Resolution. In Proceedings of the 2009 Workshop on Graph-based Methods for Natural Language Processing (TextGraphs-4), pages 54– 57. Donald Davidson, 1969. The Individuation of Events. In N. Rescher et al., eds., Essays in Honor of Carl G. Hempel, Dordrecht: Reidel. Reprinted in D. Davidson, ed., Essays on Actions and Events, 2001, Oxford: Clarendon Press. Donald Davidson, 1985. Reply to Quine on Events, pages 172–176. In E. LePore and B. McLaughlin, eds., Actions and Events: Perspectives on the Philosophy of Donald Davidson, Oxford: Blackwell. Marie-Catherine de Marneffe, Anna N. Rafferty, and Christopher D. Manning. 2008. Finding Contradictions in Text. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (ACLHLT), pages 1039–1047. Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database. MIT Press. Thomas S. Ferguson. 1973. A Bayesian Analysis of Some Nonparametric Problems. The Annals of Statistics, 1(2):209–230. Charles J. Fillmore. 1982. Frame Semantics. In Linguistics in the Morning Calm. Stuart Geman and Donald Geman. 1984. Stochastic relaxation, Gibbs distributions and the Bayesian restoration of images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 6:721–741. Zoubin Ghahramani, T. L. Griffiths, and Peter Sollich, 2007. Bayesian Statistics 8, chapter Bayesian nonparametric latent feature models, pages 201–225. Oxford University Press. Tom Griffiths and Zoubin Ghahramani. 2006. Infinite Latent Feature Models and the Indian Buffet Process. In Advances in Neural Information Processing Systems 18 (NIPS), pages 475–482. Aria Haghighi and Dan Klein. 2007. Unsupervised Coreference Resolution in a Nonparametric Bayesian Model. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics (ACL), pages 848–855. Aria Haghighi, Andrew Ng, and Christopher Manning. 2005. Robust Textual Inference via Graph Matching. In Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLTEMNLP), pages 387–394. Laura Hasler and Constantin Orasan. 2009. Do coreferential arguments make event mentions coreferential? In Proceedings of the 7th Discourse Anaphora and Anaphor Resolution Colloquium (DAARC 2009). 1421 Kevin Humphreys, Robert Gaizauskas, and Saliha Azzam. 1997. Event coreference for information extraction. In Proceedings of the Workshop on Operational Factors in Practical, Robust Anaphora Resolution for Unrestricted Texts, 35th Meeting of ACL, pages 75–81. John B. Lowe, Collin F. Baker, and Charles J. Fillmore. 1997. A frame-semantic approach to semantic annotation. In Proceedings of the SIGLEX Workshop on Tagging Text with Lexical Semantics: Why, What, and How?, pages 18–24. Xiaoqiang Luo. 2005. On coreference resolution performance metrics. In Proceedings of the Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (EMNLP-2005), pages 25–32. Jeff Malpas. 2009. Donald Davidson. In The Stanford Encyclopedia of Philosophy (Fall 2009 Edition), Edward N. Zalta (ed.), http://plato.stan ford.edu/archives/fall2009/entries/davidson/. Srini Narayanan and Sanda Harabagiu. 2004. Question Answering Based on Semantic Structures. In Proceedings of the 20th International Conference on Computational Linguistics (COLING), pages 693– 701. Radford M. Neal. 2003. Slice Sampling. The Annals of Statistics, 31:705–741. Vincent Ng. 2008. Unsupervised Models for Coreference Resolution. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 640–649. Martha Palmer, Daniel Gildea, and Paul Kingsbury. 2005. The Proposition Bank: An Annotated Corpus of Semantic Roles. Computational Linguistics, 31(1):71–105. Hoifung Poon and Pedro Domingos. 2008. Joint Unsupervised Coreference Resolution with Markov Logic. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 650–659. James Pustejovsky, Jose Castano, Bob Ingria, Roser Sauri, Rob Gaizauskas, Andrea Setzer, and Graham Katz. 2003a. TimeML: Robust Specification of Event and Temporal Expressions in Text. In Proceedings of the Fifth International Workshop on Computational Semantics (IWCS). James Pustejovsky, Patrick Hanks, Roser Sauri, Andrew See, Robert Gaizauskas, Andrea Setzer, Dragomir Radev, Beth Sundheim, David Day, Lisa Ferro, and Marcia Lazo. 2003b. The TimeBank Corpus. In Corpus Linguistics, pages 647–656. W. V. O. Quine, 1985. Events and Reification, pages 162–171. In E. LePore and B. P. McLaughlin, eds., Actions and Events: Perspectives on the philosophy of Donald Davidson, Oxford: Blackwell. Reprinted in R. Casati and A. C. Varzi, eds., Events, 1996, pages 107–116, Aldershot: Dartmouth. Lawrence R. Rabiner. 1989. A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition. In Proceedings of the IEEE, pages 257–286. Yee Whye Teh, Michael Jordan, Matthew Beal, and David Blei. 2006. Hierarchical Dirichlet Processes. Journal of the American Statistical Association, 101(476):1566–1581. Jurgen Van Gael, Y. Saatci, Yee Whye Teh, and Zoubin Ghahramani. 2008a. Beam Sampling for the Infinite Hidden Markov Model. In Proceedings of the 25th Annual International Conference on Machine Learning (ICML), pages 1088–1095. Jurgen Van Gael, Yee Whye Teh, and Zoubin Ghahramani. 2008b. The Infinite Factorial Hidden Markov Model. In Advances in Neural Information Processing Systems 21 (NIPS). 1422
|
2010
|
143
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1423–1432, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Coreference Resolution across Corpora: Languages, Coding Schemes, and Preprocessing Information Marta Recasens CLiC - University of Barcelona Gran Via 585 Barcelona, Spain [email protected] Eduard Hovy USC Information Sciences Institute 4676 Admiralty Way Marina del Rey CA, USA [email protected] Abstract This paper explores the effect that different corpus configurations have on the performance of a coreference resolution system, as measured by MUC, B3, and CEAF. By varying separately three parameters (language, annotation scheme, and preprocessing information) and applying the same coreference resolution system, the strong bonds between system and corpus are demonstrated. The experiments reveal problems in coreference resolution evaluation relating to task definition, coding schemes, and features. They also expose systematic biases in the coreference evaluation metrics. We show that system comparison is only possible when corpus parameters are in exact agreement. 1 Introduction The task of coreference resolution, which aims to automatically identify the expressions in a text that refer to the same discourse entity, has been an increasing research topic in NLP ever since MUC-6 made available the first coreferentially annotated corpus in 1995. Most research has centered around the rules by which mentions are allowed to corefer, the features characterizing mention pairs, the algorithms for building coreference chains, and coreference evaluation methods. The surprisingly important role played by different aspects of the corpus, however, is an issue to which little attention has been paid. We demonstrate the extent to which a system will be evaluated as performing differently depending on parameters such as the corpus language, the way coreference relations are defined in the corresponding coding scheme, and the nature and source of preprocessing information. This paper unpacks these issues by running the same system—a prototype entity-based architecture called CISTELL—on different corpus configurations, varying three parameters. First, we show how much language-specific issues affect performance when trained and tested on English and Spanish. Second, we demonstrate the extent to which the specific annotation scheme (used on the same corpus) makes evaluated performance vary. Third, we compare the performance using goldstandard preprocessing information with that using automatic preprocessing tools. Throughout, we apply the three principal coreference evaluation measures in use today: MUC, B3, and CEAF. We highlight the systematic preferences of each measure to reward different configurations. This raises the difficult question of why one should use one or another evaluation measure, and how one should interpret their differences in reporting changes of performance score due to ‘secondary’ factors like preprocessing information. To this end, we employ three corpora: ACE (Doddington et al., 2004), OntoNotes (Pradhan et al., 2007), and AnCora (Recasens and Mart´ı, 2009). In order to isolate the three parameters as far as possible, we benefit from a 100k-word portion (from the TDT collection) that is common to both ACE and OntoNotes. We apply the same coreference resolution system in all cases. The results show that a system’s score is not informative by itself, as different corpora or corpus parameters lead to different scores. Our goal is not to achieve the best performance to date, but rather to expose various issues raised by the choices of corpus preparation and evaluation measure and to shed light on the definition, methods, evaluation, and complexities of the coreference resolution task. The paper is organized as follows. Section 2 sets our work in context and provides the motivations for undertaking this study. Section 3 presents the architecture of CISTELL, the system used in the experimental evaluation. In Sections 4, 5, 1423 and 6, we describe the experiments on three different datasets and discuss the results. We conclude in Section 7. 2 Background The bulk of research on automatic coreference resolution to date has been done for English and used two different types of corpus: MUC (Hirschman and Chinchor, 1997) and ACE (Doddington et al., 2004). A variety of learning-based systems have been trained and tested on the former (Soon et al., 2001; Uryupina, 2006), on the latter (Culotta et al., 2007; Bengtson and Roth, 2008; Denis and Baldridge, 2009), or on both (Finkel and Manning, 2008; Haghighi and Klein, 2009). Testing on both is needed given that the two annotation schemes differ in some aspects. For example, only ACE includes singletons (mentions that do not corefer) and ACE is restricted to seven semantic types.1 Also, despite a critical discussion in the MUC task definition (van Deemter and Kibble, 2000), the ACE scheme continues to treat nominal predicates and appositive phrases as coreferential. A third coreferentially annotated corpus—the largest for English—is OntoNotes (Pradhan et al., 2007; Hovy et al., 2006). Unlike ACE, it is not application-oriented, so coreference relations between all types of NPs are annotated. The identity relation is kept apart from the attributive relation, and it also contains gold-standard morphological, syntactic and semantic information. Since the MUC and ACE corpora are annotated with only coreference information,2 existing systems first preprocess the data using automatic tools (POS taggers, parsers, etc.) to obtain the information needed for coreference resolution. However, given that the output from automatic tools is far from perfect, it is hard to determine the level of performance of a coreference module acting on gold-standard preprocessing information. OntoNotes makes it possible to separate the coreference resolution problem from other tasks. Our study adds to the previously reported evidence by Stoyanov et al. (2009) that differences in corpora and in the task definitions need to be taken into account when comparing coreference resolution systems. We provide new insights as the current analysis differs in four ways. First, Stoyanov 1The ACE-2004/05 semantic types are person, organization, geo-political entity, location, facility, vehicle, weapon. 2ACE also specifies entity types and relations. et al. (2009) report on differences between MUC and ACE, while we contrast ACE and OntoNotes. Given that ACE and OntoNotes include some of the same texts but annotated according to their respective guidelines, we can better isolate the effect of differences as well as add the additional dimension of gold preprocessing. Second, we evaluate not only with the MUC and B3 scoring metrics, but also with CEAF. Third, all our experiments use true mentions3 to avoid effects due to spurious system mentions. Finally, including different baselines and variations of the resolution model allows us to reveal biases of the metrics. Coreference resolution systems have been tested on languages other than English only within the ACE program (Luo and Zitouni, 2005), probably due to the fact that coreferentially annotated corpora for other languages are scarce. Thus there has been no discussion of the extent to which systems are portable across languages. This paper studies the case of English and Spanish.4 Several coreference systems have been developed in the past (Culotta et al., 2007; Finkel and Manning, 2008; Poon and Domingos, 2008; Haghighi and Klein, 2009; Ng, 2009). It is not our aim to compete with them. Rather, we conduct three experiments under a specific setup for comparison purposes. To this end, we use a different, neutral, system, and a dataset that is small and different from official ACE test sets despite the fact that it prevents our results from being compared directly with other systems. 3 Experimental Setup 3.1 System Description The system architecture used in our experiments, CISTELL, is based on the incrementality of discourse. As a discourse evolves, it constructs a model that is updated with the new information gradually provided. A key element in this model are the entities the discourse is about, as they form the discourse backbone, especially those that are mentioned multiple times. Most entities, however, are only mentioned once. Consider the growth of the entity Mount Popocat´epetl in (1).5 3The adjective true contrasts with system and refers to the gold standard. 4Multilinguality is one of the focuses of SemEval-2010 Task 1 (Recasens et al., 2010). 5Following the ACE terminology, we use the term mention for an instance of reference to an object, and entity for a collection of mentions referring to the same object. Entities 1424 (1) We have an update tonight on [this, the volcano in Mexico, they call El Popo]m3 ...As the sun rises over [Mt. Popo]m7 tonight, the only hint of the fire storm inside, whiffs of smoke, but just a few hours earlier, [the volcano]m11 exploding spewing rock and red-hot lava. [The fourth largest mountain in North America, nearly 18,000 feet high]m15, erupting this week with [its]m20 most violent outburst in 1,200 years. Mentions can be pronouns (m20), they can be a (shortened) string repetition using either the name (m7) or the type (m11), or they can add new information about the entity: m15 provides the supertype and informs the reader about the height of the volcano and its ranking position. In CISTELL,6 discourse entities are conceived as ‘baskets’: they are empty at the beginning of the discourse, but keep growing as new attributes (e.g., name, type, location) are predicated about them. Baskets are filled with this information, which can appear within a mention or elsewhere in the sentence. The ever-growing amount of information in a basket allows richer comparisons to new mentions encountered in the text. CISTELL follows the learning-based coreference architecture in which the task is split into classification and clustering (Soon et al., 2001; Bengtson and Roth, 2008) but combines them simultaneously. Clustering is identified with basketgrowing, the core process, and a pairwise classifier is called every time CISTELL considers whether a basket must be clustered into a (growing) basket, which might contain one or more mentions. We use a memory-based learning classifier trained with TiMBL (Daelemans and Bosch, 2005). Basket-growing is done in four different ways, explained next. 3.2 Baselines and Models In each experiment, we compute three baselines (1, 2, 3), and run CISTELL under four different models (4, 5, 6, 7). 1. ALL SINGLETONS. No coreference link is ever created. We include this baseline given the high number of singletons in the datasets, since some evaluation measures are affected by large numbers of singletons. 2. HEAD MATCH. All non-pronominal NPs that have the same head are clustered into the same entity. containing one single mention are referred to as singletons. 6‘Cistell’ is the Catalan word for ‘basket.’ 3. HEAD MATCH + PRON. Like HEAD MATCH, plus allowing personal and possessive pronouns to link to the closest noun with which they agree in gender and number. 4. STRONG MATCH. Each mention (e.g., m11) is paired with previous mentions starting from the beginning of the document (m1–m11, m2– m11, etc.).7 When a pair (e.g., m3–m11) is classified as coreferent, additional pairwise checks are performed with all the mentions contained in the (growing) entity basket (e.g., m7–m11). Only if all the pairs are classified as coreferent is the mention under consideration attached to the existing growing entity. Otherwise, the search continues.8 5. SUPER STRONG MATCH. Similar to STRONG MATCH but with a threshold. Coreference pairwise classifications are only accepted when TiMBL distance is smaller than 0.09.9 6. BEST MATCH. Similar to STRONG MATCH but following Ng and Cardie (2002)’s best link approach. Thus, the mention under analysis is linked to the most confident mention among the previous ones, using TiMBL’s confidence score. 7. WEAK MATCH. A simplified version of STRONG MATCH: not all mentions in the growing entity need to be classified as coreferent with the mention under analysis. A single positive pairwise decision suffices for the mention to be clustered into that entity.10 3.3 Features We follow Soon et al. (2001), Ng and Cardie (2002) and Luo et al. (2004) to generate most of the 29 features we use for the pairwise model. These include features that capture information from different linguistic levels: textual strings (head match, substring match, distance, frequency), morphology (mention type, coordination, possessive phrase, gender match, number match), syntax (nominal predicate, apposition, relative clause, grammatical function), and semantic match (named-entity type, is-a type, supertype). 7The opposite search direction was also tried but gave worse results. 8Taking the first mention classified as coreferent follows Soon et al. (2001)’s first-link approach. 9In TiMBL, being a memory-based learner, the closer the distance to an instance, the more confident the decision. We chose 0.09 because it appeared to offer the best results. 10STRONG and WEAK MATCH are similar to Luo et al. (2004)’s entity-mention and mention-pair models. 1425 For Spanish, we use 34 features as a few variations are needed for language-specific issues such as zero subjects (Recasens and Hovy, 2009). 3.4 Evaluation Since they sometimes provide quite different results, we evaluate using three coreference measures, as there is no agreement on a standard. • MUC (Vilain et al., 1995). It computes the number of links common between the true and system partitions. Recall (R) and precision (P) result from dividing it by the minimum number of links required to specify the true and the system partitions, respectively. • B3 (Bagga and Baldwin, 1998). R and P are computed for each mention and averaged at the end. For each mention, the number of common mentions between the true and the system entity is divided by the number of mentions in the true entity or in the system entity to obtain R and P, respectively. • CEAF (Luo, 2005). It finds the best one-toone alignment between true and system entities. Using true mentions and the φ3 similarity function, R and P are the same and correspond to the number of common mentions between the aligned entities divided by the total number of mentions. 4 Parameter 1: Language The first experiment compared the performance of a coreference resolution system on a Germanic and a Romance language—English and Spanish— to explore to what extent language-specific issues such as zero subjects11 or grammatical gender might influence a system. Although OntoNotes and AnCora are two different corpora, they are very similar in those aspects that matter most for the study’s purpose: they both include a substantial amount of texts belonging to the same genre (news) and manually annotated from the morphological to the semantic levels (POS tags, syntactic constituents, NEs, WordNet synsets, and coreference relations). More importantly, very similar coreference annotation guidelines make AnCora the ideal Spanish counterpart to OntoNotes. 11Most Romance languages are pro-drop allowing zero subject pronouns, which can be inferred from the verb. Datasets Two datasets of similar size were selected from AnCora and OntoNotes in order to rule out corpus size as an explanation of any difference in performance. Corpus statistics about the distribution of mentions and entities are shown in Tables 1 and 2. Given that this paper is focused on coreference between NPs, the number of mentions only includes NPs. Both AnCora and OntoNotes annotate only multi-mention entities (i.e., those containing two or more coreferent mentions), so singleton entities are assumed to correspond to NPs with no coreference annotation. Apart from a larger number of mentions in Spanish (Table 1), the two datasets look very similar in the distribution of singletons and multimention entities: about 85% and 15%, respectively. Multi-mention entities have an average of 3.9 mentions per entity in AnCora and 3.5 in OntoNotes. The distribution of mention types (Table 2), however, differs in two important respects: AnCora has a smaller number of personal pronouns as Spanish typically uses zero subjects, and it has a smaller number of bare NPs as the definite article accompanies more NPs than in English. Results and Discussion Table 3 presents CISTELL’s results for each dataset. They make evident problems with the evaluation metrics, namely the fact that the generated rankings are contradictory (Denis and Baldridge, 2009). They are consistent across the two corpora though: MUC rewards WEAK MATCH the most, B3 rewards HEAD MATCH the most, and CEAF is divided between SUPER STRONG MATCH and BEST MATCH. These preferences seem to reveal weaknesses of the scoring methods that make them biased towards a type of output. The model preferred by MUC is one that clusters many mentions together, thus getting a large number of correct coreference links (notice the high R for WEAK MATCH), but AnCora OntoNotes Pronouns 14.09 17.62 Personal pronouns 2.00 12.10 Zero subject pronouns 6.51 – Possessive pronouns 3.57 2.96 Demonstrative pronouns 0.39 1.83 Definite NPs 37.69 20.67 Indefinite NPs 7.17 8.44 Demonstrative NPs 1.98 3.41 Bare NPs 33.02 42.92 Misc. 6.05 6.94 Table 2: Mention types (%) in Table 1 datasets. 1426 #docs #words #mentions #entities (e) #singleton e #multi-mention e AnCora Training 955 299,014 91,904 64,535 54,991 9,544 Test 30 9,851 2,991 2,189 1,877 312 OntoNotes Training 850 301,311 74,692 55,819 48,199 7,620 Test 33 9,763 2,463 1,790 1,476 314 Table 1: Corpus statistics for the large portion of OntoNotes and AnCora. MUC B3 CEAF P R F P R F P / R / F AnCora - Spanish 1. ALL SINGLETONS – – – 100 73.32 84.61 73.32 2. HEAD MATCH 55.03 37.72 44.76 91.12 79.88 85.13 75.96 3. HEAD MATCH + PRON 48.22 44.24 46.14 86.21 80.66 83.34 76.30 4. STRONG MATCH 45.64 51.88 48.56 80.13 82.28 81.19 75.79 5. SUPER STRONG MATCH 45.68 36.47 40.56 86.10 79.09 82.45 77.20 6. BEST MATCH 43.10 35.59 38.98 85.24 79.67 82.36 75.23 7. WEAK MATCH 45.73 65.16 53.75 68.50 87.71 76.93 69.21 OntoNotes - English 1. ALL SINGLETONS – – – 100 72.68 84.18 72.68 2. HEAD MATCH 55.14 39.08 45.74 90.65 80.87 85.48 76.05 3. HEAD MATCH + PRON 47.10 53.05 49.90 82.28 83.13 82.70 75.15 4. STRONG MATCH 47.94 55.42 51.41 81.13 84.30 82.68 78.03 5. SUPER STRONG MATCH 48.27 47.55 47.90 84.00 82.27 83.13 78.24 6. BEST MATCH 50.97 46.66 48.72 86.19 82.70 84.41 78.44 7. WEAK MATCH 47.46 66.72 55.47 70.36 88.05 78.22 71.21 Table 3: CISTELL results varying the corpus language. also many spurious links that are not duly penalized. The resulting output is not very desirable.12 In contrast, B3 is more P-oriented and scores conservative outputs like HEAD MATCH and BEST MATCH first, even if R is low. CEAF achieves a better compromise between P and R, as corroborated by the quality of the output. The baselines and the system runs perform very similarly in the two corpora, but slightly better for English. It seems that language-specific issues do not result in significant differences—at least for English and Spanish—once the feature set has been appropriately adapted, e.g., including features about zero subjects or removing those about possessive phrases. Comparing the feature ranks, we find that the features that work best for each language largely overlap and are language independent, like head match, is-a match, and whether the mentions are pronominal. 5 Parameter 2: Annotation Scheme In the second experiment, we used the 100k-word portion (from the TDT collection) shared by the OntoNotes and ACE corpora (330 OntoNotes doc12Due to space constraints, the actual output cannot be shown here. We are happy to send it to interested requesters. uments occurred as 22 ACE-2003 documents, 185 ACE-2004 documents, and 123 ACE-2005 documents). CISTELL was trained on the same texts in both corpora and applied to the remainder. The three measures were then applied to each result. Datasets Since the two annotation schemes differ significantly, we made the results comparable by mapping the ACE entities (the simpler scheme) onto the information contained in OntoNotes.13 The mapping allowed us to focus exclusively on the differences expressed on both corpora: the types of mentions that were annotated, the definition of identity of reference, etc. Table 4 presents the statistics for the OntoNotes dataset merged with the ACE entities. The mapping was not straightforward due to several problems: there was no match for some mentions due to syntactic or spelling reasons (e.g., El Popo in OntoNotes vs. Ell Popo in ACE). ACE mentions for which there was no parse tree node in the OntoNotes gold-standard tree were omitted, as creating a new node could have damaged the tree. Given that only seven entity types are annotated in ACE, the number of OntoNotes mentions is al13Both ACE entities and types were mapped onto the OntoNotes dataset. 1427 #docs #words #mentions #entities (e) #singleton e #multi-mention e OntoNotes Training 297 87,068 22,127 15,983 13,587 2,396 Test 33 9,763 2,463 1,790 1,476 314 ACE Training 297 87,068 12,951 5,873 3,599 2,274 Test 33 9,763 1,464 746 459 287 Table 4: Corpus statistics for the aligned portion of ACE and OntoNotes on gold-standard data. MUC B3 CEAF P R F P R F P / R / F OntoNotes scheme 1. ALL SINGLETONS – – – 100 72.68 84.18 72.68 2. HEAD MATCH 55.14 39.08 45.74 90.65 80.87 85.48 76.05 3. HEAD MATCH + PRON 47.10 53.05 49.90 82.28 83.13 82.70 75.15 4. STRONG MATCH 46.81 53.34 49.86 80.47 83.54 81.97 76.78 5. SUPER STRONG MATCH 46.51 40.56 43.33 84.95 80.16 82.48 76.70 6. BEST MATCH 52.47 47.40 49.80 86.10 82.80 84.42 77.87 7. WEAK MATCH 47.91 64.64 55.03 71.73 87.46 78.82 71.74 ACE scheme 1. ALL SINGLETONS – – – 100 50.96 67.51 50.96 2. HEAD MATCH 82.35 39.00 52.93 95.27 64.05 76.60 66.46 3. HEAD MATCH + PRON 70.11 53.90 60.94 86.49 68.20 76.27 68.44 4. STRONG MATCH 64.21 64.21 64.21 76.92 73.54 75.19 70.01 5. SUPER STRONG MATCH 60.51 56.55 58.46 76.71 69.19 72.76 66.87 6. BEST MATCH 67.50 56.69 61.62 82.18 71.67 76.57 69.88 7. WEAK MATCH 63.52 80.50 71.01 59.76 86.36 70.64 64.21 Table 5: CISTELL results varying the annotation scheme on gold-standard data. most twice as large as the number of ACE mentions. Unlike OntoNotes, ACE mentions include premodifiers (e.g., state in state lines), national adjectives (e.g., Iraqi) and relative pronouns (e.g., who, that). Also, given that ACE entities correspond to types that are usually coreferred (e.g., people, organizations, etc.), singletons only represent 61% of all entities, while they are 85% in OntoNotes. The average entity size is 4 in ACE and 3.5 in OntoNotes. A second major difference is the definition of coreference relations, illustrated here: (2) [This] was [an all-white, all-Christian community that all the sudden was taken over ... by different groups]. (3) [ [Mayor] John Hyman] has a simple answer. (4) [Postville] now has 22 different nationalities ... For those who prefer [the old Postville], Mayor John Hyman has a simple answer. In ACE, nominal predicates corefer with their subject (2), and appositive phrases corefer with the noun they are modifying (3). In contrast, they do not fall under the identity relation in OntoNotes, which follows the linguistic understanding of coreference according to which nominal predicates and appositives express properties of an entity rather than refer to a second (coreferent) entity (van Deemter and Kibble, 2000). Finally, the two schemes frequently disagree on borderline cases in which coreference turns out to be especially complex (4). As a result, some features will behave differently, e.g., the appositive feature has the opposite effect in the two datasets. Results and Discussion From the differences pointed out above, the results shown in Table 5 might be surprising at first. Given that OntoNotes is not restricted to any semantic type and is based on a more sophisticated definition of coreference, one would not expect a system to perform better on it than on ACE. The explanation is given by the ALL SINGLETONS baseline, which is 73–84% for OntoNotes and only 51–68% for ACE. The fact that OntoNotes contains a much larger number of singletons—as Table 4 shows—results in an initial boost of performance (except with the MUC score, which ignores singletons). In contrast, the score improvement achieved by HEAD MATCH is much more noticeable on ACE than on OntoNotes, which indicates that many of its coreferent mentions share the same head. The systematic biases of the measures that were observed in Table 3 appear again in the case of 1428 MUC and B3. CEAF is divided between BEST MATCH and STRONG MATCH. The higher value of the MUC score for ACE is another indication of its tendency to reward correct links much more than to penalize spurious ones (ACE has a larger proportion of multi-mention entities). The feature rankings obtained for each dataset generally coincide as to which features are ranked best (namely NE match, is-a match, and head match), but differ in their particular ordering. It is also possible to compare the OntoNotes results in Tables 3 and 5, the only difference being that the first training set was three times larger. Contrary to expectation, the model trained on a larger dataset performs just slightly better. The fact that more training data does not necessarily lead to an increase in performance conforms to the observation that there appear to be few general rules (e.g., head match) that systematically govern coreference relationships; rather, coreference appeals to individual unique phenomena appearing in each context, and thus after a point adding more training data does not add much new generalizable information. Pragmatic information (discourse structure, world knowledge, etc.) is probably the key, if ever there is a way to encode it. 6 Parameter 3: Preprocessing The goal of the third experiment was to determine how much the source and nature of preprocessing information matters. Since it is often stated that coreference resolution depends on many levels of analysis, we again compared the two corpora, which differ in the amount and correctness of such information. However, in this experiment, entity mapping was applied in the opposite direction: the OntoNotes entities were mapped onto the automatically preprocessed ACE dataset. This exposes the shortcomings of automated preprocessing in ACE for identifying all the mentions identified and linked in OntoNotes. Datasets The ACE data was morphologically annotated with a tokenizer based on manual rules adapted from the one used in CoNLL (Tjong Kim Sang and De Meulder, 2003), with TnT 2.2, a trigram POS tagger based on Markov models (Brants, 2000), and with the built-in WordNet lemmatizer (Fellbaum, 1998). Syntactic chunks were obtained from YamCha 1.33, an SVM-based NPchunker (Kudoh and Matsumoto, 2000), and parse trees from Malt Parser 0.4, an SVM-based parser (Hall et al., 2007). Although the number of words in Tables 4 and 6 should in principle be the same, the latter contains fewer words as it lacks the null elements (traces, ellipsed material, etc.) manually annotated in OntoNotes. Missing parse tree nodes in the automatically parsed data account for the considerably lower number of OntoNotes mentions (approx. 5,700 fewer mentions).14 However, the proportions of singleton:multi-mention entities as well as the average entity size do not vary. Results and Discussion The ACE scores for the automatically preprocessed models in Table 7 are about 3% lower than those based on OntoNotes gold-standard data in Table 5, providing evidence for the advantage offered by gold-standard preprocessing information. In contrast, the similar—if not higher—scores of OntoNotes can be attributed to the use of the annotated ACE entity types. The fact that these are annotated not only for proper nouns (as predicted by an automatic NER) but also for pronouns and full NPs is a very helpful feature for a coreference resolution system. Again, the scoring metrics exhibit similar biases, but note that CEAF prefers HEAD MATCH + PRON in the case of ACE, which is indicative of the noise brought by automatic preprocessing. A further insight is offered from comparing the feature rankings with gold-standard syntax to that with automatic preprocessing. Since we are evaluating now on the ACE data, the NE match feature is also ranked first for OntoNotes. Head and is-a match are still ranked among the best, yet syntactic features are not. Instead, features like NP type have moved further up. This reranking probably indicates that if there is noise in the syntactic information due to automatic tools, then morphological and syntactic features switch their positions. Given that the noise brought by automatic preprocessing can be harmful, we tried leaving out the grammatical function feature. Indeed, the results increased about 2–3%, STRONG MATCH scoring the highest. This points out that conclusions drawn from automatically preprocessed data about the kind of knowledge relevant for coreference resolution might be mistaken. Using the most successful basic features can lead to the best results when only automatic preprocessing is available. 14In order to make the set of mentions as similar as possible to the set in Section 5, OntoNotes singletons were mapped from the ones detected in the gold-standard treebank. 1429 #docs #words #mentions #entities (e) #singleton e #multi-mention e OntoNotes Training 297 80,843 16,945 12,127 10,253 1,874 Test 33 9,073 1,931 1,403 1,156 247 ACE Training 297 80,843 13,648 6,041 3,652 2,389 Test 33 9,073 1,537 775 475 300 Table 6: Corpus statistics for the aligned portion of ACE and OntoNotes on automatically parsed data. MUC B3 CEAF P R F P R F P / R / F OntoNotes scheme 1. ALL SINGLETONS – – – 100 72.66 84.16 72.66 2. HEAD MATCH 56.76 35.80 43.90 92.18 80.52 85.95 76.33 3. HEAD MATCH + PRON 47.44 54.36 50.66 82.08 83.61 82.84 74.83 4. STRONG MATCH 52.66 58.14 55.27 83.11 85.05 84.07 78.30 5. SUPER STRONG MATCH 51.67 46.78 49.11 85.74 82.07 83.86 77.67 6. BEST MATCH 54.38 51.70 53.01 86.00 83.60 84.78 78.15 7. WEAK MATCH 49.78 64.58 56.22 75.63 87.79 81.26 74.62 ACE scheme 1. ALL SINGLETONS – – – 100 50.42 67.04 50.42 2. HEAD MATCH 81.25 39.24 52.92 94.73 63.82 76.26 65.97 3. HEAD MATCH + PRON 69.76 53.28 60.42 86.39 67.73 75.93 68.05 4. STRONG MATCH 58.85 58.92 58.89 73.36 70.35 71.82 66.30 5. SUPER STRONG MATCH 56.19 50.66 53.28 75.54 66.47 70.72 63.96 6. BEST MATCH 63.38 49.74 55.74 80.97 68.11 73.99 65.97 7. WEAK MATCH 60.22 78.48 68.15 55.17 84.86 66.87 59.08 Table 7: CISTELL results varying the annotation scheme on automatically preprocessed data. 7 Conclusion Regarding evaluation, the results clearly expose the systematic tendencies of the evaluation measures. The way each measure is computed makes it biased towards a specific model: MUC is generally too lenient with spurious links, B3 scores too high in the presence of a large number of singletons, and CEAF does not agree with either of them. It is a cause for concern that they provide contradictory indications about the core of coreference, namely the resolution models—for example, the model ranked highest by B3 in Table 7 is ranked lowest by MUC. We always assume evaluation measures provide a ‘true’ reflection of our approximation to a gold standard in order to guide research in system development and tuning. Further support to our claims comes from the results of SemEval-2010 Task 1 (Recasens et al., 2010). The performance of the six participating systems shows similar problems with the evaluation metrics, and the singleton baseline was hard to beat even by the highest-performing systems. Since the measures imply different conclusions about the nature of the corpora and the preprocessing information applied, should we use them now to constrain the ways our corpora are created in the first place, and what preprocessing we include or omit? Doing so would seem like circular reasoning: it invalidates the notion of the existence of a true and independent gold standard. But if apparently incidental aspects of the corpora can have such effects—effects rated quite differently by the various measures—then we have no fixed ground to stand on. The worrisome fact that there is currently no clearly preferred and ‘correct’ evaluation measure for coreference resolution means that we cannot draw definite conclusions about coreference resolution systems at this time, unless they are compared on exactly the same corpus, preprocessed under the same conditions, and all three measures agree in their rankings. Acknowledgments We thank Dr. M. Ant`onia Mart´ı for her generosity in allowing the first author to visit ISI to work with the second. Special thanks to Edgar Gonz`alez for his kind help with conversion issues. This work was partially supported by the Spanish Ministry of Education through an FPU scholarship (AP2006-00994) and the TEXT-MESS 2.0 Project (TIN2009-13391-C04-04). 1430 References Amit Bagga and Breck Baldwin. 1998. Algorithms for scoring coreference chains. In Proceedings of the LREC 1998 Workshop on Linguistic Coreference, pages 563–566, Granada, Spain. Eric Bengtson and Dan Roth. 2008. Understanding the value of features for coreference resolution. In Proceedings of EMNLP 2008, pages 294–303, Honolulu, Hawaii. Thorsten Brants. 2000. TnT – A statistical part-ofspeech tagger. In Proceedings of ANLP 2000, Seattle, WA. Aron Culotta, Michael Wick, Robert Hall, and Andrew McCallum. 2007. First-order probabilistic models for coreference resolution. In Proceedings of HLTNAACL 2007, pages 81–88, Rochester, New York. Walter Daelemans and Antal Van den Bosch. 2005. Memory-Based Language Processing. Cambridge University Press. Pascal Denis and Jason Baldridge. 2009. Global joint models for coreference resolution and named entity classification. Procesamiento del Lenguaje Natural, 42:87–96. George Doddington, Alexis Mitchell, Mark Przybocki, Lance Ramshaw, Stephanie Strassel, and Ralph Weischedel. 2004. The Automatic Content Extraction (ACE) Program - Tasks, Data, and Evaluation. In Proceedings of LREC 2004, pages 837–840. Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database. The MIT Press. Jenny Rose Finkel and Christopher D. Manning. 2008. Enforcing transitivity in coreference resolution. In Proceedings of ACL-HLT 2008, pages 45– 48, Columbus, Ohio. Aria Haghighi and Dan Klein. 2009. Simple coreference resolution with rich syntactic and semantic features. In Proceedings of EMNLP 2009, pages 1152–1161, Singapore. Association for Computational Linguistics. Johan Hall, Jens Nilsson, Joakim Nivre, G¨ulsen Eryigit, Be´ata Megyesi, Mattias Nilsson, and Markus Saers. 2007. Single malt or blended? A study in multilingual parser optimization. In Proceedings of the CoNLL shared task session of EMNLP-CoNLL 2007, pages 933–939. Lynette Hirschman and Nancy Chinchor. 1997. MUC7 Coreference Task Definition – Version 3.0. In Proceedings of MUC-7. Eduard Hovy, Mitchell Marcus, Martha Palmer, Lance Ramshaw, and Ralph Weischedel. 2006. OntoNotes: the 90% solution. In Proceedings of HLT-NAACL 2006, pages 57–60. Taku Kudoh and Yuji Matsumoto. 2000. Use of support vector learning for chunk identification. In Proceedings of CoNLL 2000 and LLL 2000, pages 142– 144, Lisbon, Portugal. Xiaoqiang Luo and Imed Zitouni. 2005. Multi-lingual coreference resolution with syntactic features. In Proceedings of HLT-EMNLP 2005, pages 660–667, Vancouver. Xiaoqiang Luo, Abe Ittycheriah, Hongyan Jing, Nanda Kambhatla, and Salim Roukos. 2004. A mentionsynchronous coreference resolution algorithm based on the Bell tree. In Proceedings of ACL 2004, pages 21–26, Barcelona. Xiaoqiang Luo. 2005. On coreference resolution performance metrics. In Proceedings of HLTEMNLP 2005, pages 25–32, Vancouver. Vincent Ng and Claire Cardie. 2002. Improving machine learning approaches to coreference resolution. In Proceedings of ACL 2002, pages 104–111, Philadelphia. Vincent Ng. 2009. Graph-cut-based anaphoricity determination for coreference resolution. In Proceedings of NAACL-HLT 2009, pages 575–583, Boulder, Colorado. Hoifung Poon and Pedro Domingos. 2008. Joint unsupervised coreference resolution with Markov logic. In Proceedings of EMNLP 2008, pages 650–659, Honolulu, Hawaii. Sameer S. Pradhan, Eduard Hovy, Mitch Marcus, Martha Palmer, Lance Ramshaw, and Ralph Weischedel. 2007. Ontonotes: A unified relational semantic representation. In Proceedings of ICSC 2007, pages 517–526, Washington, DC. Marta Recasens and Eduard Hovy. 2009. A Deeper Look into Features for Coreference Resolution. In S. Lalitha Devi, A. Branco, and R. Mitkov, editors, Anaphora Processing and Applications (DAARC 2009), volume 5847 of LNAI, pages 29–42. Springer-Verlag. Marta Recasens and M. Ant`onia Mart´ı. 2009. AnCoraCO: Coreferentially annotated corpora for Spanish and Catalan. Language Resources and Evaluation, DOI 10.1007/s10579-009-9108-x. Marta Recasens, Llu´ıs M`arquez, Emili Sapena, M. Ant`onia Mart´ı, Mariona Taul´e, V´eronique Hoste, Massimo Poesio, and Yannick Versley. 2010. SemEval-2010 Task 1: Coreference resolution in multiple languages. In Proceedings of the Fifth International Workshop on Semantic Evaluations (SemEval 2010), Uppsala, Sweden. Wee M. Soon, Hwee T. Ng, and Daniel C. Y. Lim. 2001. A machine learning approach to coreference resolution of noun phrases. Computational Linguistics, 27(4):521–544. 1431 Veselin Stoyanov, Nathan Gilbert, Claire Cardie, and Ellen Riloff. 2009. Conundrums in noun phrase coreference resolution: Making sense of the stateof-the-art. In Proceedings of ACL-IJCNLP 2009, pages 656–664, Singapore. Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 Shared Task: Language-independent Named Entity Recognition. In Walter Daelemans and Miles Osborne, editors, Proceedings of CoNLL 2003, pages 142–147. Edmonton, Canada. Olga Uryupina. 2006. Coreference resolution with and without linguistic knowledge. In Proceedings of LREC 2006. Kees van Deemter and Rodger Kibble. 2000. On coreferring: Coreference in MUC and related annotation schemes. Computational Linguistics, 26(4):629– 637. Marc Vilain, John Burger, John Aberdeen, Dennis Connolly, and Lynette Hirschman. 1995. A modeltheoretic coreference scoring scheme. In Proceedings of MUC-6, pages 45–52, San Francisco. 1432
|
2010
|
144
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1433–1442, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Constituency to Dependency Translation with Forests Haitao Mi and Qun Liu Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China {htmi,liuqun}@ict.ac.cn Abstract Tree-to-string systems (and their forestbased extensions) have gained steady popularity thanks to their simplicity and efficiency, but there is a major limitation: they are unable to guarantee the grammaticality of the output, which is explicitly modeled in string-to-tree systems via targetside syntax. We thus propose to combine the advantages of both, and present a novel constituency-to-dependency translation model, which uses constituency forests on the source side to direct the translation, and dependency trees on the target side (as a language model) to ensure grammaticality. Medium-scale experiments show an absolute and statistically significant improvement of +0.7 BLEU points over a state-of-the-art forest-based tree-to-string system even with fewer rules. This is also the first time that a treeto-tree model can surpass tree-to-string counterparts. 1 Introduction Linguistically syntax-based statistical machine translation models have made promising progress in recent years. By incorporating the syntactic annotations of parse trees from both or either side(s) of the bitext, they are believed better than phrasebased counterparts in reorderings. Depending on the type of input, these models can be broadly divided into two categories (see Table 1): the stringbased systems whose input is a string to be simultaneously parsed and translated by a synchronous grammar, and the tree-based systems whose input is already a parse tree to be directly converted into a target tree or string. When we also take into account the type of output (tree or string), the treebased systems can be divided into tree-to-string and tree-to-tree efforts. tree on examples (partial) fast gram. BLEU source Liu06, Huang06 + + target Galley06, Shen08 + + both Ding05, Liu09 + + both our work + + + Table 1: A classification and comparison of linguistically syntax-based SMT systems, where gram. denotes grammaticality of the output. On one hand, tree-to-string systems (Liu et al., 2006; Huang et al., 2006) have gained significant popularity, especially after incorporating packed forests (Mi et al., 2008; Mi and Huang, 2008; Liu et al., 2009; Zhang et al., 2009). Compared with their string-based counterparts, tree-based systems are much faster in decoding (linear time vs. cubic time, see (Huang et al., 2006)), do not require a binary-branching grammar as in stringbased models (Zhang et al., 2006; Huang et al., 2009), and can have separate grammars for parsing and translation (Huang et al., 2006). However, they have a major limitation that they do not have a principled mechanism to guarantee grammaticality on the target side, since there is no linguistic tree structure of the output. On the other hand, string-to-tree systems explicitly model the grammaticality of the output by using target syntactic trees. Both string-toconstituency system (e.g., (Galley et al., 2006; Marcu et al., 2006)) and string-to-dependency model (Shen et al., 2008) have achieved significant improvements over the state-of-the-art formally syntax-based system Hiero (Chiang, 2007). However, those systems also have some limitations that they run slowly (in cubic time) (Huang et al., 2006), and do not utilize the useful syntactic information on the source side. We thus combine the advantages of both tree-tostring and string-to-tree approaches, and propose 1433 a novel constituency-to-dependency model, which uses constituency forests on the source side to direct translation, and dependency trees on the target side to guarantee grammaticality of the output. In contrast to conventional tree-to-tree approaches (Ding and Palmer, 2005; Quirk et al., 2005; Xiong et al., 2007; Zhang et al., 2007; Liu et al., 2009), which only make use of a single type of trees, our model is able to combine two types of trees, outperforming both phrasebased and tree-to-string systems. Current tree-totree models (Xiong et al., 2007; Zhang et al., 2007; Liu et al., 2009) still have not outperformed the phrase-based system Moses (Koehn et al., 2007) significantly even with the help of forests.1 Our new constituency-to-dependency model (Section 2) extracts rules from word-aligned pairs of source constituency forests and target dependency trees (Section 3), and translates source constituency forests into target dependency trees with a set of features (Section 4). Medium data experiments (Section 5) show a statistically significant improvement of +0.7 BLEU points over a stateof-the-art forest-based tree-to-string system even with less translation rules, this is also the first time that a tree-to-tree model can surpass tree-to-string counterparts. 2 Model Figure 1 shows a word-aligned source constituency forest Fc and target dependency tree De, our constituency to dependency translation model can be formalized as: P(Fc, De) = X Cc∈Fc P(Cc, De) = X Cc∈Fc X o∈O P(O) = X Cc∈Fc X o∈O Y r∈o P(r), (1) where Cc is a constituency tree in Fc, o is a derivation that translates Cc to De, O is the set of derivation, r is a constituency to dependency translation rule. 1According to the reports of Liu et al. (2009), their forestbased constituency-to-constituency system achieves a comparable performance against Moses (Koehn et al., 2007), but a significant improvement of +3.6 BLEU points over the 1best tree-based constituency-to-constituency system. 2.1 Constituency Forests on the Source Side A constituency forest (in Figure 1 left) is a compact representation of all the derivations (i.e., parse trees) for a given sentence under a contextfree grammar (Billot and Lang, 1989). More formally, following Huang (2008), such a constituency forest is a pair Fc = Gf = ⟨V f, Hf⟩, where V f is the set of nodes, and Hf the set of hyperedges. For a given source sentence c1:m = c1 . . . cm, each node vf ∈V f is in the form of Xi,j, which denotes the recognition of nonterminal X spanning the substring from positions i through j (that is, ci+1 . . . cj). Each hyperedge hf ∈Hf is a pair ⟨tails(hf), head(hf)⟩, where head(hf) ∈V f is the consequent node in the deductive step, and tails(hf) ∈(V f)∗is the list of antecedent nodes. For example, the hyperedge hf 0 in Figure 1 for deduction (*) NPB0,1 CC1,2 NPB2,3 NP0,3 , (*) is notated: ⟨(NPB0,1, CC1,2, NPB2,3), NP0,3⟩. where head(hf 0) = {NP0,3}, and tails(hf 0) = {NPB0,1, CC1,2, NPB2,3}. The solid line in Figure 1 shows the best parse tree, while the dashed one shows the second best tree. Note that common sub-derivations like those for the verb VPB3,5 are shared, which allows the forest to represent exponentially many parses in a compact structure. We also denote IN (vf) to be the set of incoming hyperedges of node vf, which represents the different ways of deriving vf. Take node IP0,5 in Figure 1 for example, IN (IP0,5) = {hf 1, hf 2}. There is also a distinguished root node TOP in each forest, denoting the goal item in parsing, which is simply S0,m where S is the start symbol and m is the sentence length. 2.2 Dependency Trees on the Target Side A dependency tree for a sentence represents each word and its syntactic dependents through directed arcs, as shown in the following examples. The main advantage of a dependency tree is that it can explore the long distance dependency. 1434 1: talk blank a blan blan 2: held Bush bla blk talk a bl with b Sharon We use the lexicon dependency grammar (Hellwig, 2006) to express a projective dependency tree. Take the dependency trees above for example, they will be expressed: 1: ( a ) talk 2: ( Bush ) held ( ( a ) talk ) ( with ( Sharon ) ) where the lexicons in brackets represent the dependencies, while the lexicon out the brackets is the head. More formally, a dependency tree is also a pair De = Gd = ⟨V d, Hd⟩. For a given target sentence e1:n = e1 . . . en, each node vd ∈V d is a word ei (1 ⩽i ⩽n), each hyperedge hd ∈ Hd is a directed arc ⟨vd i , vd j ⟩from node vd i to its head node vd j . Following the formalization of the constituency forest scenario, we denote a pair ⟨tails(hd), head(hd)⟩to be a hyperedge hd, where head(hd) is the head node, tails(hd) is the node where hd leaves from. We also denote Ll(vd) and Lr(vd) to be the left and right children sequence of node vd from the nearest to the farthest respectively. Take the node vd 2 = “held” for example: Ll(vd 2) ={Bush}, Lr(vd 2) ={talk, with}. 2.3 Hypergraph Actually, both the constituency forest and the dependency tree can be formalized as a hypergraph G, a pair ⟨V, H⟩. We use Gf and Gd to distinguish them. For simplicity, we also use Fc and De to denote a constituency forest and a dependency tree respectively. Specifically, the size of tails(hd) of a hyperedge hd in a dependency tree is a constant one. IP NP x1:NPB CC yˇu x2:NPB x3:VPB →(x1) x3 (with (x2)) Figure 2: Example of the rule r1. The Chinese conjunction yˇu “and” is translated into English preposition “with”. 3 Rule Extraction We extract constituency to dependency rules from word-aligned source constituency forest and target dependency tree pairs (Figure 1). We mainly extend the tree-to-string rule extraction algorithm of Mi and Huang (2008) to our scenario. In this section, we first formalize the constituency to string translation rule (Section 3.1). Then we present the restrictions for dependency structures as well formed fragments (Section 3.2). Finally, we describe our rule extraction algorithm (Section 3.3), fractional counts computation and probabilities estimation (Section 3.4). 3.1 Constituency to Dependency Rule More formally, a constituency to dependency translation rule r is a tuple ⟨lhs(r), rhs(r), φ(r)⟩, where lhs(r) is the source side tree fragment, whose internal nodes are labeled by nonterminal symbols (like NP and VP), and whose frontier nodes are labeled by source language words ci (like “yˇu”) or variables from a set X = {x1, x2, . . .}; rhs(r) is expressed in the target language dependency structure with words ej (like “with”) and variables from the set X; and φ(r) is a mapping from X to nonterminals. Each variable xi ∈X occurs exactly once in lhs(r) and exactly once in rhs(r). For example, the rule r1 in Figure 2, lhs(r1) = IP(NP(x1 CC(yˇu) x2) x3), rhs(r1) = (x1) x3 (with (x2)), φ(r1) = {x1 7→NPB, x2 7→NPB, x3 7→VPB}. 3.2 Well Formed Dependency Fragment Following Shen et al. (2008), we also restrict rhs(r) to be well formed dependency fragment. The main difference between us is that we use more flexible restrictions. Given a dependency 1435 IP0,5 “(Bush) .. Sharon))” hf 1 NP0,3 “(Bush) ⊔(with (Sharon))” NPB0,1 “Bush” B`ush´ı hf 0 CC1,2 “with” yˇu VP1,5 “held .. Sharon))” PP1,3 “with (Sharon)” P1,2 “with” NPB2,3 “Sharon” Sh¯al´ong VPB3,5 “held ((a) talk)” VV3,4 “held ((a)*)” jˇux´ıngle NPB4,5 “talk” hu`ıt´an hf 2 Minimal rules extracted IP (NP(x1:NPB x2:CC x3:NPB) x4:VPB) →(x1) x4 (x2 (x3) ) IP (x1:NPB x2:VP) →(x1) x2 VP (x1:PP x2:VPB) →x2 (x1) PP (x1:P x2:NPB) →x1 (x2) VPB (VV(jˇux´ıngle)) x1:NPB) →held ((a) x1) NPB (B`ush´ı) →Bush NPB (hu`ıt´an) →talk CC (yˇu) →with P (yˇu) →with NPB (Sh¯al´ong) →Sharon ( Bush ) held ( ( a ) talk ) ( with ( Sharon ) ) Figure 1: Forest-based constituency to dependency rule extraction. fragment di:j composed by the words from i to j, two kinds of well formed structures are defined as follows: Fixed on one node vd one, fixed for short, if it meets the following conditions: • the head of vd one is out of [i, j], i.e.: ∀hd, if tails(hd) = vd one ⇒head(hd) /∈ei:j. • the heads of other nodes except vd one are in [i, j], i.e.: ∀k ∈[i, j] and vd k ̸= vd one, ∀hd if tails(hd) = vd k ⇒head(hd) ∈ei:j. Floating with multi nodes M, floating for short, if it meets the following conditions: • all nodes in M have a same head node, i.e.: ∃x /∈[i, j], ∀hd if tails(hd) ∈M ⇒ head(hd) = vh x. • the heads of other nodes not in M are in [i, j], i.e.: ∀k ∈[i, j] and vd k /∈M, ∀hd if tails(hd) = vd k ⇒head(hd) ∈ei:j. Take the “ (Bush) held ((a) talk))(with (Sharon)) ” for example: partial fixed examples are “ (Bush) held ” and “ held ((a) talk)”; while the partial floating examples are “ (talk) (with (Sharon)) ” and “ ((a) talk) (with (Sharon)) ”. Please note that the floating structure “ (talk) (with (Sharon)) ” can not be allowed in Shen et al. (2008)’s model. The dependency structure “ held ((a))” is not a well formed structure, since the head of word “a” is out of scope of this structure. 3.3 Rule Extraction Algorithm The algorithm shown in this Section is mainly extended from the forest-based tree-to-string extraction algorithm (Mi and Huang, 2008). We extract rules from word-aligned source constituency forest and target dependency tree pairs (see Figure 1) in three steps: (1) frontier set computation, (2) fragmentation, (3) composition. The frontier set (Galley et al., 2004) is the potential points to “cut” the forest and dependency tree pair into fragments, each of which will form a minimal rule (Galley et al., 2006). However, not every fragment can be used for rule extraction, since it may or may not respect to the restrictions, such as word alignments and well formed dependency structures. So we say a fragment is extractable if it respects to all restrictions. The root node of every extractable tree fragment corresponds to a faithful structure on the target side, in which case there is a “translational equivalence” between the subtree rooted at the node and the corresponding target structure. For example, in Figure 1, every node in the forest is annotated with its corresponding English structure. The NP0,3 node maps to a non-contiguous structure “(Bush) ⊔(with (Sharon))”, the VV3,4 node maps to a contiguous but non-faithful structure “held ((a) *)”. 1436 Algorithm 1 Forest-based constituency to dependency rule extraction. Input: Source constituency forest Fc, target dependency tree De, and alignment a Output: Minimal rule set R 1: fs ←FRONTIER(Fc, De, a) ⊲compute frontier set 2: for each vf ∈fs do 3: open ←{⟨∅, {vf}⟩} ⊲initial queue of growing fragments 4: while open ̸= ∅do 5: ⟨hs, exps⟩←open.pop() ⊲extract a fragment 6: if exps = ∅then ⊲nothing to expand? 7: generate a rule r using fragment hs ⊲generate a rule 8: R.append(r) 9: else ⊲incomplete: further expand 10: v′ ←exps.pop() ⊲a non-frontier node 11: for each hf ∈IN (v′) do 12: newexps ←exps ∪(tails(hf) \ fs) ⊲expand 13: open.append(⟨hs ∪{hf}, newexps⟩) Following Mi and Huang (2008), given a source target sentence pair ⟨c1:m, e1:n⟩with an alignment a, the span of node vf on source forest is the set of target words aligned to leaf nodes under vf: span(vf) ≜{ei ∈e1:n | ∃cj ∈yield(vf), (cj, ei) ∈a}. where the yield(vf) is all the leaf nodes under vf. For each span(vf), we also denote dep(vf) to be its corresponding dependency structure, which represents the dependency structure of all the words in span(vf). Take the span(PP1,3) ={with, Sharon} for example, the corresponding dep(PP1,3) is “with (Sharon)”. A dep(vf) is faithful structure to node vf if it meets the following restrictions: • all words in span(vf) form a continuous substring ei:j, • every word in span(vf) is only aligned to leaf nodes of vf, i.e.: ∀ei ∈span(vf), (cj, ei) ∈ a ⇒cj ∈yield(vf), • dep(vf) is a well formed dependency structure. For example, node VV3,4 has a non-faithful structure (crossed out in Figure 1), since its dep(VV3,4 = “ held ((a) *)” is not a well formed structure, where the head of word “a” lies in the outside of its words covered. Nodes with faithful structure form the frontier set (shaded nodes in Figure 1) which serve as potential cut points for rule extraction. Given the frontier set, fragmentation step is to “cut” the forest at all frontier nodes and form tree fragments, each of which forms a rule with variables matching the frontier descendant nodes. For example, the forest in Figure 1 is cut into 10 pieces, each of which corresponds to a minimal rule listed on the right. Our rule extraction algorithm is formalized in Algorithm 1. After we compute the frontier set fs (line 1). We visit each frontier node vf ∈fs on the source constituency forest Fc, and keep a queue open of growing fragments rooted at vf. We keep expanding incomplete fragments from open, and extract a rule if a complete fragment is found (line 7). Each fragment hs in open is associated with a list of expansion sites (exps in line 5) being the subset of leaf nodes of the current fragment that are not in the frontier set. So each fragment along hyperedge h is associated with exps = tails(hf) \ fs. A fragment is complete if its expansion sites is empty (line 6), otherwise we pop one expansion node v′ to grow and spin-off new fragments by following hyperedges of v′, adding new expansion sites (lines 11-13), until all active fragments are complete and open queue is empty (line 4). After we get all the minimal rules, we glue them together to form composed rules following Galley et al. (2006). For example, the composed rule r1 in Figure 2 is glued by the following two minimal rules: 1437 IP (NP(x1:NPB x2:CC x3:NPB) x4:VPB) r2 →(x1) x4 (x2 (x3) ) CC (yˇu) →with r3 where x2:CC in r2 is replaced with r3 accordingly. 3.4 Fractional Counts and Rule Probabilities Following Mi and Huang (2008), we penalize a rule r by the posterior probability of the corresponding constituent tree fragment lhs(r), which can be computed in an Inside-Outside fashion, being the product of the outside probability of its root node, the inside probabilities of its leaf nodes, and the probabilities of hyperedges involved in the fragment. αβ(lhs(r)) =α(root(r)) · Y hf ∈lhs(r) P(hf) · Y vf ∈leaves(lhs(r)) β(vf) (2) where root(r) is the root of the rule r, α(v) and β(v) are the outside and inside probabilities of node v, and leaves(lhs(r)) returns the leaf nodes of a tree fragment lhs(r). We use fractional counts to compute three conditional probabilities for each rule, which will be used in the next section: P(r | lhs(r)) = c(r) P r′:lhs(r′)=lhs(r) c(r′), (3) P(r | rhs(r)) = c(r) P r′:rhs(r′)=rhs(r) c(r′), (4) P(r | root(r)) = c(r) P r′:root(r′)=root(r) c(r′). (5) 4 Decoding Given a source forest Fc, the decoder searches for the best derivation o∗among the set of all possible derivations O, each of which forms a source side constituent tree Tc(o), a target side string e(o), and a target side dependency tree De(o): o∗= arg max Tc∈Fc,o∈O λ1 log P(o | Tc) +λ2 log Plm(e(o)) +λ3 log PDLMw(De(o)) +λ4 log PDLMp(De(o)) +λ5 log P(Tc(o)) +λ6ill(o) + λ7|o| + λ8|e(o)|, (6) where the first two terms are translation and language model probabilities, e(o) is the target string (English sentence) for derivation o, the third and forth items are the dependency language model probabilities on the target side computed with words and POS tags separately, De(o) is the target dependency tree of o, the fifth one is the parsing probability of the source side tree Tc(o) ∈Fc, the ill(o) is the penalty for the number of ill-formed dependency structures in o, and the last two terms are derivation and translation length penalties, respectively. The conditional probability P(o | Tc) is decomposes into the product of rule probabilities: P(o | Tc) = Y r∈o P(r), (7) where each P(r) is the product of five probabilities: P(r) =P(r | lhs(r))λ9 · P(r | rhs(r))λ10 · P(r | root(lhs(r)))λ11 · Plex(lhs(r) | rhs(r))λ12 · Plex(rhs(r) | lhs(r))λ13, (8) where the first three are conditional probabilities based on fractional counts of rules defined in Section 3.4, and the last two are lexical probabilities. When computing the lexical translation probabilities described in (Koehn et al., 2003), we only take into accout the terminals in a rule. If there is no terminal, we set the lexical probability to 1. The decoding algorithm works in a bottom-up search fashion by traversing each node in forest Fc. We first use pattern-matching algorithm of Mi et al. (2008) to convert Fc into a translation forest, each hyperedge of which is associated with a constituency to dependency translation rule. However, pattern-matching failure2 at a node vf will 2Pattern-matching failure at a node vf means there is no translation rule can be matched at vf or no translation hyperedge can be constructed at vf. 1438 cut the derivation path and lead to translation failure. To tackle this problem, we construct a pseudo translation rule for each parse hyperedge hf ∈ IN (vf) by mapping the CFG rule into a target dependency tree using the head rules of Magerman (1995). Take the hyperedge hf 0 in Figure1 for example, the corresponding pseudo translation rule is: NP(x1:NPB x2:CC x3:NPB) →(x1) (x2) x3, since the x3:NPB is the head word of the CFG rule: NP →NPB CC NPB. After the translation forest is constructed, we traverse each node in translation forest also in bottom-up fashion. For each node, we use the cube pruning technique (Chiang, 2007; Huang and Chiang, 2007) to produce partial hypotheses and compute all the feature scores including the dependency language model score (Section 4.1). If all the nodes are visited, we trace back along the 1-best derivation at goal item S0,m and build a target side dependency tree. For k-best search after getting 1-best derivation, we use the lazy Algorithm 3 of Huang and Chiang (2005) that works backwards from the root node, incrementally computing the second, third, through the kth best alternatives. 4.1 Dependency Language Model Computing We compute the score of a dependency language model for a dependency tree De in the same way proposed by Shen et al. (2008). For each nonterminal node vd h = eh in De and its children sequences Ll = el1, el2...eli and Lr = er1, er2...erj, the probability of a trigram is computed as follows: P(Ll, Lr | eh§) = P(Ll | eh§)·P(Lr | eh§), (9) where the P(Ll | eh§) is decomposed to be: P(Ll | eh§) =P(ell | eh§) · P(el2 | el1, eh§) ... · P(eln | eln−1, eln−2). (10) We use the suffix “§” to distinguish the head word and child words in the dependency language model. In order to alleviate the problem of data sparse, we also compute a dependency language model for POS tages over a dependency tree. We store the POS tag information on the target side for each constituency-to-dependency rule. So we will also generate a POS taged dependency tree simultaneously at the decoding time. We calculate this dependency language model by simply replacing each ei in equation 9 with its tag t(ei). 5 Experiments 5.1 Data Preparation Our training corpus consists of 239K sentence pairs with about 6.9M/8.9M words in Chinese/English, respectively. We first word-align them by GIZA++ (Och and Ney, 2000) with refinement option “grow-diag-and” (Koehn et al., 2003), and then parse the Chinese sentences using the parser of Xiong et al. (2005) into parse forests, which are pruned into relatively small forests with a pruning threshold 3. We also parse the English sentences using the parser of Charniak (2000) into 1-best constituency trees, which will be converted into dependency trees using Magerman (1995)’s head rules. We also store the POS tag information for each word in dependency trees, and compute two different dependency language models for words and POS tags in dependency tree separately. Finally, we apply translation rule extraction algorithm described in Section 3. We use SRI Language Modeling Toolkit (Stolcke, 2002) to train a 4-gram language model with Kneser-Ney smoothing on the first 1/3 of the Xinhua portion of Gigaword corpus. At the decoding step, we again parse the input sentences into forests and prune them with a threshold 10, which will direct the translation (Section 4). We use the 2002 NIST MT Evaluation test set as our development set and the 2005 NIST MT Evaluation test set as our test set. We evaluate the translation quality using the BLEU-4 metric (Papineni et al., 2002), which is calculated by the script mteval-v11b.pl with its default setting which is case-insensitive matching of n-grams. We use the standard minimum error-rate training (Och, 2003) to tune the feature weights to maximize the system’s BLEU score on development set. 5.2 Results Table 2 shows the results on the test set. Our baseline system is a state-of-the-art forest-based constituency-to-string model (Mi et al., 2008), or forest c2s for short, which translates a source forest into a target string by pattern-matching the 1439 constituency-to-string (c2s) rules and the bilingual phrases (s2s). The baseline system extracts 31.9M c2s rules, 77.9M s2s rules respectively and achieves a BLEU score of 34.17 on the test set3. At first, we investigate the influence of different rule sets on the performance of baseline system. We first restrict the target side of translation rules to be well-formed structures, and we extract 13.8M constituency-to-dependency (c2d) rules, which is 43% of c2s rules. We also extract 9.0M string-to-dependency (s2d) rules, which is only 11.6% of s2s rules. Then we convert c2d and s2d rules to c2s and s2s rules separately by removing the target-dependency structures and feed them into the baseline system. As shown in the third line in the column of BLEU score, the performance drops 1.7 BLEU points over baseline system due to the poorer rule coverage. However, when we further use all s2s rules instead of s2d rules in our next experiment, it achieves a BLEU score of 34.03, which is very similar to the baseline system. Those results suggest that restrictions on c2s rules won’t hurt the performance, but restrictions on s2s will hurt the translation quality badly. So we should utilize all the s2s rules in order to preserve a good coverage of translation rule set. The last two lines in Table 2 show the results of our new forest-based constituency-to-dependency model (forest c2d for short). When we only use c2d and s2d rules, our system achieves a BLEU score of 33.25, which is lower than the baseline system in the first line. But, with the same rule set, our model still outperform the result in the second line. This suggests that using dependency language model really improves the translation quality by less than 1 BLEU point. In order to utilize all the s2s rules and increase the rule coverage, we parse the target strings of the s2s rules into dependency fragments, and construct the pseudo s2d rules (s2s-dep). Then we use c2d and s2s-dep rules to direct the translation. With the help of the dependency language model, our new model achieves a significant improvement of +0.7 BLEU points over the forest c2s baseline system (p < 0.05, using the sign-test suggested by 3According to the reports of Liu et al. (2009), with a more larger training corpus (FBIS plus 30K) but no name entity translations (+1 BLEU points if it is used), their forest-based constituency-to-constituency model achieves a BLEU score of 30.6, which is similar to Moses (Koehn et al., 2007). So our baseline system is much better than the BLEU score (30.6+1) of the constituency-to-constituency system and Moses. System Rule Set BLEU Type # forest c2s c2s 31.9M 34.17 s2s 77.9M c2d 13.8M 32.48(↓1.7) s2d 9.0M c2d 13.8M 34.03(↓0.1) s2s 77.9M forest c2d c2d 13.8M 33.25(↓0.9) s2d 9.0M c2d 13.8M 34.88(↑0.7) s2s-dep 77.9M Table 2: Statistics of different types of rules extracted on training corpus and the BLEU scores on the test set. Collins et al. (2005)). For the first time, a tree-totree model can surpass tree-to-string counterparts significantly even with fewer rules. 6 Related Work The concept of packed forest has been used in machine translation for several years. For example, Huang and Chiang (2007) use forest to characterize the search space of decoding with integrated language models. Mi et al. (2008) and Mi and Huang (2008) use forest to direct translation and extract rules rather than 1-best tree in order to weaken the influence of parsing errors, this is also the first time to use forest directly in machine translation. Following this direction, Liu et al. (2009) and Zhang et al. (2009) apply forest into tree-to-tree (Zhang et al., 2007) and tree-sequence-to-string models(Liu et al., 2007) respectively. Different from Liu et al. (2009), we apply forest into a new constituency tree to dependency tree translation model rather than constituency tree-to-tree model. Shen et al. (2008) present a string-todependency model. They define the well-formed dependency structures to reduce the size of translation rule set, and integrate a dependency language model in decoding step to exploit long distance word relations. This model shows a significant improvement over the state-of-the-art hierarchical phrase-based system (Chiang, 2005). Compared with this work, we put fewer restrictions on the definition of well-formed dependency structures in order to extract more rules; the 1440 other difference is that we can also extract more expressive constituency to dependency rules, since the source side of our rule can encode multi-level reordering and contain more variables being larger than two; furthermore, our rules can be pattern-matched at high level, which is more reasonable than using glue rules in Shen et al. (2008)’s scenario; finally, the most important one is that our model runs very faster. Liu et al. (2009) propose a forest-based constituency-to-constituency model, they put more emphasize on how to utilize parse forest to increase the tree-to-tree rule coverage. By contrast, we only use 1-best dependency trees on the target side to explore long distance relations and extract translation rules. Theoretically, we can extract more rules since dependency tree has the best inter-lingual phrasal cohesion properties (Fox, 2002). 7 Conclusion and Future Work In this paper, we presented a novel forest-based constituency-to-dependency translation model, which combines the advantages of both tree-tostring and string-to-tree systems, runs fast and guarantees grammaticality of the output. To learn the constituency-to-dependency translation rules, we first identify the frontier set for all the nodes in the constituency forest on the source side. Then we fragment them and extract minimal rules. Finally, we glue them together to be composed rules. At the decoding step, we first parse the input sentence into a constituency forest. Then we convert it into a translation forest by patter-matching the constituency to string rules. Finally, we traverse the translation forest in a bottom-up fashion and translate it into a target dependency tree by incorporating string-based and dependency-based language models. Using all constituency-to-dependency translation rules and bilingual phrases, our model achieves +0.7 points improvement in BLEU score significantly over a state-of-the-art forest-based tree-to-string system. This is also the first time that a tree-to-tree model can surpass tree-to-string counterparts. In the future, we will do more experiments on rule coverage to compare the constituency-toconstituency model with our model. Furthermore, we will replace 1-best dependency trees on the target side with dependency forests to further increase the rule coverage. Acknowledgement The authors were supported by National Natural Science Foundation of China, Contracts 60736014 and 90920004, and 863 State Key Project No. 2006AA010108. We thank the anonymous reviewers for their insightful comments. We are also grateful to Liang Huang for his valuable suggestions. References Sylvie Billot and Bernard Lang. 1989. The structure of shared forests in ambiguous parsing. In Proceedings of ACL ’89, pages 143–151. Eugene Charniak. 2000. A maximum-entropy inspired parser. In Proceedings of NAACL, pages 132–139. David Chiang. 2005. A hierarchical phrase-based model for statistical machine translation. In Proceedings of ACL, pages 263–270, Ann Arbor, Michigan, June. David Chiang. 2007. Hierarchical phrase-based translation. Comput. Linguist., 33(2):201–228. Michael Collins, Philipp Koehn, and Ivona Kucerova. 2005. Clause restructuring for statistical machine translation. In Proceedings of ACL, pages 531–540. Yuan Ding and Martha Palmer. 2005. Machine translation using probabilistic synchronous dependency insertion grammars. In Proceedings of ACL, pages 541–548, June. Heidi J. Fox. 2002. Phrasal cohesion and statistical machine translation. In In Proceedings of EMNLP02. Michel Galley, Mark Hopkins, Kevin Knight, and Daniel Marcu. 2004. What’s in a translation rule? In Proceedings of HLT/NAACL. Michel Galley, Jonathan Graehl, Kevin Knight, Daniel Marcu, Steve DeNeefe, Wei Wang, and Ignacio Thayer. 2006. Scalable inference and training of context-rich syntactic translation models. In Proceedings of COLING-ACL, pages 961–968, July. Peter Hellwig. 2006. Parsing with Dependency Grammars, volume II. An International Handbook of Contemporary Research. Liang Huang and David Chiang. 2005. Better k-best parsing. In Proceedings of IWPT. Liang Huang and David Chiang. 2007. Forest rescoring: Faster decoding with integrated language models. In Proceedings of ACL, pages 144–151, June. Liang Huang, Kevin Knight, and Aravind Joshi. 2006. Statistical syntax-directed translation with extended domain of locality. In Proceedings of AMTA. 1441 Liang Huang, Hao Zhang, Daniel Gildea, , and Kevin Knight. 2009. Binarization of synchronous contextfree grammars. Comput. Linguist. Liang Huang. 2008. Forest reranking: Discriminative parsing with non-local features. In Proceedings of ACL. Philipp Koehn, Franz Joseph Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of HLT-NAACL, pages 127–133, Edmonton, Canada, May. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of ACL, pages 177–180, June. Yang Liu, Qun Liu, and Shouxun Lin. 2006. Treeto-string alignment template for statistical machine translation. In Proceedings of COLING-ACL, pages 609–616, Sydney, Australia, July. Yang Liu, Yun Huang, Qun Liu, and Shouxun Lin. 2007. Forest-to-string statistical translation rules. In Proceedings of ACL, pages 704–711, June. Yang Liu, Yajuan L¨u, and Qun Liu. 2009. Improving tree-to-tree translation with packed forests. In Proceedings of ACL/IJCNLP, August. David M. Magerman. 1995. Statistical decision-tree models for parsing. In Proceedings of ACL, pages 276–283, June. Daniel Marcu, Wei Wang, Abdessamad Echihabi, and Kevin Knight. 2006. Spmt: Statistical machine translation with syntactified target language phrases. In Proceedings of EMNLP, pages 44–52, July. Haitao Mi and Liang Huang. 2008. Forest-based translation rule extraction. In Proceedings of EMNLP 2008, pages 206–214, Honolulu, Hawaii, October. Haitao Mi, Liang Huang, and Qun Liu. 2008. Forestbased translation. In Proceedings of ACL-08:HLT, pages 192–199, Columbus, Ohio, June. Franz J. Och and Hermann Ney. 2000. Improved statistical alignment models. In Proceedings of ACL, pages 440–447. Franz J. Och. 2003. Minimum error rate training in statistical machine translation. In Proceedings of ACL, pages 160–167. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In Proceedings of ACL, pages 311–318, Philadephia, USA, July. Chris Quirk, Arul Menezes, and Colin Cherry. 2005. Dependency treelet translation: Syntactically informed phrasal SMT. In Proceedings of ACL, pages 271–279, June. Libin Shen, Jinxi Xu, and Ralph Weischedel. 2008. A new string-to-dependency machine translation algorithm with a target dependency language model. In Proceedings of ACL-08: HLT, June. Andreas Stolcke. 2002. SRILM - an extensible language modeling toolkit. In Proceedings of ICSLP, volume 30, pages 901–904. Deyi Xiong, Shuanglong Li, Qun Liu, and Shouxun Lin. 2005. Parsing the Penn Chinese Treebank with Semantic Knowledge. In Proceedings of IJCNLP 2005, pages 70–81. Deyi Xiong, Qun Liu, and Shouxun Lin. 2007. A dependency treelet string correspondence model for statistical machine translation. In Proceedings of SMT, pages 40–47. Hao Zhang, Liang Huang, Daniel Gildea, and Kevin Knight. 2006. Synchronous binarization for machine translation. In Proc. of HLT-NAACL. Min Zhang, Hongfei Jiang, Aiti Aw, Jun Sun, Sheng Li, and Chew Lim Tan. 2007. A tree-to-tree alignmentbased model for statistical machine translation. In Proceedings of MT-Summit. Hui Zhang, Min Zhang, Haizhou Li, Aiti Aw, and Chew Lim Tan. 2009. Forest-based tree sequence to string translation model. In Proceedings of the ACL/IJCNLP 2009. 1442
|
2010
|
145
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1443–1452, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Learning to Translate with Source and Target Syntax David Chiang USC Information Sciences Institute 4676 Admiralty Way, Suite 1001 Marina del Rey, CA 90292 USA [email protected] Abstract Statistical translation models that try to capture the recursive structure of language have been widely adopted over the last few years. These models make use of varying amounts of information from linguistic theory: some use none at all, some use information about the grammar of the target language, some use information about the grammar of the source language. But progress has been slower on translation models that are able to learn the relationship between the grammars of both the source and target language. We discuss the reasons why this has been a challenge, review existing attempts to meet this challenge, and show how some old and new ideas can be combined into a simple approach that uses both source and target syntax for significant improvements in translation accuracy. 1 Introduction Statistical translation models that use synchronous context-free grammars (SCFGs) or related formalisms to try to capture the recursive structure of language have been widely adopted over the last few years. The simplest of these (Chiang, 2005) make no use of information from syntactic theories or syntactic annotations, whereas others have successfully incorporated syntactic information on the target side (Galley et al., 2004; Galley et al., 2006) or the source side (Liu et al., 2006; Huang et al., 2006). The next obvious step is toward models that make full use of syntactic information on both sides. But the natural generalization to this setting has been found to underperform phrasebased models (Liu et al., 2009; Ambati and Lavie, 2008), and researchers have begun to explore solutions (Zhang et al., 2008; Liu et al., 2009). In this paper, we explore the reasons why treeto-tree translation has been challenging, and how source syntax and target syntax might be used together. Drawing on previous successful attempts to relax syntactic constraints during grammar extraction in various ways (Zhang et al., 2008; Liu et al., 2009; Zollmann and Venugopal, 2006), we compare several methods for extracting a synchronous grammar from tree-to-tree data. One confounding factor in such a comparison is that some methods generate many new syntactic categories, making it more difficult to satisfy syntactic constraints at decoding time. We therefore propose to move these constraints from the formalism into the model, implemented as features in the hierarchical phrasebased model Hiero (Chiang, 2005). This augmented model is able to learn from data whether to rely on syntax or not, or to revert back to monotone phrase-based translation. In experiments on Chinese-English and ArabicEnglish translation, we find that when both source and target syntax are made available to the model in an unobtrusive way, the model chooses to build structures that are more syntactically well-formed and yield significantly better translations than a nonsyntactic hierarchical phrase-based model. 2 Grammar extraction A synchronous tree-substitution grammar (STSG) is a set of rules or elementary tree pairs (γ, α), where: • γ is a tree whose interior labels are sourcelanguage nonterminal symbols and whose frontier labels are source-language nonterminal symbols or terminal symbols (words). The nonterminal-labeled frontier nodes are called substitution nodes, conventionally marked with an arrow (↓). • α is a tree of the same form except with 1443 . . . .PP . . . . . . .LCP . . . . . . .LC . . . . . 中 zhōng . . . . .NP↓ . . . ..P . . . . .在 zài . . .PP . . . . . . .NP↓ . . . ..IN . . . ..in . . . .NP . . . . . . .NP . . . . .NN . . . . .贸易 màoyì . . . . .NP . . . . . . .NP . . . . .NN . . . ..岸 àn . . . . .QP . . . . .CD . . . . . 两 liǎng . . .NP . . . . . . .PP . . . . . . .NP . . . . . . .NNS . . . . .shores . . . . . . .CD . . . . .two . . . . .DT . . . . .the . . . ..IN . . . . .between . . . . .NP . . . . .NN . . . . .trade . . . .PP . . . . . . .LCP . . . . . . .LC . . . . . 中 zhōng . . . . .NP . . . . . . .NP . . . . .NN . . . . .贸易 màoyì . . . . .NP . . . . . . .NP . . . . .NN . . . ..岸 àn . . . . .QP . . . . .CD . . . . . 两 liǎng . . . ..P . . . . .在 zài . . .PP . . . . . . .NP . . . . . . .PP . . . . . . .NP . . . . . . .NNS . . . . .shores . . . . . . .CD . . . . .two . . . . .DT . . . . .the . . . ..IN . . . . .between . . . . .NP . . . . .NN . . . . .trade . . . ..IN . . . ..in (γ1, α1) (γ2, α2) (γ3, α3) Figure 1: Synchronous tree substitution. Rule (γ2, α2) is substituted into rule (γ1, α1) to yield (γ3, α3). target-language instead of source-language symbols. • The substitution nodes of γ are aligned bijectively with those of α. • The terminal-labeled frontier nodes of γ are aligned (many-to-many) with those of α. In the substitution operation, an aligned pair of substitution nodes is rewritten with an elementary tree pair. The labels of the substitution nodes must match the root labels of the elementary trees with which they are rewritten (but we will relax this constraint below). See Figure 1 for examples of elementary tree pairs and substitution. 2.1 Exact tree-to-tree extraction The use of STSGs for translation was proposed in the Data-Oriented Parsing literature (Poutsma, 2000; Hearne and Way, 2003) and by Eisner (2003). Both of these proposals are more ambitious about handling spurious ambiguity than approaches derived from phrase-based translation usually have been (the former uses random sampling to sum over equivalent derivations during decoding, and the latter uses dynamic programming human automatic string-to-string 198,445 142,820 max nested 78,361 64,578 tree-to-string 60,939 (78%) 48,235 (75%) string-to-tree 59,274 (76%) 46,548 (72%) tree-to-tree 53,084 (68%) 39,049 (60%) Table 1: Analysis of phrases extracted from Chinese-English newswire data with human and automatic word alignments and parses. As tree constraints are added, the number of phrase pairs drops. Errors in automatic annotations also decrease the number of phrase pairs. Percentages are relative to the maximum number of nested phrase pairs. to sum over equivalent derivations during training). If we take a more typical approach, which generalizes that of Galley et al. (2004; 2006) and is similar to Stat-XFER (Lavie et al., 2008), we obtain the following grammar extraction method, which we call exact tree-to-tree extraction. Given a pair of source- and target-language parse trees with a word alignment between their leaves, identify all the phrase pairs (¯f ,¯e), i.e., those substring pairs that respect the word align1444 . . ..IP . . . . . . .VP . . . . .一百四十七亿 yībǎisìshíqī 美元 měiyuán . . . . . . .NP . . . . .NN . . . . . 顺差 shùnchā . . . . . . .PP . . . . . . .LCP . . . . . . .LC . . . . . 中 zhōng . . . . .NP . . . . . . .NP . . . . .NN . . . . .贸易 màoyì . . . . .NP . . . . . . .NP . . . . .NN . . . ..岸 àn . . . . .QP . . . . .CD . . . . . 两 liǎng . . . ..P . . . . .在 zài . . . . .NP . . . . .NR . . . . . 台湾 Táiwān . ..S . . . . . . .VP . . . . .is 14.7 billion US dollars . . . . .NP . . . . . . .PP . . . . . . .NP . . . . . . .PP . . . . . . .NP . . . . . . .NNS . . . . .shores . . . . . . .CD . . . . .two . . . . .DT . . . . .the . . . ..IN . . . . .between . . . . .NP . . . . .NN . . . . .trade . . . ..IN . . . ..in . . . . .NP . . . . . . .NN . . . . .surplus . . . . .NP . . . . . . .POS . . . ..’s . . . . .NNP . . . . .Taiwan Figure 2: Example Chinese-English sentence pair with human-annotated parse trees and word alignments. ment in the sense that at least one word in ¯f is aligned to a word in ¯e, and no word in ¯f is aligned to a word outside of ¯e, or vice versa. Then the extracted grammar is the smallest STSG G satisfying: • If (γ, α) is a pair of subtrees of a training example and the frontiers of γ and α form a phrase pair, then (γ, α) is a rule in G. • If (γ2, α2) ∈G, (γ3, α3) ∈G, and (γ1, α1) is an elementary tree pair such that substituting (γ2, α2) into (γ1, α1) results in (γ3, α3), then (γ1, α1) is a rule in G. For example, consider the training example in Figure 2, from which the elementary tree pairs shown in Figure 1 can be extracted. The elementary tree pairs (γ2, α2) and (γ3, α3) are rules in G because their yields are phrase pairs, and (γ1, α1) results from subtracting (γ2, α2) from (γ3, α3). 2.2 Fuzzy tree-to-tree extraction Exact tree-to-tree translation requires that translation rules deal with syntactic constituents on both the source and target side, which reduces the number of eligible phrases. Table 1 shows an analysis of phrases extracted from human word-aligned and parsed data and automatically word-aligned and parsed data.1 The first line shows the number of phrase-pair occurrences that are extracted in the absence of syntactic constraints,2 and the second line shows the maximum number of nested phrase-pair occurrences, which is the most that exact syntax-based extraction can achieve. Whereas tree-to-string extraction and string-to-tree extraction permit 70–80% of the maximum possible number of phrase pairs, tree-to-tree extraction only permits 60–70%. Why does this happen? We can see that moving from human annotations to automatic annotations decreases not only the absolute number of phrase pairs, but the percentage of phrases that pass the syntactic filters. Wellington et al. (2006), in a more systematic study, find that, of sentences where the tree-to-tree constraint blocks rule extraction, the majority are due to parser errors. To address this problem, Liu et al. (2009) extract rules from pairs 1The first 2000 sentences from the GALE Phase 4 Chinese Parallel Word Alignment and Tagging Part 1 (LDC2009E83) and the Chinese News Translation Text Part 1 (LDC2005T06), respectively. 2Only counting phrases that have no unaligned words at their endpoints. 1445 of packed forests instead of pairs of trees. Since a packed forest is much more likely to include the correct tree, it is less likely that parser errors will cause good rules to be filtered out. However, even on human-annotated data, treeto-tree extraction misses many rules, and many such rules would seem to be useful. For example, in Figure 2, the whole English phrase “Taiwan’s. . .shores” is an NP, but its Chinese counterpart is not a constituent. Furthermore, neither “surplus. . .shores” nor its Chinese counterpart are constituents. But both rules are arguably useful for translation. Wellington et al. therefore argue that in order to extract as many rules as possible, a more powerful formalism than synchronous CFG/TSG is required: for example, generalized multitext grammar (Melamed et al., 2004), which is equivalent to synchronous set-local multicomponent CFG/TSG (Weir, 1988). But the problem illustrated in Figure 2 does not reflect a very deep fact about syntax or crosslingual divergences, but rather choices in annotation style that interact badly with the exact treeto-tree extraction heuristic. On the Chinese side, the IP is too flat (because 台湾/Táiwān has been analyzed as a topic), whereas the more articulated structure (1) [NP Táiwān [NP [PP zaì . . .] shùnchā]] would also be quite reasonable. On the English side, the high attachment of the PP disagrees with the corresponding Chinese structure, but low attachment also seems reasonable: (2) [NP [NP Taiwan’s] [NP surplus in trade. . .]] Thus even in the gold-standard parse trees, phrase structure can be underspecified (like the flat IP above) or uncertain (like the PP attachment above). For this reason, some approaches work with a more flexible notion of constituency. Synchronous tree-sequence–substitution grammar (STSSG) allows either side of a rule to comprise a sequence of trees instead of a single tree (Zhang et al., 2008). In the substitution operation, a sequence of sister substitution nodes is rewritten with a tree sequence of equal length (see Figure 3a). This extra flexibility effectively makes the analysis (1) available to us. Any STSSG can be converted into an equivalent STSG via the creation of virtual nodes (see Figure 3b): for every elementary tree sequence with roots X1, . . . , Xn, create a new root node with a . . . .NP . . . . . . .NNP↓ . . . . . . .NNP↓ . . . . . . .NN . . . . .Minister . . . . .NN . . . . .Prime . . . . . .NNP . . . . .Ariel . . . . .NNP . . . . .Sharon . (a) . . . .NP . . . . . . .NNP ∗NNP↓ . . . . . . .NN . . . . .Minister . . . . .NN . . . . .Prime . . . . .NNP ∗NNP . . . . . . .NNP . . . . .Sharon . . . . .NNP . . . . .Ariel (b) Figure 3: (a) Example tree-sequence substitution grammar and (b) its equivalent SAMT-style treesubstitution grammar. complex label X1 ∗· · · ∗Xn immediately dominating the old roots, and replace every sequence of substitution sites X1, . . . , Xn with a single substitution site X1 ∗· · · ∗Xn. This is essentially what syntax-augmented MT (SAMT) does, in the stringto-tree setting (Zollmann and Venugopal, 2006). In addition, SAMT drops the requirement that the Xi are sisters, and uses categories X / Y (an X missing a Y on the right) and Y \ X (an X missing a Y on the left) in the style of categorial grammar (Bar-Hillel, 1953). Under this flexible notion of constituency, both (1) and (2) become available, albeit with more complicated categories. Both STSSG and SAMT are examples of what we might call fuzzy tree-to-tree extraction. We follow this approach here as well: as in STSSG, we work on tree-to-tree data, and we use the complex categories of SAMT. Moreover, we allow the product categories X1 ∗· · · ∗Xn to be of any length n, and we allow the slash categories to take any number of arguments on either side. Thus every phrase can be assigned a (possibly very complex) syntactic category, so that fuzzy tree-to-tree extraction does not lose any rules relative to stringto-string extraction. On the other hand, if several rules are extracted 1446 that differ only in their nonterminal labels, only the most-frequent rule is kept, and its count is the total count of all the rules. This means that there is a one-to-one correspondence between the rules extracted by fuzzy tree-to-tree extraction and hierarchical string-to-string extraction. 2.3 Nesting phrases Fuzzy tree-to-tree extraction (like string-to-string extraction) generates many times more rules than exact tree-to-tree extraction does. In Figure 2, we observed that the flat structure of the Chinese IP prevented exact tree-to-tree extraction from extracting a rule containing just part of the IP, for example: (3) [PP zaì . . .] [NP shùnchā] (4) [NP Táiwān] [PP zaì . . .] [NP shùnchā] (5) [PP zaì . . .] [NP shùnchā] [VP . . . měiyuán] Fuzzy tree-to-tree extraction allows any of these to be the source side of a rule. We might think of it as effectively restructuring the trees by inserting nodes with complex labels. However, it is not possible to represent this restructuring with a single tree (see Figure 4). More formally, let us say that two phrases wi · · · wj−1 and wi′ · · · wj′−1 nest if i ≤i′ < j′ ≤j or i′ ≤i < j < j′; otherwise, they cross. The two Chinese phrases (4) and (5) cross, and therefore cannot both be constituents in the same tree. In other words, exact tree-to-tree extraction commits to a single structural analysis but fuzzy tree-to-tree extraction pursues many restructured analyses at once. We can strike a compromise by continuing to allow SAMT-style complex categories, but committing to a single analysis by requiring all phrases to nest. To do this, we use a simple heuristic. Iterate through all the phrase pairs (¯f ,¯e) in the following order: 1. sort by whether ¯f and ¯e can be assigned a simple syntactic category (both, then one, then neither); if there is a tie, 2. sort by how many syntactic constituents ¯f and ¯e cross (low to high); if there is a tie, 3. give priority to (¯f ,¯e) if neither ¯f nor ¯e begins or ends with punctuation; if there is a tie, finally 4. sort by the position of ¯f in the source-side string (right to left). For each phrase pair, accept it if it does not cross any previously accepted phrase pair; otherwise, reject it. Because this heuristic produces a set of nesting phrases, we can represent them all in a single restructured tree. In Figure 4, this heuristic chooses structure (a) because the English-side counterpart of IP/VP has the simple category NP. 3 Decoding In decoding, the rules extracted during training must be reassembled to form a derivation whose source side matches the input sentence. In the exact tree-to-tree approach, whenever substitution is performed, the root labels of the substituted trees must match the labels of the substitution nodes—call this the matching constraint. Because this constraint must be satisfied on both the source and target side, it can become difficult to generalize well from training examples to new input sentences. Venugopal et al. (2009), in the string-to-tree setting, attempt to soften the data-fragmentation effect of the matching constraint: instead of trying to find the single derivation with the highest probability, they sum over derivations that differ only in their nonterminal labels and try to find the single derivation-class with the highest probability. Still, only derivations that satisfy the matching constraint are included in the summation. But in some cases we may want to soften the matching constraint itself. Some syntactic categories are similar enough to be considered compatible: for example, if a rule rooted in VBD (pasttense verb) could substitute into a site labeled VBZ (present-tense verb), it might still generate correct output. This is all the more true with the addition of SAMT-style categories: for example, if a rule rooted in ADVP ∗VP could substitute into a site labeled VP, it would very likely generate correct output. Since we want syntactic information to help the model make good translation choices, not to rule out potentially correct choices, we can change the way the information is used during decoding: we allow any rule to substitute into any site, but let the model learn which substitutions are better than others. To do this, we add the following features to the model: 1447 . . ..IP . . . . . . .VP . . . . .一百四十七亿 yībǎisìshíqī 美元 měiyuán . . . . .IP/VP . . . . . . .PP ∗NP . . . . . . .NP . . . . .NN . . . . . 顺差 shùnchā . . . . .PP . . . . .在 zài 两 liǎng 岸 àn 贸易 màoyì 中 zhōng . . . . .NP . . . . .NR . . . . . 台湾 Táiwān . . ..IP . . . . . . .IP\NP . . . . . . .VP . . . . .一百四十七亿 yībǎisìshíqī 美元 měiyuán . . . . .PP ∗NP . . . . . . .NP . . . . .NN . . . . . 顺差 shùnchā . . . . .PP . . . . .在 zài 两 liǎng 岸 àn 贸易 màoyì 中 zhōng . . . . .NP . . . . .NR . . . . . 台湾 Táiwān (a) (b) Figure 4: Fuzzy tree-to-tree extraction effectively restructures the Chinese tree from Figure 2 in two ways but does not commit to either one. • match f counts the number of substitutions where the label of the source side of the substitution site matches the root label of the source side of the rule, and ¬match f counts those where the labels do not match. • subst f X→Y counts the number of substitutions where the label of the source side of the substitution site is X and the root label of the source side of the rule is Y. • matche, ¬matche, and subste X→Y do the same for the target side. • rootX,X′ counts the number of rules whose root label on the source side is X and whose root label on the target side is X′.3 For example, in the derivation of Figure 1, the following features would fire: match f = 1 subst f NP→NP = 1 match e = 1 subst e NP→NP = 1 rootNP,NP = 1 The decoding algorithm then operates as in hierarchical phrase-based translation. The decoder has to store in each hypothesis the source and target root labels of the partial derivation, but these labels are used for calculating feature vectors only and not for checking well-formedness of derivations. This additional state does increase the search space of the decoder, but we did not change any pruning settings. 3Thanks to Adam Pauls for suggesting this feature class. 4 Experiments To compare the methods described above with hierarchical string-to-string translation, we ran experiments on both Chinese-English and ArabicEnglish translation. 4.1 Setup The sizes of the parallel texts used are shown in Table 2. We word-aligned the Chinese-English parallel text using GIZA++ followed by link deletion (Fossum et al., 2008), and the Arabic-English parallel text using a combination of GIZA++ and LEAF (Fraser and Marcu, 2007). We parsed the source sides of both parallel texts using the Berkeley parser (Petrov et al., 2006), trained on the Chinese Treebank 6 and Arabic Treebank parts 1–3, and the English sides using a reimplementation of the Collins parser (Collins, 1997). For string-to-string extraction, we used the same constraints as in previous work (Chiang, 2007), with differences shown in Table 2. Rules with nonterminals were extracted from a subset of the data (labeled “Core” in Table 2), and rules without nonterminals were extracted from the full parallel text. Fuzzy tree-to-tree extraction was performed using analogous constraints. For exact tree-to-tree extraction, we used simpler settings: no limit on initial phrase size or unaligned words, and a maximum of 7 frontier nodes on the source side. All systems used the glue rule (Chiang, 2005), which allows the decoder, working bottom-up, to stop building hierarchical structure and instead concatenate partial translations without any reordering. The model attaches a weight to the glue rule so that it can learn from data whether to build shallow or rich structures, but for efficiency’s sake the decoder has a hard limit, called the distortion 1448 Chi-Eng Ara-Eng Core training words 32+38M 28+34M initial phrase size 10 15 final rule size 6 6 nonterminals 2 2 loose source 0 ∞ loose target 0 2 Full training words 240+260M 190+220M final rule size 6 6 nonterminals 0 0 loose source ∞ ∞ loose target 1 2 Table 2: Rule extraction settings used for experiments. “Loose source/target” is the maximum number of unaligned source/target words at the endpoints of a phrase. limit, above which the glue rule must be used. We trained two 5-gram language models: one on the combined English halves of the bitexts, and one on two billion words of English. These were smoothed using modified Kneser-Ney (Chen and Goodman, 1998) and stored using randomized data structures similar to those of Talbot and Brants (2008). The base feature set for all systems was similar to the expanded set recently used for Hiero (Chiang et al., 2009), but with bigram features (source and target word) instead of trigram features (source and target word and neighboring source word). For all systems but the baselines, the features described in Section 3 were added. The systems were trained using MIRA (Crammer and Singer, 2003; Chiang et al., 2009) on a tuning set of about 3000 sentences of newswire from NIST MT evaluation data and GALE development data, disjoint from the training data. We optimized feature weights on 90% of this and held out the other 10% to determine when to stop. 4.2 Results Table 3 shows the scores on our development sets and test sets, which are about 3000 and 2000 sentences, respectively, of newswire drawn from NIST MT evaluation data and GALE development data and disjoint from the tuning data. For Chinese, we first tried increasing the distortion limit from 10 words to 20. This limit controls how deeply nested the tree structures built by the decoder are, and we want to see whether adding syntactic information leads to more complex structures. This change by itself led to an increase in the BLEU score. We then compared against two systems using tree-to-tree grammars. Using exact tree-to-tree extraction, we got a much smaller grammar, but decreased accuracy on all but the Chinese-English test set, where there was no significant change. But with fuzzy tree-to-tree extraction, we obtained an improvement of +0.6 on both Chinese-English sets, and +0.7/+0.8 on the ArabicEnglish sets. Applying the heuristic for nesting phrases reduced the grammar sizes dramatically (by a factor of 2.4 for Chinese and 4.2 for Arabic) but, interestingly, had almost no effect on translation quality: a slight decrease in BLEU on the Arabic-English development set and no significant difference on the other sets. This suggests that the strength of fuzzy tree-to-tree extraction lies in its ability to break up flat structures and to reconcile the source and target trees with each other, rather than multiple restructurings of the training trees. 4.3 Rule usage We then took a closer look at the behavior of the string-to-string and fuzzy tree-to-tree grammars (without the nesting heuristic). Because the rules of these grammars are in one-to-one correspondence, we can analyze the string-to-string system’s derivations as though they had syntactic categories. First, Table 4 shows that the system using the tree-to-tree grammar used the glue rule much less and performed more matching substitutions. That is, in order to minimize errors on the tuning set, the model learned to build syntactically richer and more well-formed derivations. Tables 5 and 6 show how the new syntax features affected particular substitutions. In general we see a shift towards more matching substitutions; correct placement of punctuation is particularly emphasized. Several changes appear to have to do with definiteness of NPs: on the English side, adding the syntax features encourages matching substitutions of type DT \ NP-C (anarthrous NP), but discourages DT \ NP-C and NN from substituting into NP-C and vice versa. For example, a translation with the rewriting NP-C → DT \ NP-C begins with “24th meeting of the Standing Committee. . .,” but the system using the fuzzy tree-to-tree grammar changes this to “The 24th meeting of the Standing Committee. . . .” The root features had a less noticeable effect on 1449 BLEU task extraction dist. lim. rules features dev test Chi-Eng string-to-string 10 440M 1k 32.7 23.4 string-to-string 20 440M 1k 33.3 23.7 ] tree-to-tree exact 20 50M 5k 32.8 23.9 tree-to-tree fuzzy 20 440M 160k 33.9 ] 24.3 ] + nesting 20 180M 79k 33.9 24.3 Ara-Eng string-to-string 10 790M 1k 48.7 48.9 tree-to-tree exact 10 38M 5k 46.6 47.5 tree-to-tree fuzzy 10 790M 130k 49.4 49.7 ] + nesting 10 190M 66k 49.2 49.8 Table 3: On both the Chinese-English and Arabic-English translation tasks, fuzzy tree-to-tree extraction outperforms exact tree-to-tree extraction and string-to-string extraction. Brackets indicate statistically insignificant differences (p ≥0.05). rule choice; one interesting change was that the frequency of rules with Chinese root VP / IP and English root VP / S-C increased from 0.2% to 0.7%: apparently the model learned that it is good to use rules that pair Chinese and English verbs that subcategorize for sentential complements. 5 Conclusion Though exact tree-to-tree translation tends to hamper translation quality by imposing too many constraints during both grammar extraction and decoding, we have shown that using both source and target syntax improves translation accuracy when the model is given the opportunity to learn from data how strongly to apply syntactic constraints. Indeed, we have found that the model learns on its own to choose syntactically richer and more wellformed structures, demonstrating that source- and target-side syntax can be used together profitably as long as they are not allowed to overconstrain the translation model. Acknowledgements Thanks to Steve DeNeefe, Adam Lopez, Jonathan May, Miles Osborne, Adam Pauls, Richard Schwartz, and the anonymous reviewers for their valuable help. This research was supported in part by DARPA contract HR0011-06-C-0022 under subcontract to BBN Technologies and DARPA contract HR0011-09-1-0028. S. D. G. frequency (%) task side kind s-to-s t-to-t Chi-Eng source glue 25 18 match 17 30 mismatch 58 52 target glue 25 18 match 9 23 mismatch 66 58 Ara-Eng source glue 36 19 match 17 34 mismatch 48 47 target glue 36 19 match 11 29 mismatch 53 52 Table 4: Moving from string-to-string (s-to-s) extraction to fuzzy tree-to-tree (t-to-t) extraction decreases glue rule usage and increases the frequency of matching substitutions. 1450 frequency (%) kind s-to-s t-to-t NP →NP 16.0 20.7 VP →VP 3.3 5.9 NN →NP 3.1 1.3 NP →VP 2.5 0.8 NP →NN 2.0 1.4 NP →entity 1.4 1.6 NN →NN 1.1 1.0 QP →entity 1.0 1.3 VV →VP 1.0 0.7 PU →NP 0.8 1.1 VV →VP ∗PU 0.2 1.2 PU →PU 0.1 3.8 Table 5: Comparison of frequency of source-side rewrites in Chinese-English translation between string-to-string (s-to-s) and fuzzy tree-to-tree (t-tot) grammars. All rewrites occurring more than 1% of the time in either system are shown. The label “entity” stands for handwritten rules for named entities and numbers. frequency (%) kind s-to-s t-to-t NP-C →NP-C 5.3 8.7 NN →NN 1.7 3.0 NP-C →entity 1.1 1.4 DT \ NP-C →DT \ NP-C 1.1 2.6 NN →NP-C 0.8 0.4 NP-C →VP 0.8 1.1 DT \ NP-C →NP-C 0.8 0.5 NP-C →DT \ NP-C 0.6 0.4 JJ →JJ 0.5 1.8 NP-C →NN 0.5 0.3 PP →PP 0.4 1.7 VP-C →VP-C 0.4 1.2 VP →VP 0.4 1.4 IN →IN 0.1 1.8 , →, 0.1 1.7 Table 6: Comparison of frequency of target-side rewrites in Chinese-English translation between string-to-string (s-to-s) and fuzzy tree-to-tree (tto-t) grammars. All rewrites occurring more than 1% of the time in either system are shown, plus a few more of interest. The label “entity” stands for handwritten rules for named entities and numbers. References Vamshi Ambati and Alon Lavie. 2008. Improving syntax driven translation models by re-structuring divergent and non-isomorphic parse tree structures. In Proc. AMTA-2008 Student Research Workshop, pages 235–244. Yehoshua Bar-Hillel. 1953. A quasi-arithmetical notation for syntactic description. Language, 29(1):47–58. Stanley F. Chen and Joshua Goodman. 1998. An empirical study of smoothing techniques for language modeling. Technical Report TR-10-98, Harvard University Center for Research in Computing Technology. David Chiang, Wei Wang, and Kevin Knight. 2009. 11,001 new features for statistical machine translation. In Proc. NAACL HLT 2009, pages 218–226. David Chiang. 2005. A hierarchical phrasebased model for statistical machine translation. In Proc. ACL 2005, pages 263–270. David Chiang. 2007. Hierarchical phrase-based translation. Computational Linguistics, 33(2):201–228. Michael Collins. 1997. Three generative lexicalised models for statistical parsing. In Proc. ACL-EACL, pages 16–23. Koby Crammer and Yoram Singer. 2003. Ultraconservative online algorithms for multiclass problems. Journal of Machine Learning Research, 3:951–991. Jason Eisner. 2003. Learning non-isomorphic tree mappings for machine translation. In Proc. ACL 2003 Companion Volume, pages 205–208. Victoria Fossum, Kevin Knight, and Steven Abney. 2008. Using syntax to improve word alignment for syntax-based statistical machine translation. In Proc. Third Workshop on Statistical Machine Translation, pages 44–52. Alexander Fraser and Daniel Marcu. 2007. Getting the structure right for word alignment: LEAF. In Proc. EMNLP 2007, pages 51–60. Michel Galley, Mark Hopkins, Kevin Knight, and Daniel Marcu. 2004. What’s in a translation rule? In Proc. HLT-NAACL 2004, pages 273–280. Michel Galley, Jonathan Graehl, Kevin Knight, Daniel Marcu, Steve DeNeefe, Wei Wang, and Ignacio Thayer. 2006. Scalable inference and training of context-rich syntactic translation models. In Proc. COLING-ACL 2006, pages 961–968. Mary Hearne and Andy Way. 2003. Seeing the wood for the trees: Data-Oriented Translation. In Proc. MT Summit IX, pages 165–172. 1451 Liang Huang, Kevin Knight, and Aravind Joshi. 2006. Statistical syntax-directed translation with extended domain of locality. In Proc. AMTA 2006, pages 65–73. Alon Lavie, Alok Parlikar, and Vamshi Ambati. 2008. Syntax-driven learning of sub-sentential translation equivalents and translation rules from parsed parallel corpora. In Proc. SSST-2, pages 87–95. Yang Liu, Qun Liu, and Shouxun Lin. 2006. Treeto-string alignment template for statistical machine translation. In Proc. COLING-ACL 2006, pages 609–616. Yang Liu, Yajuan L¨u, and Qun Liu. 2009. Improving tree-to-tree translation with packed forests. In Proc. ACL 2009, pages 558–566. I. Dan Melamed, Giorgio Satta, and Ben Wellington. 2004. Generalized multitext grammars. In Proc. ACL 2004, pages 661–668. Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. 2006. Learning accurate, compact, and interpretable tree annotation. In Proc. COLING-ACL 2006, pages 433–440. Arjen Poutsma. 2000. Data-Oriented Translation. In Proc. COLING 2000, pages 635–641. David Talbot and Thorsten Brants. 2008. Randomized language models via perfect hash functions. In Proc. ACL-08: HLT, pages 505–513. Ashish Venugopal, Andreas Zollmann, Noah A. Smith, and Stephan Vogel. 2009. Preference grammars: Softening syntactic constraints to improve statistical machine translation. In Proc. NAACL HLT 2009, pages 236–244. David J. Weir. 1988. Characterizing Mildly ContextSensitive Grammar Formalisms. Ph.D. thesis, University of Pennsylvania. Benjamin Wellington, Sonjia Waxmonsky, and I. Dan Melamed. 2006. Empirical lower bounds on the complexity of translational equivalence. In Proc. COLING-ACL 2006, pages 977–984. Min Zhang, Hongfei Jiang, Aiti Aw, Haizhou Li, Chew Lim Tan, and Sheng Li. 2008. A tree sequence alignment-based tree-to-tree translation model. In Proc. ACL-08: HLT, pages 559–567. Andreas Zollmann and Ashish Venugopal. 2006. Syntax augmented machine translation via chart parsing. In Proc. Workshop on Statistical Machine Translation, pages 138–141. 1452
|
2010
|
146
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1453–1463, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Discriminative Modeling of Extraction Sets for Machine Translation John DeNero and Dan Klein Computer Science Division University of California, Berkeley {denero,klein}@cs.berkeley.edu Abstract We present a discriminative model that directly predicts which set of phrasal translation rules should be extracted from a sentence pair. Our model scores extraction sets: nested collections of all the overlapping phrase pairs consistent with an underlying word alignment. Extraction set models provide two principle advantages over word-factored alignment models. First, we can incorporate features on phrase pairs, in addition to word links. Second, we can optimize for an extraction-based loss function that relates directly to the end task of generating translations. Our model gives improvements in alignment quality relative to state-of-the-art unsupervised and supervised baselines, as well as providing up to a 1.4 improvement in BLEU score in Chinese-to-English translation experiments. 1 Introduction In the last decade, the field of statistical machine translation has shifted from generating sentences word by word to systems that recycle whole fragments of training examples, expressed as translation rules. This general paradigm was first pursued using contiguous phrases (Och et al., 1999; Koehn et al., 2003), and has since been generalized to a wide variety of hierarchical and syntactic formalisms. The training stage of statistical systems focuses primarily on discovering translation rules in parallel corpora. Most systems discover translation rules via a two-stage pipeline: a parallel corpus is aligned at the word level, and then a second procedure extracts fragment-level rules from word-aligned sentence pairs. This paper offers a model-based alternative to phrasal rule extraction, which merges this two-stage pipeline into a single step. We present a discriminative model that directly predicts which set of phrasal translation rules should be extracted from a sentence pair. Our model predicts extraction sets: combinatorial objects that include the set of all overlapping phrasal translation rules consistent with an underlying word-level alignment. This approach provides additional discriminative power relative to word aligners because extraction sets are scored based on the phrasal rules they contain in addition to word-to-word alignment links. Moreover, the structure of our model directly reflects the purpose of alignment models in general, which is to discover translation rules. We address several challenges to training and applying an extraction set model. First, we would like to leverage existing word-level alignment resources. To do so, we define a deterministic mapping from word alignments to extraction sets, inspired by existing extraction procedures. In our mapping, possible alignment links have a precise interpretation that dictates what phrasal translation rules can be extracted from a sentence pair. This mapping allows us to train with existing annotated data sets and use the predictions from word-level aligners as features in our extraction set model. Second, our model solves a structured prediction problem, and the choice of loss function during training affects model performance. We optimize for a phrase-level F-measure in order to focus learning on the task of predicting phrasal rules rather than word alignment links. Third, our discriminative approach requires that we perform inference in the space of extraction sets. Our model does not factor over disjoint wordto-word links or minimal phrase pairs, and so existing inference procedures do not directly apply. However, we show that the dynamic program for a block ITG aligner can be augmented to score extraction sets that are indexed by underlying ITG word alignments (Wu, 1997). We also describe a 1453 2月 15日 2010年 On February 15 2010 2月 15日 2010年 On February 15 2010 σ(ei) σ(f2) σ(e1) (a) (b) Type 1: Language-specific function words omitted in the other language Type 2: Role-equivalent word pairs that are not lexical equivalents 过 地球 [go over] [Earth] over the Earth 65% 31% 被 发现 [passive marker] [discover] was discovered Distribution over possible link types σ(fj) 年 过去 两 中 [past] [two] [year] [in] Figure 1: A word alignment A (shaded grid cells) defines projections σ(ei) and σ(fj), shown as dotted lines for each word in each sentence. The extraction set R3(A) includes all bispans licensed by these projections, shown as rounded rectangles. coarse-to-fine inference approach that allows us to scale our method to long sentences. Our extraction set model outperforms both unsupervised and supervised word aligners at predicting word alignments and extraction sets. We also demonstrate that extraction sets are useful for end-to-end machine translation. Our model improves translation quality relative to state-of-theart Chinese-to-English baselines across two publicly available systems, providing total BLEU improvements of 1.2 in Moses, a phrase-based system, and 1.4 in a Joshua, a hierarchical system (Koehn et al., 2007; Li et al., 2009) 2 Extraction Set Models The input to our model is an unaligned sentence pair, and the output is an extraction set of phrasal translation rules. Word-level alignments are generated as a byproduct of inference. We first specify the relationship between word alignments and extraction sets, then define our model. 2.1 Extraction Sets from Word Alignments Rule extraction is a standard concept in machine translation: word alignment constellations license particular sets of overlapping rules, from which subsets are selected according to limits on phrase length (Koehn et al., 2003), number of gaps (Chiang, 2007), count of internal tree nodes (Galley et al., 2006), etc. In this paper, we focus on phrasal rule extraction (i.e., phrase pair extraction), upon which most other extraction procedures are based. Given a sentence pair (e, f), phrasal rule extraction defines a mapping from a set of word-to-word 2月 15日 2010年 On February 15 2010 2月 15日 2010年 On February 15 2010 σ(ei) σ(f2) σ(e1) Type 1: Language-specific function words omitted in the other language Type 2: Role-equivalent pairs that are not lexical equivalents 过 地球 [go over] [Earth] over the Earth 65% 31% 被 发现 [passive marker] [discover] was discovered Distribution over possible link types σ(fj) ast] wo] ear] n] PDT 饭 [after] [dinner] [after] [I] [sleep] [past tense] Figure 2: Examples of two types of possible alignment links (striped). These types account for 96% of the possible alignment links in our data set. alignment links A = {(i, j)} to an extraction set of bispans Rn(A) = {[g, h) ⇔[k, ℓ)}, where each bispan links target span [g, h) to source span [k, ℓ).1 The maximum phrase length n ensures that max(h −g, ℓ−k) ≤n. We can describe this mapping via word-tophrase projections, as illustrated in Figure 1. Let word ei project to the phrasal span σ(ei), where σ(ei) = min j∈Ji j , max j∈Ji j + 1 (1) Ji = {j : (i, j) ∈A} and likewise each word fj projects to a span of e. Then, Rn(A) includes a bispan [g, h) ⇔[k, ℓ) iff σ(ei) ⊆[k, ℓ) ∀i ∈[g, h) σ(fj) ⊆[g, h) ∀j ∈[k, ℓ) That is, every word in one of the phrasal spans must project within the other. This mapping is deterministic, and so we can interpret a word-level alignment A as also specifying the phrasal rules that should be extracted from a sentence pair. 2.2 Possible and Null Alignment Links We have not yet accounted for two special cases in annotated corpora: possible alignments and null alignments. To analyze these annotations, we consider a particular data set: a hand-aligned portion 1We use the fencepost indexing scheme used commonly for parsing. Words are 0-indexed. Spans are inclusive on the lower bound and exclusive on the upper bound. For example, the span [0, 2) includes the first two words of a sentence. 1454 of the NIST MT02 Chinese-to-English test set, which has been used in previous alignment experiments (Ayan et al., 2005; DeNero and Klein, 2007; Haghighi et al., 2009). Possible links account for 22% of all alignment links in these data, and we found that most of these links fall into two categories. First, possible links are used to align function words that have no equivalent in the other language, but colocate with aligned content words, such as English determiners. Second, they are used to mark pairs of words or short phrases that are not lexical equivalents, but which play equivalent roles in each sentence. Figure 2 shows examples of these two use cases, along with their corpus frequencies.2 On the other hand, null alignments are used sparingly in our annotated data. More than 90% of words participate in some alignment link. The unaligned words typically express content in one sentence that is absent in its translation. Figure 3 illustrates how we interpret possible and null links in our projection. Possible links are typically not included in extraction procedures because most aligners predict only sure links. However, we see a natural interpretation for possible links in rule extraction: they license phrasal rules that both include and exclude them. We exclude null alignments from extracted phrases because they often indicate a mismatch in content. We achieve these effects by redefining the projection operator σ. Let A(s) be the subset of A that are sure links, then let the index set Ji used for projection σ in Equation 1 be Ji = j : (i, j) ∈A(s) if ∃j : (i, j) ∈A(s) {−1, |f|} if ∄j : (i, j) ∈A {j : (i, j) ∈A} otherwise Here, Ji is a set of integers, and σ(ei) for null aligned ei will be [−1, |f| + 1) by Equation 1. Of course, the characteristics of our aligned corpus may not hold for other annotated corpora or other language pairs. However, we hope that the overall effectiveness of our modeling approach will influence future annotation efforts to build corpora that are consistent with this interpretation. 2.3 A Linear Model of Extraction Sets We now define a linear model that scores extraction sets. We restrict our model to score only co2We collected corpus frequencies of possible alignment link types ourselves on a sample of the hand-aligned data set. 2月 15日 2010年 On February 15 2010 σ(f2) σ(e1) 发现[discover] was discovered 中 PDT 在 晚饭 后 我 睡 了 [after] [dinner] [after] [I] [sleep] (past) Figure 3: Possible links constrain the word-tophrase projection of otherwise unaligned words, which in turn license overlapping phrases. In this example, σ(f2) = [1, 2) does not include the possible link at (1, 0) because of the sure link at (1, 1), but σ(e1) = [1, 2) does use the possible link because it would otherwise be unaligned. The word “PDT” is null aligned, and so its projection σ(e4) = [−1, 4) extends beyond the bounds of the sentence, excluding “PDT” from all phrase pairs. herent extraction sets Rn(A), those that are licensed by an underlying word alignment A with sure alignments A(s) ⊆A. Conditioned on a sentence pair (e, f) and maximum phrase length n, we score extraction sets via a feature vector φ(A(s), Rn(A)) that includes features on sure links (i, j) ∈A(s) and features on the bispans in Rn(A) that link [g, h) in e to [k, ℓ) in f: φ(A(s), Rn(A)) = X (i,j)∈A(s) φa(i, j) + X [g,h)⇔[k,ℓ)∈Rn(A) φb(g, h, k, ℓ) Because the projection operator Rn(·) is a deterministic function, we can abbreviate φ(A(s), Rn(A)) as φ(A) without loss of information, although we emphasize that A is a set of sure and possible alignments, and φ(A) does not decompose as a sum of vectors on individual word-level alignment links. Our model is parameterized by a weight vector θ, which scores an extraction set Rn(A) as θ · φ(A). To further limit the space of extraction sets we are willing to consider, we restrict A to block inverse transduction grammar (ITG) alignments, a space that allows many-to-many alignments through phrasal terminal productions, but otherwise enforces at-most-one-to-one phrase matchings with ITG reordering patterns (Cherry and Lin, 2007; Zhang et al., 2008). The ITG constraint 1455 On February 15 2010 2月 15日 2010年 On February 15 2010 σ(f2) σ(e1) 被 发现 [passive marker] [discover] was discovered n] PDT 饭 [after] [dinner] [after] [I] [sleep] [past tense] Figure 4: Above, we show a representative subset of the block alignment patterns that serve as terminal productions of the ITG that restricts the output space of our model. These terminal productions cover up to n = 3 words in each sentence and include a mixture of sure (filled) and possible (striped) word-level alignment links. is more computationally convenient than arbitrarily ordered phrase matchings (Wu, 1997; DeNero and Klein, 2008). However, the space of block ITG alignments is expressive enough to include the vast majority of patterns observed in handannotated parallel corpora (Haghighi et al., 2009). In summary, our model scores all Rn(A) for A ∈ITG(e, f) where A can include block terminals of size up to n. In our experiments, n = 3. Unlike previous work, we allow possible alignment links to appear in the block terminals, as depicted in Figure 4. 3 Model Estimation We estimate the weights θ of our extraction set model discriminatively using the margin-infused relaxed algorithm (MIRA) of Crammer and Singer (2003)—a large-margin, perceptron-style, online learning algorithm. MIRA has been used successfully in MT to estimate both alignment models (Haghighi et al., 2009) and translation models (Chiang et al., 2008). For each training example, MIRA requires that we find the alignment Am corresponding to the highest scoring extraction set Rn(Am) under the current model, Am = arg maxA∈ITG(e,f)θ · φ(A) (2) Section 4 describes our approach to solving this search problem for model inference. MIRA updates away from Rn(Am) and toward a gold extraction set Rn(Ag). Some handannotated alignments are outside of the block ITG model class. Hence, we update toward the extraction set for a pseudo-gold alignment Ag ∈ ITG(e, f) with minimal distance from the true reference alignment At. Ag = arg minA∈ITG(e,f)|A ∪At −A ∩At| (3) Inference details appear in Section 4.3. Given Ag and Am, we update the model parameters away from Am and toward Ag. θ ←θ + τ · (φ(Ag) −φ(Am)) where τ is the minimal step size that will ensure we prefer Ag to Am by a margin greater than the loss L(Am; Ag), capped at some maximum update size C to provide regularization. We use C = 0.01 in experiments. The step size is a closed form function of the loss and feature vectors: τ = min C, L(Am; Ag) −θ · (φ(Ag) −φ(Am)) ||φ(Ag) −φ(Am)||2 2 We train the model for 30 iterations over the training set, shuffling the order each time, and we average the weight vectors observed after each iteration to estimate our final model. 3.1 Extraction Set Loss Function In order to focus learning on predicting the right bispans, we use an extraction-level loss L(Am; Ag): an F-measure of the overlap between bispans in Rn(Am) and Rn(Ag). This measure has been proposed previously to evaluate alignment systems (Ayan and Dorr, 2006). Based on preliminary translation results during development, we chose bispan F5 as our loss: Pr(Am) = |Rn(Am) ∩Rn(Ag)|/|Rn(Am)| Rc(Am) = |Rn(Am) ∩Rn(Ag)|/|Rn(Ag)| F5(Am; Ag) = (1 + 52) · Pr(Am) · Rc(Am) 52 · Pr(Am) + Rc(Am) L(Am; Ag) = 1 −F5(Am; Ag) F5 favors recall over precision. Previous alignment work has shown improvements from adjusting the F-measure parameter (Fraser and Marcu, 2006). In particular, Lacoste-Julien et al. (2006) also chose a recall-biased objective. Optimizing for a bispan F-measure penalizes alignment mistakes in proportion to their rule extraction consequences. That is, adding a word link that prevents the extraction of many correct phrasal rules, or which licenses many incorrect rules, is strongly discouraged by this loss. 1456 3.2 Features on Extraction Sets The discriminative power of our model is driven by the features on sure word alignment links φa(i, j) and bispans φb(g, h, k, ℓ). In both cases, the most important features come from the predictions of unsupervised models trained on large parallel corpora, which provide frequency and cooccurrence information. To score word-to-word links, we use the posterior predictions of a jointly trained HMM alignment model (Liang et al., 2006). The remaining features include a dictionary feature, an identical word feature, an absolute position distortion feature, and features for numbers and punctuation. To score phrasal translation rules in an extraction set, we use a mixture of feature types. Extraction set models allow us to incorporate the same phrasal relative frequency statistics that drive phrase-based translation performance (Koehn et al., 2003). To implement these frequency features, we extract a phrase table from the alignment predictions of a jointly trained unsupervised HMM model using Moses (Koehn et al., 2007), and score bispans using the resulting features. We also include indicator features on lexical templates for the 50 most common words in each language, as in Haghighi et al. (2009). We include indicators for the number of words and Chinese characters in rules. One useful indicator feature exploits the fact that capitalized terms in English tend to align to Chinese words with three or more characters. On 1-by-n or n-by-1 phrasal rules, we include indicator features of fertility for common words.3 We also include monolingual phrase features that expose useful information to the model. For instance, English bigrams beginning with “the” are often extractable phrases. English trigrams with a hyphen as the second word are typically extractable, meaning that the first and third words align to consecutive Chinese words. When any conjugation of the word “to be” is followed by a verb, indicating passive voice or progressive tense, the two words tend to align together. Our feature set also includes bias features on phrasal rules and links, which control the number of null-aligned words and number of rules licensed. In total, our final model includes 4,249 individual features, dominated by various instantiations of lexical templates. 3Limiting lexicalized features to common words helps prevent overfitting. k l g h σ(ei) σ(fj) 年 过去 两 中 In the past two years [past] [two] [year] [in] After dinner I slept 在 晚饭 后 我 睡 了 [after] [dinner] [after] [I] [sleep] [past tense] k =2 l =4 g =1 h =3 or Figure 5: Both possible ITG decompositions of this example alignment will split one of the two highlighted bispans across constituents. 4 Model Inference Equation 2 asks for the highest scoring extraction set under our model, Rn(Am), which we also require at test time. Although we have restricted Am ∈ITG(e, f), our extraction set model does not factor over ITG productions, and so the dynamic program for a vanilla block ITG will not suffice to find Rn(Am). To see this, consider the extraction set in Figure 5. An ITG decomposition of the underlying alignment imposes a hierarchical bracketing on each sentence, and some bispan in the extraction set for this alignment will cross any such bracketing. Hence, the score of some licensed bispan will be non-local to the ITG decomposition. 4.1 A Dynamic Program for Extraction Sets If we treat the maximum phrase length n as a fixed constant, then we can define a dynamic program to search the space of extraction sets. An ITG derivation for some alignment A decomposes into two sub-derivations for AL and AR.4 The model score of A, which scores extraction set Rn(A), decomposes over AL and AR, along with any phrasal bispans licensed by adjoining AL and AR. θ · φ(A) = θ · φ(AL) + θ · φ(AR) + I(AL, AR) where I(AL, AR) is θ · P φ(g, h, k, l) summed over licensed bispans [g, h) ⇔[k, ℓ) that overlap the boundary between AL and AR.5 4We abuse notation in conflating an alignment A with its derivation. All derivations of the same alignment receive the same score, and we only compute the max, not the sum. 5We focus on the case of adjoining two aligned bispans. Our algorithm easily extends to include null alignments, but we focus on the non-null setting for simplicity. 1457 k l g h On February 15 2010 σ 中 In the past two years [in] After dinner I slept 在 晚饭 后 我 睡 了 [after] [dinner] [after] [I] [sleep] (past) k =2 l =4 g =1 h =3 Figure 6: Augmenting the ITG grammar states with the alignment configuration in an n −1 deep perimeter of the bispan allows us to score all overlapping phrasal rules introduced by adjoining two bispans. The state must encode whether a sure link appears in each edge column or row, but the specific location of edge links is not required. In order to compute I(AL, AR), we need certain information about the alignment configurations of AL and AR where they adjoin at a corner. The state must represent (a) the specific alignment links in the n −1 deep corner of each A, and (b) whether any sure alignments appear in the rows or columns extending from those corners.6 With this information, we can infer the bispans licensed by adjoining AL and AR, as in Figure 6. Applying our score recurrence yields a polynomial-time dynamic program. This dynamic program is an instance of ITG bitext parsing, where the grammar uses symbols to encode the alignment contexts described above. This context-as-symbol augmentation of the grammar is similar in character to augmenting symbols with lexical items to score language models during hierarchical decoding (Chiang, 2007). 4.2 Coarse-to-Fine Inference and Pruning Exhaustive inference under an ITG requires O(k6) time in sentence length k, and is prohibitively slow when there is no sparsity in the grammar. Maintaining the context necessary to score non-local bispans further increases running time. That is, ITG inference is organized around search states associated with a grammar symbol and a bispan; augmenting grammar symbols also augments this state space. To parse quickly, we prune away search states using predictions from the more efficient HMM 6The number of configuration states does not depend on the size of A because corners have fixed size, and because the position of links within rows or columns is not needed. alignment model (Ney and Vogel, 1996). We discard all states corresponding to bispans that are incompatible with 3 or more alignment links under an intersected HMM—a proven approach to pruning the space of ITG alignments (Zhang and Gildea, 2006; Haghighi et al., 2009). Pruning in this way reduces the search space dramatically, but only rarely prohibits correct alignments. The oracle alignment error rate for the block ITG model class is 1.4%; the oracle alignment error rate for this pruned subset of ITG is 2.0%. To take advantage of the sparsity that results from pruning, we use an agenda-based parser that orders search states from small to large, where we define the size of a bispan as the total number of words contained within it. For each size, we maintain a separate agenda. Only when the agenda for size k is exhausted does the parser proceed to process the agenda for size k + 1. We also employ coarse-to-fine search to speed up inference (Charniak and Caraballo, 1998). In the coarse pass, we search over the space of ITG alignments, but score only features on alignment links and bispans that are local to terminal blocks. This simplification eliminates the need to augment grammar symbols, and so we can exhaustively explore the (pruned) space. We then compute outside scores for bispans under a max-sum semiring (Goodman, 1996). In the fine pass with the full extraction set model, we impose a maximum size of 10,000 for each agenda. We order states on agendas by the sum of their inside score under the full model and the outside score computed in the coarse pass, pruning all states not within the fixed agenda beam size. Search states that are popped off agendas are indexed by their corner locations for fast lookup when constructing new states. For each corner and size combination, built states are maintained in sorted order according to their inside score. This ordering allows us to stop combining states early when the results are falling off the agenda beams. Similar search and beaming strategies appear in many decoders for machine translation (Huang and Chiang, 2007; Koehn and Haddow, 2009; Moore and Quirk, 2007). 4.3 Finding Pseudo-Gold ITG Alignments Equation 3 asks for the block ITG alignment Ag that is closest to a reference alignment At, which may not lie in ITG(e,f). We search for 1458 l On Febru σ(f2) After dinner I slept 在 晚饭 后 我 睡 了 [after] [dinner] [after] [I] [sleep] [past tense] k =1 l =4 g =0 h =3 Figure 7: A* search for pseudo-gold ITG alignments uses an admissible heuristic for bispans that counts the number of gold links outside of [k, ℓ) but within [g, h). Above, the heuristic is 1, which is also the minimal number of alignment errors that an ITG alignment will incur using this bispan. Ag using A* bitext parsing (Klein and Manning, 2003). Search states, which correspond to bispans [g, h) ⇔[k, ℓ), are scored by the number of errors within the bispan plus the number of (i, j) ∈At such that j ∈[k, ℓ) but i /∈[g, h) (recall errors). As an admissible heuristic for the future cost of a bispan [g, h) ⇔[k, ℓ), we count the number of (i, j) ∈At such that i ∈[g, h) but j /∈[k, ℓ), as depicted in Figure 7. These links will become recall errors eventually. A* search with this heuristic makes no errors, and the time required to compute pseudo-gold alignments is negligible. 5 Relationship to Previous Work Our model is certainly not the first alignment approach to include structures larger than words. Model-based phrase-to-phrase alignment was proposed early in the history of phrase-based translation as a method for training translation models (Marcu and Wong, 2002). A variety of unsupervised models refined this initial work with priors (DeNero et al., 2008; Blunsom et al., 2009) and inference constraints (DeNero et al., 2006; Birch et al., 2006; Cherry and Lin, 2007; Zhang et al., 2008). These models fundamentally differ from ours in that they stipulate a segmentation of the sentence pair into phrases, and only align the minimal phrases in that segmentation. Our model scores the larger overlapping phrases that result from composing these minimal phrases. Discriminative alignment is also a wellexplored area. Most work has focused on predicting word alignments via partial matching inference algorithms (Melamed, 2000; Taskar et al., 2005; Moore, 2005; Lacoste-Julien et al., 2006). Work in semi-supervised estimation has also contributed evidence that hand-annotations are useful for training alignment models (Fraser and Marcu, 2006; Fraser and Marcu, 2007). The ITG grammar formalism, the corresponding word alignment class, and inference procedures for the class have also been explored extensively (Wu, 1997; Zhang and Gildea, 2005; Cherry and Lin, 2007; Zhang et al., 2008). At the intersection of these lines of work, discriminative ITG models have also been proposed, including one-to-one alignment models (Cherry and Lin, 2006) and block models (Haghighi et al., 2009). Our model directly extends this research agenda with first-class possible links, overlapping phrasal rule features, and an extraction-level loss function. K¨a¨ari¨ainen (2009) trains a translation model discriminatively using features on overlapping phrase pairs. That work differs from ours in that it uses fixed word alignments and focuses on translation model estimation, while we focus on alignment and translate using standard relative frequency estimators. Deng and Zhou (2009) present an alignment combination technique that uses phrasal features. Our approach differs in two ways. First, their approach is tightly coupled to the input alignments, while we perform a full search over the space of ITG alignments. Also, their approach uses greedy search, while our search is optimal aside from pruning and beaming. Despite these differences, their strong results reinforce our claim that phraselevel information is useful for alignment. 6 Experiments We evaluate our extraction set model by the bispans it predicts, the word alignments it generates, and the translations generated by two end-to-end systems. Table 1 compares the five systems described below, including three baselines. All supervised aligners were optimized for bispan F5. Unsupervised Baseline: GIZA++. We trained GIZA++ (Och and Ney, 2003) using the default parameters included with the Moses training script (Koehn et al., 2007). The designated regimen concludes by Viterbi aligning under Model 4 in both directions. We combined these alignments with 1459 the grow-diag heuristic (Koehn et al., 2003). Unsupervised Baseline: Joint HMM. We trained and combined two HMM alignment models (Ney and Vogel, 1996) using the Berkeley Aligner.7 We initialized the HMM model parameters with jointly trained Model 1 parameters (Liang et al., 2006), combined word-toword posteriors by averaging (soft union), and decoded with the competitive thresholding heuristic of DeNero and Klein (2007), yielding a state-ofthe-art unsupervised baseline. Supervised Baseline: Block ITG. We discriminatively trained a block ITG aligner with only sure links, using block terminal productions up to 3 words by 3 words in size. This supervised baseline is a reimplementation of the MIRA-trained model of Haghighi et al. (2009). We use the same features and parser implementation for this model as we do for our extraction set model to ensure a clean comparison. To remain within the alignment class, MIRA updates this model toward a pseudogold alignment with only sure links. This model does not score any overlapping bispans. Extraction Set Coarse Pass. We add possible links to the output of the block ITG model by adding the mixed terminal block productions described in Section 2.3. This model scores overlapping phrasal rules contained within terminal blocks that result from including or excluding possible links. However, this model does not score bispans that cross bracketing of ITG derivations. Full Extraction Set Model. Our full model includes possible links and features on extraction sets for phrasal bispans with a maximum size of 3. Model inference is performed using the coarseto-fine scheme described in Section 4.2. 6.1 Data In this paper, we focus exclusively on Chinese-toEnglish translation. We performed our discriminative training and alignment evaluations using a hand-aligned portion of the NIST MT02 test set, which consists of 150 training and 191 test sentences (Ayan and Dorr, 2006). We trained the baseline HMM on 11.3 million words of FBIS newswire data, a comparable dataset to those used in previous alignment evaluations on our test set (DeNero and Klein, 2007; Haghighi et al., 2009). 7http://code.google.com/p/berkeleyaligner Our end-to-end translation experiments were tuned and evaluated on sentences up to length 40 from the NIST MT04 and MT05 test sets. For these experiments, we trained on a 22.1 million word parallel corpus consisting of sentences up to length 40 of newswire data from the GALE program, subsampled from a larger data set to promote overlap with the tune and test sets. This corpus also includes a bilingual dictionary. To improve performance, we retrained our aligner on a retokenized version of the hand-annotated data to match the tokenization of our corpus.8 We trained a language model with Kneser-Ney smoothing on 262 million words of newswire using SRILM (Stolcke, 2002). 6.2 Word and Phrase Alignment The first panel of Table 1 gives a word-level evaluation of all five aligners. We use the alignment error rate (AER) measure: precision is the fraction of sure links in the system output that are sure or possible in the reference, and recall is the fraction of sure links in the reference that the system outputs as sure. For this evaluation, possible links produced by our extraction set models are ignored. The full extraction set model performs the best by a small margin, although it was not tuned for word alignment. The second panel gives a phrasal rule-level evaluation, which measures the degree to which these aligners matched the extraction sets of handannotated alignments, R3(At).9 To compete fairly, all models were evaluated on the full extraction sets induced by the word alignments they predicted. Again, the extraction set model outperformed the baselines, particularly on the F5 measure for which these systems were trained. Our coarse pass extraction set model performed nearly as well as the full model. We believe these models perform similarly for two reasons. First, most of the information needed to predict an extraction set can be inferred from word links and phrasal rules contained within ITG terminal productions. Second, the coarse-to-fine inference may be constraining the full phrasal model to predict similar output to the coarse model. This similarity persists in translation experiments. 8All alignment results are reported under the annotated data set’s original tokenization. 9While pseudo-gold approximations to the annotation were used for training, the evaluation is always performed relative to the original human annotation. 1460 Word Bispan BLEU Pr Rc AER Pr Rc F1 F5 Joshua Moses Baseline GIZA++ 72.5 71.8 27.8 69.4 45.4 54.9 46.0 33.8 32.6 models Joint HMM 84.0 76.9 19.6 69.5 59.5 64.1 59.9 34.5 33.2 Block ITG 83.4 83.8 16.4 75.8 62.3 68.4 62.8 34.7 33.6 Extraction Coarse Pass 82.2 84.2 16.9 70.0 72.9 71.4 72.8 35.7 34.2 set models Full Model 84.7 84.0 15.6 69.0 74.2 71.6 74.0 35.9 34.4 Table 1: Experimental results demonstrate that the full extraction set model outperforms supervised and unsupervised baselines in evaluations of word alignment quality, extraction set quality, and translation. In word and bispan evaluations, GIZA++ did not have access to a dictionary while all other methods did. In the BLEU evaluation, all systems used a bilingual dictionary included in the training corpus. The BLEU evaluation of supervised systems also included rule counts from the Joint HMM to compensate for parse failures. 6.3 Translation Experiments We evaluate the alignments predicted by our model using two publicly available, open-source, state-of-the-art translation systems. Moses is a phrase-based system with lexicalized reordering (Koehn et al., 2007). Joshua (Li et al., 2009) is an implementation of Hiero (Chiang, 2007) using a suffix-array-based grammar extraction approach (Lopez, 2007). Both of these systems take word alignments as input, and neither of these systems accepts possible links in the alignments they consume. To interface with our extraction set models, we produced three sets of sure-only alignments from our model predictions: one that omitted possible links, one that converted all possible links to sure links, and one that includes each possible link with 0.5 probability. These three sets were aggregated and rules were extracted from all three. The training set we used for MT experiments is quite heterogenous and noisy compared to our alignment test sets, and the supervised aligners did not handle certain sentence pairs in our parallel corpus well. In some cases, pruning based on consistency with the HMM caused parse failures, which in turn caused training sentences to be skipped. To account for these issues, we added counts of phrasal rules extracted from the baseline HMM to the counts produced by supervised aligners. In Moses, our extraction set model predicts the set of phrases extracted by the system, and so the estimation techniques for the alignment model and translation model both share a common underlying representation: extraction sets. Empirically, we observe a BLEU score improvement of 1.2 over the best unsupervised baseline and 0.8 over the block ITG supervised baseline (Papineni et al., 2002). In Joshua, hierarchical rule extraction is based upon phrasal rule extraction, but abstracts away sub-phrases to create a grammar. Hence, the extraction sets we predict are closely linked to the representation that this system uses to translate. The extraction model again outperformed both unsupervised and supervised baselines, by 1.4 BLEU and 1.2 BLEU respectively. 7 Conclusion Our extraction set model serves to coordinate the alignment and translation model components of a statistical translation system by unifying their representations. Moreover, our model provides an effective alternative to phrase alignment models that choose a particular phrase segmentation; instead, we predict many overlapping phrases, both large and small, that are mutually consistent. In future work, we look forward to developing extraction set models for richer formalisms, including hierarchical grammars. Acknowledgments This project is funded in part by BBN under DARPA contract HR0011-06-C-0022 and by the NSF under grant 0643742. We thank the anonymous reviewers for their helpful comments. References Necip Fazil Ayan and Bonnie J. Dorr. 2006. Going beyond AER: An extensive analysis of word alignments and their impact on MT. In Proceedings of 1461 the Annual Conference of the Association for Computational Linguistics. Necip Fazil Ayan, Bonnie J. Dorr, and Christof Monz. 2005. Neuralign: combining word alignments using neural networks. In Proceedings of the Conference on Human Language Technology and Empirical Methods in Natural Language Processing. Alexandra Birch, Chris Callison-Burch, and Miles Osborne. 2006. Constraining the phrase-based, joint probability statistical translation model. In Proceedings of the Conference for the Association for Machine Translation in the Americas. Phil Blunsom, Trevor Cohn, Chris Dyer, and Miles Osborne. 2009. A Gibbs sampler for phrasal synchronous grammar induction. In Proceedings of the Annual Conference of the Association for Computational Linguistics. Eugene Charniak and Sharon Caraballo. 1998. New figures of merit for best-first probabilistic chart parsing. In Computational Linguistics. Colin Cherry and Dekang Lin. 2006. Soft syntactic constraints for word alignment through discriminative training. In Proceedings of the Annual Conference of the Association for Computational Linguistics. Colin Cherry and Dekang Lin. 2007. Inversion transduction grammar for joint phrasal translation modeling. In Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics Workshop on Syntax and Structure in Statistical Translation. David Chiang, Yuval Marton, and Philip Resnik. 2008. Online large-margin training of syntactic and structural translation features. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. David Chiang. 2007. Hierarchical phrase-based translation. Computational Linguistics. Koby Crammer and Yoram Singer. 2003. Ultraconservative online algorithms for multiclass problems. Journal of Machine Learning Research, 3:951–991. John DeNero and Dan Klein. 2007. Tailoring word alignments to syntactic machine translation. In Proceedings of the Annual Conference of the Association for Computational Linguistics. John DeNero and Dan Klein. 2008. The complexity of phrase alignment problems. In Proceedings of the Annual Conference of the Association for Computational Linguistics: Short Paper Track. John DeNero, Dan Gillick, James Zhang, and Dan Klein. 2006. Why generative phrase models underperform surface heuristics. In Proceedings of the NAACL Workshop on Statistical Machine Translation. John DeNero, Alexandre Bouchard-Cote, and Dan Klein. 2008. Sampling alignment structure under a bayesian translation model. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Yonggang Deng and Bowen Zhou. 2009. Optimizing word alignment combination for phrase table training. In Proceedings of the Annual Conference of the Association for Computational Linguistics: Short Paper Track. Alexander Fraser and Daniel Marcu. 2006. Semisupervised training for statistical word alignment. In Proceedings of the Annual Conference of the Association for Computational Linguistics. Alexander Fraser and Daniel Marcu. 2007. Getting the structure right for word alignment: Leaf. In Proceedings of the Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning. Michel Galley, Jonathan Graehl, Kevin Knight, Daniel Marcu, Steve DeNeefe, Wei Wang, and Ignacio Thayer. 2006. Scalable inference and training of context-rich syntactic translation models. In Proceedings of the Annual Conference of the Association for Computational Linguistics. Joshua Goodman. 1996. Parsing algorithms and metrics. In Proceedings of the Annual Meeting of the Association for Computational Linguistics. Aria Haghighi, John Blitzer, John DeNero, and Dan Klein. 2009. Better word alignments with supervised ITG models. In Proceedings of the Annual Conference of the Association for Computational Linguistics. Liang Huang and David Chiang. 2007. Forest rescoring: Faster decoding with integrated language models. In Proceedings of the Annual Conference of the Association for Computational Linguistics. Matti K¨a¨ari¨ainen. 2009. Sinuhe—statistical machine translation using a globally trained conditional exponential family translation model. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Dan Klein and Chris Manning. 2003. A* parsing: Fast exact Viterbi parse selection. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics. Philipp Koehn and Barry Haddow. 2009. Edinburghs submission to all tracks of the WMT2009 shared task with reordering and speed improvements to Moses. In Proceedings of the Workshop on Statistical Machine Translation. Philipp Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics. 1462 Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the Annual Conference of the Association for Computational Linguistics: Demonstration track. Simon Lacoste-Julien, Ben Taskar, Dan Klein, and Michael I. Jordan. 2006. Word alignment via quadratic assignment. In Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics. Zhifei Li, Chris Callison-Burch, Chris Dyer, Juri Ganitkevitch, Sanjeev Khudanpur, Lane Schwartz, Wren Thornton, Jonathan Weese, and Omar Zaidan. 2009. Joshua: An open source toolkit for parsing-based machine translation. In Proceedings of the Workshop on Statistical Machine Translation. Percy Liang, Ben Taskar, and Dan Klein. 2006. Alignment by agreement. In Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics. Adam Lopez. 2007. Hierarchical phrase-based translation with suffix arrays. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Daniel Marcu and Daniel Wong. 2002. A phrasebased, joint probability model for statistical machine translation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. I. Dan Melamed. 2000. Models of translational equivalence among words. Computational Linguistics. Robert Moore and Chris Quirk. 2007. Faster beam-search decoding for phrasal statistical machine translation. In Proceedings of MT Summit XI. Robert C. Moore. 2005. A discriminative framework for bilingual word alignment. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Hermann Ney and Stephan Vogel. 1996. HMM-based word alignment in statistical translation. In Proceedings of the Conference on Computational linguistics. Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29:19–51. Franz Josef Och, Christoph Tillmann, and Hermann Ney. 1999. Improved alignment models for statistical machine translation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. BLEU: A method for automatic evaluation of machine translation. In Proceedings of the Annual Conference of the Association for Computational Linguistics. Andreas Stolcke. 2002. Srilm an extensible language modeling toolkit. In Proceedings of the International Conference on Spoken Language Processing. Ben Taskar, Simon Lacoste-Julien, and Dan Klein. 2005. A discriminative matching approach to word alignment. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Dekai Wu. 1997. Stochastic inversion transduction grammars and bilingual parsing of parallel corpora. Computational Linguistics, 23:377–404. Hao Zhang and Daniel Gildea. 2005. Stochastic lexicalized inversion transduction grammar for alignment. In Proceedings of the Annual Conference of the Association for Computational Linguistics. Hao Zhang and Daniel Gildea. 2006. Efficient search for inversion transduction grammar. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Hao Zhang, Chris Quirk, Robert C. Moore, and Daniel Gildea. 2008. Bayesian learning of noncompositional phrases with synchronous parsing. In Proceedings of the Annual Conference of the Association for Computational Linguistics. 1463
|
2010
|
147
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1464–1472, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Detecting Experiences from Weblogs Keun Chan Park, Yoonjae Jeong and Sung Hyon Myaeng Department of Computer Science Korea Advanced Institute of Science and Technology {keunchan, hybris, myaeng}@kaist.ac.kr Abstract Weblogs are a source of human activity knowledge comprising valuable information such as facts, opinions and personal experiences. In this paper, we propose a method for mining personal experiences from a large set of weblogs. We define experience as knowledge embedded in a collection of activities or events which an individual or group has actually undergone. Based on an observation that experience-revealing sentences have a certain linguistic style, we formulate the problem of detecting experience as a classification task using various features including tense, mood, aspect, modality, experiencer, and verb classes. We also present an activity verb lexicon construction method based on theories of lexical semantics. Our results demonstrate that the activity verb lexicon plays a pivotal role among selected features in the classification performance and shows that our proposed method outperforms the baseline significantly. 1 Introduction In traditional philosophy, human beings are known to acquire knowledge mainly by reasoning and experience. Reasoning allows us to draw a conclusion based on evidence, but people tend to believe it firmly when they experience or observe it in the physical world. Despite the fact that direct experiences play a crucial role in making a firm decision and solving a problem, people often resort to indirect experiences by reading written materials or asking around. Among many sources people resort to, the Web has become the largest one for human experiences, especially with the proliferation of weblogs. While Web documents contain various types of information including facts, encyclopedic knowledge, opinions, and experiences in general, personal experiences tend to be found in weblogs more often than other web documents like news articles, home pages, and scientific papers. As such, we have begun to see some research efforts in mining experience-related attributes such as time, location, topic, and experiencer, and their relations from weblogs (Inui et al., 2008; Kurashima et al., 2009). Mined experiences can be of practical use in wide application areas. For example, a collection of experiences from the people who visited a resort area would help planning what to do and how to do things correctly without having to spend time sifting through a variety of resources or rely on commercially-oriented sources. Another example would be a public service department gleaning information about how a park is being used at a specific location and time. Experiences can be recorded around a frame like “who did what, when, where, and why” although opinions and emotions can be also linked. Therefore attributes such as location, time, and activity and their relations must be extracted by devising a method for selecting experiencecontaining sentences based on verbs that have a particular linguistics case frame or belong to a “do” class (Kurashima et al., 2009). However, this kind of method may extract the following sentences as containing an experience: [1] If Jason arrives on time, I’ll buy him a drink. [2] Probably, she will laugh and dance in his funeral. [3] Can anyone explain what is going on here? [4] Don’t play soccer on the roads! None of the sentences contain actual experiences because hypotheses, questions, and orders have not actually happened in the real world. For experience mining, it is important to ensure a sentence mentions an event or passes a factuality test to contain experience (Inui et al., 2008). In this paper, we focus on the problem of detecting experiences from weblogs. We formulate 1464 Class Examples State like, know, believe Activity run, swim, walk Achievement recognize, realize Accomplishment paint (a picture), build (a house) Table 1. Vendler class examples the problem as a classification task using various linguistic features including tense, mood, aspect, modality, experiencer, and verb classes. Based on our observation that experiencerevealing sentences tend to have a certain linguistic style (Jijkoun et al., 2010), we investigate on the roles of various features. The ability to detect experience-revealing sentences should be a precursor for ensuring the quality of extracting various elements of actual experiences. Another issue addressed in this paper is automatic construction of a lexicon for verbs related to activities and events. While there have been well-known studies about classifying verbs based on aspectual features (Vendler, 1967), thematic roles and selectional restrictions (Fillmore, 1968; Somers, 1987; Kipper et al., 2008), valence alternations and intuitions (Levin, 1993) and conceptual structures (Fillmore and Baker, 2001), we found that none of the existing lexical resources such as Framenet (Baker et al., 2003) and Verbnet (Kipper et al., 2008) are sufficient for identifying experience-revealing verbs. We introduce a method for constructing an activity/event verb lexicon based on Vendler’s theory and statistics obtained by utilizing a web search engine. We define experience as knowledge embedded in a collection of activities or events which an individual or group has actually undergone1. It can be subjective as in opinions as well as objective, but our focus in this article lies in objective knowledge. The following sentences contain objective experiences: [5] I ran with my wife 3 times a week until we moved to Washington, D.C. [6] Jane and I hopped on a bus into the city centre. [7] We went to a restaurant near the central park. Whereas sentences like the following contain subjective knowledge: [8] I like your new style. You’re beautiful! [9] The food was great, the interior too. Subject knowledge has been studied extensively for various functions such as identification, po 1 http://en.wikipedia.org/wiki/Experience_(disambiguation) larity detection, and holder extraction under the names of opinion mining and sentiment analysis (Pang and Lee, 2008). In summary, our contribution lies in three aspects: 1) conception of experience detection, which is a precursor for experience mining, and specific related tasks that can be tackled with a high performance machine learning based solution; 2) examination and identification of salient linguistic features for experience detection; 3) a novel lexicon construction method with identification of key features to be used for verb classification. The remainder of the paper is organized as follows. Section 2 presents our lexicon construction method with experiments. Section 3 describes the experience detection method, including experimental setup, evaluation, and results. In Section 4, we discuss related work, before we close with conclusion and future work in Section 5. 2 Lexicon Construction Since our definition of experience is based on activities and events, it is critical to determine whether a sentence contains a predicate describing an activity or an event. To this end, it is quite conceivable that a lexicon containing activity / event verbs would play a key role. Given that our ultimate goal is to extract experiences from a large amount of weblogs, we opt for increased coverage by automatically constructing a lexicon rather than high precision obtainable by manually crafted lexicon. Based on the theory of Vendler (1967), we classify a given verb or a verb phrase into one of the two categories: activity and state. We consider all the verbs and verb phrases in WordNet (Fellbaum, 1998) which is the largest electronic lexical database. In addition to the linguistic schemata features based on Vendler’s theory, we used thematic role features and an external knowledge feature. 2.1 Background Vendler (1967) proposes that verb meanings can be categorized into four basic classes, states, activities, achievements, and accomplishments, depending on interactions between the verbs and their aspectual and temporal modifiers. Table 1 shows some examples for the classes. Vendler (1967) and Dowty (1979) introduce linguistic schemata that serve as evidence for the classes. 1465 Linguistic Schemata bs prs prp pts ptp No schema ■ ■ ■ ■ ■ Progressive ■ Force ■ Persuade ■ Stop ■ For ■ ■ ■ ■ ■ Carefully ■ ■ ■ ■ ■ Table 2. Query matrix. The “■” indicates that the query is applied. No Schema indicates that no schema is applied when the word itself is a query. bs, prs, prp, pts, ptp correspond to base form, present simple (3rd person singular), present participle, past simple and past participle, respectfully. Below are the six schemata we chose because they can be tested automatically: progressive, force, persuade, stop, for, and carefully (An asterisk denotes that the statement is awkward). • States cannot occur in progressive tense: John is running. John is liking.* • States cannot occur as complements of force and persuade: John forced harry to run. John forced harry to know.* John persuaded harry to know.* • Achievements cannot occur as complements of stop: John stopped running. John stopped realizing.* • Achievements cannot occur with time adverbial for: John ran for an hour. John realized for an hour.* • State and achievement cannot occur with adverb carefully: John runs carefully. John knows carefully.* The schemata are not perfect because verbs can shift classes due to various contextual factors such as arguments and senses. However, a verb certainly has its fundamental class that is its most natural category at least in its dominant use. The four classes can further be grouped into two genuses: a genus of processes going on in time and the other that refers to non-processes. Activity and accomplishment belong to the former whereas state and achievement belong to the latter. As can be seen in table 1, states are rather immanent operations and achievements are those occur in a single moment or operations related to perception level. On the other hand, activity and accomplishment are processes (transeunt operations) in traditional philosophy. We henceforth call the first genus activity and the latter state. Our aim is to classify verbs into the two genuses. 2.2 Features based on Linguistic Schemata We developed a relatively simple computational testing method for the schemata. Assuming that an awkward expression like, “John is liking something” won’t occur frequently, for example, we generated a co-occurrence based test for the first linguistic schema using the Web as a corpus. By issuing a search query, ((be OR am OR is OR was OR were OR been) and ? ing) where ‘?’ represents the verb at hand, to a search engine, we can get an estimate about how the verb is likely to belong to state. A test can be generated for each of the schemata in a similar way. For completeness, we considered all the verb forms (i.e., 3rd person singular present, present participle, simple past, past participle) available. However, some of the patterns cannot be applied to some forms. For example, other forms except the base form cannot come as a complement of force (e.g., force to runs.*). Therefore, we created a query matrix which represents all query patterns we have applied, in table 2. Based on the query matrix in table 2, we issued queries for all the verbs and verb phrases from WordNet to a search engine. We used the Google news archive search for two reasons. First, since news articles are written rather formally compared to weblogs and other web pages, the statistics obtained for a test would be more reliable. Second, Google provides an advanced option to retrieve snippets containing the query word. Normally, a snippet is composed of 3~5 sentences. The basic statistics we consider are hit count, candidate sentence count and correct sentence count which we use the notations Hij(w), Sij(w), and Cij(w), respectfully, where w is a word, i the linguistic schema and j the verb form from the query matrix in table 2. Hij(w) was directly gathered from the Google search engine. Sij(w) is the number of sentences containing the word w in the search result snippets. Cij(w) is the number of correct sentences matching the query pattern among the candidate sentences. For example, the progressive schema for a verb “build” can retrieve the following sentences. [10] …, New-York, is building one of the largest … [11] Is building an artifact? 1466 “Building” in the first example is a progressive verb, but the one in second is a noun, which does not satisfy the linguistic schema. For a POS and grammatical check of a candidate sentence, we used the Stanford POS tagger (Toutanova et al., 2003) and Stanford dependency parser (Klein and Manning, 2003). For each linguistic schema, we derived three features: Absolute hit ratio, Relative hit ratio and Valid ratio for which we use the notations Ai(w), Ri(w) and Vi(w), respectfully, where w is a word and i a linguistic schema. The index j for summations represents the j-th verb form. They are computed as follows. ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) * ij j i i ij j i NoScheme j ij j i ij j H w A w H H w R w H w C w V w S w = = = ∑ ∑ ∑ ∑ ∑ (1) Absolute hit ratio is computes the extent to which the target word w occurs with the i-th schema over all occurrences of the schema. The denominator is the hit count of wild card “*” matching any single word with the schema pattern from Google (e.g., H1(*), the progressive test hit count is 3.82 × 108). Relative hit ratio computes the extent to which the target word w occurs with the i-th schema over all occurrences of the word. The denominator is the sum of all verb forms. Valid ratio means the fraction of correct sentences among candidate sentences. The weight of a linguistic schema increases as the valid ratio gets high. With the three different ratios, Ai(w), Ri(w) and Vi(w), for each test, we can generate a total of 18 features. 2.3 Features based on case frames Since the hit count via Google API sometimes returns unreliable results (e.g., when the query becomes too long in case of long verb phrases), we also consider additional features. While our initial observation indicated that the existing lexical resources would not be sufficient for our goal, it occurred to us that the linguistic theory behind them would be worth exploring as generating additional features for categorizing verbs for the two classes. Consider the following examples: [12] John(D) believed(V) the story(O). [13] John(A) hit(V) him(O) with a bat(I). The subject of a state verb is dative (D) as in [12] whereas the subject for an action verb takes the agent (A) role. In addition, a verb with the instrument (I) role tends to be an action verb. From these observations, we can use the distribution of cases (thematic roles) for a verb in a corpus. Activity verbs are expected to have high frequency of agent and instrument roles than state verbs. Although a verb may have more than one case frame, it is possible to determine which thematic roles used more dominantly. We utilize two major resources of lexical semantics, Verbnet (Kipper et al., 2008) based on the theory of Levin (1993), and Framenet (Baker et al., 2003), which is based on Fillmore (1968). Levin (1993) demonstrated that syntactic alternations can be the basis for groupings of verbs semantically and accord reasonably well with linguistic intuitions. Verbnet provides 274 verb classes with 23 thematic roles covering 3,769 verbs based on their alternation behaviors with thematic roles annotated. Framenet defines 978 semantic frames with 7,124 unique semantic roles, covering 11,583 words including verbs, nouns, adverbs, etc. Using Verbnet alone does not suit our needs because it has a relatively small number of example sentences. Framenet contains a much larger number of examples but the vast number of semantic roles presents a problem. In order to get meaningful distributions for a manageable number of thematic roles, we used Semlink (Loper et al., 2007) that provides a mapping between Framenet and Verbnet and uses a total of 23 thematic roles of Verbnet for the annotated corpora of the two resources. By the mapping, we obtained distributions of the thematic roles for 2,868 unique verbs that exist in both of the resources. For example, the verb “construct” has high frequencies with agent, material and product roles. 2.4 Features based on how-to instructions Ryu et al. (2010) presented a method for extracting action steps for how-to goals from eHow2 a website containing a large number of how-to instructions. The authors attempted to extract actions comprising a verb and some ingredients like an object entity from the documents based on syntactic patterns and a CRF based model. Since each extracted action has its probability, we can use the value as a feature for state / activity verb classification. However, a verb may appear in different contexts and can have multiple 2 http://www.ehow.com 1467 Feature ME SVM Prec. Recall Prec. Recall All 43 68% 50% 83% 75% Top 30 72% 52% 83% 75% Top 20 83% 76% 85% 77% Top 10 89% 88% 91% 78% Table 3. Classification Performance Class Examples Activity act, battle, build, carry, chase, drive, hike, jump, kick, sky dive, tap dance, walk, … State admire, believe, know, like, love, … Table 4. Classified Examples probability values. To generate a single value for a verb, we combine multiple probability values using the following sigmoid function: 1 ( ) 1 ( ) w t d d D E w e t P w − ∈ = + =∑ (2) Evidence of a word w being an action in eHow is denoted as E(w) where variable t is the sum of individual action probability values in Dw the set of documents from which the word w has been extracted as an action. The higher probability a word gets and the more frequent the word has been extracted as an action, the more evidence we get. 2.5 Classification For training, we selected 80 seed verbs from Dowty’s list (1979) which are representative verbs for each Vendler (1967) class. The selection was based on the lack of word sense ambiguity. One of our classifiers is based on Maximum Entropy (ME) models that implement the intuition that the best model will be the one that is consistent with the set of constraints imposed by the evidence, but otherwise is as uniform as possible (Berger et al., 1996). ME models are widely used in natural language processing tasks for its flexibility to incorporate a diverse range of features. The other one is based on Support Vector Machine (Chang and Lin, 2001) which is the state-of-the-art algorithm for many classification tasks. We used RBF kernel with the default settings (Hsu et al., 2009) because it is been known to show moderate performance using multiple feature compositions. The features we considered are a total of 42 real values: 18 from linguistic schemata, 23 thematic role distributions, and one from eHow. In order to examine which features are discriminative for the classification, we used two well known feature selection methods, Chi-square and information gain. 2.6 Results Table 3 shows the classification performance values for different feature selection methods. The evaluation was done on the training data with 10-fold cross validation. Note that the precision and recall are macroaveraged values across the two classes, activity and state. The most discriminative features were absolute ratio and relative ratio in conjunction with the force, stop, progressive, and persuade schemata, the role distribution of experiencer, and the eHow evidence. It is noteworthy that eHow evidence and the distribution of experiencer got into the top 10. Other thematic roles did not perform well because of the data sparseness. Only a few roles (e.g., experience, agent, topic, location) among the 23 had frequency values other than 0 for many verbs. Data sparseness affected the linguistic schemata as well. Many of the verbs had zero hit counts for the for and carefully schemata. It is also interesting that the validity ratio Vi(w) was not shown to be a good feature-generating statistic. We finally trained our model with the top 10 features and classified all WordNet verbs and verb phrases. For actual construction of the lexicon, 11,416 verbs and verb phrases were classified into the two classes roughly equally. We randomly sampled 200 items and examined how accurately the classification was done. A total of 164 items were correctly classified, resulting in 82% accuracy. Some examples from the classification are shown in table 4. A further analysis of the results show that most of the errors occurred with domain-specific verbs (e.g., ablactate, alkalify, and transaminate in chemistry) and multi-word verb phrases (e.g., turn a nice dime; keep one’s shoulder to the wheel). Since many features are computed based on Web resources, rare verbs cannot be classified correctly when their hit rations are very low. The domain-specific words rarely appear in Framenet or e-how, either. 3 Experience Detection As mentioned earlier, experience-revealing sentences tend to have a certain linguistic style. 1468 Having converted the problem of experience detection for sentences to a classification task, we focus on the extent to which various linguistic features contribute to the performance of the binary classifier for sentences. We also explain the experimental setting for evaluation, including the classifier and the test corpus. 3.1 Linguistic features In addition to the verb class feature available in the verb lexicon constructed automatically, we used tense, mood, aspect, modality, and experiencer features. Verb class: The feature comes directly from the lexicon since a verb has been classified into a state or activity verb. The predicate part of the sentence to be classified for experience is looked up in the lexicon without sense disambiguation. Tense: The tense of a sentence is important since an experience-revealing sentence tends to use past and present tense. Future tenses are not experiences in most cases. We use POS tagging (Toutanova et al., 2003) for tense determination, but since the Penn tagset provides no future tenses, they are determined by exploiting modal verbs such as “will” and future expressions such “going to”. Mood: It is one of distinctive forms that are used to signal the modal status of a sentence. We consider three mood categories: indicative, imperative and subjunctive. We determine the mood of a sentence by a small set of heuristic rules using the order of POS occurrences and punctuation marks. Aspect: It defines the temporal flow of a verb in the activity or state. Two categories are used: progressive and perfective. This feature is determined by the POS of the predicate in a sentence. Modality: In linguistics, modals are expressions broadly associated with notions of possibility. While modality can be classified at a fine level (e.g., epistemic and deontic), we simply determine whether or not a sentence includes a modal marker that is involved in the main predicate of the sentence. In other words, this binary feature is determined based on the existence of a model verb like “can”, “shall”, “must”, and “may” or a phrase like “have to” or “need to”. The dependency parser is used to ensure a modal marker is indeed associated with the main predicate. Experiencer: A sentence can or cannot be treated as containing an experience depending on the subject or experiencer of the verb (note that this is different from the experiencer role in a case frame). Consider the following sentences: [14] The stranger messed up the entire garden. [15] His presence messed up the whole situation. The first sentence is considered an experience since the subject is a person. However, the second sentence with the same verb is not, because the subject is a non-animate abstract concept. That is, a non-animate noun can hardly constitute an experience. In order to make a distinction, we use the dependency parser and a named-entity recognizer (Finkel et al., 2005) that can recognize person pronouns and person names. 3.2 Classification To train our classifier, we first crawled weblogs from Wordpress3, one of the most popular blog sites in use today. Worpress provides an interface to search blog posts with queries. In selecting experience-containing blog pots, we used location names such as Central Park, SOHO, Seoul and general place names such as airport, subway station, and restaurant because blog posts with some places are expected to describe experiences rather than facts or thoughts. We crawled 6,000 blog posts. After deleting non-English and multi-media blog posts for which we could not obtain any meaningful text data, the number became 5,326. We randomly sampled 1,000 sentences4 and asked three annotators to judge whether or not individual sentences are considered containing an experience based on our definition. For maximum accuracy, we decided to use only those sentences all the three annotators agreed, resulting in a total of 568 sentences. While we tested several classifiers, we chose to use two different classifiers based on SVM and Logistic Regression for the final experimental results because they showed the best performance. 3.3 Results For comparison purposes, we take the method of Kurashima et al. (2005) as our baseline because the method was used in subsequent studies (Kurashima et al., 2006; Kurashima et al., 2009) where experience attributes are extracted. We briefly describe the method and present how we implemented it. The method first extracts all verbs and their dependent phrasal unit from candidate sentences. 3 http://wordpress.com 4 It was due to the limited human resources, but when we increased the number at a later stage, the performance increase was almost negligible. 1469 Feature Logistic Regression SVM Prec. Recall Prec. Recall Baseline 32.0% 55.1% 25.3% 44.4% Lexicon 77.5% 76.0% 77.5% 76.0% Tense 75.1% 75.1% 75.1% 75.1% Mood 75.8% 60.3% 75.8% 60.3% Aspect 26.7% 51.7% 26.7% 51.7% Modality 79.8% 70.5% 79.8% 70.5% Experiencer 54.3% 53.5% 54.3% 53.5% All included 91.9% 91.7% 91.7% 91.4% Table 5. Experience Detection Performance The candidate goes through three filters before it is treated as experience-containing sentence. First, the candidates that do not have an objective case (Fillmore, 1968) are eliminated because their definition of experience as “action + object”. This was done by identifying the objectindicating particle (case marker) in Japanese. Next, the candidates belonging to “become” and “be” statements based on Japanese verb types are filtered out. Finally, the candidate sentences including a verb that indicates a movement are eliminated because the main interest was to identify an activity in a place. Although their definition of experience is somewhat different from ours (i.e., “action + object”), they used the method to generate candidate sentences from which various experience attributes are extracted. From this perspective, the method functioned like our experience detection. Put differently, the definition and the method by which it is determined were much cruder than the one we are using, which seems close to our general understanding.5 The three filtering steps were implemented as follows. We used the dependency parser for extracting objective cases using the direct object relation. The second step, however, could not be applied because there is no grammatical distinction among “do, be, become” statements in English. We had to alter this step by adopting the approach of Inui et al. (2008). The authors propose a lexicon of experience expression by collecting hyponyms from a hierarchically structured dictionary. We collected all hyponyms of words “do” and “act”, from WordNet (Fellbaum, 1998). Lastly, we removed all the verbs that are under the hierarchy of “move” from WordNet. We not only compared our results with the baseline in terms of precision and recall but also 5 This is based on our observation that the three annotators found their task of identifying experience sentences not difficulty, resulting in a high degree of agreements. Feature Logistic Regression SVM Prec. Recall Prec. Recall Baseline 32.0% 55.1% 25.3% 44.4% -Lexicon 84.6% 84.6% 83.1% 81.2% -Tense 87.3% 87.1% 86.8% 86.5% -Mood 89.5% 89.5% 89.3% 89.2% -Aspect 90.8% 90.5% 89.0% 88.6% -Modality 89.5% 89.5% 82.8% 82.8% -Experiencer 91.5% 91.4% 91.1% 90.8% All included 91.9% 91.7% 91.7% 91.4% Table 6. Experience Detection Performance without Individual Features evaluated individual features for their importance in experience detection (classification). The evaluation was conducted with 10-fold cross validation. The results are shown in table 5. The performance, especially precision, of the baseline is much lower than those of the others. The method devised for Japanese doesn’t seem suitable for English. It seems that the linguistic styles shown in experience expressions are different from each other. In addition, the lexicon we constructed for the baseline (i.e., using the WordNet) contains more errors than our activity lexicon for activity verbs. Some hyponyms of an activity verb may not be activity verbs. (e.g., “appear” is a hyponym of “do”). There is almost no difference between the Logistic Regression and SVM classifiers for our methods although SVM was inferior for the baseline. The performance for the best case with all the features included is very promising, closed to 92% precision and recall. Among the features, the lexicon, i.e., verb classes, gave the best result when each is used alone, followed by modality, tense, and mood. Aspect was the worst but close to the baseline. This result is very encouraging for the automatic lexicon construction work because the lexicon plays a pivotal role in the overall performance. In order to see the effect of including individual features in the feature set, precision and recall were measured after eliminating a particular feature from the full set. The results are shown in table 6. Although the absence of the lexicon feature hurt the performance most badly, still the performance was reasonably high (roughly 84 % in precision and recall for the Logistic Regression case). Similar to table 5, the aspect and experience features were the least contributors as the performance drops are almost negligible. 1470 4 Related Work Experience mining in its entirety is a relatively new area where various natural language processing and text mining techniques can play a significant role. While opinion mining or sentiment analysis, which can be considered an important part of experience mining, has been studied quite extensively (see Pang and Lee’s excellent survey (2008)), another sub-area, factuality analysis, begins to gain some popularity (Inui et al., 2008; Saurí, 2008). Very few studies have focused explicitly on extracting various entities that constitute experiences (Kurashima et al., 2009) or detecting experience-containing parts of text although many NLP research areas such as named entity recognition and verb classification are strongly related. The previous work on experience detection relies on a handcrafted lexicon. There have been a number of studies for verb classification (Fillmore, 1968; Vendler, 1967; Somers, 1982; Levin, 1993; Fillmore and Baker, 2001; Kipper et al., 2008) that are essential for construction of an activity verb lexicon, which in turn is important for experience detection. Most similar to our work was done by Siegel and McKeown (2000), who attempted to categorize verbs into state or event classes based on 14 tests similar to those of Vendler’s. They attempted to compute co-occurrence statistics from a corpus. The event class, however, includes activity, accomplishment, and achievement. Similarly, Zacrone and Lenci (2008) attempted to categorize verbs in Italian into the four Vendler classes using the Vendler tests by using a tagged corpus. They focused on existence of arguments such as subject and object that should co-occur with the linguistic features in the tests. The main difference between the previous work and ours lies in the goal and scope of the work. Since our work is specifically geared toward domain-independent experience detection, we attempted to maximize the coverage by using all the verbs in WordNet, as opposed to the verbs appearing in a particular domain-specific corpus (e.g., medicine domain) as done in the previous work. Another difference is that while we are not limited to a particular domain, we did not use extensive human-annotated corpus other than using the 80 seed verbs and existing lexical resources. 5 Conclusion and Future Work We defined experience detection as an essential task for experience mining, which is restated as determining whether individual sentences contain experience or not. Viewing the task as a classification problem, we focused on identification and examination of various linguistic features such as verb class, tense, aspect, mood, modality, and experience, all of which were computed automatically. For verb classes, in particular, we devised a method for classifying all the verbs and verb phrases in WordNet into the activity and state classes. The experimental results show that verb and verb phrase classification method is reasonably accurate with 91% precision and 78% recall with manually constructed gold standard consisting of 80 verbs and 82% accuracy for a random sample of all the WordNet entries. For experience detection, the performance was very promising, closed to 92% in precision and recall when all the features were used. Among the features, the verb classes, or the lexicon we constructed, contributed the most. In order to increase the coverage even further and reduce the errors in lexicon construction, i.e., verb classification, caused by data sparseness, we need to devise a different method, perhaps using domain specific resources. Given that experience mining is a relatively new research area, there are many areas to explore. In addition to refinements of our work, our next step is to develop a method for representing and extracting actual experiences from experience-revealing sentences. Furthermore, considering that only 13% of the blog data we processed contain experiences, an interesting extension is to apply the methodology to extract other types of knowledge such as facts, which are not necessarily experiences. Acknowledgments This research was supported by the IT R&D program of MKE/KEIT under grant KI001877 [Locational/Societal Relation-Aware Social Media Service Technology], and by the MKE (The Ministry of Knowledge Economy), Korea, under the ITRC (Information Technology Research Center) support program supervised by the NIPA (National IT Industry Promotion Agency) [NIPA-2010-C1090-1011-0008]. Reference Eiji Aramaki, Yasuhide Miura, Masatsugu Tonoike, Tomoko Ohkuma, Hiroshi Mashuichi, and Kazuhiko Ohe. 2009. TEXT2TABLE: Medical Text Summarization System based on Named Entity 1471 Recognition and Modality Identification. In Proceedings of the Workshop on BioNLP. Collin F. Baker, Charles J. Fillmore, and Beau Cronin. 2003. The Structure of the Framenet Database. International Journal of Lexicography. Adam L. Berger, Stephen A. Della Pietra, and Vincent J. Della Pietra. 1996. A Mximum Entropy Approach to Natural Language Processing. Computational Linguistics. Chih-Chung Chang and Chih-Jen Lin. 2001. LIBSVM : a Library for Support Vector Machines. http://www.csie.ntu.edu.tw/~cjlin/libsvm. David R. Dowty. 1979. Word meaning and Montague Grammar. Reidel, Dordrecht. Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database. MIT Press. Charles J. Fillmore. 1968. The Case for Case. In Bach and Harms (Ed.): Universals in Linguistic Theory. Charles J. Fillmore and Collin F. Baker. 2001. Frame Semantics for Text Understanding. In Proceedings of WordNet and Other Lexical Resources Workshop, NAACL. Jenny R. Finkel, Trond Grenager, and Christopher D. Manning. 2005. Incorporating Non-local Information into Information Extraction Systems by Gibbs Sampling. In Proceedings of ACL. Chih-Wei Hsu, Chih-Chung Chang, and Chih-Jen Lin. 2009. A Practical Guide to Support Vector Classification. http://www.csie.ntu.edu.tw/~cjlin/libsvm. Kentaro Inui, Shuya Abe, Kazuo Hara, Hiraku Morita, Chitose Sao, Megumi Eguchi, Asuka Sumida, Koji Murakami, and Suguru Matsuyoshi. 2008. Experience Mining: Building a Large-Scale Database of Personal Experiences and Opinions from Web Documents. In Proceedings of the International Conference on Web Intelligence. Valentin Jijkoun, Maarten de Rijke, Wouter Weerkamp, Paul Ackermans and Gijs Geleijnse. 2010. Mining User Experiences from Online Forums: An Exploration. In Proceedings of NAACL HLT Workshop on Computational Linguistics in a World of Social Media. Karin Kipper, Anna Korhonen, Neville Ryant, and Martha Palmer. 2008. A Large-scale Classification of English Verbs. Language Resources and Evaluation Journal. Dan Klein and Christopher D. Manning. 2003. Accurate Unlexicalized Parsing. In Proceedings of ACL. Takeshi Kurashima, Ko Fujimura, and Hidenori Okuda. 2009. Discovering Association Rules on Experiences from Large-Scale Blog Entries. In Proceedings of ECIR. Takeshi Kurashima, Taro Tezuka, and Katsumi Tanaka. 2005. Blog Map of Experiences: Extracting and Geographically Mapping Visitor Experiences from Urban Blogs. In Proceedings of WISE. Takeshi Kurashima, Taro Tezuka, and Katsumi Tanaka. 2006. Mining and Visualizing Local Experiences from Blog Entries. In Proceedings of DEXA. John Lafferty, Andew McCallum, and Fernando Pereira. 2001. Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data. In Proceedings of ICML. Beth Levin. 1993. English verb classes and alternations: A Preliminary investigation. University of Chicago press. Edward Loper, Szu-ting Yi, and Martha Palmer. 2007. Combining Lexical Resources: Mapping Between PropBank and Verbnet. In Proceedings of the International Workshop on Computational Linguistics. Bo Pang and Lillian Lee. 2008. Opinion Mining and Sentiment Analysis, Foundations and Trends in Information Retrieval. Jihee Ryu, Yuchul Jung, Kyung-min Kim and Sung H. Myaeng. 2010. Automatic Extraction of Human Activity Knowledge from Method-Describing Web Articles. In Proceedings of the 1st Workshop on Automated Knowledge Base Construction. Roser Saurí. 2008. A Factuality Profiler for Eventualities in Text. PhD thesis, Brandeis University. Eric V. Siegel and Kathleen R. McKeown. 2000. Learing Methods to Combine Linguistic Indicators: Improving Aspectual Classification and Revealing Linguistic Insights. In Computational Linguistics. Harold L. Somers. 1987. Valency and Case in Computational Linguistics. Edinburgh University Press. Kristina Toutanova, Dan Klein, Christopher D. Manning, and Yoram Singer. 2003. Feature-Rich Partof-Speech Tagging with a Cyclic Dependency Network. In Proceedings of HLT-NAACL. Zeno Vendler. 1967. Linguistics in Philosophy. Cornell University Press. Alessandra Zarcone and Alessandro Lenci. 2008. Computational Models of Event Type Classification in Context. In Proceedings of LREC. 1472
|
2010
|
148
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1473–1481, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Experiments in Graph-based Semi-Supervised Learning Methods for Class-Instance Acquisition Partha Pratim Talukdar∗ Search Labs, Microsoft Research Mountain View, CA 94043 [email protected] Fernando Pereira Google, Inc. Mountain View, CA 94043 [email protected] Abstract Graph-based semi-supervised learning (SSL) algorithms have been successfully used to extract class-instance pairs from large unstructured and structured text collections. However, a careful comparison of different graph-based SSL algorithms on that task has been lacking. We compare three graph-based SSL algorithms for class-instance acquisition on a variety of graphs constructed from different domains. We find that the recently proposed MAD algorithm is the most effective. We also show that class-instance extraction can be significantly improved by adding semantic information in the form of instance-attribute edges derived from an independently developed knowledge base. All of our code and data will be made publicly available to encourage reproducible research in this area. 1 Introduction Traditionally, named-entity recognition (NER) has focused on a small number of broad classes such as person, location, organization. However, those classes are too coarse to support important applications such as sense disambiguation, semantic matching, and textual inference in Web search. For those tasks, we need a much larger inventory of specific classes and accurate classification of terms into those classes. While supervised learning methods perform well for traditional NER, they are impractical for fine-grained classification because sufficient labeled data to train classifiers for all the classes is unavailable and would be very expensive to obtain. ∗Research carried out while at the University of Pennsylvania, Philadelphia, PA, USA. To overcome these difficulties, seed-based information extraction methods have been developed over the years (Hearst, 1992; Riloff and Jones, 1999; Etzioni et al., 2005; Talukdar et al., 2006; Van Durme and Pas¸ca, 2008). Starting with a few seed instances for some classes, these methods, through analysis of unstructured text, extract new instances of the same class. This line of work has evolved to incorporate ideas from graph-based semi-supervised learning in extraction from semi-structured text (Wang and Cohen, 2007), and in combining extractions from free text and from structured sources (Talukdar et al., 2008). The benefits of combining multiple sources have also been demonstrated recently (Pennacchiotti and Pantel, 2009). We make the following contributions: • Even though graph-based SSL algorithms have achieved early success in class-instance acquisition, there is no study comparing different graph-based SSL methods on this task. We address this gap with a series of experiments comparing three graph-based SSL algorithms (Section 2) on graphs constructed from several sources (Metaweb Technologies, 2009; Banko et al., 2007). • We investigate whether semantic information in the form of instance-attribute edges derived from an independent knowledge base (Suchanek et al., 2007) can improve class-instance acquisition. The intuition behind this is that instances that share attributes are more likely to belong to the same class. We demonstrate that instance-attribute edges significantly improve the accuracy of classinstance extraction. In addition, useful classattribute relationships are learned as a byproduct of this process. • In contrast to previous studies involving pro1473 prietary datasets (Van Durme and Pas¸ca, 2008; Talukdar et al., 2008; Pennacchiotti and Pantel, 2009), all of our experiments use publicly available datasets and we plan to release our code1. In Section 2, we review three graph-based SSL algorithms that are compared for the classinstance acquisition task in Section 3. In Section 3.6, we show how additional instance-attribute based semantic constraints can be used to improve class-instance acquisition performance. We summarize the results and outline future work in Section 4. 2 Graph-based SSL We now review the three graph-based SSL algorithms for class inference over graphs that we have evaluated. 2.1 Notation All the algorithms compute a soft assignment of labels to the nodes of a graph G = (V, E, W), where V is the set of nodes with |V | = n, E is the set of edges, and W is an edge weight matrix. Out of the n = nl + nu nodes in G, nl nodes are labeled, while the remaining nu nodes are unlabeled. If edge (u, v) ̸∈E, Wuv = 0. The (unnormalized) Laplacian, L, of G is given by L = D−W, where D is an n×n diagonal degree matrix with Duu = P v Wuv. Let S be an n × n diagonal matrix with Suu = 1 iff node u ∈V is labeled. That is, S identifies the labeled nodes in the graph. C is the set of labels, with |C| = m representing the total number of labels. Y is the n × m matrix storing training label information, if any. ˆY is an n × m matrix of soft label assignments, with ˆYvl representing the score of label l on node v. A graph-based SSL computes ˆY from {G, SY }. 2.2 Label Propagation (LP-ZGL) The label propagation method presented by Zhu et al. (2003), which we shall refer to as LP-ZGL in this paper, is one of the first graph-based SSL methods. The objective minimized by LP-ZGL is: min ˆY X l∈C ˆY ⊤ l L ˆYl, s.t. SYl = S ˆYl (1) 1http://www.talukdar.net/datasets/class inst/ where ˆYl of size n × 1 is the lth column of ˆY . The constraint SY = S ˆY makes sure that the supervised labels are not changed during inference. The above objective can be rewritten as: X l∈C ˆY ⊤ l L ˆYl = X u,v∈V,l∈C Wuv( ˆYul −ˆYvl)2 From this, we observe that LP-ZGL penalizes any label assignment where two nodes connected by a highly weighted edge are assigned different labels. In other words, LP-ZGL prefers smooth labelings over the graph. This property is also shared by the two algorithms we shall review next. LP-ZGL has been the basis for much subsequent work in the graph-based SSL area, and is still one of the most effective graph-based SSL algorithms. 2.3 Adsorption Adsorption (Baluja et al., 2008) is a graph-based SSL algorithm which has been used for opendomain class-instance acquisition (Talukdar et al., 2008). Adsorption is an iterative algorithm, where label estimates on node v in the (t + 1)th iteration are updated using estimates from the tth iteration: ˆY (t+1) v ←pinj v ×Yv+pcont v ×B(t) v +pabnd v ×r (2) where, B(t) v = X u Wuv P u′ Wu′v ˆY (t) u In (2), pinj v , pcont v , and pabnd v are three probabilities defined on each node v ∈V by Adsorption; and r is a vector used by Adsorption to express label uncertainty at a node. On each node v, the three probabilities sum to one, i.e., pinj v + pcont v + pabnd v = 1, and they are based on the random-walk interpretation of the Adsorption algorithm (Talukdar et al., 2008). The main idea of Adsorption is to control label propagation more tightly by limiting the amount of information that passes through a node. For instance, Adsorption can reduce the importance of a high-degree node v during the label inference process by increasing pabnd v on that node. For more details on these, please refer to Section 2 of (Talukdar and Crammer, 2009). In contrast to LP-ZGL, Adsorption allows labels on labeled (seed) nodes to change, which is desirable in case of noisy input labels. 1474 2.4 Modified Adsorption (MAD) Talukdar and Crammer (2009) introduced a modification of Adsorption called MAD, which shares Adsorption’s desirable properties but can be expressed as an unconstrained optimization problem: min ˆY X l∈C µ1 Yl −ˆYl ⊤S Yl −ˆYl + µ2 ˆY ⊤ l L ′ ˆYl + µ3 ˆYl −Rl 2 (3) where µ1, µ2, and µ3 are hyperparameters; L ′ is the Laplacian of an undirected graph derived from G, but with revised edge weights; and R is an n × m matrix of per-node label prior, if any, with Rl representing the lth column of R. As in Adsorption, MAD allows labels on seed nodes to change. In case of MAD, the three random-walk probabilities, pinj v , pcont v , and pabnd v , defined by Adsorption on each node are folded inside the matrices S, L ′, and R, respectively. The optimization problem in (3) can be solved with an efficient iterative algorithm described in detail by Talukdar and Crammer (2009). These three algorithms are all easily parallelizable in a MapReduce framework (Talukdar et al., 2008; Rao and Yarowsky, 2009), which makes them suitable for SSL on large datasets. Additionally, all three algorithms have similar space and time complexity. 3 Experiments We now compare the experimental performance of the three graph-based SSL algorithms reviewed in the previous section, using graphs constructed from a variety of sources described below. Following previous work (Talukdar et al., 2008), we use Mean Reciprocal Rank (MRR) as the evaluation metric in all experiments: MRR = 1 |Q| X v∈Q 1 rv (4) where Q ⊆V is the set of test nodes, and rv is the rank of the gold label among the labels assigned to node v. Higher MRR reflects better performance. We used iterative implementations of the graphbased SSL algorithms, and the number of iterations was treated as a hyperparameter which was tuned, along with other hyperparameters, on separate held-out sets, as detailed in a longer version of this paper. Statistics of the graphs used during experiments in this section are presented in Table 1. 3.1 Freebase-1 Graph with Pantel Classes Table ID: people-person Name Place of Birth Gender · · · · · · · · · Isaac Newton Lincolnshire Male Bob Dylan Duluth Male Johnny Cash Kingsland Male · · · · · · · · · Table ID: film-music contributor Name Film Music Credits · · · · · · Bob Dylan No Direction Home · · · · · · Figure 1: Examples of two tables from Freebase, one table is from the people domain while the other is from the film domain. 0.5 0.575 0.65 0.725 0.8 23 x 2 23 x 10 Freebase-1 Graph, 23 Pantel Classes Mean Reciprocal Rank (MRR) Amount of Supervision (# classes x seeds per class) LP-ZGL Adsorption MAD Figure 3: Comparison of three graph transduction methods on a graph constructed from the Freebase dataset (see Section 3.1), with 23 classes. All results are averaged over 4 random trials. In each group, MAD is the rightmost bar. Freebase (Metaweb Technologies, 2009)2 is a large collaborative knowledge base. The knowledge base harvests information from many open data sets (for instance Wikipedia and MusicBrainz), as well as from user contributions. For our current purposes, we can think of the Freebase 2http://www.freebase.com/ 1475 Graph Vertices Edges Avg. Min. Max. Deg. Deg. Deg. Freebase-1 (Section 3.1) 32970 957076 29.03 1 13222 Freebase-2 (Section 3.2) 301638 2310002 7.66 1 137553 TextRunner (Section 3.3) 175818 529557 3.01 1 2738 YAGO (Section 3.6) 142704 777906 5.45 0 74389 TextRunner + YAGO (Section 3.6) 237967 1307463 5.49 1 74389 Table 1: Statistics of various graphs used in experiments in Section 3. Some of the test instances in the YAGO graph, added for fair comparison with the TextRunner graph in Section 3.6, had no attributes in YAGO KB, and hence these instance nodes had degree 0 in the YAGO graph. Bob Dylan film-music_contributor-name Johnny Cash people-person-name Isaac Newton Bob Dylan film-music_contributor-name Johnny Cash people-person-name Isaac Newton has_attribute:albums (a) (b) Figure 2: (a) Example of a section of the graph constructed from the two tables in Figure 1. Rectangular nodes are properties, oval nodes are entities or cell values. (b) The graph in part (a) augmented with an attribute node, has attribue:albums, along with the edges incident on it. This results is additional constraints for the nodes Johnny Cash and Bob Dylan to have similar labels (see Section 3.6). dataset as a collection of relational tables, where each table is assigned a unique ID. A table consists of one or more properties (column names) and their corresponding cell values (column entries). Examples of two Freebase tables are shown in Figure 1. In this figure, Gender is a property in the table people-person, and Male is a corresponding cell value. We use the following process to convert the Freebase data tables into a single graph: • Create a node for each unique cell value • Create a node for each unique property name, where unique property name is obtained by prefixing the unique table ID to the property name. For example, in Figure 1, peopleperson-gender is a unique property name. • Add an edge of weight 1.0 from cell-value node v to unique property node p, iff value v is present in the column corresponding to property p. Similarly, add an edge in the reverse direction. By applying this graph construction process on the first column of the two tables in Figure 1, we end up with the graph shown in Figure 2 (a). We note that even though the resulting graph consists of edges connecting nodes of different types: cell value nodes to property nodes; the graph-based SSL methods (Section 2) can still be applied on such graphs as a cell value node and a property node connected by an edge should be assigned same or similar class labels. In other words, the label smoothness assumption (see Section 2.2) holds on such graphs. We applied the same graph construction process on a subset of the Freebase dataset consisting of topics from 18 randomly selected domains: astronomy, automotive, biology, book, business, 1476 chemistry, comic books, computer, film, food, geography, location, people, religion, spaceflight, tennis, travel, and wine. The topics in this subset were further filtered so that only cell-value nodes with frequency 10 or more were retained. We call the resulting graph Freebase-1 (see Table 1). Pantel et al. (2009) have made available a set of gold class-instance pairs derived from Wikipedia, which is downloadable from http://ow.ly/13B57. From this set, we selected all classes which had more than 10 instances overlapping with the Freebase graph constructed above. This resulted in 23 classes, which along with their overlapping instances were used as the gold standard set for the experiments in this section. Experimental results with 2 and 10 seeds (labeled nodes) per class are shown in Figure 3. From the figure, we see that that LP-ZGL and Adsorption performed comparably on this dataset, with MAD significantly outperforming both methods. 3.2 Freebase-2 Graph with WordNet Classes 0.25 0.285 0.32 0.355 0.39 192 x 2 192 x 10 Freebase-2 Graph, 192 WordNet Classes Mean Reciprocal Rank (MRR) Amount of Supervision (# classes x seeds per class) LP-ZGL Adsorption MAD Figure 4: Comparison of graph transduction methods on a graph constructed from the Freebase dataset (see Section 3.2). All results are averaged over 10 random trials. In each group, MAD is the rightmost bar. To evaluate how the algorithms scale up, we construct a larger graph from the same 18 domains as in Section 3.1, and using the same graph construction process. We shall call the resulting graph Freebase-2 (see Table 1). In order to scale up the number of classes, we selected all Wordnet (WN) classes, available in the YAGO KB (Suchanek et al., 2007), that had more than 100 instances overlapping with the larger Freebase graph constructed above. This resulted in 192 WN classes which we use for the experiments in this section. The reason behind imposing such frequency constraints during class selection is to make sure that each class is left with a sufficient number of instances during testing. Experimental results comparing LP-ZGL, Adsorption, and MAD with 2 and 10 seeds per class are shown in Figure 4. A total of 292k test nodes were used for testing in the 10 seeds per class condition, showing that these methods can be applied to large datasets. Once again, we observe MAD outperforming both LP-ZGL and Adsorption. It is interesting to note that MAD with 2 seeds per class outperforms LP-ZGL and adsorption even with 10 seeds per class. 3.3 TextRunner Graph with WordNet Classes 0.15 0.2 0.25 0.3 0.35 170 x 2 170 x 10 TextRunner Graph, 170 WordNet Classes Mean Reciprocal Rank (MRR) Amount of Supervision (# classes x seeds per class) LP-ZGL Adsorption MAD Figure 5: Comparison of graph transduction methods on a graph constructed from the hypernym tuples extracted by the TextRunner system (Banko et al., 2007) (see Section 3.3). All results are averaged over 10 random trials. In each group, MAD is the rightmost bar. In contrast to graph construction from structured tables as in Sections 3.1, 3.2, in this section we use hypernym tuples extracted by TextRunner (Banko et al., 2007), an open domain IE system, to construct the graph. Example of a hypernym tuple extracted by TextRunner is (http, protocol, 0.92), where 0.92 is the extraction confidence. To convert such a tuple into a graph, we create a node for the instance (http) and a node for the class (protocol), and then connect the nodes with two 1477 directed edges in both directions, with the extraction confidence (0.92) as edge weights. The graph created with this process from TextRunner output is called the TextRunner Graph (see Table 1). As in Section 3.2, we use WordNet class-instance pairs as the gold set. In this case, we considered all WordNet classes, once again from YAGO KB (Suchanek et al., 2007), which had more than 50 instances overlapping with the constructed graph. This resulted in 170 WordNet classes being used for the experiments in this section. Experimental results with 2 and 10 seeds per class are shown in Figure 5. The three methods are comparable in this setting, with MAD achieving the highest overall MRR. 3.4 Discussion If we correlate the graph statistics in Table 1 with the results of sections 3.1, 3.2, and 3.3, we see that MAD is most effective for graphs with high average degree, that is, graphs where nodes tend to connect to many other nodes. For instance, the Freebase-1 graph has a high average degree of 29.03, with a corresponding large advantage for MAD over the other methods. Even though this might seem mysterious at first, it becomes clearer if we look at the objectives minimized by different algorithms. We find that the objective minimized by LP-ZGL (Equation 1) is underregularized, i.e., its model parameters ( ˆY ) are not constrained enough, compared to MAD (Equation 3, specifically the third term), resulting in overfitting in case of highly connected graphs. In contrast, MAD is able to avoid such overfitting because of its minimization of a well regularized objective (Equation 3). Based on this, we suggest that average degree, an easily computable structural property of the graph, may be a useful indicator in choosing which graph-based SSL algorithm should be applied on a given graph. Unlike MAD, Adsorption does not optimize any well defined objective (Talukdar and Crammer, 2009), and hence any analysis along the lines described above is not possible. The heuristic choices made in Adsorption may have lead to its sub-optimal performance compared to MAD; we leave it as a topic for future investigation. 3.5 Effect of Per-Node Class Sparsity For all the experiments in Sections 3.1, 3.2, and 3.6, each node was allowed to have a maximum of 15 classes during inference. After each update 0.3 0.33 0.36 0.39 0.42 5 15 25 35 45 Effect of Per-node Sparsity Constraint Mean Reciprocal Rank (MRR) Maximum Allowed Classes per Node Figure 6: Effect of per node class sparsity (maximum number of classes allowed per node) during MAD inference in the experimental setting of Figure 4 (one random split). on a node, all classes except for the top scoring 15 classes were discarded. Without such sparsity constraints, a node in a connected graph will end up acquiring all the labels injected into the graph. This is undesirable for two reasons: (1) for experiments involving a large numbers of classes (as in the previous section and in the general case of open domain IE), this increases the space requirement and also slows down inference; (2) a particular node is unlikely to belong to a large number of classes. In order to estimate the effect of such sparsity constraints, we varied the number of classes allowed per node from 5 to 45 on the graph and experimental setup of Figure 4, with 10 seeds per class. The results for MAD inference over the development split are shown in Figure 6. We observe that performance can vary significantly as the maximum number of classes allowed per node is changed, with the performance peaking at 25. This suggests that sparsity constraints during graph based SSL may have a crucial role to play, a question that needs further investigation. 3.6 TextRunner Graph with additional Semantic Constraints from YAGO Recently, the problem of instance-attribute extraction has started to receive attention (Probst et al., 2007; Bellare et al., 2007; Pasca and Durme, 2007). An example of an instance-attribute pair is (Bob Dylan, albums). Given a set of seed instance-attribute pairs, these methods attempt to extract more instance-attribute pairs automatically 1478 0.18 0.23 0.28 0.33 0.38 LP-ZGL Adsorption MAD 170 WordNet Classes, 2 Seeds per Class Mean Reciprocal Rank (MRR) Algorithms TextRunner Graph YAGO Graph TextRunner + YAGO Graph 0.3 0.338 0.375 0.413 0.45 LP-ZGL Adsorption MAD 170 WordNet Classes, 10 Seeds per Class Mean Reciprocal Rank (MRR) Algorithms TextRunner Graph YAGO Graph TextRunner + YAGO Graph Figure 7: Comparison of class-instance acquisition performance on the three different graphs described in Section 3.6. All results are averaged over 10 random trials. Addition of YAGO attributes to the TextRunner graph significantly improves performance. YAGO Top-2 WordNet Classes Assigned by MAD Attribute (example instances for each class are shown in brackets) has currency wordnet country 108544813 (Burma, Afghanistan) wordnet region 108630039 (Aosta Valley, Southern Flinders Ranges) works at wordnet scientist 110560637 (Aage Niels Bohr, Adi Shamir) wordnet person 100007846 (Catherine Cornelius, Jamie White) has capital wordnet state 108654360 (Agusan del Norte, Bali) wordnet region 108630039 (Aosta Valley, Southern Flinders Ranges) born in wordnet boxer 109870208 (George Chuvalo, Fernando Montiel) wordnet chancellor 109906986 (Godon Brown, Bill Bryson) has isbn wordnet book 106410904 (Past Imperfect, Berlin Diary) wordnet magazine 106595351 (Railway Age, Investors Chronicle) Table 2: Top 2 (out of 170) WordNet classes assigned by MAD on 5 randomly chosen YAGO attribute nodes (out of 80) in the TextRunner + YAGO graph used in Figure 7 (see Section 3.6), with 10 seeds per class used. A few example instances of each WordNet class is shown within brackets. Top ranked class for each attribute is shown in bold. from various sources. In this section, we explore whether class-instance assignment can be improved by incorporating new semantic constraints derived from (instance, attribute) pairs. In particular, we experiment with the following type of constraint: two instances with a common attribute are likely to belong to the same class. For example, in Figure 2 (b), instances Johnny Cash and Bob Dylan are more likely to belong to the same class as they have a common attribute, albums. Because of the smooth labeling bias of graph-based SSL methods (see Section 2.2), such constraints are naturally captured by the methods reviewed in Section 2. All that is necessary is the introduction of bidirectional (instance, attribute) edges to the graph, as shown in Figure 2 (b). In Figure 7, we compare class-instance acquisition performance of the three graph-based SSL methods (Section 2) on the following three graphs (also see Table 1): TextRunner Graph: Graph constructed from the hypernym tuples extracted by TextRunner, as in Figure 5 (Section 3.3), with 175k vertices and 529k edges. YAGO Graph: Graph constructed from the (instance, attribute) pairs obtained from the YAGO KB (Suchanek et al., 2007), with 142k nodes and 777k edges. TextRunner + YAGO Graph: Union of the 1479 two graphs above, with 237k nodes and 1.3m edges. In all experimental conditions with 2 and 10 seeds per class in Figure 7, we observe that the three methods consistently achieved the best performance on the TextRunner + YAGO graph. This suggests that addition of attribute based semantic constraints from YAGO to the TextRunner graph results in a better connected graph which in turn results in better inference by the graphbased SSL algorithms, compared to using either of the sources, i.e., TextRunner output or YAGO attributes, in isolation. This further illustrates the advantage of aggregating information across sources (Talukdar et al., 2008; Pennacchiotti and Pantel, 2009). However, we are the first, to the best of our knowledge, to demonstrate the effectiveness of attributes in class-instance acquisition. We note that this work is similar in spirit to the recent work by Carlson et al. (2010) which also demonstrates the benefits of additional constraints in SSL. Because of the label propagation behavior, graph-based SSL algorithms assign classes to all nodes reachable in the graph from at least one of the labeled instance nodes. This allows us to check the classes assigned to nodes corresponding to YAGO attributes in the TextRunner + YAGO graph, as shown in Table 2. Even though the experiments were designed for classinstance acquisition, it is encouraging to see that the graph-based SSL algorithm (MAD in Table 2) is able to learn class-attribute relationships, an important by-product that has been the focus of recent studies (Reisinger and Pasca, 2009). For example, the algorithm is able to learn that works at is an attribute of the WordNet class wordnet scientist 110560637, and thereby its instances (e.g. Aage Niels Bohr, Adi Shamir). 4 Conclusion We have started a systematic experimental comparison of graph-based SSL algorithms for classinstance acquisition on a variety of graphs constructed from different domains. We found that MAD, a recently proposed graph-based SSL algorithm, is consistently the most effective across the various experimental conditions. We also showed that class-instance acquisition performance can be significantly improved by incorporating additional semantic constraints in the class-instance acquisition process, which for the experiments in this paper were derived from instance-attribute pairs available in an independently developed knowledge base. All the data used in these experiments was drawn from publicly available datasets and we plan to release our code3 to foster reproducible research in this area. Topics for future work include the incorporation of other kinds of semantic constraint for improved class-instance acquisition, further investigation into per-node sparsity constraints in graph-based SSL, and moving beyond bipartite graph constructions. Acknowledgments We thank William Cohen for valuable discussions, and Jennifer Gillenwater, Alex Kulesza, and Gregory Malecha for detailed comments on a draft of this paper. We are also very grateful to the authors of (Banko et al., 2007), Oren Etzioni and Stephen Soderland in particular, for providing TextRunner output. This work was supported in part by NSF IIS-0447972 and DARPA HRO1107-1-0029. References S. Baluja, R. Seth, D. Sivakumar, Y. Jing, J. Yagnik, S. Kumar, D. Ravichandran, and M. Aly. 2008. Video suggestion and discovery for youtube: taking random walks through the view graph. Proceedings of WWW-2008. M. Banko, M.J. Cafarella, S. Soderland, M. Broadhead, and O. Etzioni. 2007. Open information extraction from the web. Procs. of IJCAI. K. Bellare, P. Talukdar, G. Kumaran, F. Pereira, M. Liberman, A. McCallum, and M. Dredze. 2007. Lightly-Supervised Attribute Extraction. NIPS 2007 Workshop on Machine Learning for Web Search. A. Carlson, J. Betteridge, R.C. Wang, E.R. Hruschka Jr, and T.M. Mitchell. 2010. Coupled Semi-Supervised Learning for Information Extraction. In Proceedings of the Third ACM International Conference on Web Search and Data Mining (WSDM), volume 2, page 110. O. Etzioni, Michael Cafarella, Doug Downey, AnaMaria Popescu, Tal Shaked, Stephen Soderland, Daniel S. Weld, and Alexander Yates. 2005. Unsupervised named-entity extraction from the web - an experimental study. Artificial Intelligence Journal. M. Hearst. 1992. Automatic acquisition of hyponyms from large text corpora. In Fourteenth International 3http://www.talukdar.net/datasets/class inst/ 1480 Conference on Computational Linguistics, Nantes, France. Metaweb Technologies. 2009. Freebase data dumps. http://download.freebase.com/datadumps/. P. Pantel, E. Crestan, A. Borkovsky, A.M. Popescu, and V. Vyas. 2009. Web-scale distributional similarity and entity set expansion. Proceedings of EMNLP09, Singapore. M. Pasca and Benjamin Van Durme. 2007. What you seek is what you get: Extraction of class attributes from query logs. In IJCAI-07. Ferbruary, 2007. M. Pennacchiotti and P. Pantel. 2009. Entity Extraction via Ensemble Semantics. Proceedings of EMNLP-09, Singapore. K. Probst, R. Ghani, M. Krema, A. Fano, and Y. Liu. 2007. Semi-supervised learning of attribute-value pairs from product descriptions. In IJCAI-07, Ferbruary, 2007. D. Rao and D. Yarowsky. 2009. Ranking and Semisupervised Classification on Large Scale Graphs Using Map-Reduce. TextGraphs. J. Reisinger and M. Pasca. 2009. Bootstrapped extraction of class attributes. In Proceedings of the 18th international conference on World wide web, pages 1235–1236. ACM. E. Riloff and R. Jones. 1999. Learning dictionaries for information extraction by multi-level bootstrapping. In Proceedings of the 16th National Conference on Artificial Intelligence (AAAI-99), pages 474–479, Orlando, Florida. F.M. Suchanek, G. Kasneci, and G. Weikum. 2007. Yago: a core of semantic knowledge. In Proceedings of the 16th international conference on World Wide Web, page 706. ACM. P. P. Talukdar and Koby Crammer. 2009. New regularized algorithms for transductive learning. In ECMLPKDD. P. P. Talukdar, T. Brants, F. Pereira, and M. Liberman. 2006. A context pattern induction method for named entity extraction. In Tenth Conference on Computational Natural Language Learning, page 141. P. P. Talukdar, J. Reisinger, M. Pasca, D. Ravichandran, R. Bhagat, and F. Pereira. 2008. WeaklySupervised Acquisition of Labeled Class Instances using Graph Random Walks. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 581–589. B. Van Durme and M. Pas¸ca. 2008. Finding cars, goddesses and enzymes: Parametrizable acquisition of labeled instances for open-domain information extraction. Twenty-Third AAAI Conference on Artificial Intelligence. R. Wang and W. Cohen. 2007. Language-Independent Set Expansion of Named Entities Using the Web. Data Mining, 2007. ICDM 2007. Seventh IEEE International Conference on, pages 342–350. X. Zhu, Z. Ghahramani, and J. Lafferty. 2003. Semisupervised learning using gaussian fields and harmonic functions. ICML-03, 20th International Conference on Machine Learning. 1481
|
2010
|
149
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 138–147, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Extracting Social Networks from Literary Fiction David K. Elson Dept. of Computer Science Columbia University [email protected] Nicholas Dames English Department Columbia University [email protected] Kathleen R. McKeown Dept. of Computer Science Columbia University [email protected] Abstract We present a method for extracting social networks from literature, namely, nineteenth-century British novels and serials. We derive the networks from dialogue interactions, and thus our method depends on the ability to determine when two characters are in conversation. Our approach involves character name chunking, quoted speech attribution and conversation detection given the set of quotes. We extract features from the social networks and examine their correlation with one another, as well as with metadata such as the novel’s setting. Our results provide evidence that the majority of novels in this time period do not fit two characterizations provided by literacy scholars. Instead, our results suggest an alternative explanation for differences in social networks. 1 Introduction Literary studies about the nineteenth-century British novel are often concerned with the nature of the community that surrounds the protagonist. Some theorists have suggested a relationship between the size of a community and the amount of dialogue that occurs, positing that “face to face time” diminishes as the number of characters in the novel grows. Others suggest that as the social setting becomes more urbanized, the quality of dialogue also changes, with more interactions occurring in rural communities than urban communities. Such claims have typically been made, however, on the basis of a few novels that are studied in depth. In this paper, we aim to determine whether an automated study of a much larger sample of nineteenth century novels supports these claims. The research presented here is concerned with the extraction of social networks from literature. We present a method to automatically construct a network based on dialogue interactions between characters in a novel. Our approach includes components for finding instances of quoted speech, attributing each quote to a character, and identifying when certain characters are in conversation. We then construct a network where characters are vertices and edges signify an amount of bilateral conversation between those characters, with edge weights corresponding to the frequency and length of their exchanges. In contrast to previous approaches to social network construction, ours relies on a novel combination of patternbased detection, statistical methods, and adaptation of standard natural language tools for the literary genre. We carried out this work on a corpus of 60 nineteenth-century novels and serials, including 31 authors such as Dickens, Austen and Conan Doyle. In order to evaluate the literary claims in question, we compute various characteristics of the dialogue-based social network and stratify these results by categories such as the novel’s setting. For example, the density of the network provides evidence about the cohesion of a large or small community, and cliques may indicate a social fragmentation. Our results surprisingly provide evidence that the majority of novels in this time period do not fit the suggestions provided by literary scholars, and we suggest an alternative explanation for our observations of differences across novels. In the following sections, we survey related work on social networks as well as computational studies of literature. We then present the literary hypotheses in more detail. We describe the methods we use to extract dialogue and construct conversational networks, along with our approach to analyzing their characteristics. After we present the statistical results, we analyze their significance from a literary perspective. 138 2 Related Work Computer-assisted literary analysis has typically occurred at the word level. This level of granularity lends itself to studies of authorial style based on patterns of word use (Burrows, 2004), and researchers have successfully “outed” the writers of anonymous texts by comparing their style to that of a corpus of known authors (Mostellar and Wallace, 1984). Determining instances of “text reuse,” a type of paraphrasing, is also a form of analysis at the lexical level, and it has recently been used to validate theories about the lineage of ancient texts (Lee, 2007). Analysis of literature using more semanticallyoriented techniques has been rare, most likely because of the difficulty in automatically determining meaningful interpretations. Some exceptions include recent work on learning common event sequences in news stories (Chambers and Jurafsky, 2008), an approach based on statistical methods, and the development of an event calculus for characterizing stories written by children (Halpin et al., 2004), a knowledge-based strategy. On the other hand, literary theorists, linguists and others have long developed symbolic but non-computational models for novels. For example, Moretti (2005) has graphically mapped out texts according to geography, social connections and other variables. While researchers have not attempted the automatic construction of social networks representing connections between characters in a corpus of novels, the ACE program has involved entity and relation extraction in unstructured text (Doddington et al., 2004). Other recent work in social network construction has explored the use of structured data such as email headers (McCallum et al., 2007) and U.S. Senate bill cosponsorship (Cho and Fowler, 2010). In an analysis of discussion forums, Gruzd and Haythornthwaite (2008) explored the use of message text as well as posting data to infer who is talking to whom. In this paper, we also explore how to build a network based on conversational interaction, but we analyze the reported dialogue found in novels to determine the links. The kinds of language that is used to signal such information is quite different in the two media. In discussion forums, people tend to use addresses such as “Hi Tom,” while in novels, a system must determine both the speaker of a quotation and then the intended recipient of the dialogue act. This is a significantly different problem. 3 Hypotheses It is commonly held that the novel is a literary form which tries to produce an accurate representation of the social world. Within literary studies, the recurring problem is how that representation is achieved. Theories about the relation between novelistic form (the workings of plot, characters, and dialogue, to take the most basic categories) and changes to real-world social milieux abound. Many of these theories center on nineteenth-century European fiction; innovations in novelistic form during this period, as well as the rapid social changes brought about by revolution, industrialization, and transport development, have traditionally been linked. These theories, however, have used only a select few representative novels as proof. By using statistical methods of analysis, it is possible to move beyond this small corpus of proof texts. We believe these methods are essential to testing the validity of some core theories about social interaction and their representation in literary genres like the novel. Major versions of the theories about the social worlds of nineteenth-century fiction tend to center on characters, in two specific ways: how many characters novels tend to have, and how those characters interact with one another. These two “formal” facts about novels are usually explained with reference to a novel’s setting. From the influential work of the Russian critic Mikhail Bakhtin to the present, a consensus emerged that as novels are increasingly set in urban areas, the number of characters and the quality of their interaction change to suit the setting. Bakhtin’s term for this causal relationship was chronotope: the “intrinsic interconnectedness of temporal and spatial relationships that are artistically expressed in literature,” in which “space becomes charged and responsive to movements of time, plot, and history” (Bakhtin, 1981, 84). In Bakhtin’s analysis, different spaces have different social and emotional potentialities, which in turn affect the most basic aspects of a novel’s aesthetic technique. After Bakhtin’s invention of the chronotope, much literary criticism and theory devoted itself to filling in, or describing, the qualities of specific chronotopes, particularly those of the village or rural environment and the city or urban environment. Following a suggestion of Bakhtin’s that the population of village or rural fictions is modeled on the world of the family, made up of 139 Author/Title/Year Persp. Setting Author/Title/Year Persp. Setting Ainsworth, Jack Sheppard (1839) 3rd urban Gaskell, North and South (1854) 3rd urban Austen, Emma (1815) 3rd rural Gissing, In the Year of Jubilee (1894) 3rd urban Austen, Mansfield Park (1814) 3rd rural Gissing, New Grub Street (1891) 3rd urban Austen, Persuasion (1817) 3rd rural Hardy, Jude the Obscure (1894) 3rd mixed Austen, Pride and Prejudice (1813) 3rd rural Hardy, The Return of the Native (1878) 3rd rural Braddon, Lady Audley’s Secret (1862) 3rd mixed Hardy, Tess of the d’Ubervilles (1891) 3rd rural Braddon, Aurora Floyd (1863) 3rd rural Hughes, Tom Brown’s School Days (1857) 3rd rural Bront¨e, Anne, The Tenant of Wildfell Hall (1848) 1st rural James, The Portrait of a Lady (1881) 3rd urban Bront¨e, Charlotte, Jane Eyre (1847) 1st rural James, The Ambassadors (1903) 3rd urban Bront¨e, Charlotte, Villette (1853) 1st mixed James, The Wings of the Dove (1902) 3rd urban Bront¨e, Emily, Wuthering Heights (1847) 1st rural Kingsley, Alton Locke (1860) 1st mixed Bulwer-Lytton, Paul Clifford (1830) 3rd urban Martineau, Deerbrook (1839) 3rd rural Collins, The Moonstone (1868) 1st urban Meredith, The Egoist (1879) 3rd rural Collins, The Woman in White (1859) 1st urban Meredith, The Ordeal of Richard Feverel (1859) 3rd rural Conan Doyle, The Sign of the Four (1890) 1st urban Mitford, Our Village (1824) 1st rural Conan Doyle, A Study in Scarlet (1887) 1st urban Reade, Hard Cash (1863) 3rd urban Dickens, Bleak House (1852) mixed urban Scott, The Bride of Lammermoor (1819) 3rd rural Dickens, David Copperfield (1849) 1st mixed Scott, The Heart of Mid-Lothian (1818) 3rd rural Dickens, Little Dorrit (1855) 3rd urban Scott, Waverley (1814) 3rd rural Dickens, Oliver Twist (1837) 3rd urban Stevenson, The Strange Case of Dr. Jekyll and Mr. Hyde (1886) 1st urban Dickens, The Pickwick Papers (1836) 3rd mixed Stoker, Dracula (1897) 1st urban Disraeli, Sybil, or the Two Nations (1845) 3rd mixed Thackeray, History of Henry Esmond (1852) 1st urban Edgeworth, Belinda (1801) 3rd rural Thackeray, History of Pendennis (1848) 1st urban Edgeworth, Castle Rackrent (1800) 3rd rural Thackeray, Vanity Fair (1847) 3rd urban Eliot, Adam Bede (1859) 3rd rural Trollope, Barchester Towers (1857) 3rd rural Eliot, Daniel Deronda (1876) 3rd urban Trollope, Doctor Thorne (1858) 3rd rural Eliot, Middlemarch (1871) 3rd rural Trollope, Phineas Finn (1867) 3rd urban Eliot, The Mill on the Floss (1860) 3rd rural Trollope, The Way We Live Now (1874) 3rd urban Galt, Annals of the Parish (1821) 1st rural Wilde, The Picture of Dorian Gray (1890) 3rd urban Gaskell, Mary Barton (1848) 3rd urban Wood, East Lynne (1860) 3rd mixed Table 1: Properties of the nineteenth-century British novels and serials included in our study. an intimately related set of characters, many critics analyzed the formal expression of this world as constituted by a small set of characters who express themselves conversationally. Raymond Williams used the term “knowable communities” to describe this world, in which face-to-face relations of a restricted set of characters are the primary mode of social interaction (Williams, 1975, 166). By contrast, the urban world, in this traditional account, is both larger and more complex. To describe the social-psychological impact of the city, Franco Moretti argues, protagonists of urban novels “change overnight from ‘sons’ into ‘young men’: their affective ties are no longer vertical ones (between successive generations), but horizontal, within the same generation. They are drawn towards those unknown yet congenial faces seen in gardens, or at the theater; future friends, or rivals, or both” (Moretti, 1999, 65). The result is two-fold: more characters, indeed a mass of characters, and more interactions, although less actual conversation; as literary critic Terry Eagleton argues, the city is where “most of our encounters consist of seeing rather than speaking, glimpsing each other as objects rather than conversing as fellow subjects” (Eagleton, 2005, 145). Moretti argues in similar terms. For him, the difference in number of characters is “not just a matter of quantity... it’s a qualitative, morphological one” (Moretti, 1999, 68). As the number of characters increases, Moretti argues (following Bakhtin in his logic), social interactions of different kinds and durations multiply, displacing the family-centered and conversational logic of village or rural fictions. “The narrative system becomes complicated, unstable: the city turns into a gigantic roulette table, where helpers and antagonists mix in unpredictable combinations” (Moretti, 1999, 68). This argument about how novelistic setting produces different forms of social interaction is precisely what our method seeks to evaluate. Our corpus of 60 novels was selected for its representativeness, particularly in the following categories: authorial (novels from the major canoni140 cal authors of the period), historical (novels from each decade), generic (from the major sub-genres of nineteenth-century fiction), sociological (set in rural, urban, and mixed locales), and technical (narrated in first-person and third-person form). The novels, as well as important metadata we assigned to them (the perspective and setting), are shown in Table 1. We define urban to mean set in a metropolitan zone, characterized by multiple forms of labor (not just agricultural). Here, social relations are largely financial or commercial in character. We conversely define rural to describe texts that are set in a country or village zone, where agriculture is the primary activity, and where land-owning, non-productive, rentcollecting gentry are socially predominant. Social relations here are still modeled on feudalism (relations of peasant-lord loyalty and family tie) rather than the commercial cash nexus. We also explored other properties of the texts, such as literary genre, but focus on the results found with setting and perspective. We obtained electronic encodings of the texts from Project Gutenberg. All told, these texts total more than 10 million words. We assembled this representative corpus in order to test two hypotheses, which are derived from the aforementioned theories: 1. That there is an inverse correlation between the amount of dialogue in a novel and the number of characters in that novel. One basic, shared assumption of these theorists is that as the network of characters expands– as, in Moretti’s words, a quantitative change becomes qualitative– the importance, and in fact amount, of dialogue decreases. With a method for extracting conversation from a large corpus of texts, it is possible to test this hypothesis against a wide range of data. 2. That a significant difference in the nineteenth-century novel’s representation of social interaction is geographical: novels set in urban environments depict a complex but loose social network, in which numerous characters share little conversational interaction, while novels set in rural environments inhabit more tightly bound social networks, with fewer characters sharing much more conversational interaction. This hypothesis is based on the contrast between Williams’s rural “knowable communities” and the sprawling, populous, less conversational urban fictions or Moretti’s and Eagleton’s analyses. If true, it would suggest that the inverse relationship of hypothesis #1 (more characters means less conversation) can be correlated to, and perhaps even caused by, the geography of a novel’s setting. The claims about novelistic geography and social interaction have usually been based on comparisons of a selected few novelists (Jane Austen and Charles Dickens preeminently). Do they remain valid when tested against a larger corpus? 4 Extracting Conversational Networks from Literature In order to test these hypotheses, we developed a novel approach to extracting social networks from literary texts themselves, building on existing analysis tools. We defined “social network” as “conversational network” for purposes of evaluating these literary theories. In a conversational network, vertices represent characters (assumed to be named entities) and edges indicate at least one instance of dialogue interaction between two characters over the course of the novel. The weight of each edge is proportional to the amount of interaction. We define a conversation as a continuous span of narrative time featuring a set of characters in which the following conditions are met: 1. The characters are in the same place at the same time; 2. The characters take turns speaking; and 3. The characters are mutually aware of each other and each character’s speech is mutually intended for the other to hear. In the following subsections, we discuss the methods we devised for the three problems in text processing invoked by this approach: identifying the characters present in a literary text, assigning a “speaker” (if any) to each instance of quoted speech from among those characters, and constructing a social network by detecting conversations from the set of dialogue acts. 4.1 Character Identification The first challenge was to identify the candidate speakers by “chunking” names (such as Mr. Holmes) from the text. We processed each novel 141 with the Stanford NER tagger (Finkel et al., 2005) and extracted noun phrases that were categorized as persons or organizations. We then clustered the noun phrases into coreferents for the same entity (person or organization). The clustering process is as follows: 1. For each named entity, we generate variations on the name that we would expect to see in a coreferent. Each variation omits certain parts of multi-word names, respecting titles and first/last name distinctions, similar to work by Davis et al. (2003). For example, Mr. Sherlock Holmes may refer to the same character as Mr. Holmes, Sherlock Holmes, Sherlock and Holmes. 2. For each named entity, we compile a list of other named entities that may be coreferents, either because they are identical or because one is an expected variation on the other. 3. We then match each named entity to the most recent of its possible coreferents. In aggregate, this creates a cluster of mentions for each character. We also pre-processed the texts to normalize formatting, detect headings and chapter breaks, remove metadata, and identify likely instances of quoted speech (that is, mark up spans of text that fall between quotation marks, assumed to be a superset of the quoted speech present in the text). 4.2 Quoted Speech Attribution In order to programmatically assign a speaker to each instance of quoted speech, we applied a highprecision subset of a general approach we describe elsewhere (Elson and McKeown, 2010). The first step of this approach was to compile a separate training and testing corpus of literary texts from British, American and Russian authors of the nineteenth and twentieth centuries. The training corpus consisted of about 111,000 words including 3,176 instances of quoted speech. To obtain goldstandard annotations, we conducted an online survey via Amazon’s Mechanical Turk program. For each quote, we asked three annotators to independently choose a speaker from the list of contextual candidates– or, choose “spoken by an unlisted character” if the answer was not available, or “not spoken by any character” for non-dialogue cases such as sneer quotes. We divided this corpus into training and testing sets, and used the training set to develop a categorizer that assigned one of five syntactic categories to each quote. For example, if a quote is followed by a verb that indicates verbal expression (such as “said”), and then a character mention, a category called Character trigram is assigned to the quote. The fifth category is a catch-all for quotes that do not fall into the other four. In many cases, the answer can be reliably determined based solely on its syntactic category. For instance, in the Character trigram category, the mentioned character is the quote’s speaker in 99% of both the training and testing sets. In all, we were able to determine the speaker of 57% of the testing set with 96% accuracy just on the basis of syntactic categorization. This is the technique we used to construct our conversational networks. In another study, we applied machine learning tools to the data (one model for each syntactic category) and achieved an overall accuracy of 83% over the entire test set (Elson and McKeown, 2010). The other 43% of quotes are left here as “unknown” speakers; however, in the present study, we are interested in conversations rather than individual quotes. Each conversation is likely to consist of multiple quotes by each speaker, increasing the chances of detecting the interaction. Moreover, this design decision emphasizes the precision of the social networks over their recall. This tilts “in favor” of hypothesis #1 (that there are fewer social interactions in larger communities); however, we shall see that despite the emphasis of precision over recall, we identify a sufficient mass of interactions in the texts to constitute evidence against this hypothesis. 4.3 Constructing social networks We then applied the results from our character identification and quoted speech attribution methods toward the construction of conversational networks from literature. We derived one network from each text in our corpus. We first assigned vertices to character entities that are mentioned repeatedly throughout the novel. Coreferents for the same name (such as Mr. Darcy and Darcy) were grouped into the same vertex. We found that a network that included incidental or single-mention named entities became too noisy to function effectively, so we filtered out the entities that are mentioned fewer than three 142 times in the novel or are responsible for less than 1% of the named entity mentions in the novel. We assigned undirected edges between vertices that represent adjacency in quoted speech fragments. Specifically, we set the weight of each undirected edge between two character vertices to the total length, in words, of all quotes that either character speaks from among all pairs of adjacent quotes in which they both speak– implying face to face conversation. We empirically determined that the most accurate definition of “adjacency” is one where the two characters’ quotes fall within 300 words of one another with no attributed quotes in between. When such an adjacency is found, the length of the quote is added to the edge weight, under the hypothesis that the significance of the relationship between two individuals is proportional to the length of the dialogue that they exchange. Finally, we normalized each edge’s weight by the length of the novel. An example network, automatically constructed in this manner from Jane Austen’s Mansfield Park, is shown in Figure 1. The width of each vertex is drawn to be proportional to the character’s share of all the named entity mentions in the book (so that protagonists, who are mentioned frequently, appear in larger ovals). The width of each edge is drawn to be proportional to its weight (total conversation length). We also experimented with two alternate methods for identifying edges, for purposes of a baseline: 1. The “correlation” method divides the text into 10-paragraph segments and counts the number of mentions of each character in each segment (excluding mentions inside quoted speech). It then computes the Pearson product-moment correlation coefficient for the distributions of mentions for each pair of characters. These coefficients are used for the edge weights. Characters that tend to appear together in the same areas of the novel are taken to be more socially connected, and have a higher edge weight. 2. The “spoken mention” method counts occurrences when one character refers to another in his or her quoted speech. These counts, normalized by the length of the text, are used as edge weights. The intuition is that characters who refer to one another are likely to be in conversation. Figure 1: Automatically extracted conversation network for Jane Austen’s Mansfield Park. 4.4 Evaluation To check the accuracy of our method for extracting conversational networks, we conducted an evaluation involving four of the novels (The Sign of the Four, Emma, David Copperfield and The Portrait of a Lady). We did not use these texts when developing our method for identifying conversations. For each book, we randomly selected 4-5 chapters from among those with significant amounts of quoted speech, so that all excerpts from each novel amounted to at least 10,000 words. We then asked three annotators to identity all the conversations that occur in all 44,000 words. We requested that the annotators include both direct and indirect (unquoted) speech, and define “conversation” as in the beginning of Section 4, but exclude “retold” conversations (those that occur within other dialogue). We processed the annotation results by breaking down each multi-way conversation into all of its unique two-character interactions (for example, a conversation between four people indicates six bilateral interactions). To calculate inter-annotator agreement, we first compiled a list of all possible interactions between all characters in each text. In this model, each annotator contributed a set of “yes” or “no” decisions, one for every character pair. We then applied the kappa measurement for agreement in a binary classification problem (Co143 Method Precision Recall F Speech adjacency .95 .51 .67 Correlation .21 .65 .31 Spoken-mention .45 .49 .47 Table 2: Precision, recall, and F-measure of three methods for detecting bilateral conversations in literary texts. hen, 1960). In 95% of character pairs, annotators were unanimous, which is a high agreement of k = .82. The precision and recall of our method for detecting conversations is shown in Table 2. Precision was .95; this indicates that we can be confident in the specificity of the conversational networks that we automatically construct. Recall was .51, indicating a sensitivity of slightly more than half. There were several reasons that we did not detect the missing links, including indirect speech, quotes attributed to anaphoras or coreferents, and “diffuse” conversations in which the characters do not speak in turn with one another. To calculate precision and recall for the two baseline social networks, we set a threshold t to derive a binary prediction from the continuous edge weights. The precision and recall values shown for the baselines in Table 2 represent the highest performance we achieved by varying t between 0 and 1 (maximizing F-measure over t). Both baselines performed significantly worse in precision and F-measure than our quoted speech adjacency method for detecting conversations. 5 Data Analysis 5.1 Feature extraction We extracted features from the conversational networks that emphasize the complexity of the social interactions found in each novel: 1. The number of characters and the number of speaking characters 2. The variance of the distribution of quoted speech (specifically, the proportion of quotes spoken by the n most frequent speakers, for 1 ≤n ≤5) 3. The number of quotes, and proportion of words in the novel that are quoted speech 4. The number of 3-cliques and 4-cliques in the social network 5. The average degree of the graph, defined as P v∈V |Ev| |V | = 2|E| |V | (1) where |Ev| is the number of edges incident on a vertex v, and |V | is the number of vertices. In other words, this determines the average number of characters connected to each character in the conversational network (“with how many people on average does a character converse?”). 6. A variation on graph density that normalizes the average degree feature by the number of characters: P v∈V |Ev| |V |(|V | −1) = 2|E| |V |(|V | −1) (2) By dividing again by |V | −1, we use this as a metric for the overall connectedness of the graph: “with what percent of the entire network (besides herself) does each character converse, on average?” The weight of the edge, as long as it is greater than 0, does not affect either the network’s average degree or graph density. 5.2 Results We derived results from the data in two ways. First, we examined the strengths of the correlations between the features that we extracted (for example, between number of character vertices and the average degree of each vertex). We used Pearson’s product-moment correlation coefficient in these calculations. Second, we compared the extracted features to the metadata we previously assigned to each text (e.g., urban vs. rural). Hypothesis #1, which we described in Section 3, claims that there is an inverse correlation between the amount of dialogue in a nineteenthcentury novel and the number of characters in that novel. We did not find this to be the case. Rather, we found a weak but positive correlation (r=.16) between the number of quotes in a novel and the number of characters (normalizing the quote count for text length). There was a stronger positive correlation (r=.50) between the number of unique speakers (those characters who speak at least once) and the normalized number of quotes, suggesting that larger networks have more conversations than smaller ones. But because the first 144 correlation is weak, we investigated whether further analysis could identify other evidence that confirms or contradicts the hypothesis. Another way to interpret hypothesis #1 is that social networks with more characters tend to break apart and be less connected. However, we found the opposite to be true. The correlation between the number of characters in each graph and the average degree (number of conversation partners) for each character was a positive, moderately strong r=.42. This is not a given; a network can easily, for example, break into minimally connected or mutually exclusive subnetworks when more characters are involved. Instead, we found that networks tend to stay close-knit regardless of their size: even the density of the graph (the percentage of the community that each character talks to) grows with the total population size at r=.30. Moreover, as the population of speakers grows, the density is likely to increase at r=.49. A higher number of characters (speaking or non-speaking) is also correlated with a higher rate of 3-cliques per character (r=.38), as well as with a more balanced distribution of dialogue (the share of dialogue spoken by the top three speakers decreases at r=−.61). This evidence suggests that in nineteenth-century British literature, it is the small communities, rather than the large ones, that tend to be disconnected. Hypothesis #2, meanwhile, posited that a novel’s setting (urban or rural) would have an effect on the structure of its social network. After defining “social network” as a conversational network, we did not find this to be the case. Surprisingly, the numbers of characters and speakers found in the urban novel were not significantly greater than those found in the rural novel. Moreover, each of the features we extracted, such as the rate of cliques, average degree, density, and rate of characters’ mentions of other characters, did not change in a statistically significant manner between the two genres. For example, Figure 2 shows the mean over all texts of each network’s average degree, with confidence intervals, separated by setting into urban and rural. The increase in degree seen in urban texts is not significant. Rather, the only type of metadata variable that did impact the average degree with any significance was the text’s perspective. Figure 2 also separates texts into first- and third-person tellings and shows the means and confidence intervals for the 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 3rd 1st urban rural Average Degree Setting / Perspective Figure 2: The average degree for each character as a function of the novel’s setting and its perspective. Figure 3: Conversational networks for first-person novels like Collins’s The Woman in White are less connected due to the structure imposed by the perspective. average degree measure. Stories told in the third person had much more connected networks than stories told in the first person: not only did the average degree increase with statistical significance (by the homoscedastic t-test to p < .005), so too did the graph density (p < .05) and the rate of 3-cliques per character (p < .05). We believe the reason for this can be intuited with a visual inspection of a first-person graph. Figure 3 shows the conversational network extracted for Collins’s The Woman in White, which is told in the first person. Not surprisingly, the most oft-repeated named entity in the text is I, referring to the narrator. More surprising is the lack of conversation connections between the auxiliary characters. The story’s structure revolves around the narrator and each character is understood in terms of his or her relationship to the narrator. Private conversations between auxiliary characters would not include the narrator, and thus do not appear in a 145 first-hand account. An “omniscient” third person narrator, by contrast, can eavesdrop on any pair of characters conversing. This highlights the importance of detecting reported and indirect speech in future work, as a first-person narrator may hear about other connections without witnessing them. 6 Literary Interpretation of Results Our data, therefore, markedly do not confirm hypothesis #1. They also suggest, in relation to hypothesis #2 (also not confirmed by the data), a strong reason why. One of the basic assumptions behind hypothesis #2– that urban novels contain more characters, mirroring the masses of nineteenth-century cities– is not borne out by our data. Our results do, however, strongly correlate a point of view (thirdperson narration) with more frequently connected characters, implying tighter and more talkative social networks. We would propose that this suggests that the form of a given novel– the standpoint of the narrative voice, whether the voice is “omniscient” or not– is far more determinative of the kind of social network described in the novel than where it is set or even the number of characters involved. Whereas standard accounts of nineteenth-century fiction, following Bakhtin’s notion of the “chronotope,” emphasize the content of the novel as determinative (where it is set, whether the novel fits within a genre of “village” or “urban” fiction), we have found that content to be surprisingly irrelevant to the shape of social networks within. Bakhtin’s influential theory, and its detailed reworkings by Williams, Moretti, and others, suggests that as the novel becomes more urban, more centered in (and interested in) populous urban settings, the novel’s form changes to accommodate the looser, more populated, less conversational networks of city life. Our data suggests the opposite: that the “urban novel” is not as strongly distinctive a form as has been asserted, and that in fact it can look much like the village fictions of the century, as long as the same method of narration is used. This conclusion leads to some further considerations. We are suggesting that the important element of social networks in nineteenth-century fiction is not where the networks are set, but from what standpoint they are imagined or narrated. Narrative voice, that is, trumps setting. 7 Conclusion In this paper, we presented a method for characterizing a text of literary fiction by extracting the network of social conversations that occur between its characters. This allowed us to take a systematic and wide look at a large corpus of texts, an approach which complements the narrower and deeper analysis performed by literary scholars and can provide evidence for or against some of their claims. In particular, we described a high-precision method for detecting face-to-face conversations between two named characters in a novel, and showed that as the number of characters in a novel grows, so too do the cohesion, interconnectedness and balance of their social network. In addition, we showed that the form of the novel (first- or third-person) is a stronger predictor of these features than the setting (urban or rural). Our results thus far suggest further review of our methods, our corpus and our results for more insights into the social networks found in this and other genres of fiction. 8 Acknowledgments This material is based on research supported in part by the U.S. National Science Foundation (NSF) under IIS-0935360. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the NSF. References Mikhail Bakhtin. 1981. Forms of time and of the chronotope in the novel. In Trans. Michael Holquist and Caryl Emerson, editors, The Dialogic Imagination: Four Essays, pages 84–258. University of Texas Press, Austin. John Burrows. 2004. Textual analysis. In Susan Schreibman, Ray Siemens, and John Unsworth, editors, A Companion to Digital Humanities. Blackwell, Oxford. Nathanael Chambers and Dan Jurafsky. 2008. Unsupervised learning of narrative event chains. In In Proceedings of the 46th Annual Meeting of the Association of Com- putational Linguistics (ACL-08), pages 789–797, Columbus, Ohio. Wendy K. Tam Cho and James H. Fowler. 2010. Legislative success in a small world: Social network analysis and the dynamics of congressional legislation. The Journal of Politics, 72(1):124–135. 146 Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1):37–46. Peter T. Davis, David K. Elson, and Judith L. Klavans. 2003. Methods for precise named entity matching in digital collections. In Proceedings of the Third ACM/IEEE Joint Conference on Digital Libraries (JCDL ’03), Houston, Texas. George Doddington, Alexis Mitchell, Mark Przybocki, Lance Ramshaw, Stephanie Strassel, and Ralph Weischedel. 2004. The automatic content extraction (ace) program tasks, data, and evaluation. In Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC 2004), pages 837–840, Lisbon. Terry Eagleton. 2005. The English Novel: An Introduction. Blackwell, Oxford. David K. Elson and Kathleen R. McKeown. 2010. Automatic attribution of quoted speech in literary narrative. In Proceedings of the Twenty-Fourth AAAI Conference on Artificial Intelligence (AAAI 2010), Atlanta, Georgia. Jenny Rose Finkel, Trond Grenager, and Christopher D. Manning. 2005. Incorporating non-local information into information extraction systems by gibbs sampling. In Proceedings of the 43nd Annual Meeting of the Association for Computational Linguistics (ACL 2005), pages 363–370. Anatoliy Gruzd and Caroline Haythornthwaite. 2008. Automated discovery and analysis of social networks from threaded discussions. In International Network of Social Network Analysis (INSNA) Conference, St. Pete Beach, Florida. Harry Halpin, Johanna D. Moore, and Judy Robertson. 2004. Automatic analysis of plot for story rewriting. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP ’04), Barcelona. John Lee. 2007. A computational model of text reuse in ancient literary texts. In In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics (ACL 2007), pages 472–479, Prague. Andrew McCallum, Xuerui Wang, and Andr´es Corrada-Emmanual. 2007. Topic and role discovery in social networks with experiments on enron and academic email. Journal of Artificial Intelligence Research, 30:249–272. Franco Moretti. 1999. Atlas of the European Novel, 1800-1900. Verso, London. Franco Moretti. 2005. Graphs, Maps, Trees: Abstract Models for a Literary History. Verso, London. Frederick Mostellar and David L. Wallace. 1984. Applied Bayesian and Classical Inference: The Case of The Federalist Papers. Springer, New York. Raymond Williams. 1975. The Country and The City. Oxford University Press, Oxford. 147
|
2010
|
15
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1482–1491, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Learning Arguments and Supertypes of Semantic Relations using Recursive Patterns Zornitsa Kozareva and Eduard Hovy USC Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292-6695 {kozareva,hovy}@isi.edu Abstract A challenging problem in open information extraction and text mining is the learning of the selectional restrictions of semantic relations. We propose a minimally supervised bootstrapping algorithm that uses a single seed and a recursive lexico-syntactic pattern to learn the arguments and the supertypes of a diverse set of semantic relations from the Web. We evaluate the performance of our algorithm on multiple semantic relations expressed using “verb”, “noun”, and “verb prep” lexico-syntactic patterns. Humanbased evaluation shows that the accuracy of the harvested information is about 90%. We also compare our results with existing knowledge base to outline the similarities and differences of the granularity and diversity of the harvested knowledge. 1 Introduction Building and maintaining knowledge-rich resources is of great importance to information extraction, question answering, and textual entailment. Given the endless amount of data we have at our disposal, many efforts have focused on mining knowledge from structured or unstructured text, including ground facts (Etzioni et al., 2005), semantic lexicons (Thelen and Riloff, 2002), encyclopedic knowledge (Suchanek et al., 2007), and concept lists (Katz et al., 2003). Researchers have also successfully harvested relations between entities, such as is-a (Hearst, 1992; Pasca, 2004) and part-of (Girju et al., 2003). The kinds of knowledge learned are generally of two kinds: ground instance facts (New York is-a city, Rome is the capital of Italy) and general relational types (city is-a location, engines are part-of cars). A variety of NLP tasks involving inference or entailment (Zanzotto et al., 2006), including QA (Katz and Lin, 2003) and MT (Mt et al., 1988), require a slightly different form of knowledge, derived from many more relations. This knowledge is usually used to support inference and is expressed as selectional restrictions (Wilks, 1975) (namely, the types of arguments that may fill a given relation, such as person live-in city and airline fly-to location). Selectional restrictions constrain the possible fillers of a relation, and hence the possible contexts in which the patterns expressing that relation can participate in, thereby enabling sense disambiguation of both the fillers and the expression itself. To acquire this knowledge two common approaches are employed: clustering and patterns. While clustering has the advantage of being fully unsupervised, it may or may not produce the types and granularity desired by a user. In contrast pattern-based approaches are more precise, but they typically require a handful to dozens of seeds and lexico-syntactic patterns to initiate the learning process. In a closed domain these approaches are both very promising, but when tackling an unbounded number of relations they are unrealistic. The quality of clustering decreases as the domain becomes more continuously varied and diverse, and it has proven difficult to create collections of effective patterns and high-yield seeds manually. In addition, the output of most harvesting systems is a flat list of lexical semantic expressions such as “New York is-a city” and “virus causes flu”. However, using this knowledge in inference requires it to be formulated appropriately and organized in a semantic repository. (Pennacchiotti and Pantel, 2006) proposed an algorithm for automatically ontologizing semantic relations into WordNet. However, despite its high precision entries, WordNet’s limited coverage makes it impossible for relations whose arguments are not present in WordNet to be incorporated. One would like a procedure that dynamically organizes and extends 1482 its semantic repository in order to be able to accommodate all newly-harvested information, and thereby become a global semantic repository. Given these considerations, we address in this paper the following question: How can the selectional restrictions of semantic relations be learned automatically from the Web with minimal effort using lexico-syntactic recursive patterns? The contributions of the paper are as follows: • A novel representation of semantic relations using recursive lexico-syntactic patterns. • An automatic procedure to learn the selectional restrictions (arguments and supertypes) of semantic relations from Web data. • An exhaustive human-based evaluation of the harvested knowledge. • A comparison of the results with some large existing knowledge bases. The rest of the paper is organized as follows. In the next section, we review related work. Section 3 addresses the representation of semantic relations using recursive patterns. Section 4 describes the bootstrapping mechanism that learns the selectional restrictions of the relations. Section 5 describes data collection. Section 6 discusses the obtained results. Finally, we conclude in Section 7. 2 Related Work A substantial body of work has been done in attempts to harvest bits of semantic information, including: semantic lexicons (Riloff and Shepherd, 1997), concept lists (Lin and Pantel, 2002), isa relations (Hearst, 1992; Etzioni et al., 2005; Pasca, 2004; Kozareva et al., 2008), part-of relations (Girju et al., 2003), and others. Knowledge has been harvested with varying success both from structured text such as Wikipedia’s infoboxes (Suchanek et al., 2007) or unstructured text such as the Web (Pennacchiotti and Pantel, 2006; Yates et al., 2007). A variety of techniques have been employed, including clustering (Lin and Pantel, 2002), co-occurrence statistics (Roark and Charniak, 1998), syntactic dependencies (Pantel and Ravichandran, 2004), and lexico-syntactic patterns (Riloff and Jones, 1999; Fleischman and Hovy, 2002; Thelen and Riloff, 2002). When research focuses on a particular relation, careful attention is paid to the pattern(s) that express it in various ways (as in most of the work above, notably (Riloff and Jones, 1999)). But it has proven a difficult task to manually find effectively different variations and alternative patterns for each relation. In contrast, when research focuses on any relation, as in TextRunner (Yates et al., 2007), there is no standardized manner for re-using the pattern learned. TextRunner scans sentences to obtain relation-independent lexico-syntactic patterns to extract triples of the form (John, fly to, Prague). The middle string denotes some (unspecified) semantic relation while the first and third denote the learned arguments of this relation. But TextRunner does not seek specific semantic relations, and does not re-use the patterns it harvests with different arguments in order to extend their yields. Clearly, it is important to be able to specify both the actual semantic relation sought and use its textual expression(s) in a controlled manner for maximal benefit. The objective of our research is to combine the strengths of the two approaches, and, in addition, to provide even richer information by automatically mapping each harvested argument to its supertype(s) (i.e., its semantic concepts). For instance, given the relation destination and the pattern X flies to Y, automatically determining that John, Prague) and (John, conference) are two valid filler instance pairs, that (RyanAir, Prague) is another, as well as that person and airline are supertypes of the first argument and city and event of the second. This information provides the selectional restrictions of the given semantic relation, indicating that living things like people can fly to cities and events, while non-living things like airlines fly mainly to cities. This is a significant improvement over systems that output a flat list of lexical semantic knowledge (Thelen and Riloff, 2002; Yates et al., 2007; Suchanek et al., 2007). Knowing the sectional restrictions of a semantic relation supports inference in many applications, for example enabling more accurate information extraction. (Igo and Riloff, 2009) report that patterns like “attack on ⟨NP⟩” can learn undesirable words due to idiomatic expressions and parsing errors. Over time this becomes problematic for the bootstrapping process and leads to significant deterioration in performance. (Thelen and Riloff, 2002) address this problem by learning multiple semantic categories simultaneously, relying on the often unrealistic assumption that a word cannot belong to more than one semantic category. How1483 ever, if we have at our disposal a repository of semantic relations with their selectional restrictions, the problem addressed in (Igo and Riloff, 2009) can be alleviated. In order to obtain selectional restriction classes, (Pennacchiotti and Pantel, 2006) made an attempt to ontologize the harvested arguments of is-a, part-of, and cause relations. They mapped each argument of the relation into WordNet and identified the senses for which the relation holds. Unfortunately, despite its very high precision entries, WordNet is known to have limited coverage, which makes it impossible for algorithms to map the content of a relation whose arguments are not present in WordNet. To surmount this limitation, we do not use WordNet, but employ a different method of obtaining superclasses of a filler term: the inverse doubly-anchored patterns DAP−1 (Hovy et al., 2009), which, given two arguments, harvests its supertypes from the source corpus. (Hovy et al., 2009) show that DAP−1 is reliable and it enriches WordNet with additional hyponyms and hypernyms. 3 Recursive Patterns A singly-anchored pattern contains one example of the seed term (the anchor) and one open position for the term to be learned. Most researchers use singly-anchored patterns to harvest semantic relations. Unfortunately, these patterns run out of steam very quickly. To surmount this obstacle, a handful of seeds is generally used, and helps to guarantee diversity in the extraction of new lexicosyntactic patterns (Riloff and Jones, 1999; Snow et al., 2005; Etzioni et al., 2005). Some algorithms require ten seeds (Riloff and Jones, 1999; Igo and Riloff, 2009), while others use a variation of 5, 10, to even 25 seeds (Talukdar et al., 2008). Seeds may be chosen at random (Davidov et al., 2007; Kozareva et al., 2008), by picking the most frequent terms of the desired class (Igo and Riloff, 2009), or by asking humans (Pantel et al., 2009). As (Pantel et al., 2009) show, picking seeds that yield high numbers of different terms is difficult. Thus, when dealing with unbounded sets of relations (Banko and Etzioni, 2008), providing many seeds becomes unrealistic. Interestingly, recent work reports a class of patterns that use only one seed to learn as much information with only one seed. (Kozareva et al., 2008; Hovy et al., 2009) introduce the so-called doublyanchored pattern (DAP) that has two anchor seed positions “⟨type⟩such as ⟨seed⟩and *”, plus one open position for the terms to be learned. Learned terms can then be replaced into the seed position automatically, creating a recursive procedure that is reportedly much more accurate and has much higher final yield. (Kozareva et al., 2008; Hovy et al., 2009) have successfully applied DAP for the learning of hyponyms and hypernyms of is-a relations and report improvements over (Etzioni et al., 2005) and (Pasca, 2004). Surprisingly, this work was limited to the semantic relation is-a. No other study has described the use or effect of recursive patterns for different semantic relations. Therefore, going beyond (Kozareva et al., 2008; Hovy et al., 2009), we here introduce recursive patterns other than DAP that use only one seed to harvest the arguments and supertypes of a wide variety of relations. (Banko and Etzioni, 2008) show that semantic relations can be expressed using a handful of relation-independent lexico-syntactic patterns. Practically, we can turn any of these patterns into recursive form by giving as input only one of the arguments and leaving the other one as an open slot, allowing the learned arguments to replace the initial seed argument directly. For example, for the relation “fly to”, the following recursive patterns can be built: “* and ⟨seed⟩fly to *”, “⟨seed⟩ and * fly to *”, “* fly to ⟨seed⟩and *”, “* fly to * and ⟨seed⟩”, “⟨seed⟩fly to *” or “* fly to ⟨seed⟩”, where ⟨seed⟩is an example like John or Ryanair, and (∗) indicates the position on which the arguments are learned. Conjunctions like and, or are useful because they express list constructions and extract arguments similar to the seed. Potentially, one can explore all recursive pattern variations when learning a relation and compare their yield, however this study is beyond the scope of this paper. We are particularly interested in the usage of recursive patterns for the learning of semantic relations not only because it is a novel method, but also because recursive patterns of the DAP fashion are known to: (1) learn concepts with high precision compared to singly-anchored patterns (Kozareva et al., 2008), (2) use only one seed instance for the discovery of new previously unknown terms, and (3) harvest knowledge with minimal supervision. 1484 4 Bootstrapping Recursive Patterns 4.1 Problem Formulation The main goal of our research is: Task Definition: Given a seed and a semantic relation expressed using a recursive lexico-syntactic pattern, learn in bootstrapping fashion the selectional restrictions (i.e., the arguments and supertypes) of the semantic relation from an unstructured corpus such as the Web. Figure 1 shows an example of the task and the types of information learned by our algorithm. * and John fly to * seed = John relation = fly to BrianKate politicians people artists Delta Alaska airlines carriers bees animals party event Italy France countries New York city flowers trees plants Figure 1: Bootstrapping Recursive Patterns. Given a seed John and a semantic relation fly to expressed using the recursive pattern “* and John fly to *”, our algorithm learns the left side arguments {Brian, Kate, bees, Delta, Alaska} and the right side arguments {flowers, trees, party, New York, Italy, France}. For each argument, the algorithm harvests supertypes such as {people, artists, politicians, airlines, city, countries, plants, event} among others. The colored links between the right and left side concepts denote the selectional restrictions of the relation. For instance, people fly to events and countries, but never to trees or flowers. 4.2 System Architecture We propose a minimally supervised bootstrapping algorithm based on the framework adopted in (Kozareva et al., 2008; Hovy et al., 2009). The algorithm has two phases: argument harvesting and supertype harvesting. The final output is a ranked list of interlinked concepts which captures the selectional restrictions of the relation. 4.2.1 Argument Harvesting In the argument extraction phase, the first bootstrapping iteration is initiated with a seed Y and a recursive pattern “X∗and Y verb+prep|verb|noun Z∗”, where X∗and Z∗are the placeholders for the arguments to be learned. The pattern is submitted to Yahoo! as a web query and all unique snippets matching the query are retrieved. The newly learned and previously unexplored arguments on the X∗position are used as seeds in the subsequent iteration. The arguments on the Z∗position are stored at each iteration, but never used as seeds since the recursivity is created using the terms on X and Y . The bootstrapping process is implemented as an exhaustive breadth-first algorithm which terminates when all arguments are explored. We noticed that despite the specific lexicosyntactic structure of the patterns, erroneous information can be acquired due to part-of-speech tagging errors or flawed facts on the Web. The challenge is to identify and separate the erroneous from the true arguments. We incorporate the harvested arguments on X and Y positions in a directed graph G = (V, E), where each vertex v ∈V is a candidate argument and each edge (u, v) ∈E indicates that the argument v is generated by the argument u. An edge has weight w corresponding to the number of times the pair (u, v) is extracted from different snippets. A node u is ranked by u= P ∀(u,v)∈E w(u,v)+P ∀(v,u)∈E w(v,u) |V |−1 which represents the weighted sum of the outgoing and incoming edges normalized by the total number of nodes in the graph. Intuitively, our confidence in a correct argument u increases when the argument (1) discovers and (2) is discovered by many different arguments. Similarly, to rank the arguments standing on the Z position, we build a bipartite graph G′ = (V ′, E′) that has two types of vertices. One set of vertices represents the arguments found on the Y position in the recursive pattern. We will call these Vy. The second set of vertices represents the arguments learned on the Z position. We will call these Vz. We create an edge e′(u′, v′) ∈E′ between u′ ∈Vy and v′ ∈Vz when the argument on the Z position represented by v′ was harvested by the argument on the Y position represented by u′. The weight w′ of the edge indicates the number of times an argument on the Y position found Z. Vertex v′ is ranked as v′= P ∀(u′,v′)∈E′ w(u′,v′) |V ′|−1 . In a very large corpus, like the Web, we assume that a correct argument Z is the one that is frequently discovered by various arguments Y . 1485 4.2.2 Supertype Harvesting In the supertype extraction phase, we take all <X,Y> argument pairs collected during the argument harvesting stage and instantiate them in the inverse DAP−1 pattern “* such as X and Y”. The query is sent to Yahoo! as a web query and all 1000 snippets matching the pattern are retrieved. For each <X,Y> pair, the terms on the (*) position are extracted and considered as candidate supertypes. To avoid the inclusion of erroneous supertypes, again we build a bipartite graph G′′ = (V ′′, E′′). The set of vertices Vsup represents the supertypes, while the set of vertices Vp corresponds to the ⟨X,Y⟩pair that produced the supertype. An edge e′′(u′′, v′′) ∈E′′, where u′′ ∈Vp and v′′ ∈Vsup shows that the pair ⟨X,Y⟩denoted as u′′ harvested the supertype represented by v′′. For example, imagine that the argument X∗= Ryanair was harvested in the previous phase by the recursive pattern “X∗and EasyJet fly to Z∗”. Then the pair ⟨Ryanair,EasyJet⟩forms a new Web query “* such as Ryanair and EasyJet” which learns the supertypes “airlines” and “carriers”. The bipartite graph has two vertices v′′ 1 and v′′ 2 for the supertypes “airlines” and “carriers”, one vertex u′′ 3 for the argument pair ⟨Ryanair, EasyJet⟩, and two edges e′′ 1(u′′ 3, v′′ 1) and e′′ 2(u′′ 3, v′′ 1). A vertex v′′ ∈Vsup is ranked by v′′= P ∀(u′′,v′′)∈E′′ w(u′′,v′′) |V ′′|−1 . Intuitively, a supertype which is discovered multiple times by various argument pairs is ranked highly. However, it might happen that a highly ranked supertype actually does not satisfy the selectional restrictions of the semantic relation. To avoid such situations, we further instantiate each supertype concept in the original pattern1. For example, “aircompanies fly to *” and “carriers fly to *”. If the candidate supertype produces many web hits for the query, then this suggests that the term is a relevant supertype. Unfortunately, to learn the supertypes of the Z arguments, currently we have to form all possible combinations among the top 150 highly ranked concepts, because these arguments have not been learned through pairing. For each pair of Z arguments, we repeat the same procedure as described above. 1Except for the “dress” and “person” relations, where the targeted arguments are adjectives, and the supertypes are nouns. 5 Semantic Relations So far, we have described the mechanism that learns from one seed and a recursive pattern the selectional restrictions of any semantic relation. Now, we are interested in evaluating the performance of our algorithm. A natural question that arises is: “How many patterns are there?”. (Banko and Etzioni, 2008) found that 95% of the semantic relations can be expressed using eight lexico-syntactic patterns. Space prevents us from describing all of them, therefore we focus on the three most frequent patterns which capture a large diversity of semantic relations. The relative frequency of these patterns is 37.80% for “verbs”, 22.80% for “noun prep”, and 16.00% for “verb prep”. 5.1 Data Collection Table 1 shows the lexico-syntactic pattern and the initial seed we used to express each semantic relation. To collect data, we ran our knowledge harvesting algorithm until complete exhaustion. For each query submitted to Yahoo!, we retrieved the top 1000 web snippets and kept only the unique ones. In total, we collected 30GB raw data which was part-of-speech tagged and used for the argument and supertype extraction. Table 1 shows the obtained results. recursive pattern seed X arg Z arg #iter X and Y work for Z Charlie 2949 3396 20 X and Y fly to Z EasyJet 772 1176 19 X and Y go to Z Rita 18406 27721 13 X and Y work in Z John 4142 4918 13 X and Y work on Z Mary 4126 5186 7 X and Y work at Z Scott 1084 1186 14 X and Y live in Z Harry 8886 19698 15 X and Y live at Z Donald 1102 1175 15 X and Y live with Z Peter 1344 834 11 X and Y cause Z virus 12790 52744 19 X and Y celebrate Jim 6033 – 12 X and Y drink Sam 1810 – 13 X and Y dress nice 1838 – 8 X and Y person scared 2984 – 17 Table 1: Total Number of Harvested Arguments. An interesting characteristic of the recursive patterns is the speed of leaning which can be measured in terms of the number of unique arguments acquired during each bootstrapping iteration. Figure 2 shows the bootstrapping process for the “cause” and “dress” relations. Although both relations differ in terms of the total number of iterations and harvested items, the overall behavior of the learning curves is similar. Learning starts of very slowly and as bootstrapping progresses a 1486 rapid growth is observed until a saturation point is reached. 0 10000 20000 30000 40000 50000 60000 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 #Items Learned Iterations X and Y Cause Z X Z 0 500 1000 1500 2000 1 2 3 4 5 6 7 8 #Items Learned Iterations X and Y Dress X Figure 2: Items extracted in 10 iterations. The speed of leaning is related to the connectivity behavior of the arguments of the relation. Intuitively, a densely connected graph takes shorter time (i.e., fewer iterations) to be learned, as in the “work on” relation, while a weakly connected network takes longer time to harvest the same amount of information, as in the “work for” relation. 6 Results In this section, we evaluate the results of our knowledge harvesting algorithm. Initially, we decided to conduct an automatic evaluation comparing our results to knowledge bases that have been extracted in a similar way (i.e., through pattern application over unstructured text). However, it is not always possible to perform a complete comparison, because either researchers have not fully explored the same relations we have studied, or for those relations that overlap, the gold standard data was not available. The online demo of TextRunner2 (Yates et al., 2007) actually allowed us to collect the arguments for all our semantic relations. However, due to Web based query limitations, TextRunner returns only the first 1000 snippets. Since we do not have the complete and ranked output of TextRunner, comparing results in terms of recall and precision is impossible. Turning instead to results obtained from structured sources (which one expects to have high correctness), we found that two of our relations overlap with those of the freely available ontology Yago (Suchanek et al., 2007), which was harvested from the Infoboxes tables in Wikipedia. In addition, we also had two human annotators judge as many results as we could afford, to obtain Precision. We conducted two evaluations, one for the arguments and one for the supertypes. 2http://www.cs.washington.edu/research/textrunner/ 6.1 Human-Based Argument Evaluation In this section, we discuss the results of the harvested arguments. For each relation, we selected the top 200 highly ranked arguments. We hired two annotators to judge their correctness. We created detailed annotation guidelines that define the labels for the arguments of the relations, as shown in Table 2. (Previously, for the same task, researchers have not conducted such an exhaustive and detailed human-based evaluation.) The annotation was conducted using the CAT system3. TYPE LABEL EXAMPLES Correct Person John, Mary Role mother, president Group team, Japanese Physical yellow, shabby NonPhysical ugly, thought NonLiving airplane Organization IBM, parliament Location village, New York, in the house Time at 5 o’clock Event party, prom, earthquake State sick, anrgy Manner live in happiness Medium work on Linux, Word Fixed phrase go to war Incorrect Error wrong part-of-speech tag Other none of the above Table 2: Annotation Labels. We allow multiple labels to be assigned to the same concept, because sometimes the concept can appear in different contexts that carry various conceptual representations. Although the labels can be easily collapsed to judge correct and incorrect terms, the fine-grained annotation shown here provides a better overview of the information learned by our algorithm. We measured the inter-annotator agreement for all labels and relations considering that a single entry can be tagged with multiple labels. The Kappa score is around 0.80. This judgement is good enough to warrant using these human judgements to estimate the accuracy of the algorithm. We compute Accuracy as the number of examples tagged as Correct divided by the total number of examples. Table 4 shows the obtained results. The overall accuracy of the argument harvesting phase is 91%. The majority of the occurred errors are due to part-of-speech tagging. Table 3 shows a sample of 10 randomly selected examples from the top 200 ranked and manually annotated arguments. 3http://cat.ucsur.pitt.edu/default.aspx 1487 Relation Arguments (X) Dress: stylish, comfortable, expensive, shabby, gorgeous silver, clean, casual, Indian, black (X) Person: honest, caring, happy, intelligent, gifted friendly, responsible, mature, wise, outgoing (X) Cause: pressure, stress, fire, bacteria, cholesterol flood, ice, cocaine, injuries, wars GoTo (Z): school, bed, New York, the movies, the park, a bar the hospital, the church, the mall, the beach LiveIn (Z): peace, close proximity, harmony, Chicago, town New York, London, California, a house, Australia WorkFor (Z): a company, the local prison, a gangster, the show a boss, children, UNICEF, a living, Hispanics Table 3: Examples of Harvested Arguments. 6.2 Comparison against Existing Resources In this section, we compare the performance of our approach with the semantic knowledge base Yago4 that contains 2 million entities5, 95% of which were manually confirmed to be correct. In this study, we compare only the unique arguments of the “live in” and “work at” relations. We provide Precision scores using the following measures: P rY ago = #terms found in Y ago #terms harvested by system P rHuman = #terms judged correct by human #terms harvested by system NotInY ago = #terms judged correct by human but not in Y ago Table 5 shows the obtained results. We carefully analyzed those arguments that were found by one of the systems but were missing in the other. The recursive patterns learn information about non-famous entities like Peter and famous entities like Michael Jordan. In contrast, Yago contains entries mostly about famous entities, because this is the predominant knowledge in Wikipedia. For the “live in” relation, both repositories contain the same city and country names. However, the recursive pattern learned arguments like pain, effort which express a manner of living, and locations like slums, box. This information is missing from Yago. Similarly for the “work at” relation, both systems learned that people work at universities. In addition, the recursive pattern learned a diversity of company names absent from Yago. While it is expected that our algorithm finds many terms not contained in Yago—specifically, the information not deemed worthy of inclusion in Wikipedia—we are interested in the relatively large number of terms contained in Yago but not found by our algorithm. To our knowledge, no 4http://www.mpi-inf.mpg.de/yago-naga/yago/ 5Names of cities, people, organizations among others. X WorkFor A1 A2 WorkFor Z A1 A2 Person 148 152 Organization 111 110 Role 5 7 Person 60 60 Group 12 14 Event 4 2 Organization 8 7 Time 4 5 NonPhysical 22 23 NonPhysical 18 19 Other 5 5 Other 3 4 Acc. .98 .98 Acc. .99 .98 X Cause A1 A2 Cause Z A1 A2 PhysicalObj 82 75 PhysicalObj 15 20 NonPhysicalObj 69 66 NonPhysicalObj 89 91 Event 21 24 Event 72 72 State 29 31 State 50 50 Other 3 4 Other 5 4 Acc. .99 .98 Acc. .98 .98 X GoTo A1 A2 GoTo Z A1 A2 Person 190 188 Location 163 155 Role 4 4 Event 21 30 Group 3 3 Person 11 13 NonPhysical 1 3 NonPhysical 2 1 Other 2 2 Other 3 1 Acc. .99 .99 Acc. .99 .99 X FlyTo A1 A2 FlyTo Z A1 A2 Person 140 139 Location 199 198 Organization 54 57 Event 1 2 NonPhysical 2 2 Person 0 0 Other 4 2 Other 0 0 Acc. .98 .99 Acc. 1 1 X WorkOn A1 A2 WorkOn Z A1 A2 Person 173 172 Location 110 108 Role 2 3 Organization 27 25 Group 4 5 Manner 38 40 Organization 6 6 Time 4 4 NonPhysical 15 14 NonPhysical 18 21 Error 1 1 Medium 8 8 Other 1 1 Other 13 15 Acc. .99 .99 Acc. .94 .93 X WorkIn A1 A2 WorkIn Z A1 A2 Person 117 118 Location 104 111 Group 10 9 Organization 10 25 Organization 3 3 Manner 39 40 Fixed 3 1 Time 4 4 NonPhysical 55 59 NonPhysical 22 21 Error 12 10 Medium 8 8 Other 0 0 Error 13 15 Acc. .94 .95 Acc. .94 .93 X WorkAt A1 A2 WorkAt Z A1 A2 Person 193 192 Organization 189 190 Role 1 1 Manner 5 4 Group 1 1 Time 3 3 Organization 0 0 Error 3 2 Other 5 6 Other 0 1 Acc. .98 .97 Acc. .99 .99 X LiveIn A1 A2 LiveIn Z A1 A2 Person 185 185 Location 182 186 Role 3 4 Manner 6 8 Group 9 8 Time 1 2 NonPhysical 1 2 Fixed 5 2 Other 2 1 Other 6 2 Acc. .99 .99 Acc. .97 .99 X LiveAt A1 A2 LiveAt Z A1 A2 Person 196 195 Location 158 157 Role 1 1 Person 5 7 NonPhysical 0 1 Manner 1 2 Other 3 3 Error 36 34 Acc. .99 .99 Acc. .82 .83 X LiveWith A1 A2 LiveWith Z A1 A2 Person 188 187 Person 165 163 Role 6 6 Animal 2 4 Group 2 2 Manner 15 15 NonPhysical 2 3 NonPhysical 15 15 Other 2 2 Other 3 3 Acc. .99 .99 Acc. .99 .99 X Dress A1 A2 X Person A1 A2 Physical 72 59 Physical 8 2 NonPhysical 120 136 NonPhysical 188 194 Other 8 5 Other 4 4 Acc .96 .98 Acc. .98 .98 X Drink A1 A2 X Celebrate A1 A2 Living 165 174 Living 157 164 NonLiving 8 2 NonLiving 42 35 Error 27 24 Error 1 1 Acc .87 .88 Acc. .99 .99 Table 4: Harvested Arguments. 1488 P rY ago P rHuman NotInYago X LiveIn .19 (2863/14705) .58 (5165)/8886 2302 LiveIn Z .10 (495/4754) .72 (14248)/19698 13753 X WorkAt .12(167/1399) .88 (959)/1084 792 WorkAt Z .3(15/525) .95 (1128)/1186 1113 Table 5: Comparison against Yago. other automated harvesting algorithm has ever been compared to Yago, and our results here form a baseline that we aim to improve upon. And in the future, one can build an extensive knowledge harvesting system combining the wisdom of the crowd and Wikipedia. 6.3 Human-Based Supertype Evaluation In this section, we discuss the results of harvesting the supertypes of the learned arguments. Figure 3 shows the top 100 ranked supertypes for the “cause” and “work on” relations. The x-axis indicates a supertype, the y-axis denotes the number of different argument pairs that lead to the discovery of the supertype. 0 100 200 300 400 500 600 700 800 900 1000 10 20 30 40 50 60 70 80 90 100 #Pairs Discovering the Supertype Supertype WorkOn Cause Figure 3: Ranked Supertypes. The decline of the curve indicates that certain supertypes are preferred and shared among different argument pairs. It is interesting to note that the text on the Web prefers a small set of supertypes, and to see what they are. These most-popular harvested types tend to be the more descriptive terms. The results indicate that one does not need an elaborate supertype hierarchy to handle the selectional restrictions of semantic relations. Since our problem definition differs from available related work, and WordNet does not contain all harvested arguments as shown in (Hovy et al., 2009), it is not possible to make a direct comparison. Instead, we conduct a manual evaluation of the most highly ranked supertypes which normally are the top 20. The overall accuracy of the supertypes for all relations is 92%. Table 6 shows the Relation Arguments (Supx) Celebrate: men, people, nations, angels, workers, children countries, teams, parents, teachers (Supx) Dress: colors, effects, color tones, activities, patterns styles, materials, size, languages, aspects (Supx) FlyTo: airlines, carriers, companies, giants, people competitors, political figures, stars, celebs Cause (Supz): diseases, abnormalities, disasters, processes, isses disorders, discomforts, emotions, defects, symptoms WorkFor (Supz) organizations, industries, people, markets, men automakers, countries, departments, artists, media GoTo (Supz) : countries, locations, cities, people, events men, activities, games, organizations, FlyTo (Supz) places, countries, regions, airports, destinations locations, cities, area, events Table 6: Examples of Harvested Supertypes. top 10 highly ranked supertypes for six of our relations. 7 Conclusion We propose a minimally supervised algorithm that uses only one seed example and a recursive lexicosyntactic pattern to learn in bootstrapping fashion the selectional restrictions of a large class of semantic relations. The principal contribution of the paper is to demonstrate that this kind of pattern can be applied to almost any kind of semantic relation, as long as it is expressible in a concise surface pattern, and that the recursive mechanism that allows each newly acquired term to restart harvesting automatically is a significant advance over patterns that require a handful of seeds to initiate the learning process. It also shows how one can combine free-form but undirected pattern-learning approaches like TextRunner with more-controlled but effort-intensive approaches like commonly used. In our evaluation, we show that our algorithm is capable of extracting high quality non-trivial information from unstructured text given very restricted input (one seed). To measure the performance of our approach, we use various semantic relations expressed with three lexico-syntactic patterns. For two of the relations, we compare results with the freely available ontology Yago, and conduct a manual evaluation of the harvested terms. We will release the annotated and the harvested data to the public to be used for comparison by other knowledge harvesting algorithms. The success of the proposed framework opens many challenging directions. We plan to use the algorithm described in this paper to learn the selectional restrictions of numerous other relations, in order to build a rich knowledge repository 1489 that can support a variety of applications, including textual entailment, information extraction, and question answering. Acknowledgments This research was supported by DARPA contract number FA8750-09-C-3705. References Michele Banko and Oren Etzioni. 2008. The tradeoffs between open and traditional relation extraction. In Proceedings of ACL-08: HLT, pages 28–36, June. Dmitry Davidov, Ari Rappoport, and Moshel Koppel. 2007. Fully unsupervised discovery of conceptspecific relationships by web mining. In Proc. of the 45th Annual Meeting of the Association of Computational Linguistics, pages 232–239, June. Oren Etzioni, Michael Cafarella, Doug Downey, AnaMaria Popescu, Tal Shaked, Stephen Soderland, Daniel S. Weld, and Alexander Yates. 2005. Unsupervised named-entity extraction from the web: an experimental study. Artificial Intelligence, 165(1):91–134, June. Michael Fleischman and Eduard Hovy. 2002. Fine grained classification of named entities. In Proceedings of the 19th international conference on Computational linguistics, pages 1–7. Roxana Girju, Adriana Badulescu, and Dan Moldovan. 2003. Learning semantic constraints for the automatic discovery of part-whole relations. In Proc. of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology, pages 1–8. Marti Hearst. 1992. Automatic acquisition of hyponyms from large text corpora. In Proc. of the 14th conference on Computational linguistics, pages 539–545. Eduard Hovy, Zornitsa Kozareva, and Ellen Riloff. 2009. Toward completeness in concept extraction and classification. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 948–957. Sean Igo and Ellen Riloff. 2009. Corpus-based semantic lexicon induction with web-based corroboration. In Proceedings of the Workshop on Unsupervised and Minimally Supervised Learning of Lexical Semantics. Boris Katz and Jimmy Lin. 2003. Selectively using relations to improve precision in question answering. In In Proceedings of the EACL-2003 Workshop on Natural Language Processing for Question Answering, pages 43–50. Boris Katz, Jimmy Lin, Daniel Loreto, Wesley Hildebrandt, Matthew Bilotti, Sue Felshin, Aaron Fernandes, Gregory Marton, and Federico Mora. 2003. Integrating web-based and corpus-based techniques for question answering. In Proceedings of the twelfth text retrieval conference (TREC), pages 426– 435. Zornitsa Kozareva, Ellen Riloff, and Eduard Hovy. 2008. Semantic class learning from the web with hyponym pattern linkage graphs. In Proceedings of ACL-08: HLT, pages 1048–1056. Dekang Lin and Patrick Pantel. 2002. Concept discovery from text. In Proc. of the 19th international conference on Computational linguistics, pages 1–7. Characteristics Of Mt, John Lehrberger, Laurent Bourbeau, Philadelphia John Benjamins, and Rita Mccardell. 1988. Machine Translation: Linguistic Characteristics of Mt Systems and General Methodology of Evaluation. John Benjamins Publishing Co(1988-03). Patrick Pantel and Deepak Ravichandran. 2004. Automatically labeling semantic classes. In Proc. of Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics, pages 321–328. Patrick Pantel, Eric Crestan, Arkady Borkovsky, AnaMaria Popescu, and Vishnu Vyas. 2009. Webscale distributional similarity and entity set expansion. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 938–947, August. Marius Pasca. 2004. Acquisition of categorized named entities for web search. In Proc. of the thirteenth ACM international conference on Information and knowledge management, pages 137–145. Marco Pennacchiotti and Patrick Pantel. 2006. Ontologizing semantic relations. In ACL-44: Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the Association for Computational Linguistics, pages 793–800. Ellen Riloff and Rosie Jones. 1999. Learning dictionaries for information extraction by multi-level bootstrapping. In AAAI ’99/IAAI ’99: Proceedings of the Sixteenth National Conference on Artificial intelligence. Ellen Riloff and Jessica Shepherd. 1997. A CorpusBased Approach for Building Semantic Lexicons. In Proc. of the Second Conference on Empirical Methods in Natural Language Processing, pages 117–124. Brian Roark and Eugene Charniak. 1998. Nounphrase co-occurrence statistics for semiautomatic semantic lexicon construction. In Proceedings of the 17th international conference on Computational linguistics, pages 1110–1116. 1490 Rion Snow, Daniel Jurafsky, and Andrew Y. Ng. 2005. Learning syntactic patterns for automatic hypernym discovery. In Advances in Neural Information Processing Systems 17, pages 1297–1304. MIT Press. Fabian M. Suchanek, Gjergji Kasneci, and Gerhard Weikum. 2007. Yago: a core of semantic knowledge. In WWW ’07: Proceedings of the 16th international conference on World Wide Web, pages 697– 706. Partha Pratim Talukdar, Joseph Reisinger, Marius Pasca, Deepak Ravichandran, Rahul Bhagat, and Fernando Pereira. 2008. Weakly-supervised acquisition of labeled class instances using graph random walks. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP 2008, pages 582–590. Michael Thelen and Ellen Riloff. 2002. A Bootstrapping Method for Learning Semantic Lexicons Using Extraction Pattern Contexts. In Proc. of the 2002 Conference on Empirical Methods in Natural Language Processing, pages 214–221. Yorick Wilks. 1975. A preferential pattern-seeking, semantics for natural language inference. Artificial Intelligence, 6(1):53–74. Alexander Yates, Michael Cafarella, Michele Banko, Oren Etzioni, Matthew Broadhead, and Stephen Soderland. 2007. Textrunner: open information extraction on the web. In NAACL ’07: Proceedings of Human Language Technologies: The Annual Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations on XX, pages 25–26. Fabio Massimo Zanzotto, Marco Pennacchiotti, and Maria Teresa Pazienza. 2006. Discovering asymmetric entailment relations between verbs using selectional preferences. In ACL-44: Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the Association for Computational Linguistics, pages 849– 856. 1491
|
2010
|
150
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1492–1501, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics A Transition-Based Parser for 2-Planar Dependency Structures Carlos G´omez-Rodr´ıguez Departamento de Computaci´on Universidade da Coru˜na, Spain [email protected] Joakim Nivre Department of Linguistics and Philology Uppsala University, Sweden [email protected] Abstract Finding a class of structures that is rich enough for adequate linguistic representation yet restricted enough for efficient computational processing is an important problem for dependency parsing. In this paper, we present a transition system for 2-planar dependency trees – trees that can be decomposed into at most two planar graphs – and show that it can be used to implement a classifier-based parser that runs in linear time and outperforms a stateof-the-art transition-based parser on four data sets from the CoNLL-X shared task. In addition, we present an efficient method for determining whether an arbitrary tree is 2-planar and show that 99% or more of the trees in existing treebanks are 2-planar. 1 Introduction Dependency-based syntactic parsing has become a widely used technique in natural language processing, and many different parsing models have been proposed in recent years (Yamada and Matsumoto, 2003; Nivre et al., 2004; McDonald et al., 2005a; Titov and Henderson, 2007; Martins et al., 2009). One of the unresolved issues in this area is the proper treatment of non-projective dependency trees, which seem to be required for an adequate representation of predicate-argument structure, but which undermine the efficiency of dependency parsing (Neuhaus and Br¨oker, 1997; BuchKromann, 2006; McDonald and Satta, 2007). Caught between the Scylla of linguistically inadequate projective trees and the Charybdis of computationally intractable non-projective trees, some researchers have sought a middle ground by exploring classes of mildly non-projective dependency structures that strike a better balance between expressivity and complexity (Nivre, 2006; Kuhlmann and Nivre, 2006; Kuhlmann and M¨ohl, 2007; Havelka, 2007). Although these proposals seem to have a very good fit with linguistic data, in the sense that they often cover 99% or more of the structures found in existing treebanks, the development of efficient parsing algorithms for these classes has met with more limited success. For example, while both Kuhlmann and Satta (2009) and G´omez-Rodr´ıguez et al. (2009) have shown how well-nested dependency trees with bounded gap degree can be parsed in polynomial time, the best time complexity for lexicalized parsing of this class remains a prohibitive O(n7), which makes the practical usefulness questionable. In this paper, we explore another characterization of mildly non-projective dependency trees based on the notion of multiplanarity. This was originally proposed by Yli-Jyr¨a (2003) but has so far played a marginal role in the dependency parsing literature, because no algorithm was known for determining whether an arbitrary tree was mplanar, and no parsing algorithm existed for any constant value of m. The contribution of this paper is twofold. First, we present a procedure for determining the minimal number m such that a dependency tree is m-planar and use it to show that the overwhelming majority of sentences in dependency treebanks have a tree that is at most 2planar. Secondly, we present a transition-based parsing algorithm for 2-planar dependency trees, developed in two steps. We begin by showing how the stack-based algorithm of Nivre (2003) can be generalized from projective to planar structures. We then extend the system by adding a second stack and show that the resulting system captures exactly the set of 2-planar structures. Although the contributions of this paper are mainly theoretical, we also present an empirical evaluation of the 2planar parser, showing that it outperforms the projective parser on four data sets from the CoNLL-X shared task (Buchholz and Marsi, 2006). 1492 2 Preliminaries 2.1 Dependency Graphs Let w = w1 . . . wn be an input string.1 An interval (with endpoints i and j) of the string w is a set of the form [i, j] = {wk | i ≤k ≤j}. Definition 1. A dependency graph for w is a directed graph G = (Vw, E), where Vw = [1, n] and E ⊆Vw × Vw. We call an edge (wi, wj) in a dependency graph G a dependency link2 from wi to wj. We say that wi is the parent (or head) of wj and, conversely, that wj is a syntactic child (or dependent) of wi. For convenience, we write wi →wj ∈E if the link (wi, wj) exists; wi ↔wj ∈E if there is a link from wi to wj or from wj to wi; wi →∗wj ∈E if there is a (possibly empty) directed path from wi to wj; and wi ↔∗wj ∈E if there is a (possibly empty) path between wi and wj in the undirected graph underlying G (omitting reference to E when clear from the context). The projection of a node wi, denoted ⌊wi⌋, is the set of reflexive-transitive dependents of wi: ⌊wi⌋= {wj ∈V | wi →∗wj}. Most dependency representations do not allow arbitrary dependency graphs but typically require graphs to be acyclic and have at most one head per node. Such a graph is called a dependency forest. Definition 2. A dependency graph G for a string w1 . . . wn is said to be a forest iff it satisfies: 1. Acyclicity: If wi →∗wj, then not wj →wi. 2. Single-head: If wj →wi, then not wk →wi (for every k ̸= j). Nodes in a forest that do not have a head are called roots. Some frameworks require that dependency forests have a unique root (i.e., are connected). Such a forest is called a dependency tree. 2.2 Projectivity For reasons of computational efficiency, many dependency parsers are restricted to work with projective dependency structures, that is, forests in which the projection of each node corresponds to a contiguous substring of the input: 1For notational convenience, we will assume throughout the paper that all symbols in an input string are distinct, i.e., i ̸= j ⇔wi ̸= wj. This can be guaranteed in practice by annotating each terminal symbol with its position in the input. 2In practice, dependency links are usually labeled, but to simplify the presentation we will ignore labels throughout most of the paper. However, all the results and algorithms presented can be applied to labeled dependency graphs and will be so applied in the experimental evaluation. Definition 3. A dependency forest G for a string w1 . . . wn is projective iff ⌊wi⌋is an interval for every word wi ∈[1, n]. Projective dependency trees correspond to the set of structures that can be induced from lexicalised context-free derivations (Kuhlmann, 2007; Gaifman, 1965). Like context-free grammars, projective dependency trees are not sufficient to represent all the linguistic phenomena observed in natural languages, but they have the advantage of being efficiently parsable: their parsing problem can be solved in cubic time with chart parsing techniques (Eisner, 1996; G´omez-Rodr´ıguez et al., 2008), while in the case of general non-projective dependency forests, it is only tractable under strong independence assumptions (McDonald et al., 2005b; McDonald and Satta, 2007). 2.3 Planarity The concept of planarity (Sleator and Temperley, 1993) is closely related to projectivity3 and can be informally defined as the property of a dependency forest whose links can be drawn above the words without crossing.4 To define planarity more formally, we first define crossing links as follows: let (wi, wk) and (wj, wl) be dependency links in a dependency graph G. Without loss of generality, we assume that min(i, k) ≤min(j, l). Then, the links are said to be crossing if min(i, k) < min(j, l) < max(i, k) < max(j, l). Definition 4. A dependency graph is planar iff it does not contain a pair of crossing links. 2.4 Multiplanarity The concept of planarity on its own does not seem to be very relevant as an extension of projectivity for practical dependency parsing. According to the results by Kuhlmann and Nivre (2006), most non-projective structures in dependency treebanks are also non-planar, so being able to parse planar structures will only give us a modest improvement in coverage with respect to a projective parser. However, our interest in planarity is motivated by the fact that it can be generalised to multiplanarity (Yli-Jyr¨a, 2003): 3For dependency forests that are extended with a unique artificial root located at position 0, as is commonly done, the two notions are equivalent. 4Planarity in the context of dependency structures is not to be confused with the homonymous concept in graph theory, which does not restrict links to be drawn above the nodes. 1493 Figure 1: A 2-planar dependency structure with two different ways of distributing its links into two planes (represented by solid and dotted lines). Definition 5. A dependency graph G = (V, E) is m-planar iff there exist planar dependency graphs G1 = (V, E1), . . . , Gm = (V, Em) (called planes) such that E = E1 ∪· · · ∪Em. Intuitively, we can associate planes with colours and say that a dependency graph G is m-planar if it is possible to assign one of m colours to each of its links in such a way that links with the same colour do not cross. Note that there may be multiple ways of dividing an m-planar graph into planes, as shown in the example of Figure 1. 3 Determining Multiplanarity Several constraints on non-projective dependency structures have been proposed recently that seek a good balance between parsing efficiency and coverage of non-projective phenomena present in natural language treebanks. For example, Kuhlmann and Nivre (2006) and Havelka (2007) have shown that the vast majority of structures present in existing treebanks are well-nested and have a small gap degree (Bodirsky et al., 2005), leading to an interest in parsers for these kinds of structures (G´omezRodr´ıguez et al., 2009). No similar analysis has been performed for m-planar structures, although Yli-Jyr¨a (2003) provides evidence that all except two structures in the Danish dependency treebank are at most 3-planar. However, his analysis is based on constraints that restrict the possible ways of assigning planes to dependency links, and he is not guaranteed to find the minimal number m for which a given structure is m-planar. In this section, we provide a procedure for finding the minimal number m such that a dependency graph is m-planar and use it to show that the vast majority of sentences in dependency treebanks are Figure 2: The crossings graph corresponding to the dependency structure of Figure 1. at most 2-planar, with a coverage comparable to that of well-nestedness. The idea is to reduce the problem of determining whether a dependency graph G = (V, E) is m-planar, for a given value of m, to a standard graph colouring problem. Consider first the following undirected graph: U(G) = (E, C) where C = {{ei, ej} | ei, ej are crossing links in G} This graph, which we call the crossings graph of G, has one node corresponding to each link in the dependency graph G, with an undirected link between two nodes if they correspond to crossing links in G. Figure 2 shows the crossings graph of the 2-planar structure in Figure 1. As noted in Section 2.4, a dependency graph G is m-planar if each of its links can be assigned one of m colours in such a way that links with the same colours do not cross. In terms of the crossings graph, this means that G is m-planar if each of the nodes of U(G) can be assigned one of m colours such that no two neighbours have the same colour. This amounts to solving the well-known kcolouring problem for U(G), where k = m. For k = 1 the problem is trivial: a graph is 1colourable only if it has no edges. For k = 2, the problem can be solved in time linear in the size of the graph by simple breadth-first search. Given a graph U = (V, E), we pick an arbitrary node v and give it one of two colours. This forces us to give the other colour to all its neighbours, the first colour to the neighbours’ neighbours, and so on. This process continues until we have processed all the nodes in the connected component of v. If this has resulted in assigning two different colours to the same node, the graph is not 2-colourable. Otherwise, we have obtained a 2-colouring of the connected component of U that contains v. If there are still unprocessed nodes, we repeat the process by arbitrarily selecting one of them, continue with the rest of the connected components, and in this way obtain a 2-colouring of the whole graph if it 1494 Language Structures Non-Projective Not Planar Not 2-Planar Not 3-Pl. Not 4-pl. Ill-nested Arabic 2995 205 ( 6.84%) 158 ( 5.28%) 0 (0.00%) 0 (0.00%) 0 (0.00%) 1 (0.03%) Czech 87889 20353 (23.16%) 16660 (18.96%) 82 (0.09%) 0 (0.00%) 0 (0.00%) 96 (0.11%) Danish 5512 853 (15.48%) 827 (15.00%) 1 (0.02%) 1 (0.02%) 0 (0.00%) 6 (0.11%) Dutch 13349 4865 (36.44%) 4115 (30.83%) 162 (1.21%) 1 (0.01%) 0 (0.00%) 15 (0.11%) German 39573 10927 (27.61%) 10908 (27.56%) 671 (1.70%) 0 (0.00%) 0 (0.00%) 419 (1.06%) Portuguese 9071 1718 (18.94%) 1713 (18.88%) 8 (0.09%) 0 (0.00%) 0 (0.00%) 7 (0.08%) Swedish 6159 293 ( 4.76%) 280 ( 4.55%) 5 (0.08%) 0 (0.00%) 0 (0.00%) 14 (0.23%) Turkish 5510 657 (11.92%) 657 (11.92%) 10 (0.18%) 0 (0.00%) 0 (0.00%) 20 (0.36%) Table 1: Proportion of dependency trees classified by projectivity, planarity, m-planarity and illnestedness in treebanks for Arabic (Hajiˇc et al., 2004), Czech (Hajiˇc et al., 2006), Danish (Kromann, 2003), Dutch (van der Beek et al., 2002), German (Brants et al., 2002), Portuguese (Afonso et al., 2002), Swedish (Nilsson et al., 2005) and Turkish (Oflazer et al., 2003; Atalay et al., 2003). exists. Since this process can be completed by visiting each node and edge of the graph U once, its complexity is O(V + E). The crossings graph of a dependency graph with n nodes can trivially be built in time O(n2) by checking each pair of dependency links to determine if they cross, and cannot contain more than n2 edges, which means that we can check if the dependency graph for a sentence of length n is 2-planar in O(n2) time. For k > 2, the k-colouring problem is known to be NP-complete (Karp, 1972). However, we have found this not to be a problem when measuring multiplanarity in natural language treebanks, since the effective problem size can be reduced by noting that each connected component of the crossings graph can be treated separately, and that nodes that are not part of a cycle need not be considered.5 Given that non-projective sentences in natural language tend to have a small proportion of non-projective links (Nivre and Nilsson, 2005), the connected components of their crossings graphs are very small, and k-colourings for them can quickly be found by brute-force search. By applying these techniques to dependency treebanks of several languages, we obtain the data shown in Table 1. As we can see, the coverage provided by the 2-planarity constraint is comparable to that of well-nestedness. In most of the treebanks, well over 99% of the sentences are 2planar, and 3-planarity has almost total coverage. As we will see below, the class of 2-planar dependency structures not only has good coverage of linguistic phenomena in existing treebanks but is also efficiently parsable with transition-based parsing methods, making it a practically interesting subclass of non-projective dependency structures. 5If we have a valid colouring for all the cycles in the graph, the rest of the nodes can be safely coloured by breadthfirst search as in the k = 2 case. 4 Parsing 1-Planar Structures In this section, we present a deterministic lineartime parser for planar dependency structures. The parser is a variant of Nivre’s arc-eager projective parser (Nivre, 2003), modified so that it can also handle graphs that are planar but not projective. As seen in Table 1, this only gives a modest improvement in coverage compared to projective parsing, so the main interest of this algorithm lies in the fact that it can be generalised to deal with 2-planar structures, as shown in the next section. 4.1 Transition Systems In the transition-based framework of Nivre (2008), a deterministic dependency parser is defined by a non-deterministic transition system, specifying a set of elementary operations that can be executed during the parsing process, and an oracle that deterministically selects a single transition at each choice point of the parsing process. Definition 6. A transition system for dependency parsing is a quadruple S = (C, T, cs, Ct) where 1. C is a set of possible parser configurations, 2. T is a set of transitions, each of which is a partial function t : C →C, 3. cs is a function that maps each input sentence w to an initial configuration cs(w) ∈C, 4. Ct ⊆C is a set of terminal configurations. Definition 7. An oracle for a transition system S = (C, T, cs, Ct) is a function o : C →T. An input sentence w can be parsed using a transition system S = (C, T, cs, Ct) and an oracle o by starting in the initial configuration cs(w), calling the oracle function on the current configuration c, and updating the configuration by applying the transition o(c) returned by the oracle. This process is repeated until a terminal configuration is 1495 Initial configuration: cs(w1 . . . wn) = ⟨[], [w1 . . . wn], ∅⟩ Terminal configurations: Cf = {⟨Σ, [], A⟩∈C} Transitions: SHIFT ⟨Σ, wi|B, A⟩⇒⟨Σ|wi, B, A⟩ REDUCE ⟨Σ|wi, B, A⟩⇒⟨Σ, B, A⟩ LEFT-ARC ⟨Σ|wi, wj|B, A⟩⇒⟨Σ|wi, wj|B, A ∪{(wj, wi)}⟩ only if ̸ ∃k|(wk, wi) ∈A (single-head) and not wi ↔∗wj ∈A (acyclicity). RIGHT-ARC ⟨Σ|wi, wj|B, A⟩⇒⟨Σ|wi, wj|B, A ∪{(wi, wj)}⟩ only if ̸ ∃k|(wk, wj) ∈A (single-head) and not wi ↔∗wj ∈A (acyclicity). Figure 3: Transition system for planar dependency parsing. reached, and the dependency analysis of the sentence is defined by the terminal configuration. Each sequence of configurations that the parser can traverse from an initial configuration to a terminal configuration for some input w is called a transition sequence. If we associate each configuration c of a transition system S = (C, T, cs, Ct) with a dependency graph g(c), we can say that S is sound for a class of dependency graphs G if, for every sentence w and transition sequence (cs(w), c1, . . . , cf) of S, g(cf) is in G, and that S is complete for G if, for every sentence w and dependency graph G ∈G for w, there is a transition sequence (cs(w), c1, . . . , cf) such that g(cf) = G. A transition system that is sound and complete for G is said to be correct for G. Note that, apart from a correct transition system, a practical parser needs a good oracle to achieve the desired results, since a transition system only specifies how to reach all the possible dependency graphs that could be associated to a sentence, but not how to select the correct one. Oracles for practical parsers can be obtained by training classifiers on treebank data (Nivre et al., 2004). 4.2 A Transition System for Planar Structures A correct transition system for the class of planar dependency forests can be obtained as a variant of the arc-eager projective system by Nivre (2003). As in that system, the set of configurations of the planar transition system is the set of all triples c = ⟨Σ, B, A⟩such that Σ and B are disjoint lists of words from Vw (for some input w), and A is a set of dependency links over Vw. The list B, called the buffer, is initialised to the input string and is used to hold the words that are still to be read from the input. The list Σ, called the stack, is initially empty and holds words that have dependency links pending to be created. The system is shown in Figure 3, where we use the notation Σ|wi for a stack with top wi and tail Σ, and we invert the notation for the buffer for clarity (i.e., wi|B is a buffer with top wi and tail B). The system reads the input from left to right and creates links in a left-to-right order by executing its four transitions: 1. SHIFT: pops the first (leftmost) word in the buffer, and pushes it to the stack. 2. LEFT-ARC: adds a link from the first word in the buffer to the top of the stack. 3. RIGHT-ARC: adds a link from the top of the stack to the first word in the buffer. 4. REDUCE: pops the top word from the stack, implying that we have finished building links to or from it. Note that the planar parser’s transitions are more fine-grained than those of the arc-eager projective parser by Nivre (2003), which pops the stack as part of its LEFT-ARC transition and shifts a word as part of its RIGHT-ARC transition. Forcing these actions after creating dependency links rules out structures whose root is covered by a dependency link, which are planar but not projective. In order to support these structures, we therefore simplify the ARC transitions (LEFT-ARC and RIGHT-ARC) so that they only create an arc. For the same reason, we remove the constraint in Nivre’s parser by which words without a head cannot be reduced. This has the side effect of making the parser able to output cyclic graphs. Since we are interested in planar dependency forests, which do not contain cycles, we only apply ARC transitions after checking that there is no undirected path between the nodes to be linked. This check can be done without affecting the linear-time complexity of the 1496 parser by storing the weakly connected component of each node in g(c). The fine-grained transitions used by this parser have also been used by Sagae and Tsujii (2008) to parse DAGs. However, the latter parser differs from ours in the constraints, since it does not allow the reduction of words without a head (disallowing forests with covered roots) and does not enforce the acyclicity constraint (which is guaranteed by post-processing the graphs to break cycles). 4.3 Correctness and Complexity For reasons of space, we can only give a sketch of the correctness proof. We wish to prove that the planar transition system is sound and complete for the set Fp of all planar dependency forests. To prove soundness, we have to show that, for every sentence w and transition sequence (cs(w), c1, . . . , cf), the graph g(cf) associated with cf is in Fp. We take the graph associated with a configuration c = (Σ, B, A) to be g(c) = (Vw, A). With this, we prove the stronger claim that g(c) ∈Fp for every configuration c that belongs to some transition sequence starting with cs(w). This amounts to showing that in every configuration c reachable from cs(w), g(c) meets the following three conditions that characterise a planar dependency forest: (1) g(c) does not contain nodes with more than one head; (2) g(c) is acyclic; and (3) g(c) contains no crossing links. (1) is trivially guaranteed by the single-head constraint; (2) follows from (1) and the acyclicity constraint; and (3) can be established by proving that there is no transition sequence that will invoke two ARC transitions on node pairs that would create crossing links. At the point when a link from wi to wj is created, we know that all the words strictly located between wi and wj are not in the stack or in the buffer, so no links can be created to or from them. To prove completeness, we show that every planar dependency forest G = (V, E) ∈Fp for a sentence w can be produced by applying the oracle function that maps a configuration ⟨Σ|wi, wj|B, A⟩to: 1. LEFT-ARC if wj →wi ∈(E \ A), 2. RIGHT-ARC if wi →wj ∈(E \ A), 3. REDUCE if ∃x[x<i][wx ↔wj ∈(E \ A)], 4. SHIFT otherwise. We show completeness by setting the following invariants on transitions traversed by the application of the oracle: 1. ∀a, b[a,b<j][wa ↔wb ∈E ⇒wa ↔wb ∈A] 2. [wi ↔wj ∈A ⇒ ∀k[i<k<j][wk ↔wj ∈E ⇒wk ↔wj ∈A]] 3. ∀k[k<j][wk ̸∈Σ ⇒ ∀l[l>k][wk ↔wl ∈E ⇒wk ↔wl ∈A]] We can show that each branch of the oracle function keeps these invariants true. When we reach a terminal configuration (which always happens after a finite number of transitions, since every transition generating a configuration c = ⟨Σ, B, A⟩ decreases the value of the variant function |E| + |Σ| + 2|B| −|A|), it can be deduced from the invariant that A = E, which proves completeness. The worst-case complexity of a deterministic transition-based parser is given by an upper bound on transition sequence length (Nivre, 2008). For the planar system, like its projective counterpart, the length is clearly O(n) (where n is the number of input words), since there can be no more than n SHIFT transitions, n REDUCE transitions, and n ARC transitions in a transition sequence. 5 Parsing 2-Planar Structures The planar parser introduced in the previous section can be extended to parse all 2-planar dependency structures by adding a second stack to the system and making REDUCE and ARC transitions apply to only one of the stacks at a time. This means that the set of links created in the context of each individual stack will be planar, but pairs of links created in different stacks are allowed to cross. In this way, the parser will build a 2-planar dependency forest by using each of the stacks to construct one of its two planes. The 2-planar transition system, shown in Figure 4, has configurations of the form ⟨Σ0, Σ1, B, A⟩, where we call Σ0 the active stack and Σ1 the inactive stack, and the following transitions: 1. SHIFT: pops the first (leftmost) word in the buffer, and pushes it to both stacks. 2. LEFT-ARC: adds a link from the first word in the buffer to the top of the active stack. 3. RIGHT-ARC: adds a link from the top of the active stack to the first word in the buffer. 4. REDUCE: pops the top word from the active stack, implying that we have added all links to or from it on the plane tied to that stack. 5. SWITCH: makes the active stack inactive and vice versa, changing the plane the parser is working with. 1497 Initial configuration: cs(w1 . . . wn) = ⟨[], [], [w1 . . . wn], ∅⟩ Terminal configurations: Cf = {⟨Σ0, Σ1, [], A⟩∈C} Transitions: SHIFT ⟨Σ0, Σ1, wi|B, A⟩⇒⟨Σ0|wi, Σ1|wi, B, A⟩ REDUCE ⟨Σ0|wi, Σ1, B, A⟩⇒⟨Σ0, Σ1, B, A⟩ LEFT-ARC ⟨Σ0|wi, Σ1, wj|B, A⟩⇒⟨Σ0|wi, Σ1, wj|B, A ∪{(wj, wi)}⟩ only if ̸ ∃k | (wk, wi) ∈A (single-head) and not wi ↔∗wj ∈A (acyclicity). RIGHT-ARC ⟨Σ0|wi, Σ1, wj|B, A⟩⇒⟨Σ0|wi, Σ1, wj|B, A ∪{(wi, wj)}⟩ only if ̸ ∃k|(wk, wj) ∈A (single-head) and not wi ↔∗wj ∈A (acyclicity). SWITCH ⟨Σ0, Σ1, B, A⟩⇒⟨Σ1, Σ0, B, A⟩ Figure 4: Transition system for 2-planar dependency parsing. 5.1 Correctness and Complexity As in the planar case, we provide a brief sketch of the proof that the transition system in Figure 4 is correct for the set F2p of 2-planar dependency forests. Soundness follows from a reasoning analogous to the planar case, but applying the proof of planarity separately to each stack. In this way, we prove that the sets of dependency links created by linking to or from the top of each of the two stacks are always planar graphs, and thus their union (which is the dependency graph stored in A) is 2-planar. This, together with the single-head and acyclicity constraints, guarantees that the dependency graphs associated with reachable configurations are always 2-planar dependency forests. For completeness, we assume an extended form of the transition system where transitions take the form ⟨Σ0, Σ1, B, A, p⟩, where p is a flag taking values in {0, 1} which equals 0 for initial configurations and gets flipped by each application of a SWITCH transition. Then we show that every 2planar dependency forest G ∈F2p, with planes G0 = (V, E0) and G1 = (V, E1), can be produced by this system by applying the oracle function that maps a configuration ⟨Σ0|wi, Σ1, wj|B, A, p⟩to: 1. LEFT-ARC if wj →wi ∈(Ep \ A), 2. RIGHT-ARC if wi →wj ∈(Ep \ A), 3. REDUCE if ∃x[x<i][wx ↔wj ∈(Ep \ A) ∧ ¬∃y[x<y≤i][wy ↔wj ∈(Ep \ A)]], 4. SWITCH if ∃x<j : (wx, wj) or (wj, wx) ∈(Ep\A), 5. SHIFT otherwise. This can be shown by employing invariants analogous to the planar case, with the difference that the third invariant applies to each stack and its corresponding plane: if Σy is associated with the plane Ex,6 we have: 3. ∀k[k<j][wk ̸∈Σy] ⇒ ∀l[l>k][wk ↔wl ∈Ex] ⇒[wk ↔wl ∈A] Since the presence of the flag p in configurations does not affect the set of dependency graphs generated by the system, the completeness of the system extended with the flag p implies that of the system in Figure 4. We can show that the complexity of the 2-planar system is O(n) by the same kind of reasoning as for the 1-planar system, with the added complication that we must constrain the system to prevent two adjacent SWITCH transitions. In fact, without this restriction, the parser is not even guaranteed to terminate. 5.2 Implementation In practical settings, oracles for transition-based parsers can be approximated by classifiers trained on treebank data (Nivre, 2008). To do this, we need an oracle that will generate transition sequences for gold-standard dependency graphs. In the case of the planar parser of Section 4.2, the oracle of 4.3 is suitable for this purpose. However, in the case of the 2-planar parser, the oracle used for the completeness proof in Section 5.1 cannot be used directly, since it requires the gold-standard trees to be divided into two planes in order to generate a transition sequence. Of course, it is possible to use the algorithm presented in Section 3 to obtain a division of sentences into planes. However, for training purposes and to obtain a robust behaviour if non-2-planar 6The plane corresponding to each stack in a configuration changes with each SWITCH transition: Σx is associated with Ex in configurations where p = 0, and with Ex in those where p = 1. 1498 Czech Danish German Portuguese Parser LAS UAS NPP NPR LAS UAS NPP NPR LAS UAS NPP NPR LAS UAS NPP NPR 2-planar 79.24 85.30 68.9 60.7 83.81 88.50 66.7 20.0 86.50 88.84 57.1 45.8 87.04 90.82 82.8 33.8 Malt P 78.18 84.12 – – 83.31 88.30 – – 85.36 88.06 – – 86.60 90.20 – – Malt PP 79.80 85.70 76.7 56.1 83.67 88.52 41.7 25.0 85.76 88.66 58.1 40.7 87.08 90.66 83.3 46.2 Table 2: Parsing accuracy for 2-planar parser in comparison to MaltParser with (PP) and without (P) pseudo-projective transformations. LAS = labeled attachment score; UAS = unlabeled attachment score; NPP = precision on non-projective arcs; NPR = recall on non-projective arcs. sentences are found, it is more convenient that the oracle can distribute dependency links into the planes incrementally, and that it produces a distribution of links that only uses SWITCH transitions when it is strictly needed to account for nonplanarity. Thus we use a more complex version of the oracle which performs a search in the crossings graph to check if a dependency link can be built on the plane of the active stack, and only performs a switch when this is not possible. This has proved to work well in practice, as will be observed in the results in the next section. 6 Empirical Evaluation In order to get a first estimate of the empirical accuracy that can be obtained with transition-based 2-planar parsing, we have evaluated the parser on four data sets from the CoNLL-X shared task (Buchholz and Marsi, 2006): Czech, Danish, German and Portuguese. As our baseline, we take the strictly projective arc-eager transition system proposed by Nivre (2003), as implemented in the freely available MaltParser system (Nivre et al., 2006a), with and without the pseudo-projective parsing technique for recovering non-projective dependencies (Nivre and Nilsson, 2005). For the two baseline systems, we use the parameter settings used by Nivre et al. (2006b) in the original shared task, where the pseudo-projective version of MaltParser was one of the two top performing systems (Buchholz and Marsi, 2006). For our 2planar parser, we use the same kernelized SVM classifiers as MaltParser, using the LIBSVM package (Chang and Lin, 2001), with feature models that are similar to MaltParser but extended with features defined over the second stack.7 In Table 2, we report labeled (LAS) and unlabeled (UAS) attachment score on the four languages for all three systems. For the two systems that are capable of recovering non-projective de7Complete information about experimental settings can be found at http://stp.lingfil.uu.se/ nivre/exp/. pendencies, we also report precision (NPP) and recall (NPR) specifically on non-projective dependency arcs. The results show that the 2-planar parser outperforms the strictly projective variant of MaltParser on all metrics for all languages, and that it performs on a par with the pseudoprojective variant with respect to both overall attachment score and precision and recall on nonprojective dependencies. These results look very promising in view of the fact that very little effort has been spent on optimizing the training oracle and feature model for the 2-planar parser so far. It is worth mentioning that the 2-planar parser has two advantages over the pseudo-projective parser. The first is simplicity, given that it is based on a single transition system and makes a single pass over the input, whereas the pseudo-projective parsing technique involves preprocessing of training data and post-processing of parser output (Nivre and Nilsson, 2005). The second is the fact that it parses a well-defined class of dependency structures, with known coverage8, whereas no formal characterization exists of the class of structures parsable by the pseudo-projective parser. 7 Conclusion In this paper, we have presented an efficient algorithm for deciding whether a dependency graph is 2-planar and a transition-based parsing algorithm that is provably correct for 2-planar dependency forests, neither of which existed in the literature before. In addition, we have presented empirical results showing that the class of 2-planar dependency forests includes the overwhelming majority of structures found in existing treebanks and that a deterministic classifier-based implementation of the 2-planar parser gives state-of-the-art accuracy on four different languages. 8If more coverage is desired, the 2-planar parser can be generalised to m-planar structures for larger values of m by adding additional stacks. However, this comes at the cost of more complex training models, making the practical interest of increasing m beyond 2 dubious. 1499 Acknowledgments The first author has been partially supported by Ministerio de Educaci´on y Ciencia and FEDER (HUM2007-66607-C04) and Xunta de Galicia (PGIDIT07SIN005206PR, Rede Galega de Procesamento da Linguaxe e Recuperaci´on de Informaci´on, Rede Galega de Ling¨u´ıstica de Corpus, Bolsas Estad´ıas INCITE/FSE cofinanced). References Susana Afonso, Eckhard Bick, Renato Haber, and Diana Santos. 2002. “Floresta sint´a(c)tica”: a treebank for Portuguese. In Proceedings of the 3rd International Conference on Language Resources and Evaluation (LREC 2002), pages 1968–1703, Paris, France. ELRA. Nart B. Atalay, Kemal Oflazer, and Bilge Say. 2003. The annotation process in the Turkish treebank. In Proceedings of EACL Workshop on Linguistically Interpreted Corpora (LINC-03), pages 243– 246, Morristown, NJ, USA. Association for Computational Linguistics. Leonoor van der Beek, Gosse Bouma, Robert Malouf, and Gertjan van Noord. 2002. The Alpino dependency treebank. In Language and Computers, Computational Linguistics in the Netherlands 2001. Selected Papers from the Twelfth CLIN Meeting, pages 8–22, Amsterdam, the Netherlands. Rodopi. Manuel Bodirsky, Marco Kuhlmann, and Mathias M¨ohl. 2005. Well-nested drawings as models of syntactic structure. In 10th Conference on Formal Grammar and 9th Meeting on Mathematics of Language, Edinburgh, Scotland, UK. Sabine Brants, Stefanie Dipper, Silvia Hansen, Wolfgang Lezius, and George Smith. 2002. The tiger treebank. In Proceedings of the Workshop on Treebanks and Linguistic Theories, September 20-21, Sozopol, Bulgaria. Matthias Buch-Kromann. 2006. Discontinuous Grammar: A Model of Human Parsing and Language Acquisition. Ph.D. thesis, Copenhagen Business School. Sabine Buchholz and Erwin Marsi. 2006. CoNLLX shared task on multilingual dependency parsing. In Proceedings of the 10th Conference on Computational Natural Language Learning (CoNLL), pages 149–164. Chih-Chung Chang and Chih-Jen Lin, 2001. LIBSVM: A Library for Support Vector Machines. Software available at http://www.csie.ntu.edu.tw/∼cjlin/libsvm. Jason Eisner. 1996. Three new probabilistic models for dependency parsing: An exploration. In Proceedings of the 16th International Conference on Computational Linguistics (COLING-96), pages 340–345, San Francisco, CA, USA, August. ACL / Morgan Kaufmann. Haim Gaifman. 1965. Dependency systems and phrase-structure systems. Information and Control, 8:304–337. Carlos G´omez-Rodr´ıguez, John Carroll, and David Weir. 2008. A deductive approach to dependency parsing. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (ACL’08:HLT), pages 968–976, Morristown, NJ, USA. Association for Computational Linguistics. Carlos G´omez-Rodr´ıguez, David Weir, and John Carroll. 2009. Parsing mildly non-projective dependency structures. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 291– 299. Jan Hajiˇc, Otakar Smrˇz, Petr Zem´anek, Jan ˇSnaidauf, and Emanuel Beˇska. 2004. Prague Arabic dependency treebank: Development in data and tools. In Proceedings of the NEMLAR International Conference on Arabic Language Resources and Tools, pages 110–117. Jan Hajiˇc, Jarmila Panevov´a, Eva Hajiˇcov´a, Jarmila Panevov´a, Petr Sgall, Petr Pajas, Jan ˇStˇep´anek, Jiˇr´ı Havelka, and Marie Mikulov´a. 2006. Prague Dependency Treebank 2.0. CDROM CAT: LDC2006T01, ISBN 1-58563-370-4. Linguistic Data Consortium. Jiri Havelka. 2007. Beyond projectivity: Multilingual evaluation of constraints and measures on nonprojective structures. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 608–615. Richard M. Karp. 1972. Reducibility among combinatorial problems. In R. Miller and J. Thatcher, editors, Complexity of Computer Computations, pages 85–103. Plenum Press. Matthias T. Kromann. 2003. The Danish dependency treebank and the underlying linguistic theory. In Proceedings of the 2nd Workshop on Treebanks and Linguistic Theories (TLT), pages 217–220, V¨axj¨o, Sweden. V¨axj¨o University Press. Marco Kuhlmann and Mathias M¨ohl. 2007. Mildly context-sensitive dependency languages. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 160–167. Marco Kuhlmann and Joakim Nivre. 2006. Mildly non-projective dependency structures. In Proceedings of the COLING/ACL 2006 Main Conference Poster Sessions, pages 507–514. 1500 Marco Kuhlmann and Giorgio Satta. 2009. Treebank grammar techniques for non-projective dependency parsing. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 478–486. Marco Kuhlmann. 2007. Dependency Structures and Lexicalized Grammars. Doctoral dissertation, Saarland University, Saarbr¨ucken, Germany. Andre Martins, Noah Smith, and Eric Xing. 2009. Concise integer linear programming formulations for dependency parsing. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP (ACLIJCNLP), pages 342–350. Ryan McDonald and Giorgio Satta. 2007. On the complexity of non-projective data-driven dependency parsing. In Proceedings of the 10th International Conference on Parsing Technologies (IWPT), pages 122–131. Ryan McDonald, Koby Crammer, and Fernando Pereira. 2005a. Online large-margin training of dependency parsers. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL), pages 91–98. Ryan McDonald, Fernando Pereira, Kiril Ribarov, and Jan Hajiˇc. 2005b. Non-projective dependency parsing using spanning tree algorithms. In HLT/EMNLP 2005: Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, pages 523–530, Morristown, NJ, USA. Association for Computational Linguistics. Peter Neuhaus and Norbert Br¨oker. 1997. The complexity of recognition of linguistically adequate dependency grammars. In Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics (ACL) and the 8th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 337–343. Jens Nilsson, Johan Hall, and Joakim Nivre. 2005. MAMBA meets TIGER: Reconstructing a Swedish treebank from antiquity. In Proceedings of NODALIDA 2005 Special Session on Treebanks, pages 119– 132. Samfundslitteratur, Frederiksberg, Denmark, May. Joakim Nivre and Jens Nilsson. 2005. Pseudoprojective dependency parsing. In ACL ’05: Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, pages 99–106, Morristown, NJ, USA. Association for Computational Linguistics. Joakim Nivre, Johan Hall, and Jens Nilsson. 2004. Memory-based dependency parsing. In Proceedings of the 8th Conference on Computational Natural Language Learning (CoNLL-2004), pages 49– 56, Morristown, NJ, USA. Association for Computational Linguistics. Joakim Nivre, Johan Hall, and Jens Nilsson. 2006a. MaltParser: A data-driven parser-generator for dependency parsing. In Proceedings of the 5th International Conference on Language Resources and Evaluation (LREC), pages 2216–2219. Joakim Nivre, Johan Hall, Jens Nilsson, G¨ulsen Eryi˘git, and Svetoslav Marinov. 2006b. Labeled pseudo-projective dependency parsing with support vector machines. In Proceedings of the 10th Conference on Computational Natural Language Learning (CoNLL), pages 221–225. Joakim Nivre. 2003. An efficient algorithm for projective dependency parsing. In Proceedings of the 8th International Workshop on Parsing Technologies (IWPT), pages 149–160. Joakim Nivre. 2006. Constraints on non-projective dependency graphs. In Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 73– 80. Joakim Nivre. 2008. Algorithms for Deterministic Incremental Dependency Parsing. Computational Linguistics, 34(4):513–553. Kemal Oflazer, Bilge Say, Dilek Zeynep Hakkani-T¨ur, and G¨okhan T¨ur. 2003. Building a Turkish treebank. In A. Abeille (ed.), Building and Exploiting Syntactically-annotated Corpora, pages 261–277, Dordrecht, the Netherlands. Kluwer. Kenji Sagae and Jun’ichi Tsujii. 2008. Shift-reduce dependency DAG parsing. In COLING ’08: Proceedings of the 22nd International Conference on Computational Linguistics, pages 753–760, Morristown, NJ, USA. Association for Computational Linguistics. Daniel Sleator and Davy Temperley. 1993. Parsing English with a Link Grammar. In Proceedings of the Third International Workshop on Parsing Technologies (IWPT’93), pages 277–292. ACL/SIGPARSE. Ivan Titov and James Henderson. 2007. A latent variable model for generative dependency parsing. In Proceedings of the 10th International Conference on Parsing Technologies (IWPT), pages 144–155. Hiroyasu Yamada and Yuji Matsumoto. 2003. Statistical dependency analysis with support vector machines. In Proceedings of the 8th International Workshop on Parsing Technologies (IWPT), pages 195–206. Anssi Mikael Yli-Jyr¨a. 2003. Multiplanarity – a model for dependency structures in treebanks. In Joakim Nivre and Erhard Hinrichs, editors, TLT 2003. Proceedings of the Second Workshop on Treebanks and Linguistic Theories, volume 9 of Mathematical Modelling in Physics, Engineering and Cognitive Sciences, pages 189–200, V¨axj¨o, Sweden, 1415 November. V¨axj¨o University Press. 1501
|
2010
|
151
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1502–1511, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Viterbi Training for PCFGs: Hardness Results and Competitiveness of Uniform Initialization Shay B. Cohen and Noah A. Smith School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA {scohen,nasmith}@cs.cmu.edu Abstract We consider the search for a maximum likelihood assignment of hidden derivations and grammar weights for a probabilistic context-free grammar, the problem approximately solved by “Viterbi training.” We show that solving and even approximating Viterbi training for PCFGs is NP-hard. We motivate the use of uniformat-random initialization for Viterbi EM as an optimal initializer in absence of further information about the correct model parameters, providing an approximate bound on the log-likelihood. 1 Introduction Probabilistic context-free grammars are an essential ingredient in many natural language processing models (Charniak, 1997; Collins, 2003; Johnson et al., 2006; Cohen and Smith, 2009, inter alia). Various algorithms for training such models have been proposed, including unsupervised methods. Many of these are based on the expectationmaximization (EM) algorithm. There are alternatives to EM, and one such alternative is Viterbi EM, also called “hard” EM or “sparse” EM (Neal and Hinton, 1998). Instead of using the parameters (which are maintained in the algorithm’s current state) to find the true posterior over the derivations, Viterbi EM algorithm uses a posterior focused on the Viterbi parse of those parameters. Viterbi EM and variants have been used in various settings in natural language processing (Yejin and Cardie, 2007; Wang et al., 2007; Goldwater and Johnson, 2005; DeNero and Klein, 2008; Spitkovsky et al., 2010). Viterbi EM can be understood as a coordinate ascent procedure that locally optimizes a function; we call this optimization goal “Viterbi training.” In this paper, we explore Viterbi training for probabilistic context-free grammars. We first show that under the assumption that P ̸= NP, solving and even approximating the Viterbi training problem is hard. This result holds even for hidden Markov models. We extend the main hardness result to the EM algorithm (giving an alternative proof to this known result), as well as the problem of conditional Viterbi training. We then describe a “competitiveness” result for uniform initialization of Viterbi EM: we show that initialization of the trees in an E-step which uses uniform distributions over the trees is optimal with respect to a certain approximate bound. The rest of this paper is organized as follows. §2 gives background on PCFGs and introduces some notation. §3 explains Viterbi training, the declarative form of Viterbi EM. §4 describes a hardness result for Viterbi training. §5 extends this result to a hardness result of approximation and §6 further extends these results for other cases. §7 describes the advantages in using uniform-at-random initialization for Viterbi training. We relate these results to work on the k-means problem in §8. 2 Background and Notation We assume familiarity with probabilistic contextfree grammars (PCFGs). A PCFG G consists of: • A finite set of nonterminal symbols N; • A finite set of terminal symbols Σ; • For each A ∈N, a set of rewrite rules R(A) of the form A →α, where α ∈(N ∪Σ)∗, and R = ∪A∈NR(A); • For each rule A →α, a probability θA→α. The collection of probabilities is denoted θ, and they are constrained such that: ∀(A →α) ∈R(A), θA→α ≥ 0 ∀A ∈N, X α:(A→α)∈R(A) θA→α = 1 That is, θ is grouped into |N| multinomial distributions. 1502 Under the PCFG, the joint probability of a string x ∈Σ∗and a grammatical derivation z is1 p(x, z | θ) = Y (A→α)∈R (θA→α)fA→α(z) (1) = exp X (A→α)∈R fA→α(z) log θA→α where fA→α(z) is a function that “counts” the number of times the rule A →α appears in the derivation z. fA(z) will similarly denote the number of times that nonterminal A appears in z. Given a sample of derivations z = ⟨z1, . . . , zn⟩, let: FA→α(z) = n X i=1 fA→α(zi) (2) FA(z) = n X i=1 fA(zi) (3) We use the following notation for G: • L(G) is the set of all strings (sentences) x that can be generated using the grammar G (the “language of G”). • D(G) is the set of all possible derivations z that can be generated using the grammar G. • D(G, x) is the set of all possible derivations z that can be generated using the grammar G and have the yield x. 3 Viterbi Training Viterbi EM, or “hard” EM, is an unsupervised learning algorithm, used in NLP in various settings (Yejin and Cardie, 2007; Wang et al., 2007; Goldwater and Johnson, 2005; DeNero and Klein, 2008; Spitkovsky et al., 2010). In the context of PCFGs, it aims to select parameters θ and phrasestructure trees z jointly. It does so by iteratively updating a state consisting of (θ, z). The state is initialized with some value, then the algorithm alternates between (i) a “hard” E-step, where the strings x1, . . . , xn are parsed according to a current, fixed θ, giving new values for z, and (ii) an M-step, where the θ are selected to maximize likelihood, with z fixed. With PCFGs, the E-step requires running an algorithm such as (probabilistic) CKY or Earley’s 1Note that x = yield(z); if the derivation is known, the string is also known. On the other hand, there may be many derivations with the same yield, perhaps even infinitely many. algorithm, while the M-step normalizes frequency counts FA→α(z) to obtain the maximum likelihood estimate’s closed-form solution. We can understand Viterbi EM as a coordinate ascent procedure that approximates the solution to the following declarative problem: Problem 1. ViterbiTrain Input: G context-free grammar, x1, . . . , xn training instances from L(G) Output: θ and z1, . . . , zn such that (θ, z1, . . . , zn) = argmax θ,z n Y i=1 p(xi, zi | θ) (4) The optimization problem in Eq. 4 is nonconvex and, as we will show in §4, hard to optimize. Therefore it is necessary to resort to approximate algorithms like Viterbi EM. Neal and Hinton (1998) use the term “sparse EM” to refer to a version of the EM algorithm where the E-step finds the modes of hidden variables (rather than marginals as in standard EM). Viterbi EM is a variant of this, where the Estep finds the mode for each xi’s derivation, argmaxz∈D(G,xi) p(xi, z | θ). We will refer to L(θ, z) = n Y i=1 p(xi, zi | θ) (5) as “the objective function of ViterbiTrain.” Viterbi training and Viterbi EM are closely related to self-training, an important concept in semi-supervised NLP (Charniak, 1997; McClosky et al., 2006a; McClosky et al., 2006b). With selftraining, the model is learned with some seed annotated data, and then iterates by labeling new, unannotated data and adding it to the original annotated training set. McClosky et al. consider selftraining to be “one round of Viterbi EM” with supervised initialization using labeled seed data. We refer the reader to Abney (2007) for more details. 4 Hardness of Viterbi Training We now describe hardness results for Problem 1. We first note that the following problem is known to be NP-hard, and in fact, NP-complete (Sipser, 2006): Problem 2. 3-SAT Input: A formula φ = Vm i=1 (ai ∨bi ∨ci) in conjunctive normal form, such that each clause has 3 1503 Sφ2 ccccccccccccccccccccccccccccccc T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T T Sφ1 A1 eeeeeeeeeeeeeeeeeee Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y A2 eeeeeeeeeeeeeeeeeee Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y UY1,0 qqqqqqq M M M M M M M UY2,1 qqqqqqq M M M M M M M UY4,0 qqqqqqq M M M M M M M UY1,0 qqqqqqq M M M M M M M UY2,1 qqqqqqq M M M M M M M UY3,1 qqqqqqq M M M M M M M V ¯ Y1 VY1 VY2 V ¯ Y2 V ¯ Y4 VY4 V ¯ Y1 VY1 VY2 V ¯ Y2 VY3 V ¯ Y3 1 0 1 0 1 0 1 0 1 0 1 0 Figure 1: An example of a Viterbi parse tree which represents a satisfying assignment for φ = (Y1 ∨Y2 ∨¯Y4)∧( ¯Y1 ∨¯Y2 ∨Y3). In θφ, all rules appearing in the parse tree have probability 1. The extracted assignment would be Y1 = 0, Y2 = 1, Y3 = 1, Y4 = 0. Note that there is no usage of two different rules for a single nonterminal. literals. Output: 1 if there is a satisfying assignment for φ and 0 otherwise. We now describe a reduction of 3-SAT to Problem 1. Given an instance of the 3-SAT problem, the reduction will, in polynomial time, create a grammar and a single string such that solving the ViterbiTrain problem for this grammar and string will yield a solution for the instance of the 3-SAT problem. Let φ = Vm i=1 (ai ∨bi ∨ci) be an instance of the 3-SAT problem, where ai, bi and ci are literals over the set of variables {Y1, . . . , YN} (a literal refers to a variable Yj or its negation, ¯Yj). Let Cj be the jth clause in φ, such that Cj = aj ∨bj ∨cj. We define the following context-free grammar Gφ and string to parse sφ: 1. The terminals of Gφ are the binary digits Σ = {0, 1}. 2. We create N nonterminals VYr, r ∈ {1, . . . , N} and rules VYr →0 and VYr →1. 3. We create N nonterminals V ¯ Yr, r ∈ {1, . . . , N} and rules V ¯ Yr →0 and V ¯ Yr →1. 4. We create UYr,1 →VYrV ¯ Yr and UYr,0 → V ¯ YrVYr. 5. We create the rule Sφ1 →A1. For each j ∈ {2, . . . , m}, we create a rule Sφj →Sφj−1Aj where Sφj is a new nonterminal indexed by φj ≜Vj i=1 Ci and Aj is also a new nonterminal indexed by j ∈{1, . . . , m}. 6. Let Cj = aj ∨bj ∨cj be clause j in φ. Let Y (aj) be the variable that aj mentions. Let (y1, y2, y3) be a satisfying assignment for Cj where yk ∈{0, 1} and is the value of Y (aj), Y (bj) and Y (cj) respectively for k ∈{1, 2, 3}. For each such clause-satisfying assignment, we add the rule: Aj →UY (aj),y1UY (bj),y2UY (cj),y3 (6) For each Aj, we would have at most 7 rules of that form, since one rule will be logically inconsistent with aj ∨bj ∨cj. 7. The grammar’s start symbol is Sφn. 8. The string to parse is sφ = (10)3m, i.e. 3m consecutive occurrences of the string 10. A parse of the string sφ using Gφ will be used to get an assignment by setting Yr = 0 if the rule VYr →0 or V ¯ Yr →1 are used in the derivation of the parse tree, and 1 otherwise. Notice that at this point we do not exclude “contradictions” coming from the parse tree, such as VY3 →0 used in the tree together with VY3 →1 or V ¯ Y3 →0. The following lemma gives a condition under which the assignment is consistent (so contradictions do not occur in the parse tree): Lemma 1. Let φ be an instance of the 3-SAT problem, and let Gφ be a probabilistic CFG based on the above grammar with weights θφ. If the (multiplicative) weight of the Viterbi parse of sφ is 1, then the assignment extracted from the parse tree is consistent. Proof. Since the probability of the Viterbi parse is 1, all rules of the form {VYr, V ¯ Yr} →{0, 1} which appear in the parse tree have probability 1 as well. There are two possible types of inconsistencies. We show that neither exists in the Viterbi parse: 1504 1. For any r, an appearance of both rules of the form VYr →0 and VYr →1 cannot occur because all rules that appear in the Viterbi parse tree have probability 1. 2. For any r, an appearance of rules of the form VYr →1 and V ¯ Yr →1 cannot occur, because whenever we have an appearance of the rule VYr →0, we have an adjacent appearance of the rule V ¯ Yr →1 (because we parse substrings of the form 10), and then again we use the fact that all rules in the parse tree have probability 1. The case of VYr →0 and V ¯ Yr →0 is handled analogously. Thus, both possible inconsistencies are ruled out, resulting in a consistent assignment. Figure 1 gives an example of an application of the reduction. Lemma 2. Define φ, Gφ as before. There exists θφ such that the Viterbi parse of sφ is 1 if and only if φ is satisfiable. Moreover, the satisfying assignment is the one extracted from the parse tree with weight 1 of sφ under θφ. Proof. (=⇒) Assume that there is a satisfying assignment. Each clause Cj = aj ∨bj ∨cj is satisfied using a tuple (y1, y2, y3) which assigns value for Y (aj), Y (bj) and Y (cj). This assignment corresponds the following rule Aj →UY (aj),y1UY (bj),y2UY (cj),y3 (7) Set its probability to 1, and set all other rules of Aj to 0. In addition, for each r, if Yr = y, set the probabilities of the rules VYr →y and V ¯ Yr →1−y to 1 and V ¯ Yr →y and VYr →1 −y to 0. The rest of the weights for Sφj →Sφj−1Aj are set to 1. This assignment of rule probabilities results in a Viterbi parse of weight 1. (⇐=) Assume that the Viterbi parse has probability 1. From Lemma 1, we know that we can extract a consistent assignment from the Viterbi parse. In addition, for each clause Cj we have a rule Aj →UY (aj),y1UY (bj),y2UY (cj),y3 (8) that is assigned probability 1, for some (y1, y2, y3). One can verify that (y1, y2, y3) are the values of the assignment for the corresponding variables in clause Cj, and that they satisfy this clause. This means that each clause is satisfied by the assignment we extracted. In order to show an NP-hardness result, we need to “convert” ViterbiTrain to a decision problem. The natural way to do it, following Lemmas 1 and 2, is to state the decision problem for ViterbiTrain as “given G and x1, . . . , xn and α ≥0, is the optimized value of the objective function L(θ, z) ≥α?” and use α = 1 together with Lemmas 1 and 2. (Naturally, an algorithm for solving ViterbiTrain can easily be used to solve its decision problem.) Theorem 3. The decision version of the ViterbiTrain problem is NP-hard. 5 Hardness of Approximation A natural path of exploration following the hardness result we showed is determining whether an approximation of ViterbiTrain is also hard. Perhaps there is an efficient approximation algorithm for ViterbiTrain we could use instead of coordinate ascent algorithms such as Viterbi EM. Recall that such algorithms’ main guarantee is identifying a local maximum; we know nothing about how far it will be from the global maximum. We next show that approximating the objective function of ViterbiTrain with a constant factor of ρ is hard for any ρ ∈(1 2, 1] (i.e., 1/2+ϵ approximation is hard for any ϵ ≤1/2). This means that, under the P ̸= NP assumption, there is no efficient algorithm that, given a grammar G and a sample of sentences x1, . . . , xn, returns θ′ and z′ such that: L(θ′, z′) ≥ ρ · max θ,z n Y i=1 p(xi, zi | θ) (9) We will continue to use the same reduction from §4. Let sφ be the string from that reduction, and let (θ, z) be the optimal solution for ViterbiTrain given Gφ and sφ. We first note that if p(sφ, z | θ) < 1 (implying that there is no satisfying assignment), then there must be a nonterminal which appears along with two different rules in z. This means that we have a nonterminal B ∈N with some rule B →α that appears k times, while the nonterminal appears in the parse r ≥ k + 1 times. Given the tree z, the θ that maximizes the objective function is the maximum likelihood estimate (MLE) for z (counting and normalizing the rules).2 We therefore know that the ViterbiTrain objective function, L(θ, z), is at 2Note that we can only make p(z | θ, x) greater by using θ to be the MLE for the derivation z. 1505 most k r k , because it includes a factor equal to fB→α(z) fB(z) fB→α(z) , where fB(z) is the number of times nonterminal B appears in z (hence fB(z) = r) and fB→α(z) is the number of times B →α appears in z (hence fB→α(z) = k). For any k ≥1, r ≥k + 1: k r k ≤ k k + 1 k ≤1 2 (10) This means that if the value of the objective function of ViterbiTrain is not 1 using the reduction from §4, then it is at most 1 2. If we had an efficient approximate algorithm with approximation coefficient ρ > 1 2 (Eq. 9 holds), then in order to solve 3-SAT for formula φ, we could run the algorithm on Gφ and sφ and check whether the assignment to (θ, z) that the algorithm returns satisfies φ or not, and return our response accordingly. If φ were satisfiable, then the true maximal value of L would be 1, and the approximation algorithm would return (θ, z) such that L(θ, z) ≥ ρ > 1 2. z would have to correspond to a satisfying assignment, and in fact p(z | θ) = 1, because in any other case, the probability of a derivation which does not represent a satisfying assignment is smaller than 1 2. If φ were not satisfiable, then the approximation algorithm would never return a (θ, z) that results in a satisfying assignment (because such a (θ, z) does not exist). The conclusion is that an efficient algorithm for approximating the objective function of ViterbiTrain (Eq. 4) within a factor of 1 2 + ϵ is unlikely to exist. If there were such an algorithm, we could use it to solve 3-SAT using the reduction from §4. 6 Extensions of the Hardness Result An alternative problem to Problem 1, a variant of Viterbi-training, is the following (see, for example, Klein and Manning, 2001): Problem 3. ConditionalViterbiTrain Input: G context-free grammar, x1, . . . , xn training instances from L(G) Output: θ and z1, . . . , zn such that (θ, z1, . . . , zn) = argmax θ,z n Y i=1 p(zi | θ, xi) (11) Here, instead of maximizing the likelihood, we maximize the conditional likelihood. Note that there is a hidden assumption in this problem definition, that xi can be parsed using the grammar G. Otherwise, the quantity p(zi | θ, xi) is not well-defined. We can extend ConditionalViterbiTrain to return ⊥in the case of not having a parse for one of the xi—this can be efficiently checked using a run of a cubic-time parser on each of the strings xi with the grammar G. An approximate technique for this problem is similar to Viterbi EM, only modifying the Mstep to maximize the conditional, rather than joint, likelihood. This new M-step will not have a closed form and may require auxiliary optimization techniques like gradient ascent. Our hardness result for ViterbiTrain applies to ConditionalViterbiTrain as well. The reason is that if p(z, sφ | θφ) = 1 for a φ with a satisfying assignment, then L(G) = {sφ} and D(G) = {z}. This implies that p(z | θφ, sφ) = 1. If φ is unsatisfiable, then for the optimal θ of ViterbiTrain we have z and z′ such that 0 < p(z, sφ | θφ) < 1 and 0 < p(z′, sφ | θφ) < 1, and therefore p(z | θφ, sφ) < 1, which means the conditional objective function will not obtain the value 1. (Note that there always exist some parameters θφ that generate sφ.) So, again, given an algorithm for ConditionalViterbiTrain, we can discern between a satisfiable formula and an unsatisfiable formula, using the reduction from §4 with the given algorithm, and identify whether the value of the objective function is 1 or strictly less than 1. We get the result that: Theorem 4. The decision problem of ConditionalViterbiTrain problem is NP-hard. where the decision problem of ConditionalViterbiTrain is defined analogously to the decision problem of ViterbiTrain. We can similarly show that finding the global maximum of the marginalized likelihood: max θ 1 n n X i=1 log X z p(xi, z | θ) (12) is NP-hard. The reasoning follows. Using the reduction from before, if φ is satisfiable, then Eq. 12 gets value 0. If φ is unsatisfiable, then we would still get value 0 only if L(G) = {sφ}. If Gφ generates a single derivation for (10)3m, then we actually do have a satisfying assignment from 1506 Lemma 1. Otherwise (more than a single derivation), the optimal θ would have to give fractional probabilities to rules of the form VYr →{0, 1} (or V ¯ Yr →{0, 1}). In that case, it is no longer true that (10)3m is the only generated sentence, which is a contradiction. The quantity in Eq. 12 can be maximized approximately using algorithms like EM, so this gives a hardness result for optimizing the objective function of EM for PCFGs. Day (1983) previously showed that maximizing the marginalized likelihood for hidden Markov models is NP-hard. We note that the grammar we use for all of our results is not recursive. Therefore, we can encode this grammar as a hidden Markov model, strengthening our result from PCFGs to HMMs.3 7 Uniform-at-Random Initialization In the previous sections, we showed that solving Viterbi training is hard, and therefore requires an approximation algorithm. Viterbi EM, which is an example of such algorithm, is dependent on an initialization of either θ to start with an E-step or z to start with an M-step. In the absence of a betterinformed initializer, it is reasonable to initialize z using a uniform distribution over D(G, xi) for each i. If D(G, xi) is finite, it can be done efficiently by setting θ = 1 (ignoring the normalization constraint), running the inside algorithm, and sampling from the (unnormalized) posterior given by the chart (Johnson et al., 2007). We turn next to an analysis of this initialization technique that suggests it is well-motivated. The sketch of our result is as follows: we first give an asymptotic upper bound for the loglikelihood of derivations and sentences. This bound, which has an information-theoretic interpretation, depends on a parameter λ, which depends on the distribution from which the derivations were chosen. We then show that this bound is minimized when we pick λ such that this distribution is (conditioned on the sentence) a uniform distribution over derivations. Let q(x) be any distribution over L(G) and θ some parameters for G. Let f(z) be some feature function (such as the one that counts the number of appearances of a certain rule in a derivation), and then: Eq,θ[f] ≜ X x∈L(G) q(x) X z∈D(G,x) p(z | θ, x)f(z) 3We thank an anonymous reviewer for pointing this out. which gives the expected value of the feature function f(z) under the distribution q(x)×p(z | θ, x). We will make the following assumption about G: Condition 1. There exists some θI such that ∀x ∈L(G), ∀z ∈D(G, x), p(z | θI, x) = 1/|D(G, x)|. This condition is satisfied, for example, when G is in Chomsky normal form and for all A, A′ ∈N, |R(A)| = |R(A′)|. Then, if we set θA→α = 1/|R(A)|, we get that all derivations of x will have the same number of rules and hence the same probability. This condition does not hold for grammars with unary cycles because |D(G, x)| may be infinite for some derivations. Such grammars are not commonly used in NLP. Let us assume that some “correct” parameters θ∗exist, and that our data were drawn from a distribution parametrized by θ∗. The goal of this section is to motivate the following initialization for θ, which we call UniformInit: 1. Initialize z by sampling from the uniform distribution over D(G, xi) for each xi. 2. Update the grammar parameters using maximum likelihood estimation. 7.1 Bounding the Objective To show our result, we require first the following definition due to Freund et al. (1997): Definition 5. A distribution p1 is within λ ≥1 of a distribution p2 if for every event A, we have 1 λ ≤p1(A) p2(A) ≤λ (13) For any feature function f(z) and any two sets of parameters θ2 and θ1 for G and for any marginal q(x), if p(z | θ1, x) is within λ of p(z | θ2, x) for all x then: Eq,θ1[f] λ ≤Eq,θ2[f] ≤λEq,θ1[f] (14) Let θ0 be a set of parameters such that we perform the following procedure in initializing Viterbi EM: first, we sample from the posterior distribution p(z | θ0, x), and then update the parameters with maximum likelihood estimate, in a regular M-step. Let λ be such that p(z | θ0, x) is within λ of p(z | θ∗, x) (for all x ∈L(G)). (Later we will show that UniformInit is a wise choice for making λ small. Note that UniformInit is equivalent to the procedure mentioned above with θ0 = θI.) 1507 Consider ˜pn(x), the empirical distribution over x1, . . . , xn. As n →∞, we have that ˜pn(x) → p∗(x), almost surely, where p∗is: p∗(x) = X z p∗(x, z | θ∗) (15) This means that as n →∞we have E˜pn,θ[f] → Ep∗,θ[f]. Now, let z0 = (z0,1, . . . , z0,n) be samples from p(z | θ0, xi) for i ∈{1, . . . , n}. Then, from simple MLE computation, we know that the value max θ′ n Y i=1 p(xi, z0,i | θ′) (16) = Y (A→α)∈R FA→α(z0) FA(z0) FA→α(z0) We also know that for θ0, from the consistency of MLE, for large enough samples: FA→α(z0) FA(z0) ≈ E˜pn,θ0[fA→α] E˜pn,θ0[fA] (17) which means that we have the following as n grows (starting from the ViterbiTrain objective with initial state z = z0): max θ′ n Y i=1 p(xi, z0,i | θ′) (18) (Eq. 16) = Y (A→α)∈R FA→α(z0) FA(z0) FA→α(z0) (19) (Eq. 17) ≈ Y (A→α)∈R E˜pn,θ0[fA→α] E˜pn,θ0[fA] FA→α(z0) (20) We next use the fact that ˜pn(x) ≈p∗(x) for large n, and apply Eq. 14, noting again our assumption that p(z | θ0, x) is within λ of p(z | θ∗, x). We also let B = X i |zi|, where |zi| is the number of nodes in the derivation zi. Note that FA(zi) ≤ B. The above quantity (Eq. 20) is approximately bounded above by Y (A→α)∈R 1 λ2B Ep∗,θ∗[fA→α] Ep∗,θ∗[fA] FA→α(z0) (21) = 1 λ2|R|B Y (A→α)∈R (θ∗ A→α)FA→α(z0) (22) Eq. 22 follows from: θ∗ A→α = Ep∗,θ∗[fA→α] Ep∗,θ∗[fA] (23) If we continue to develop Eq. 22 and apply Eq. 17 and Eq. 23 again, we get that: 1 λ2|R|B Y (A→α)∈R (θ∗ A→α)FA→α(z0) = 1 λ2|R|B Y (A→α)∈R (θ∗ A→α)FA→α(z0)· FA(z0) FA(z0) ≈ 1 λ2|R|B Y (A→α)∈R (θ∗ A→α) Ep∗,θ0 [fA→α] Ep∗,θ0 [fA] ·FA(z0) ≥ 1 λ2|R|B Y (A→α)∈R (θ∗ A→α)λ2θ∗ A→αFA(z0) ≥ 1 λ2|R|B Y (A→α)∈R (θ∗ A→α)nθ∗ A→α | {z } T(θ∗,n) Bλ2/n (24) = 1 λ2|R|B T(θ∗, n)Bλ2/n (25) ≜ d(λ; θ∗, |R|, B) (26) where Eq. 24 is the result of FA(z0) ≤B. For two series {an} and {bn}, let “an ⪆bn” denote that limn→∞an ≥limn→∞bn. In other words, an is asymptotically larger than bn. Then, if we changed the representation of the objective function of the ViterbiTrain problem to loglikelihood, for θ′ that maximizes Eq. 18 (with some simple algebra) we have: 1 n n X i=1 log2 p(xi, z0,i | θ′) (27) ⪆ −2|R|B n log2 λ + Bλ2 n 1 n log2 T(θ∗, n) = −2|R|B n log2 λ −|N| Bλ2 |N|n X A∈N H(θ∗, A) (28) where H(θ∗, A) = − X (A→α)∈R(A) θ∗ A→α log2 θ∗ A→α (29) is the entropy of the multinomial for nonterminal A. H(θ∗, A) can be thought of as the minimal number of bits required to encode a choice of a rule from A, if chosen independently from the other rules. All together, the quantity B |N|n P A∈N H(θ∗, A) is the average number of bits required to encode a tree in our sample using 1508 θ∗, while removing dependence among all rules and assuming that each node at the tree is chosen uniformly.4 This means that the log-likelihood, for large n, is bounded from above by a linear function of the (average) number of bits required to optimally encode n trees of total size B, while assuming independence among the rules in a tree. We note that the quantity B/n will tend toward the average size of a tree, which, under Condition 1, must be finite. Our final approximate bound from Eq. 28 relates the choice of distribution, from which sample z0, to λ. The lower bound in Eq. 28 is a monotonedecreasing function of λ. We seek to make λ as small as possible to make the bound tight. We next show that the uniform distribution optimizes λ in that sense. 7.2 Optimizing λ Note that the optimal choice of λ, for a single x and for candidate initializer θ′, is λopt(x, θ∗; θ0) = sup z∈D(G,x) p(z | θ0, x) p(z | θ∗, x)(30) In order to avoid degenerate cases, we will add another condition on the true model, θ∗: Condition 2. There exists τ > 0 such that, for any x ∈L(G) and for any z ∈D(G, x), p(z | θ∗, x) ≥τ. This is a strong condition, forcing the cardinality of D(G) to be finite, but it is not unreasonable if natural language sentences are effectively bounded in length. Without further information about θ∗(other than that it satisfies Condition 2), we may want to consider the worst-case scenario of possible λ, hence we seek initializer θ0 such that Λ(x; θ0) ≜ sup θ λopt(x, θ; θ0) (31) is minimized. If θ0 = θI, then we have that p(z | θI, x) = |D(G, x)|−1 ≜µx. Together with Condition 2, this implies that p(z | θI, x) p(z | θ∗, x) ≤ µx τ (32) 4We note that Grenander (1967) describes a (linear) relationship between the derivational entropy and H(θ∗, A). The derivational entropy is defined as h(θ∗, A) = −P x,z p(x, z | θ∗) log p(x, z | θ∗), where z ranges over trees that have nonterminal A as the root. It follows immediately from Grenander’s result that P A H(θ∗, A) ≤ P A h(θ∗, A). and hence λopt(x, θ∗) ≤µx/τ for any θ∗, hence Λ(x; θI) ≤µx/τ. However, if we choose θ0 ̸= θI, we have that p(z′ | θ0, x) > µx for some z′, hence, for θ∗such that it assigns probability τ on z′, we have that sup z∈D(G,x) p(z | θ0, x) p(z | θ∗, x) > µx τ (33) and hence λopt(x, θ∗; θ′) > µx/τ, so Λ(x; θ′) > µx/τ. So, to optimize for the worst-case scenario over true distributions with respect to λ, we are motivated to choose θ0 = θI as defined in Condition 1. Indeed, UniformInit uses θI to initialize the state of Viterbi EM. We note that if θI was known for a specific grammar, then we could have used it as a direct initializer. However, Condition 1 only guarantees its existence, and does not give a practical way to identify it. In general, as mentioned above, θ = 1 can be used to obtain a weighted CFG that satisfies p(z | θ, x) = 1/|D(G, x)|. Since we require a uniform posterior distribution, the number of derivations of a fixed length is finite. This means that we can converted the weighted CFG with θ = 1 to a PCFG with the same posterior (Smith and Johnson, 2007), and identify the appropriate θI. 8 Related Work Viterbi training is closely related to the k-means clustering problem, where the objective is to find k centroids for a given set of d-dimensional points such that the sum of distances between the points and the closest centroid is minimized. The analog for Viterbi EM for the k-means problem is the k-means clustering algorithm (Lloyd, 1982), a coordinate ascent algorithm for solving the k-means problem. It works by iterating between an E-likestep, in which each point is assigned the closest centroid, and an M-like-step, in which the centroids are set to be the center of each cluster. “k” in k-means corresponds, in a sense, to the size of our grammar. k-means has been shown to be NP-hard both when k varies and d is fixed and when d varies and k is fixed (Aloise et al., 2009; Mahajan et al., 2009). An open problem relating to our hardness result would be whether ViterbiTrain (or ConditionalViterbiTrain) is hard even if we do not permit grammars of arbitrarily large size, or at least, constrain the number of rules that do not rewrite to terminals (in our current reduction, the 1509 size of the grammar grows as the size of the 3-SAT formula grows). On a related note to §7, Arthur and Vassilvitskii (2007) described a greedy initialization algorithm for initializing the centroids of k-means, called k-means++. They show that their initialization is O(log k)-competitive; i.e., it approximates the optimal clusters assignment by a factor of O(log k). In §7.1, we showed that uniform-at-random initialization is approximately O(|N|Lλ2/n)-competitive (modulo an additive constant) for CNF grammars, where n is the number of sentences, L is the total length of sentences and λ is a measure for distance between the true distribution and the uniform distribution.5 Many combinatorial problems in NLP involving phrase-structure trees, alignments, and dependency graphs are hard (Sima’an, 1996; Goodman, 1998; Knight, 1999; Casacuberta and de la Higuera, 2000; Lyngsø and Pederson, 2002; Udupa and Maji, 2006; McDonald and Satta, 2007; DeNero and Klein, 2008, inter alia). Of special relevance to this paper is Abe and Warmuth (1992), who showed that the problem of finding maximum likelihood model of probabilistic automata is hard even for a single string and an automaton with two states. Understanding the complexity of NLP problems, we believe, is crucial as we seek effective practical approximations when necessary. 9 Conclusion We described some properties of Viterbi training for probabilistic context-free grammars. We showed that Viterbi training is NP-hard and, in fact, NP-hard to approximate. We gave motivation for uniform-at-random initialization for derivations in the Viterbi EM algorithm. Acknowledgments We acknowledge helpful comments by the anonymous reviewers. This research was supported by NSF grant 0915187. References N. Abe and M. Warmuth. 1992. On the computational complexity of approximating distributions by prob5Making the assumption that the grammar is in CNF permits us to use L instead of B, since there is a linear relationship between them in that case. abilistic automata. Machine Learning, 9(2–3):205– 260. S. Abney. 2007. Semisupervised Learning for Computational Linguistics. CRC Press. D. Aloise, A. Deshpande, P. Hansen, and P. Popat. 2009. NP-hardness of Euclidean sum-of-squares clustering. Machine Learning, 75(2):245–248. D. Arthur and S. Vassilvitskii. 2007. k-means++: The advantages of careful seeding. In Proc. of ACMSIAM symposium on Discrete Algorithms. F. Casacuberta and C. de la Higuera. 2000. Computational complexity of problems on probabilistic grammars and transducers. In Proc. of ICGI. E. Charniak. 1997. Statistical parsing with a contextfree grammar and word statistics. In Proc. of AAAI. S. B. Cohen and N. A. Smith. 2009. Shared logistic normal distributions for soft parameter tying in unsupervised grammar induction. In Proc. of HLTNAACL. M. Collins. 2003. Head-driven statistical models for natural language processing. Computational Linguistics, 29(4):589–637. W. H. E. Day. 1983. Computationally difficult parsimony problems in phylogenetic systematics. Journal of Theoretical Biology, 103. J. DeNero and D. Klein. 2008. The complexity of phrase alignment problems. In Proc. of ACL. Y. Freund, H. Seung, E. Shamir, and N. Tishby. 1997. Selective sampling using the query by committee algorithm. Machine Learning, 28(2–3):133–168. S. Goldwater and M. Johnson. 2005. Bias in learning syllable structure. In Proc. of CoNLL. J. Goodman. 1998. Parsing Inside-Out. Ph.D. thesis, Harvard University. U. Grenander. 1967. Syntax-controlled probabilities. Technical report, Brown University, Division of Applied Mathematics. M. Johnson, T. L. Griffiths, and S. Goldwater. 2006. Adaptor grammars: A framework for specifying compositional nonparameteric Bayesian models. In Advances in NIPS. M. Johnson, T. L. Griffiths, and S. Goldwater. 2007. Bayesian inference for PCFGs via Markov chain Monte Carlo. In Proc. of NAACL. D. Klein and C. Manning. 2001. Natural language grammar induction using a constituentcontext model. In Advances in NIPS. K. Knight. 1999. Decoding complexity in wordreplacement translation models. Computational Linguistics, 25(4):607–615. S. P. Lloyd. 1982. Least squares quantization in PCM. In IEEE Transactions on Information Theory. R. B. Lyngsø and C. N. S. Pederson. 2002. The consensus string problem and the complexity of comparing hidden Markov models. Journal of Computing and System Science, 65(3):545–569. M. Mahajan, P. Nimbhorkar, and K. Varadarajan. 2009. The planar k-means problem is NP-hard. In Proc. of International Workshop on Algorithms and Computation. 1510 D. McClosky, E. Charniak, and M. Johnson. 2006a. Effective self-training for parsing. In Proc. of HLTNAACL. D. McClosky, E. Charniak, and M. Johnson. 2006b. Reranking and self-training for parser adaptation. In Proc. of COLING-ACL. R. McDonald and G. Satta. 2007. On the complexity of non-projective data-driven dependency parsing. In Proc. of IWPT. R. M. Neal and G. E. Hinton. 1998. A view of the EM algorithm that justifies incremental, sparse, and other variants. In Learning and Graphical Models, pages 355–368. Kluwer Academic Publishers. K. Sima’an. 1996. Computational complexity of probabilistic disambiguation by means of tree-grammars. In In Proc. of COLING. M. Sipser. 2006. Introduction to the Theory of Computation, Second Edition. Thomson Course Technology. N. A. Smith and M. Johnson. 2007. Weighted and probabilistic context-free grammars are equally expressive. Computational Linguistics, 33(4):477– 491. V. I. Spitkovsky, H. Alshawi, D. Jurafsky, and C. D. Manning. 2010. Viterbi training improves unsupervised dependency parsing. In Proc. of CoNLL. R. Udupa and K. Maji. 2006. Computational complexity of statistical machine translation. In Proc. of EACL. M. Wang, N. A. Smith, and T. Mitamura. 2007. What is the Jeopardy model? a quasi-synchronous grammar for question answering. In Proc. of EMNLP. C. Yejin and C. Cardie. 2007. Structured local training and biased potential functions for conditional random fields with application to coreference resolution. In Proc. of HLT-NAACL. 1511
|
2010
|
152
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1512–1521, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics A Generalized-Zero-Preserving Method for Compact Encoding of Concept Lattices Matthew Skala School of Computer Science University of Waterloo [email protected] Victoria Krakovna J´anos Kram´ar Dept. of Mathematics University of Toronto {vkrakovna,jkramar}@gmail.com Gerald Penn Dept. of Computer Science University of Toronto [email protected] Abstract Constructing an encoding of a concept lattice using short bit vectors allows for efficient computation of join operations on the lattice. Join is the central operation any unification-based parser must support. We extend the traditional bit vector encoding, which represents join failure using the zero vector, to count any vector with less than a fixed number of one bits as failure. This allows non-joinable elements to share bits, resulting in a smaller vector size. A constraint solver is used to construct the encoding, and a variety of techniques are employed to find near-optimal solutions and handle timeouts. An evaluation is provided comparing the extended representation of failure with traditional bit vector techniques. 1 Introduction The use of bit vectors is almost as old as HPSG parsing itself. Since they were first suggested in the programming languages literature (A¨ıt-Kaci et al., 1989) as a method for computing the unification of two types without table lookup, bit vectors have been attractive because of three speed advantages: • The classical bit vector encoding uses bitwise AND to calculate type unification. This is hard to beat. • Hash tables, the most common alternative, involve computing the Dedekind-MacNeille completion (DMC) at compile time if the input type hierarchy is not a bounded-complete partial order. That is exponential time in the worst case; most bit vector methods avoid explicitly computing it. • With large type signatures, the table that indexes unifiable pairs of types may be so large that it pushes working parsing memory into swap. This loss of locality of reference costs time. Why isn’t everyone using bit vectors? For the most part, the reason is their size. The classical encoding given by A¨ıt-Kaci et al. (1989) is at least as large as the number of meet-irreducible types, which in the parlance of HPSG type signatures is the number of unary-branching types plus the number of maximally specific types. For the English Resource Grammar (ERG) (Copestake and Flickinger, 2000), these are 314 and 2474 respectively. While some systems use them nonetheless (PET (Callmeier, 2000) does, as a very notable exception), it is clear that the size of these codes is a source of concern. Again, it has been so since the very beginning: A¨ıt-Kaci et al. (1989) devoted several pages to a discussion of how to “modularize” type codes, which typically achieves a smaller code in exchange for a larger-time operation than bitwise AND as the implementation of type unification. However, in this and later work on the subject (e.g. (Fall, 1996)), one constant has been that we know our unification has failed when the implementation returns the zero vector. Zero preservation (Mellish, 1991; Mellish, 1992), i.e., detecting a type unification failure, is just as important as obtaining the right answer quickly when it succeeds. The approach of the present paper borrows from recent statistical machine translation research, which addresses the problem of efficiently representing large-scale language models using a mathematical construction called a Bloom filter (Talbot and Osborne, 2007). The approach is best combined with modularization in order to further reduce the size of the codes, but its novelty lies in 1512 the observation that counting the number of one bits in an integer is implemented in the basic instruction sets of many CPUs. The question then arises whether smaller codes would be obtained by relaxing zero preservation so that any resulting vector with at most λ bits is interpreted as failure, with λ ≥1. Penn (2002) generalized join-preserving encodings of partial orders to the case where more than one code can be used to represent the same object, but the focus there was on codes arising from successful unifications; there was still only one representative for failure. To our knowledge, the present paper is the first generalization of zero preservation in CL or any other application domain of partial order encodings. We note at the outset that we are not using Bloom filters as such, but rather a derandomized encoding scheme that shares with Bloom filters the essential insight that λ can be greater than zero without adverse consequences for the required algebraic properties of the encoding. Deterministic variants of Bloom filters may in turn prove to be of some value in language modelling. 1.1 Notation and definitions A partial order ⟨X, ⊑⟩consists of a set X and a reflexive, antisymmetric, and transitive binary relation ⊑. We use u ⊔v to denote the unique least upper bound or join of u, v ∈X, if one exists, and u ⊓v for the greatest lower bound or meet. If we need a second partial order, we use ⪯for its order relation and ⋎for its join operation. We are especially interested in a class of partial orders called meet semilattices, in which every pair of elements has a unique meet. In a meet semilattice, the join of two elements is unique when it exists at all, and there is a unique globally least element ⊥(“bottom”). A successor of an element u ∈X is an element v ̸= u ∈X such that u ⊑v and there is no w ∈X with w ̸= u, w ̸= v, and u ⊑w ⊑v, i.e., v follows u in X with no other elements in between. A maximal element has no successor. A meet irreducible element is an element u ∈X such that for any v, w ∈X, if u = v ⊓w then u = v or u = w. A meet irreducible has at most one successor. Given two partial orders ⟨X, ⊑⟩and ⟨Y, ⪯⟩, an embedding of X into Y is a pair of functions f : X →Y and g : (Y × Y ) →{0, 1}, which may have some of the following properties for all u, v ∈X: u ⊑v ⇒f(u) ⪯f(v) (1) defined(u ⊔v) ⇒g(f(u), f(v)) = 1 (2) ¬defined(u ⊔v) ⇒g(f(u), f(v)) = 0 (3) u ⊔v = w ⇔f(u) ⋎f(v) = f(w) (4) With property (1), the embedding is said to preserve order; with property (2), it preserves success; with property (3), it preserves failure; and with property (4), it preserves joins. 2 Bit-vector encoding Intuitively, taking the join of two types in a type hierarchy is like taking the intersection of two sets. Types often represent sets of possible values, and the type represented by the join really does represent the intersection of the sets that formed the input. So it seems natural to embed a partial order of types ⟨X, ⊑⟩into a partial order (in fact, a lattice) of sets ⟨Y, ⪯⟩, where Y is the power set of some set Z, and ⪯is the superset relation ⊇. Then join ⋎is simply set intersection ∩. The embedding function g, which indicates whether a join exists, can be naturally defined by g(f(u), f(v)) = 0 if and only if f(u) ∩f(v) = ∅. It remains to choose the underlying set Z and embedding function f. A¨ıt-Kaci et al. (1989) developed what has become the standard technique of this type. They set Z to be the set of all meet irreducible elements in X; and f(u) = {v ∈Z|v ⊒u}, that is, the meet irreducible elements greater than or equal to u. The resulting embedding preserves order, success, failure, and joins. If Z is chosen to be the maximal elements of X instead, then join preservation is lost but the embedding still preserves order, success, and failure. The sets can be represented efficiently by vectors of bits. We hope to minimize the size of the largest set f(⊥), which determines the vector length. It follows from the work of Markowsky (1980) that the construction of A¨ıt-Kaci et al. is optimal among encodings that use sets with intersection for meet and empty set for failure: with Y defined as the power set of some set Z, ⊑as ⊇, ⊔as ∩, and g(f(u), f(v)) = 0 if and only if f(u) ∩f(v) = ∅, then the smallest Z that will preserve order, success, failure, and joins is the set of all meet irreducible elements of X. No shorter bit vectors are possible. We construct shorter bit vectors by modifying the definition of g, so that the minimality results 1513 no longer apply. In the following discussion we present first an intuitive and then a technical description of our approach. 2.1 Intuition from Bloom filters Vectors generated by the above construction tend to be quite sparse, or if not sparse, at least boring. Consider a meet semilattice containing only the bottom element ⊥and n maximal elements all incomparable to each other. Then each bit vector would consist of either all ones, or all zeroes except for a single one. We would thus be spending n bits to represent a choice among n + 1 alternatives, which should fit into a logarithmic number of bits. The meet semilattices that occur in practice are more complicated than this example, but they tend to contain things like it as a substructure. With the traditional bit vector construction, each of the maximal elements consumes its own bit, even though those bits are highly correlated. The well-known technique called Bloom filtering (Bloom, 1970) addresses a similar issue. There, it is desired to store a large array of bits subject to two considerations. First, most of the bits are zeroes. Second, we are willing to accept a small proportion of one-sided errors, where every query that should correctly return one does so, but some queries that should correctly return zero might actually return one instead. The solution proposed by Bloom and widely used in the decades since is to map the entries in the large bit array pseudorandomly (by means of a hash function) into the entries of a small bit array. To store a one bit we find its hashed location and store it there. If we query a bit for which the answer should be zero but it happens to have the same hashed location as another query with the answer one, then we return a one and that is one of our tolerated errors. To reduce the error rate we can elaborate the construction further: with some fixed k, we use k hash functions to map each bit in the large array to several locations in the small one. Figure 1 illustrates the technique with k = 3. Each bit has three hashed locations. On a query, we check all three; they must all contain ones for the query to return a one. There will be many collisions of individual hashed locations, as shown; but the chances are good that when we query a bit we did not intend to store in the filter, at least one of its hashed locations will still be empty, and so the query will 1 1 1 1 1 1 1 1 1 ? 1 Figure 1: A Bloom filter return zero. Bloom describes how to calculate the optimal value of k, and the necessary length of the hashed array, to achieve any desired bound on the error rate. In general, the hashed array can be much smaller than the original unhashed array (Bloom, 1970). Classical Bloom filtering applied to the sparse vectors of the embedding would create some percentage of incorrect join results, which would then have to be handled by other techniques. Our work described here combines the idea of using k hash functions to reduce the error rate, with perfect hashes designed in a precomputation step to bring the error rate to zero. 2.2 Modified failure detection In the traditional bit vector construction, types map to sets, join is computed by intersection of sets, and the empty set corresponds to failure (where no join exists). Following the lead of Bloom filters, we change the embedding function g(f(u), f(v)) to be 0 if and only if |f(u)∩f(v)| ≤ λ for some constant λ. With λ = 0 this is the same as before. Choosing greater values of λ allows us to re-use set elements in different parts of the type hierarchy while still avoiding collisions. Figure 2 shows an example meet semilattice. In the traditional construction, to preserve joins we must assign one bit to each of the meet-irreducible elements {d, e, f, g, h, i, j, k, l, m}, for a total of ten bits. But we can use eight bits and still preserve joins by setting g(f(u), f(v)) = 0 if and only if |f(u) ∩f(v)| ≤λ = 1, and f as follows. f(⊥) = {1, 2, 3, 4, 5, 6, 7, 8} f(a) = {1, 2, 3, 4, 5} f(b) = {1, 6, 7, 8} f(c) = {1, 2, 3} f(d) = {2, 3, 4, 5} f(e) = {1, 6} f(f) = {1, 7} f(g) = {1, 8} f(h) = {6, 7} f(i) = {6, 8} f(j) = {1, 2} f(k) = {1, 3} f(l) = {2, 3} f(m) = {2, 3, 4} (5) 1514 a c d b e f g h i j k l m Figure 2: An example meet semilattice; ⊥is the most general type. As a more general example, consider the very simple meet semilattice consisting of just a least element ⊥with n maximal elements incomparable to each other. For a given λ we can represent this in b bits by choosing the smallest b such that b λ+1 ≥n and assigning each maximal element a distinct choice of the bits. With optimal choice of λ, b is logarithmic in n. 2.3 Modules As A¨ıt-Kaci et al. (1989) described, partial orders encountered in practice often resemble trees. Both their technique and ours are at a disadvantage when applied to large trees; in particular, if the bottom of the partial order has successors which are not joinable with each other, then those will be assigned large sets with little overlap, and bits in the vectors will tend to be wasted. To avoid wasting bits, we examine the partial order X in a precomputation step to find the modules, which are the smallest upward-closed subsets of X such that for any x ∈X, if x has at least two joinable successors, then x is in a module. This is similar to ALE’s definition of module (Penn, 1999), but not the same. The definition of A¨ıt-Kaci et al. (1989) also differs from ours. Under our definition, every module has a unique least element, and not every type is in a module. For instance, in Figure 2, the only module has a as its least element. In the ERG’s type hierarchy, there are 11 modules, with sizes ranging from 10 to 1998 types. To find the join of two types in the same module, we find the intersection of their encodings and check whether it is of size greater than λ. If the types belong to two distinct modules, there is no join. For the remaining cases, where at least one of the types lacks a module, we observe that the module bottoms and non-module types form a tree, and the join can be computed in that tree. If x is a type in the module whose bottom is y, and z has no module, then x ⊔z = y ⊔z unless y ⊔z = y in which case x ⊔z = x; so it only remains to compute joins within the tree. Our implementation does that by table lookup. More sophisticated approaches could be appropriate on larger trees. 3 Set programming Ideally, we would like to have an efficient algorithm for finding the best possible encoding of any given meet semilattice. The encoding can be represented as a collection of sets of integers (representing bit indices that contain ones), and an optimal encoding is the collection of sets whose overall union is smallest subject to the constraint that the collection forms an encoding at all. This combinatorial optimization problem is a form of set programming; and set programming problems are widely studied. We begin by defining the form of set programming we will use. Definition 1 Choose set variables S1, S2, . . . , Sn to minimize b = | Sn i=1 Si| subject to some constraints of the forms |Si| ≥ri, Si ⊆Sj, Si ⊉Sj, |Si ∩Sj| ≤λ, and Si ∩Sj = Sk. The constant λ is the same for all constraints. Set elements may be arbitrary, but we generally assume they are the integers {1 . . . b} for convenience. The reduction of partial order representation to set programming is clear: we create a set variable for every type, force the maximal types’ sets to contain at least λ + 1 elements, and then use subset to enforce that every type is a superset of all its successors (preserving order and success). We limit the maximum intersection of incomparable types to preserve failure. To preserve joins, if that property is desired, we add a constraint Si ⊉Sj for every pair of types xi ̸⊑xj and one of the form Si ∩Sj = Sk for every xi, xj, xk such that xi ⊔xj = xk.. Given a constraint satisfaction problem like this one, we can ask two questions: is there a feasible solution, assigning values to the variables so all constraints are satisfied; and if so what is the optimal solution, producing the best value of the objective while remaining feasible? In our problem, there is always a feasible solution we can find by the generalized A¨ıt-Kaci et al. construction (GAK), which consists of assigning λ bits 1515 shared among all types; adding enough unshared new bits to maximal elements to satisfy cardinality constraints; adding one new bit to each nonmaximal meet irreducible type; and propagating all the bits down the hierarchy to satisfy the subset constraints. Since the GAK solution is feasible, it provides a useful upper bound on the result of the set programming. Ongoing research on set programming has produced a variety of software tools for solving these problems. However, at first blush our instances are much too large for readily-available set programming tools. Grammars like ERG contain thousands of types. We use binary constraints between every pair of types, for a total of millions of constraints—and these are variables and constraints over a domain of sets, not integers or reals. General-purpose set programming software cannot handle such instances. 3.1 Simplifying the instances First of all, we only use minimum cardinality constraints |Si| ≥ri for maximal types; and every ri ≥λ + 1. Given a feasible bit assignment for a maximal type with more than ri elements in its set Si, we can always remove elements until it has exactly ri elements, without violating the other constraints. As a result, instead of using constraints |Si| ≥ri we can use constraints |Si| = ri. Doing so reduces the search space. Subset is transitive; so if we have constraints Si ⊆Sj and Sj ⊆Sk, then Si ⊆Sk is implied and we need not specify it as a constraint. Similarly, if we have Si ⊆Sj and Si ⊈Sk, then we have Sj ⊈Sk. Furthermore, if Si and Sj have maximum intersection λ, then any subset of Si also has maximum intersection λ with any subset of Sk, and we need not specify those constraints either. Now, let a choke-vertex in the partial order ⟨X, ⊑⟩be an element u ∈X such that for every v, w ∈X where v is a successor of w and u ⊑v, we have u ⊑w. That is, any chain of successors from elements not after u to elements after u, must pass through u. Figure 2 shows chokevertices as squares. We call these choke-vertices by analogy with the graph theoretic concept of cut-vertices in the Hasse diagram of the partial order; but note that some vertices (like j and k) can be choke-vertices without being cut-vertices, and some vertices (like c) can be cut-vertices without being choke-vertices. Maximal and minimal elements are always choke-vertices. Choke-vertices are important because the optimal bit assignment for elements after a chokevertex u is almost independent of the bit assignment elsewhere in the partial order. Removing the redundant constraints means there are no constraints between elements after u and elements before, or incomparable with, u. All constraints across u must involve u directly. As a result, we can solve a smaller instance consisting of u and everything after it, to find the minimal number of bits ru for representing u. Then we solve the rest of the problem with a constraint |Su| = ru, excluding all partial order elements after u, and then combine the two solutions with any arbitrary bijection between the set elements assigned to u in each solution. Assuming optimal solutions to both sub-problems, the result is an optimal solution to the original problem. 3.2 Splitting into components If we cut the partial order at every choke-vertex, we reduce the huge and impractical encoding problem to a collection of smaller ones. The cutting expresses the original partial order as a tree of components, each of which corresponds to a set programming instance. Components are shown by the dashed lines in Figure 2. We can find an optimal encoding for the entire partial order by optimally encoding the components, starting with the leaves of that tree and working our way back to the root. The division into components creates a collection of set programming instances with a wide range of sizes and difficulty; we examine each instance and choose appropriate techniques for each one. Table 1 summarizes the rules used to solve an instance, and shows the number of times each rule was applied in a typical run with the modules extracted from ERG, a ten-minute timeout, and each λ from 0 to 10. In many simple cases, GAK is provably optimal. These include when λ = 0 regardless of the structure of the component; when the component consists of a bottom and zero, one, or two nonjoinable successors; and when there is one element (a top) greater than all other elements in the component. We can easily recognize these cases and apply GAK to them. Another important special case is when the 1516 Condition Succ. Fail. Method λ = 0 216 GAK (optimal) ∃top 510 GAK (optimal) 2 successors 850 GAK (optimal) 3 or 4 successors 70 exponential variable only ULs 420 b-choose-(λ+1) special case before UL removal 251 59 ic_sets after UL removal 9 50 ic_sets remaining 50 GAK Table 1: Rules for solving an instance in the ERG component consists of a bottom and some number k of pairwise non-joinable successors, and the successors all have required cardinality λ + 1. Then the optimal encoding comes from finding the smallest b such that b λ+1 is at least k, and giving each successor a distinct combination of the b bits. 3.3 Removing unary leaves For components that do not have one of the special forms described above, it becomes necessary to solve the set programming problem. Some of our instances are small enough to apply constraint solving software directly; but for larger instances, we have one more technique to bring them into the tractable range. Definition 2 A unary leaf (UL) is an element x in a partial order ⟨X, ⊑⟩such that x is maximal and x is the successor of exactly one other element. ULs are special because their set programming constraints always take a particular form: if x is a UL and a successor of y, then the constraints on its set Sx are exactly that |Sx| = λ + 1, Sx ⊆Sy, and Sx has intersection of size at most λ with the set for any other successor of y. Other constraints disappear by the simplifications described earlier. Furthermore, ULs occur frequently in the partial orders we consider in practice; and by increasing the number of sets in an instance, they have a disproportionate effect on the difficulty of solving the set programming problem. We therefore implement a special solution process for instances containing ULs: we remove them all, solve the resulting instance, and then add them back one at a time while attempting to increase the overall number of elements as little as possible. This process of removing ULs, solving, and adding them back in, may in general produce suboptimal solutions, so we use it only when the solver cannot find a solution on the full-sized problem. In practical experiments, the solver generally either produces an optimal or very nearly optimal solution within a time limit on the order of ten minutes; or fails to produce a feasible solution at all, even with a much longer limit. Testing whether it finds a solution is then a useful way to determine whether UL removal is worthwhile. Recall that in an instance consisting of k ULs and a bottom, an optimal solution consists of finding the smallest b such that b λ+1 is at least k; that is the number of bits for the bottom, and we can choose any k distinct subsets of size λ + 1 for the ULs. Augmenting an existing solution to include additional ULs involves a similar calculation. To add a UL x as the successor of an element y without increasing the total number of bits, we must find a choice of λ + 1 of the bits already assigned to y, sharing at most λ bits with any of y’s other successors. Those successors are in general sets of arbitrary size, but all that matters for assigning x is how many subsets of size λ + 1 they already cover. The UL can use any such subset not covered by an existing successor of y. Our algorithm counts the subsets already covered, and compares that with the number of choices of λ+1 bits from the bits assigned to y. If enough choices remain, we use them; otherwise, we add bits until there are enough choices. 3.4 Solving For instances with a small number of sets and relatively large number of elements in the sets, we use an exponential variable solver. This encodes the set programming problem into integer programming. For each element x ∈{1, 2, . . . , b}, let c(x) = {i|x ∈Si}; that is, c(x) represents the indices of all the sets in the problem that contain the element x. There are 2n −1 possible values of c(x), because each element must be in at least one set. We create an integer variable for each of those values. Each element is counted once, so the sum of the integer variables is b. The constraints translate into simple inequalities on sums of the variables; and the system of constraints can be solved with standard integer programming techniques. After solving the integer programming problem we can then assign elements arbitrarily 1517 to the appropriate combinations of sets. Where applicable, the exponential variable approach works well, because it breaks all the symmetries between set elements. It also continues to function well even when the sets are large, since nothing in the problem directly grows when we increase b. The wide domains of the variables may be advantageous for some integer programming solvers as well. However, it creates an integer programming problem of size exponential in the number of sets. As a result, it is only applicable to instances with a very few set variables. For more general set programming instances, we feed the instance directly into a solver designed for such problems. We used the ECLiPSe logic programming system (Cisco Systems, 2008), which offers several set programming solvers as libraries, and settled on the ic sets library. This is a straightforward set programming solver based on containment bounds. We extended the solver by adding a lightweight not-subset constraint, and customized heuristics for variable and value selection designed to guide the solver to a feasible solution as soon as possible. We choose variables near the top of the instance first, and prefer to assign values that share exactly λ bits with existing assigned values. We also do limited symmetry breaking, in that whenever we assign a bit not shared with any current assignment, the choice of bit is arbitrary so we assume it must be the lowestindex bit. That symmetry breaking speeds up the search significantly. The present work is primarily on the benefits of nonzero λ, and so a detailed study of general set programming techniques would be inappropriate; but we made informal tests of several other set-programming solvers. We had hoped that a solver using containment-lexicographic hybrid bounds as described by Sadler and Gervet (Sadler and Gervet, 2008) would offer good performance, and chose the ECLiPSe framework partly to gain access to its ic hybrid sets implementation of such bounds. In practice, however, ic hybrid sets gave consistently worse performance than ic sets (typically by an approximate factor of two). It appears that in intuitive terms, the lexicographic bounds rarely narrowed the domains of variables much until the variables were almost entirely labelled anyway, at which point containment bounds were almost as good; and meanwhile the increased overhead of maintaining the extra bounds slowed down the entire process to more than compensate for the improved propagation. We also evaluated the Cardinal solver included in ECLiPSe, which offers stronger propagation of cardinality information; it lacked other needed features and seemed no more efficient than ic sets. Among these three solvers, the improvements associated with our custom variable and value heuristics greatly outweighed the baseline differences between the solvers; and the differences were in optimization time rather than quality of the returned solutions. Solvers with available source code were preferred for ease of customization, and free solvers were preferred for economy, but a license for ILOG CPLEX (IBM, 2008) was available and we tried using it with the natural encoding of sets as vectors of binary variables. It solved small instances to optimality in time comparable to that of ECLiPSe. However, for medium to large instances, CPLEX proved impractical. An instance with n sets of up to b bits, dense with pairwise constraints like subset and maximum intersection, requires Θ(n2b) variables when encoded into integer programming in the natural way. CPLEX stores a copy of the relaxed problem, with significant bookkeeping information per variable, for every node in the search tree. It is capable of storing most of the tree in compressed form on disk, but in our larger instances even a single node is too large; CPLEX exhausts memory while loading its input. The ECLiPSe solver also stores each set variable in a data structure that increases linearly with the number of elements, so that the size of the problem as stored by ECLiPSe is also Θ(n2b); but the constant for ECLiPSe appears to be much smaller, and its search algorithm stores only incremental updates (with nodes per set instead of per element) on a stack as it explores the tree. As a result, the ECLiPSe solver can process much larger instances than CPLEX without exhausting memory. Encoding into SAT would allow use of the sophisticated solvers available for that problem. Unfortunately, cardinality constraints are notoriously difficult to encode in Boolean logic. The obvious encoding of our problem into CNFSAT would require O(n2bλ) clauses and variables. Encodings into Boolean variables with richer constraints than CNFSAT (we tried, for instance, the SICStus Prolog clp(FD) implementation (Carlsson et al., 1997)) generally exhausted memory on much smaller instances than those handled by the set1518 Module n b0 λ bλ mrs_min 10 7 0 7 conj 13 8 1 7 list 27 15 1 11 local_min 27 21 1 10 cat_min 30 17 1 14 individual 33 15 0 15 head_min 247 55 0 55 *sort* 247 129 3 107 synsem_min 612 255 0 255 sign_min 1025 489 3 357 mod_relation 1998 1749 6 284 entire ERG 4305 2788 140 985 Table 2: Best encodings of the ERG and its modules: n is number of types, b0 is vector length with λ = 0, and λ is parameter that gives the shortest vector length bλ. variable solvers, while offering no improvement in speed. 4 Evaluation Table 2 shows the size of our smallest encodings to date for the entire ERG without modularization, and for each of its modules. These were found by running the optimization process of the previous section on Intel Xeon servers with a timeout of 30 minutes for each invocation of the solver (which may occur several times per module). Under those conditions, some modules take a long time to optimize—as much as two hours per tested value of λ for sign_min. The Xeon’s hyperthreading feature makes reproducibility of timing results difficult, but we found that results almost never improved with additional time allowance beyond the first few seconds in any case, so the practical effect of the timing variations should be minimal. These results show some significant improvements in vector length for the larger modules. However, they do not reveal the entire story. In particular, the apparent superiority of λ = 0 for the synsem_min module should not be taken as indicating that no higher λ could be better: rather, that module includes a very difficult set programming instance on which the solver failed and fell back to GAK. For the even larger modules, nonzero λ proved helpful despite solver failures, because of the bits saved by UL removal. UL removal is clearly a significant advantage, but only Encoding length time space Lookup table n/a 140 72496 Modular, best λ 0–357 321 203 Modular, λ = 0 0–1749 747 579 Non-mod, λ = 0 2788 4651 1530 Non-mod, λ = 1 1243 2224 706 Non-mod, λ = 2 1140 2008 656 Non-mod, λ = 9 1069 1981 622 Non-mod, λ = 140 985 3018 572 Table 3: Query performance. Vector length in bits, time in milliseconds, space in Kbytes. for the modules where the solver is failing anyway. One important lesson seems to be that further work on set programming solvers would be beneficial: any future more capable set programming solver could be applied to the unsolved instances and would be expected to save more bits. Table 3 and Figure 3 show the performance of the join query with various encodings. These results are from a simple implementation in C that tests all ordered pairs of types for joinability. As well as testing the non-modular ERG encoding for different values of λ, we tested the modularized encoding with λ = 0 for all modules (to show the effect of modularization alone) and with λ chosen per-module to give the shortest vectors. For comparison, we also tested a simple lookup table. The same implementation sufficed for all these tests, by means of putting all types in one module for the non-modular bit vectors or no types in any module for the pure lookup table. The times shown are milliseconds of user CPU time to test all join tests (roughly 18.5 million of them), on a non-hyperthreading Intel Pentium 4 with a clock speed of 2.66GHz and 1G of RAM, running Linux. Space consumption shown is the total amount of dynamically-allocated memory used to store the vectors and lookup table. The non-modular encoding with λ = 0 is the basic encoding of A¨ıt-Kaci et al. (1989). As Table 3 shows, we achieved more than a factor of two improvement from that, in both time and vector length, just by setting λ = 1. Larger values offered further small improvements in length up to λ = 140, which gave the minimum vector length of 985. That is a shallow minimum; both λ = 120 and λ = 160 gave vector lengths of 986, and the length slowly increased with greater λ. However, the fastest bit-count on this architec1519 1500 2000 2500 3000 3500 4000 4500 5000 0 50 100 150 200 user CPU time (ms) lambda (bits) Figure 3: Query performance for the ERG without modularization. ture, using a technique first published by Wegner (1960), requires time increasing with the number of nonzero bits it counts; and a similar effect would appear on a word-by-word basis even if we used a constant-time per-word count. As a result, there is a time cost associated with using larger λ, so that the fastest value is not necessarily the one that gives the shortest vectors. In our experiments, λ = 9 gave the fastest joins for the non-modular encoding of the ERG. As shown in Figure 3, all small nonzero λ gave very similar times. Modularization helps a lot, both with λ = 0, and when we choose the optimal λ per module. Here, too, the use of optimal λ improves both time and space by more than a factor of two. Our best bit-vector encoding, the modularized one with permodule optimal λ, is only a little less than half the speed of the lookup table; and this test favours the lookup table by giving it a full word for every entry (no time spent shifting and masking bits) and testing the pairs in a simple two-level loop (almost purely sequential access). 5 Conclusion We have described a generalization of conventional bit vector concept lattice encoding techniques to the case where all vectors with λ or fewer one bits represent failure; traditional encodings are the case λ = 0. Increasing λ can reduce the overall storage space and improve speed. A good encoding requires a kind of perfect hash, the design of which maps naturally to constraint programming over sets of integers. We have described a practical framework for solving the instances of constraint programming thus created, in which we can apply existing or future constraint solvers to the subproblems for which they are best suited; and a technique for modularizing practical type hierarchies to get better value from the bit vector encodings. We have evaluated the resulting encodings on the ERG’s type system, and examined the performance of the associated unification test. Modularization, and the use of nonzero λ, each independently provide significant savings in both time and vector length. The modified failure detection concept suggests several directions for future work, including evaluation of the new encodings in the context of a large-scale HPSG parser; incorporation of further developments in constraint solvers; and the possibility of approximate encodings that would permit one-sided errors as in traditional Bloom filtering. References Hassan A¨ıt-Kaci, Robert S. Boyer, Patrick Lincoln, and Roger Nasr. 1989. Efficient implementation of lattice operations. ACM Transactions on Programming Languages and Systems, 11(1):115–146, January. 1520 Burton H. Bloom. 1970. Space/time trade-offs in hash coding with allowable errors. Communications of the ACM, 13(7):422–426, July. Ulrich Callmeier. 2000. PET – a platform for experimentation with efficient HPSG processing techniques. Natural Language Engineering, 6(1):99– 107. Mats Carlsson, Greger Ottosson, and Bj¨orn Carlson. 1997. An open-ended finite domain constraint solver. In H. Glaser, P. Hartel, and H. Kucken, editors, Programming Languages: Implementations, Logics, and Programming, volume 1292 of Lecture Notes in Computer Science, pages 191–206. Springer-Verlag, September. Cisco Systems. 2008. ECLiPSe 6.0. Computer software. Online http://eclipse-clp.org/. Ann Copestake and Dan Flickinger. 2000. An open-source grammar development environment and broad-coverage English grammar using HPSG. In Proceedings of the Second Conference on Language Resources and Evaluation (LREC 2000). Andrew Fall. 1996. Reasoning with Taxonomies. Ph.D. thesis, Simon Fraser University. IBM. 2008. ILOG CPLEX 11. Computer software. George Markowsky. 1980. The representation of posets and lattices by sets. Algebra Universalis, 11(1):173–192. Chris Mellish. 1991. Graph-encodable description spaces. Technical report, University of Edinburgh Department of Artificial Intelligence. DYANA Deliverable R3.2B. Chris Mellish. 1992. Term-encodable description spaces. In D.R. Brough, editor, Logic Programming: New Frontiers, pages 189–207. Kluwer. Gerald Penn. 1999. An optimized prolog encoding of typed feature structures. In D. De Schreye, editor, Logic programming: proceedings of the 1999 International Conference on Logic Programming (ICLP), pages 124–138. Gerald Penn. 2002. Generalized encoding of description spaces and its application to typed feature structures. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL 2002), pages 64–71. Andrew Sadler and Carmen Gervet. 2008. Enhancing set constraint solvers with lexicographic bounds. Journal of Heuristics, 14(1). David Talbot and Miles Osborne. 2007. Smoothed Bloom filter language models: Tera-scale LMs on the cheap. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pages 468–476. Peter Wegner. 1960. A technique for counting ones in a binary computer. Communications of the ACM, 3(5):322. 1521
|
2010
|
153
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1522–1531, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Knowledge-rich Word Sense Disambiguation Rivaling Supervised Systems Simone Paolo Ponzetto Department of Computational Linguistics Heidelberg University [email protected] Roberto Navigli Dipartimento di Informatica Sapienza Universit`a di Roma [email protected] Abstract One of the main obstacles to highperformance Word Sense Disambiguation (WSD) is the knowledge acquisition bottleneck. In this paper, we present a methodology to automatically extend WordNet with large amounts of semantic relations from an encyclopedic resource, namely Wikipedia. We show that, when provided with a vast amount of high-quality semantic relations, simple knowledge-lean disambiguation algorithms compete with state-of-the-art supervised WSD systems in a coarse-grained all-words setting and outperform them on gold-standard domain-specific datasets. 1 Introduction Knowledge lies at the core of Word Sense Disambiguation (WSD), the task of computationally identifying the meanings of words in context (Navigli, 2009b). In the recent years, two main approaches have been studied that rely on a fixed sense inventory, i.e., supervised and knowledgebased methods. In order to achieve high performance, supervised approaches require large training sets where instances (target words in context) are hand-annotated with the most appropriate word senses. Producing this kind of knowledge is extremely costly: at a throughput of one sense annotation per minute (Edmonds, 2000) and tagging one thousand examples per word, dozens of person-years would be required for enabling a supervised classifier to disambiguate all the words in the English lexicon with high accuracy. In contrast, knowledge-based approaches exploit the information contained in wide-coverage lexical resources, such as WordNet (Fellbaum, 1998). However, it has been demonstrated that the amount of lexical and semantic information contained in such resources is typically insufficient for high-performance WSD (Cuadros and Rigau, 2006). Several methods have been proposed to automatically extend existing resources (cf. Section 2) and it has been shown that highlyinterconnected semantic networks have a great impact on WSD (Navigli and Lapata, 2010). However, to date, the real potential of knowledge-rich WSD systems has been shown only in the presence of either a large manually-developed extension of WordNet (Navigli and Velardi, 2005) or sophisticated WSD algorithms (Agirre et al., 2009). The contributions of this paper are two-fold. First, we relieve the knowledge acquisition bottleneck by developing a methodology to extend WordNet with millions of semantic relations. The relations are harvested from an encyclopedic resource, namely Wikipedia. Wikipedia pages are automatically associated with WordNet senses, and topical, semantic associative relations from Wikipedia are transferred to WordNet, thus producing a much richer lexical resource. Second, two simple knowledge-based algorithms that exploit our extended WordNet are applied to standard WSD datasets. The results show that the integration of vast amounts of semantic relations in knowledge-based systems yields performance competitive with state-of-the-art supervised approaches on open-text WSD. In addition, we support previous findings from Agirre et al. (2009) that in a domain-specific WSD scenario knowledge-based systems perform better than supervised ones, and we show that, given enough knowledge, simple algorithms perform better than more sophisticated ones. 2 Related Work In the last three decades, a large body of work has been presented that concerns the development of automatic methods for the enrichment of existing resources such as WordNet. These in1522 clude proposals to extract semantic information from dictionaries (e.g. Chodorow et al. (1985) and Rigau et al. (1998)), approaches using lexicosyntactic patterns (Hearst, 1992; Cimiano et al., 2004; Girju et al., 2006), heuristic methods based on lexical and semantic regularities (Harabagiu et al., 1999), taxonomy-based ontologization (Pennacchiotti and Pantel, 2006; Snow et al., 2006). Other approaches include the extraction of semantic preferences from sense-annotated (Agirre and Martinez, 2001) and raw corpora (McCarthy and Carroll, 2003), as well as the disambiguation of dictionary glosses based on cyclic graph patterns (Navigli, 2009a). Other works rely on the disambiguation of collocations, either obtained from specialized learner’s dictionaries (Navigli and Velardi, 2005) or extracted by means of statistical techniques (Cuadros and Rigau, 2008), e.g. based on the method proposed by Agirre and de Lacalle (2004). But while most of these methods represent state-of-the-art proposals for enriching lexical and taxonomic resources, none concentrates on augmenting WordNet with associative semantic relations for many domains on a very large scale. To overcome this limitation, we exploit Wikipedia, a collaboratively generated Web encyclopedia. The use of collaborative contributions from volunteers has been previously shown to be beneficial in the Open Mind Word Expert project (Chklovski and Mihalcea, 2002). However, its current status indicates that the project remains a mainly academic attempt. In contrast, due to its low entrance barrier and vast user base, Wikipedia provides large amounts of information at practically no cost. Previous work aimed at transforming its content into a knowledge base includes opendomain relation extraction (Wu and Weld, 2007), the acquisition of taxonomic (Ponzetto and Strube, 2007a; Suchanek et al., 2008; Wu and Weld, 2008) and other semantic relations (Nastase and Strube, 2008), as well as lexical reference rules (Shnarch et al., 2009). Applications using the knowledge contained in Wikipedia include, among others, text categorization (Gabrilovich and Markovitch, 2006), computing semantic similarity of texts (Gabrilovich and Markovitch, 2007; Ponzetto and Strube, 2007b; Milne and Witten, 2008a), coreference resolution (Ponzetto and Strube, 2007b), multi-document summarization (Nastase, 2008), and text generation (Sauper and Barzilay, 2009). In our work we follow this line of research and show that knowledge harvested from Wikipedia can be used effectively to improve the performance of a WSD system. Our proposal builds on previous insights from Bunescu and Pas¸ca (2006) and Mihalcea (2007) that pages in Wikipedia can be taken as word senses. Mihalcea (2007) manually maps Wikipedia pages to WordNet senses to perform lexical-sample WSD. We extend her proposal in three important ways: (1) we fully automatize the mapping between Wikipedia pages and WordNet senses; (2) we use the mappings to enrich an existing resource, i.e. WordNet, rather than annotating text with sense labels; (3) we deploy the knowledge encoded by this mapping to perform unrestricted WSD, rather than apply it to a lexical sample setting. Knowledge from Wikipedia is injected into a WSD system by means of a mapping to WordNet. Previous efforts aimed at automatically linking Wikipedia to WordNet include full use of the first WordNet sense heuristic (Suchanek et al., 2008), a graph-based mapping of Wikipedia categories to WordNet synsets (Ponzetto and Navigli, 2009), a model based on vector spaces (RuizCasado et al., 2005) and a supervised approach using keyword extraction (Reiter et al., 2008). These latter methods rely only on text overlap techniques and neither they take advantage of the input from Wikipedia being semi-structured, e.g. hyperlinked, nor they propose a high-performing probabilistic formulation of the mapping problem, a task to which we turn in the next section. 3 Extending WordNet Our approach consists of two main phases: first, a mapping is automatically established between Wikipedia pages and WordNet senses; second, the relations connecting Wikipedia pages are transferred to WordNet. As a result, an extended version of WordNet is produced, that we call WordNet++. We present the two resources used in our methodology in Section 3.1. Sections 3.2 and 3.3 illustrate the two phases of our approach. 3.1 Knowledge Resources WordNet. Being the most widely used computational lexicon of English in Natural Language Processing, WordNet is an essential resource for WSD. A concept in WordNet is represented as a synonym set, or synset, i.e. the set of words which share a common meaning. For instance, the con1523 cept of soda drink is expressed as: { pop2 n, soda2 n, soda pop1 n, soda water2 n, tonic2 n } where each word’s subscripts and superscripts indicate their parts of speech (e.g. n stands for noun) and sense number1, respectively. For each synset, WordNet provides a textual definition, or gloss. For example, the gloss of the above synset is: “a sweet drink containing carbonated water and flavoring”. Wikipedia. Our second resource, Wikipedia, is a collaborative Web encyclopedia composed of pages2. A Wikipedia page (henceforth, Wikipage) presents the knowledge about a specific concept (e.g. SODA (SOFT DRINK)) or named entity (e.g. FOOD STANDARDS AGENCY). The page typically contains hypertext linked to other relevant Wikipages. For instance, SODA (SOFT DRINK) is linked to COLA, FLAVORED WATER, LEMONADE, and many others. The title of a Wikipage (e.g. SODA (SOFT DRINK)) is composed of the lemma of the concept defined (e.g. soda) plus an optional label in parentheses which specifies its meaning in case the lemma is ambiguous (e.g. SOFT DRINK vs. SODIUM CARBONATE). Finally, some Wikipages are redirections to other pages, e.g. SODA (SODIUM CARBONATE) redirects to SODIUM CARBONATE. 3.2 Mapping Wikipedia to WordNet During the first phase of our methodology we aim to establish links between Wikipages and WordNet senses. Formally, given the entire set of pages SensesWiki and WordNet senses SensesWN, we aim to acquire a mapping: µ : SensesWiki →SensesWN, such that, for each Wikipage w ∈SensesWiki: µ(w) = s ∈SensesWN(w) if a link can be established, ϵ otherwise, where SensesWN(w) is the set of senses of the lemma of w in WordNet. For example, if our 1We use WordNet version 3.0. We use word senses to unambiguously denote the corresponding synsets (e.g. plane1 n for { airplane1 n, aeroplane1 n, plane1 n }). 2http://download.wikipedia.org. We use the English Wikipedia database dump from November 3, 2009, which includes 3,083,466 articles. Throughout this paper, we use Sans Serif for words, SMALL CAPS for Wikipedia pages and CAPITALS for Wikipedia categories. mapping methodology linked SODA (SOFT DRINK) to the corresponding WordNet sense soda2 n, we would have µ(SODA (SOFT DRINK)) = soda2 n. In order to establish a mapping between the two resources, we first identify different kinds of disambiguation contexts for Wikipages (Section 3.2.1) and WordNet senses (Section 3.2.2). Next, we intersect these contexts to perform the mapping (see Section 3.2.3). 3.2.1 Disambiguation Context of a Wikipage Given a target Wikipage w which we aim to map to a WordNet sense of w, we use the following information as a disambiguation context: • Sense labels: e.g. given the page SODA (SOFT DRINK), the words soft and drink are added to the disambiguation context. • Links: the titles’ lemmas of the pages linked from the Wikipage w (outgoing links). For instance, the links in the Wikipage SODA (SOFT DRINK) include soda, lemonade, sugar, etc. • Categories: Wikipages are classified according to one or more categories, which represent meta-information used to categorize them. For instance, the Wikipage SODA (SOFT DRINK) is categorized as SOFT DRINKS. Since many categories are very specific and do not appear in WordNet (e.g., SWEDISH WRITERS or SCIENTISTS WHO COMMITTED SUICIDE), we use the lemmas of their syntactic heads as disambiguation context (i.e. writer and scientist). To this end, we use the category heads provided by Ponzetto and Navigli (2009). Given a Wikipage w, we define its disambiguation context Ctx(w) as the set of words obtained from some or all of the three sources above. 3.2.2 Disambiguation Context of a WordNet Sense Given a WordNet sense s and its synset S, we use the following information as disambiguation context to provide evidence for a potential link in our mapping µ: • Synonymy: all synonyms of s in synset S. For instance, given the synset of soda2 n, all its synonyms are included in the context (that is, tonic, soda pop, pop, etc.). 1524 • Hypernymy/Hyponymy: all synonyms in the synsets H such that H is either a hypernym (i.e., a generalization) or a hyponym (i.e., a specialization) of S. For example, given soda2 n, we include the words from its hypernym { soft drink1 n }. • Sisterhood: words from the sisters of S. A sister synset S′ is such that S and S′ have a common direct hypernym. For example, given soda2 n, it can be found that bitter lemon1 n and soda2 n are sisters. Thus the words bitter and lemon are included in the disambiguation context of s. • Gloss: the set of lemmas of the content words occurring within the gloss of s. For instance, given s = soda2 n, defined as “a sweet drink containing carbonated water and flavoring”, we add to the disambiguation context of s the following lemmas: sweet, drink, contain, carbonated, water, flavoring. Given a WordNet sense s, we define its disambiguation context Ctx(s) as the set of words obtained from some or all of the four sources above. 3.2.3 Mapping Algorithm In order to link each Wikipedia page to a WordNet sense, we developed a novel algorithm, whose pseudocode is shown in Algorithm 1. The following steps are performed: • Initially (lines 1-2), our mapping µ is empty, i.e. it links each Wikipage w to ϵ. • For each Wikipage w whose lemma is monosemous both in Wikipedia and WordNet (i.e. |SensesWiki(w)| = |SensesWN(w)| = 1) we map w to its only WordNet sense w1 n (lines 3-5). • Finally, for each remaining Wikipage w for which no mapping was previously found (i.e., µ(w) = ϵ, line 7), we do the following: – lines 8-10: for each Wikipage d which is a redirection to w, for which a mapping was previously found (i.e. µ(d) ̸= ϵ, that is, d is monosemous in both Wikipedia and WordNet) and such that it maps to a sense µ(d) in a synset S that also contains a sense of w, we map w to the corresponding sense in S. – lines 11-14: if a Wikipage w has not been linked yet, we assign the most likely sense to w based on the maximization of the conditional probabilities p(s|w) over the senses Algorithm 1 The mapping algorithm Input: SensesWiki, SensesWN Output: a mapping µ : SensesWiki →SensesWN 1: for each w ∈SensesWiki 2: µ(w) := ϵ 3: for each w ∈SensesWiki 4: if |SensesWiki(w)| = |SensesWN(w)| = 1 then 5: µ(w) := w1 n 6: for each w ∈SensesWiki 7: if µ(w) = ϵ then 8: for each d ∈SensesWiki s.t. d redirects to w 9: if µ(d) ̸= ϵ and µ(d) is in a synset of w then 10: µ(w) := sense of w in synset of µ(d); break 11: for each w ∈SensesWiki 12: if µ(w) = ϵ then 13: if no tie occurs then 14: µ(w) := argmax s∈SensesWN(w) p(s|w) 15: return µ s ∈SensesWN(w) (no mapping is established if a tie occurs, line 13). As a result of the execution of the algorithm, the mapping µ is returned (line 15). At the heart of the mapping algorithm lies the calculation of the conditional probability p(s|w) of selecting the WordNet sense s given the Wikipage w. The sense s which maximizes this probability can be obtained as follows: µ(w) = argmax s∈SensesWN(w) p(s|w) = argmax s p(s, w) p(w) = argmax s p(s, w) The latter formula is obtained by observing that p(w) does not influence our maximization, as it is a constant independent of s. As a result, the most appropriate sense s is determined by maximizing the joint probability p(s, w) of sense s and page w. We estimate p(s, w) as: p(s, w) = score(s, w) X s′∈SensesWN(w), w′∈SensesWiki(w) score(s′, w′) , where score(s, w) = |Ctx(s)∩Ctx(w)|+1 (we add 1 as a smoothing factor). Thus, in our algorithm we determine the best sense s by computing the intersection of the disambiguation contexts of s and w, and normalizing by the scores summed over all senses of w in Wikipedia and WordNet. 3.2.4 Example We illustrate the execution of our mapping algorithm by way of an example. Let us focus on the 1525 Wikipage SODA (SOFT DRINK). The word soda is polysemous both in Wikipedia and WordNet, thus lines 3–5 of the algorithm do not concern this Wikipage. Lines 6–14 aim to find a mapping µ(SODA (SOFT DRINK)) to an appropriate WordNet sense of the word. First, we check whether a redirection exists to SODA (SOFT DRINK) that was previously disambiguated (lines 8–10). Next, we construct the disambiguation context for the Wikipage by including words from its label, links and categories (cf. Section 3.2.1). The context includes, among others, the following words: soft, drink, cola, sugar. We now construct the disambiguation context for the two WordNet senses of soda (cf. Section 3.2.2), namely the sodium carbonate (#1) and the drink (#2) senses. To do so, we include words from their synsets, hypernyms, hyponyms, sisters, and glosses. The context for soda1 n includes: salt, acetate, chlorate, benzoate. The context for soda2 n contains instead: soft, drink, cola, bitter, etc. The sense with the largest intersection is #2, so the following mapping is established: µ(SODA (SOFT DRINK)) = soda2 n. 3.3 Transferring Semantic Relations The output of the algorithm presented in the previous section is a mapping between Wikipages and WordNet senses (that is, implicitly, synsets). Our insight is to use this alignment to enable the transfer of semantic relations from Wikipedia to WordNet. In fact, given a Wikipage w we can collect all Wikipedia links occurring in that page. For any such link from w to w′, if the two Wikipages are mapped to WordNet senses (i.e., µ(w) ̸= ϵ and µ(w′) ̸= ϵ), we can transfer the corresponding edge (µ(w), µ(w′)) to WordNet. Note that µ(w) and µ(w′) are noun senses, as Wikipages describe nominal concepts or named entities. We refer to this extended resource as WordNet++. For instance, consider the Wikipage SODA (SOFT DRINK). This page contains, among others, a link to the Wikipage SYRUP. Assuming µ(SODA (SODA DRINK)) = soda2 n and µ(SYRUP) = syrup1 n, we can add the corresponding semantic relation (soda2 n, syrup1 n) to WordNet3. Thus, WordNet++ represents an extension of WordNet which includes semantic associative relations between synsets. These are originally 3Note that such relations are unlabeled. However, for our purposes this has no impact, since our algorithms do not distinguish between is-a and other kinds of relations in the lexical knowledge base (cf. Section 4.2). found in Wikipedia and then integrated into WordNet by means of our mapping. In turn, WordNet++ represents the English-only subset of a larger multilingual resource, BabelNet (Navigli and Ponzetto, 2010), where lexicalizations of the synsets are harvested for many languages using the so-called Wikipedia inter-language links and applying a machine translation system. 4 Experiments We perform two sets of experiments: we first evaluate the intrinsic quality of our mapping (Section 4.1) and then quantify the impact of WordNet++ for coarse-grained (Section 4.2) and domainspecific WSD (Section 4.3). 4.1 Evaluation of the Mapping Experimental setting. We first conducted an evaluation of the mapping quality. To create a gold standard for evaluation, we started from the set of all lemmas contained both in WordNet and Wikipedia: the intersection between the two resources includes 80,295 lemmas which correspond to 105,797 WordNet senses and 199,735 Wikipedia pages. The average polysemy is 1.3 and 2.5 for WordNet senses and Wikipages, respectively (2.8 and 4.7 when excluding monosemous words). We selected a random sample of 1,000 Wikipages and asked an annotator with previous experience in lexicographic annotation to provide the correct WordNet sense for each page title (an empty sense label was given if no correct mapping was possible). 505 non-empty mappings were found, i.e. Wikipedia pages with a corresponding WordNet sense. In order to quantify the quality of the annotations and the difficulty of the task, a second annotator sense tagged a subset of 200 pages from the original sample. We computed the inter-annotator agreement using the kappa coefficient (Carletta, 1996) and found out that our annotators achieved an agreement coefficient κ of 0.9, indicating almost perfect agreement. Table 1 summarizes the performance of our disambiguation algorithm against the manually annotated dataset. Evaluation is performed in terms of standard measures of precision (the ratio of correct sense labels to the non-empty labels output by the mapping algorithm), recall (the ratio of correct sense labels to the total of non-empty labels in the gold standard) and F1-measure ( 2P R P +R). We also calculate accuracy, which accounts for 1526 P R F1 A Structure 82.2 68.1 74.5 81.1 Gloss 81.1 64.2 71.7 78.8 Structure + Gloss 81.9 77.5 79.6 84.4 MFS BL 24.3 47.8 32.2 24.3 Random BL 23.8 46.8 31.6 23.9 Table 1: Performance of the mapping algorithm. empty sense labels (that is, calculated on all 1,000 test instances). As baseline we use the most frequent WordNet sense (MFS), as well as a random sense assignment. We evaluate the mapping methodology described in Section 3.2 against different disambiguation contexts for the WordNet senses (cf. Section 3.2.2), i.e. structure-based (including synonymy, hypernymy/hyponymy and sisterhood), gloss-derived evidence, and a combination of the two. As disambiguation context of a Wikipage (Section 3.2.1) we use all information available, i.e. sense labels, links and categories4. Results and discussion. The results show that our method improves on the baseline by a large margin and that higher performance can be achieved by using more disambiguation information. That is, using a richer disambiguation context helps to better choose the most appropriate WordNet sense for a Wikipedia page. The combination of structural and gloss information attains a slight variation in terms of precision (−0.3% and +0.8% compared to Structure and Gloss respectively), but a significantly high increase in recall (+9.4% and +13.3%). This implies that the different disambiguation contexts only partially overlap and, when used separately, each produces different mappings with a similar level of precision. In the joint approach, the harmonic mean of precision and recall, i.e. F1, is in fact 5 and 8 points higher than when separately using structural and gloss information, respectively. As for the baselines, the most frequent sense is just 0.6% and 0.4% above the random baseline in terms of F1 and accuracy, respectively. A χ2 test reveals in fact no statistically significant difference at p < 0.05. This is related to the random distribution of senses in our dataset and the Wikipedia unbiased coverage of WordNet senses. So select4We leave out the evaluation of different contexts for a Wikipage for the sake of brevity. During prototyping we found that the best results were given by using the largest context available, as reported in Table 1. ing the most frequent sense rather than any other sense for each target page represents a choice as arbitrary as picking a sense at random. The final mapping contains 81,533 pairs of Wikipages and word senses they map to, covering 55.7% of the noun senses in WordNet. Using our best performing mapping we are able to extend WordNet with 1,902,859 semantic edges: of these, 97.93% are deemed novel, i.e. no direct edge could previously be found between the synsets. In addition, we performed a stricter evaluation of the novelty of our relations by checking whether these can still be found indirectly by searching for a connecting path between the two synsets of interest. Here we found that 91.3%, 87.2% and 78.9% of the relations are novel to WordNet when performing a graph search of maximum depth of 2, 3 and 4, respectively. 4.2 Coarse-grained WSD Experimental setting. We extrinsically evaluate the impact of WordNet++ on the Semeval2007 coarse-grained all-words WSD task (Navigli et al., 2007). Performing experiments in a coarse-grained setting is a natural choice for several reasons: first, it has been argued that the fine granularity of WordNet is one of the main obstacles to accurate WSD (cf. the discussion in Navigli (2009b)); second, the meanings of Wikipedia pages are intuitively coarser than those in WordNet5. For instance, mapping TRAVEL to the first or the second sense in WordNet is an arbitrary choice, as the Wikipage refers to both senses. Finally, given their different nature, WordNet and Wikipedia do not fully overlap. Accordingly, we expect the transfer of semantic relations from Wikipedia to WordNet to have sometimes the side effect to penalize some fine-grained senses of a word. We experiment with two simple knowledgebased algorithms that are set to perform coarsegrained WSD on a sentence-by-sentence basis: • Simplified Extended Lesk (ExtLesk): The first algorithm is a simplified version of the Lesk 5Note that our polysemy rates from Section 4.1 also include Wikipages whose lemma is contained in WordNet, but which have out-of-domain meanings, i.e. encyclopedic entries referring to specialized named entities such as e.g., DISCOVERY (SPACE SHUTTLE) or FIELD ARTILLERY (MAGAZINE). We computed the polysemy rate for a random sample of 20 polysemous words by manually removing these NEs and found that Wikipedia’s polysemy rate is indeed lower than that of WordNet – i.e. average polysemy of 2.1 vs. 2.8. 1527 algorithm (Lesk, 1986), that performs WSD based on the overlap between the context surrounding the target word to be disambiguated and the definitions of its candidate senses (Kilgarriff and Rosenzweig, 2000). Given a target word w, this method assigns to w the sense whose gloss has the highest overlap (i.e. most words in common) with the context of w, namely the set of content words co-occurring with it in a pre-defined window (a sentence in our case). Due to the limited context provided by the WordNet glosses, we follow Banerjee and Pedersen (2003) and expand the gloss of each sense s to include words from the glosses of those synsets in a semantic relation with s. These include all WordNet synsets which are directly connected to s, either by means of the semantic pointers found in WordNet or through the unlabeled links found in WordNet++. • Degree Centrality (Degree): The second algorithm is a graph-based approach that relies on the notion of vertex degree (Navigli and Lapata, 2010). Starting from each sense s of the target word, it performs a depth-first search (DFS) of the WordNet(++) graph and collects all the paths connecting s to senses of other words in context. As a result, a sentence graph is produced. A maximum search depth is established to limit the size of this graph. The sense of the target word with the highest vertex degree is selected. We follow Navigli and Lapata (2010) and run Degree in a weakly supervised setting where the system attempts no sense assignment if the highest degree score is below a certain (empirically estimated) threshold. The optimal threshold and maximum search depth are estimated by maximizing Degree’s F1 on a development set of 1,000 randomly chosen noun instances from the SemCor corpus (Miller et al., 1993). Experiments on the development dataset using Degree on WordNet++ revealed a performance far lower than expected. Error analysis showed that many instances were incorrectly disambiguated, due to the noise from weak semantic links, e.g. the links from SODA (SOFT DRINK) to EUROPE or AUSTRALIA. Accordingly, in order to improve the disambiguation performance, we developed a filter to rule out weak semantic relations from WordNet++. Given a WordNet++ edge (µ(w), µ(w′)) where w and w′ are both Wikipages and w links to w′, Resource Algorithm Nouns only P R F1 WordNet ExtLesk 83.6 57.7 68.3 Degree 86.3 65.5 74.5 Wikipedia ExtLesk 82.3 64.1 72.0 Degree 96.2 40.1 57.4 WordNet++ ExtLesk 82.7 69.2 75.4 Degree 87.3 72.7 79.4 MFS BL 77.4 77.4 77.4 Random BL 63.5 63.5 63.5 Table 2: Performance on Semeval-2007 coarsegrained all-words WSD (nouns only subset). we first collect all words from the category labels of w and w′ into two bags of words. We remove stopwords and lemmatize the remaining words. We then compute the degree of overlap between the two sets of categories as the number of words in common between the two bags of words, normalized in the [0, 1] interval. We finally retain the link for the DFS if such score is above an empirically determined threshold. The optimal value for this category overlap threshold was again estimated by maximizing Degree’s F1 on the development set. The final graph used by Degree consists of WordNet, together with 152,944 relations from our semantic relation enrichment method (cf. Section 3.3). Results and discussion. We report our results in terms of precision, recall and F1-measure on the Semeval-2007 coarse-grained all-words dataset (Navigli et al., 2007). We first evaluated ExtLesk and Degree using three different resources: (1) WordNet only; (2) Wikipedia only, i.e. only those relations harvested from the links found within Wikipedia pages; (3) their union, i.e. WordNet++. In Table 2 we report the results on nouns only. As common practice, we compare with random sense assignment and the most frequent sense (MFS) from SemCor as baselines. Enriching WordNet with encyclopedic relations from Wikipedia yields a consistent improvement against using WordNet (+7.1% and +4.9% F1 for ExtLesk and Degree) or Wikipedia (+3.4% and +22.0%) alone. The best results are obtained by using Degree with WordNet++. The better performance of Wikipedia against WordNet when using ExtLesk (+3.7%) highlights the quality of the relations extracted. However, no such improvement is found with De1528 Algorithm Nouns only All words P/R/F1 P/R/F1 ExtLesk 81.0 79.1 Degree 85.5 81.7 SUSSX-FR 81.1 77.0 TreeMatch N/A 73.6 NUS-PT 82.3 82.5 SSI 84.1 83.2 MFS BL 77.4 78.9 Random BL 63.5 62.7 Table 3: Performance on Semeval-2007 coarsegrained all-words WSD with MFS as a back-off strategy when no sense assignment is attempted. gree, due to its lower recall. Interestingly, Degree on WordNet++ beats the MFS baseline, which is notably a difficult competitor for unsupervised and knowledge-lean systems. We finally compare our two algorithms using WordNet++ with state-of-the-art WSD systems, namely the best unsupervised (Koeling and McCarthy, 2007, SUSSX-FR) and supervised (Chan et al., 2007, NUS-PT) systems participating in the Semeval-2007 coarse-grained all-words task. We also compare with SSI (Navigli and Velardi, 2005) – a knowledge-based system that participated out of competition – and the unsupervised proposal from Chen et al. (2009, TreeMatch). Table 3 shows the results for nouns (1,108) and all words (2,269 words): we use the MFS as a back-off strategy when no sense assignment is attempted. Degree with WordNet++ achieves the best performance in the literature6. On the nounonly subset of the data, its performance is comparable with SSI and significantly better than the best supervised and unsupervised systems (+3.2% and +4.4% F1 against NUS-PT and SUSSX-FR). On the entire dataset, it outperforms SUSSX-FR and TreeMatch (+4.7% and +8.1%) and its recall is not statistically different from that of SSI and NUS-PT. This result is particularly interesting, given that WordNet++ is extended only with relations between nominals, and, in contrast to SSI, it does not rely on a costly annotation effort to engineer the set of semantic relations. Last but not least, we achieve state-of-the-art performance with a much simpler algorithm that is based on the notion of vertex degree in a graph. 6The differences between the results in bold in each column of the table are not statistically significant at p < 0.05. Algorithm Sports Finance P/R/F1 P/R/F1 k-NN † 30.3 43.4 Static PR † 20.1 39.6 Personalized PR † 35.6 46.9 ExtLesk 40.1 45.6 Degree 42.0 47.8 MFS BL 19.6 37.1 Random BL 19.5 19.6 Table 4: Performance on the Sports and Finance sections of the dataset from Koeling et al. (2005): † indicates results from Agirre et al. (2009). 4.3 Domain WSD The main strength of Wikipedia is to provide wide coverage for many specific domains. Accordingly, on the Semeval dataset our system achieves the best performance on a domain-specific text, namely d004, a document on computer science where we achieve 82.9% F1 (+6.8% when compared with the best supervised system, namely NUS-PT). To test whether our performance on the Semeval dataset is an artifact of the data, i.e. d004 coming from Wikipedia itself, we evaluated our system on the Sports and Finance sections of the domain corpora from Koeling et al. (2005). In Table 4 we report our results on these datasets and compare them with Personalized PageRank, the state-of-the-art system from Agirre et al. (2009)7, as well as Static PageRank and a k-NN supervised WSD system trained on SemCor. The results we obtain on the two domains with our best configuration (Degree using WordNet++) outperform by a large margin k-NN, thus supporting the findings from Agirre et al. (2009) that knowledge-based systems exhibit a more robust performance than their supervised alternatives when evaluated across different domains. In addition, our system achieves better results than Static and Personalized PageRank, indicating that competitive disambiguation performance can still be achieved by a less sophisticated knowledgebased WSD algorithm when provided with a rich amount of high-quality knowledge. Finally, the results show that WordNet++ enables competitive performance also in a fine-grained domain setting. 7We compare only with those system configurations performing token-based WSD, i.e. disambiguating each instance of a target word separately, since our aim is not to perform type-based disambiguation. 1529 5 Conclusions In this paper, we have presented a large-scale method for the automatic enrichment of a computational lexicon with encyclopedic relational knowledge8. Our experiments show that the large amount of knowledge injected into WordNet is of high quality and, more importantly, it enables simple knowledge-based WSD systems to perform as well as the highest-performing supervised ones in a coarse-grained setting and to outperform them on domain-specific text. Thus, our results go one step beyond previous findings (Cuadros and Rigau, 2006; Agirre et al., 2009; Navigli and Lapata, 2010) and prove that knowledge-rich disambiguation is a competitive alternative to supervised systems, even when relying on a simple algorithm. We note, however, that the present contribution does not show which knowledge-rich algorithm performs best with WordNet++. In fact, more sophisticated approaches, such as Personalized PageRank (Agirre and Soroa, 2009), could be still applied to yield even higher performance. We leave such exploration to future work. Moreover, while the mapping has been used to enrich WordNet with a large amount of semantic edges, the method can be reversed and applied to the encyclopedic resource itself, that is Wikipedia, to perform disambiguation with the corresponding sense inventory (cf. the task of wikification proposed by Mihalcea and Csomai (2007) and Milne and Witten (2008b)). In this paper, we focused on English Word Sense Disambiguation. However, since WordNet++ is part of a multilingual semantic network (Navigli and Ponzetto, 2010), we plan to explore the impact of this knowledge in a multilingual setting. References Eneko Agirre and Oier Lopez de Lacalle. 2004. Publicly available topic signatures for all WordNet nominal senses. In Proc. of LREC ’04. Eneko Agirre and David Martinez. 2001. Learning class-to-class selectional preferences. In Proceedings of CoNLL-01, pages 15–22. Eneko Agirre and Aitor Soroa. 2009. Personalizing PageRank for Word Sense Disambiguation. In Proc. of EACL-09, pages 33–41. Eneko Agirre, Oier Lopez de Lacalle, and Aitor Soroa. 2009. Knowledge-based WSD on specific domains: 8The resulting resource, WordNet++, is freely available at http://lcl.uniroma1.it/wordnetplusplus for research purposes. performing better than generic supervised WSD. In Proc. of IJCAI-09, pages 1501–1506. Satanjeev Banerjee and Ted Pedersen. 2003. Extended gloss overlap as a measure of semantic relatedness. In Proc. of IJCAI-03, pages 805–810. Razvan Bunescu and Marius Pas¸ca. 2006. Using encyclopedic knowledge for named entity disambiguation. In Proc. of EACL-06, pages 9–16. Jean Carletta. 1996. Assessing agreement on classification tasks: The kappa statistic. Computational Linguistics, 22(2):249–254. Yee Seng Chan, Hwee Tou Ng, and Zhi Zhong. 2007. NUS-ML: Exploiting parallel texts for Word Sense Disambiguation in the English all-words tasks. In Proc. of SemEval-2007, pages 253–256. Ping Chen, Wei Ding, Chris Bowes, and David Brown. 2009. A fully unsupervised Word Sense Disambiguation method using dependency knowledge. In Proc. of NAACL-HLT-09, pages 28–36. Tim Chklovski and Rada Mihalcea. 2002. Building a sense tagged corpus with Open Mind Word Expert. In Proceedings of the ACL-02 Workshop on WSD: Recent Successes and Future Directions at ACL-02. Martin Chodorow, Roy Byrd, and George E. Heidorn. 1985. Extracting semantic hierarchies from a large on-line dictionary. In Proc. of ACL-85, pages 299– 304. Philipp Cimiano, Siegfried Handschuh, and Steffen Staab. 2004. Towards the self-annotating Web. In Proc. of WWW-04, pages 462–471. Montse Cuadros and German Rigau. 2006. Quality assessment of large scale knowledge resources. In Proc. of EMNLP-06, pages 534–541. Montse Cuadros and German Rigau. 2008. KnowNet: building a large net of knowledge from the Web. In Proc. of COLING-08, pages 161–168. Philip Edmonds. 2000. Designing a task for SENSEVAL-2. Technical report, University of Brighton, U.K. Christiane Fellbaum, editor. 1998. WordNet: An Electronic Database. MIT Press, Cambridge, MA. Evgeniy Gabrilovich and Shaul Markovitch. 2006. Overcoming the brittleness bottleneck using Wikipedia: Enhancing text categorization with encyclopedic knowledge. In Proc. of AAAI-06, pages 1301–1306. Evgeniy Gabrilovich and Shaul Markovitch. 2007. Computing semantic relatedness using Wikipediabased explicit semantic analysis. In Proc. of IJCAI07, pages 1606–1611. Roxana Girju, Adriana Badulescu, and Dan Moldovan. 2006. Automatic discovery of part-whole relations. Computational Linguistics, 32(1):83–135. Sanda M. Harabagiu, George A. Miller, and Dan I. Moldovan. 1999. WordNet 2 – a morphologically and semantically enhanced resource. In Proceedings of the SIGLEX99 Workshop on Standardizing Lexical Resources, pages 1–8. 1530 Marti A. Hearst. 1992. Automatic acquisition of hyponyms from large text corpora. In Proc. of COLING-92, pages 539–545. Adam Kilgarriff and Joseph Rosenzweig. 2000. Framework and results for English SENSEVAL. Computers and the Humanities, 34(1-2). Rob Koeling and Diana McCarthy. 2007. Sussx: WSD using automatically acquired predominant senses. In Proc. of SemEval-2007, pages 314–317. Rob Koeling, Diana McCarthy, and John Carroll. 2005. Domain-specific sense distributions and predominant sense acquisition. In Proc. of HLTEMNLP-05, pages 419–426. Michael Lesk. 1986. Automatic sense disambiguation using machine readable dictionaries: How to tell a pine cone from an ice cream cone. In Proceedings of the 5th Annual Conference on Systems Documentation, Toronto, Ontario, Canada, pages 24–26. Diana McCarthy and John Carroll. 2003. Disambiguating nouns, verbs and adjectives using automatically acquired selectional preferences. Computational Linguistics, 29(4):639–654. Rada Mihalcea and Andras Csomai. 2007. Wikify! Linking documents to encyclopedic knowledge. In Proc. of CIKM-07, pages 233–242. Rada Mihalcea. 2007. Using Wikipedia for automatic Word Sense Disambiguation. In Proc. of NAACLHLT-07, pages 196–203. George A. Miller, Claudia Leacock, Randee Tengi, and Ross Bunker. 1993. A semantic concordance. In Proceedings of the 3rd DARPA Workshop on Human Language Technology, pages 303–308, Plainsboro, N.J. David Milne and Ian H. Witten. 2008a. An effective, low-cost measure of semantic relatedness obtained from Wikipedia links. In Proceedings of the Workshop on Wikipedia and Artificial Intelligence: An Evolving Synergy at AAAI-08, pages 25–30. David Milne and Ian H. Witten. 2008b. Learning to link with Wikipedia. In Proc. of CIKM-08, pages 509–518. Vivi Nastase and Michael Strube. 2008. Decoding Wikipedia category names for knowledge acquisition. In Proc. of AAAI-08, pages 1219–1224. Vivi Nastase. 2008. Topic-driven multi-document summarization with encyclopedic knowledge and activation spreading. In Proc. of EMNLP-08, pages 763–772. Roberto Navigli and Mirella Lapata. 2010. An experimental study on graph connectivity for unsupervised Word Sense Disambiguation. IEEE Transactions on Pattern Anaylsis and Machine Intelligence, 32(4):678–692. Roberto Navigli and Simone Paolo Ponzetto. 2010. BabelNet: Building a very large multilingual semantic network. In Proc. of ACL-10. Roberto Navigli and Paola Velardi. 2005. Structural Semantic Interconnections: a knowledge-based approach to Word Sense Disambiguation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(7):1075–1088. Roberto Navigli, Kenneth C. Litkowski, and Orin Hargraves. 2007. Semeval-2007 task 07: Coarsegrained English all-words task. In Proc. of SemEval2007, pages 30–35. Roberto Navigli. 2009a. Using cycles and quasicycles to disambiguate dictionary glosses. In Proc. of EACL-09, pages 594–602. Roberto Navigli. 2009b. Word Sense Disambiguation: A survey. ACM Computing Surveys, 41(2):1–69. Marco Pennacchiotti and Patrick Pantel. 2006. Ontologizing semantic relations. In Proc. of COLINGACL-06, pages 793–800. Simone Paolo Ponzetto and Roberto Navigli. 2009. Large-scale taxonomy mapping for restructuring and integrating Wikipedia. In Proc. of IJCAI-09, pages 2083–2088. Simone Paolo Ponzetto and Michael Strube. 2007a. Deriving a large scale taxonomy from Wikipedia. In Proc. of AAAI-07, pages 1440–1445. Simone Paolo Ponzetto and Michael Strube. 2007b. Knowledge derived from Wikipedia for computing semantic relatedness. Journal of Artificial Intelligence Research, 30:181–212. Nils Reiter, Matthias Hartung, and Anette Frank. 2008. A resource-poor approach for linking ontology classes to Wikipedia articles. In Johan Bos and Rodolfo Delmonte, editors, Semantics in Text Processing, volume 1 of Research in Computational Semantics, pages 381–387. College Publications, London, England. German Rigau, Horacio Rodr´ıguez, and Eneko Agirre. 1998. Building accurate semantic taxonomies from monolingual MRDs. In Proc. of COLING-ACL-98, pages 1103–1109. Maria Ruiz-Casado, Enrique Alfonseca, and Pablo Castells. 2005. Automatic assignment of Wikipedia encyclopedic entries to WordNet synsets. In Advances in Web Intelligence, volume 3528 of Lecture Notes in Computer Science. Springer Verlag. Christina Sauper and Regina Barzilay. 2009. Automatically generating Wikipedia articles: A structureaware approach. In Proc. of ACL-IJCNLP-09, pages 208–216. Eyal Shnarch, Libby Barak, and Ido Dagan. 2009. Extracting lexical reference rules from Wikipedia. In Proc. of ACL-IJCNLP-09, pages 450–458. Rion Snow, Dan Jurafsky, and Andrew Ng. 2006. Semantic taxonomy induction from heterogeneous evidence. In Proc. of COLING-ACL-06, pages 801– 808. Fabian M. Suchanek, Gjergji Kasneci, and Gerhard Weikum. 2008. Yago: A large ontology from Wikipedia and WordNet. Journal of Web Semantics, 6(3):203–217. Fei Wu and Daniel Weld. 2007. Automatically semantifying Wikipedia. In Proc. of CIKM-07, pages 41–50. Fei Wu and Daniel Weld. 2008. Automatically refining the Wikipedia infobox ontology. In Proc. of WWW08, pages 635–644. 1531
|
2010
|
154
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1532–1541, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics All Words Domain Adapted WSD: Finding a Middle Ground between Supervision and Unsupervision Mitesh M. Khapra Anup Kulkarni Saurabh Sohoney Pushpak Bhattacharyya Indian Institute of Technology Bombay, Mumbai - 400076, India. {miteshk,anup,saurabhsohoney,pb}@cse.iitb.ac.in Abstract In spite of decades of research on word sense disambiguation (WSD), all-words general purpose WSD has remained a distant goal. Many supervised WSD systems have been built, but the effort of creating the training corpus - annotated sense marked corpora - has always been a matter of concern. Therefore, attempts have been made to develop unsupervised and knowledge based techniques for WSD which do not need sense marked corpora. However such approaches have not proved effective, since they typically do not better Wordnet first sense baseline accuracy. Our research reported here proposes to stick to the supervised approach, but with far less demand on annotation. We show that if we have ANY sense marked corpora, be it from mixed domain or a specific domain, a small amount of annotation in ANY other domain can deliver the goods almost as if exhaustive sense marking were available in that domain. We have tested our approach across Tourism and Health domain corpora, using also the well known mixed domain SemCor corpus. Accuracy figures close to self domain training lend credence to the viability of our approach. Our contribution thus lies in finding a convenient middle ground between pure supervised and pure unsupervised WSD. Finally, our approach is not restricted to any specific set of target words, a departure from a commonly observed practice in domain specific WSD. 1 Introduction Amongst annotation tasks, sense marking surely takes the cake, demanding as it does high level of language competence, topic comprehension and domain sensitivity. This makes supervised approaches to WSD a difficult proposition (Agirre et al., 2009b; Agirre et al., 2009a; McCarthy et al., 2007). Unsupervised and knowledge based approaches have been tried with the hope of creating WSD systems with no need for sense marked corpora (Koeling et al., 2005; McCarthy et al., 2007; Agirre et al., 2009b). However, the accuracy figures of such systems are low. Our work here is motivated by the desire to develop annotation-lean all-words domain adapted techniques for supervised WSD. It is a common observation that domain specific WSD exhibits high level of accuracy even for the all-words scenario (Khapra et al., 2010) - provided training and testing are on the same domain. Also domain adaptation - in which training happens in one domain and testing in another - often is able to attain good levels of performance, albeit on a specific set of target words (Chan and Ng, 2007; Agirre and de Lacalle, 2009). To the best of our knowledge there does not exist a system that solves the combined problem of all words domain adapted WSD. We thus propose the following: a. For any target domain, create a small amount of sense annotated corpus. b. Mix it with an existing sense annotated corpus – from a mixed domain or specific domain – to train the WSD engine. This procedure tested on four adaptation scenarios, viz., (i) SemCor (Miller et al., 1993) to Tourism, (ii) SemCor to Health, (iii) Tourism to Health and (iv) Health to Tourism has consistently yielded good performance (to be explained in sections 6 and 7). The remainder of this paper is organized as follows. In section 2 we discuss previous work in the area of domain adaptation for WSD. In section 3 1532 we discuss three state of art supervised, unsupervised and knowledge based algorithms for WSD. Section 4 discusses the injection strategy for domain adaptation. In section 5 we describe the dataset used for our experiments. We then present the results in section 6 followed by discussions in section 7. Section 8 examines whether there is any need for intelligent choice of injections. Section 9 concludes the paper highlighting possible future directions. 2 Related Work Domain specific WSD for selected target words has been attempted by Ng and Lee (1996), Agirre and de Lacalle (2009), Chan and Ng (2007), Koeling et al. (2005) and Agirre et al. (2009b). They report results on three publicly available lexical sample datasets, viz., DSO corpus (Ng and Lee, 1996), MEDLINE corpus (Weeber et al., 2001) and the corpus made available by Koeling et al. (2005). Each of these datasets contains a handful of target words (41-191 words) which are sense marked in the corpus. Our main inspiration comes from the targetword specific results reported by Chan and Ng (2007) and Agirre and de Lacalle (2009). The former showed that adding just 30% of the target data to the source data achieved the same performance as that obtained by taking the entire source and target data. Agirre and de Lacalle (2009) reported a 22% error reduction when source and target data were combined for training a classifier, as compared to the case when only the target data was used for training the classifier. However, both these works focused on target word specific WSD and do not address all-words domain specific WSD. In the unsupervised setting, McCarthy et al. (2007) showed that their predominant sense acquisition method gives good results on the corpus of Koeling et al. (2005). In particular, they showed that the performance of their method is comparable to the most frequent sense obtained from a tagged corpus, thereby making a strong case for unsupervised methods for domain-specific WSD. More recently, Agirre et al. (2009b) showed that knowledge based approaches which rely only on the semantic relations captured by the Wordnet graph outperform supervised approaches when applied to specific domains. The good results obtained by McCarthy et al. (2007) and Agirre et al. (2009b) for unsupervised and knowledge based approaches respectively have cast a doubt on the viability of supervised approaches which rely on sense tagged corpora. However, these conclusions were drawn only from the performance on certain target words, leaving open the question of their utility in all words WSD. We believe our work contributes to the WSD research in the following way: (i) it shows that there is promise in supervised approach to allword WSD, through the instrument of domain adaptation; (ii) it places in perspective some very recently reported unsupervised and knowledge based techniques of WSD; (ii) it answers some questions arising out of the debate between supervision and unsupervision in WSD; and finally (iv) it explores a convenient middle ground between unsupervised and supervised WSD – the territory of “annotate-little and inject” paradigm. 3 WSD algorithms employed by us In this section we describe the knowledge based, unsupervised and supervised approaches used for our experiments. 3.1 Knowledge Based Approach Agirre et al. (2009b) showed that a graph based algorithm which uses only the relations between concepts in a Lexical Knowledge Base (LKB) can outperform supervised approaches when tested on specific domains (for a set of chosen target words). We employ their method which involves the following steps: 1. Represent Wordnet as a graph where the concepts (i.e., synsets) act as nodes and the relations between concepts define edges in the graph. 2. Apply a context-dependent Personalized PageRank algorithm on this graph by introducing the context words as nodes into the graph and linking them with their respective synsets. 3. These nodes corresponding to the context words then inject probability mass into the synsets they are linked to, thereby influencing the final relevance of all nodes in the graph. We used the publicly available implementation of this algorithm1 for our experiments. 1http://ixa2.si.ehu.es/ukb/ 1533 3.2 Unsupervised Approach McCarthy et al. (2007) used an untagged corpus to construct a thesaurus of related words. They then found the predominant sense (i.e., the most frequent sense) of each target word using pair-wise Wordnet based similarity measures by pairing the target word with its top-k neighbors in the thesaurus. Each target word is then disambiguated by assigning it its predominant sense – the motivation being that the predominant sense is a powerful, hard-to-beat baseline. We implemented their method using the following steps: 1. Obtain a domain-specific untagged corpus (we crawled a corpus of approximately 9M words from the web). 2. Extract grammatical relations from this text using a dependency parser2 (Klein and Manning, 2003). 3. Use the grammatical relations thus extracted to construct features for identifying the k nearest neighbors for each word using the distributional similarity score described in (Lin, 1998). 4. Rank the senses of each target word in the test set using a weighted sum of the distributional similarity scores of the neighbors. The weights in the sum are based on Wordnet Similarity scores (Patwardhan and Pedersen, 2003). 5. Each target word in the test set is then disambiguated by simply assigning it its predominant sense obtained using the above method. 3.3 Supervised approach Khapra et al. (2010) proposed a supervised algorithm for domain-specific WSD and showed that it beats the most frequent corpus sense and performs on par with other state of the art algorithms like PageRank. We implemented their iterative algorithm which involves the following steps: 1. Tag all monosemous words in the sentence. 2. Iteratively disambiguate the remaining words in the sentence in increasing order of their degree of polysemy. 3. At each stage rank the candidate senses of a word using the scoring function of Equation (1) which combines corpus based parameters (such as, sense distributions and corpus co-occurrence) and Wordnet based parameters 2We used the Stanford parser http://nlp. stanford.edu/software/lex-parser.shtml (such as, semantic similarity, conceptual distance, etc.) S∗= arg max i (θiVi + X j∈J Wij ∗Vi ∗Vj) (1) where, i ∈Candidate Synsets J = Set of disambiguated words θi = BelongingnessToDominantConcept(Si) Vi = P(Si|word) Wij = CorpusCooccurrence(Si, Sj) ∗1/WNConceptualDistance(Si, Sj) ∗1/WNSemanticGraphDistance(Si, Sj) 4. Select the candidate synset with maximizes the above score as the winner sense. 4 Injections for Supervised Adaptation This section describes the main interest of our work i.e. adaptation using injections. For supervised adaptation, we use the supervised algorithm described above (Khapra et al., 2010) in the following 3 settings as proposed by Agirre et al. (2009a): a. Source setting: We train the algorithm on a mixed-domain corpus (SemCor) or a domainspecific corpus (say, Tourism) and test it on a different domain (say, Health). A good performance in this setting would indicate robustness to domain-shifts. b. Target setting: We train and test the algorithm using data from the same domain. This gives the skyline performance, i.e., the best performance that can be achieved if sense marked data from the target domain were available. c. Adaptation setting: This setting is the main focus of interest in the paper. We augment the training data which could be from one domain or mixed domain with a small amount of data from the target domain. This combined data is then used for training. The aim here is to reach as close to the skyline performance using as little data as possible. For injecting data from the target domain we randomly select some sense marked words from the target domain and add 1534 Polysemous words Monosemous words Category Tourism Health Tourism Health Noun 53133 15437 23665 6979 Verb 15528 7348 1027 356 Adjective 19732 5877 10569 2378 Adverb 6091 1977 4323 1694 All 94484 30639 39611 11407 Avg. no. of instances perpolysemous word Category Health Tourism SemCor Noun 7.06 12.56 10.98 Verb 7.47 9.76 11.95 Adjective 5.74 12.07 8.67 Adverb 9.11 19.78 25.44 All 6.94 12.17 11.25 Table 1: Polysemous and Monosemous words per category in each domain Table 2: Average number of instances per polysemous word per category in the 3 domains Avg. degree of Wordnet polysemy for polysemous words Category Health Tourism SemCor Noun 5.24 4.95 5.60 Verb 10.60 10.10 9.89 Adjective 5.52 5.08 5.40 Adverb 3.64 4.16 3.90 All 6.49 5.77 6.43 Avg. degree of Corpus polysemy for polysemous words Category Health Tourism SemCor Noun 1.92 2.60 3.41 Verb 3.41 4.55 4.73 Adjective 2.04 2.57 2.65 Adverb 2.16 2.82 3.09 All 2.31 2.93 3.56 Table 3: Average degree of Wordnet polysemy of polysemous words per category in the 3 domains Table 4: Average degree of Corpus polysemy of polysemous words per category in the 3 domains them to the training data. An obvious question which arises at this point is “Why were the words selected at random?” or “Can selection of words using some active learning strategy yield better results than a random selection?” We discuss this question in detail in Section 7 and show that a random set of injections performs no worse than a craftily selected set of injections. 5 DataSet Preparation Due to the lack of any publicly available all-words domain specific sense marked corpora we set upon the task of collecting data from two domains, viz., Tourism and Health. The data for Tourism domain was downloaded from Indian Tourism websites whereas the data for Health domain was obtained from two doctors. This data was manually sense annotated by two lexicographers adept in English. Princeton Wordnet 2.13 (Fellbaum, 1998) was used as the sense inventory. A total of 1,34,095 words from the Tourism domain and 42,046 words from the Health domain were manually sense marked. Some files were sense marked by both the lexicographers and the Inter Tagger Agreement (ITA) calculated from these files was 83% which is comparable to the 78% ITA reported on the SemCor corpus considering the domainspecific nature of the corpus. We now present different statistics about the corpora. Table 1 summarizes the number of polysemous and monosemous words in each category. 3http://wordnetweb.princeton.edu/perl/webwn Note that we do not use the monosemous words while calculating precision and recall of our algorithms. Table 2 shows the average number of instances per polysemous word in the 3 corpora. We note that the number of instances per word in the Tourism domain is comparable to that in the SemCor corpus whereas the number of instances per word in the Health corpus is smaller due to the overall smaller size of the Health corpus. Tables 3 and 4 summarize the average degree of Wordnet polysemy and corpus polysemy of the polysemous words in the corpus. Wordnet polysemy is the number of senses of a word as listed in the Wordnet, whereas corpus polysemy is the number of senses of a word actually appearing in the corpus. As expected, the average degree of corpus polysemy (Table 4) is much less than the average degree of Wordnet polysemy (Table 3). Further, the average degree of corpus polysemy (Table 4) in the two domains is less than that in the mixed-domain SemCor corpus, which is expected due to the domain specific nature of the corpora. Finally, Table 5 summarizes the number of unique polysemous words per category in each domain. No. of unique polysemous words Category Health Tourism SemCor Noun 2188 4229 5871 Verb 984 1591 2565 Adjective 1024 1635 2640 Adverb 217 308 463 All 4413 7763 11539 Table 5: Number of unique polysemous words per category in each domain. 1535 The data is currently being enhanced by manually sense marking more words from each domain and will be soon freely available4 for research purposes. 6 Results We tested the 3 algorithms described in section 4 using SemCor, Tourism and Health domain corpora. We did a 2-fold cross validation for supervised adaptation and report the average performance over the two folds. Since the knowledge based and unsupervised methods do not need any training data we simply test it on the entire corpus from the two domains. 6.1 Knowledge Based approach The results obtained by applying the Personalized PageRank (PPR) method to Tourism and Health data are summarized in Table 6. We also report the Wordnet first sense baseline (WFS). Domain Algorithm P(%) R(%) F(%) Tourism PPR 53.1 53.1 53.1 WFS 62.5 62.5 62.5 Health PPR 51.1 51.1 51.1 WFS 65.5 65.5 65.5 Table 6: Comparing the performance of Personalized PageRank (PPR) with Wordnet First Sense Baseline (WFS) 6.2 Unsupervised approach The predominant sense for each word in the two domains was calculated using the method described in section 4.2. McCarthy et al. (2004) reported that the best results were obtained using k = 50 neighbors and the Wordnet Similarity jcn measure (Jiang and Conrath, 1997). Following them, we used k = 50 and observed that the best results for nouns and verbs were obtained using the jcn measure and the best results for adjectives and adverbs were obtained using the lesk measure (Banerjee and Pedersen, 2002). Accordingly, we used jcn for nouns and verbs and lesk for adjectives and adverbs. Each target word in the test set is then disambiguated by simply assigning it its predominant sense obtained using the above method. We tested this approach only on Tourism domain due to unavailability of large 4http://www.cfilt.iitb.ac.in/wsd/annotated corpus untagged Health corpus which is needed for constructing the thesaurus. The results are summarized in Table 7. Domain Algorithm P(%) R(%) F(%) Tourism McCarthy 51.85 49.32 50.55 WFS 62.50 62.50 62.50 Table 7: Comparing the performance of unsupervised approach with Wordnet First Sense Baseline (WFS) 6.3 Supervised adaptation We report results in the source setting, target setting and adaptation setting as described earlier using the following four combinations for source and target data: 1. SemCor to Tourism (SC→T) where SemCor is used as the source domain and Tourism as the target (test) domain. 2. SemCor to Health (SC→H) where SemCor is used as the source domain and Health as the target (test) domain. 3. Tourism to Health (T→H) where Tourism is used as the source domain and Health as the target (test) domain. 4. Health to Tourism (H→T) where Health is used as the source domain and Tourism as the target (test) domain. In each case, the target domain data was divided into two folds. One fold was set aside for testing and the other for injecting data in the adaptation setting. We increased the size of the injected target examples from 1000 to 14000 words in increments of 1000. We then repeated the same experiment by reversing the role of the two folds. Figures 1, 2, 3 and 4 show the graphs of the average F-score over the 2-folds for SC→T, SC→H, T→H and H→T respectively. The x-axis represents the amount of training data (in words) injected from the target domain and the y-axis represents the F-score. The different curves in each graph are as follows: a. only random : This curve plots the performance obtained using x randomly selected sense tagged words from the target domain and zero sense tagged words from the source domain (x was varied from 1000 to 14000 words in increments of 1000). 1536 35 40 45 50 55 60 65 70 75 80 0 2000 4000 6000 8000 10000 12000 14000 F-score (%) Injection Size (words) Injection Size v/s F-score wfs srcb tsky only_random random+semcor 35 40 45 50 55 60 65 70 75 80 0 2000 4000 6000 8000 10000 12000 14000 F-score (%) Injection Size (words) Injection Size v/s F-score wfs srcb tsky only_random random+semcor Figure 1: Supervised adaptation from SemCor to Tourism using injections Figure 2: Supervised adaptation from SemCor to Health using injections 35 40 45 50 55 60 65 70 75 80 0 2000 4000 6000 8000 10000 12000 14000 F-score (%) Injection Size (words) Injection Size v/s F-score wfs srcb tsky only_random random+tourism 35 40 45 50 55 60 65 70 75 80 0 2000 4000 6000 8000 10000 12000 14000 F-score (%) Injection Size (words) Injection Size v/s F-score wfs srcb tsky only_random random+health Figure 3: Supervised adaptation from Tourism to Health using injections Figure 4: Supervised adaptation from Health to Tourism using injections b. random+source : This curve plots the performance obtained by mixing x randomly selected sense tagged words from the target domain with the entire training data from the source domain (again x was varied from 1000 to 14000 words in increments of 1000). c. source baseline (srcb) : This represents the Fscore obtained by training on the source data alone without mixing any examples from the target domain. d. wordnet first sense (wfs) : This represents the F-score obtained by selecting the first sense from Wordnet, a typically reported baseline. e. target skyline (tsky) : This represents the average 2-fold F-score obtained by training on one entire fold of the target data itself (Health: 15320 polysemous words; Tourism: 47242 polysemous words) and testing on the other fold. These graphs along with other results are discussed in the next section. 7 Discussions We discuss the performance of the three approaches. 7.1 Knowledge Based and Unsupervised approaches It is apparent from Tables 6 and 7 that knowledge based and unsupervised approaches do not perform well when compared to the Wordnet first sense (which is freely available and hence can be used for disambiguation). Further, we observe that the performance of these approaches is even less than the source baseline (i.e., the case when training data from a source domain is applied as it is to a target domain - without using any injections). These observations bring out the weaknesses of these approaches when used in an all-words setting and clearly indicate that they come nowhere close to replacing a supervised system. 1537 7.2 Supervised adaptation 1. The F-score obtained by training on SemCor (mixed-domain corpus) and testing on the two target domains without using any injections (srcb) – F-score of 61.7% on Tourism and Fscore of 65.5% on Health – is comparable to the best result reported on the SEMEVAL datasets (65.02%, where both training and testing happens on a mixed-domain corpus (Snyder and Palmer, 2004)). This is in contrast to previous studies (Escudero et al., 2000; Agirre and Martinez, 2004) which suggest that instead of adapting from a generic/mixed domain to a specific domain, it is better to completely ignore the generic examples and use hand-tagged data from the target domain itself. The main reason for the contrasting results is that the earlier work focused only on a handful of target words whereas we focus on all words appearing in the corpus. So, while the behavior of a few target words would change drastically when the domain changes, a majority of the words will exhibit the same behavior (i.e., same predominant sense) even when the domain changes. We agree that the overall performance is still lower than that obtained by training on the domainspecific corpora. However, it is still better than the performance of unsupervised and knowledge based approaches which tilts the scale in favor of supervised approaches even when only mixed domain sense marked corpora is available. 2. Adding injections from the target domain improves the performance. As the amount of injection increases the performance approaches the skyline, and in the case of SC→H and T→H it even crosses the skyline performance showing that combining the source and target data can give better performance than using the target data alone. This is consistent with the domain adaptation results reported by Agirre and de Lacalle (2009) on a specific set of target words. 3. The performance of random+source is always better than only random indicating that the data from the source domain does help to improve performance. A detailed analysis showed that the gain obtained by using the source data is attributable to reducing recall errors by increasing the coverage of seen words. 4. Adapting from one specific domain (Tourism or Health) to another specific domain (Health or Tourism) gives the same performance as that obtained by adapting from a mixed-domain (SemCor) to a specific domain (Tourism, Health). This is an interesting observation as it suggests that as long as data from one domain is available it is easy to build a WSD engine that works for other domains by injecting a small amount of data from these domains. To verify that the results are consistent, we randomly selected 5 different sets of injections from fold-1 and tested the performance on fold-2. We then repeated the same experiment by reversing the roles of the two folds. The results were indeed consistent irrespective of the set of injections used. Due to lack of space we have not included the results for these 5 different sets of injections. 7.3 Quantifying the trade-off between performance and corpus size To correctly quantify the benefit of adding injections from the target domain, we calculated the amount of target data (peak size) that is needed to reach the skyline F-score (peak F) in the absence of any data from the source domain. The peak size was found to be 35000 (Tourism) and 14000 (Health) corresponding to peak F values of 74.2% (Tourism) and 73.4% (Health). We then plotted a graph (Figure 5) to capture the relation between the size of injections (expressed as a percentage of the peak size) and the F-score (expressed as a percentage of the peak F). 80 85 90 95 100 105 0 20 40 60 80 100 % peak_F % peak_size Size v/s Performance SC --> H T --> H SC --> T H --> T Figure 5: Trade-off between performance and corpus size We observe that by mixing only 20-40% of the peak size with the source domain we can obtain up to 95% of the performance obtained by using the 1538 entire target data (peak size). In absolute terms, the size of the injections is only 7000-9000 polysemous words which is a very small price to pay considering the performance benefits. 8 Does the choice of injections matter? An obvious question which arises at this point is “Why were the words selected at random?” or “Can selection of words using some active learning strategy yield better results than a random selection?” An answer to this question requires a more thorough understanding of the sensebehavior exhibited by words across domains. In any scenario involving a shift from domain D1 to domain D2, we will always encounter words belonging to the following 4 categories: a. WD1 : This class includes words which are encountered only in the source domain D1 and do not appear in the target domain D2. Since we are interested in adapting to the target domain and since these words do not appear in the target domain, it is quite obvious that they are not important for the problem of domain adaptation. b. WD2 : This class includes words which are encountered only in the target domain D2 and do not appear in the source domain D1. Again, it is quite obvious that these words are important for the problem of domain adaptation. They fall in the category of unseen words and need handling from that point of view. c. WD1D2conformists : This class includes words which are encountered in both the domains and exhibit the same predominant sense in both the domains. Correct identification of these words is important so that we can use the predominant sense learned from D1 for disambiguating instances of these words appearing in D2. d. WD1D2non−conformists : This class includes words which are encountered in both the domains but their predominant sense in the target domain D2 does not conform to the predominant sense learned from the source domain D1. Correct identification of these words is important so that we can ignore the predominant senses learned from D1 while disambiguating instances of these words appearing in D2. Table 8 summarizes the percentage of words that fall in each category in each of the three adaptation scenarios. The fact that nearly 50-60% of the words fall in the “conformist” category once again makes a strong case for reusing sense tagged data from one domain to another domain. Category SC→T SC→H T→H WD2 7.14% 5.45% 13.61% Conformists 49.54% 60.43% 54.31% Non-Conformists 43.30% 34.11% 32.06% Table 8: Percentage of Words belonging to each category in the three settings. The above characterization suggests that an ideal domain adaptation strategy should focus on injecting WD2 and WD1D2non−conformists as these would yield maximum benefits if injected into the training data. While it is easy to identify the WD2 words, “identifying non-conformists” is a hard problem which itself requires some type of WSD5. However, just to prove that a random injection strategy does as good as an ideal strategy we assume the presence of an oracle which identifies the WD1D2non−conformists. We then augment the training data with 5-8 instances for WD2 and WD1D2non−conformists words thus identified. We observed that adding more than 5-8 instances per word does not improve the performance. This is due to the “one sense per domain” phenomenon – seeing only a few instances of a word is sufficient to identify the predominant sense of the word. Further, to ensure a better overall performance, the instances of the most frequent words are injected first followed by less frequent words till we exhaust the total size of the injections (1000, 2000 and so on). We observed that there was a 7580% overlap between the words selected by random strategy and oracle strategy. This is because oracle selects the most frequent words which also have a high chance of getting selected when a random sampling is done. Figures 6, 7, 8 and 9 compare the performance of the two strategies. We see that the random strategy does as well as the oracle strategy thereby supporting our claim that if we have sense marked corpus from one domain then simply injecting ANY small amount of data from the target domain will 5Note that the unsupervised predominant sense acquisition method of McCarthy et al. (2007) implicitly identifies conformists and non-conformists 1539 35 40 45 50 55 60 65 70 75 80 0 2000 4000 6000 8000 10000 12000 14000 F-score (%) Injection Size (words) Injection Size v/s F-score wfs srcb tsky random+semcor oracle+semcor 35 40 45 50 55 60 65 70 75 80 0 2000 4000 6000 8000 10000 12000 14000 F-score (%) Injection Size (words) Injection Size v/s F-score wfs srcb tsky random+semcor oracle+semcor Figure 6: Comparing random strategy with oracle based ideal strategy for SemCor to Tourism adaptation Figure 7: Comparing random strategy with oracle based ideal strategy for SemCor to Health adaptation 35 40 45 50 55 60 65 70 75 80 0 2000 4000 6000 8000 10000 12000 14000 F-score (%) Injection Size (words) Injection Size v/s F-score wfs srcb tsky random+tourism oracle+tourism 35 40 45 50 55 60 65 70 75 80 0 2000 4000 6000 8000 10000 12000 14000 F-score (%) Injection Size (words) Injection Size v/s F-score wfs srcb tsky random+health oracle+health Figure 8: Comparing random strategy with oracle based ideal strategy for Tourism to Health adaptation Figure 9: Comparing random strategy with oracle based ideal strategy for Health to Tourism adaptation do the job. 9 Conclusion and Future Work Based on our study of WSD in 4 domain adaptation scenarios, we make the following conclusions: 1. Supervised adaptation by mixing small amount of data (7000-9000 words) from the target domain with the source domain gives nearly the same performance (F-score of around 70% in all the 4 adaptation scenarios) as that obtained by training on the entire target domain data. 2. Unsupervised and knowledge based approaches which use distributional similarity and Wordnet based similarity measures do not compare well with the Wordnet first sense baseline performance and do not come anywhere close to the performance of supervised adaptation. 3. Supervised adaptation from a mixed domain to a specific domain gives the same performance as that from one specific domain (Tourism) to another specific domain (Health). 4. Supervised adaptation is not sensitive to the type of data being injected. This is an interesting finding with the following implication: as long as one has sense marked corpus - be it from a mixed or specific domain - simply injecting ANY small amount of data from the target domain suffices to beget good accuracy. As future work, we would like to test our work on the Environment domain data which was released as part of the SEMEVAL 2010 shared task on “Allwords Word Sense Disambiguation on a Specific Domain”. 1540 References Eneko Agirre and Oier Lopez de Lacalle. 2009. Supervised domain adaption for wsd. In EACL ’09: Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics, pages 42–50, Morristown, NJ, USA. Association for Computational Linguistics. Eneko Agirre and David Martinez. 2004. The effect of bias on an automatically-built word sense corpus. In Proceedings of the 4rd International Conference on Languages Resources and Evaluations (LREC). Eneko Agirre, Oier Lopez de Lacalle, Christiane Fellbaum, Andrea Marchetti, Antonio Toral, and Piek Vossen. 2009a. Semeval-2010 task 17: all-words word sense disambiguation on a specific domain. In DEW ’09: Proceedings of the Workshop on Semantic Evaluations: Recent Achievements and Future Directions, pages 123–128, Morristown, NJ, USA. Association for Computational Linguistics. Eneko Agirre, Oier Lopez De Lacalle, and Aitor Soroa. 2009b. Knowledge-based wsd on specific domains: Performing better than generic supervised wsd. In In Proceedings of IJCAI. Satanjeev Banerjee and Ted Pedersen. 2002. An adapted lesk algorithm for word sense disambiguation using wordnet. In CICLing ’02: Proceedings of the Third International Conference on Computational Linguistics and Intelligent Text Processing, pages 136–145, London, UK. Springer-Verlag. Yee Seng Chan and Hwee Tou Ng. 2007. Domain adaptation with active learning for word sense disambiguation. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 49–56, Prague, Czech Republic, June. Association for Computational Linguistics. Gerard Escudero, Llu´ıs M`arquez, and German Rigau. 2000. An empirical study of the domain dependence of supervised word sense disambiguation systems. In Proceedings of the 2000 Joint SIGDAT conference on Empirical methods in natural language processing and very large corpora, pages 172–180, Morristown, NJ, USA. Association for Computational Linguistics. C. Fellbaum. 1998. WordNet: An Electronic Lexical Database. J.J. Jiang and D.W. Conrath. 1997. Semantic similarity based on corpus statistics and lexical taxonomy. In Proc. of the Int’l. Conf. on Research in Computational Linguistics, pages 19–33. Mitesh Khapra, Sapan Shah, Piyush Kedia, and Pushpak Bhattacharyya. 2010. Domain-specific word sense disambiguation combining corpus based and wordnet based parameters. In 5th International Conference on Global Wordnet (GWC2010). Dan Klein and Christopher D. Manning. 2003. Accurate unlexicalized parsing. In IN PROCEEDINGS OF THE 41ST ANNUAL MEETING OF THE ASSOCIATION FOR COMPUTATIONAL LINGUISTICS, pages 423–430. Rob Koeling, Diana McCarthy, and John Carroll. 2005. Domain-specific sense distributions and predominant sense acquisition. In HLT ’05: Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, pages 419–426, Morristown, NJ, USA. Association for Computational Linguistics. Dekang Lin. 1998. Automatic retrieval and clustering of similar words. In Proceedings of the 17th international conference on Computational linguistics, pages 768–774, Morristown, NJ, USA. Association for Computational Linguistics. Diana McCarthy, Rob Koeling, Julie Weeds, and John Carroll. 2004. Finding predominant word senses in untagged text. In ACL ’04: Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics, page 279, Morristown, NJ, USA. Association for Computational Linguistics. Diana McCarthy, Rob Koeling, Julie Weeds, and John Carroll. 2007. Unsupervised acquisition of predominant word senses. Comput. Linguist., 33(4):553– 590. George A. Miller, Claudia Leacock, Randee Tengi, and Ross T. Bunker. 1993. A semantic concordance. In HLT ’93: Proceedings of the workshop on Human Language Technology, pages 303–308, Morristown, NJ, USA. Association for Computational Linguistics. Hwee Tou Ng and Hian Beng Lee. 1996. Integrating multiple knowledge sources to disambiguate word sense: an exemplar-based approach. In Proceedings of the 34th annual meeting on Association for Computational Linguistics, pages 40–47, Morristown, NJ, USA. Association for Computational Linguistics. Siddharth Patwardhan and Ted Pedersen. 2003. The cpan wordnet::similarity package. http://search .cpan.org/ sid/wordnet-similarity/. Benjamin Snyder and Martha Palmer. 2004. The english all-words task. In Rada Mihalcea and Phil Edmonds, editors, Senseval-3: Third International Workshop on the Evaluation of Systems for the Semantic Analysis of Text, pages 41–43, Barcelona, Spain, July. Association for Computational Linguistics. Marc Weeber, James G. Mork, and Alan R. Aronson. 2001. Developing a test collection for biomedical word sense disambiguation. In In Proceedings of the AMAI Symposium, pages 746–750. 1541
|
2010
|
155
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1542–1551, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Combining Orthogonal Monolingual and Multilingual Sources of Evidence for All Words WSD Weiwei Guo Computer Science Department Columbia University New York, NY, 10115 [email protected] Mona Diab Center for Computational Learning Systems Columbia University New York, NY, 10115 [email protected] Abstract Word Sense Disambiguation remains one of the most complex problems facing computational linguists to date. In this paper we present a system that combines evidence from a monolingual WSD system together with that from a multilingual WSD system to yield state of the art performance on standard All-Words data sets. The monolingual system is based on a modification of the graph based state of the art algorithm In-Degree. The multilingual system is an improvement over an AllWords unsupervised approach, SALAAM. SALAAM exploits multilingual evidence as a means of disambiguation. In this paper, we present modifications to both of the original approaches and then their combination. We finally report the highest results obtained to date on the SENSEVAL 2 standard data set using an unsupervised method, we achieve an overall F measure of 64.58 using a voting scheme. 1 Introduction Despite advances in natural language processing (NLP), Word Sense Disambiguation (WSD) is still considered one of the most challenging problems in the field. Ever since the field’s inception, WSD has been perceived as one of the central problems in NLP. WSD is viewed as an enabling technology that could potentially have far reaching impact on NLP applications in general. We are starting to see the beginnings of a positive effect of WSD in NLP applications such as Machine Translation (Carpuat and Wu, 2007; Chan et al., 2007). Advances in WSD research in the current millennium can be attributed to several key factors: the availability of large scale computational lexical resources such as WordNets (Fellbaum, 1998; Miller, 1990), the availability of large scale corpora, the existence and dissemination of standardized data sets over the past 10 years through different testbeds such as SENSEVAL and SEMEVAL competitions,1 devising more robust computing algorithms to handle large scale data sets, and simply advancement in hardware machinery. In this paper, we address the problem of WSD of all content words in a sentence, All-Words data. In this framework, the task is to associate all tokens with their contextually relevant meaning definitions from some computational lexical resource. Our work hinges upon combining two high quality WSD systems that rely on essentially different sources of evidence. The two WSD systems are a monolingual system RelCont and a multilingual system TransCont. RelCont is an enhancement on an existing graph based algorithm, In-Degree, first described in (Navigli and Lapata, 2007). TransCont is an enhancement over an existing approach that leverages multilingual evidence through projection, SALAAM, described in detail in (Diab and Resnik, 2002). Similar to the leveraged systems, the current combined approach is unsupervised, namely it does not rely on training data from the onset. We show that by combining both sources of evidence, our approach yields the highest performance for an unsupervised system to date on standard All-Words data sets. This paper is organized as follows: Section 2 delves into the problem of WSD in more detail; Section 3 explores some of the relevant related work; in Section 4, we describe the two WSD systems in some detail emphasizing the improvements to the basic systems in addition to a description of our combination approach; we present our experimental set up and results in Section 5; we discuss the results and our overall observations with error analysis in Section 6; Finally, we con1http://www.semeval.org 1542 clude in Section 7. 2 Word Sense Disambiguation The definition of WSD has taken on several different practical meanings in recent years. In the latest SEMEVAL 2010 workshop, there are 18 tasks defined, several of which are on different languages, however we recognize the widening of the definition of the task of WSD. In addition to the traditional All-Words and Lexical Sample tasks, we note new tasks on word sense discrimination (no sense inventory needed, the different senses are merely distinguished), lexical substitution using synonyms of words as substitutes both monolingually and multilingually, as well as meaning definitions obtained from different languages namely using words in translation. Our paper is about the classical All-Words (AW) task of WSD. In this task, all content bearing words in running text are disambiguated from a static lexical resource. For example a sentence such as ‘I walked by the bank and saw many beautiful plants there.’ will have the verbs ‘walked, saw’, the nouns ‘bank, plants’, the adjectives ‘many, beautiful’, and the adverb ‘there’, be disambiguated from a standard lexical resource. Hence, using WordNet,2 ‘walked’ will be assigned the corresponding meaning definitions of: to use one’s feet to advance; to advance by steps, ‘saw’ will be assigned the meaning definition of: to perceive by sight or have the power to perceive by sight, the noun ‘bank’ will be assigned the meaning definition of: sloping land especially the slope beside a body of water, and so on. 3 Related Works Many systems over the years have been proposed for the task. A thorough review of the state of the art through the late 1990s (Ide and Veronis, 1998) and more recently in (Navigli, 2009). Several techniques have been used to tackle the problem ranging from rule based/knowledge based approaches to unsupervised and supervised machine learning techniques. To date, the best approaches that solve the AW WSD task are supervised as illustrated in the different SenseEval and SEMEVAL AW task (Palmer et al., 2001; Snyder and Palmer, 2004; Pradhan et al., 2007). In this paper, we present an unsupervised combination approach to the AW WSD problem that 2http://wordnet.princeton.edu relies on WN similarity measures in conjunction with evidence obtained through exploiting multilingual evidence. We will review the closely relevant related work on which this current investigation is based.3 4 Our Approach Our current investigation exploits two basic unsupervised approaches that perform at state-of-theart for the AW WSD task in an unsupervised setting. Crucially the two systems rely on different sources of evidence allowing them to complement each other to a large extent leading to better performance than for each system independently. Given a target content word and co-occurring contextual clues, the monolingual system RelCont attempts to assign the approporiate meaning definition to the target word. Such words by definition are semantically related words. TransCont, on the other hand, is the multilingual system. TransCont defines the notion of context in the translational space using a foreign word as a filter for defining the contextual content words for a given target word. In this multilingual setting, all the words that are mapped to (aligned with) the same orthographic form in a foreign language constitute the context. In the next subsections we describe the two approaches RelCont and TransCont in some detail, then we proceed to describe two combination methods for the two approaches: MERGE and VOTE. 4.1 Monolingual System RelCont RelCont is based on an extension of a stateof-the-art WSD approach by (Sinha and Mihalcea, 2007), henceforth (SM07). In the basic SM07 work, the authors combine different semantic similarity measures with different graph based algorithms as an extension to work in (Mihalcea, 2005). Given a sequence of words W = {w1, w2...wn}, each word wi with several senses {si1, si2...sim}. A graph G = (V,E) is defined such that there exists a vertex v for each sense. Two senses of two different words may be connected by an edge e, depending on their distance. That two senses are connected suggests they should have influence on each other, accordingly a maximum 3We acknowledge the existence of many research papers that tackled the AW WSD problem using unsupervised approaches, yet for lack of space we will not be able to review most of them. 1543 allowable distance is set. They explore 4 different graph based algorithms. The highest yielding algorithm in their work is the In-Degree algorithm combining different WN similarity measures depending on POS. They used the Jiang and Conrath (JCN) (Jiang and Conrath., 1997) similarity measure within nouns, the Leacock & Chodorow (LCH) (Leacock and Chodorow, 1998) similarity measure within verbs, and the Lesk (Lesk, 1986) similarity measure within adjectives, within adverbs, and among different POS tag pairings. They evaluate their work against the SENSEVAL 2 AW test data (SV2AW). They tune the parameters of their algorithm – namely, the normalization ratio for some of these measures – on the SENSEVAL 3 data set. They report a state-ofthe-art unsupervised system that yields an overall performance across all AW POS sets of 57.2%. In our current work, we extend the SM07 work in some interesting ways. A detailed narrative of our approach is described in (Guo and Diab, 2009). Briefly, we focus on the In-Degree graph based algorithm since it is the best performer in the SM07 work. The In-Degree algorithm presents the problem as a weighted graph with senses as nodes and the similarity between senses as weights on edges. The In-Degree of a vertex refers to the number of edges incident on that vertex. In the weighted graph, the In-Degree for each vertex is calculated by summing the weights on the edges that are incident on it. After all the In-Degree values for each sense are computed, the sense with maximum value is chosen as the final sense for that word. In this paper, we use the In-Degree algorithm while applying some modifications to the basic similarity measures exploited and the WN lexical resource tapped into. Similar to the original In-Degree algorithm, we produce a probabilistic ranked list of senses. Our modifications are described as follows: JCN for Verb-Verb Similarity In our implementation of the In-Degree algorithm, we use the JCN similarity measure for both Noun-Noun similarity calculation similar to SM07. However, different from SM07, instead of using LCH for Verb-Verb similarity, we use the JCN metric as it yields better performance in our experimentations. Expand Lesk Following the intuition in (Pedersen et al., 2005), henceforth (PEA05), we expand the basic Lesk similarity measure to take into account the glosses for all the relations for the synsets on the contextual words and compare them with the glosses of the target word senses, therefore going beyond the is-a relation. We exploit the observation that WN senses are too fine-grained, accordingly the neighbors would be slightly varied while sharing significant semantic meaning content. To find similar senses, we use the relations: hypernym, hyponym, similar attributes, similar verb group, pertinym, holonym, and meronyms.4 The algorithm assumes that the words in the input are POS tagged. In PEA05, the authors retrieve all the relevant neighbors to form a bag of words for both the target sense and the surrounding senses of the context words, they specifically focus on the Lesk similarity measure. In our current work, we employ the neighbors in a disambiguation strategy using different similarity measures one pair at a time. Our algorithm takes as input a target sense and a sense pertaining to a word in the surrounding context, and returns a sense similarity score. We do not apply the WN relations expansion to the target sense. It is only applied to the contextual word.5 For the monolingual system, we employ the same normalization values used in SM07 for the different similarity measures. Namely for the Lesk and Expand-Lesk, we use the same cut-off value of 240, accordingly, if the Lesk or Expand-Lesk similarity value returns 0 <= 240 it is converted to a real number in the interval [0,1], any similarity over 240 is by default mapped to 1. We will refer to the Expand-Lesk with this threshold as Lesk2. We also experimented with different thresholds for the Lesk and Expand-Lesk similarity measure using the SENSEVAL 3 data as a tuning set. We found that a cut-off threshold of 40 was also useful. We will refer to this variant of Expand-Lesk with a cut off threshold of 40 as Lesk3. For JCN, similar to SM07, the values are from 0.04 to 0.2, we mapped them to the interval [0,1]. We did not run any calibration studies beyond the what was reported in SM07. 4In our experiments, we varied the number of relations to employ and they all yielded relatively similar results. Hence in this paper, we report results using all the relations listed above. 5We experimented with expanding both the contextual sense and the target sense and we found that the unreliability of some of the relations is detrimental to the algorithm’s performance. Hence we decided empirically to expand only the contextual word. 1544 SemCor Expansion of WN A part of the RelCont approach relies on using the Lesk algorithm. Accordingly, the availability of glosses associated with the WN entries is extremely beneficial. Therefore, we expand the number of glosses available in WN by using the SemCor data set, thereby adding more examples to compare. The SemCor corpus is a corpus that is manually sense tagged (Miller, 1990).6 In this expansion, depending on the version of WN, we use the sense-index file in the WN Database to convert the SemCor data to the appropriate version sense annotations. We augment the sense entries for the different POS WN databases with example usages from SemCor. The augmentation is done as a look up table external to WN proper since we did not want to dabble with the WN offsets. We set a cap of 30 additional examples per synset. We used the first 30 examples with no filtering criteria. Many of the synsets had no additional examples. WN1.7.1 comprises a total of 26875 synsets, of which 25940 synsets are augmented with SemCor examples.7 4.2 Multilingual System TransCont TransCont is based on the WSD system SALAAM (Diab and Resnik, 2002), henceforth (DR02). The SALAAM system leverages word alignments from parallel corpora to perform WSD. The SALAAM algorithm exploits the word correspondence cross linguistically to tag word senses on words in running text. It relies on several underlying assumptions. The first assumption is that senses of polysemous words in one language could be lexicalized differently in other languages. For example, ‘bank’ in English would be translated as banque or rive de fleuve in French, depending on context. The other assumption is that if Language 1 (L1) words are translated to the same orthographic form in Language 2 (L2), then they share the some element of meaning, they are semantically similar.8 The SALAAM algorithm can be described as follows. Given a parallel corpus of L1-L2 that 6Using SemCor in this setting to augment WN does hint of using supervised data in the WSD process, however, since our approach does not rely on training data and SemCor is not used in our algorithm directly to tag data, but to augment a rich knowledge resource, we contend that this does not affect our system’s designation as an unsupervised system. 7Some example sentences are repeated across different synsets and POS since the SemCor data is annotated as an All-Words tagged data set. 8We implicitly make the underlying simplifying assumption that the L2 words are less ambiguous than the L1 words. is sentence and word aligned, group all the word types in L1 that map to same word in L2 creating clusters referred to as typesets. Then perform disambiguation on the typeset clusters using WN. Once senses are identified for each word in the cluster, the senses are propagated back to the original word instances in the corpus. In the SALAAM algorithm, the disambiguation step is carried out as follows: within each of these target sets consider all possible sense tags for each word and choose sense tags informed by semantic similarity with all the other words in the whole group. The algorithm is a greedy algorithm that aims at maximizing the similarity of the chosen sense across all the words in the set. The SALAAM disambiguation algorithm used the noun groupings (NounGroupings) algorithm described in DR02. The algorithm applies disambiguation within POS tag. The authors report only results on the nouns only since NounGroupings heavily exploits the hierarchy structure of the WN noun taxonomy, which does not exist for adjectives and adverbs, and is very shallow for verbs. Essentially SALAAM relies on variability in translation as it is important to have multiple words in a typeset to allow for disambiguation. In the original SALAAM system, the authors automatically translated several balanced corpora in order to render more variable data for the approach to show it’s impact. The corpora that were translated are: the WSJ, the Brown corpus and all the SENSEVAL data. The data were translated to different languages (Arabic, French and Spanish) using state of art MT systems. They employed the automatic alignment system GIZA++ (Och and Ney, 2003) to obtain word alignments in a single direction from L1 to L2. For TransCont we use the basic SALAAM approach with some crucial modifications that lead to better performance. We still rely on parallel corpora, we extract typesets based on the intersection of word alignments in both alignment directions using more advanced GIZA++ machinery. In contrast to DR02, we experiment with all four POS: Verbs (V), Nouns (N), Adjectives (A) and Adverbs (R). Moreover, we modified the underlying disambiguation method on the typesets. We still employ WN similarity, however, we do not use the NounGroupings algorithm. Our disambiguation method relies on calculating the sense pair similarity exhaustively across all the 1545 word types in a typeset and choosing the combination that yields the highest similarity. We experimented with all the WN similarity measures in the WN similarity package.9 We also experiment with Lesk2 and Lesk3 as well as other measures, however we do not use SemCor examples with TransCont. We found that the best results are yielded using the Lesk2/Lesk3 similarity measure for N, A and R POS tagsets, while the Lin and JCN measures yield the best performance for the verbs. In contrast to the DR02 approach, we modify the internal WSD process to use the In-Degree algorithm on the typeset, so each sense obtains a confidence, and the sense(s) with the highest confidences are returned. 4.3 Combining RelCont and TransCont Our objective is to combine the different sources of evidence for the purposes of producing an effective overall global WSD system that is able to disambiguate all content words in running text. We combine the two systems in two different ways. 4.3.1 MERGE In this combination scheme, the words in the typeset that result from the TransCont approach are added to the context of the target word in the RelCont approach. However the typeset words are not treated the same as the words that come from the surrounding context in the In-Degree algorithm as we recognize that words that are yielded in the typesets are semantically similar in terms of content rather than being co-occurring words as is the case for contextual words in RelCont. Heeding this difference, we proceed to calculate similarity for words in the typesets using different similarity measures. In the case of noun-noun similarity, in the original RelCont experiments we use JCN, however with the words present in the TransCont typesets we use one of the Lesk variants, Lesk2 or Lesk3. Our observation is that the JCN measure is relatively coarser grained, compared to Lesk measures, therefore it is sufficient in case of lexical relatedness therefore works well in case of the context words. Yet for the words yielded in the TransCont typesets a method that exploits the underlying rich relations in the noun hierarchy captures the semantic similarity more aptly. In the case of verbs we still maintain the JCN similarity as it most effective 9http://wn-similarity.sourceforge.net/ given the shallowness of the verb hierarchy and the inherent nature of the verbal synsets which are differentiated along syntactic rather than semantic dimensions. We employ the Lesk algorithm still with A-A and R-R similarity and when comparing across different POS tag pairings. 4.3.2 VOTE In this combination scheme, the output of the global disambiguation system is simply an intersection of the two outputs from the two underlying systems RelCont and TransCont. Specifically, we sum up the confidence ranging from 0 to 1 of the two system In-Degree algorithm outputs to obtain a final confidence for each sense, choosing the sense(s) that yields the highest confidences. The fact that TransCont uses In-Degree internally allows for a seamless integration. 5 Experiments and Results 5.1 Data The parallel data we experiment with are the same standard data sets as in (Diab and Resnik, 2002), namely, Senseval 2 English AW data sets (SV2AW) (Palmer et al., 2001), and Seneval 3 English AW (SV3AW) data set. We use the true POS tag sets in the test data as rendered in the Penn Tree Bank.10 We present our results on WordNet 1.7.1 for ease of comparison with previous results. 5.2 Evaluation Metrics We use the scorer2 software to report finegrained (P)recision and (R)ecall and (F)-measure. 5.3 Baselines We consider here several baselines. 1. A random baseline (RAND) is the most appropriate baseline for an unsupervised approach.2. We include the most frequent sense baseline (MFBL), though we note that we consider the most frequent sense or first sense baseline to be a supervised baseline since it depends crucially on SemCor in ranking the senses within WN.11 3. The SM07 results as a 10We exclude the data points that have a tag of ”U” in the gold standard for both baselines and our system. 11From an application standpoint, we do not find the first sense baseline to be of interest since it introduces a strong level of uniformity – removing semantic variability – which is not desirable. Even if the first sense achieves higher results in data sets, it is an artifact of the size of the data and the very limited number of documents under investigation. 1546 monolingual baseline. 4. The DR02 results as the multilingual baseline. 5.4 Experimental Results 5.4.1 RelCont We present the results for 4 different experimental conditions for RelCont: JCN-V which uses JCN instead of LCH for verb-verb similarity comparison, we consider this our base condition; +ExpandL is adding the Lesk Expansion to the base condition, namely Lesk2;12 +SemCor adds the SemCor expansion to the base condition; and finally +ExpandL SemCor, adds the latter both conditions simultaneously. Table 1 illustrates the obtained results for the SV2AW using WordNet 1.7.1 since it is the most studied data set and for ease of comparison with previous studies. We break the results down by POS tag (N)oun, (V)erb, (A)djective, and Adve(R)b. The coverage for SV2AW is 98.17% losing some of the verb and adverb target words. Our overall results on all the data sets clearly outperform the baseline as well as state-of-theart performance using an unsupervised system (SM07) in overall f-measure across all the data sets. We are unable to beat the most frequent baseline (MFBL) which is obtained using the first sense. However MFBL is a supervised baseline and our approach is unsupervised. Our implementation of SM07 is slightly higher than those reported in (Sinha and Mihalcea, 2007) (57.12% ) is probably due to the fact that we do not consider the items tagged as ”U” and also we resolve some of the POS tag mismatches between the gold set and the test data. We note that for the SV2AW data set our coverage is not 100% due to some POS tag mismatches that could not have been resolved automatically. These POS tag problems have to do mainly with multiword expressions. In observing the performance of the overall RelCont, we note that using JCN for verbs clearly outperforms using the LCH similarity measure. Using SemCor to augment WN examples seems to have the biggest impact. Combining SemCor with ExpandL yields the best results. Observing the results yielded per POS in Table 1, ExpandL seems to have the biggest impact on the Nouns only. This is understandable since the noun hierarchy has the most dense relations and the most consistent ones. SemCor augmen12Using Lesk3 yields almost the same results tation of WN seemed to benefit all POS significantly except for nouns. In fact the performance on the nouns deteriorated from the base condition JCN-V from 68.7 to 68.3%. This maybe due to inconsistencies in the annotations of nouns in SemCor or the very fine granularity of the nouns in WN. We know that 72% of the nouns, 74% of the verbs, 68.9% of the adjectives, and 81.9% of the adverbs directly exploited the use of SemCor augmented examples. Combining SemCor and ExpandL seems to have a positive impact on the verbs and adverbs, but not on the nouns and adjectives. These trends are not held consistently across data sets. For example, we see that SemCor augmentation helps all POS tag sets over using ExpandL alone or even when combined with SemCor. We note the similar trends in performance for the SV3AW data. Compared to state of the art systems, RelCont with an overall F-measure performance of 62.13% outperforms the best unsupervised system of 57.5% UNED-AW-U2 for SV2 (Navigli, 2009). It is worth noting that it is higher than several of the supervised systems. Moreover, RelCont yields better overall results on SV3 at 59.87 compared to the best unsupervised system IRST-DDD-U which yielded an F-measure of 58.3% (Navigli, 2009). 5.4.2 TransCont For the TransCont results we illustrate the original SALAAM results as our baseline. Similar to the DR02 work, we actually use the same SALAAM parallel corpora comprising more than 5.5M English tokens translated using a single machine translation system GlobalLink. Therefore our parallel corpus is the French English translation condition mentioned in DR02 work as FrGl. We have 4 experimental conditions: FRGL using Lesk2 for all POS tags in the typeset disambiguation (Lesk2); FRGL using Lesk3 for all POS tags (Lesk3); using Lesk3 for N, A and R but LIN similarity measure for verbs (Lesk3 Lin); using Lesk3 for N, A and R but JCN for verbs (Lesk3 JCN). In Table 3 we note the the Lesk3 JCN followed immediately by Lesk3 Lin yield the best performance. The trend holds for both SV2AW and SV3AW. Essentially our new implementation of the multilingual system significantly outperforms the original DR02 implementation for all experimental conditions. 1547 Condition N V A R Global F Measure RAND 43.7 21 41.2 57.4 39.9 MFBL 71.8 41.45 67.7 81.8 65.35 SM07 68.7 33.01 65.2 63.1 59.2 JCN-V 68.7 35.46 65.2 63.1 59.72 +ExpandL 70.2 35.86 65.4 62.45 60.48 +SemCor 68.5 38.66 69.2 67.75 61.79 +ExpandL SemCor 69.0 38.66 68.8 69.45 62.13 Table 1: RelCont F-measure results per POS tag per condition for SV2AW using WN 1.7.1. Condition N V A R Global F Measure RAND 39.67 19.34 41.85 92.31 32.97 MFBL 70.4 54.15 66.7 92.88 63.96 SM07 60.9 43.4 57 92.88 53.98 JCN-V 60.9 48.5 57 92.88 55.87 +ExpandL 59.9 48.55 57.95 92.88 55.62 +SemCor 66 48.95 65.55 92.88 59.87 +ExpandL SemCor 65 49.2 65.55 92.88 59.52 Table 2: RelCont F-measure results per POS tag per condition for SV3AW using WN 1.7.1. 5.4.3 Global Combined WSD In this section we present the results of the global combined WSD system. All the combined experimental conditions have the same percentage coverage.13 We present the results combining using MERGE and using VOTE. We have chosen 4 baseline systems: (1) SM07; (2) the our baseline monolingual system using JCN for verb-verb comparisons (RelCont-BL), so as to distinguish the level of improvement that could be attributed to the multilingual system in the combination results; as well as (3) and (4) our best individual system results from RelCont (ExpandL SemCor) referred to in the tables below as (RelCont-Final) and TransCont using the best experimental condition (Lesk3 JCN). Table 5 and 6 illustrates the overall performance of our combined approach. In Table 5 we note that the combined conditions outperform the two base systems independently, using TransCont is always helpful for any of the 3 monolingual systems, no matter we use VOTE or MERGE. In general the trend is that VOTE outperforms MERGE, however they exhibit different behaviors with respect to what works for each POS. In Table 6 the combined result is not always better than the corresponding monolingual system. When applying to our baseline monolin13We do not back off in any of our systems to a default sense, hence the coverage is not at a 100%. gual system, the combined result is still better. However, we observed worse results for ExpandL Semcor, RelCont-Final. There may be 2 main reasons for the loss: (1) SV3 is the tuning set in SM07, and we inherit the thresholds for similarity metrics from that study. Accordingly, an overfitting of the thresholds is probably happening in this case; (2) TransCont results are not good enough on the SV3AW data. Comparing the RelCont and TransCont system results, we find a drop in f-measure of −1.37% in SV2AW, in contrast to a much larger drop in performance for the SV3AW data set where the drop in performance is −6.38% when comparing RelCont-BL to TransCont and nearly −10% comparing against RelCont-Final. 6 Discussion We looked closely at the data in the combined conditions attempting to get a feel for the data and understand what was captured and what was not. Some of the good examples that are captured in the combined system that are not tagged in RelCont is the case of ringer in Like most of the other 6,000 churches in Britain with sets of bells , St. Michael once had its own “ band ” of ringers , who would herald every Sunday morning and evening service .. The RelCont answer is ringer sense number 4: (horseshoes) the successful throw of a horseshoe 1548 Condition N V A R Global F Measure RAND 43.7 21 41.2 57.4 39.9 DR02-FRGL 54.5 SALAAM 65.48 31.77 56.87 67.4 57.23 Lesk2 67.05 30 59.69 68.01 57.27 Lesk3 67.15 30 60.2 68.01 57.41 Lesk3 Lin 67.15 29.27 60.2 68.01 57.61 Lesk3 JCN 67.15 33.88 60.2 68.01 58.35 Table 3: TransCont F-measure results per POS tag per condition for SV2AW using WN 1.7.1. Condition N V A R Global F Measure RAND 39.67 19.34 41.85 92.31 32.93 SALAAM 52.42 29.27 54.14 88.89 45.63 Lesk2 53.57 33.58 53.63 88.89 47 Lesk3 53.77 33.30 56.48 88.89 47.5 Lesk3 Lin 53.77 29.24 56.48 88.89 46.37 Lesk3 JCN 53.77 38.43 56.48 88.89 49.29 Table 4: TransCont F-measure results per POS tag per condition for SV3AW using WN 1.7.1. or quoit so as to encircle a stake or peg. When the merged system is employed we see the correct sense being chosen as sense number 1 in the MERGE condition: defined in WN as a person who rings church bells (as for summoning the congregation) resulting from a corresponding translation into French as sonneur. We did some basic data analysis on the items we are incapable of capturing. Several of them are cases of metonymy in examples such as ”the English are known...”, the sense of English here is clearly in reference to the people of England, however, our WSD system preferred the language sense of the word. These cases are not gotten by any of our systems. If it had access to syntactic/semantic roles we assume it could capture that this sense of the word entails volition for example. Other types of errors resulted from the lack of a way to explicitly identify multiwords. Looking at the performance of TransCont we note that much of the loss is a result of the lack of variability in the translations which is a key factor in the performance of the algorithm. For example for the 157 adjective target test words in SV2AW, there was a single word alignment for 51 of the cases, losing any tagging for these words. 7 Conclusions and Future Directions In this paper we present a framework that combines orthogonal sources of evidence to create a state-of-the-art system for the task of WSD disambiguation for AW. Our approach yields an overall global F measure of 64.58 for the standard SV2AW data set combining monolingual and multilingual evidence. The approach can be further refined by adding other types of orthogonal features such as syntactic features and semantic role label features. Adding SemCor examples to TransCont should have a positive impact on performance. Also adding more languages as illustrated by the DR02 work should also yield much better performance. References Marine Carpuat and Dekai Wu. 2007. Improving statistical machine translation using word sense disambiguation. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pages 61–72, Prague, Czech Republic, June. Association for Computational Linguistics. Yee Seng Chan, Hwee Tou Ng, and David Chiang. 2007. Word sense disambiguation improves statistical machine translation. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 33–40, Prague, Czech Republic, June. Association for Computational Linguistics. Mona Diab and Philip Resnik. 2002. An unsupervised method for word sense tagging using parallel corpora. In Proceedings of 40th Annual Meeting of the Association for Computational Linguistics, 1549 Condition N V A R Global F Measure SM07 68.7 33.01 65.2 63.1 59.2 RelCont-BL 68.7 35.46 65.2 63.1 59.72 RelCont-Final 69.0 38.66 68.8 69.45 62.13 TransCont 67.15 33.88 60.2 68.01 58.35 MERGE: RelCont-BL+TransCont 69.3 36.91 66.7 64.45 60.82 VOTE: RelCont-BL+TransCont 71 37.71 66.5 66.1 61.92 MERGE: RelCont-Final+TransCont 70.7 38.66 69.5 70.45 63.14 VOTE: RelCont-Final+TransCont 74.2 38.26 68.6 71.45 64.58 Table 5: F-measure % for all Combined experimental conditions on SV2AW Condition N V A R Global F Measure SM07 60.9 43.4 57 92.88 53.98 RelCont-BL 60.9 48.5 57 92.88 55.87 RelCont-Final 65 49.2 65.55 92.88 59.52 TransCont 53.77 38.43 56.48 88.89 49.29 MERGE: RelCont-BL+TransCont 60.6 49.5 58.85 92.88 56.47 VOTE: RelCont-BL+TransCont 59.3 49.5 59.1 92.88 55.92 MERGE: RelCont-Final+TransCont 63.2 50.3 65.25 92.88 59.07 VOTE: RelCont-Final+TransCont 62.4 49.65 65.25 92.88 58.47 Table 6: F-measure % for all Combined experimental conditions on SV3AW pages 255–262, Philadelphia, Pennsylvania, USA, July. Association for Computational Linguistics. Christiane Fellbaum. 1998. ”wordnet: An electronic lexical database”. MIT Press. Weiwei Guo and Mona Diab. 2009. Improvements to monolingual english word sense disambiguation. In Proceedings of the Workshop on Semantic Evaluations: Recent Achievements and Future Directions (SEW-2009), pages 64–69, Boulder, Colorado, June. Association for Computational Linguistics. N. Ide and J. Veronis. 1998. Word sense disambiguation: The state of the art. In Computational Linguistics, pages 1–40, 24:1. J. Jiang and D. Conrath. 1997. Semantic similarity based on corpus statistics and lexical taxonomy. In Proceedings of the International Conference on Research in Computational Linguistics, Taiwan. C. Leacock and M. Chodorow. 1998. Combining local context and wordnet sense similarity for word sense identification. In WordNet, An Electronic Lexical Database. The MIT Press. M. Lesk. 1986. Automatic sense disambiguation using machine readable dictionaries: How to tell a pine cone from an ice cream cone. In In Proceedings of the SIGDOC Conference, Toronto, June. Rada Mihalcea. 2005. Unsupervised large-vocabulary word sense disambiguation with graph-based algorithms for sequence data labeling. In Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing, pages 411–418, Vancouver, British Columbia, Canada, October. Association for Computational Linguistics. George A. Miller. 1990. Wordnet: a lexical database for english. In Communications of the ACM, pages 39–41. Roberto Navigli and Mirella Lapata. 2007. Graph connectivity measures for unsupervised word sense disambiguation. In Proceedings of the 20th International Joint Conference on Artificial Intelligence (IJCAI), pages 1683–1688, Hyderabad, India. Roberto Navigli. 2009. Word sense disambiguation: a survey. In ACM Computing Surveys, pages 1–69. ACM Press. Franz Joseph Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1):19–51. M. Palmer, C. Fellbaum, S. Cotton, L. Delfs, , and H. Dang. 2001. English tasks: all-words and verb lexical sample. In In Proceedings of ACL/SIGLEX Senseval-2, Toulouse, France, June. Ted Pedersen, Satanjeev Banerjee, and Siddharth Patwardhan. 2005. Maximizing semantic relatedness to perform word sense disambiguation. In University of Minnesota Supercomputing Institute Research Report UMSI 2005/25, Minnesotta, March. 1550 Sameer Pradhan, Edward Loper, Dmitriy Dligach, and Martha Palmer. 2007. Semeval-2007 task-17: English lexical sample, srl and all words. In Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007), pages 87–92, Prague, Czech Republic, June. Association for Computational Linguistics. Ravi Sinha and Rada Mihalcea. 2007. Unsupervised graph-based word sense disambiguation using measures of word semantic similarity. In Proceedings of the IEEE International Conference on Semantic Computing (ICSC 2007), Irvine, CA. Benjamin Snyder and Martha Palmer. 2004. The english all-words task. In Rada Mihalcea and Phil Edmonds, editors, Senseval-3: Third International Workshop on the Evaluation of Systems for the Semantic Analysis of Text, pages 41–43, Barcelona, Spain, July. Association for Computational Linguistics. 1551
|
2010
|
156
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1552–1561, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Phrase-based Statistical Language Generation using Graphical Models and Active Learning Franc¸ois Mairesse, Milica Gaˇsi´c, Filip Jurˇc´ıˇcek, Simon Keizer, Blaise Thomson, Kai Yu and Steve Young∗ Cambridge University Engineering Department, Trumpington Street, Cambridge, CB2 1PZ, UK {f.mairesse, mg436, fj228, sk561, brmt2, ky219, sjy}@eng.cam.ac.uk Abstract Most previous work on trainable language generation has focused on two paradigms: (a) using a statistical model to rank a set of generated utterances, or (b) using statistics to inform the generation decision process. Both approaches rely on the existence of a handcrafted generator, which limits their scalability to new domains. This paper presents BAGEL, a statistical language generator which uses dynamic Bayesian networks to learn from semantically-aligned data produced by 42 untrained annotators. A human evaluation shows that BAGEL can generate natural and informative utterances from unseen inputs in the information presentation domain. Additionally, generation performance on sparse datasets is improved significantly by using certainty-based active learning, yielding ratings close to the human gold standard with a fraction of the data. 1 Introduction The field of natural language generation (NLG) is one of the last areas of computational linguistics to embrace statistical methods. Over the past decade, statistical NLG has followed two lines of research. The first one, pioneered by Langkilde and Knight (1998), introduces statistics in the generation process by training a model which reranks candidate outputs of a handcrafted generator. While their HALOGEN system uses an n-gram language model trained on news articles, other systems have used hierarchical syntactic models (Bangalore and Rambow, 2000), models trained on user ratings of ∗This research was partly funded by the UK EPSRC under grant agreement EP/F013930/1 and funded by the EU FP7 Programme under grant agreement 216594 (CLASSiC project: www.classic-project.org). utterance quality (Walker et al., 2002), or alignment models trained on speaker-specific corpora (Isard et al., 2006). A second line of research has focused on introducing statistics at the generation decision level, by training models that find the set of generation parameters maximising an objective function, e.g. producing a target linguistic style (Paiva and Evans, 2005; Mairesse and Walker, 2008), generating the most likely context-free derivations given a corpus (Belz, 2008), or maximising the expected reward using reinforcement learning (Rieser and Lemon, 2009). While such methods do not suffer from the computational cost of an overgeneration phase, they still require a handcrafted generator to define the generation decision space within which statistics can be used to find an optimal solution. This paper presents BAGEL (Bayesian networks for generation using active learning), an NLG system that can be fully trained from aligned data. While the main requirement of the generator is to produce natural utterances within a dialogue system domain, a second objective is to minimise the overall development effort. In this regard, a major advantage of data-driven methods is the shift of the effort from model design and implementation to data annotation. In the case of NLG systems, learning to produce paraphrases can be facilitated by collecting data from a large sample of annotators. Our meaning representation should therefore (a) be intuitive enough to be understood by untrained annotators, and (b) provide useful generalisation properties for generating unseen inputs. Section 2 describes BAGEL’s meaning representation, which satisfies both requirements. Section 3 then details how our meaning representation is mapped to a phrase sequence, using a dynamic Bayesian network with backoff smoothing. Within a given domain, the same semantic concept can occur in different utterances. Section 4 details how BAGEL exploits this redundancy 1552 to improve generation performance on sparse datasets, by guiding the data collection process using certainty-based active learning (Lewis and Catlett, 1994). We train BAGEL in the information presentation domain, from a corpus of utterances produced by 42 untrained annotators (see Section 5.1). An automated evaluation metric is used to compare preliminary model and training configurations in Section 5.2, while Section 5.3 shows that the resulting system produces natural and informative utterances, according to 18 human judges. Finally, our human evaluation shows that training using active learning significantly improves generation performance on sparse datasets, yielding results close to the human gold standard using a fraction of the data. 2 Phrase-based generation from semantic stacks BAGEL uses a stack-based semantic representation to constrain the sequence of semantic concepts to be searched. This representation can be seen as a linearised semantic tree similar to the one previously used for natural language understanding in the Hidden Vector State model (He and Young, 2005). A stack representation provides useful generalisation properties (see Section 3.1), while the resulting stack sequences are relatively easy to align (see Section 5.1). In the context of dialogue systems, Table 1 illustrates how the input dialogue act is first mapped to a set of stacks of semantic concepts, and then aligned with a word sequence. The bottom concept in the stack will typically be a dialogue act type, e.g. an utterance providing information about the object under discussion (inform) or specifying that the request of the user cannot be met (reject). Other concepts include attributes of that object (e.g., food, area), values for those attributes (e.g., Chinese, riverside), as well as special symbols for negating underlying concepts (e.g., not) or specifying that they are irrelevant (e.g., dontcare). The generator’s goal is thus finding the most likely realisation given an unordered set of mandatory semantic stacks Sm derived from the input dialogue act. For example, s =inform(area(centre)) is a mandatory stack associated with the dialogue act in Table 1 (frame 8). While mandatory stacks must all be conveyed in the output realisation, Sm does not contain the optional intermediary stacks Si that can refer to (a) general attributes of the object under discussion (e.g., inform(area) in Table 1), or (b) to concepts that are not in the input at all, which are associated with the singleton stack inform (e.g., phrases expressing the dialogue act type, or clause aggregation operations). For example, the stack sequence in Table 1 contains 3 intermediary stacks for t = 2, 5 and 7. BAGEL’s granularity is defined by the semantic annotation in the training data, rather than external linguistic knowledge about what constitutes a unit of meaning, i.e. contiguous words belonging to the same semantic stack are modelled as an atomic observation unit or phrase.1 In contrast with wordlevel models, a major advantage of phrase-based generation models is that they can model longrange dependencies and domain-specific idiomatic phrases with fewer parameters. 3 Dynamic Bayesian networks for NLG Dynamic Bayesian networks have been used successfully for speech recognition, natural language understanding, dialogue management and text-tospeech synthesis (Rabiner, 1989; He and Young, 2005; Lef`evre, 2006; Thomson and Young, 2010; Tokuda et al., 2000). Such models provide a principled framework for predicting elements in a large structured space, such as required for nontrivial NLG tasks. Additionally, their probabilistic nature makes them suitable for modelling linguistic variation, i.e. there can be multiple valid paraphrases for a given input. BAGEL models the generation task as finding the most likely sequence of realisation phrases R∗= (r1...rL) given an unordered set of mandatory semantic stacks Sm, with |Sm| ≤L. BAGEL must thus derive the optimal sequence of semantic stacks S∗that will appear in the utterance given Sm, i.e. by inserting intermediary stacks if needed and by performing content ordering. Any number of intermediary stacks can be inserted between two consecutive mandatory stacks, as long as all their concepts are included in either the previous or following mandatory stack, and as long as each stack transition leads to a different stack (see example in Table 1). Let us define the set of possible stack sequences matching these constraints as Seq(Sm) ⊆{S = (s1...sL) s.t. st ∈Sm ∪Si}. We propose a model which estimates the dis1The term phrase is thus defined here as any sequence of one or more words. 1553 Charlie Chan is a Chinese restaurant near Cineworld in the centre of town Charlie Chan Chinese restaurant Cineworld centre name food type near near area area inform inform inform inform inform inform inform inform t = 1 t = 2 t = 3 t = 4 t = 5 t = 6 t = 7 t = 8 Table 1: Example semantic stacks aligned with an utterance for the dialogue act inform(name(Charlie Chan) type(restaurant) area(centre) food(Chinese) near(Cineworld)). Mandatory stacks are in bold. tribution P(R|Sm) from a training set of realisation phrases aligned with semantic stack sequences, by marginalising over all stack sequences in Seq(Sm): P(R|Sm) = X S∈Seq(Sm) P(R, S|Sm) = X S∈Seq(Sm) P(R|S, Sm)P(S|Sm) = X S∈Seq(Sm) P(R|S)P(S|Sm) (1) Inference over the model defined in (1) requires the decoding algorithm to consider all possible orderings over Seq(Sm) together with all possible realisations, which is intractable for non-trivial domains. We thus make the additional assumption that the most likely sequence of semantic stacks S∗given Sm is the one yielding the optimal realisation phrase sequence: P(R|Sm) ≈P(R|S∗)P(S∗|Sm) (2) with S∗= argmax S∈Seq(Sm) P(S|Sm) (3) The semantic stacks are therefore decoded first using the model in Fig. 1 to solve the argmax in (3). The decoded stack sequence S∗is then treated as observed in the realisation phase, in which the model in Fig. 2 is used to find the realisation phrase sequence R∗maximising P(R|S∗) over all phrase sequences of length L = |S∗| in our vocabulary: R∗= argmax R=(r1...rL) P(R|S∗)P(S∗|Sm) (4) = argmax R=(r1...rL) P(R|S∗) (5) In order to reduce model complexity, we factorise our model by conditioning the realisation phrase at time t on the previous phrase rt−1, and the previous, current, and following semantic stacks. The semantic stack st at time t is assumed last mandatory stack stack set validator first frame semantic stack s stack set tracker repeated frame final frame validator Figure 1: Graphical model for the semantic decoding phase. Plain arrows indicate smoothed probability distributions, dashed arrows indicate deterministic relations, and shaded nodes are observed. The generation of the end semantic stack symbol deterministically triggers the final frame. to depend only on the previous two stacks and the last mandatory stack su ∈Sm with 1 ≤u < t: P(S|Sm) = QT t=1 P(st|st−1, st−2, su) if S ∈Seq(Sm) 0 otherwise (6) P(R|S∗) = T Y t=1 P(rt|rt−1, s∗ t−1, s∗ t , s∗ t+1) (7) While dynamic Bayesian networks typically take sequential inputs, mapping a set of semantic stacks to a sequence of phrases is achieved by keeping track of the mandatory stacks that were visited in the current sequence (see stack set tracker variable in Fig. 1), and pruning any sequence that has not included all mandatory input stacks on reaching the final frame (see observed stack set validator variable in Fig. 1). Since the number of intermediary stacks is not known at decoding time, the network is unrolled for a fixed number of frames T defining the maximum number of phrases that can be generated (e.g., T = 50). The end of the stack sequence is then determined by a special end symbol, which can only be emitted within the T frames once all mandatory stacks have been visited. The probability of the resulting utterance is thus computed over all frames up to the end symbol, which determines the length 1554 L of S∗and R∗. While the decoding constraints enforce that L > |Sm|, the search for S∗requires comparing sequences of different lengths. A consequence is that shorter sequences containing only mandatory stacks are likely to be favoured. While future work should investigate length normalisation strategies, we find that the learned transition probabilities are skewed enough to favour stack sequences including intermediary stacks. Once the topology and the decoding constraints of the network have been defined, any inference algorithm can be used to search for S∗and R∗. We use the junction tree algorithm implemented in the Graphical Model ToolKit (GMTK) for our experiments (Bilmes and Zweig, 2002), however both problems can be solved using a standard Viterbi search given the appropriate state representation. In terms of computational complexity, it is important to note that the number of stack sequences Seq(Sm) to search over increases exponentially with the number of input mandatory stacks. Nevertheless, we find that real-time performance can be achieved by pruning low probability sequences, without affecting the quality of the solution. 3.1 Generalisation to unseen semantic stacks In order to generalise to semantic stacks which have not been observed during training, the realisation phrase r is made dependent on underspecified stack configurations, i.e. the tail l and the head h of the stack. For example, the last stack in Table 1 is associated with the head centre and the tail inform(area). As a result, BAGEL assigns non-zero probabilities to realisation phrases in unseen semantic contexts, by backing off to the head and the tail of the stack. A consequence is that BAGEL’s lexical realisation can generalise across contexts. For example, if reject(area(centre)) was never observed at training time, P(r = centre of town|s = reject(area(centre))) will be estimated by backing off to P(r = centre of town|h = centre). BAGEL can thus generate ‘there are no venues in the centre of town’ if the phrase ‘centre of town’ was associated with the concept centre in a different context, such as inform(area(centre)). The final realisation model is illustrated in Fig. 2: realisation phrase r repeated frame final frame first frame stack head h semantic stack s stack tail l Figure 2: Graphical model for the realisation phase. Dashed arrows indicate deterministic relations, and shaded node are observed. !"#$%&& '(")*+ 1 1 1 1 1 , , , , , , , | + − + − − t t t t t t t t t s s s l l r l h r t t t t t t t s l l r l h r , , , , , | 1 1 1 + − − 1 1 1 , , , , | + − − t t t t t t l l r l h r t t t l h r , | 2 1, | − − t t t s s s u t t t s s s s , , | 2 1 − − t t h r | 1 | − t t s s tr ts Figure 3: Backoff graphs for the semantic decoding and realisation models. P(R|S∗) = L Y t=1 P(rt|rt−1, ht, lt−1, lt, lt+1, s∗ t−1, s∗ t , s∗ t+1) (8) Conditional probability distributions are represented as factored language models smoothed using Witten-Bell interpolated backoff smoothing (Bilmes and Kirchhoff, 2003), according to the backoff graphs in Fig. 3. Variables which are the furthest away in time are dropped first, and partial stack variables are dropped last as they are observed the most. It is important to note that generating unseen semantic stacks requires all possible mandatory semantic stacks in the target domain to be predefined, in order for all stack unigrams to be assigned a smoothed non-zero probability. 3.2 High cardinality concept abstraction While one should expect a trainable generator to learn multiple lexical realisations for lowcardinality semantic concepts, learning lexical realisations for high-cardinality database entries (e.g., proper names) would increase the number of model parameters prohibitively. We thus divide pre-terminal concepts in the semantic stacks into two types: (a) enumerable attributes whose values are associated with distinct semantic stacks in 1555 our model (e.g., inform(pricerange(cheap))), and (b) non-enumerable attributes whose values are replaced by a generic symbol before training in both the utterance and the semantic stack (e.g., inform(name(X)). These symbolic values are then replaced in the surface realisation by the corresponding value in the input specification. A consequence is that our model can only learn synonymous lexical realisations for enumerable attributes. 4 Certainty-based active learning A major issue with trainable NLG systems is the lack of availability of domain-specific data. It is therefore essential to produce NLG models that minimise the data annotation cost. BAGEL supports the optimisation of the data collection process through active learning, in which the next semantic input to annotate is determined by the current model. The probabilistic nature of BAGEL allows the use of certaintybased active learning (Lewis and Catlett, 1994), by querying the k semantic inputs for which the model is the least certain about its output realisation. Given a finite semantic input space I representing all possible dialogue acts in our domain (i.e., the set of all sets of mandatory semantic stacks Sm), BAGEL’s active learning training process iterates over the following steps: 1. Generate an utterance for each semantic input Sm ∈I using the current model.2 2. Annotate the k semantic inputs {S1 m...Sk m} yielding the lowest realisation probability, i.e. for q ∈(1..k) Sq m = argmin Sm∈I\{S1m...Sq−1 m } (max R P(R|Sm)) (9) with P(R|Sm) defined in (2). 3. Retrain the model with the additional k data points. The number of utterances to be queried k should depend on the flexibility of the annotators and the time required for generating all possible utterances in the domain. 5 Experimental method BAGEL’s factored language models are trained using the SRILM toolkit (Stolcke, 2002), and decoding is performed using GMTK’s junction tree inference algorithm (Bilmes and Zweig, 2002). 2Sampling methods can be used if I is infinite or too large. Since each active learning iteration requires generating all training utterances in our domain, they are generated using a larger clique pruning threshold than the test utterances used for evaluation. 5.1 Corpus collection We train BAGEL in the context of a dialogue system providing information about restaurants in Cambridge. The domain contains two dialogue act types: (a) inform: presenting information about a restaurant (see Table 1), and (b) reject: informing that the user’s constraints cannot be met (e.g., ‘There is no cheap restaurant in the centre’). Our domain contains 8 restaurant attributes: name, food, near, pricerange, postcode, phone, address, and area, out of which food, pricerange, and area are treated as enumerable.3 Our input semantic space is approximated by the set of information presentation dialogue acts produced over 20,000 simulated dialogues between our statistical dialogue manager (Young et al., 2010) and an agenda-based user simulator (Schatzmann et al., 2007), which results in 202 unique dialogue acts after replacing nonenumerable values by a generic symbol. Each dialogue act contains an average of 4.48 mandatory semantic stacks. As one of our objectives is to test whether BAGEL can learn from data provided by a large sample of untrained annotators, we collected a corpus of semantically-aligned utterances using Amazon’s Mechanical Turk data collection service. A crucial aspect of data collection for NLG is to ensure that the annotators understand the meaning of the semantics to be conveyed. Annotators were first asked to provide an utterance matching an abstract description of the dialogue act, regardless of the order in which the constraints are presented (e.g., Offer the venue Taj Mahal and provide the information type(restaurant), area(riverside), food(Indian), near(The Red Lion)). The order of the constraints in the description was randomised to reduce the effect of priming. The annotators were then asked to align the attributes (e.g., Indicate the region of the utterance related to the concept ‘area’), and the attribute values (e.g., Indicate only the words related to the concept ‘riverside’). Two paraphrases were collected for each dialogue act in our domain, resulting in a total of 404 aligned ut3With the exception of areas defined as proper nouns. 1556 rt st ht lt <s> START START START The Rice Boat inform(name(X)) X inform(name) is a inform inform EMPTY restaurant inform(type(restaurant)) restaurant inform(type) in the inform(area) area inform riverside inform(area(riverside)) riverside inform(area) area inform(area) area inform that inform inform EMPTY serves inform(food) food inform French inform(food(French)) French inform(food) food inform(food) food inform </s> END END END Table 2: Example utterance annotation used to estimate the conditional probability distributions of the models in Figs. 1 and 2 ( rt=realisation phrase, st=semantic stack, ht=stack head, lt=stack tail). terances produced by 42 native speakers of English. After manually checking and normalising the dataset,4 the layered annotations were automatically mapped to phrase-level semantic stacks by splitting the utterance into phrases at annotation boundaries. Each annotated utterance is then converted into a sequence of symbols such as in Table 2, which are used to estimate the conditional probability distributions defined in (6) and (8). The resulting vocabulary consists of 52 distinct semantic stacks and 109 distinct realisation phrases, with an average of 8.35 phrases per utterance. 5.2 BLEU score evaluation We first evaluate BAGEL using the BLEU automated metric (Papineni et al., 2002), which measures the word n-gram overlap between the generated utterances and the 2 reference paraphrases over a test corpus (with n up to 4). While BLEU suffers from known issues such as a bias towards statistical NLG systems (Reiter and Belz, 2009), it provides useful information when comparing similar systems. We evaluate BAGEL for different training set sizes, model dependencies, and active learning parameters. Our results are averaged over a 10-fold cross-validation over distinct dialogue acts, i.e. dialogue acts used for testing are not seen at training time,5 and all systems are tested on the same folds. The training and test sets respectively contain an average of 181 and 21 distinct dialogue acts, and each dialogue act is associated with two paraphrases, resulting in 362 training utterances. 4The normalisation process took around 4 person-hour for 404 utterances. 5We do not evaluate performance on dialogue acts used for training, as the training examples can trivially be used as generation templates. !"#$ !"% !"%$ !"# !"#$ !"% !"%$ !"#$%&'()%*+,-" !"$ !"$$ !"# !"#$ !"% !"%$ !"#$%&'()%*+,-" !"$ !"$$ !"# !"#$ !"% !"%$ !"#$%&'()%*+,-" &'(()*+,-( !". !".$ !"$ !"$$ !"# !"#$ !"% !"%$ !"#$%&'()%*+,-" &'(()*+,-( /+)01234)5234+66 /+)01234)5234+667)8+)6'1'9-)0-*281:30 !";$ !". !".$ !"$ !"$$ !"# !"#$ !"% !"%$ <! =! .! #! >! <!! <=! <$! =!! =$! ;!! ;#= !"#$%&'()%*+,-" .-#/$/$0%*"1%*/2" &'(()*+,-( /+)01234)5234+66 /+)01234)5234+667)8+)6'1'9-)0-*281:30 !";$ !". !".$ !"$ !"$$ !"# !"#$ !"% !"%$ <! =! .! #! >! <!! <=! <$! =!! =$! ;!! ;#= !"#$%&'()%*+,-" .-#/$/$0%*"1%*/2" &'(()*+,-( /+)01234)5234+66 /+)01234)5234+667)8+)6'1'9-)0-*281:30 !";$ !". !".$ !"$ !"$$ !"# !"#$ !"% !"%$ <! =! .! #! >! <!! <=! <$! =!! =$! ;!! ;#= !"#$%&'()%*+,-" .-#/$/$0%*"1%*/2" &'(()*+,-( /+)01234)5234+66 /+)01234)5234+667)8+)6'1'9-)0-*281:30 Figure 4: BLEU score averaged over a 10-fold cross-validation for different training set sizes and network topologies, using random sampling. Results: Fig. 4 shows that adding a dependency on the future semantic stack improves performances for all training set sizes, despite the added model complexity. Backing off to partial stacks also improves performance, but only for sparse training sets. Fig. 5 compares the full model trained using random sampling in Fig. 4 with the same model trained using certainty-based active learning, for different values of k. As our dataset only contains two paraphrases per dialogue act, the same dialogue act can only be queried twice during the active learning procedure. A consequence is that the training set used for active learning converges towards the randomly sampled set as its size increases. Results show that increasing the training set one utterance at a time using active learning (k = 1) significantly outperforms random sampling when using 40, 80, and 100 utterances (p < .05, two-tailed). Increasing the number of utterances to be queried at each iteration to k = 10 results in a smaller performance increase. A possi1557 !"# !"## !"$ !"$# !"% !"%# !"#$%&'()%*+,-" &'()*+,-'+./0(1 !"2# !"3 !"3# !"# !"## !"$ !"$# !"% !"%# 4! 5! 3! $! 6! 4!! 45! 4#! 5!! 5#! 2!! 2$5 !"#$%&'()%*+,-" .-#/$/$0%*"1%*/2" &'()*+,-'+./0(1 7890:;,/;'<(0(1,=>4 7890:;,/;'<(0(1,=>4! Figure 5: BLEU score averaged over a 10-fold cross-validation for different numbers of queries per iteration, using the full model with the query selection criterion (9). !"# !"## !"$ !"$# !"% !"%# !"#$%&'()%*+,-" &'(()(*+,-*. !"/# !"0 !"0# !"# !"## !"$ !"$# !"% !"%# 1! 2! 0! $! 3! 1!! 12! 1#! 2!! 2#! /!! /$2 !"#$%&'()%*+,-" .-#/$/$0%*"1%*/2" &'(()(*+,-*. 4*+,-*.),5-)6-785 4*9+5:;)<9,';)6<-:; Figure 6: BLEU score averaged over a 10-fold cross-validation for different query selection criteria, using the full model with k = 1. ble explanation is that the model is likely to assign low probabilities to similar inputs, thus any value above k = 1 might result in redundant queries within an iteration. As the length of the semantic stack sequence is not known before decoding, the active learning selection criterion presented in (9) is biased towards longer utterances, which tend to have a lower probability. However, Fig. 6 shows that normalising the log probability by the number of semantic stacks does not improve overall learning performance. Although a possible explanation is that longer inputs tend to contain more information to learn from, Fig. 6 shows that a baseline selecting the largest remaining semantic input at each iteration performs worse than the active learning scheme for training sets above 20 utterances. The full log probability selection criterion defined in (9) is therefore used throughout the rest of the paper (with k = 1). 5.3 Human evaluation While automated metrics provide useful information for comparing different systems, human feedback is needed to assess (a) the quality of BAGEL’s outputs, and (b) whether training models using active learning has a significant impact on user perceptions. We evaluate BAGEL through a largescale subjective rating experiment using Amazon’s Mechanical Turk service. For each dialogue act in our domain, participants are presented with a ‘gold standard’ human utterance from our dataset, which they must compare with utterances generated by models trained with and without active learning on a set of 20, 40, 100, and 362 utterances (full training set), as well as with the second human utterance in our dataset. See example utterances in Table 3. The judges are then asked to evaluate the informativeness and naturalness of each of the 8 utterances on a 5 point likert-scale. Naturalness is defined as whether the utterance could have been produced by a human, and informativeness is defined as whether it contains all the information in the gold standard utterance. Each utterance is taken from the test folds of the cross-validation experiment presented in Section 5.2, i.e. the models are trained on up to 90% of the data and the training set does not contain the dialogue act being tested. Results: Figs. 7 and 8 compare the naturalness and informativeness scores of each system averaged over all 202 dialogue acts. A paired t-test shows that models trained on 40 utterances or less produce utterances that are rated significantly lower than human utterances for both naturalness and informativeness (p < .05, two-tailed). However, models trained on 100 utterances or more do not perform significantly worse than human utterances for both dimensions, with a mean difference below .10 over 202 comparisons. Given the large sample size, this result suggests that BAGEL can successfully learn our domain using a fraction of our initial dataset. As far as the learning method is concerned, a paired t-test shows that models trained on 20 and 40 utterances using active learning significantly outperform models trained using random sampling, for both dimensions (p < .05). The largest increase is observed using 20 utterances, i.e. the naturalness increases by .49 and the informativeness by .37. When training on 100 utterances, the effect of active learning becomes insignificant. In1558 Input inform(name(the Fountain) near(the Arts Picture House) area(centre) pricerange(cheap)) Human There is an inexpensive restaurant called the Fountain in the centre of town near the Arts Picture House Rand-20 The Fountain is a restaurant near the Arts Picture House located in the city centre cheap price range Rand-40 The Fountain is a restaurant in the cheap city centre area near the Arts Picture House AL-20 The Fountain is a restaurant near the Arts Picture House in the city centre cheap AL-40 The Fountain is an affordable restaurant near the Arts Picture House in the city centre Full set The Fountain is a cheap restaurant in the city centre near the Arts Picture House Input reject(area(Barnwell) near(Saint Mary′s Church)) Human I am sorry but I know of no venues near Saint Mary’s Church in the Barnwell area Full set I am sorry but there are no venues near Saint Mary’s Church in the Barnwell area Input inform(name(the Swan)area(Castle Hill) pricerange(expensive)) Human The Swan is a restaurant in Castle Hill if you are seeking something expensive Full set The Swan is an expensive restaurant in the Castle Hill area Input inform(name(Browns) area(centre) near(the Crowne Plaza) near(El Shaddai) pricerange(cheap)) Human Browns is an affordable restaurant located near the Crowne Plaza and El Shaddai in the centre of the city Full set Browns is a cheap restaurant in the city centre near the Crowne Plaza and El Shaddai Table 3: Example utterances for different input dialogue acts and system configurations. AL-20 = active learning with 20 utterances, Rand = random sampling. !"## !"$% !"&' !"(% !")* *"%% *"%# *"%' + +"$ ! !"$ * *"$ $ !"#$%$#&'(#)$"**%*+,(" ,-./01 !"## !"$% !"&' !"(% !")* *"%% *"%# *"%' # #"$ + +"$ ! !"$ * *"$ $ +% *% #%% !(+ !"#$%$#&'(#)$"**%*+,(" -(#.$.$/%*"&%*.0" ,-./01 234567897-:.5.; <=1-.8=447:-.378>8*"%' Figure 7: Naturalness mean opinion scores for different training set sizes, using random sampling and active learning. Differences for training set sizes of 20 and 40 are all significant (p < .05). terestingly, while models trained on 100 utterances outperform models trained on 40 utterances using random sampling (p < .05), they do not significantly outperform models trained on 40 utterances using active learning (p = .15 for naturalness and p = .41 for informativeness). These results suggest that certainty-based active learning is beneficial for training a generator from a limited amount of data given the domain size. Looking back at the results presented in Section 5.2, we find that the BLEU score correlates with a Pearson correlation coefficient of .42 with the mean naturalness score and .35 with the mean informativeness score, over all folds of all systems tested (n = 70, p < .01). This is lower than previous correlations reported by Reiter and Belz (2009) in the shipping forecast domain with nonexpert judges (r = .80), possibly because our domain is larger and more open to subjectivity. !"## !"$$ #"%& !"'& !"() #"%$ #"%# #"&! * *"+ ! !"+ # #"+ + !"#$%&$'()*#+&,"$"--%-.()" ,-./01 !"## !"$$ #"%& !"'& !"() #"%$ #"%# #"&! & &"+ * *"+ ! !"+ # #"+ + *% #% &%% !)* !"#$%&$'()*#+&,"$"--%-.()" /)#&$&$0%-"+%-&1" ,-./01 234567897-:.5.; <=1-.8=447:-.378>8#"&! Figure 8: Informativeness mean opinion scores for different training set sizes, using random sampling and active learning. Differences for training set sizes of 20 and 40 are all significant (p < .05). 6 Related work While most previous work on trainable NLG relies on a handcrafted component (see Section 1), recent research has started exploring fully datadriven NLG models. Factored language models have recently been used for surface realisation within the OpenCCG framework (White et al., 2007; Espinosa et al., 2008). More generally, chart generators for different grammatical formalisms have been trained from syntactic treebanks (White et al., 2007; Nakanishi et al., 2005), as well as from semantically-annotated treebanks (Varges and Mellish, 2001). However, a major difference with our approach is that BAGEL uses domain-specific data to generate a surface form directly from semantic concepts, without any syntactic annotation (see Section 7 for further discussion). 1559 This work is strongly related to Wong and Mooney’s WASP−1 generation system (2007), which combines a language model with an inverted synchronous CFG parsing model, effectively casting the generation task as a translation problem from a meaning representation to natural language. WASP−1 relies on GIZA++ to align utterances with derivations of the meaning representation (Och and Ney, 2003). Although early experiments showed that GIZA++ did not perform well on our data—possibly because of the coarse granularity of our semantic representation—future work should evaluate the generalisation performance of synchronous CFGs in a dialogue system domain. Although we do not know of any work on active learning for NLG, previous work has used active learning for semantic parsing and information extraction (Thompson et al., 1999; Tang et al., 2002), spoken language understanding (Tur et al., 2003), speech recognition (Hakkani-T¨ur et al., 2002), word alignment (Sassano, 2002), and more recently for statistical machine translation (Bloodgood and Callison-Burch, 2010). While certaintybased methods have been widely used, future work should investigate the performance of committeebased active learning for NLG, in which examples are selected based on the level of disagreement between models trained on subsets of the data (Freund et al., 1997). 7 Discussion and conclusion This paper presents and evaluates BAGEL, a statistical language generator that can be trained entirely from data, with no handcrafting required beyond the semantic annotation. All the required subtasks—i.e. content ordering, aggregation, lexical selection and realisation—are learned from data using a unified model. To train BAGEL in a dialogue system domain, we propose a stack-based semantic representation at the phrase level, which is expressive enough to generate natural utterances from unseen inputs, yet simple enough for data to be collected from 42 untrained annotators with a minimal normalisation step. A human evaluation over 202 dialogue acts does not show any difference in naturalness and informativeness between BAGEL’s outputs and human utterances. Additionally, we find that the data collection process can be optimised using active learning, resulting in a significant increase in performance when training data is limited, according to ratings from 18 human judges.6 These results suggest that the proposed framework can largely reduce the development time of NLG systems. While this paper only evaluates the most likely realisation given a dialogue act, we believe that BAGEL’s probabilistic nature and generalisation capabilities are well suited to model the linguistic variation resulting from the diversity of annotators. Our first objective is thus to evaluate the quality of BAGEL’s n-best outputs, and test whether sampling from the output distribution can improve naturalness and user satisfaction within a dialogue. Our results suggest that explicitly modelling syntax is not necessary for our domain, possibly because of the lack of syntactic complexity compared with formal written language. Nevertheless, future work should investigate whether syntactic information can improve performance in more complex domains. For example, the realisation phrase can easily be conditioned on syntactic constructs governing that phrase, and the recursive nature of syntax can be modelled by keeping track of the depth of the current embedded clause. While syntactic information can be included with no human effort by using syntactic parsers, their robustness to dialogue system utterances must first be evaluated. Finally, recent years have seen HMM-based synthesis models become competitive with unit selection methods (Tokuda et al., 2000). Our long term objective is to take advantage of those advances to jointly optimise the language generation and the speech synthesis process, by combining both components into a unified probabilistic concept-to-speech generation model. References S. Bangalore and O. Rambow. Exploiting a probabilistic hierarchical model for generation. In Proceedings of the 18th International Conference on Computational Linguistics (COLING), pages 42–48, 2000. A. Belz. Automatic generation of weather forecast texts using comprehensive probabilistic generation-space models. Natural Language Engineering, 14(4):431–455, 2008. J. Bilmes and K. Kirchhoff. Factored language models and generalized parallel backoff. In Proceedings of HLTNAACL, short papers, 2003. J. Bilmes and G. Zweig. The Graphical Models ToolKit: An open source software system for speech and time-series processing. In Proceedings of ICASSP, 2002. 6The full training corpus and the generated utterances used for evaluation are available at http://mi.eng.cam.ac.uk/∼farm2/bagel. 1560 M. Bloodgood and C. Callison-Burch. Bucking the trend: Large-scale cost-focused active learning for statistical machine translation. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics (ACL), 2010. D. Espinosa, M. White, and D. Mehay. Hypertagging: Supertagging for surface realization with CCG. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics (ACL), 2008. Y. Freund, H. S. Seung, E.Shamir, and N. Tishby. Selective sampling using the query by committee algorithm. Machine Learning, 28:133–168, 1997. D. Hakkani-T¨ur, G. Riccardi, and A. Gorin. Active learning for automatic speech recognition. In Proceedings of ICASSP, 2002. Y. He and S. Young. Semantic processing using the Hidden Vector State model. Computer Speech & Language, 19 (1):85–106, 2005. A. Isard, C. Brockmann, and J. Oberlander. Individuality and alignment in generated dialogues. In Proceedings of the 4th International Natural Language Generation Conference (INLG), pages 22–29, 2006. I. Langkilde and K. Knight. Generation that exploits corpusbased statistical knowledge. In Proceedings of the 36th Annual Meeting of the Association for Computational Linguistics (ACL), pages 704–710, 1998. F. Lef`evre. A DBN-based multi-level stochastic spoken language understanding system. In Proceedings of the IEEE Workshop on Spoken Language Technology (SLT), 2006. D. D. Lewis and J. Catlett. Heterogeneous uncertainty ampling for supervised learning. In Proceedings of ICML, 1994. F. Mairesse and M. A. Walker. Trainable generation of BigFive personality styles through data-driven parameter estimation. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics (ACL), 2008. H. Nakanishi, Y. Miyao, , and J. Tsujii. Probabilistic methods for disambiguation of an HPSG-based chart generator. In Proceedings of the IWPT, 2005. F. J. Och and H. Ney. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1):19–51, 2003. D. S. Paiva and R. Evans. Empirically-based control of natural language generation. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL), pages 58–65, 2005. K. Papineni, S. Roukos, T. Ward, and W. J. Zhu. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), 2002. L. R. Rabiner. Tutorial on Hidden Markov Models and selected applications in speech recognition. Proceedings of the IEEE, 77(2):257–285, 1989. E. Reiter and A. Belz. An investigation into the validity of some metrics for automatically evaluating natural language generation systems. Computational Linguistics, 25: 529–558, 2009. V. Rieser and O. Lemon. Natural language generation as planning under uncertainty for spoken dialogue systems. In Proceedings of the Annual Meeting of the European Chapter of the ACL (EACL), 2009. M. Sassano. An empirical study of active learning with support vector machines for japanese word segmentation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), 2002. J. Schatzmann, B. Thomson, K. Weilhammer, H. Ye, and S. Young. Agenda-based user simulation for bootstrapping a POMDP dialogue system. In Proceedings of HLTNAACL, short papers, pages 149–152, 2007. A. Stolcke. SRILM – an extensible language modeling toolkit. In Proceedings of the International Conference on Spoken Language Processing, 2002. M. Tang, X. Luo, and S. Roukos. Active learning for statistical natural language parsing. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), 2002. C. Thompson, M. E. Califf, and R. J. Mooney. Active learning for natural language parsing and information extraction. In Proceedings of ICML, 1999. B. Thomson and S. Young. Bayesian update of dialogue state: A POMDP framework for spoken dialogue systems. Computer Speech & Language, 24(4):562–588, 2010. Y. Tokuda, T. Yoshimura, T. Masuko, T. Kobayashi, and T. Kitamura. Speech parameter generation algorithms for HMM-based speech synthesis. In Proceedings of ICASSP, 2000. G. Tur, R. E. Schapire, and D. Hakkani-T¨ur. Active learning for spoken language understanding. In Proceedings of ICASSP, 2003. S. Varges and C. Mellish. Instance-based natural language generation. In Proceedings of the Annual Meeting of the North American Chapter of the ACL (NAACL), 2001. M. A. Walker, O. Rambow, and M. Rogati. Training a sentence planner for spoken dialogue using boosting. Computer Speech and Language, 16(3-4), 2002. M. White, R. Rajkumar, and S. Martin. Towards broad coverage surface realization with CCG. In Proceedings of the Workshop on Using Corpora for NLG: Language Generation and Machine Translation, 2007. Y. W. Wong and R. Mooney. Generation by inverting a semantic parser that uses statistical machine translation. In Proceedings of HLT-NAACL, 2007. S. Young, M. Gaˇsi´c, S. Keizer, F. Mairesse, J. Schatzmann, B. Thomson, and K. Yu. The Hidden Information State model: a practical framework for POMDP-based spoken dialogue management. Computer Speech and Language, 24(2):150–174, 2010. 1561
|
2010
|
157
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1562–1572, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Plot Induction and Evolutionary Search for Story Generation Neil McIntyre and Mirella Lapata School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh, EH8 9AB, UK [email protected], [email protected] Abstract In this paper we develop a story generator that leverages knowledge inherent in corpora without requiring extensive manual involvement. A key feature in our approach is the reliance on a story planner which we acquire automatically by recording events, their participants, and their precedence relationships in a training corpus. Contrary to previous work our system does not follow a generate-and-rank architecture. Instead, we employ evolutionary search techniques to explore the space of possible stories which we argue are well suited to the story generation task. Experiments on generating simple children’s stories show that our system outperforms previous data-driven approaches. 1 Introduction Computer story generation has met with fascination since the early days of artificial intelligence. Indeed, over the years, several generators have been developed capable of creating stories that resemble human output. To name only a few, TALE-SPIN (Meehan, 1977) generates stories through problem solving, MINSTREL (Turner, 1992) relies on an episodic memory scheme, essentially a repository of previous hand-coded stories, to solve the problems in the current story, and MAKEBELIEVE (Liu and Singh, 2002) uses commonsense knowledge to generate short stories from an initial seed story (supplied by the user). A large body of more recent work views story generation as a form of agent-based planning (Swartjes and Theune, 2008; Pizzi et al., 2007). The agents act as characters with a list of goals. They form plans of action and try to fulfill them. Interesting stories emerge as plans interact and cause failures and possible replanning. The broader appeal of computational story generation lies in its application potential. Examples include the entertainment industry and the development of tools that produce large numbers of plots automatically that might provide inspiration to professional screen writers (Agudo et al., 2004); rendering video games more interesting by allowing the plot to adapt dynamically to the players’ actions (Barros and Musse, 2007); and assisting teachers to create or personalize stories for their students (Riedl and Young, 2004). A major stumbling block for the widespread use of computational story generators is their reliance on expensive, manually created resources. A typical story generator will make use of a knowledge base for providing detailed domain-specific information about the characters and objects involved in the story and their relations. It will also have a story planner that specifies how these characters interact, what their goals are and how their actions result in different story plots. Finally, a sentence planner (coupled with a surface realizer) will render an abstract story specification into natural language text. Traditionally, most of this knowledge is created by hand, and the effort must be repeated for new domains, new characters and plot elements. Fortunately, recent work in natural language processing has taken significant steps towards developing algorithms that learn some of this knowledge automatically from natural language corpora. Chambers and Jurafsky (2009, 2008) propose an unsupervised method for learning narrative schemas, chains of events whose arguments are filled with participant semantic roles defined over words. An example schema is {X arrest, X charge, X raid, X seize, X confiscate, X detain, X deport}, where X stands for the argument types {police, agent, authority, government}. Their approach relies on the intuition that in a coherent text events that are about the same participants are 1562 likely to be part of the same story or narrative. Their model extracts narrative chains, essentially events that share argument slots and merges them into schemas. The latter could be used to construct or enrich the knowledge base of a story generator. In McIntyre and Lapata (2009) we presented a story generator that leverages knowledge inherent in corpora without requiring extensive manual involvement. The generator operates over predicateargument and predicate-predicate co-occurrence tuples gathered from training data. These are used to produce a large set of candidate stories which are subsequently ranked based on their interestingness and coherence. The approach is unusual in that it does not involve an explicit story planning component. Stories are created stochastically by selecting entities and the events they are most frequently attested with. In this work we develop a story generator that is also data-driven but crucially relies on a story planner for creating meaningful stories. Inspired by Chambers and Jurafsky (2009) we acquire story plots automatically by recording events, their participants, and their precedence relationships as attested in a training corpus. Entities give rise to different potential plots which in turn generate multiple stories. Contrary to our previous work (McIntyre and Lapata, 2009), we do not follow a generate-and-rank architecture. Instead, we search the space of possible stories using Genetic Algorithms (GAs) which we argue are advantageous in the story generation setting, as they can search large fitness landscapes while greatly reducing the risk of getting stuck in local optima. By virtue of exploring the search space more broadly, we are able to generate creative stories without an explicit interest scoring module. In the remainder of this paper we give a brief overview of the system described in McIntyre and Lapata (2009) and discuss previous applications of GAs in natural language generation (Section 2). Next, we detail our approach, specifically how plots are created and used in conjunction with genetic search (Sections 3 and 4). Finally, we present our experimental results (Sections 6 and 7) and conclude the paper with discussion of future work. 2 Related Work Our work builds on and extends the story generator developed in McIntyre and Lapata (2009). The system creates simple children’s stories in an interactive context: the user supplies the topic of the story and its desired length (number of sentences). The generator creates a story following a pipeline architecture typical of natural language generation systems (Reiter and Dale, 2000) consisting of content selection, sentence planning, and surface realization. The content of a story is determined by consulting a data-driven knowledge base that records the entities (i.e., nouns) appearing in a corpus and the actions they perform. These are encoded as dependency relations (e.g., subj-verb, verb-obj). In order to promote between-sentence coherence the generator also make use of an action graph that contains action-role pairs and the likelihood of transitioning from one to another. The sentence planner aggregates together entities and their actions into a sentence using phrase structure rules. Finally, surface realization is performed by interfacing RealPro (Lavoie and Rambow, 1997) with a language model. The system searches for the best story overall as well as the best sentences that can be generated from the knowledge base. Unlikely stories are pruned using beam search. In addition, stories are reranked using two scoring functions based on coherence and interest. These are learnt from training data, i.e., stories labeled with numeric values for interest and coherence. Evolutionary search techniques have been previously employed in natural language generation, especially in the context of document planning. Structuring a set of facts into a coherent text is effectively a search problem that may lead to combinatorial explosion for large domains. Mellish et al. (1998) (and subsequently Karamanis and Manurung 2002) advocate genetic algorithms as an alternative to exhaustively searching for the optimal ordering of descriptions of museum artefacts. Rather than requiring a global optimum to be found, the genetic algorithm selects an order (based on coherence) that is good enough for people to understand. Cheng and Mellish (2000) focus on the interaction of aggregation and text planning and use genetic algorithms to search for the best aggregated document that satisfies coherence constraints. The application of genetic algorithms to story generation is novel to our knowledge. Our work also departs from McIntyre and Lapata (2009) in two important ways. Firstly, our generator does not rely on a knowledge base of seemingly unrelated entities and relations. Rather, we employ 1563 a document planner to create and structure a plot for a story. The planner is built automatically from a training corpus and creates plots dynamically depending on the protagonists of the story. Secondly, our search procedure is simpler and more global; instead of searching for the best story twice (i.e., by first finding the n-best stories and then subsequently reranking them based on coherence and interest), our genetic algorithm explores the space of possible stories once. 3 Plot Generation Following previous work (e.g., Shim and Kim 2002; McIntyre and Lapata 2009) we assume that the user supplies a sentence (e.g., the princess loves the prince) from which the system creates a story. Each entity in this sentence (e.g., princess, prince) is associated with its own narrative schema, a set of key events and actors cooccurring with it in the training corpus. Our narrative schemas differ slightly from Chambers and Jurafsky (2009). They acquire schematic representations of situations akin to FrameNet (Fillmore et al., 2003): schemas consists of semantically similar predicates and the entities evoked by them. In our setting, every entity has its own schema, and predicates associated with it are ordered. Plots are generated by merging the entity-specific narrative schemas which subsequently serve as the input to the genetic algorithm. In the following we describe how the narrative schemas are extracted and plots merged, and then discuss our evolutionary search procedure. Entity-based Schema Extraction Before we can generate a plot for a story we must have an idea of the actions associated with the entities in the story, the order in which these actions are performed and also which other entities can participate. This information is stored in a directed graph which we explain below. Our algorithm processes each document at a time, it operates over dependency structures and assumes that entity mentions have been resolved. In our experiments we used Rasp (Briscoe and Carroll, 2002), a broad coverage dependency parser, and the OpenNLP1 coreference resolution engine.2 However, any dependency parser or coreference tool could serve our 1See http://opennlp.sourceforge.net/. 2The coreference resolution tool we employ is not error-free and on occasion will fail to resolve a pronoun. We map unresolved pronouns to the generic labels person or object. purpose. We also assume that the actions associated with a given entity are ordered and that linear order corresponds to temporal order. This is a gross simplification as it is well known that temporal relationships between events are not limited to precedence, they may overlap, occur simultaneously, or be temporally unrelated. We could have obtained a more accurate ordering using a temporal classifier (see Chambers and Jurafsky 2008), however we leave this to future work. For each entity e in the corpus we build a directed graph G = (V,E) whose nodes V denote predicate argument relationships, and edges E represent transitions from node Vi to node Vj. As an example of our schema construction process, consider a very small corpus consisting of the two documents shown in Figure 1. The schema for princess after processing the first document is given on the left hand side. Each node in this graph corresponds to an action attested with princess (we also record who performs it and where or how). Nodes are themselves dependency trees (see Figure 4a), but are linearized in the figure for the sake of brevity. Edges in the graph indicate ordering and are weighted using the mutual information metric proposed in Lin (1998) (the weights are omitted from the example).3 The first sentence in the text gives rise to the first node in the graph, the second sentence to the second node, and so on. Note that the third sentence is not present in the graph as it is not about the princess. When processing the second document, we simply expand this graph. Before inserting a new node, we check if it can be merged with an already existing one. Nodes are merged only if they have the same verb and similar arguments, with the focal entity (i.e., princess) appearing in the same argument slot. In our example, the nodes “prince marry princess in castle” and “prince marry princess in temple” can be merged as they contain the same verb and number of similar arguments. The nodes “princess have influence” and “princess have baby” cannot be merged as influence and baby are semantically unrelated. We compute argument similarity using WordNet (Fellbaum, 1998) and the measure proposed by Wu and Palmer (1994) which is based on path length. We merge nodes with related arguments only if their similarity exceeds a threshold (determined empirically). 3We use mutual information to identify event sequences strongly associated with the graph entity. 1564 The goblin holds the princess in a lair. The prince rescues the princess and marries her in a castle. The ceremony is beautiful. The princess has influence as the prince rules the country. The dragon holds the princess in a cave. The prince slays the dragon. The princess loves the prince. The prince asks the king’s permission. The prince marries the princess in the temple. The princess has a baby. goblin hold princess in lair prince rescue princess prince marry princess in castle princess have influence goblin dragon hold princess in lair cave prince rescue princess princess love prince prince marry princess in castle temple princess have influence princess have baby Figure 1: Example of schema construction for the entity princess The schema construction algorithm terminates when graphs like the ones shown in Figure 1 (right hand side) have been created for all entities in the corpus. Building a Story Plot Our generator takes an input sentence and uses it to instantiate several plots. We achieve this by merging the schemas associated with the entities in the sentence into a plot graph. As an example, consider again the sentence the princess loves the prince which requires combing the schemas representing prince and princess shown in Figures 2 and 1 (right hand side), respectively. Again, we look for nodes that can be merged based on the identity of the actions involved and the (WordNet) similarity of their arguments. However, we disallow the merging of nodes with focal entities appearing in the same argument slot (e.g., “[prince, princess] cries”). Once the plot graph is created, a depth first search starting from the node corresponding to the input sentence, finds all paths with length matching the desired story length (cycles are disallowed). Assuming we wish to generate a story consisting of three sentences, the graph in Figure 3 would create four plots. These are (princess love prince, prince marry princess in [castle, temple], princess have influence), (princess love prince, prince marry princess in [castle, temple], princess have baby), (princess love prince, prince marry princess in [castle, temple], prince rule country), and (princess love prince, prince ask king’s permission prince marry princess in [castle, temple]). Each of these plots represents two different stories one with castle and one with temple in it. Sentence Planning The sentence planner is interleaved with the story planner and influences the final structure of each sentence in the story. To avoid generating short sentences — note that nodes in the plot graph consist of a single action and would otherwise correspond to a sentence with a single clause — we combine pairs of nodes within the same graph by looking at intrasentential verb-verb co-occurrences in the training corpus. For example, the nodes (prince have problem, prince keep secret) could become the sentence the prince has a problem keeping a secret. We leave it up to the sentence planner to decide how the two actions should be combined.4 The sentence planner will also insert adverbs and adjectives, using co-occurrence likelihoods acquired from the training corpus. It is essentially a phrase structure grammar compiled from the lexical resources made available by Korhonen and Briscoe (2006) and Grishman et al. (1994). The grammar rules act as templates for combining clauses and filling argument slots. 4We only turn an action into a subclause if its subject entity is same as that of the previous action. 1565 prince slay dragon prince rescue princess princess love prince prince marry princess in castle temple prince ask king’s permission prince rule country Figure 2: Narrative schema for the entity prince. 4 Genetic Algorithms The example shown in Figure 3 is a simplified version of a plot graph. The latter would normally contain hundreds of nodes and give rise to thousands of stories once lexical variables have been expanded. Searching the story space is a difficult optimization problem, that must satisfy several constraints: the story should be of a certain length, overall coherent, creative, display some form of event progression, and generally make sense. We argue that evolutionary search is appealing here, as it can find global optimal solutions in a more efficient way than traditional optimization methods. In this study we employ genetic algorithms (GAs) a well-known search technique for finding approximate (or exact) solutions to optimization problems. The basic idea behind GAs is based on “natural selection” and the Darwinian principle of the survival of the fittest (Mitchell, 1998). An initial population is randomly created containing a predefined number of individuals (or solutions), each represented by a genetic string (e.g., a population of chromosomes). Each individual is evaluated according to an objective function (also called a fitness function). A number of individuals are then chosen as parents from the population according to their fitness, and undergo crossover (also called recombination) and mutation in order to develop the new population. Offspring with better fitness are then inserted into the population, replacing the inferior individuals in the previous generation. The algorithm thus identifies the individuals with the optimizing fitness values, and those with lower fitness will naturally get discarded from the population. This cycle is repeated for a given number of generations, or stopped when the solution goblin dragon hold princess in lair cave prince rescue princess princess love prince prince marry princess in castle temple princess have influence princess have baby prince slay dragon prince ask king’s permission prince rule country Figure 3: Plot graph for the input sentence the princess loves the prince. obtained is considered optimal. This process leads to the evolution of a population in which the individuals are more and more suited to their environment, just as natural adaptation. We describe below how we developed a genetic algorithm for our story generation problem. Initial Population Rather than start with a random population, we seed the initial population with story plots generated from our plot graph. For an input sentence, we generate all possible plots. The latter are then randomly sampled until a population of the desired size is created. Contrary to McIntyre and Lapata (2009), we initialize the search with complete stories, rather than generate one sentence at a time. The genetic algorithm will thus avoid the pitfall of selecting early on a solution that will later prove detrimental. Crossover Each plot is represented as an ordered graph of dependency trees (corresponding to sentences). We have decided to use crossover of a single point between two selected parents. The children will therefore contain sentences up to the crossover point of the first parent and sentences after that point of the second. Figure 4a shows two parents (prince rescue princess, prince marry princess in castle, princess have baby) and (prince rescue princess, prince love princess, princess kiss prince) and how two new plots are created by swapping their last nodes. 1566 a) rescue prince princess marry prince princess castle have princess baby rescue prince princess love prince princess kiss princess prince =⇒ rescue prince princess marry prince princess castle kiss prince princess rescue prince princess love prince princess have princess baby in in b) marry prince princess castle hall temple forest kingdom c) rescue prince princess marry prince princess castle kiss prince princess in rescue prince princess marry prince princess castle kiss prince princess in d) rescue prince princess marry prince princess castle kiss prince princess in =⇒ hold prince princess e) knows prince loves princess child =⇒ escape princess dragon Figure 4: Example of genetic algorithm operators as they are applied to plot structures: a) crossover of two plots on a single point, indicated by the dashed line, resulting in two children which are a recombination of the parents; b) mutation of a lexical node, church can be replaced from a list of semantically related candidates; c) sentences can be switched under mutation to create a potentially more coherent structure; d) if the matrix verb undergoes mutation then, a random sentence is generated to replace it; e) if the verb chosen for mutation is the head of a subclause, then a random subclause replaces it. Mutation Mutation can occur on any verb, noun, adverb, or adjective in the plot. If a noun, adverb or adjective is chosen to undergo mutation, then we simply substitute it with a new lexical item that is sufficiently similar (see Figure 4b for an example). Verbs, however, have structural importance in the stories and we cannot simply replace them without taking account of their arguments. If a matrix verb is chosen to undergo mutation, then a new random sentence is generated to replace the entire sentence (see Figure 4d). If it is a subclause, then it is replaced with a randomly generated clause, headed by a verb that has been seen in the corpus to co-occur with the matrix verb (Figure 4e). The sentence planner selects and fills template trees for generating random clauses. Mutation may also change the order of any two nodes in the graph in the hope that this will increase the story’s coherence or create some element of surprise (see Figure 4c). Selection To choose the plots for the next generation, we used fitness proportional selection (also know as roulette-wheel selection, Goldberg 1989) which chooses candidates randomly but with a bias towards those with a larger proportion of the population’s combined fitness. We do not want to always select the fittest candidates as there may be valid partial solutions held within less fit members of the population. However, we did employ some elitism by allowing the top 1% of solutions to be copied straight from one generation to the next. Note that our candidates may also represent invalid solutions. For instance, through crossover it is possible to create a plot in which all or some nodes are identical. If any such candidates are identified, they are assigned a low fitness, without however being eliminated from the population as some could be used to create fitter solutions. In a traditional GA, the fitness function deals with one optimization objective. It is possible to optimize several objectives either using a vot1567 ing model or more sophisticated methods such as Pareto ranking (Goldberg, 1989). Following previous work (Mellish et al., 1998) we used a single fitness function that scored candidates based on their coherence. Our function was learned from training data using the Entity Grid document representation proposed in Barzilay and Lapata (2007). An entity grid is a two-dimensional array in which columns correspond to entities and rows to sentences. Each cell indicates whether an entity appears in a given sentence or not and whether it is a subject, object or neither. For training, this representation is converted into a feature vector of entity transition sequences and a model is learnt from examples of coherent and incoherent stories. The latter can be easily created by permuting the sentences of coherent stories (assuming that the original story is more coherent than its permutations). In addition to coherence, in McIntyre and Lapata (2009) we used a scoring function based on interest which we approximated with lexical and syntactic features such as the number of noun/verb tokens/types, the number of subjects/objects, the number of letters, word familiarity, imagery, and so on. An interest-based scoring function made sense in our previous setup as a means of selecting unusual stories. However, in the context of genetic search such a function seems redundant as interesting stories emerge naturally through the operations of crossover and mutation. 5 Surface Realization Once the final generation of the population has been reached, the fittest story is selected for surface realization. The realizer takes each sentence in the story and reformulates it into input compatible with the RealPro (Lavoie and Rambow, 1997) text generation engine. Realpro creates several variants of the same story differing in the choice of determiners, number (singular or plural), and prepositions. A language model is then used to select the most probable realization (Knight and Hatzivassiloglou, 1995). Ideally, the realizer should also select an appropriate tense for the sentence. However, we make the simplifying assumption that all sentences are in the present tense. 6 Experimental Setup In this section we present our experimental set-up for assessing the performance of our story generator. We give details on our training corpus, system, parameters (such as the population size for the GA search), the baselines used for comparison, and explain how our system output was evaluated. Corpus The generator was trained on the same corpus used in McIntyre and Lapata (2009), 437 stories from the Andrew Lang fairy tales collection.5 The average story length is 125.18 sentences. The corpus contains 15,789 word tokens. Following McIntyre and Lapata, we discarded tokens that did not appear in the Children’s Printed Word Database6, a database of printed word frequencies as read by children aged between five and nine. From this corpus we extracted narrative schemas for 667 entities in total. We disregarded any graph that contained less than 10 nodes as too small. The graphs had on average 61.04 nodes, with an average clustering rate7 of 0.027 which indicates that they are substantially connected. Parameter Setting Considerable latitude is available when selecting parameters for the GA. These involve the population size, crossover, and mutation rates. To evaluate which setting was best, we asked two human evaluators to rate (on a 1–5 scale) stories produced with a population size ranging from 1,000 to 10,000, crossover rate of 0.1 to 0.6 and mutation rate of 0.001 to 0.1. For each run of the system a limit was set to 5,000 generations. The human ratings revealed that the best stories were produced for a population size of 10,000, a crossover rate of 0.1% and a mutation rate of 0.1%. Compared to previous work (e.g., Karamanis and Manurung 2002) our crossover rate may seem low and the mutation rate high. However, it makes intuitively sense, as high crossover may lead to incoherence by disrupting canonical action sequences found in the plots. On the other hand, a higher mutation will raise the likelihood of a lexical item being swapped for another and may improve overall coherence and interest. The fitness function was trained on 200 documents from the fairy tales collection using Joachims’s (2002) SVMlight package and entity transition sequences of length 2. The realizer was interfaced with a trigram language model trained on the British National Corpus with the SRI toolkit. 5Available from http://homepages.inf.ed.ac.uk/ s0233364/McIntyreLapata09/. 6http://www.essex.ac.uk/psychology/cpwd/ 7Clustering rate (or transitivity) is the number of triangles in the graph — sets of three vertices each of which is connected to each of the others. 1568 Evaluation We compared the stories generated by the GA against those produced by the rank-based system described in McIntyre and Lapata (2009) and a system that creates stories from the plot graph, without any stochastic search. Since plot graphs are weighted, we can simply select the graph with the highest weight. After expanding all lexical variables, the chosen plot graph will give rise to different stories (e.g., castle or temple in the example above). We select the story ranked highest according to our coherence function. In addition, we included a baseline which randomly selects sentences from the training corpus provided they contain either of the story protagonists (i.e., entities in the input sentence). Sentence length was limited to 12 words or less as this was on average the length of the sentences generated by our GA system. Each system created stories for 12 input sentences, resulting in 48 (4×12) stories for evaluation. The sentences used commonly occurring entities in the fairy tales corpus (e.g., The child watches the bird, The emperor rules the kingdom., The wizard casts the spell.). The stories were split into three sets containing four stories from each system but with only one story from each input sentence. All stories had the same length, namely five sentences. Human judges were presented with one of the three sets and asked to rate the stories on a scale of 1 to 5 for fluency (was the sentence grammatical?), coherence (does the story make sense overall?) and interest (how interesting is the story?). The stories were presented in random order and participants were told that all of them were generated by a computer program. They were instructed to rate more favorably interesting stories, stories that were comprehensible and overall grammatical. The study was conducted over the Internet using WebExp (Keller et al., 2009) and was completed by 56 volunteers, all self reported native English speakers. 7 Results Our results are summarized in Table 1 which lists the average human ratings for the four systems. We performed an Analysis of Variance (ANOVA) to examine the effect of system type on the story generation task. Statistical tests were carried out on the mean of the ratings shown in Table 1 for fluency, coherence, and interest. In terms of interest, the GA-based system is sigSystem Fluency Coherence Interest GA-based 3.09 2.48 2.36 Plot-based 3.03 2.36 2.14∗ Rank-based 1.96∗∗ 1.65∗ 1.85∗ Random 3.10 2.23∗ 2.20∗ Table 1: Human evaluation results: mean story ratings for four story generators; ∗: p < 0.05, ∗∗: p < 0.01, significantly different from GA-based system. nificantly better than the Rank-based, Plot-based and Random ones (using a Post-hoc Tukey test, α < 0.05). With regard to fluency, the Rankbased system is significantly worse than the rest (α < 0.01). Interestingly, the sentences generated by the GA and Plot-based systems are as fluent as those created by humans. Recall that the Random system, simply selects sentences from the training corpus. Finally, the GA system is significantly more coherent than the Rank-based and Random systems (α < 0.05), but not the Plot-based one. This is not surprising, the GA and Plot-based systems rely on similar plots to create a coherent story. The performance of the Random system is also inferior as it does not have any explicit coherence enforcing mechanism. The Rank-based system is perceived overall worse. As this system is also the least fluent, we conjecture that participants are influenced in their coherence judgments by the grammaticality of the stories. Overall our results indicate that an explicit story planner improves the quality of the generated stories, especially when coupled with a search mechanism that advantageously explores the search space. It is worth noting that the Plot-based system is relatively simple, however the explicit use of a story plot, seems to make up for the lack of sophisticated search and more elaborate linguistic information. Example stories generated by the four systems are shown in Table 2 for the input sentences The emperor rules the kingdom and The child watches the bird. Possible extensions and improvements to the current work are many and varied. Firstly, we could improve the quality of our plot graphs by taking temporal knowledge into account and making use of knowledge bases such as WordNet and ConceptNet (Liu and Davenport, 2004), a freely available commonsense knowledge base. Secondly, our fitness function optimizes one ob1569 PlotGA The emperor rules the kingdom. The kingdom holds on to the emperor. The emperor rides out of the kingdom. The kingdom speaks out against the emperor. The emperor lies. The child watches the bird. The bird weeps for the child. The child begs the bird to listen.The bird dresses up the child. The child grows up. Plot The emperor rules the kingdom. The emperor takes over. The emperor goes on to feel for the kingdom. Possibly the emperor sleeps. The emperor steals. The child watches the bird. The bird comes to eat away at the child. The child does thoroughly. The bird sees the child. The child sits down. Rank The emperor rules the kingdom. The kingdom lives from the reign to the emperor. The emperor feels that the brothers tempt a beauty into the game. The kingdom saves the life from crumbling the earth into the bird. The kingdom forces the whip into wiping the tears on the towel. The child watches the bird. The bird lives from the reign to the child. The child thanks the victory for blessing the thought. The child loves to hate the sun with the thought. The child hopes to delay the duty from the happiness. Random Exclaimed the emperor when Petru had put his question. In the meantime, mind you take good care of our kingdom. At first the emperor felt rather distressed. The dinner of an emperor! Thus they arrived at the court of the emperor. They cried, “what a beautiful child!” “No, that I cannot do, my child” he said at last. “What is the matter, dear child?” “You wicked child,” cried the Witch. Well, I will watch till the bird comes. Table 2: Stories generated by a system that uses plots and genetic search (PlotGA), a system that uses only plots (Plot), McIntyre and Lapata (2009)’s rank-based system (Rank) and a system that randomly pastes together sentences from the training corpus (Random). jective, namely coherence. In the future we plan to explore multiple objectives, such as whether the story is verbose, readable (using existing readability metrics), has two many or two few protagonists, and so on. Thirdly, our stories would benefit from some explicit modeling of discourse structure. Although the plot graph captures the progression of the actions in a story, we would also like to know where in the story these actions are likely to occur— some tend to appear in the beginning and others in the end. Such information would allow us to structure the stories better and render them more natural sounding. For example, an improvement would be the inclusion of proper endings, as the stories are currently cut off at an arbitrary point when the desired maximum length is reached. Finally, the fluency of the stories would benefit from generating referring expressions, multiple tense forms, indirect speech, aggregation and generally more elaborate syntactic structure. References Agudo, Bel´en Di´az, Pablo Gerv´as, and Frederico Peinado. 2004. A case based reasoning approach to story plot generation. In Proceedings of the 7th European Conference on Case-Based Reasoning. Springer, Madrid, Spain, pages 142–156. Barros, Leandro Motta and Soraia Raupp Musse. 2007. Planning algorithms for interactive storytelling. In Computers in Entertainment (CIE), Association for Computing Machinery (ACM), volume 5. Barzilay, Regina and Mirella Lapata. 2007. Modeling local coherence: An entity-based approach. Computational Linguistics 34(1):1–34. Briscoe, E. and J. Carroll. 2002. Robust accurate statistical annotation of general text. In Proceedings of the 3rd International Conference on Language Resources and Evaluation. Las Palmas, Gran Canaria, pages 1499–1504. Chambers, Nathanael and Dan Jurafsky. 2008. Unsupervised learning of narrative event chains. In Proceedings of 46th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies. Columbus, Ohio, pages 789–797. Chambers, Nathanael and Dan Jurafsky. 2009. 1570 Unsupervised learning of narrative schemas and their participants. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP. Singapore, pages 602–610. Cheng, Hua and Chris Mellish. 2000. Capturing the interaction between aggregation and text planning in two generation systems. In Proceedings of the 1st International Conference on Natural Language Generation. Mitzpe Ramon, Israel, pages 186–193. Fellbaum. 1998. WordNet: An Electronic Lexical Database (Language, Speech, and Communication). The MIT Press, Cambridge, Massachusetts. Fillmore, Charles J., Christopher R. Johnson, and Miriam R. L. Petruck. 2003. Background to FrameNet. International Journal of Lexicography 16:235–250. Goldberg, David E. 1989. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Longman Publishing Co., Inc., Boston, Massachusetts. Grishman, Ralph, Catherine Macleod, and Adam Meyers. 1994. COMLEX syntax: Building a computational lexicon. In Proceedings of the 15th COLING. Kyoto, Japan, pages 268–272. Joachims, Thorsten. 2002. Optimizing search engines using clickthrough data. In Proceedings of the 8th Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining. Edmonton, Alberta, pages 133–142. Karamanis, Nikiforos and Hisar Maruli Manurung. 2002. Stochastic text structuring using the principle of continuity. In Proceedings of the 2nd International Natural Language Generation Conference (INLG’02). pages 81–88. Keller, Frank, Subahshini Gunasekharan, Neil Mayo, and Martin Corley. 2009. Timing accuracy of web experiments: A case study using the WebExp software package. Behavior Research Methods 41(1):1–12. Knight, Kevin and Vasileios Hatzivassiloglou. 1995. Two-level, many-paths generation. In Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics (ACL’95). Cambridge, Massachusetts, pages 252–260. Korhonen, Y. Krymolowski, A. and E.J. Briscoe. 2006. A large subcategorization lexicon for natural language processing applications. In Proceedings of the 5th LREC. Genova, Italy. Lavoie, Benoit and Owen Rambow. 1997. A fast and portable realizer for text generation systems. In Proceedings of the 5th Conference on Applied Natural Language Processing. Washington, D.C., pages 265–268. Lin, Dekang. 1998. Automatic retrieval and clustering of similar words. In Proceedings of the 17th International Conference on Computational Linguistic. Montreal, Quebec, pages 768–774. Liu, Hugo and Glorianna Davenport. 2004. ConceptNet: a practical commonsense reasoning toolkit. BT Technology Journal 22(4):211–226. Liu, Hugo and Push Singh. 2002. Using commonsense reasoning to generate stories. In Proceedings of the 18th National Conference on Artificial Intelligence. Edmonton, Alberta, pages 957–958. McIntyre, Neil and Mirella Lapata. 2009. Learning to tell tales: A data-driven approach to story generation. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP. Singapore, pages 217–225. Meehan, James. 1977. An interactive program that writes stories. In Proceedings of the 5th International Joint Conference on Artificial Intelligence. Cambridge, Massachusetts, pages 91– 98. Mellish, Chris, Alisdair Knott, Jon Oberlander, and Mick O’Donnell. 1998. Experiments using stochastic search for text planning. In Proceedings of the 9th International Conference on Natural Language Generation. New Brunswick, New Jersey, pages 98–107. Mitchell, Melanie. 1998. An Introduction to Genetic Algorithms. MIT Press, Cambridge, Massachusetts. Pizzi, David, Fred Charles, Jean-Luc Lugrin, and Marc Cavazza. 2007. Interactive storytelling with literary feelings. In Proceedings of the 2nd International Conference on Affective Computing and Intelligent Interaction. Lisbon, Portugal, pages 630–641. 1571 Reiter, E and R Dale. 2000. Building NaturalLanguage Generation Systems. Cambridge University Press, Cambridge, UK. Riedl, Mark O. and R. Michael Young. 2004. A planning approach to story generation and history education. In Proceedings of the 3rd International Conference on Narrative and Interactive Learning Environments. Edinburgh, UK, pages 41–48. Shim, Yunju and Minkoo Kim. 2002. Automatic short story generator based on autonomous agents. In Proceedings of the 5th Pacific Rim International Workshop on Multi Agents. Tokyo, pages 151–162. Swartjes, I.M.T. and M. Theune. 2008. The virtual storyteller: story generation by simulation. In Proceedings of the 20th Belgian-Netherlands Conference on Artificial Intelligence, BNAIC 2008. Enschede, the Netherlands, pages 257– 264. Turner, Scott R. 1992. Ministrel: A Computer Model of Creativity and Sotrytelling. University of California, Los Angeles, California. Wu, Zhibiao and Martha Palmer. 1994. Verb semantics and lexical selection. In Proceedings of the 32nd Annual Meeting of the Association for Computational Linguistics. Las Cruces, New Mexico, pages 133–138. 1572
|
2010
|
158
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1573–1582, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Automated planning for situated natural language generation Konstantina Garoufiand Alexander Koller Cluster of Excellence “Multimodal Computing and Interaction” Saarland University, Saarbr¨ucken, Germany {garoufi,koller}@mmci.uni-saarland.de Abstract We present a natural language generation approach which models, exploits, and manipulates the non-linguistic context in situated communication, using techniques from AI planning. We show how to generate instructions which deliberately guide the hearer to a location that is convenient for the generation of simple referring expressions, and how to generate referring expressions with context-dependent adjectives. We implement and evaluate our approach in the framework of the Challenge on Generating Instructions in Virtual Environments, finding that it performs well even under the constraints of realtime generation. 1 Introduction The problem of situated natural language generation (NLG)—i.e., of generating natural language in the context of a physical (or virtual) environment—has received increasing attention in the past few years. On the one hand, this is because it is the foundation of various emerging applications, including human-robot interaction and mobile navigation systems, and is the focus of a current evaluation effort, the Challenges on Generating Instructions in Virtual Environments (GIVE; (Koller et al., 2010b)). On the other hand, situated generation comes with interesting theoretical challenges: Compared to the generation of pure text, the interpretation of expressions in situated communication is sensitive to the non-linguistic context, and this context can change as easily as the user can move around in the environment. One interesting aspect of situated communication from an NLG perspective is that this nonlinguistic context can be manipulated by the speaker. Consider the following segment of discourse between an instruction giver (IG) and an instruction follower (IF), which is adapted from the SCARE corpus (Stoia et al., 2008): (1) IG: Walk forward and then turn right. IF: (walks and turns) IG: OK. Now hit the button in the middle. In this example, the IG plans to refer to an object (here, a button); and in order to do so, gives a navigation instruction to guide the IF to a convenient location at which she can then use a simple referring expression (RE). That is, there is an interaction between navigation instructions (intended to manipulate the non-linguistic context in a certain way) and referring expressions (which exploit the non-linguistic context). Although such subdialogues are common in SCARE, we are not aware of any previous research that can generate them in a computationally feasible manner. This paper presents an approach to generation which is able to model the effect of an utterance on the non-linguistic context, and to intentionally generate utterances such as the above as part of a process of referring to objects. Our approach builds upon the CRISP generation system (Koller and Stone, 2007), which translates generation problems into planning problems and solves these with an AI planner. We extend the CRISP planning operators with the perlocutionary effects that uttering a particular word has on the physical environment if it is understood correctly; more specifically, on the position and orientation of the hearer. This allows the planner to predict the nonlinguistic context in which a later part of the utterance will be interpreted, and therefore to search for contexts that allow the use of simple REs. As a result, the work of referring to an object gets distributed over multiple utterances of low cognitive load rather than a single complex noun phrase. A second contribution of our paper is the generation of REs involving context-dependent adjectives: A button can be described as “the left blue 1573 button” even if there is a red button to its left. We model adjectives whose interpretation depends on the nominal phrases they modify, as well as on the non-linguistic context, by keeping track of the distractors that remain after uttering a series of modifiers. Thus, unlike most other RE generation approaches, we are not restricted to building an RE by simply intersecting lexically specified sets representing the extensions of different attributes, but can correctly generate expressions whose meaning depends on the context in a number of ways. In this way we are able to refer to objects earlier and more flexibly. We implement and evaluate our approach in the context of a GIVE NLG system, by using the GIVE-1 software infrastructure and a GIVE-1 evaluation world. This shows that our system generates an instruction-giving discourse as in (1) in about a second. It outperforms a mostly nonsituated baseline significantly, and compares well against a second baseline based on one of the top-performing systems of the GIVE-1 Challenge. Next to the practical usefulness this evaluation establishes, we argue that our approach to jointly modeling the grammatical and physical effects of a communicative action can also inform new models of the pragmatics of speech acts. Plan of the paper. We discuss related work in Section 2, and review the CRISP system, on which our work is based, in Section 3. We then show in Section 4 how we extend CRISP to generate navigation-and-reference discourses as in (1), and add context-dependent adjectives in Section 5. We evaluate our system in Section 6; Section 7 concludes and points to future work. 2 Related work The research reported here can be seen in the wider context of approaches to generating referring expressions. Since the foundational work of Dale and Reiter (1995), there has been a considerable amount of literature on this topic. Our work departs from the mainstream in two ways. First, it exploits the situated communicative setting to deliberately modify the context in which an RE is generated. Second, unlike most other RE generation systems, we allow the contribution of a modifier to an RE to depend both on the context and on the rest of the RE. We are aware of only one earlier study on generation of REs with focus on interleaving navigation and referring (Stoia et al., 2006). In this machine learning approach, Stoia et al. train classifiers that signal when the context conditions (e.g. visibility of target and distractors) are appropriate for the generation of an RE. This method can be then used as part of a content selection component of an NLG system. Such a component, however, can only inform a system on whether to choose navigation over RE generation at a given point of the discourse, and is not able to help it decide what kind of navigational instructions to generate so that subsequent REs become simple. To our knowledge, the only previous research on generating REs with context-dependent modifiers is van Deemter’s (2006) algorithm for generating vague adjectives. Unlike van Deemter, we integrate the RE generation process tightly with the syntactic realization, which allows us to generate REs with more than one context-dependent modifier and model the effect of their linear order on the meaning of the phrase. In modeling the context, we focus on the non-linguistic context and the influence of each of the RE’s words; this is in contrast to previous research on contextsensitive generation of REs, which mainly focused on the discourse context (Krahmer and Theune, 2002). Our interpretation of context-dependent modifiers picks up ideas by Kamp and Partee (1995) and implements them in a practical system, while our method of ordering modifiers is linguistically informed by the class-based paradigm (e.g., Mitchell (2009)). On the other hand, our work also stands in a tradition of NLG research that is based on AI planning. Early approaches (Perrault and Allen, 1980; Appelt, 1985) provided compelling intuitions for this connection, but were not computationally viable. The research we report here can be seen as combining Appelt’s idea of using planning for sentence-level NLG with a computationally benign variant of Perrault et al.’s approach of modeling the intended perlocutionary effects of a speech act as the effects of a planning operator. Our work is linked to a growing body of very recent work that applies modern planning research to various problems in NLG (Steedman and Petrick, 2007; Brenner and Kruijff-Korbayov´a, 2008; Benotti, 2009). It is directly based on Koller and Stone’s (2007) reimplementation of the SPUD generator (Stone et al., 2003) with planning. As far as we know, ours is the first system in the SPUD tradi1574 S:self NP:subj ↓ VP:self V:self pushes NP:obj ↓ semcontent: {push(self,subj,obj)} John NP:self semcontent: {John(self)} NP:self the N:self button semcontent: {button(self)} N:self red N * semcontent: {red(self)} (a) S:e NP:j ↓ VP:e V:e pushes NP:b1 ↓ (b) John NP:j NP:b1 the N:b1 button N:b1 red N * Figure 1: (a) An example grammar; (b) a derivation of “John pushes the red button” using (a). tion that explicitly models the context change effects of an utterance. While nothing in our work directly hinges on this, we implemented our approach in the context of an NLG system for the GIVE Challenge (Koller et al., 2010b), that is, as an instruction giving system for virtual worlds. This makes our system comparable with other approaches to instruction giving implemented in the GIVE framework. 3 Sentence generation as planning Our work is based on the CRISP system (Koller and Stone, 2007), which encodes sentence generation with tree-adjoining grammars (TAG; (Joshi and Schabes, 1997)) as an AI planning problem and solves that using efficient planners. It then decodes the resulting plan into a TAG derivation, from which it can read off a sentence. In this section, we briefly recall how this works. For space reasons, we will present primarily examples instead of definitions. 3.1 TAG sentence generation The CRISP generation problem (like that of SPUD (Stone et al., 2003)) assumes a lexicon of entries consisting of a TAG elementary tree annotated with semantic and pragmatic information. An example is shown in Fig. 1a. In addition to the elementary tree, each lexicon entry specifies its semantic content and possibly a semantic requirement, which can express certain presuppositions triggered by this entry. The nodes in the tree may be labeled with argument names such as semantic roles, which specify the participants in the relation expressed by the lexicon entry; in the example, every entry uses the semantic role self representing the event or individual itself, and the entry for “pushes” furthermore uses subj and obj for the subject and object argument, respectively. We combine here for simplicity the entries for “the” and “button” into “the button”. For generation, we assume as input a knowledge base and a communicative goal in addition to the grammar. The goal is to compute a derivation that expresses the communicative goal in a sentence that is grammatically correct and complete; whose meaning is justified by the knowledge base; and in which all REs can be resolved to unique individuals in the world by the hearer. Let’s say, for example, that we have a knowledge base {push(e, j, b1), John(j), button(b1), button(b2), red(b1)}. Then we can combine instances of the trees for “John”, “pushes”, and “the button” into a grammatically complete derivation. However, because both b1 and b2 satisfy the semantic content of “the button”, we must adjoin “red” into the derivation to make the RE refer uniquely to b1. The complete derivation is shown in Fig. 1b; we can read off the output sentence “John pushes the red button” from the leaves of the derived tree we build in this way. 3.2 TAG generation as planning In the CRISP system, Koller and Stone (2007) show how this generation problem can be solved by converting it into a planning problem (Nau et al., 2004). The basic idea is to encode the partial derivation in the planning state, and to encode the action of adding each elementary tree in the planning operators. The encoding of our example as a planning problem is shown in Fig. 2. In the example, we start with an initial state which contains the entire knowledge base, plus atoms subst(S, root) and ref(root, e) expressing that we want to generate a sentence about the event e. We can then apply the (instantiated) action pushes(root, n1, n2, n3, e, j, b1), which models the act of substituting the elementary tree for “pushes” 1575 pushes(u, u1, u2, un, x, x1, x2): Precond: subst(S, u), ref(u, x), push(x, x1, x2), current(u1), next(u1, u2), next(u2, un) Effect: ¬subst(S, u), subst(NP, u1), subst(NP, u2), ref(u1, x1), ref(u2, x2), ∀y.distractor(u1, y), ∀y.distractor(u2, y) John(u, x): Precond: subst(NP, u), ref(u, x), John(x) Effect: ¬subst(NP, u), ∀y.¬John(y) →¬distractor(u, y) the-button(u, x): Precond: subst(NP, u), ref(u, x), button(x) Effect: ¬subst(NP, u), canadjoin(N, u), ∀y.¬button(y) →¬distractor(u, y) red(u, x): Precond: canadjoin(N, u), ref(u, x), red(x) Effect: ∀y.¬red(y) →¬distractor(u, y) Figure 2: CRISP planning operators for the elementary trees in Fig. 1. into the substitution node root: It can only be applied because root is an unfilled substitution node (precondition subst(S, root)), and its effect is to remove subst(S, root) from the planning state while adding two new atoms subst(NP, n1) and subst(NP, n2) for the substitution nodes of the “pushes” tree. The planning state maintains information about which individual each node refers to in the ref atoms. The current and next atoms are needed to select unused names for newly introduced syntax nodes.1 Finally, the action introduces a number of distractor atoms including distractor(n2, e) and distractor(n2, b2), expressing that the RE at n2 can still be misunderstood by the hearer as e or b2. In this new state, all subst and distractor atoms for n1 can be eliminated with the action John(n1, j). We can also apply the action the-button(n2, b1) to eliminate subst(NP, n2) and distractor(n2, e), since e is not a button. However distractor(n2, b2) remains. Now because the action the-button also introduced the atom canadjoin(N, n2), we can remove the final distractor atom by applying red(n2, b1). This brings us into a goal state, and we are done. Goal states in CRISP planning problems are characterized by axioms such as ∀A∀u.¬subst(A, u) (encoding grammatical completeness) and ∀u∀x.¬distractor(u, x) (requiring unique reference). 1This is a different solution to the name-selection problem than in Koller and Stone (2007). It is simpler and improves computational efficiency. 1 2 3 4 1 2 3 4 b1 b2 b3 f1 north Figure 3: An example map for instruction giving. 3.3 Decoding the plan An AI planner such as FF (Hoffmann and Nebel, 2001) can compute a plan for a planning problem that consists of the planning operators in Fig. 2 and a specification of the initial state and the goal. We can then decode this plan into the TAG derivation shown in Fig. 1b. The basic idea of this decoding step is that an action with a precondition subst(A, u) fills the substitution node u, while an action with a precondition canadjoin(A, u) adjoins into a node of category A in the elementary tree that was substituted into u. CRISP allows multiple trees to adjoin into the same node. In this case, the decoder executes the adjunctions in the order in which they occur in the plan. 4 Context manipulation We are now ready to describe our NLG approach, SCRISP (“Situated CRISP”), which extends CRISP to take the non-linguistic context of the generated utterance into account, and deliberately manipulate it to simplify RE generation. As a simplified version of our introductory instruction giving example (1), consider the map in Fig. 3. The instruction follower (IF), who is located on the map at position pos3,2 facing north, sees the scene from the first-person perspective as in Fig. 7. Now an instruction giver (IG) could instruct the IF to press the button b1 in this scene by saying “push the button on the wall to your left”. Interpreting this instruction is difficult for the IF because it requires her to either memorize the RE until she has turned to see the button, or to perform a mental rotation task to visualize b1 internally. Alternatively, the IG can first instruct the IF to “turn left”; once the IF has done this, the IG can then simply say “now push the button in front 1576 S:self V:self push NP:obj ↓ semreq: visible(p, o, obj) nonlingcon: player–pos(p), player–ori(o) impeff: push(obj) S:self V:self turn Adv left nonlingcon: player–ori(o1), next–ori–left(o1, o2) nonlingeff: ¬player–ori(o1), player–ori(o2) impeff: turnleft S:self S:self * S:other ↓ and Figure 4: An example SCRISP lexicon. of you”. This lowers the cognitive load on the IF, and presumably improves the rate of correctly interpreted REs. SCRISP is capable of deliberately generating such context-changing navigation instructions. The key idea of our approach is to extend the CRISP planning operators with preconditions and effects that describe the (simulated) physical environment: A “turn left” action, for example, modifies the IF’s orientation in space and changes the set of visible objects; a “push” operator can then pick up this changed set and restrict the distractors of the forthcoming RE it introduces (i.e. “the button”) to only objects that are visible in the changed context. We also extend CRISP to generate imperative rather than declarative sentences. 4.1 Situated CRISP We define a lexicon for SCRISP to be a CRISP lexicon in which every lexicon entry may also describe non-linguistic conditions, non-linguistic effects and imperative effects. Each of these is a set of atoms over constants, semantic roles, and possibly some free variables. Non-linguistic conditions specify what must be true in the world so a particular instance of a lexicon entry can be uttered felicitously; non-linguistic effects specify what changes uttering the word brings about in the world; and imperative effects contribute to the IF’s “to-do list” (Portner, 2007) by adding the properties they denote. A small lexicon for our example is shown in Fig. 4. This lexicon specifies that saying “push X” puts pushing X on the IF’s to-do list, and carries the presupposition that X must be visible from the location where “push X” is uttered; this reflects our simplifying assumption that the IG can turnleft(u, x, o1, o2): Precond: subst(S, u), ref(u, x), player–ori(o1), next–ori–left(o1, o2), . . . Effect: ¬subst(S, u), ¬player–ori(o1), player–ori(o2), to–do(turnleft), . . . push(u, u1, un, x, x1, p, o): Precond: subst(S, u), ref(u, x), player–pos(p), player–ori(o), visible(p, o, x1), . . . Effect: ¬subst(S, u), subst(NP, u1), ref(u1, x1), ∀y.(y ̸= x1 ∧visible(p, o, y) →distractor(u1, y)), to–do(push(x1)), canadjoin(S, u), . . . and(u, u1, un, e1, e2): Precond: canadjoin(S, u), ref(u, e1), . . . Effect: subst(S, u1), ref(u1, e2), . . . Figure 5: SCRISP planning operators for the lexicon in Fig. 4. only refer to objects that are currently visible. Similarly, “turn left” puts turning left on the IF’s agenda. In addition, the lexicon entry for “turn left” specifies that, under the assumption that the IF understands and follows the instruction, they will turn 90 degrees to the left after hearing it. The planning operators are written in a way that assumes that the intended (perlocutionary) effects of an utterance actually come true. This assumption is crucial in connecting the non-linguistic effects of one SCRISP action to the non-linguistic preconditions of another, and generalizes to a scalable model of planning perlocutionary acts. We discuss this in more detail in Koller et al. (2010a). We then translate a SCRISP generation problem into a planning problem. In addition to what CRISP does, we translate all non-linguistic conditions into preconditions and all non-linguistic effects into effects of the planning operator, adding any free variables to the operator’s parameters. An imperative effect P is translated into an effect to–do(P). The operators for the example lexicon of Fig. 4 are shown in Fig. 5. Finally, we add information about the situated environment to the initial state, and specify the planning goal by adding to–do(P) atoms for each atom P that is to be placed on the IF’s agenda. 4.2 An example Now let’s look at how this generates the appropriate instructions for our example scene of Fig. 3. We encode the state of the world as depicted in the map in an initial state which contains, among others, the atoms player–pos(pos3,2), player–ori(north), next–ori–left(north, west), 1577 visible(pos3,2, west, b1), etc.2 We want the IF to press b1, so we add to–do(push(b1)) to the goal. We can start by applying the action turnleft(root, e, north, west) to the initial state. Next to the ordinary grammatical effects from CRISP, this action makes player–ori(west) true. The new state does not contain any subst atoms, but we can continue the sentence by adjoining “and”, i.e. by applying the action and(root, n1, n2, e, e1). This produces a new atom subst(S, e1), which satisfies one precondition of push(n1, n2, n3, e1, b1, pos3,2, west). Because turnleft changed the player orientation, the visible precondition of push is now satisfied too (unlike in the initial state, in which b1 was not visible). Applying the action push now introduces the need to substitute a noun phrase for the object, which we can eliminate with an application of the-button(n2, b1) as in Subsection 3.2. Since there are no other visible buttons from pos3,2 facing west, there are no remaining distractor atoms at this point, and a goal state has been reached. Together, this four-step plan decodes into the sentence “turn left and push the button”. The final state contains the atoms to–do(push(b1)) and to–do(turnleft), indicating that an IF that understands and accepts this instruction also accepts these two commitments into their to-do list. 5 Generating context-dependent adjectives Now consider if we wanted to instruct the IF to press b2 in Fig. 3 instead of b1, say with the instruction “push the left button”. This is still challenging, because (like most other approaches to RE generation) CRISP interprets adjectives by simply intersecting all their extensions. In the case of “left”, the most reasonable way to do this would be to interpret it as “leftmost among all visible objects”; but this is f1 in the example, and so there is no distinguishing RE for b2. In truth, spatial adjectives like “left” and “upper” depend on the context in two different ways. On the one hand, they are interpreted with respect to the current spatio-visual context, in that what is on the left depends on the current position and orientation of the hearer. On the other hand, they also 2In a more complex situation, it may be infeasible to exhaustively model visibility in this way. This could be fixed by connecting the planner to an external spatial reasoner (Dornhege et al., 2009). left(u, x): Precond: ∀y.¬(distractor(u, y) ∧left–of(y, x)), canadjoin(N, u), ref(u, x) Effect: ∀y.(left–of(x, y) →¬distractor(u, y)), premod–index(u, 2), . . . red(u, x): Precond: red(x), canadjoin(N, u), ref(u, x), ¬premod–index(u, 2) Effect: ∀y.(¬red(y) →¬distractor(u, y)), premod–index(u, 1), . . . Figure 6: SCRISP operators for contextdependent and context-independent adjectives. depend on the meaning of the phrase they modify: “the left button” is not necessarily both a button and further to the left than all other objects, it is only the leftmost object among the buttons. We will now show how to extend SCRISP so it can generate REs that use such context-dependent adjectives. 5.1 Context-dependence of adjectives in SCRISP As a planning-based approach to NLG, SCRISP is not limited to simply intersecting sets of potential referents that only depend on the attributes that contribute to an RE: Distractors are removed by applying operators which may have contextsensitive conditions depending on the referent and the distractors that are still left. Our encoding of context-dependent adjectives as planning operators is shown in Fig. 6. We only show the operators here for lack of space; they can of course be computed automatically from lexicon entries. In addition to the ordinary CRISP preconditions, the left operator has a precondition requiring that no current distractor for the RE u is to the left of x, capturing a presupposition of the adjective. Its effect is that everything that is to the right of x is no longer a distractor for u. Notice that we allow that there may still be distractors after left has been applied (above or below x); we only require unique reference in the goal state. (Ignore the premod–index part of the effect for now; we will get to that in a moment.) Let’s say that we are computing a plan for referring to b2 in the example map of Fig. 3, starting with push(root, n1, n2, e, b2, pos3,1, north) and the-button(n1, b2). The state after these two actions is not a goal state, because it still contains the atom distractor(n1, b3) (the plant f1 was removed as a distractor by the action the-button). 1578 Now assume that we have modeled the spatial relations between all objects in the initial state in left–of and above atoms; in particular, we have left–of(b2, b3). Then the action instance left(n1, b2) is applicable in this state, as there is no other object that is still a distractor in this state and that is to the left of b2. Applying left removes distractor(n1, b3) from the state. Thus we have reached a goal state; the complete plan decodes to the sentence “push the left button”. This system is sensitive to the order in which operators for context-dependent adjectives are applied. To generate the RE “the upper left button”, for instance, we first apply the left action and then the upper action, and therefore upper only needs to remove distractors in the leftmost position. On the other hand, the RE “the left upper button” corresponds to first applying upper and then left. These action sequences succeed in removing all distractors for different context states, which is consistent with the difference in meaning between the two REs. Furthermore, notice that the adjective operators themselves do not interact directly with the encoding of the context in atoms like visible and player–pos, just like the noun operators in Section 4 didn’t. The REs to which the adjectives and nouns contribute are introduced by verb operators; it is these verb operators that inspect the current context and initialize the distractor set for the new RE appropriately. This makes the correctness of the generated sentence independent of the order in which noun and adjective operators occur in the plan. We only need to ensure that the verbs are ordered correctly, and the workload of modeling interactions with the non-linguistic context is limited to a single place in the encoding. 5.2 Adjective word order One final challenge that arises in our system is to generate the adjectives in the correct order, which on top of semantically valid must be linguistically acceptable. In particular, it is known that some types of adjectives are limited with respect to the word order in which they can occur in a noun phrase. For instance, “large foreign financial firms” sounds perfectly acceptable, but “? foreign large financial firms” sounds odd (Shaw and Hatzivassiloglou, 1999). In our setting, some adjective orders are forbidden because only one order produces a correct and distinguishing descripFigure 7: The IF’s view of the scene in Fig. 3, as rendered by the GIVE client. tion of the target referent (cf. “upper left” vs. “left upper” example above). However, there are also other constraints at work: “? the red left button” is rather odd even when it is a semantically correct description, whereas “the left red button” is fine. To ensure that SCRISP chooses to generate these adjectives correctly, we follow a class-based approach to the premodifier ordering problem (Mitchell, 2009). In our lexicon we assign adjectives denoting spatial relations (“left”) to one class and adjectives denoting color (“red”) to another; then we require that spatial adjectives must always precede color adjectives. We enforce this by keeping track of the current premodifier index of the RE in atoms of the form premod–index. Any newly generated RE node starts off with a premodifier index of zero; adjoining an adjective of a certain class then raises this number to the index for that class. As the operators in Fig. 6 illustrate, color adjectives such as “red” have index one and can only be used while the index is not higher; once an adjective from a higher class (such as “left”, of a class with index two) is used, the premod–index precondition of the “red” operator will fail. For this reason, we can generate a plan for “the left red button”, but not for “? the red left button”, as desired. 6 Evaluation To establish the quality of the generated instructions, we implemented SCRISP as part of a generation system in the GIVE-1 framework, and evaluated it against two baselines. GIVE-1 was the First Challenge on Generating Instructions in Virtual Environments, which was completed in 2009 1579 SCRISP 1. Turn right and move one step. 2. Push the right red button. Baseline A 1. Press the right red button on the wall to your right. Baseline B 1. Turn right. 2. Walk forward 3 steps. 3. Turn right. 4. Walk forward 1 step. 5. Turn left. 6. Good! Now press the left button. Table 1: Example system instructions generated in the same scene. REs for the target are typeset in boldface. (Koller et al., 2010b). In this challenge, systems must generate real-time instructions that help users perform a task in a treasure-hunt virtual environment such as the one shown in Fig. 7. We conducted our evaluation in World 2 from GIVE-1, which was deliberately designed to be challenging for RE generation. The world consists of one room filled with several objects and buttons, most of which cannot be distinguished by simple descriptions. Moreover, some of those may activate an alarm and cause the player to lose the game. The player’s moves and turns are discrete and the NLG system has complete and accurate real-time information about the state of the world. Instructions that each of the three systems under comparison generated in an example scene of the evaluation world are presented in Table 1. The evaluation took place online via the Amazon Mechanical Turk, where we collected 25 games for each system. We focus on four measures of evaluation: success rates for solving the task and resolving the generated REs, average task completion time (in seconds) for successful games, and average distance (in steps) between the IF and the referent at the time when the RE was generated. As in the challenge, the task is considered as solved if the player has correctly been led through manipulating all target objects required to discover and collect the treasure; in World 2, the minimum number of such targets is eight. An RE is successfully resolved if it results in the manipulation of the referent, whereas manipulation of an alarm-triggering distractor ends the game unsuccessfully. 6.1 The SCRISP system Our system receives as input a plan for what the IF should do to solve the task, and successively takes object-manipulating actions as the commusuccess RE rate time success distance SCRISP 69% 306 71% 2.49 Baseline A 16%** 230 49%** 1.97* Baseline B 84% 288 81%* 2.00* Table 2: Evaluation results. Differences to SCRISP are significant at *p < .05, **p < .005 (Pearson’s chi-square test for system success rates; unpaired two-sample t-test for the rest). nicative goals for SCRISP. Then, for each of the communicative goals, it generates instructions using SCRISP, segments them into navigation and action parts, and presents these to the user as separate instructions sequentially (see Table 1). For each instruction, SCRISP thus draws from a knowledge base of about 1500 facts and a grammar of about 30 lexicon entries. We use the FF planner (Hoffmann and Nebel, 2001; Koller and Hoffmann, 2010) to solve the planning problems. The maximum planning time for any instruction is 1.03 seconds on a 3.06 GHz Intel Core 2 Duo CPU. So although our planning-based system tackles a very difficult search problem, FF is very good at solving it—fast enough to generate instructions in real time. 6.2 Comparison with Baseline A Baseline A is a very basic system designed to simulate the performance of a classical RE generation module which does not attempt to manipulate the visual context. We hand-coded a correct distinguishing RE for each target button in the world; the only way in which Baseline A reacts to changes of the context is to describe on which wall the button is with respect to the user’s current orientation (e.g. “Press the right red button on the wall to your right”). As Table 2 shows, our system guided 69% of users to complete the task successfully, compared to only 16% for Baseline A (difference is statistically significant at p < .005; Pearson’s chisquare test). This is primarily because only 49% of the REs generated by Baseline A were successful. This comparison illustrates the importance of REs that minimize the cognitive load on the IF to avoid misunderstandings. 6.3 Comparison with Baseline B Baseline B is a corrected and improved version of the “Austin” system (Chen and Karpov, 2009), 1580 one of the best-performing systems of the GIVE-1 Challenge. Baseline B, like the original “Austin” system, issues navigation instructions by precomputing the shortest path from the IF’s current location to the target, and generates REs using the description logic based algorithm of Areces et al. (2008). Unlike the original system, which inflexibly navigates the user all the way to the target, Baseline B starts off with navigation, and opportunistically instructs the IF to push a button once it has become visible and can be described by a distinguishing RE. We fixed bugs in the original implementation of the RE generation module, so that Baseline B generates only unambiguous REs. The module nonetheless naively treats all adjectives as intersective and is not sensitive to the context of their comparison set. Specifically, a button cannot be referred to as “the right red button” if it is not the rightmost of all visible objects—which explains the long chain of navigational instructions the system produced in Table 1. We did not find any significant differences in the success rates or task completion times between this system and SCRISP, but the former achieved a higher RE success rate (see Table 2). However, a closer analysis shows that SCRISP was able to generate REs from significantly further away. This means that SCRISP’s RE generator solves a harder problem, as it typically has to deal with more visible distractors. Furthermore, because of the increased distance, the system’s execution monitoring strategies (e.g. for detection and repair of misunderstandings) become increasingly important, and this was not a focus of this work. In summary, then, we take the results to mean that SCRISP performs quite capably in comparison to a top-ranked GIVE-1 system. 7 Conclusion In this paper, we have shown how situated instructions can be generated using AI planning. We exploited the planner’s ability to model the perlocutionary effects of communicative actions for efficient generation. We showed how this made it possible to generate instructions that manipulate the non-linguistic context in convenient ways, and to generate correct REs with context-dependent adjectives. We believe that this illustrates the power of a planning-based approach to NLG to flexibly model very different phenomena. An interesting topic for future work, for instance, is to expand our notion of context by taking visual and discourse salience into account when generating REs. In addition, we plan to experiment with assigning costs to planning operators in a metric planning problem (Hoffmann, 2002) in order to model the cognitive cost of an RE (Krahmer et al., 2003) and compute minimal-cost instruction sequences. On a more theoretical level, the SCRISP actions model the physical effects of a correctly understood and grounded instruction directly as effects of the planning operator. This is computationally much less complex than classical speech act planning (Perrault and Allen, 1980), in which the intended physical effect comes at the end of a long chain of inferences. But our approach is also very optimistic in estimating the perlocutionary effects of an instruction, and must be complemented by an appropriate model of execution monitoring. What this means for a novel scalable approach to the pragmatics of speech acts (Koller et al., 2010a) is, we believe, an interesting avenue for future research. Acknowledgments. We are grateful to J¨org Hoffmann for improving the efficiency of FF in the SCRISP domain at a crucial time, and to Margaret Mitchell, Matthew Stone and Kees van Deemter for helping us expand our view of the contextdependent adjective generation problem. We also thank Ines Rehbein and Josef Ruppenhofer for testing early implementations of our system, and Andrew Gargett as well as the reviewers for their helpful comments. References Douglas E. Appelt. 1985. Planning English sentences. Cambridge University Press, Cambridge, England. Carlos Areces, Alexander Koller, and Kristina Striegnitz. 2008. Referring expressions as formulas of description logic. In Proceedings of the 5th International Natural Language Generation Conference, pages 42–49, Salt Fork, Ohio, USA. Luciana Benotti. 2009. Clarification potential of instructions. In Proceedings of the SIGDIAL 2009 Conference, pages 196–205, London, UK. Michael Brenner and Ivana Kruijff-Korbayov´a. 2008. A continual multiagent planning approach to situated dialogue. In Proceedings of the 12th Workshop on the Semantics and Pragmatics of Dialogue, London, UK. 1581 David Chen and Igor Karpov. 2009. The GIVE-1 Austin system. In The First GIVE Challenge: System descriptions. http://www.give-challenge.org/ research/files/GIVE-09-Austin.pdf. Robert Dale and Ehud Reiter. 1995. Computational interpretations of the Gricean maxims in the generation of referring expressions. Cognitive Science, 19. Christian Dornhege, Patrick Eyerich, Thomas Keller, Sebastian Tr¨ug, Michael Brenner, and Bernhard Nebel. 2009. Semantic attachments for domainindependent planning systems. In Proceedings of the 19th International Conference on Automated Planning and Scheduling, pages 114–121. J¨org Hoffmann and Bernhard Nebel. 2001. The FF planning system: Fast plan generation through heuristic search. Journal of Artificial Intelligence Research, 14:253–302. J¨org Hoffmann. 2002. Extending FF to numerical state variables. In Proceedings of the 15th European Conference on Artificial Intelligence, Lyon, France. Aravind K. Joshi and Yves Schabes. 1997. TreeAdjoining Grammars. In G. Rozenberg and A. Salomaa, editors, Handbook of Formal Languages, volume 3, pages 69–123. Springer-Verlag, Berlin, Germany. Hans Kamp and Barbara Partee. 1995. Prototype theory and compositionality. Cognition, 57(2):129 – 191. Alexander Koller and J¨org Hoffmann. 2010. Waking up a sleeping rabbit: On natural-language sentence generation with FF. In Proceedings of the 20th International Conference on Automated Planning and Scheduling, Toronto, Canada. Alexander Koller and Matthew Stone. 2007. Sentence generation as planning. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, Prague, Czech Republic. Alexander Koller, Andrew Gargett, and Konstantina Garoufi. 2010a. A scalable model of planning perlocutionary acts. In Proceedings of the 14th Workshop on the Semantics and Pragmatics of Dialogue, Poznan, Poland. Alexander Koller, Kristina Striegnitz, Donna Byron, Justine Cassell, Robert Dale, Johanna Moore, and Jon Oberlander. 2010b. The First Challenge on Generating Instructions in Virtual Environments. In M. Theune and E. Krahmer, editors, Empirical Methods in Natural Language Generation, volume 5790 of LNCS, pages 337–361. Springer, Berlin/Heidelberg. To appear. Emiel Krahmer and Mariet Theune. 2002. Efficient context-sensitive generation of referring expressions. In Kees van Deemter and Rodger Kibble, editors, Information Sharing: Reference and Presupposition in Language Generation and Interpretation, pages 223–264. CSLI Publications. Emiel Krahmer, Sebastiaan van Erk, and Andr´e Verleg. 2003. Graph-based generation of referring expressions. Computational Linguistics, 29(1):53–72. Margaret Mitchell. 2009. Class-based ordering of prenominal modifiers. In Proceedings of the 12th European Workshop on Natural Language Generation, pages 50–57, Athens, Greece. Dana Nau, Malik Ghallab, and Paolo Traverso. 2004. Automated Planning: Theory and Practice. Morgan Kaufmann. C. Raymond Perrault and James F. Allen. 1980. A plan-based analysis of indirect speech acts. American Journal of Computational Linguistics, 6(3– 4):167–182. Paul Portner. 2007. Imperatives and modals. Natural Language Semantics, 15(4):351–383. James Shaw and Vasileios Hatzivassiloglou. 1999. Ordering among premodifiers. In Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics, pages 135–143, College Park, Maryland, USA. Mark Steedman and Ronald P. A. Petrick. 2007. Planning dialog actions. In Proceedings of the 8th SIGdial Workshop on Discourse and Dialogue, pages 265–272, Antwerp, Belgium. Laura Stoia, Donna K. Byron, Darla Magdalene Shockley, and Eric Fosler-Lussier. 2006. Sentence planning for realtime navigational instructions. In NAACL ’06: Proceedings of the Human Language Technology Conference of the NAACL, pages 157– 160, Morristown, NJ, USA. Laura Stoia, Darla M. Shockley, Donna K. Byron, and Eric Fosler-Lussier. 2008. SCARE: A situated corpus with annotated referring expressions. In Proceedings of the 6th International Conference on Language Resources and Evaluation, Marrakech, Morocco. Matthew Stone, Christine Doran, Bonnie Webber, Tonia Bleam, and Martha Palmer. 2003. Microplanning with communicative intentions: The SPUD system. Computational Intelligence, 19(4):311– 381. Kees van Deemter. 2006. Generating referring expressions that involve gradable properties. Computational Linguistics, 32(2). 1582
|
2010
|
159
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 148–156, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Pseudo-word for Phrase-based Machine Translation Xiangyu Duan Min Zhang Haizhou Li Institute for Infocomm Research, A-STAR, Singapore {Xduan, mzhang, hli}@i2r.a-star.edu.sg Abstract The pipeline of most Phrase-Based Statistical Machine Translation (PB-SMT) systems starts from automatically word aligned parallel corpus. But word appears to be too fine-grained in some cases such as non-compositional phrasal equivalences, where no clear word alignments exist. Using words as inputs to PBSMT pipeline has inborn deficiency. This paper proposes pseudo-word as a new start point for PB-SMT pipeline. Pseudo-word is a kind of basic multi-word expression that characterizes minimal sequence of consecutive words in sense of translation. By casting pseudo-word searching problem into a parsing framework, we search for pseudo-words in a monolingual way and a bilingual synchronous way. Experiments show that pseudo-word significantly outperforms word for PB-SMT model in both travel translation domain and news translation domain. 1 Introduction The pipeline of most Phrase-Based Statistical Machine Translation (PB-SMT) systems starts from automatically word aligned parallel corpus generated from word-based models (Brown et al., 1993), proceeds with step of induction of phrase table (Koehn et al., 2003) or synchronous grammar (Chiang, 2007) and with model weights tuning step. Words are taken as inputs to PB-SMT at the very beginning of the pipeline. But there is a deficiency in such manner that word is too finegrained in some cases such as non-compositional phrasal equivalences, where clear word alignments do not exist. For example in Chinese-toEnglish translation, “想” and “would like to” constitute a 1-to-n phrasal equivalence, “多少 钱” and “how much is it” constitute a m-to-n phrasal equivalence. No clear word alignments are there in such phrasal equivalences. Moreover, should basic translational unit be word or coarsegrained multi-word is an open problem for optimizing SMT models. Some researchers have explored coarsegrained translational unit for machine translation. Marcu and Wong (2002) attempted to directly learn phrasal alignments instead of word alignments. But computational complexity is prohibitively high for the exponentially large number of decompositions of a sentence pair into phrase pairs. Cherry and Lin (2007) and Zhang et al. (2008) used synchronous ITG (Wu, 1997) and constraints to find non-compositional phrasal equivalences, but they suffered from intractable estimation problem. Blunsom et al. (2008; 2009) induced phrasal synchronous grammar, which aimed at finding hierarchical phrasal equivalences. Another direction of questioning word as basic translational unit is to directly question word segmentation on languages where word boundaries are not orthographically marked. In Chineseto-English translation task where Chinese word boundaries are not marked, Xu et al. (2004) used word aligner to build a Chinese dictionary to resegment Chinese sentence. Xu et al. (2008) used a Bayesian semi-supervised method that combines Chinese word segmentation model and Chinese-to-English translation model to derive a Chinese segmentation suitable for machine translation. There are also researches focusing on the impact of various segmentation tools on machine translation (Ma et al. 2007; Chang et al. 2008; Zhang et al. 2008). Since there are many 1-to-n phrasal equivalences in Chinese-to-English translation (Ma and Way. 2009), only focusing on Chinese word as basic translational unit is not adequate to model 1-to-n translations. Ma and Way (2009) tackle this problem by using word aligner to bootstrap bilingual segmentation suitable for machine translation. Lambert and Banchs (2005) detect bilingual multi-word ex148 pressions by monotonically segmenting a given Spanish-English sentence pair into bilingual units, where word aligner is also used. IBM model 3, 4, 5 (Brown et al., 1993) and Deng and Byrne (2005) are another kind of related works that allow 1-to-n alignments, but they rarely questioned if such alignments exist in word units level, that is, they rarely questioned word as basic translational unit. Moreover, m-ton alignments were not modeled. This paper focuses on determining the basic translational units on both language sides without using word aligner before feeding them into PBSMT pipeline. We call such basic translational unit as pseudo-word to differentiate with word. Pseudo-word is a kind of multi-word expression (includes both unary word and multi-word). Pseudo-word searching problem is the same to decomposition of a given sentence into pseudowords. We assume that such decomposition is in the Gibbs distribution. We use a measurement, which characterizes pseudo-word as minimal sequence of consecutive words in sense of translation, as potential function in Gibbs distribution. Note that the number of decomposition of one sentence into pseudo-words grows exponentially with sentence length. By fitting decomposition problem into parsing framework, we can find optimal pseudo-word sequence in polynomial time. Then we feed pseudo-words into PB-SMT pipeline, and find that pseudo-words as basic translational units improve translation performance over words as basic translational units. Further experiments of removing the power of higher order language model and longer max phrase length, which are inherent in pseudowords, show that pseudo-words still improve translational performance significantly over unary words. This paper is structured as follows: In section 2, we define the task of searching for pseudowords and its solution. We present experimental results and analyses of using pseudo-words in PB-SMT model in section 3. The conclusion is presented at section 4. 2 Searching for Pseudo-words Pseudo-word searching problem is equal to decomposition of a given sentence into pseudowords. We assume that the distribution of such decomposition is in the form of Gibbs distribution as below: ) exp( 1 ) | ( ∑ = y Sig X Y P where X denotes the sentence, Y denotes a decomposition of X. Sig function acts as potential function on each multi-word yk, and ZX acts as partition function. Note that the number of yk is not fixed given X because X can be decomposed into various number of multi-words. Given X, ZX is fixed, so searching for optimal decomposition is as below: ∑ = = k y Y Y k K Sig ARGMAX X Y P ARGMAX Y 1 ) | ( ˆ (2) where Y1 K denotes K multi-word units from decomposition of X. A multi-word sequence with maximal sum of Sig function values is the search target — pseudo-word sequence. From (2) we can see that Sig function is vital for pseudo-word searching. In this paper Sig function calculates sequence significance which is proposed to characterize pseudo-word as minimal sequence of consecutive words in sense of translation. The detail of sequence significance is described in the following section. 2.1 Sequence Significance Two kinds of definitions of sequence significance are proposed. One is monolingual sequence significance. X and Y are monolingual sentence and monolingual multi-words respectively in this monolingual scenario. The other is bilingual sequence significance. X and Y are sentence pair and multi-word pairs respectively in this bilingual scenario. 2.1.1 Monolingual Sequence Significance Given a sentence w1, …, wn, where wi denotes unary word, monolingual sequence significance is defined as: 1 ,1 , , + − = j i j i j i Freq Freq Sig (3) where Freqi, j (i≤j) represents frequency of word sequence wi, …, wj in the corpus, Sigi, j represents monolingual sequence significance of a word sequence wi, …, wj. We also denote word sequence wi, …, wj as span[i, j], whole sentence as span[1, n]. Each span is also a multi-word expression. Monolingual sequence significance of span[i, j] is proportional to span[i, j]’s frequency, while is inversely proportion to frequency of expanded span (span[i-1, j+1]). Such definition characterizes minimal sequence of consecutive words which we are looking for. Our target is to find pseudo-word sequence which has maximal sum of spans’ significances: k X k Z (1) 149 k (4) ∑ = = K k span span K K Sig ARGMAX pw 1 1 1 where pw denotes pseudo-word, K is equal to or less than sentence’s length. spank is the kth span of K spans span1 K. Equation (4) is the rewrite of equation (2) in monolingual scenario. Searching for pseudo-words pw1 K is the same to finding optimal segmentation of a sentence into K segments span1 K (K is a variable too). Details of searching algorithm are described in section 2.2.1. We firstly search for monolingual pseudowords on source and target side individually. Then we apply word alignment techniques to build pseudo-word alignments. We argue that word alignment techniques will work fine if nonexistent word alignments in such as noncompositional phrasal equivalences have been filtered by pseudo-words. 2.1.2 Bilingual Sequence Significance Bilingual sequence significance is proposed to characterize pseudo-word pairs. Co-occurrence of sequences on both language sides is used to define bilingual sequence significance. Given a bilingual sequence pair: span-pair[is, js, it, jt] (source side span[is, js] and target side span[it, jt]), bilingual sequence significance is defined as below: 1 k ,1 ,1 ,1 , , , , , , + − + − = t t s s t t s s t t s s j i j i j i j i j i j i Freq Freq Sig (5) where Freq denotes the frequency of a span-pair. Bilingual sequence significance is an extension of monolingual sequence significance. Its value is proportional to frequency of span-pair[is, js, it, jt], while is inversely proportional to frequency of expanded span-pair[is-1, js+1, it-1, jt+1]. Pseudo-word pairs of one sentence pair are such pairs that maximize the sum of span-pairs’ bilingual sequence significances: ∑= − − = K k pair span pair span K K Sig ARGMAX pwp 1 1 1 (6) pwp represents pseudo-word pair. Equation (6) is the rewrite of equation (2) in bilingual scenario. Searching for pseudo-word pairs pwp1 K is equal to bilingual segmentation of a sentence pair into optimal span-pair1 K. Details of searching algorithm are presented in section 2.2.2. 2.2 Algorithms of Searching for Pseudowords Pseudo-word searching problem is equal to decomposition of a sentence into pseudo-words. But the number of possible decompositions of the sentence grows exponentially with the sentence length in both monolingual scenario and bilingual scenario. By casting such decomposition problem into parsing framework, we can find pseudo-word sequence in polynomial time. According to the two scenarios, searching for pseudo-words can be performed in a monolingual way and a synchronous way. Details of the two kinds of searching algorithms are described in the following two sections. 2.2.1 Algorithm of Searching for Monolingual Pseudo-words (SMP) Searching for monolingual pseudo-words is based on the computation of monolingual sequence significance. Figure 1 presents the search algorithm. It is performed in a way similar to CKY (Cocke-Kasami-Younger) parser. Initialization: Wi, i = Sigi, i; Wi, j = 0, (i≠j); 1: for d = 2 … n do 2: for all i, j s.t. j-i=d-1 do 3: for k = i … j – 1 do 4: v = Wi, k + Wk+1, j 5: if v > Wi, j then 6: Wi, j = v; 7: u = Sigi, j 8: if u > Wi, j then 9: Wi, j = u; Figure 1. Algorithm of searching for monolingual pseudo-words (SMP). In this algorithm, Wi, j records maximal sum of monolingual sequence significances of sub spans of span[i, j]. During initialization, Wi, i is initialized as Sigi,i (note that this sequence is word wi only). For all spans that have more than one word (i≠j), Wi, j is initialized as zero. In the main algorithm, d represents span’s length, ranging from 2 to n, i represents start position of a span, j represents end position of a span, k represents decomposition position of span[i,j]. For span[i, j], Wi, j is updated if higher sum of monolingual sequence significances is found. The algorithm is performed in a bottom-up way. Small span’s computation is first. After maximal sum of significances is found in small spans, big span’s computation, which uses small spans’ maximal sum, is continued. Maximal sum of significances for whole sentence (W1,n, n is sentence’s length) is guaranteed in this way, and optimal decomposition is obtained correspondingly. 150 The method of fitting the decomposition problem into CKY parsing framework is located at steps 7-9. After steps 3-6, all possible decompositions of span[i, j] are explored and Wi, j of optimal decomposition of span[i, j] is recorded. Then monolingual sequence significance Sigi,j of span[i, j] is computed at step 7, and it is compared to Wi, j at step 8. Update of Wi, j is taken at step 9 if Sigi,j is bigger than Wi, j, which indicates that span[i, j] is non-decomposable. Thus whether span[i, j] should be non-decomposable or not is decided through steps 7-9. 2.2.2 Algorithm of Synchronous Searching for Pseudo-words (SSP) Synchronous searching for pseudo-words utilizes bilingual sequence significance. Figure 2 presents the search algorithm. It is similar to ITG (Wu, 1997), except that it has no production rules and non-terminal nodes of a synchronous grammar. What it cares about is the span-pairs that maximize the sum of bilingual sequence significances. Initialization: if is = js or it = jt then t t s s t t s s t t s s j i j i j i j i Sig W , , , , , , = ; else 0 , , , = j i j i W ; 1: for ds = 2 … ns, dt = 2 … nt do 2: for all is, js, it, jt s.t. js-is=ds-1 and jt-it=dt-1 do 3: for ks = is … js – 1, kt = it … jt – 1 do 4: v = max{ , t t s s t t s s j k j k k i k i W W ,1 , ,1 , , , + + + t t s s t tj i j i , , , tj , , , tj, , , j i j i , , , t s s k i j k j k k i W W , , ,1 , 1 , , + + + } 5: if v > W then t s s 6: W = v; t s s i j i 7: u = t t s s j i j i Sig , , , 8: if u > W then t s s i j i 9: W = u; t t s s Figure 2. Algorithm of Synchronous Searching for Pseudo-words(SSP). In the algorithm, records maximal sum of bilingual sequence significances of sub span-pairs of span-pair[i t t s s j i j i W , , , s, js, it, jt]. For 1-to-m span-pairs, Ws are initialized as bilingual sequence significances of such span-pairs. For other span-pairs, Ws are initialized as zero. In the main algorithm, ds/dt denotes the length of a span on source/target side, ranging from 2 to ns/nt (source/target sentence’s length). is/it is the start position of a span-pair on source/target side, js/jt is the end position of a span-pair on source/target side, ks/kt is the decomposition position of a span-pair[is, js, it, jt] on source/target side. Update steps in Figure 2 are similar to that of Figure 1, except that the update is about spanpairs, not monolingual spans. Reversed and nonreversed alignments inside a span-pair are compared at step 4. For span-pair[is, js, it, jt], is updated at step 6 if higher sum of bilingual sequence significances is found. t t s s j i j i W , , , Fitting the bilingually searching for pseudowords into ITG framework is located at steps 7-9. Steps 3-6 have explored all possible decompositions of span-pair[is, js, it, jt] and have recorded maximal t t s s of these decompositions. Then bilingual sequence significance of span-pair[i j i j i W , , , s, js, it, jt] is computed at step 7. It is compared to t t s s at step 8. Update is taken at step 9 if bilingual sequence significance of span-pair[i j i j i W , , , s, js, it, jt] is bigger than t t s s , which indicates that span-pair[i j i j i W , , , s, js, it, jt] is non-decomposable. Whether the span-pair[is, js, it, jt] should be nondecomposable or not is decided through steps 79. In addition to the initialization step, all spanpairs’ bilingual sequence significances are computed. Maximal sum of bilingual sequence significances for one sentence pair is guaranteed through this bottom-up way, and the optimal decomposition of the sentence pair is obtained correspondingly. z Algorithm of Excluded Synchronous Searching for Pseudo-words (ESSP) The algorithm of SSP in Figure 2 explores all span-pairs, but it neglects NULL alignments, where words and “empty” word are aligned. In fact, SSP requires that all parts of a sentence pair should be aligned. This requirement is too strong because NULL alignments are very common in many language pairs. In SSP, words that should be aligned to “empty” word are programmed to be aligned to real words. Unlike most word alignment methods (Och and Ney, 2003) that add “empty” word to account for NULL alignment entries, we propose a method to naturally exclude such NULL alignments. We call this method as Excluded Synchronous Searching for Pseudo-words (ESSP). The main difference between ESSP and SSP is in steps 3-6 in Figure 3. We illustrate Figure 3’s span-pair configuration in Figure 4. 151 Initialization: if is = js or it = jt then t t s s t t s s j i j i j i j i , , , , , , , , , j i j i W Sig W = ; else 0 = t t s s ; 1: for ds = 2 … ns, dt = 2 … nt do 2: for all is, js, it, jt s.t. js-is=ds-1 and jt-it=dt-1 do 3: for ks1=is+1 … js, ks2=ks1-1 … js-1 kt1=it+1 … jt, kt2=kt1-1 … jt-1 do 4: v = max{W , t t s s t t s s j k j k k i k i W , 1 , , 1 1 , , 1 , 2 2 1 1 + + − − + 1 , , , 1 , 1 , 1 2 2 − + + + t t s s t t k i j k j k W t t j, , , tj , , , Sig t t j i , , , t t s s j i j i , , , 1 , 1− s s k i W } 5: if v > W then s s i j i 6: W = v; t s s i j i 7: u = t t s s j i j i , , , 8: if u > W then s s j i 9: W = u; Figure 3. Algorithm of Excluded Synchronous Searching for Pseudo-words (ESSP). The solid boxes in Figure 4 represent excluded parts of span-pair[is, js, it, jt] in ESSP. Note that, in SSP, there is no excluded part, that is, ks1=ks2 and kt1=kt2. We can see that in Figure 4, each monolingual span is configured into three parts, for example: span[is, ks1-1], span[ks1, ks2] and span[ks2+1, js] on source language side. ks1 and ks2 are two new variables gliding between is and js, span[ks1, ks2] is source side excluded part of span-pair[is, js, it, jt]. Bilingual sequence significance is computed only on pairs of blank boxes, solid boxes are excluded in this computation to represent NULL alignment cases. Figure 4. Illustration of excluded configuration. Note that, in Figure 4, solid box on either language side can be void (i.e., length is zero) if there is no NULL alignment on its side. If all solid boxes are shrunk into void, algorithm of ESSP is the same to SSP. Generally, span length of NULL alignment is not very long, so we can set a length threshold for NULL alignments, eg. ks2-ks1≤EL, where EL denotes Excluded Length threshold. Computational complexity of the ESSP remains the same to SSP’s complexity O(ns 3.nt 3), except multiply a constant EL2. There is one kind of NULL alignments that ESSP can not consider. Since we limit excluded parts in the middle of a span-pair, the algorithm will end without considering boundary parts of a sentence pair as NULL alignments. 3 Experiments and Results In our experiments, pseudo-words are fed into PB-SMT pipeline. The pipeline uses GIZA++ model 4 (Brown et al., 1993; Och and Ney, 2003) for pseudo-word alignment, uses Moses (Koehn et al., 2007) as phrase-based decoder, uses the SRI Language Modeling Toolkit to train language model with modified Kneser-Ney smoothing (Kneser and Ney 1995; Chen and Goodman 1998). Note that MERT (Och, 2003) is still on original words of target language. In our experiments, pseudo-word length is limited to no more than six unary words on both sides of the language pair. We conduct experiments on Chinese-toEnglish machine translation. Two data sets are adopted, one is small corpus of IWSLT-2008 BTEC task of spoken language translation in travel domain (Paul, 2008), the other is large corpus in news domain, which consists Hong Kong News (LDC2004T08), Sinorama Magazine (LDC2005T10), FBIS (LDC2003E14), Xinhua (LDC2002E18), Chinese News Translation (LDC2005T06), Chinese Treebank (LDC2003E07), Multiple Translation Chinese (LDC2004T07). Table 1 lists statistics of the corpus used in these experiments. is ks1 ks2 js it kt1 kt2 jt is ks1 ks2 js it kt1 kt2 jt a) non-reversed b) reversed small large Ch → En Ch → En Sent. 23k 1,239k word 190k 213k 31.7m 35.5m ASL 8.3 9.2 25.6 28.6 Table 1. Statistics of corpora, “Ch” denotes Chinese, “En” denotes English, “Sent.” row is the number of sentence pairs, “word” row is the number of words, “ASL” denotes average sentence length. 152 For small corpus, we use CSTAR03 as development set, use IWSLT08 official test set for test. A 5-gram language model is trained on English side of parallel corpus. For large corpus, we use NIST02 as development set, use NIST03 as test set. Xinhua portion of the English Gigaword3 corpus is used together with English side of large corpus to train a 4-gram language model. Experimental results are evaluated by caseinsensitive BLEU-4 (Papineni et al., 2001). Closest reference sentence length is used for brevity penalty. Additionally, NIST score (Doddington, 2002) and METEOR (Banerjee and Lavie, 2005) are also used to check the consistency of experimental results. Statistical significance in BLEU score differences was tested by paired bootstrap re-sampling (Koehn, 2004). 3.1 Baseline Performance Our baseline system feeds word into PB-SMT pipeline. We use GIZA++ model 4 for word alignment, use Moses for phrase-based decoding. The setting of language model order for each corpus is not changed. Baseline performances on test sets of small corpus and large corpus are reported in table 2. small Large BLEU 0.4029 0.3146 NIST 7.0419 8.8462 METEOR 0.5785 0.5335 Table 2. Baseline performances on test sets of small corpus and large corpus. 3.2 Pseudo-word Unpacking Because pseudo-word is a kind of multi-word expression, it has inborn advantage of higher language model order and longer max phrase length over unary word. To see if such inborn advantage is the main contribution to the performance or not, we unpack pseudo-word into words after GIZA++ aligning. Aligned pseudowords are unpacked into m×n word alignments. PB-SMT pipeline is executed thereafter. The advantage of longer max phrase length is removed during phrase extraction, and the advantage of higher order of language model is also removed during decoding since we use language model trained on unary words. Performances of pseudoword unpacking are reported in section 3.3.1 and 3.4.1. Ma and Way (2009) used the unpacking after phrase extraction, then re-estimated phrase translation probability and lexical reordering model. The advantage of longer max phrase length is still used in their method. 3.3 Pseudo-word Performances on Small Corpus Table 3 presents performances of SMP, SSP, ESSP on small data set. pwchpwen denotes that pseudo-words are on both language side of training data, and they are input strings during development and testing, and translations are also pseudo-words, which will be converted to words as final output. wchpwen/pwchwen denotes that pseudo-words are adopted only on English/Chinese side of the data set. We can see from table 3 that, ESSP attains the best performance, while SSP attains the worst performance. This shows that excluding NULL alignments in synchronous searching for pseudowords is effective. SSP puts overly strong alignment constraints on parallel corpus, which impacts performance dramatically. ESSP is superior to SMP indicating that bilingually motivated searching for pseudo-words is more effective. Both SMP and ESSP outperform baseline consistently in BLEU, NIST and METEOR. There is a common phenomenon among SMP, SSP and ESSP. wchpwen always performs better than the other two cases. It seems that Chinese word prefers to have English pseudo-word equivalence which has more than or equal to one word. pwchpwen in ESSP performs similar to the baseline, which reflects that our direct pseudoword pairs do not work very well with GIZA++ alignments. Such disagreement is weakened by using pseudo-words on only one language side (wchpwen or pwchwen), while the advantage of pseudo-words is still leveraged in the alignments. Best ESSP (wchpwen) is significantly better than baseline (p<0.01) in BLEU score, best SMP (wchpwen) is significantly better than baseline (p<0.05) in BLEU score. This indicates that pseudo-words, through either monolingual searching or synchronous searching, are more effective than words as to being basic translational units. Figure 5 illustrates examples of pseudo-words of one Chinese-to-English sentence pair. Gold standard word alignments are shown at the bottom of figure 5. We can see that “front desk” is recognized as one pseudo-word in ESSP. Because SMP performs monolingually, it can not consider “前台” and “front desk” simultaneously. SMP only detects frequent monolingual multiwords as pseudo-words. SSP has a strong constraint that all parts of a sentence pair should be aligned, so source sentence and target sentence have same length after merging words into 153 Table 3. Performance of using pseudo-words on small data. pseudo-words. We can see that too many pseudowords are detected by SSP. Figure 5. Outputs of the three algorithms ESSP, SMP and SSP on one sentence pair and gold standard word alignments. Words in one pseudo-word are concatenated by “_”. 3.3.1 Pseudo-word Unpacking Performances on Small Corpus We test pseudo-word unpacking in ESSP. Table 4 presents its performances on small corpus. unpackingESSP pwchpwen wchpwen pwchwen baseline BLEU 0.4097 0.4182 0.4031 0.4029 NIST 7.5547 7.2893 7.2670 7.0419 METEOR 0.5951 0.5874 0.5846 0.5785 Table 4. Performances of pseudo-word unpacking on small corpus. We can see that pseudo-word unpacking significantly outperforms baseline. wchpwen is significantly better than baseline (p<0.04) in BLEU score. Unpacked pseudo-word performs comparatively with pseudo-word without unpacking. There is no statistical difference between them. It shows that the improvement derives from pseudo-word itself as basic translational unit, does not rely very much on higher language model order or longer max phrase length setting. 3.4 Pseudo-word Performances on Large Corpus Table 5 lists the performance of using pseudowords on large corpus. We apply SMP on this task. ESSP is not applied because of its high computational complexity. Table 5 shows that all three configurations (pwchpwen, wchpwen, pwchwen) of SMP outperform the baseline. If we go back to the definition of sequence significance, we can see that it is a data-driven definition that utilizes corpus frequencies. Corpus scale has an influence on computation of sequence significance in long sentences which appear frequently in news domain. SMP benefits from large corpus, and wchpwen is significantly better than baseline (p<0.01). Similar to performances on small corpus, wchpwen always performs better than the other two cases, which indicates that Chinese word prefers to have English pseudo-word equivalence which has more than or equal to one word. SMP pwchpwen wchpwen pwchwen baseline BLEU 0.3185 0.3230 0.3166 0.3146 NIST 8.9216 9.0447 8.9210 8.8462 METEOR 0.5402 0.5489 0.5435 0.5335 Table 5. Performance of using pseudo-words on large corpus. 3.4.1 Pseudo-word Unpacking Performances on Large Corpus Table 6 presents pseudo-word unpacking performances on large corpus. All three configurations improve performance over baseline after pseudo-word unpacking. pwchpwen attains the best BLEU among the three configurations, and is significantly better than baseline (p<0.03). wchpwen is also significantly better than baseline (p<0.04). By comparing table 6 with table 5, we can see that unpacked pseudo-word performs comparatively with pseudo-word without unpacking. There is no statistical difference beSMP SSP ESSP pwchpwen wchpwen pwchwen pwchpwen wchpwen pwchwen pwchpwen wchpwen pwchwen baseline BLEU 0.3996 0.4155 0.4024 0.3184 0.3661 0.3552 0.3998 0.4229 0.4147 0.4029 NIST 7.4711 7.6452 7.6186 6.4099 6.9284 6.8012 7.1665 7.4373 7.4235 7.0419 METEOR 0.5900 0.6008 0.6000 0.5255 0.5569 0.5454 0.5739 0.5963 0.5891 0.5785 前台 的 那个 人 真 粗鲁 。 The guy at the front desk is pretty rude . 前台 的 那个 人 真 粗鲁 。 The guy_at the front_desk is pretty_rude . 前台 的 那个 人 真 粗鲁 。 The guy at the front_desk is pretty rude . ESSP 前台 的 那个 人 真 粗鲁 。 The guy at the front desk is pretty rude . Gold standard word alignments SMP SSP 154 tween them. It shows that the improvement derives from pseudo-word itself as basic translational unit, does not rely very much on higher language model order or longer max phrase length setting. In fact, slight improvement in pwchpwen and pwchwen is seen after pseudo-word unpacking, which indicates that higher language model order and longer max phrase length impact the performance in these two configurations. UnpackingSMP pwchpwen wchpwen pwchwen Baseline BLEU 0.3219 0.3192 0.3187 0.3146 NIST 8.9458 8.9325 8.9801 8.8462 METEOR 0.5429 0.5424 0.5411 0.5335 Table 6. Performance of pseudo-word unpacking on large corpus. 3.5 Comparison to English Chunking English chunking is experimented to compare with pseudo-word. We use FlexCRFs (XuanHieu Phan et al., 2005) to get English chunks. Since there is no standard Chinese chunking data and code, only English chunking is executed. The experimental results show that English chunking performs far below baseline, usually 8 absolute BLEU points below. It shows that simple chunks are not suitable for being basic translational units. 4 Conclusion We have presented pseudo-word as a novel machine translational unit for phrase-based machine translation. It is proposed to replace too finegrained word as basic translational unit. Pseudoword is a kind of basic multi-word expression that characterizes minimal sequence of consecutive words in sense of translation. By casting pseudo-word searching problem into a parsing framework, we search for pseudo-words in polynomial time. Experimental results of Chinese-toEnglish translation task show that, in phrasebased machine translation model, pseudo-word performs significantly better than word in both spoken language translation domain and news domain. Removing the power of higher order language model and longer max phrase length, which are inherent in pseudo-words, shows that pseudo-words still improve translational performance significantly over unary words. References S. Banerjee, and A. Lavie. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization (ACL’05). 65–72. P. Blunsom, T. Cohn, C. Dyer, M. Osborne. 2009. A Gibbs Sampler for Phrasal Synchronous Grammar Induction. In Proceedings of ACLIJCNLP, Singapore. P. Blunsom, T. Cohn, M. Osborne. 2008. Bayesian synchronous grammar induction. In Proceedings of NIPS 21, Vancouver, Canada. P. Brown, S. Della Pietra, V. Della Pietra, and R. Mercer. 1993. The mathematics of machine translation: Parameter estimation. Computational Linguistics, 19:263–312. P.-C. Chang, M. Galley, and C. D. Manning. 2008. Optimizing Chinese word segmentation for machine translation performance. In Proceedings of the 3rd Workshop on Statistical Machine Translation (SMT’08). 224–232. Chen, Stanley F. and Joshua Goodman. 1998. An empirical study of smoothing techniques for language modeling. Technical Report TR-10-98, Harvard University Center for Research in Computing Technology. C. Cherry, D. Lin. 2007. Inversion transduction grammar for joint phrasal translation modeling. In Proc. of the HLTNAACL Workshop on Syntax and Structure in Statistical Translation (SSST 2007), Rochester, USA. D. Chiang. 2007. Hierarchical phrase-based translation.Computational Linguistics, 33(2):201– 228. Y. Deng and W. Byrne. 2005. HMM word and phrase alignment for statistical machine translation. In Proc. of HLT-EMNLP, pages 169–176. G. Doddington. 2002. Automatic evaluation of machine translation quality using n-gram cooccurrence statistics. In Proceedings of the 2nd International Conference on Human Language Technology (HLT’02). 138–145. Kneser, Reinhard and Hermann Ney. 1995. Improved backing-off for M-gram language modeling. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, pages 181–184, Detroit, MI. P. Koehn, H. Hoang, A. Birch, C. Callison-Burch, M. Federico, N. Bertoldi, B. Cowan,W. Shen, C. Moran, R. Zens, C. Dyer, O. Bojar, A. Constantin, E. Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proc. of the 155 45th Annual Meeting of the ACL (ACL-2007), Prague. P. Koehn, F. J. Och, D. Marcu. 2003. Statistical phrasebased translation. In Proc. of the 3rd International conference on Human Language Technology Research and 4th Annual Meeting of the NAACL (HLT-NAACL 2003), 81–88, Edmonton, Canada. P. Koehn. 2004. Statistical Significance Tests for Machine Translation Evaluation. In Proceedings of EMNLP. P. Lambert and R. Banchs. 2005. Data Inferred Multi-word Expressions for Statistical Machine Translation. In Proceedings of MT Summit X. Y. Ma, N. Stroppa, and A. Way. 2007. Bootstrapping word alignment via word packing. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics (ACL’07). 304–311. Y. Ma, and A. Way. 2009. Bilingually Motivated Word Segmentation for Statistical Machine Translation. In ACM Transactions on Asian Language Information Processing, 8(2). D. Marcu,W.Wong. 2002. A phrase-based, joint probability model for statistical machine translation. In Proc. of the 2002 Conference on Empirical Methods in Natural Language Processing (EMNLP-2002), 133–139, Philadelphia. Association for Computational Linguistics. F. J. Och. 2003. Minimum error rate training in statistical machine translation. In Proc. of ACL, pages 160–167. F. J. Och and H. Ney. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1):19–51. Xuan-Hieu Phan, Le-Minh Nguyen, and Cam-Tu Nguyen. 2005. FlexCRFs: Flexible Conditional Random Field Toolkit, http://flexcrfs.sourceforge. net K. Papineni, S. Roukos, T. Ward, W. Zhu. 2001. Bleu: a method for automatic evaluation of machine translation, 2001. M. Paul, 2008. Overview of the IWSLT 2008 evaluation campaign. In Proc. of Internationa Workshop on Spoken Language Translation, 20-21 October 2008. A. Stolcke. (2002). SRILM - an extensible language modeling toolkit. In Proceedings of ICSLP, Denver, Colorado. D. Wu. 1997. Stochastic inversion transduction grammars and bilingual parsing of parallel corpora. Computational Linguistics, 23(3):377– 403. J. Xu, Zens., and H. Ney. 2004. Do we need Chinese word segmentation for statistical machine translation? In Proceedings of the ACL Workshop on Chinese Language Processing SIGHAN’04). 122–128. J. Xu, J. Gao, K. Toutanova, and H. Ney. 2008. Bayesian semi-supervised chinese word segmentation for statistical machine translation. In Proceedings of the 22nd International Conference on Computational Linguistics (COLING’08). 1017–1024. H. Zhang, C. Quirk, R. C. Moore, D. Gildea. 2008. Bayesian learning of non-compositional phrases with synchronous parsing. In Proc. of the 46th Annual Conference of the Association for Computational Linguistics: Human Language Technologies (ACL-08:HLT), 97–105, Columbus, Ohio. R. Zhang, K. Yasuda, and E. Sumita. 2008. Improved statistical machine translation by multiple Chinese word segmentation. In Proceedings of the 3rd Workshop on Statistical Machine Translation (SMT’08). 216–223. 156
|
2010
|
16
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1583–1592, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Beyond NomBank: A Study of Implicit Arguments for Nominal Predicates Matthew Gerber and Joyce Y. Chai Department of Computer Science Michigan State University East Lansing, Michigan, USA {gerberm2,jchai}@cse.msu.edu Abstract Despite its substantial coverage, NomBank does not account for all withinsentence arguments and ignores extrasentential arguments altogether. These arguments, which we call implicit, are important to semantic processing, and their recovery could potentially benefit many NLP applications. We present a study of implicit arguments for a select group of frequent nominal predicates. We show that implicit arguments are pervasive for these predicates, adding 65% to the coverage of NomBank. We demonstrate the feasibility of recovering implicit arguments with a supervised classification model. Our results and analyses provide a baseline for future work on this emerging task. 1 Introduction Verbal and nominal semantic role labeling (SRL) have been studied independently of each other (Carreras and M`arquez, 2005; Gerber et al., 2009) as well as jointly (Surdeanu et al., 2008; Hajiˇc et al., 2009). These studies have demonstrated the maturity of SRL within an evaluation setting that restricts the argument search space to the sentence containing the predicate of interest. However, as shown by the following example from the Penn TreeBank (Marcus et al., 1993), this restriction excludes extra-sentential arguments: (1) [arg0 The two companies] [pred produce] [arg1 market pulp, containerboard and white paper]. The goods could be manufactured closer to customers, saving [pred shipping] costs. The first sentence in Example 1 includes the PropBank (Kingsbury et al., 2002) analysis of the verbal predicate produce, where arg0 is the agentive producer and arg1 is the produced entity. The second sentence contains an instance of the nominal predicate shipping that is not associated with arguments in NomBank (Meyers, 2007). From the sentences in Example 1, the reader can infer that The two companies refers to the agents (arg0) of the shipping predicate. The reader can also infer that market pulp, containerboard and white paper refers to the shipped entities (arg1 of shipping).1 These extra-sentential arguments have not been annotated for the shipping predicate and cannot be identified by a system that restricts the argument search space to the sentence containing the predicate. NomBank also ignores many within-sentence arguments. This is shown in the second sentence of Example 1, where The goods can be interpreted as the arg1 of shipping. These examples demonstrate the presence of arguments that are not included in NomBank and cannot easily be identified by systems trained on the resource. We refer to these arguments as implicit. This paper presents our study of implicit arguments for nominal predicates. We began our study by annotating implicit arguments for a select group of predicates. For these predicates, we found that implicit arguments add 65% to the existing role coverage of NomBank.2 This increase has implications for tasks (e.g., question answering, information extraction, and summarization) that benefit from semantic analysis. Using our annotations, we constructed a feature-based model for automatic implicit argument identification that unifies standard verbal and nominal SRL. Our results indicate a 59% relative (15-point absolute) gain in F1 over an informed baseline. Our analyses highlight strengths and weaknesses of the approach, providing insights for future work on this emerging task. 1In PropBank and NomBank, the interpretation of each role (e.g., arg0) is specific to a predicate sense. 2Role coverage indicates the percentage of roles filled. 1583 In the following section, we review related research, which is historically sparse but recently gaining traction. We present our annotation effort in Section 3, and follow with our implicit argument identification model in Section 4. In Section 5, we describe the evaluation setting and present our experimental results. We analyze these results in Section 6 and conclude in Section 7. 2 Related work Palmer et al. (1986) made one of the earliest attempts to automatically recover extra-sentential arguments. Their approach used a fine-grained domain model to assess the compatibility of candidate arguments and the slots needing to be filled. A phenomenon similar to the implicit argument has been studied in the context of Japanese anaphora resolution, where a missing case-marked constituent is viewed as a zero-anaphoric expression whose antecedent is treated as the implicit argument of the predicate of interest. This behavior has been annotated manually by Iida et al. (2007), and researchers have applied standard SRL techniques to this corpus, resulting in systems that are able to identify missing case-marked expressions in the surrounding discourse (Imamura et al., 2009). Sasano et al. (2004) conducted similar work with Japanese indirect anaphora. The authors used automatically derived nominal case frames to identify antecedents. However, as noted by Iida et al., grammatical cases do not stand in a one-to-one relationship with semantic roles in Japanese (the same is true for English). Fillmore and Baker (2001) provided a detailed case study of implicit arguments (termed null instantiations in that work), but did not provide concrete methods to account for them automatically. Previously, we demonstrated the importance of filtering out nominal predicates that take no local arguments (Gerber et al., 2009); however, this work did not address the identification of implicit arguments. Burchardt et al. (2005) suggested approaches to implicit argument identification based on observed coreference patterns; however, the authors did not implement and evaluate such methods. We draw insights from all three of these studies. We show that the identification of implicit arguments for nominal predicates leads to fuller semantic interpretations when compared to traditional SRL methods. Furthermore, motivated by Burchardt et al., our model uses a quantitative analysis of naturally occurring coreference patterns to aid implicit argument identification. Most recently, Ruppenhofer et al. (2009) conducted SemEval Task 10, “Linking Events and Their Participants in Discourse”, which evaluated implicit argument identification systems over a common test set. The task organizers annotated implicit arguments across entire passages, resulting in data that cover many distinct predicates, each associated with a small number of annotated instances. In contrast, our study focused on a select group of nominal predicates, each associated with a large number of annotated instances. 3 Data annotation and analysis 3.1 Data annotation Implicit arguments have not been annotated within the Penn TreeBank, which is the textual and syntactic basis for NomBank. Thus, to facilitate our study, we annotated implicit arguments for instances of nominal predicates within the standard training, development, and testing sections of the TreeBank. We limited our attention to nominal predicates with unambiguous role sets (i.e., senses) that are derived from verbal role sets. We then ranked this set of predicates using two pieces of information: (1) the average difference between the number of roles expressed in nominal form (in NomBank) versus verbal form (in PropBank) and (2) the frequency of the nominal form in the corpus. We assumed that the former gives an indication as to how many implicit roles an instance of the nominal predicate might have. The product of (1) and (2) thus indicates the potential prevalence of implicit arguments for a predicate. To focus our study, we ranked the predicates in NomBank according to this product and selected the top ten, shown in Table 1. We annotated implicit arguments document-bydocument, selecting all singular and plural nouns derived from the predicates in Table 1. For each missing argument position of each predicate instance, we inspected the local discourse for a suitable implicit argument. We limited our attention to the current sentence as well as all preceding sentences in the document, annotating all mentions of an implicit argument within this window. In the remainder of this paper, we will use iargn to refer to an implicit argument position n. We will use argn to refer to an argument provided by PropBank or NomBank. We will use p to mark 1584 Pre-annotation Post-annotation Role average Predicate # Role coverage (%) Noun Verb Role coverage (%) Noun role average price 217 42.4 1.7 1.7 55.3 2.2 sale 185 24.3 1.2 2.0 42.0 2.1 investor 160 35.0 1.1 2.0 54.6 1.6 fund 109 8.7 0.4 2.0 21.6 0.9 loss 104 33.2 1.3 2.0 46.9 1.9 plan 102 30.9 1.2 1.8 49.3 2.0 investment 102 15.7 0.5 2.0 33.3 1.0 cost 101 26.2 1.1 2.3 47.5 1.9 bid 88 26.9 0.8 2.2 72.0 2.2 loan 85 22.4 1.1 2.5 41.2 2.1 Overall 1,253 28.0 1.1 2.0 46.2 1.8 Table 1: Predicates targeted for annotation. The second column gives the number of predicate instances annotated. Pre-annotation numbers only include NomBank annotations, whereas Post-annotation numbers include NomBank and implicit argument annotations. Role coverage indicates the percentage of roles filled. Role average indicates how many roles, on average, are filled for an instance of a predicate’s noun form or verb form within the TreeBank. Verbal role averages were computed using PropBank. predicate instances. Below, we give an example annotation for an instance of the investment predicate: (2) [iarg0 Participants] will be able to transfer [iarg1 money] to [iarg2 other investment funds]. The [p investment] choices are limited to [iarg2 a stock fund and a money-market fund]. NomBank does not associate this instance of investment with any arguments; however, we were able to identify the investor (iarg0), the thing invested (iarg1), and two mentions of the thing invested in (iarg2). Our data set was also independently annotated by an undergraduate linguistics student. For each missing argument position, the student was asked to identify the closest acceptable implicit argument within the current and preceding sentences. The argument position was left unfilled if no acceptable constituent could be found. For a missing argument position, the student’s annotation agreed with our own if both identified the same constituent or both left the position unfilled. Analysis indicated an agreement of 67% using Cohen’s kappa coefficient (Cohen, 1960). 3.2 Annotation analysis Role coverage for a predicate instance is equal to the number of filled roles divided by the number of roles in the predicate’s lexicon entry. Role coverage for the marked predicate in Example 2 is 0/3 for NomBank-only arguments and 3/3 when the annotated implicit arguments are also considered. Returning to Table 1, the third column gives role coverage percentages for NomBank-only arguments. The sixth column gives role coverage percentages when both NomBank arguments and the annotated implicit arguments are considered. Overall, the addition of implicit arguments created a 65% relative (18-point absolute) gain in role coverage across the 1,253 predicate instances that we annotated. The predicates in Table 1 are typically associated with fewer arguments on average than their corresponding verbal predicates. When considering NomBank-only arguments, this difference (compare columns four and five) varies from zero (for price) to a factor of five (for fund). When implicit arguments are included in the comparison, these differences are reduced and many nominal predicates express approximately the same number of arguments on average as their verbal counterparts (compare the fifth and seventh columns). In addition to role coverage and average count, we examined the location of implicit arguments. Figure 1 shows that approximately 56% of the implicit arguments in our data can be resolved within the sentence containing the predicate. The remaining implicit arguments require up to forty-six sen1585 0.4 0.5 0.6 0.7 0.8 0.9 1 0 2 4 6 8 10 12 18 28 46 Sentences prior Implicit arguments resolved Figure 1: Location of implicit arguments. For missing argument positions with an implicit filler, the y-axis indicates the likelihood of the filler being found at least once in the previous x sentences. tences for resolution; however, a vast majority of these can be resolved within the previous few sentences. Section 6 discusses implications of this skewed distribution. 4 Implicit argument identification 4.1 Model formulation In our study, we assumed that each sentence in a document had been analyzed for PropBank and NomBank predicate-argument structure. NomBank includes a lexicon listing the possible argument positions for a predicate, allowing us to identify missing argument positions with a simple lookup. Given a nominal predicate instance p with a missing argument position iargn, the task is to search the surrounding discourse for a constituent c that fills iargn. Our model conducts this search over all constituents annotated by either PropBank or NomBank with non-adjunct labels. A candidate constituent c will often form a coreference chain with other constituents in the discourse. Consider the following abridged sentences, which are adjacent in their Penn TreeBank document: (3) [Mexico] desperately needs investment. (4) Conservative Japanese investors are put off by [Mexico’s] investment regulations. (5) Japan is the fourth largest investor in [c Mexico], with 5% of the total [p investments]. NomBank does not associate the labeled instance of investment with any arguments, but it is clear from the surrounding discourse that constituent c (referring to Mexico) is the thing being invested in (the iarg2). When determining whether c is the iarg2 of investment, one can draw evidence from other mentions in c’s coreference chain. Example 3 states that Mexico needs investment. Example 4 states that Mexico regulates investment. These propositions, which can be derived via traditional SRL analyses, should increase our confidence that c is the iarg2 of investment in Example 5. Thus, the unit of classification for a candidate constituent c is the three-tuple ⟨p, iargn, c′⟩, where c′ is a coreference chain comprising c and its coreferent constituents.3 We defined a binary classification function Pr(+| ⟨p, iargn, c′⟩) that predicts the probability that the entity referred to by c fills the missing argument position iargn of predicate instance p. In the remainder of this paper, we will refer to c as the primary filler, differentiating it from other mentions in the coreference chain c′. In the following section, we present the feature set used to represent each three-tuple within the classification function. 4.2 Model features Starting with a wide range of features, we performed floating forward feature selection (Pudil et al., 1994) over held-out development data comprising implicit argument annotations from section 24 of the Penn TreeBank. As part of the feature selection process, we conducted a grid search for the best per-class cost within LibLinear’s logistic regression solver (Fan et al., 2008). This was done to reduce the negative effects of data imbalance, which is severe even when selecting candidates from the current and previous few sentences. Table 2 shows the selected features, which are quite different from those used in our previous work to identify traditional semantic arguments (Gerber et al., 2009).4 Below, we give further explanations for some of the features. Feature 1 models the semantic role relationship between each mention in c′ and the missing argument position iargn. To reduce data sparsity, this feature generalizes predicates and argument positions to their VerbNet (Kipper, 2005) classes and 3We used OpenNLP for coreference identification: http://opennlp.sourceforge.net 4We have omitted many of the lowest-ranked features. Descriptions of these features can be obtained by contacting the authors. 1586 # Feature value description 1* For every f, the VerbNet class/role of pf/argf concatenated with the class/role of p/iargn. 2* Average pointwise mutual information between ⟨p, iargn⟩and any ⟨pf, argf⟩. 3 Percentage of all f that are definite noun phrases. 4 Minimum absolute sentence distance from any f to p. 5* Minimum pointwise mutual information between ⟨p, iargn⟩and any ⟨pf, argf⟩. 6 Frequency of the nominal form of p within the document that contains it. 7 Nominal form of p concatenated with iargn. 8 Nominal form of p concatenated with the sorted integer argument indexes from all argn of p. 9 Number of mentions in c′. 10* Head word of p’s right sibling node. 11 For every f, the synset (Fellbaum, 1998) for the head of f concatenated with p and iargn. 12 Part of speech of the head of p’s parent node. 13 Average absolute sentence distance from any f to p. 14* Discourse relation whose two discourse units cover c (the primary filler) and p. 15 Number of left siblings of p. 16 Whether p is the head of its parent node. 17 Number of right siblings of p. Table 2: Features for determining whether c fills iargn of predicate p. For each mention f (denoting a filler) in the coreference chain c′, we define pf and argf to be the predicate and argument position of f. Features are sorted in descending order of feature selection gain. Unless otherwise noted, all predicates were normalized to their verbal form and all argument positions (e.g., argn and iargn) were interpreted as labels instead of word content. Features marked with an asterisk are explained in Section 4.2. semantic roles using SemLink.5 For explanation purposes, consider again Example 1, where we are trying to fill the iarg0 of shipping. Let c′ contain a single mention, The two companies, which is the arg0 of produce. As described in Table 2, feature 1 is instantiated with a value of create.agentsend.agent, where create and send are the VerbNet classes that contain produce and ship, respectively. In the conversion to LibLinear’s instance representation, this instantiation is converted into a single binary feature create.agent-send.agent whose value is one. Features 1 and 11 are instantiated once for each mention in c′, allowing the model to consider information from multiple mentions of the same entity. Features 2 and 5 are inspired by the work of Chambers and Jurafsky (2008), who investigated unsupervised learning of narrative event sequences using pointwise mutual information (PMI) between syntactic positions. We used a similar PMI score, but defined it with respect to semantic arguments instead of syntactic dependencies. Thus, the values for features 2 and 5 are computed as follows (the notation is explained in 5http://verbs.colorado.edu/semlink the caption for Table 2): pmi(⟨p, iargn⟩, ⟨pf, argf⟩) = log Pcoref(⟨p, iargn⟩, ⟨pf, argf⟩) Pcoref(⟨p, iargn⟩, ∗)Pcoref(⟨pf, argf⟩, ∗) (6) To compute Equation 6, we first labeled a subset of the Gigaword corpus (Graff, 2003) using the verbal SRL system of Punyakanok et al. (2008) and the nominal SRL system of Gerber et al. (2009). We then identified coreferent pairs of arguments using OpenNLP. Suppose the resulting data has N coreferential pairs of argument positions. Also suppose that M of these pairs comprise ⟨p, argn⟩ and ⟨pf, argf⟩. The numerator in Equation 6 is defined as M N . Each term in the denominator is obtained similarly, except that M is computed as the total number of coreference pairs comprising an argument position (e.g., ⟨p, argn⟩) and any other argument position. Like Chambers and Jurafsky, we also used the discounting method suggested by Pantel and Ravichandran (2004) for lowfrequency observations. The PMI score is somewhat noisy due to imperfect output, but it provides information that is useful for classification. 1587 Feature 10 does not depend on c′ and is specific to each predicate. Consider the following example: (7) Statistics Canada reported that its [arg1 industrial-product] [p price] index dropped 2% in September. The “[p price] index” collocation is rarely associated with an arg0 in NomBank or with an iarg0 in our annotations (both argument positions denote the seller). Feature 10 accounts for this type of behavior by encoding the syntactic head of p’s right sibling. The value of feature 10 for Example 7 is price:index. Contrast this with the following: (8) [iarg0 The company] is trying to prevent further [p price] drops. The value of feature 10 for Example 8 is price:drop. This feature captures an important distinction between the two uses of price: the former rarely takes an iarg0, whereas the latter often does. Features 12 and 15-17 account for predicatespecific behaviors in a similar manner. Feature 14 identifies the discourse relation (if any) that holds between the candidate constituent c and the filled predicate p. Consider the following example: (9) [iarg0 SFE Technologies] reported a net loss of $889,000 on sales of $23.4 million. (10) That compared with an operating [p loss] of [arg1 $1.9 million] on sales of $27.4 million in the year-earlier period. In this case, a comparison discourse relation (signaled by the underlined text) holds between the first and sentence sentence. The coherence provided by this relation encourages an inference that identifies the marked iarg0 (the loser). Throughout our study, we used gold-standard discourse relations provided by the Penn Discourse TreeBank (Prasad et al., 2008). 5 Evaluation We trained the feature-based logistic regression model over 816 annotated predicate instances associated with 650 implicitly filled argument positions (not all predicate instances had implicit arguments). During training, a candidate three-tuple ⟨p, iargn, c′⟩was given a positive label if the candidate implicit argument c (the primary filler) was annotated as filling the missing argument position. To factor out errors from standard SRL analyses, the model used gold-standard argument labels provided by PropBank and NomBank. As shown in Figure 1 (Section 3.2), implicit arguments tend to be located in close proximity to the predicate. We found that using all candidate constituents c within the current and previous two sentences worked best on our development data. We compared our supervised model with the simple baseline heuristic defined below:6 Fill iargn for predicate instance p with the nearest constituent in the twosentence candidate window that fills argn for a different instance of p, where all nominal predicates are normalized to their verbal forms. The normalization allows an existing arg0 for the verb invested to fill an iarg0 for the noun investment. We also evaluated an oracle model that made gold-standard predictions for candidates within the two-sentence prediction window. We evaluated these models using the methodology proposed by Ruppenhofer et al. (2009). For each missing argument position of a predicate instance, the models were required to either (1) identify a single constituent that fills the missing argument position or (2) make no prediction and leave the missing argument position unfilled. We scored predictions using the Dice coefficient, which is defined as follows: 2 ∗|Predicted T True| |Predicted| + |True| (11) Predicted is the set of tokens subsumed by the constituent predicted by the model as filling a missing argument position. True is the set of tokens from a single annotated constituent that fills the missing argument position. The model’s prediction receives a score equal to the maximum Dice overlap across any one of the annotated fillers. Precision is equal to the summed prediction scores divided by the number of argument positions filled by the model. Recall is equal to the summed prediction scores divided by the number of argument positions filled in our annotated data. Predictions not covering the head of a true filler were assigned a score of zero. 6This heuristic outperformed a more complicated heuristic that relied on the PMI score described in section 4.2. 1588 Baseline Discriminative Oracle # Imp. # P R F1 P R F1 p R F1 sale 64 60 50.0 28.3 36.2 47.2 41.7 44.2 0.118 80.0 88.9 price 121 53 24.0 11.3 15.4 36.0 32.6 34.2 0.008 88.7 94.0 investor 78 35 33.3 5.7 9.8 36.8 40.0 38.4 < 0.001 91.4 95.5 bid 19 26 100.0 19.2 32.3 23.8 19.2 21.3 0.280 57.7 73.2 plan 25 20 83.3 25.0 38.5 78.6 55.0 64.7 0.060 82.7 89.4 cost 25 17 66.7 23.5 34.8 61.1 64.7 62.9 0.024 94.1 97.0 loss 30 12 71.4 41.7 52.6 83.3 83.3 83.3 0.020 100.0 100.0 loan 11 9 50.0 11.1 18.2 42.9 33.3 37.5 0.277 88.9 94.1 investment 21 8 0.0 0.0 0.0 40.0 25.0 30.8 0.182 87.5 93.3 fund 43 6 0.0 0.0 0.0 14.3 16.7 15.4 0.576 50.0 66.7 Overall 437 246 48.4 18.3 26.5 44.5 40.4 42.3 < 0.001 83.1 90.7 Table 3: Evaluation results. The second column gives the number of predicate instances evaluated. The third column gives the number of ground-truth implicitly filled argument positions for the predicate instances (not all instances had implicit arguments). P, R, and F1 indicate precision, recall, and Fmeasure (β = 1), respectively. p-values denote the bootstrapped significance of the difference in F1 between the baseline and discriminative models. Oracle precision (not shown) is 100% for all predicates. Our evaluation data comprised 437 predicate instances associated with 246 implicitly filled argument positions. Table 3 presents the results. Predicates with the highest number of implicit arguments - sale and price - showed F1 increases of 8 points and 18.8 points, respectively. Overall, the discriminative model increased F1 performance 15.8 points (59.6%) over the baseline. We measured human performance on this task by running our undergraduate assistant’s annotations against the evaluation data. Our assistant achieved an overall F1 score of 58.4% using the same candidate window as the baseline and discriminative models. The difference in F1 between the discriminative and human results had an exact p-value of less than 0.001. All significance testing was performed using a two-tailed bootstrap method similar to the one described by Efron and Tibshirani (1993). 6 Discussion 6.1 Feature ablation We conducted an ablation study to measure the contribution of specific feature sets. Table 4 presents the ablation configurations and results. For each configuration, we retrained and retested the discriminative model using the features described. As shown, we observed significant losses when excluding features that relate the semantic roles of mentions in c′ to the semantic role Percent change (p-value) Configuration P R F1 Remove 1,2,5 -35.3 (< 0.01) -36.1 (< 0.01) -35.7 (< 0.01) Use 1,2,5 only -26.3 (< 0.01) -11.9 (0.05) -19.2 (< 0.01) Remove 14 0.2 (0.95) 1.0 (0.66) 0.7 (0.73) Table 4: Feature ablation results. The first column lists the feature configurations. All changes are percentages relative to the full-featured discriminative model. p-values for the changes are indicated in parentheses. of the missing argument position (first configuration). The second configuration tested the effect of using only the SRL-based features. This also resulted in significant performance losses, suggesting that the other features contribute useful information. Lastly, we tested the effect of removing discourse relations (feature 14), which are likely to be difficult to extract reliably in a practical setting. As shown, this feature did not have a statistically significant effect on performance and could be excluded in future applications of the model. 6.2 Unclassified true implicit arguments Of all the errors made by the system, approximately 19% were caused by the system’s failure to 1589 generate a candidate constituent c that was a correct implicit argument. Without such a candidate, the system stood no chance of identifying a correct implicit argument. Two factors contributed to this type of error, the first being our assumption that implicit arguments are also core (i.e., argn) arguments to traditional SRL structures. Approximately 8% of the overall error was due to a failure of this assumption. In many cases, the true implicit argument filled a non-core (i.e., adjunct) role within PropBank or NomBank. More frequently, however, true implicit arguments were missed because the candidate window was too narrow. This accounts for 12% of the overall error. Oracle recall (second-to-last column in Table 3) indicates the nominals that suffered most from windowing errors. For example, the sale predicate was associated with the highest number of true implicit arguments, but only 80% of those could be resolved within the two-sentence candidate window. Empirically, we found that extending the candidate window uniformly for all predicates did not increase performance on the development data. The oracle results suggest that predicate-specific window settings might offer some advantage. 6.3 The investment and fund predicates In Section 4.2, we discussed the price predicate, which frequently occurs in the “[p price] index” collocation. We observed that this collocation is rarely associated with either an overt arg0 or an implicit iarg0. Similar observations can be made for the investment and fund predicates. Although these two predicates are frequent, they are rarely associated with implicit arguments: investment takes only eight implicit arguments across its 21 instances, and fund takes only six implicit arguments across its 43 instances. This behavior is due in large part to collocations such as “[p investment] banker”, “stock [p fund]”, and “mutual [p fund]”, which use predicate senses that are not eventive. Such collocations also violate our assumption that differences between the PropBank and NomBank argument structure for a predicate are indicative of implicit arguments (see Section 3.1 for this assumption). Despite their lack of implicit arguments, it is important to account for predicates such as investment and fund because incorrect prediction of implicit arguments for them can lower precision. This is precisely what happened for the fund predicate, where the model incorrectly identified many implicit arguments for “stock [p fund]” and “mutual [p fund]”. The left context of fund should help the model avoid this type of error; however, our feature selection process did not identify any overall gains from including this information. 6.4 Improvements versus the baseline The baseline heuristic covers the simple case where identical predicates share arguments in the same position. Thus, it is interesting to examine cases where the baseline heuristic failed but the discriminative model succeeded. Consider the following sentence: (12) Mr. Rogers recommends that [p investors] sell [iarg2 takeover-related stock]. Neither NomBank nor the baseline heuristic associate the marked predicate in Example 12 with any arguments; however, the feature-based model was able to correctly identify the marked iarg2 as the entity being invested in. This inference captured a tendency of investors to sell the things they have invested in. We conclude our discussion with an example of an extra-sentential implicit argument: (13) [iarg0 Olivetti] has denied that it violated the rules, asserting that the shipments were properly licensed. However, the legality of these [p sales] is still an open question. As shown in Example 13, the system was able to correctly identify Olivetti as the agent in the selling event of the second sentence. This inference involved two key steps. First, the system identified coreferent mentions of Olivetti that participated in exporting and supplying events (not shown). Second, the system identified a tendency for exporters and suppliers to also be sellers. Using this knowledge, the system extracted information that could not be extracted by the baseline heuristic or a traditional SRL system. 7 Conclusions and future work Current SRL approaches limit the search for arguments to the sentence containing the predicate of interest. Many systems take this assumption a step further and restrict the search to the predicate’s local syntactic environment; however, predicates and the sentences that contain them rarely 1590 exist in isolation. As shown throughout this paper, they are usually embedded in a coherent and semantically rich discourse that must be taken into account. We have presented a preliminary study of implicit arguments for nominal predicates that focused specifically on this problem. Our contribution is three-fold. First, we have created gold-standard implicit argument annotations for a small set of pervasive nominal predicates.7 Our analysis shows that these annotations add 65% to the role coverage of NomBank. Second, we have demonstrated the feasibility of recovering implicit arguments for many of the predicates, thus establishing a baseline for future work on this emerging task. Third, our study suggests a few ways in which this research can be moved forward. As shown in Section 6, many errors were caused by the absence of true implicit arguments within the set of candidate constituents. More intelligent windowing strategies in addition to alternate candidate sources might offer some improvement. Although we consistently observed development gains from using automatic coreference resolution, this process creates errors that need to be studied more closely. It will also be important to study implicit argument patterns of non-verbal predicates such as the partitive percent. These predicates are among the most frequent in the TreeBank and are likely to require approaches that differ from the ones we pursued. Finally, any extension of this work is likely to encounter a significant knowledge acquisition bottleneck. Implicit argument annotation is difficult because it requires both argument and coreference identification (the data produced by Ruppenhofer et al. (2009) is similar). Thus, it might be productive to focus future work on (1) the extraction of relevant knowledge from existing resources (e.g., our use of coreference patterns from Gigaword) or (2) semi-supervised learning of implicit argument models from a combination of labeled and unlabeled data. Acknowledgments We would like to thank the anonymous reviewers for their helpful questions and comments. We would also like to thank Malcolm Doering for his annotation effort. This work was supported in part by NSF grants IIS-0347548 and IIS-0840538. 7Our annotation data can be freely downloaded at http://links.cse.msu.edu:8000/lair/projects/semanticrole.html References Aljoscha Burchardt, Anette Frank, and Manfred Pinkal. 2005. Building text meaning representations from contextually related frames - a case study. In Proceedings of the Sixth International Workshop on Computational Semantics. Xavier Carreras and Llu´ıs M`arquez. 2005. Introduction to the CoNLL-2005 shared task: Semantic role labeling. Nathanael Chambers and Dan Jurafsky. 2008. Unsupervised learning of narrative event chains. In Proceedings of the Association for Computational Linguistics, pages 789–797, Columbus, Ohio, June. Association for Computational Linguistics. Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1):3746. Bradley Efron and Robert J. Tibshirani. 1993. An Introduction to the Bootstrap. Chapman & Hall, New York. Rong-En Fan, Kai-Wei Chang, Cho-Jui Hsieh, XiangRui Wang, and Chih-Jen Lin. 2008. LIBLINEAR: A Library for Large Linear Classification. Journal of Machine Learning Research, 9:1871–1874. Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database (Language, Speech, and Communication). The MIT Press, May. C.J. Fillmore and C.F. Baker. 2001. Frame semantics for text understanding. In Proceedings of WordNet and Other Lexical Resources Workshop, NAACL. Matthew Gerber, Joyce Y. Chai, and Adam Meyers. 2009. The role of implicit argumentation in nominal SRL. In Proceedings of the North American Chapter of the Association for Computational Linguistics, pages 146–154, Boulder, Colorado, USA, June. David Graff. 2003. English Gigaword. Linguistic Data Consortium, Philadelphia. Jan Hajiˇc, Massimiliano Ciaramita, Richard Johansson, Daisuke Kawahara, Maria Ant`onia Mart´ı, Llu´ıs M`arquez, Adam Meyers, Joakim Nivre, Sebastian Pad´o, Jan ˇStˇep´anek, Pavel Straˇn´ak, Mihai Surdeanu, Nianwen Xue, and Yi Zhang. 2009. The CoNLL2009 shared task: Syntactic and semantic dependencies in multiple languages. In Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL 2009): Shared Task, pages 1–18, Boulder, Colorado, June. Association for Computational Linguistics. Ryu Iida, Mamoru Komachi, Kentaro Inui, and Yuji Matsumoto. 2007. Annotating a Japanese text corpus with predicate-argument and coreference relations. In Proceedings of the Linguistic Annotation Workshop in ACL-2007, page 132139. 1591 Kenji Imamura, Kuniko Saito, and Tomoko Izumi. 2009. Discriminative approach to predicateargument structure analysis with zero-anaphora resolution. In Proceedings of the ACL-IJCNLP 2009 Conference Short Papers, pages 85–88, Suntec, Singapore, August. Association for Computational Linguistics. P. Kingsbury, M. Palmer, and M. Marcus. 2002. Adding semantic annotation to the Penn TreeBank. In Proceedings of the Human Language Technology Conference (HLT’02). Karin Kipper. 2005. VerbNet: A broad-coverage, comprehensive verb lexicon. Ph.D. thesis, Department of Computer and Information Science University of Pennsylvania. Mitchell Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: the Penn TreeBank. Computational Linguistics, 19:313–330. Adam Meyers. 2007. Annotation guidelines for NomBank - noun argument structure for PropBank. Technical report, New York University. Martha S. Palmer, Deborah A. Dahl, Rebecca J. Schiffman, Lynette Hirschman, Marcia Linebarger, and John Dowding. 1986. Recovering implicit information. In Proceedings of the 24th annual meeting on Association for Computational Linguistics, pages 10–19, Morristown, NJ, USA. Association for Computational Linguistics. Patrick Pantel and Deepak Ravichandran. 2004. Automatically labeling semantic classes. In Daniel Marcu Susan Dumais and Salim Roukos, editors, HLT-NAACL 2004: Main Proceedings, pages 321–328, Boston, Massachusetts, USA, May 2 May 7. Association for Computational Linguistics. Rashmi Prasad, Alan Lee, Nikhil Dinesh, Eleni Miltsakaki, Geraud Campion, Aravind Joshi, and Bonnie Webber. 2008. Penn discourse treebank version 2.0. Linguistic Data Consortium, February. P. Pudil, J. Novovicova, and J. Kittler. 1994. Floating search methods in feature selection. Pattern Recognition Letters, 15:1119–1125. Vasin Punyakanok, Dan Roth, and Wen-tau Yih. 2008. The importance of syntactic parsing and inference in semantic role labeling. Comput. Linguist., 34(2):257–287. Josef Ruppenhofer, Caroline Sporleder, Roser Morante, Collin Baker, and Martha Palmer. 2009. Semeval-2010 task 10: Linking events and their participants in discourse. In Proceedings of the Workshop on Semantic Evaluations: Recent Achievements and Future Directions (SEW-2009), pages 106–111, Boulder, Colorado, June. Association for Computational Linguistics. Ryohei Sasano, Daisuke Kawahara, and Sadao Kurohashi. 2004. Automatic construction of nominal case frames and its application to indirect anaphora resolution. In Proceedings of Coling 2004, pages 1201–1207, Geneva, Switzerland, Aug 23–Aug 27. COLING. Mihai Surdeanu, Richard Johansson, Adam Meyers, Llu´ıs M`arquez, and Joakim Nivre. 2008. The CoNLL 2008 shared task on joint parsing of syntactic and semantic dependencies. In CoNLL 2008: Proceedings of the Twelfth Conference on Computational Natural Language Learning, pages 159–177, Manchester, England, August. Coling 2008 Organizing Committee. 1592
|
2010
|
160
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 157–166, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Hierarchical Search for Word Alignment Jason Riesa and Daniel Marcu Information Sciences Institute Viterbi School of Engineering University of Southern California {riesa, marcu}@isi.edu Abstract We present a simple yet powerful hierarchical search algorithm for automatic word alignment. Our algorithm induces a forest of alignments from which we can efficiently extract a ranked k-best list. We score a given alignment within the forest with a flexible, linear discriminative model incorporating hundreds of features, and trained on a relatively small amount of annotated data. We report results on Arabic-English word alignment and translation tasks. Our model outperforms a GIZA++ Model-4 baseline by 6.3 points in F-measure, yielding a 1.1 BLEU score increase over a state-of-the-art syntax-based machine translation system. 1 Introduction Automatic word alignment is generally accepted as a first step in training any statistical machine translation system. It is a vital prerequisite for generating translation tables, phrase tables, or syntactic transformation rules. Generative alignment models like IBM Model-4 (Brown et al., 1993) have been in wide use for over 15 years, and while not perfect (see Figure 1), they are completely unsupervised, requiring no annotated training data to learn alignments that have powered many current state-of-the-art translation system. Today, there exist human-annotated alignments and an abundance of other information for many language pairs potentially useful for inducing accurate alignments. How can we take advantage of all of this data at our fingertips? Using feature functions that encode extra information is one good way. Unfortunately, as Moore (2005) points out, it is usually difficult to extend a given generative model with feature functions without changing the entire generative story. This difficulty Sentence 1 the five previous tests have been limited to the target missile and one other body . !"#$%!&!'() "* +,-* !&.( /01234( !5!67*,8.( 9:; <)+,=.( 1>?@!A8BC( DEFG* ) # 1G( ?H() * 1 Figure 1: Model-4 alignment vs. a gold standard. Circles represent links in a human-annotated alignment, and black boxes represent links in the Model-4 alignment. Bold gray boxes show links gained after fully connecting the alignment. has motivated much recent work in discriminative modeling for word alignment (Moore, 2005; Ittycheriah and Roukos, 2005; Liu et al., 2005; Taskar et al., 2005; Blunsom and Cohn, 2006; LacosteJulien et al., 2006; Moore et al., 2006). We present in this paper a discriminative alignment model trained on relatively little data, with a simple, yet powerful hierarchical search procedure. We borrow ideas from both k-best parsing (Klein and Manning, 2001; Huang and Chiang, 2005; Huang, 2008) and forest-based, and hierarchical phrase-based translation (Huang and Chiang, 2007; Chiang, 2007), and apply them to word alignment. Using a foreign string and an English parse tree as input, we formulate a bottom-up search on the parse tree, with the structure of the tree as a backbone for building a hypergraph of possible alignments. Our algorithm yields a forest of 157 the man ate the NP VP S NP the ﺍﻛﻞ ﺍﻟﺮﺟﻞ the ﺍﻛﻞ ﺍﻟﺮﺟﻞ the ﺍﻛﻞ ﺍﻟﺮﺍﺟﻞ man ﺍﻛﻞ ﺍﻟﺮﺟﻞ the man ate the bread ﺍﳋﺒﺰ ﺍﳋﺒﺰ ﺍﳋﺒﺰ ﺍﳋﺒﺰ bread bread ﺍﻛﻞ ﺍﻟﺮﺟﻞ ﺍﳋﺒﺰ Figure 2: Example of approximate search through a hypergraph with beam size = 5. Each black square implies a partial alignment. Each partial alignment at each node is ranked according to its model score. In this figure, we see that the partial alignment implied by the 1-best hypothesis at the leftmost NP node is constructed by composing the best hypothesis at the terminal node labeled “the” and the 2ndbest hypothesis at the terminal node labeled “man”. (We ignore terminal nodes in this toy example.) Hypotheses at the root node imply full alignment structures. word alignments, from which we can efficiently extract the k-best. We handle an arbitrary number of features, compute them efficiently, and score alignments using a linear model. We train the parameters of the model using averaged perceptron (Collins, 2002) modified for structured outputs, but can easily fit into a max-margin or related framework. Finally, we use relatively little training data to achieve accurate word alignments. Our model can generate arbitrary alignments and learn from arbitrary gold alignments. 2 Word Alignment as a Hypergraph Algorithm input The input to our alignment algorithm is a sentence-pair (en 1, f m 1 ) and a parse tree over one of the input sentences. In this work, we parse our English data, and for each sentence E = en 1, let T be its syntactic parse. To generate parse trees, we use the Berkeley parser (Petrov et al., 2006), and use Collins head rules (Collins, 2003) to head-out binarize each tree. Overview We present a brief overview here and delve deeper in Section 2.1. Word alignments are built bottom-up on the parse tree. Each node v in the tree holds partial alignments sorted by score. 158 u11 u12 u13 u21 2.2 4.1 5.5 u22 2.4 3.5 7.2 u23 3.2 4.5 11.4 u11 u12 u13 u21 2.2 4.1 5.5 u22 2.4 3.5 7.2 u23 3.2 4.5 11.4 u11 u12 u13 u21 2.2 4.1 5.5 u22 2.4 3.5 7.2 u23 3.2 4.5 11.4 (a) Score the left corner alignment first. Assume it is the 1best. Numbers in the rest of the boxes are hidden at this point. u11 u12 u13 u21 2.2 4.1 5.5 u22 2.4 3.5 7.2 u23 3.2 4.5 11.4 u11 u12 u13 u21 2.2 4.1 5.5 u22 2.4 3.5 7.2 u23 3.2 4.5 11.4 u11 u12 u13 u21 2.2 4.1 5.5 u22 2.4 3.5 7.2 u23 3.2 4.5 11.4 (b) Expand the frontier of alignments. We are now looking for the 2nd best. u11 u12 u13 u21 2.2 4.1 5.5 u22 2.4 3.5 7.2 u23 3.2 4.5 11.4 u11 u12 u13 u21 2.2 4.1 5.5 u22 2.4 3.5 7.2 u23 3.2 4.5 11.4 u11 u12 u13 u21 2.2 4.1 5.5 u22 2.4 3.5 7.2 u23 3.2 4.5 11.4 (c) Expand the frontier further. After this step we have our top k alignments. Figure 3: Cube pruning with alignment hypotheses to select the top-k alignments at node v with children ⟨u1, u2⟩. In this example, k = 3. Each box represents the combination of two partial alignments to create a larger one. The score in each box is the sum of the scores of the child alignments plus a combination cost. Each partial alignment comprises the columns of the alignment matrix for the e-words spanned by v, and each is scored by a linear combination of feature functions. See Figure 2 for a small example. Initial partial alignments are enumerated and scored at preterminal nodes, each spanning a single column of the word alignment matrix. To speed up search, we can prune at each node, keeping a beam of size k. In the diagram depicted in Figure 2, the beam is size k = 5. From here, we traverse the tree nodes bottomup, combining partial alignments from child nodes until we have constructed a single full alignment at the root node of the tree. If we are interested in the k-best, we continue to populate the root node until we have k alignments.1 We use one set of feature functions for preterminal nodes, and another set for nonterminal nodes. This is analogous to local and nonlocal feature functions for parse-reranking used by Huang (2008). Using nonlocal features at a nonterminal node emits a combination cost for composing a set of child partial alignments. Because combination costs come into play, we use cube pruning (Chiang, 2007) to approximate the k-best combinations at some nonterminal node v. Inference is exact when only local features are used. Assumptions There are certain assumptions related to our search algorithm that we must make: 1We use approximate dynamic programming to store alignments, keeping only scored lists of pointers to initial single-column spans. Each item in the list is a derivation that implies a partial alignment. (1) that using the structure of 1-best English syntactic parse trees is a reasonable way to frame and drive our search, and (2) that F-measure approximately decomposes over hyperedges. We perform an oracle experiment to validate these assumptions. We find the oracle for a given (T,e, f) triple by proceeding through our search algorithm, forcing ourselves to always select correct links with respect to the gold alignment when possible, breaking ties arbitrarily. The the F1 score of our oracle alignment is 98.8%, given this “perfect” model. 2.1 Hierarchical search Initial alignments We can construct a word alignment hierarchically, bottom-up, by making use of the structure inherent in syntactic parse trees. We can think of building a word alignment as filling in an M×N matrix (Figure 1), and we begin by visiting each preterminal node in the tree. Each of these nodes spans a single e word. (Line 2 in Algorithm 1). From here we can assign links from each e word to zero or more f words (Lines 6–14). At this level of the tree the span size is 1, and the partial alignment we have made spans a single column of the matrix. We can make many such partial alignments depending on the links selected. Lines 5 through 9 of Algorithm 1 enumerate either the null alignment, single-link alignments, or two-link alignments. Each partial alignment is scored and stored in a sorted heap (Lines 9 and 13). In practice enumerating all two-link alignments can be prohibitive for long sentence pairs; we set a practical limit and score only pairwise combina159 Algorithm 1: Hypergraph Alignment Input: Source sentence en 1 Target sentence f m 1 Parse tree T over en 1 Set of feature functions h Weight vector w Beam size k Output: A k-best list of alignments over en 1 and f m 1 1 function A(en 1, f m 1 , T) 2 for v ∈T in bottom-up order do 3 αv ←∅ 4 if -PN(v) then 5 i ←index-of(v) 6 for j = 0 to m do 7 links ←(i, j) 8 score ←w · h(links, v, en 1, f m 1 ) 9 P(αv, ⟨score, links⟩, k ) 10 for k = j + 1 to m do 11 links ←(i, j), (i, k) 12 score ←w · h(links, v, en 1, f m 1 ) 13 P(αv, ⟨score, links⟩, k ) 14 end 15 end 16 else 17 αv ←GS(children(v), k) 18 end 19 end 20 end 21 function GS(⟨u1, u2⟩, k) 22 return CP(⟨αu1, αu2⟩, k,w,h) 23 end tions of the top n = max n |f| 2 , 10 o scoring singlelink alignments. We limit the number of total partial alignments αv kept at each node to k. If at any time we wish to push onto the heap a new partial alignment when the heap is full, we pop the current worst offthe heap and replace it with our new partial alignment if its score is better than the current worst. Building the hypergraph We now visit internal nodes (Line 16) in the tree in bottom-up order. At each nonterminal node v we wish to combine the partial alignments of its children u1, . . . , uc. We use cube pruning (Chiang, 2007; Huang and Chiang, 2007) to select the k-best combinations of the partial alignments of u1, . . . , uc (Line 19). Note Sentence 1 TOP1 S2 NP-C1 NPB2 DT NPB-BAR2 CD NPB-BAR2 JJ NNS S-BAR1 VP1 VBP VP-C1 VBN VP-C1 VBN PP1 IN NP-C1 NP-C-BAR1 NP1 NPB2 DT NPB-BAR2 NN NN CC NP1 NPB2 CD NPB-BAR2 JJ NN . the five previous tests have been limited to the target missile and one other body . !"#$%!&!'() "* +,-* !&.( /01234( !5!67*,8.( 9:; <)+,=.( 1>?@!A8BC( DEFG* ) # 1G( ?H() * Figure 4: Correct version of Figure 1 after hypergraph alignment. Subscripts on the nonterminal labels denote the branch containing the head word for that span. that Algorithm 1 assumes a binary tree2, but is not necessary. In the general case, cube pruning will operate on a d-dimensional hypercube, where d is the branching factor of node v. We cannot enumerate and score every possibility; without the cube pruning approximation, we will have kc possible combinations at each node, exploding the search space exponentially. Figure 3 depicts how we select the top-k alignments at a node v from its children ⟨u1, u2 ⟩. 3 Discriminative training We incorporate all our new features into a linear model and learn weights for each using the online averaged perceptron algorithm (Collins, 2002) with a few modifications for structured outputs inspired by Chiang et al. (2008). We define: 2We find empirically that using binarized trees reduces search errors in cube pruning. 160 in in !" !" . . . ... ... Figure 5: A common problem with GIZA++ Model 4 alignments is a weak distortion model. The second English “in” is aligned to the wrong Arabic token. Circles show the gold alignment. γ(y) = ℓ(yi, y) + w · (h(yi) −h(y)) (1) where ℓ(yi,y) is a loss function describing how bad it is to guess y when the correct answer is yi. In our case, we define ℓ(yi,y) as 1−F1(yi,y). We select the oracle alignment according to: y+ = arg min y∈(x) γ(y) (2) where (x) is a set of hypothesis alignments generated from input x. Instead of the traditional oracle, which is calculated solely with respect to the loss ℓ(yi,y), we choose the oracle that jointly minimizes the loss and the difference in model score to the true alignment. Note that Equation 2 is equivalent to maximizing the sum of the Fmeasure and model score of y: y+ = arg max y∈(x) (F1(yi, y) + w · h(y)) (3) Let ˆy be the 1-best alignment according to our model: ˆy = arg max y∈(x) w · h(y) (4) Then, at each iteration our weight update is: w ←w + η(h(y+) −h(ˆy)) (5) where η is a learning rate parameter.3 We find that this more conservative update gives rise to a much more stable search. After each iteration, we expect y+ to get closer and closer to the true yi. 4 Features Our simple, flexible linear model makes it easy to throw in many features, mapping a given complex 3We set η to 0.05 in our experiments. alignment structure into a single high-dimensional feature vector. Our hierarchical search framework allows us to compute these features when needed, and affords us extra useful syntactic information. We use two classes of features: local and nonlocal. Huang (2008) defines a feature h to be local if and only if it can be factored among the local productions in a tree, and non-local otherwise. Analogously for alignments, our class of local features are those that can be factored among the local partial alignments competing to comprise a larger span of the matrix, and non-local otherwise. These features score a set of links and the words connected by them. Feature development Our features are inspired by analysis of patterns contained among our gold alignment data and automatically generated parse trees. We use both local lexical and nonlocal structural features as described below. 4.1 Local features These features fire on single-column spans. • From the output of GIZA++ Model 4, we compute lexical probabilities p(e | f) and p(f | e), as well as a fertility table φ(e). From the fertility table, we fire features φ0(e), φ1(e), and φ2+(e) when a word e is aligned to zero, one, or two or more words, respectively. Lexical probability features p(e | f) and p(f | e) fire when a word e is aligned to a word f. • Based on these features, we include a binary lexical-zero feature that fires if both p(e | f) and p(f | e) are equal to zero for a given word pair (e, f). Negative weights essentially penalize alignments with links never seen before in the Model 4 alignment, and positive weights encourage such links. We employ a separate instance of this feature for each English part-of-speech tag: p(f | e, t). We learn a different feature weight for each. Critically, this feature tells us how much to trust alignments involving nouns, verbs, adjectives, function words, punctuation, etc. from the Model 4 alignments from which our p(e | f) and p(f | e) tables are built. Table 1 shows a sample of learned weights. Intuitively, alignments involving English partsof-speech more likely to be content words (e.g. NNPS, NNS, NN) are more trustworthy 161 PP IN NP eprep ehead ... f NP DT NP edet ehead ... f VP VBD VP everb ehead ... f Figure 6: Features PP-NP-head, NP-DT-head, and VP-VP-head fire on these tree-alignment patterns. For example, PP-NP-head fires exactly when the head of the PP is aligned to exactly the same f words as the head of it’s sister NP. Penalty NNPS −1.11 NNS −1.03 NN −0.80 NNP −0.62 VB −0.54 VBG −0.52 JJ −0.50 JJS −0.46 VBN −0.45 ... ... POS −0.0093 EX −0.0056 RP −0.0037 WP$ −0.0011 TO 0.037 Reward Table 1: A sampling of learned weights for the lexical zero feature. Negative weights penalize links never seen before in a baseline alignment used to initialize lexical p(e | f) and p(f | e) tables. Positive weights outright reward such links. than those likely to be function words (e.g. TO, RP, EX), where the use of such words is often radically different across languages. • We also include a measure of distortion. This feature returns the distance to the diagonal of the matrix for any link in a partial alignment. If there is more than one link, we return the distance of the link farthest from the diagonal. • As a lexical backoff, we include a tag probability feature, p(t | f) that fires for some link (e, f) if the part-of-speech tag of e is t. The conditional probabilities in this table are computed from our parse trees and the baseline Model 4 alignments. • In cases where the lexical probabilities are too strong for the distortion feature to overcome (see Figure 5), we develop the multiple-distortion feature. Although local features do not know the partial alignments at other spans, they do have access to the entire English sentence at every step because our input is constant. If some e exists more than once in en 1 we fire this feature on all links containing word e, returning again the distance to the diagonal for that link. We learn a strong negative weight for this feature. • We find that binary identity and punctuation-mismatch features are important. The binary identity feature fires if e = f, and proves useful for untranslated numbers, symbols, names, and punctuation in the data. Punctuation-mismatch fires on any link that causes nonpunctuation to be aligned to punctuation. Additionally, we include fine-grained versions of the lexical probability, fertility, and distortion features. These fire for for each link (e, f) and partof-speech tag. That is, we learn a separate weight for each feature for each part-of-speech tag in our data. Given the tag of e, this affords the model the ability to pay more or less attention to the features described above depending on the tag given to e. Arabic-English specific features We describe here language specific features we implement to exploit shallow Arabic morphology. 162 PP IN NP from ... !" ... Figure 7: This figure depicts the tree/alignment structure for which the feature PP-from-prep fires. The English preposition “from” is aligned to Arabic word áÓ. Any aligned words in the span of the sister NP are aligned to words following áÓ. English preposition structure commonly matches that of Arabic in our gold data. This family of features captures these observations. • We observe the Arabic prefixð, transliterated w- and generally meaning and, to prepend to most any word in the lexicon, so we define features p¬w(e | f) and p¬w(f | e). If f begins with w-, we strip offthe prefix and return the values of p(e | f) and p(f | e). Otherwise, these features return 0. • We also include analogous feature functions for several functional and pronominal prefixes and suffixes.4 4.2 Nonlocal features These features comprise the combination cost component of a partial alignment score and may fire when concatenating two partial alignments to create a larger span. Because these features can look into any two arbitrary subtrees, they are considered nonlocal features as defined by Huang (2008). • Features PP-NP-head, NP-DT-head, and VP-VP-head (Figure 6) all exploit headwords on the parse tree. We observe English prepositions and determiners to often align to the headword of their sister. Likewise, we observe the head of a VP to align to the head of an immediate sister VP. 4Affixes used by our model are currently: K., Ë, Ë@, ËAK., ù , Õº, AÒº, Ñê, AÒê. Others either we did not experiment with, or seemed to provide no significant benefit, and are not included. In Figure 4, when the search arrives at the left-most NPB node, the NP-DT-head feature will fire given this structure and links over the span [the ... tests]. When search arrives at the second NPB node, it will fire given the structure and links over the span [the ... missle], but will not fire at the right-most NPB node. • Local lexical preference features compete with the headword features described above. However, we also introduce nonlocal lexicalized features for the most common types of English and foreign prepositions to also compete with these general headword features. PP features PP-of-prep, PP-from-prep, PPto-prep, PP-on-prep, and PP-in-prep fire at any PP whose left child is a preposition and right child is an NP. The head of the PP is one of the enumerated English prepositions and is aligned to any of the three most common foreign words to which it has also been observed aligned in the gold alignments. The last constraint on this pattern is that all words under the span of the sister NP, if aligned, must align to words following the foreign preposition. Figure 7 illustrates this pattern. • Finally, we have a tree-distance feature to avoid making too many many-to-one (from many English words to a single foreign word) links. This is a simplified version of and similar in spirit to the tree distance metric used in (DeNero and Klein, 2007). For any pair of links (ei, f) and (e j, f) in which the e words differ but the f word is the same token in each, return the tree height of first common ancestor of ei and ej. This feature captures the intuition that it is much worse to align two English words at different ends of the tree to the same foreign word, than it is to align two English words under the same NP to the same foreign word. To see why a string distance feature that counts only the flat horizontal distance from ei to e j is not the best strategy, consider the following. We wish to align a determiner to the same f word as its sister head noun under the same NP. Now suppose there are several intermediate adjectives separating the determiner and noun. A string distance met163 ric, with no knowledge of the relationship between determiner and noun will levy a much heavier penalty than its tree distance analog. 5 Related Work Recent work has shown the potential for syntactic information encoded in various ways to support inference of superior word alignments. Very recent work in word alignment has also started to report downstream effects on BLEU score. Cherry and Lin (2006) introduce soft syntactic ITG (Wu, 1997) constraints into a discriminative model, and use an ITG parser to constrain the search for a Viterbi alignment. Haghighi et al. (2009) confirm and extend these results, showing BLEU improvement for a hierarchical phrasebased MT system on a small Chinese corpus. As opposed to ITG, we use a linguistically motivated phrase-structure tree to drive our search and inform our model. And, unlike ITG-style approaches, our model can generate arbitrary alignments and learn from arbitrary gold alignments. DeNero and Klein (2007) refine the distortion model of an HMM aligner to reflect tree distance instead of string distance. Fossum et al. (2008) start with the output from GIZA++ Model-4 union, and focus on increasing precision by deleting links based on a linear discriminative model exposed to syntactic and lexical information. Fraser and Marcu (2007) take a semi-supervised approach to word alignment, using a small amount of gold data to further tune parameters of a headword-aware generative model. They show a significant improvement over a Model-4 union baseline on a very large corpus. 6 Experiments We evaluate our model and and resulting alignments on Arabic-English data against those induced by IBM Model-4 using GIZA++ (Och and Ney, 2003) with both the union and grow-diagfinal heuristics. We use 1,000 sentence pairs and gold alignments from LDC2006E86 to train model parameters: 800 sentences for training, 100 for testing, and 100 as a second held-out development set to decide when to stop perceptron training. We also align the test data using GIZA++5 along with 50 million words of English. 5We use a standard training procedure: 5 iterations of Model-1, 5 iterations of HMM, 3 iterations of Model-3, and 3 iterations of Model-4. 0 5 10 15 20 25 30 35 40 0.73 0.735 0.74 0.745 0.75 0.755 0.76 0.765 0.77 0.775 Training epoch Training F−measure Figure 8: Learning curves for 10 random restarts over time for parallel averaged perceptron training. These plots show the current F-measure on the training set as time passes. Perceptron training here is quite stable, converging to the same general neighborhood each time. 0.67 0.68 0.69 0.70 0.71 0.72 0.73 0.74 0.75 0.76 Model 1 HMM Model 4 F-measure Initial alignments Figure 9: Model robustness to the initial alignments from which the p(e | f) and p(f | e) features are derived. The dotted line indicates the baseline accuracy of GIZA++ Model 4 alone. 6.1 Alignment Quality We empirically choose our beam size k from the results of a series of experiments, setting k=1, 2, 4, 8, 16, 32, and 64. We find setting k = 16 to yield the highest accuracy on our held-out test data. Using wider beams results in higher F-measure on training data, but those gains do not translate into higher accuracy on held-out data. The first three columns of Table 2 show the balanced F-measure, Precision, and Recall of our alignments versus the two GIZA++ Model-4 baselines. We report an F-measure 8.6 points over Model-4 union, and 6.3 points over Model-4 growdiag-final. 164 F P R Arabic/English # Unknown BLEU Words M4 (union) .665 .636 .696 45.1 2,538 M4 (grow-diag-final) .688 .702 .674 46.4 2,262 Hypergraph alignment .751 .780 .724 47.5 1,610 Table 2: F-measure, Precision, Recall, the resulting BLEU score, and number of unknown words on a held-out test corpus for three types of alignments. BLEU scores are case-insensitive IBM BLEU. We show a 1.1 BLEU increase over the strongest baseline, Model-4 grow-diag-final. This is statistically significant at the p < 0.01 level. Figure 8 shows the stability of the search procedure over ten random restarts of parallel averaged perceptron training with 40 CPUs. Training examples are randomized at each epoch, leading to slight variations in learning curves over time but all converge into the same general neighborhood. Figure 9 shows the robustness of the model to initial alignments used to derive lexical features p(e | f) and p(f | e). In addition to IBM Model 4, we experiment with alignments from Model 1 and the HMM model. In each case, we significantly outperform the baseline GIZA++ Model 4 alignments on a heldout test set. 6.2 MT Experiments We align a corpus of 50 million words with GIZA++ Model-4, and extract translation rules from a 5.4 million word core subset. We align the same core subset with our trained hypergraph alignment model, and extract a second set of translation rules. For each set of translation rules, we train a machine translation system and decode a held-out test corpus for which we report results below. We use a syntax-based translation system for these experiments. This system transforms Arabic strings into target English syntax trees Translation rules are extracted from (e-tree, f-string, alignment) triples as in (Galley et al., 2004; Galley et al., 2006). We use a randomized language model (similar to that of Talbot and Brants (2008)) of 472 million English words. We tune the the parameters of the MT system on a held-out development corpus of 1,172 parallel sentences, and test on a heldout parallel corpus of 746 parallel sentences. Both corpora are drawn from the NIST 2004 and 2006 evaluation data, with no overlap at the document or segment level with our training data. Columns 4 and 5 in Table 2 show the results of our MT experiments. Our hypergraph alignment algorithm allows us a 1.1 BLEU increase over the best baseline system, Model-4 grow-diag-final. This is statistically significant at the p < 0.01 level. We also report a 2.4 BLEU increase over a system trained with alignments from Model-4 union. 7 Conclusion We have opened up the word alignment task to advances in hypergraph algorithms currently used in parsing and machine translation decoding. We treat word alignment as a parsing problem, and by taking advantage of English syntax and the hypergraph structure of our search algorithm, we report significant increases in both F-measure and BLEU score over standard baselines in use by most state-of-the-art MT systems today. Acknowledgements We would like to thank our colleagues in the Natural Language Group at ISI for many meaningful discussions and the anonymous reviewers for their thoughtful suggestions. This research was supported by DARPA contract HR0011-06-C-0022 under subcontract to BBN Technologies, and a USC CREATE Fellowship to the first author. References Phil Blunsom and Trevor Cohn. 2006. Discriminative Word Alignment with Conditional Random Fields. In Proceedings of the 44th Annual Meeting of the ACL. Sydney, Australia. Peter F. Brown, Stephen A. Della Pietra, Vincent Della J. Pietra, and Robert L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19(2):263– 312. MIT Press. Camrbidge, MA. USA. 165 Colin Cherry and Dekang Lin. 2006. Soft Syntactic Constraints for Word Alignment through Discriminative Training. In Proceedings of the 44th Annual Meeting of the ACL. Sydney, Australia. David Chiang. 2007. Hierarchical phrase-based translation. Computational Linguistics. 33(2):201–228. MIT Press. Cambridge, MA. USA. David Chiang, Yuval Marton, and Philip Resnik. 2008. Online Large-Margin Training of Syntactic and Structural Translation Features. In Proceedings of EMNLP. Honolulu, HI. USA. Michael Collins. 2003. Head-Driven Statistical Models for Natural Language Parsing. Computational Linguistics. 29(4):589–637. MIT Press. Cambridge, MA. USA. Michael Collins 2002. Discriminative training methods for hidden markov models: Theory and experiments with perceptron algorithms. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. John DeNero and Dan Klein. 2007. Tailoring Word Alignments to Syntactic Machine Translation. In Proceedings of the 45th Annual Meeting of the ACL. Prague, Czech Republic. Alexander Fraser and Daniel Marcu. 2007. Getting the Structure Right for Word Alignment: LEAF. In Proceedings of EMNLP-CoNLL. Prague, Czech Republic. Victoria Fossum, Kevin Knight, and Steven Abney. 2008. Using Syntax to Improve Word Alignment Precision for Syntax-Based Machine Translation. In Proceedings of the Third Workshop on Statistical Machine Translation. Columbus, Ohio. Dan Klein and Christopher D. Manning. 2001. Parsing and Hypergraphs. In Proceedings of the 7th International Workshop on Parsing Technologies. Beijing, China. Aria Haghighi, John Blitzer, and Dan Klein. 2009. Better Word Alignments with Supervised ITG Models. In Proceedings of ACL-IJCNLP 2009. Singapore. Liang Huang and David Chiang. 2005. Better k-best Parsing. In Proceedings of the 9th International Workshop on Parsing Technologies. Vancouver, BC. Canada. Liang Huang and David Chiang. 2007. Forest Rescoring: Faster Decoding with Integrated Language Models. In Proceedings of the 45th Annual Meeting of the ACL. Prague, Czech Republic. Liang Huang. 2008. Forest Reranking: Discriminative Parsing with Non-Local Features. In Proceedings of the 46th Annual Meeting of the ACL. Columbus, OH. USA. Michel Galley, Mark Hopkins, Kevin Knight, and Daniel Marcu. 2004. What’s in a Translation Rule? In Proceedings of NAACL. Michel Galley, Jonathan Graehl, Kevin Knight, Daniel Marcu, Steve DeNeefe, Wei Wang, and Ignacio Thayer. 2006. Scalable Inference and Training of Context-Rich Syntactic Models In Proceedings of the 44th Annual Meeting of the ACL. Sydney, Australia. Abraham Ittycheriah and Salim Roukos. 2005. A maximum entropy word aligner for Arabic-English machine translation. In Proceedings of HLT-EMNLP. Vancouver, BC. Canada. Simon Lacoste-Julien, Ben Taskar, Dan Klein, and Michael I. Jordan. 2006. Word alignment via Quadratic Assignment. In Proceedings of HLTEMNLP. New York, NY. USA. Yang Liu, Qun Liu, and Shouxun Lin. 2005. Loglinear Models for Word Alignment In Proceedings of the 43rd Annual Meeting of the ACL. Ann Arbor, Michigan. USA. Robert C. Moore. 2005. A Discriminative Framework for Word Alignment. In Proceedings of EMNLP. Vancouver, BC. Canada. Robert C. Moore, Wen-tau Yih, and Andreas Bode. 2006. Improved Discriminative Bilingual Word Alignment In Proceedings of the 44th Annual Meeting of the ACL. Sydney, Australia. Franz Josef Och and Hermann Ney. 2003. A Systematic Comparison of Various Statistical Alignment Models. Computational Linguistics. 29(1):19–52. MIT Press. Cambridge, MA. USA. Slav Petrov, Leon Barrett, Romain Thibaux and Dan Klein 2006. Learning Accurate, Compact, and Interpretable Tree Annotation In Proceedings of the 44th Annual Meeting of the ACL. Sydney, Australia. Kishore Papineni, Salim Roukos, T. Ward, and W-J. Zhu. 2002. BLEU: A Method for Automatic Evaluation of Machine Translation In Proceedings of the 40th Annual Meeting of the ACL. Philadelphia, PA. USA. Ben Taskar, Simon Lacoste-Julien, and Dan Klein. 2005. A Discriminative Matching Approach to Word Alignment. In Proceedings of HLT-EMNLP. Vancouver, BC. Canada. David Talbot and Thorsten Brants. 2008. Randomized Language Models via Perfect Hash Functions. In Proceedings of ACL-08: HLT. Columbus, OH. USA. Dekai Wu. 1997. Stochastic inversion transduction grammars and bilingual parsing of parallel corpora. Computational Linguistics. 23(3):377–404. MIT Press. Cambridge, MA. USA. 166
|
2010
|
17
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 167–176, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics “Was it good? It was provocative.” Learning the meaning of scalar adjectives Marie-Catherine de Marneffe, Christopher D. Manning and Christopher Potts Linguistics Department Stanford University Stanford, CA 94305 {mcdm,manning,cgpotts}@stanford.edu Abstract Texts and dialogues often express information indirectly. For instance, speakers’ answers to yes/no questions do not always straightforwardly convey a ‘yes’ or ‘no’ answer. The intended reply is clear in some cases (Was it good? It was great!) but uncertain in others (Was it acceptable? It was unprecedented.). In this paper, we present methods for interpreting the answers to questions like these which involve scalar modifiers. We show how to ground scalar modifier meaning based on data collected from the Web. We learn scales between modifiers and infer the extent to which a given answer conveys ‘yes’ or ‘no’. To evaluate the methods, we collected examples of question–answer pairs involving scalar modifiers from CNN transcripts and the Dialog Act corpus and use response distributions from Mechanical Turk workers to assess the degree to which each answer conveys ‘yes’ or ‘no’. Our experimental results closely match the Turkers’ response data, demonstrating that meanings can be learned from Web data and that such meanings can drive pragmatic inference. 1 Introduction An important challenge for natural language processing is how to learn not only basic linguistic meanings but also how those meanings are systematically enriched when expressed in context. For instance, answers to polar (yes/no) questions do not always explicitly contain a ‘yes’ or ‘no’, but rather give information that the hearer can use to infer such an answer in a context with some degree of certainty. Hockey et al. (1997) find that 27% of answers to polar questions do not contain a direct ‘yes’ or ‘no’ word, 44% of which they regard as failing to convey a clear ‘yes’ or ‘no’ response. In some cases, interpreting the answer is straightforward (Was it bad? It was terrible.), but in others, what to infer from the answer is unclear (Was it good? It was provocative.). It is even common for the speaker to explicitly convey his own uncertainty with such answers. In this paper, we focus on the interpretation of answers to a particular class of polar questions: ones in which the main predication involves a gradable modifier (e.g., highly unusual, not good, little) and the answer either involves another gradable modifier or a numerical expression (e.g., seven years old, twenty acres of land). Interpreting such question–answer pairs requires dealing with modifier meanings, specifically, learning context-dependent scales of expressions (Horn, 1972; Fauconnier, 1975) that determine how, and to what extent, the answer as a whole resolves the issue raised by the question. We propose two methods for learning the knowledge necessary for interpreting indirect answers to questions involving gradable adjectives, depending on the type of predications in the question and the answer. The first technique deals with pairs of modifiers: we hypothesized that online, informal review corpora in which people’s comments have associated ratings would provide a general-purpose database for mining scales between modifiers. We thus use a large collection of online reviews to learn orderings between adjectives based on contextual entailment (good < excellent), and employ this scalar relationship to infer a yes/no answer (subject to negation and other qualifiers). The second strategy targets numerical answers. Since it is unclear what kind of corpora would contain the relevant information, we turn to the Web in general: we use distributional information retrieved via Web searches to assess whether the numerical measure counts as a posi167 tive or negative instance of the adjective in question. Both techniques exploit the same approach: using texts (the Web) to learn meanings that can drive pragmatic inference in dialogue. This paper demonstrates to some extent that meaning can be grounded from text in this way. 2 Related work Indirect speech acts are studied by Clark (1979), Perrault and Allen (1980), Allen and Perrault (1980) and Asher and Lascarides (2003), who identify a wide range of factors that govern how speakers convey their intended messages and how hearers seek to uncover those messages from uncertain and conflicting signals. In the computational literature, Green and Carberry (1994, 1999) provide an extensive model that interprets and generates indirect answers to polar questions. They propose a logical inference model which makes use of discourse plans and coherence relations to infer categorical answers. However, to adequately interpret indirect answers, the uncertainty inherent in some answers needs to be captured (de Marneffe et al., 2009). While a straightforward ‘yes’ or ‘no’ response is clear in some indirect answers, such as in (1), the intended answer is less certain in other cases (2):1 (1) A: Do you think that’s a good idea, that we just begin to ignore these numbers? B: I think it’s an excellent idea. (2) A: Is he qualified? B: I think he’s young. In (2), it might be that the answerer does not know about qualifications or does not want to talk about these directly, and therefore shifts the topic slightly. As proposed by Zeevat (1994) in his work on partial answers, the speaker’s indirect answer might indicate that he is deliberately leaving the original question only partially addressed, while giving a fully resolving answer to another one. The hearer must then interpret the answer to work out the other question. In (2) in context, we get a sense that the speaker would resolve the issue to ‘no’, but that he is definitely not committed to that in any strong sense. Uncertainty can thus reside both on the speaker and the hearer sides, and the four following scenarios are attested in conversation: 1Here and throughout, the examples come from the corpus described in section 3. a. The speaker is certain of ‘yes’ or ‘no’ and conveys that directly and successfully to the hearer. b. The speaker is certain of ‘yes’ or ‘no’ but conveys this only partially to the hearer. c. The speaker is uncertain of ‘yes’ or ‘no’ and conveys this uncertainty to the hearer. d. The speaker is uncertain of ‘yes’ or ‘no’, but the hearer infers one of those with confidence. The uncertainty is especially pressing for predications built around scalar modifiers, which are inherently vague and highly context-dependent (Kamp and Partee, 1995; Kennedy and McNally, 2005; Kennedy, 2007). For example, even if we fix the basic sense for little to mean ‘young for a human’, there is a substantial amount of gray area between the clear instances (babies) and the clear non-instances (adults). This is the source of uncertainty in (3), in which B’s children fall into the gray area. (3) A: Are your kids little? B: I have a seven year-old and a ten year-old. 3 Corpus description Since indirect answers are likely to arise in interviews, to gather instances of question–answer pairs involving gradable modifiers (which will serve to evaluate the learning techniques), we use online CNN interview transcripts from five different shows aired between 2000 and 2008 (Anderson Cooper, Larry King Live, Late Edition, Lou Dobbs Tonight, The Situation Room). We also searched the Switchboard Dialog Act corpus (Jurafsky et al., 1997). We used regular expressions and manual filtering to find examples of twoutterance dialogues in which the question and the reply contain some kind of gradable modifier. 3.1 Types of question–answer pairs In total, we ended up with 224 question–answer pairs involving gradable adjectives. However our collection contains different types of answers, which naturally fall into two categories: (I) in 205 dialogues, both the question and the answer contain a gradable modifier; (II) in 19 dialogues, the reply contains a numerical measure (as in (3) above and (4)). 168 Modification in answer Example Count I Other adjective (1), (2) 125 Adverb - same adjective (5) 55 Negation - same adjective (6), (7) 21 Omitted adjective (8) 4 II Numerical measure (3), (4) 19 Table 1: Types of question–answer pairs, and counts in the corpus. I Modification in answer Mean SD Other adjective 1.1 0.6 Adverb - same adjective 0.8 0.6 Negation - same adjective 1.0 0.5 Omitted adjective 1.1 0.2 II Numerical measure 1.5 0.2 Table 2: Mean entropy values and standard deviation obtained in the Mechanical Turk experiment for each question–answer pair category. (4) A: Have you been living there very long? B: I’m in here right now about twelve and a half years. Category I, which consists of pairs of modifiers, can be further divided. In most dialogues, the answer contains another adjective than the one used in the question, such as in (1). In others, the answer contains the same adjective as in the question, but modified by an adverb (e.g., very, basically, quite) as in (5) or a negation as in (6). (5) A: That seems to be the biggest sign of progress there. Is that accurate? B: That’s absolutely accurate. (6) A: Are you bitter? B: I’m not bitter because I’m a soldier. The negation can be present in the main clause when the adjectival predication is embedded, as in example (7). (7) A: [. . . ] Is that fair? B: I don’t think that’s a fair statement. In a few cases, when the question contains an adjective modifying a noun, the adjective is omitted in the answer: (8) A: Is that a huge gap in the system? B: It is a gap. Table 1 gives the distribution of the types appearing in the corpus. 3.2 Answer assignment To assess the degree to which each answer conveys ‘yes’ or ‘no’ in context, we use response distributions from Mechanical Turk workers. Given a written dialogue between speakers A and B, Turkers were asked to judge what B’s answer conveys: ‘definite yes’, ‘probable yes’, ‘uncertain’, ‘probable no’, ‘definite no’. Within each of the two ‘yes’ and ‘no’ pairs, there is a scalar relationship, but the pairs themselves are not in a scalar relationship with each other, and ‘uncertain’ is arguably a separate judgment. Figure 1 shows the exact formulation used in the experiment. For each dialogue, we got answers from 30 Turkers, and we took the dominant response as the correct one though we make extensive use of the full response distributions in evaluating our approach.2 We also computed entropy values for the distribution of answers for each item. Overall, the agreement was good: 21 items have total agreement (entropy of 0.0 — 11 in the “adjective” category, 9 in the “adverb-adjective” category and 1 in the “negation” category), and 80 items are such that a single response got chosen 20 or more times (entropy < 0.9). The dialogues in (1) and (9) are examples of total agreement. In contrast, (10) has response entropy of 1.1, and item (11) has the highest entropy of 2.2. (9) A: Advertisements can be good or bad. Was it a good ad? B: It was a great ad. (10) A: Am I clear? B: I wish you were a little more forthright. (11) A: 91 percent of the American people still express confidence in the long-term prospect of the U.S. economy; only 8 percent are not confident. Are they overly optimistic, in your professional assessment? 2120 Turkers were involved (the median number of items done was 28 and the mean 56.5). The Fleiss’ Kappa score for the five response categories is 0.46, though these categories are partially ordered. For the three-category response system used in section 5, which arguably has no scalar ordering, the Fleiss’ Kappa is 0.63. Despite variant individual judgments, aggregate annotations done with Mechanical Turk have been shown to be reliable (Snow et al., 2008; Sheng et al., 2008; Munro et al., 2010). Here, the relatively low Kappa scores also reflect the uncertainty inherent in many of our examples, uncertainty that we seek to characterize and come to grips with computationally. 169 Indirect Answers to Yes/No Questions In the following dialogue, speaker A asks a simple Yes/No question, but speaker B answers with something more indirect and complicated. dialoguehere Which of the following best captures what speaker B meant here: • B definitely meant to convey “Yes”. • B probably meant to convey “Yes”. • B definitely meant to convey “No”. • B probably meant to convey “No”. • (I really can’t tell whether B meant to convey “Yes” or “No”.) Figure 1: Design of the Mechanical Turk experiment. B: I think it shows how wise the American people are. Table 2 shows the mean entropy values for the different categories identified in the corpus. Interestingly, the pairs involving an adverbial modification in the answer all received a positive answer (‘yes’ or ‘probable yes’) as dominant response. All 19 dialogues involving a numerical measure had either ‘probable yes’ or ‘uncertain’ as dominant response. There is thus a significant bias for positive answers: 70% of the category I items and 74% of the category II items have a positive answer as dominant response. Examining a subset of the Dialog Act corpus, we found that 38% of the yes/no questions receive a direct positive answers, whereas 21% have a direct negative answer. This bias probably stems from the fact that people are more likely to use an overt denial expression where they need to disagree, whether or not they are responding indirectly. 4 Methods In this section, we present the methods we propose for grounding the meanings of scalar modifiers. 4.1 Learning modifier scales and inferring yes/no answers The first technique targets items such as the ones in category I of our corpus. Our central hypothesis is that, in polar question dialogues, the semantic relationship between the main predication PQ in the question and the main predication PA in the answer is the primary factor in determining whether, and to what extent, ‘yes’ or ‘no’ was intended. If PA is at least as strong as PQ, the intended answer is ‘yes’; if PA is weaker than PQ, the intended answer is ‘no’; and, where no reliable entailment relationship exists between PA and PQ, the result is uncertainty. For example, good is weaker (lower on the relevant scale) than excellent, and thus speakers infer that the reply in example (1) above is meant to convey ‘yes’. In contrast, if we reverse the order of the modifiers — roughly, Is it a great idea?; It’s a good idea — then speakers infer that the answer conveys ‘no’. Had B replied with It’s a complicated idea in (1), then uncertainty would likely have resulted, since good and complicated are not in a reliable scalar relationship. Negation reverses scales (Horn, 1972; Levinson, 2000), so it flips ‘yes’ and ‘no’ in these cases, leaving ‘uncertain’ unchanged. When both the question and the answer contain a modifier (such as in (9–11)), the yes/no response should correlate with the extent to which the pair of modifiers can be put into a scale based on contextual entailment. To ground such scales from text, we collected a large corpus of online reviews from IMDB. Each of the reviews in this collection has an associated star rating: one star (most negative) to ten stars (most positive). Table 3 summarizes the distribution of reviews as well as the number of words and vocabulary across the ten rating categories. As is evident from table 3, there is a significant bias for ten-star reviews. This is a common feature of such corpora of informal, userprovided reviews (Chevalier and Mayzlin, 2006; Hu et al., 2006; Pang and Lee, 2008). However, since we do not want to incorporate the linguistically uninteresting fact that people tend to write a lot of ten-star reviews, we assume uniform priors for the rating categories. Let count(w, r) be the number of tokens of word w in reviews in rating category r, and let count(r) be the total word count for all words in rating category r. The probability of w given a rating category r is simply Pr(w|r) = count(w, r)/ count(r). Then under the assumption of uniform priors, we get Pr(r|w) = Pr(w|r)/ P r′∈R Pr(w|r′). In reasoning about our dialogues, we rescale the rating categories by subtracting 5.5 from each, to center them at 0. This yields the scale R = 170 Rating Reviews Words Vocabulary Average words per review 1 124,587 25,389,211 192,348 203.79 2 51,390 11,750,820 133,283 228.66 3 58,051 13,990,519 148,530 241.00 4 59,781 14,958,477 156,564 250.22 5 80,487 20,382,805 188,461 253.24 6 106,145 27,408,662 225,165 258.22 7 157,005 40,176,069 282,530 255.89 8 195,378 48,706,843 313,046 249.30 9 170,531 40,264,174 273,266 236.11 10 358,441 73,929,298 381,508 206.25 Total 1,361,796 316,956,878 1,160,072 206.25 Table 3: Numbers of reviews, words and vocabulary size per rating category in the IMDB review corpus, as well as the average number of words per review. enjoyable 0.0 0.1 0.2 0.3 0.4 -4.5 -3.5 -2.5 -1.5 -0.5 0.5 1.5 2.5 3.5 4.5 ER = 0.74 best 0.0 0.1 0.2 0.3 0.4 -4.5 -3.5 -2.5 -1.5 -0.5 0.5 1.5 2.5 3.5 4.5 ER = 1.08 great 0.0 0.1 0.2 0.3 0.4 -4.5 -3.5 -2.5 -1.5 -0.5 0.5 1.5 2.5 3.5 4.5 ER = 1.1 superb 0.0 0.1 0.2 0.3 0.4 -4.5 -3.5 -2.5 -1.5 -0.5 0.5 1.5 2.5 3.5 4.5 ER = 2.18 disappointing 0.0 0.1 0.2 0.3 0.4 -4.5 -3.5 -2.5 -1.5 -0.5 0.5 1.5 2.5 3.5 4.5 ER = -1.1 bad 0.0 0.1 0.2 0.3 0.4 -4.5 -3.5 -2.5 -1.5 -0.5 0.5 1.5 2.5 3.5 4.5 ER = -1.47 awful 0.0 0.1 0.2 0.3 0.4 -4.5 -3.5 -2.5 -1.5 -0.5 0.5 1.5 2.5 3.5 4.5 ER = -2.5 worst 0.0 0.1 0.2 0.3 0.4 -4.5 -3.5 -2.5 -1.5 -0.5 0.5 1.5 2.5 3.5 4.5 ER = -2.56 Pr(Rating|Word) Rating (centered at 0) Figure 2: The distribution of some scalar modifiers across the ten rating categories. The vertical lines mark the expected ratings, defined as a weighted sum of the probability values (black dots). ⟨−4.5, −3.5, −2.5, −1.5, −0.5, 0.5, 1.5, 2.5, 3.5, 4.5⟩. Our rationale for this is that modifiers at the negative end of the scale (bad, awful, terrible) are not linguistically comparable to those at the positive end of the scale (good, excellent, superb). Each group forms its own qualitatively different scale (Kennedy and McNally, 2005). Rescaling allows us to make a basic positive vs. negative distinction. Once we have done that, an increase in absolute value is an increase in strength. In our experiments, we use expected rating values to characterize the polarity and strength of modifiers. The expected rating value for a word w is ER(w) = P r∈R r Pr(r|w). Figure 2 plots these values for a number of scalar terms, both positive and negative, across the rescaled ratings, with the vertical lines marking their ER values. The weak scalar modifiers all the way on the left are most common near the middle of the scale, with a slight positive bias in the top row and a slight negative bias in the bottom row. As we move from left to right, the bias for one end of the scale grows more extreme, until the words in question are almost never used outside of the most extreme rating category. The resulting scales correspond well with linguistic intuitions and thus provide an initial indication that the rating categories are a reliable guide to strength and polarity for scalar modifiers. We put this information to use in our dialogue corpus via the decision procedure 171 Let D be a dialogue consisting of (i) a polar question whose main predication is based on scalar predicate PQ and (ii) an indirect answer whose main predication is based on scalar predicate PA. Then: 1. if PA or PQ is missing from our data, infer ‘Uncertain’; 2. else if ER(PQ) and ER(PA) have different signs, infer ‘No’; 3. else if abs(ER(PQ)) ⩽abs(ER(PA)), infer ‘Yes’; 4. else infer ‘No’. 5. In the presence of negation, map ‘Yes’ to ‘No’, ‘No’ to ‘Yes’, and ‘Uncertain’ to ‘Uncertain’. Figure 3: Decision procedure for using the word frequencies across rating categories in the review corpus to decide what a given answer conveys. described in figure 3. 4.2 Interpreting numerical answers The second technique aims at determining whether a numerical answer counts as a positive or negative instance of the adjective in the question (category II in our corpus). Adjectives that can receive a conventional unit of measure, such as little or long, inherently possess a degree of vagueness (Kamp and Partee, 1995; Kennedy, 2007): while in the extreme cases, judgments are strong (e.g., a six foot tall woman can clearly be called “a tall woman” whereas a five foot tall woman cannot), there are borderline cases for which it is difficult to say whether the adjectival predication can truthfully be ascribed to them. A logistic regression model can capture these facts. To build this model, we gather distributional information from the Web. For instance, in the case of (3), we can retrieve from the Web positive and negative examples of age in relation to the adjective and the modified entity “little kids”. The question contains the adjective and the modified entity. The reply contains the unit of measure (here “year-old”) and the numerical answer. Specifically we query the Web using Yahoo! BOSS (Academic) for “little kids” yearold (positive instances) as well as for “not little kids” year-old (negative instances). Yahoo! BOSS is an open search services platform that provides a query API for Yahoo! Web search. We then extract ages from the positive and negative snippets obtained, and we fit a logistic regression to these data. To remove noise, we discard low counts (positive and negative instances for a given unit < 5). Also, for some adjectives, such as little or young, there is an inherent ambiguity between absolute and relative uses. Ideally, a word sense disambiguation system would be used to filter these cases. For now, we extract the largest contiguous range for which the data counts are over the noise threshold.3 When not enough data is retrieved for the negative examples, we expand the query by moving the negation outside the search phrase. We also replace the negation and the adjective by the antonyms given in WordNet (using the first sense). The logistic regression thus has only one factor — the unit of measure (age in the case of little kids). For a given answer, the model assigns a probability indicating the extent to which the adjectival property applies to that answer. If the factor is a significant predictor, we can use the probabilities from the model to decide whether the answer qualifies as a positive or negative instance of the adjective in the question, and thus interpret the indirect response as a ‘yes’ or a ‘no’. The probabilistic nature of this technique adheres perfectly to the fact that indirect answers are intimately tied up with uncertainty. 5 Evaluation and results Our primary goal is to evaluate how well we can learn the relevant scalar and entailment relationships from the Web. In the evaluation, we thus applied our techniques to a manually coded corpus version. For the adjectival scales, we annotated each example for its main predication (modifier, or adverb–modifier bigram), including whether that predication was negated. For the numerical cases, we manually constructed the initial queries: we identified the adjective and the modified entity in the question, and the unit of measure in the answer. However, we believe that identifying the requisite predications and recognizing the presence of negation or embedding could be done automatically using dependency graphs.4 3Otherwise, our model is ruined by references to “young 80-year olds”, using the relative sense of young, which are moderately frequent on the Web. 4As a test, we transformed our corpus into the Stanford dependency representation (de Marneffe et al., 2006), using the Stanford parser (Klein and Manning, 2003) and were able to automatically retrieve all negated modifier predications, except one (We had a view of it, not a particularly good one), 172 Modification in answer Precision Recall I Other adjective 60 60 Adverb - same adjective 95 95 Negation - same adjective 100 100 Omitted adjective 100 100 II Numerical 89 40 Total 75 71 Table 4: Summary of precision and recall (%) by type. Response Precision Recall F1 I Yes 87 76 81 No 57 71 63 II Yes 100 36 53 Uncertain 67 40 50 Table 5: Precision, recall, and F1 (%) per response category. In the case of the scalar modifiers experiment, there were just two examples whose dominant response from the Turkers was ‘Uncertain’, so we have left that category out of the results. In the case of the numerical experiment, there were not any ‘No’ answers. To evaluate the techniques, we pool the Mechanical Turk ‘definite yes’ and ‘probable yes’ categories into a single category ‘Yes’, and we do the same for ‘definite no’ and ‘probable no’. Together with ‘uncertain’, this makes for threeresponse categories. We count an inference as successful if it matches the dominant Turker response category. To use the three-response scheme in the numerical experiment, we simply categorize the probabilities as follows: 0–0.33 = ‘No’, 0.33–0.66 = ‘Uncertain’, 0.66–1.00 = ‘Yes’. Table 4 gives a breakdown of our system’s performance on the various category subtypes. The overall accuracy level is 71% (159 out of 224 inferences correct). Table 5 summarizes the results per response category, for the examples in which both the question and answer contain a gradable modifier (category I), and for the numerical cases (category II). 6 Analysis and discussion Performance is extremely good on the “Adverb – same adjective” and “Negation – same adjective” cases because the ‘Yes’ answer is fairly direct for them (though adverbs like basically introduce an interesting level of uncertainty). The results are because of a parse error which led to wrong dependencies. Response Precision Recall F1 WordNet-based Yes 82 83 82.5 (items I) No 60 56 58 Table 6: Precision, recall, and F1 (%) per response category for the WordNet-based approach. somewhat mixed for the “Other adjective” category. Inferring the relation between scalar adjectives has some connection with work in sentiment detection. Even though most of the research in that domain focuses on the orientation of one term using seed sets, techniques which provide the orientation strength could be used to infer a scalar relation between adjectives. For instance, BlairGoldensohn et al. (2008) use WordNet to develop sentiment lexicons in which each word has a positive or negative value associated with it, representing its strength. The algorithm begins with seed sets of positive, negative, and neutral terms, and then uses the synonym and antonym structure of WordNet to expand those initial sets and refine the relative strength values. Using our own seed sets, we built a lexicon using Blair-Goldensohn et al. (2008)’s method and applied it as in figure 3 (changing the ER values to sentiment scores). Both approaches achieve similar results: for the “Other adjective” category, the WordNet-based approach yields 56% accuracy, which is not significantly different from our performance (60%); for the other types in category I, there is no difference in results between the two methods. Table 6 summarizes the results per response category for the WordNet-based approach (which can thus be compared to the category I results in table 5). However in contrast to the WordNet-based approach, we require no hand-built resources: the synonym and antonym structures, as well as the strength values, are learned from Web data alone. In addition, the WordNet-based approach must be supplemented with a separate method for the numerical cases. In the “Other adjective” category, 31 items involve oppositional terms: canonical antonyms (e.g., right/wrong, good/bad) as well as terms that are “statistically oppositional” (e.g., ready/ premature, true/preposterous, confident/nervous). “Statistically oppositional” terms are not oppositional by definition, but as a matter of contingent fact. Our technique accurately deals with most 173 0 10 20 30 40 50 60 0.0 0.2 0.4 0.6 0.8 little kids Age Probability of being "little" 0 10 20 30 40 50 60 0.2 0.4 0.6 0.8 young kids Age Probability of being "young" 0 20 40 60 80 100 120 0.3 0.4 0.5 0.6 0.7 0.8 warm weather Degree Probability of being "warm" Figure 4: Probabilities of being appropriately described as “little”, “young” or “warm”, fitted on data retrieved when querying the Web for “little kids”, “young kids” and “warm weather”. of the canonical antonyms, and also finds some contingent oppositions (qualified/young, wise/ neurotic) that are lacking in antonymy resources or automatically generated antonymy lists (Mohammad et al., 2008). Out of these 31 items, our technique correctly marks 18, whereas Mohammad et al.’s list of antonyms only contains 5 and BlairGoldensohn et al. (2008)’s technique finds 11. Our technique is solely based on unigrams, and could be improved by adding context: making use of dependency information, as well as moving beyond unigrams. In the numerical cases, precision is high but recall is low. For roughly half of the items, not enough negative instances can be gathered from the Web and the model lacks predictive power (as for items (4) or (12)). (12) A: Do you happen to be working for a large firm? B: It’s about three hundred and fifty people. Looking at the negative hits for item (12), one sees that few give an indication about the number of people in the firm, but rather qualifications about colleagues or employees (great people, people’s productivity), or the hits are less relevant: “Most of the people I talked to were actually pretty optimistic. They were rosy on the job market and many had jobs, although most were not large firm jobs”. The lack of data comes from the fact that the queries are very specific, since the adjective depends on the product (e.g., “expensive exercise bike”, “deep pond”). However when we do get a predictive model, the probabilities correEntropy of response distribution Probability of correct inference by our system 0.0 0.5 1.0 1.5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Figure 5: Correlation between agreement among Turkers and whether the system gets the correct answer. For each dialogue, we plot a circle at Turker response entropy and either 1 = correct inference or 0 = incorrect inference, except the points are jittered a little vertically to show where the mass of data lies. As the entropy rises (i.e., as agreement levels fall), the system’s inferences become less accurate. The fitted logistic regression model (black line) has a statistically significant coefficient for response entropy (p < 0.001). 174 late almost perfectly with the Turkers’ responses. This happens for 8 items: “expensive to call (50 cents a minute)”, “little kids (7 and 10 year-old)”, “long growing season (3 months)”, “lot of land (80 acres)”, “warm weather (80 degrees)”, “young kids (5 and 2 year-old)”, “young person (31 yearold)” and “large house (2450 square feet)”. In the latter case only, the system output (uncertain) doesn’t correlate with the Turkers’ judgment (where the dominant answer is ‘probable yes’ with 15 responses, and 11 answers are ‘uncertain’). The logistic curves in figure 4 capture nicely the intuitions that people have about the relation between age and “little kids” or “young kids”, as well as between Fahrenheit degrees and “warm weather”. For “little kids”, the probabilities of being little or not are clear-cut for ages below 7 and above 15, but there is a region of vagueness in between. In the case of “young kids”, the probabilities drop less quickly with age increasing (an 18 year-old can indeed still be qualified as a “young kid”). In sum, when the data is available, this method delivers models which fit humans’ intuitions about the relation between numerical measure and adjective, and can handle pragmatic inference. If we restrict attention to the 66 examples on which the Turkers completely agreed about which of these three categories was intended (again pooling ‘probable’ and ‘definite’), then the percentage of correct inferences rises to 89% (59 correct inferences). Figure 5 plots the relationship between the response entropy and the accuracy of our decision procedure, along with a fitted logistic regression model using response entropy to predict whether our system’s inference was correct. The handful of empirical points in the lower left of the figure show cases of high agreement between Turkers but incorrect inference from the system. The few points in the upper right indicate low agreement between Turkers and correct inference from the system. Three of the high-agreement/incorrect-inference cases involve the adjectives right–correct. For lowagreement/correct-inference, the disparity could trace to context dependency: the ordering is clear in the context of product reviews, but unclear in a television interview. The analysis suggests that overall agreement is positively correlated with our system’s chances of making a correct inference: our system’s accuracy drops as human agreement levels drop. 7 Conclusion We set out to find techniques for grounding basic meanings from text and enriching those meanings based on information from the immediate linguistic context. We focus on gradable modifiers, seeking to learn scalar relationships between their meanings and to obtain an empirically grounded, probabilistic understanding of the clear and fuzzy cases that they often give rise to (Kamp and Partee, 1995). We show that it is possible to learn the requisite scales between modifiers using review corpora, and to use that knowledge to drive inference in indirect responses. When the relation in question is not too specific, we show that it is also possible to learn the strength of the relation between an adjective and a numerical measure. Acknowledgments This paper is based on work funded in part by ONR award N00014-10-1-0109 and ARO MURI award 548106, as well as by the Air Force Research Laboratory (AFRL) under prime contract no. FA8750-09-C-0181. Any opinions, findings, and conclusion or recommendations expressed in this material are those of the authors and do not necessarily reflect the view of the Air Force Research Laboratory (AFRL), ARO or ONR. References James F. Allen and C. Raymond Perrault. 1980. Analyzing intention in utterances. Artificial Intelligence, 15:143–178. Nicholas Asher and Alex Lascarides. 2003. Logics of Conversation. Cambridge University Press, Cambridge. Sasha Blair-Goldensohn, Kerry Hannan, Ryan McDonald, Tyler Neylon, George A. Reis, and JeffReynar. 2008. Building a sentiment summarizer for local service reviews. In WWW Workshop on NLP in the Information Explosion Era (NLPIX). Judith A. Chevalier and Dina Mayzlin. 2006. The effect of word of mouth on sales: Online book reviews. Journal of Marketing Research, 43(3):345– 354. Herbert H. Clark. 1979. Responding to indirect speech acts. Cognitive Psychology, 11:430–477. Marie-Catherine de Marneffe, Bill MacCartney, and Christopher D. Manning. 2006. Generating typed 175 dependency parses from phrase structure parses. In Proceedings of the 5th International Conference on Language Resources and Evaluation (LREC-2006). Marie-Catherine de Marneffe, Scott Grimm, and Christopher Potts. 2009. Not a simple ‘yes’ or ‘no’: Uncertainty in indirect answers. In Proceedings of the 10th Annual SIGDIAL Meeting on Discourse and Dialogue. Gilles Fauconnier. 1975. Pragmatic scales and logical structure. Linguistic Inquiry, 6(3):353–375. Nancy Green and Sandra Carberry. 1994. A hybrid reasoning model for indirect answers. In Proceedings of the 32nd Annual Meeting of the Association for Computational Linguistics, pages 58–65. Nancy Green and Sandra Carberry. 1999. Interpreting and generating indirect answers. Computational Linguistics, 25(3):389–435. Beth Ann Hockey, Deborah Rossen-Knill, Beverly Spejewski, Matthew Stone, and Stephen Isard. 1997. Can you predict answers to Y/N questions? Yes, No and Stuff. In Proceedings of Eurospeech 1997, pages 2267–2270. Laurence R Horn. 1972. On the Semantic Properties of Logical Operators in English. Ph.D. thesis, UCLA, Los Angeles. Nan Hu, Paul A. Pavlou, and Jennifer Zhang. 2006. Can online reviews reveal a product’s true quality?: Empirical findings and analytical modeling of online word-of-mouth communication. In Proceedings of Electronic Commerce (EC), pages 324–330. Daniel Jurafsky, Elizabeth Shriberg, and Debra Biasca. 1997. Switchboard SWBD-DAMSL shallowdiscourse-function annotation coders manual, draft 13. Technical Report 97-02, University of Colorado, Boulder Institute of Cognitive Science. Hans Kamp and Barbara H. Partee. 1995. Prototype theory and compositionality. Cognition, 57(2):129– 191. Christopher Kennedy and Louise McNally. 2005. Scale structure and the semantic typology of gradable predicates. Language, 81(2):345–381. Christopher Kennedy. 2007. Vagueness and grammar: The semantics of relative and absolute gradable adjectives. Linguistics and Philosophy, 30(1):1–45. Dan Klein and Christopher D. Manning. 2003. Accurate unlexicalized parsing. In Proceedings of the 41st Meeting of the Association of Computational Linguistics. Stephen C. Levinson. 2000. Presumptive Meanings: The Theory of Generalized Conversational Implicature. MIT Press, Cambridge, MA. Saif Mohammad, Bonnie Dorr, and Graeme Hirst. 2008. Computing word-pair antonymy. In Proceedings of the Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-2008). Robert Munro, Steven Bethard, Victor Kuperman, Vicky Tzuyin Lai, Robin Melnick, Christopher Potts, Tyler Schnoebelen, and Harry Tily. 2010. Crowdsourcing and language studies: The new generation of linguistic data. In NAACL 2010 Workshop on Creating Speech and Language Data With Amazon’s Mechanical Turk. Bo Pang and Lillian Lee. 2008. Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval, 2(1):1–135. C. Raymond Perrault and James F. Allen. 1980. A plan-based analysis of indirect speech acts. American Journal of Computational Linguistics, 6(34):167–182. Victor S. Sheng, Foster Provost, and Panagiotis G. Ipeirotis. 2008. Get another label? improving data quality and data mining using multiple, noisy labelers. In Proceedings of KDD-2008. Rion Snow, Brendan O’Connor, Daniel Jurafsky, and Andrew Y. Ng. 2008. Cheap and fast – but is it good? evaluating non-expert annotations for natural language tasks. In Proceedings of the Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-2008). Henk Zeevat. 1994. Questions and exhaustivity in update semantics. In Harry Bunt, Reinhard Muskens, and Gerrit Rentier, editors, Proceedings of the International Workshop on Computational Semantics, pages 211–221. 176
|
2010
|
18
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 177–185, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Importance-Driven Turn-Bidding for Spoken Dialogue Systems Ethan O. Selfridge and Peter A. Heeman Center for Spoken Language Understanding Oregon Health & Science University 20000 NW Walker Rd., Beaverton, OR, 97006 [email protected], [email protected] Abstract Current turn-taking approaches for spoken dialogue systems rely on the speaker releasing the turn before the other can take it. This reliance results in restricted interactions that can lead to inefficient dialogues. In this paper we present a model we refer to as Importance-Driven Turn-Bidding that treats turn-taking as a negotiative process. Each conversant bids for the turn based on the importance of the intended utterance, and Reinforcement Learning is used to indirectly learn this parameter. We find that Importance-Driven Turn-Bidding performs better than two current turntaking approaches in an artificial collaborative slot-filling domain. The negotiative nature of this model creates efficient dialogues, and supports the improvement of mixed-initiative interaction. 1 Introduction As spoken dialogue systems are designed to perform ever more elaborate tasks, the need for mixed-initiative interaction necessarily grows. Mixed-initiative interaction, where agents (both artificial and human) may freely contribute to reach a solution efficiently, has long been a focus of dialogue systems research (Allen et al., 1999; Guinn, 1996). Simple slot-filling tasks might not require the flexible environment that mixedinitiative interaction brings but those of greater complexity, such as collaborative task completion or long-term planning, certainly do (Ferguson et al., 1996). However, translating this interaction into working systems has proved problematic (Walker et al., 1997), in part to issues surrounding turn-taking: the transition from one speaker to another. Many computational turn-taking approaches seek to minimize silence and utterance overlap during transitions. This leads to the speaker controlling the turn transition. For example, systems using the Keep-Or-Release approach will not attempt to take the turn unless it is sure the user has released it. One problem with this approach is that the system might have important information to give but will be unable to get the turn. The speaker-centric nature of current approaches does not enable mixed-initiative interaction and results in inefficient dialogues. Primarily, these approaches have been motivated by smooth transitions reported in the human turn-taking studies of Sacks et al. (1974) among others. Sacks et al. also acknowledge the negotiative nature of turn-taking, stating that the “the turn as unit is interactively determined”(p. 727). Other studies have supported this, suggesting that humans negotiate the turn assignment through the use of cues and that these cues are motivated by the importance of what the conversant wishes to contribute (Duncan and Niederehe, 1974; Yang and Heeman, 2010; Schegloff, 2000). Given this, any dialogue system hoping to interact with humans efficiently and naturally should have a negotiative and importance-driven quality to its turn-taking protocol. We believe that, by focusing on the rationale of human turn-taking behavior, a more effective turn-taking system may be achieved. We propose the Importance-Driven Turn-Bidding (IDTB) model where conversants bid for the turn based on the importance of their utterance. We use Reinforcement Learning to map a given situation to the optimal utterance and bidding behavior. By allowing conversants to bid for the turn, the IDTB model enables negotiative turntaking and supports true mixed-initiative interaction, and with it, greater dialogue efficiency. We compare the IDTB model to current turntaking approaches. Using an artificial collaborative dialogue task, we show that the IDTB model enables the system and user to complete 177 the task more efficiently than the other approaches. Though artificial dialogues are not ideal, they allow us to test the validity of the IDTB model before embarking on costly and time-consuming human studies. Since our primary evaluation criteria is model comparison, consistent user simulations provide a constant needed for such measures and increase the external validity of our results. 2 Current Turn-Taking Approaches Current dialogue systems focus on the release-turn as the most important aspect of turn-taking, in which a listener will only take the turn after the speaker has released it. The simplest of these approaches only allows a single utterance per turn, after which the turn necessarily transitions to the next speaker. This Single-Utterance (SU) model has been extended to allow the speaker to keep the turn for multiple utterances: the Keep-Or-Release (KR) approach. Since the KR approach gives the speaker sole control of the turn, it is overwhelmingly speaker-centric, and so necessarily unnegotiative. This restriction is meant to encourage smooth turn-transitions, and is inspired by the order, smoothness, and predictability reported in human turn-taking studies (Duncan, 1972; Sacks et al., 1974). Systems using the KR approach differ on how they detect the user’s release-turn. Turn releases are commonly identified in two ways: either using a silence-threshold (Sutton et al., 1996), or the predictive nature of turn endings (Sacks et al., 1974) and the cues associated with them (e.g. Gravano and Hirschberg, 2009). Raux and Eskenazi (2009) used decision theory with lexical cues to predict appropriate places to take the turn. Similarly, Jonsdottir, Thorisson, and Nivel (2008) used Reinforcement Learning to reduce silences between turns and minimize overlap between utterances by learning the specific turn-taking patterns of individual speakers. Skantze and Schlangan (2009) used incremental processing of speech and prosodic turn-cues to reduce the reaction time of the system, finding that that users rated this approach as more human-like than a baseline system. In our view, systems built using the KR turntaking approach suffer from two deficits. First, the speaker-centricity leads to inefficient dialogues since the speaker may continue to hold the turn even when the listener has vital information to give. In addition, the lack of negotiation forces the turn to necessarily transition to the listener after the speaker releases it. The possibility that the dialogue may be better served if the listener does not get the turn is not addressed by current approaches. Barge-in, which generally refers to allowing users to speak at any time (Str¨om and Seneff, 2000), has been the primary means to create a more flexible turn-taking environment. Yet, since barge-in recasts speaker-centric systems as usercentric, the system’s contributions continue to be limited. System barge-in has also been investigated. Sato et al. (2002) used decision trees to determine whether the system should take the turn or not when the user pauses. An incremental method by DeVault, Sagae, and Traum (2009) found possible points that a system could interrupt without loss of user meaning, but failed to supply a reasonable model as to when to use such information. Despite these advances, barge-in capable systems lack a negotiative turn-taking method, and continue to be deficient for reasons similar to those described above. 3 Importance-Driven Turn-Bidding (IDTB) We introduce the IDTB model to overcome the deficiencies of current approaches. The IDTB model has two foundational components: (1) The importance of speaking is the primary motivation behind turn-taking behavior, and (2) conversants use turncue strength to bid for the turn based on this importance. Importance may be broadly defined as how well the utterance leads to some predetermined conversational success, be it solely task completion or encompassing a myriad of social etiquette components. Importance-Driven Turn-Bidding is motivated by empirical studies of human turn-conflict resolution. Yang and Heeman (2010) found an increase of turn conflicts during tighter time constraints, which suggests that turn-taking is influenced by the importance of task completion. Schlegoff (2000) proposed that persistent utterance overlap was indicative of conversants having a strong interest in holding the turn. Walker and Whittaker (1990) show that people will interrupt to remedy some understanding discrepancy, which is certainly important to the conversation’s success. People communicate the importance of their utterance through turn-cues. Duncan and 178 Niederehe (1974) found that turn-cue strength was the best predictor of who won the turn, and this finding is consistent with the use of volume to win turns found by Yang and Heeman (2010). The IDTB model uses turn-cue strength to bid for the turn based on the importance of the utterance. Stronger turn-cues should be used when the intended utterance is important to the overall success of the dialogue, and weaker ones when it is not. In the prototype described in Section 5, both the system and user agents bid for the turn after every utterance and the bids are conceptualized here as utterance onset: conversants should be quick to speak important utterances but slow with less important ones. This is relatively consistent with Yang and Heeman (2010). A mature version of our work will use cues in addition to utterance onset, such as those recently detailed in Gravano and Hirshberg (2009).1 A crucial element of our model is the judgment and quantization of utterance importance. We use Reinforcement Learning (RL) to determine importance by conceptualizing it as maximizing the reward over an entire dialogue. Whatever actions lead to a higher return may be thought of as more important than ones that do not.2 By using RL to learn both the utterance and bid behavior, the system can find an optimal pairing between them, and choose the best combination for a given conversational situation. 4 Information State Update and Reinforcement Learning We build our dialogue system using the Information State Update approach (Larsson and Traum, 2000) and use Reinforcement Learning for action selection (Sutton and Barto, 1998). The system architecture consists of an Information State (IS) that represents the agent’s knowledge and is updated using a variety of rules. The IS also uses rules to propose possible actions. A condensed and compressed subset of the IS — the Reinforcement Learning State — is used to learn which proposed action to take (Heeman, 2007). It has been shown that using RL to learn dialogue polices is generally more effective than “hand crafted” di1Our work (present and future) is distinct from some recent work on user pauses (Sato et al., 2002) since we treat turn-taking as an integral piece of dialogue success. 2We gain an inherent flexibility in using RL since the reward can be computed by a wide array of components. This is consistent with the broad definition of importance. alogue policies since the learning algorithm may capture environmental dynamics that are unattended to by human designers (Levin et al., 2000). Reinforcement Learning learns an optimal policy, a mapping between a state s and action a, where performing a in s leads to the lowest expected cost for the dialogue (we use minimum cost instead of maximum reward). An ϵ-greedy search is used to estimate Q-scores, the expected cost of some state–action pair, where the system chooses a random action with ϵ probability and the argminaQ(s, a) action with 1-ϵ probability. For Q-learning, a popular RL algorithm and the one used here, ϵ is commonly set at 0.2 (Sutton and Barto, 1998). Q-learning updates Q(s, a) based on the best action of the next state, given by the following equation, with the step size parameter α = 1/ p N(s, a) where N(s, a) is the number of times the s, a pair has been seen since the beginning of training. Q(st, at) = Q(st, at) + α[costt+1 + argminaQ(st+1, a) −Q(st, at)] The state space should be formulated as a Markov Decision Process (MDP) for Q-learning to update Q-scores properly. An MDP relies on a first-order Markov assumption in that the transition and reward probability from some st, at pair is completely contained by that pair and is unaffected by the history st−1at−1, st−2at−2, . . .. For this assumption to be met, care is required when deciding which features to include for learning. The RL State features we use are described in the following section. 5 Domain and Turn-Taking Models In this section, we show how the IDTB approach can be implemented for a collaborative slot filling domain. We also describe the SingleUtterance and Keep-Or-Release domain implementations that we use for comparison. 5.1 Domain Task We use a food ordering domain with two participants, the system and a user, and three slots: drink, burger, and side. The system’s objective is to fill all three slots with the available fillers as quickly as possible. The user’s role is to specify its desired filler for each slot, though that specific filler may not be available. The user simulation, while intended to be realistic, is not based on empirical data. Rather, it is designed to provide a rich turn179 taking domain to evaluate the performance of different turn-taking designs. We consider this a collaborative slot-filling task since both conversants must supply information to determine the intersection of available and desired fillers. Users have two fillers for each slot.3 A user’s top choice is either available, in which case we say that the user has adequate filler knowledge, or their second choice will be available, in which we say it has inadequate filler knowledge. This assures that at least one of the user’s filler is available. Whether a user has adequate or inadequate filler knowledge is probabilistically determined based on user type, which will be described in Section 5.2. Table 1: Agent speech acts Agent Actions System query slot, inform [yes/no], inform avail. slot fillers, inform filler not available, bye User inform slot filler, query filler availability We model conversations at the speech act level, shown in Table 1, and so do not model the actual words that the user and system might say. Each agent has an Information State that proposes possible actions. The IS is made up of a number of variables that model the environment and is slightly different for the system and the user. Shared variables include QUD, a stack which manages the questions under discussion; lastUtterance, the previous utterance, and slotList, a list of the slot names. The major system specific IS variables that are not included in the RL State are availSlotFillers, the available fillers for each slot; and three slotFiller variables that hold the fillers given by the user. The major user specific IS variables are three desiredSlotFiller variables that hold an ordered list of fillers, and unvisitedSlots, a list of slots that the user believes are unfilled. The system has a variety of speech actions: inform [yes/no], to answer when the user has asked a filler availability question; inform filler not available, to inform the user when they have specified an unavailable filler; three query slot actions (one for each slot), a query which asks the user for a filler and is proposed if that specific slot is unfilled; 3We use two fillers so as to minimize the length of training. This can be increased without substantial effort. three inform available slot fillers actions, which lists the available fillers for that slot and is proposed if that specific slot is unfilled or filled with an unavailable filler; and bye, which is always proposed. The user has two actions. They can inform the system of a desired slot filler, inform slot filler, or query the availability of a slot’s top filler, query filler availability. A user will always respond with the same slot as a system query, but may change slots entirely for all other situations. Additional details on user action selection are given in Section 5.2. Specific information is used to produce an instantiated speech action, what we refer to as an utterance. For example, the speech action inform slot filler results in the utterance of ”inform drink d1.” A sample dialogue fragment using the SingleUtterance approach is shown in Table 2. Notice that in Line 3 the system informs the user that their first filler, d1, is unavailable. The user then asks asks about the availability of its second drink choice, d2 (Line 4), and upon receiving an affirmative response (Line 5), informs the system of that filler preference (Line 6). Table 2: Single-Utterance dialogue Spkr Speech Action Utterance 1 S: q. slot q. drink 2 U: i. slot filler i. drink d1 3 S: i. filler not avail i. not have d1 4 U: q. filler avail q. drink have d2 5 S: i. slot i. yes 6 U: i. slot filler i. drink d2 7 S: i. avail slot fillers i. burger have b1 Implementation in RL: The system uses RL to learn which of the IS proposed actions to take. In this domain we use a cost function based on dialogue length and the number of slots filled with an available filler: C = Number of Utterances + 25 · unavailablyFilledSlots. In the present implementation the system’s bye utterance is costless. The system chooses the action that minimizes the expected cost of the entire dialogue from the current state. The RL state for the speaker has seven variables:4 QUD-speaker, the stack of speakers who have unresolved questions; Incorrect-Slot-Fillers, 4We experimented with a variety of RL States and this one proved to be both small and effective. 180 a list of slot fillers (ordered chronologically on when the user informed them) that are unavailable and have not been resolved; Last-Sys-SpeechAction, the last speech action the system performed; Given-Slot-Fillers, a list of slots that the system has performed the inform available slot filler action on; and three booleans variables, slotRL, that specify whether a slot has been filled correctly or not (e.g. Drink-RL). 5.2 User Types We define three different types of users — Experts, Novices, and Intermediates. User types differ probabilistically on two dimensions: slot knowledge, and slot belief strength. We define experts to have a 90 percent chance of having adequate filler knowledge, intermediates a 50 percent chance, and novices a 10 percent chance. These probabilities are independent between slots. Slot belief strength represents the user’s confidence that it has adequate domain knowledge for the slot (i.e. the top choice for that slot is available). It is either a strong, warranted, or weak belief (Chu-Carroll and Carberry, 1995). The intuition is that experts should know when their top choice is available, and novices should know that they do not know the domain well. Initial slot belief strength is dependent on user type and whether their filler knowledge is adequate (their initial top choice is available). Experts with adequate filler knowledge have a 70, 20, and 10 percent chance of having Strong, Warranted, and Weak beliefs respectfully. Similarly, intermediates with adequate knowledge have a 50, 25, and 25 percent chance of the respective belief strengths. When these user types have inadequate filler knowledge the probabilities are reversed to determine belief strength (e.g. Experts with inadequate domain knowledge for a slot have a 70% chance of having a weak belief). Novice users always have a 10, 10, and 80 percent chance of the respective belief strengths. The user choses whether to use the query or inform speech action based on the slot’s belief strength. A strong belief will always result in an inform, a warranted belief resulting in an inform with p = 0.5, and weak belief will result in an inform with p = 0.25. If the user is informed of the correct fillers by the system’s inform, that slot’s belief strength is set to strong. If the user is informed that a filler is not available, than that filler is removed from the desired filler list and the belief remains the same.5 5.3 Turn-Taking Models We now discuss how turn-taking works for the IDTB model and the two competing models that we use to evaluate our approach. The system chooses its turn action based on the RL state and we add a boolean variable turn-action to the RL State to indicate when the system is performing a turn action or a speech action. The user uses belief to choose its turn action. Turn-Bidding: Agents bid for the turn at the end of each utterance to determine who will speak next. Each bid is represented as a value between 0 and 1, and the agent with the lower value (stronger bid) wins the turn. This is consistent with the use of utterance onset. There are 5 types of bids, highest, high, middle, low, and lowest, which are spread over a portion of the range as shown in Figure 1. The system uses RL to choose a bid and a random number (uniform distribution) is generated from that bid’s range. The users’ bids are determined by their belief strength, which specifies the mean of a Gaussian distribution, as shown in Figure 1 (e.g Strong belief implies a µ = 0.35). Computing bids in this fashion leads to, on average, users with strong beliefs bidding highest, warranted beliefs bidding in the middle, and weak beliefs bidding lowest. The use of the probability distributions allows us to randomly decide ties between system and user bids. Figure 1: Bid Value Probability Distribution Single-Utterance: The Single-Utterance (SU) approach, as described in Section 2, has a rigid 5In this simple domain the next filler is guaranteed to be available if the first is not. We do not model this with belief strength since it is probably not representative of reality. 181 turn-taking mechanism. After a speaker makes a single utterance the turn transitions to the listener. Since the turn transitions after every utterance the system must only choose appropriate utterances, not turn-taking behavior. Similarly, user agents do not have any turn-taking behavior and slot beliefs are only used to choose between a query and an inform. Keep-Or-Release Model: The Keep-OrRelease (KR) model, as described in Section 2, allows the speaker to either keep the turn to make multiple utterances or release it. Taking the same approach as English and Heeman (2005), the system learns to keep or release the turn after each utterance that it makes. We also use RL to determine which conversant should begin the dialogue. While the use of RL imparts some importance onto the turn-taking behavior, it is not influencing whether the system gets the turn when it did not already have it. This is an crucial distinction between KR and IDTB. IDTB allows the conversants to negotiate the turn using turn-bids motivated by importance, whereas in KR only the speaker determines when the turn can transition. Users in the KR environment choose whether to keep or release the turn similarly to bid decisions.6 After a user performs an utterance, it chooses the slot that would be in the next utterance. A number, k, is generated from a Gaussian distribution using belief strength in the same manner as the IDTB users’ bids are chosen. If k ≤0.55 then the user keeps the turn, otherwise it releases it. 5.4 Preliminary Turn-Bidding System We described a preliminary turn-bidding system in earlier work presented at a workshop (Selfridge and Heeman, 2009). A major limitation was an overly simplified user model. We used two user types, expert and novice, who had fixed bids. Experts always bid high and had complete domain knowledge, and the novices always bid low and had incomplete domain knowledge. The system, using all five bid types, was always able to out bid and under bid the simulated users. Among other things, this situation gives the system complete control of the turn, which is at odds with the negotiative nature of IDTB. The present contribution is a more realistic and mature implementation. 6We experimented with a few different KR decision strategies, and chose the one that performed the best. 6 Evaluation and Discussion We now evaluate the IDTB approach by comparing it against the two competing models: SingleUtterance and Keep-Or-Release. The three turntaking approaches are trained and tested in four user conditions: novice, intermediate, expert, and combined. In the combined condition, one of the three user types is randomly selected for each dialogue. We train ten policies for each condition and turn-taking approach. Policies are trained using Qlearning, and ϵ−greedy search for 10000 epochs (1 epoch = 100 dialogues, after which the Q-scores are updated) with ϵ = 0.2. Each policy is then ran over 10000 test dialogues with no exploration (ϵ = 0), and the mean dialogue cost for that policy is determined. The 10 separate policy values are then averaged to create the mean policy cost. The mean policy cost between the turn-taking approaches and user conditions are shown in Table 3. Lower numbers are indicative of shorter dialogues, since the system learns to successfully complete the task in all cases. Table 3: Mean Policy Cost for Model and User condition7 Model Novice Int. Expert Combined SU 7.61 7.09 6.43 7.05 KR 6.00 6.35 4.46 6.01 IDTB 6.09 5.77 4.35 5.52 Single User Conditions: Single user conditions show how well each turn-taking approach can optimize its behavior for specific user populations and handle slight differences found in those populations. Table 3 shows that the mean policy cost of the SU model is higher than the other two models which indicates longer dialogues on average. Since the SU system must respond to every user utterance and cannot learn a turn-taking strategy to utilize user knowledge, the dialogues are necessarily longer. For example, in the expert condition the best possible dialogue for a SU interaction will have a cost of five (three user utterances for each slot, two system utterances in response). This cost is in contrast to the best expert dialogue cost of three (three user utterances) for KR and IDTB interactions. The IDTB turn-taking approach outperforms the KR design in all single user conditions ex7SD between policies ≤0.04 182 cept for novice (6.09 vs. 6.00). In this condition, the KR system takes the turn first, informs the available fillers for each slot, and then releases the turn. The user can then inform its filler easily. The IDTB system attempts a similar dialogue strategy by using highest bids but sometimes loses the turn when users also bid highest. If the user uses the turn to query or inform an unavailable filler the dialogue grows longer. However, this is quite rare as shown by small difference in performance between the two models. In all other single user conditions, the IDTB approach has shorter dialogues than the KR approach (5.77 and 4.35 vs. 6.35 and 4.46). A detailed explanation of IDTB’s performance will be given in Section 6.1. Combined User Condition: We next measure performance on the combined condition that mixes all three user types. This condition is more realistic than the other three, as it better mimics how a system will be used in actual practice. The IDTB approach (mean policy cost = 5.52) outperforms the KR (mean policy cost = 6.01) and SU (mean policy cost = 7.05) approaches. We also observe that KR outperforms SU. These results suggest that the more a turn-taking design can be flexible and negotiative, the more efficient the dialogues can be. Exploiting User bidding differences: It follows that IDTB’s performance stems from its negotiative turn transitions. These transitions are distinctly different than KR transitions in that there is information inherent in the users bids. A user that has a stronger belief strength is more likely to be have a higher bid and inform an available filler. Policy analysis shows that the IDTB system takes advantage of this information by using moderate bids —neither highest nor lowest bids— to filter users based on their turn behavior. The distribution of bids used over the ten learned policies is shown in Table 4. The initial position refers to the first bid of the dialogue; final position, the last bid of the dialogue; and medial position, all other bids. Notice that the system uses either the low or mid bids as its initial policy and that 67.2% of dialogue medial bids are moderate. These distributions show that the system has learned to use the entire bid range to filter the users, and is not seeking to win or lose the turn outright. This behavior is impossible in the KR approach. Table 4: Bid percentages over ten policies in the Combined User condition for IDTB Position H-est High Mid Low L-est Initial 0.0 0.0 70.0 30.0 0.0 Medial 20.5 19.4 24.5 23.3 12.3 Final 49.5 41.0 9.5 0.0 0.0 6.1 IDTB Performance: In our domain, performance is measured by dialogue length and solution quality. However, since solution quality never affects the dialogue cost for a trained system, dialogue length is the only component influencing the mean policy cost. The primary cause of longer dialogues are unavailable filler inform and query (UFI–Q) utterances by the user, which are easily identified. These utterances lengthen the dialogue since the system must inform the user of the available fillers (the user would otherwise not know that the filler was unavailable) and then the user must then inform the system of its second choice. The mean number of UFI–Q utterance for each dialogue over the ten learned policies are shown for all user conditions in Table 5. Notice that these numbers are inversely related to performance: the more UFI– Q utterances, the worse the performance. For example, in the combined condition the IDTB users perform 0.38 UFI–Q utterances per dialogue (u/d) compared to the 0.94 UFI–Q u/d for KR users. While a KR user will release the turn if its planned Table 5: Mean number of UFI–Q utterances over policies Model Novice Int. Expert Combined KR 0.0 1.15 0.53 0.94 IDTB 0.1 0.33 0.39 0.38 utterance has a weak belief, it may select that weak utterance when first getting the turn (either after a system utterance or at the start of the dialogue). This may lead to a UFI–Q utterance. The IDTB system, however, will outbid the same user, resulting in a shorter dialogue. This situation is shown in Tables 6 and 7. The dialogue is the same until utterance 3, where the IDTB system wins the turn with a mid bid over the user’s low bid. In the KR environment however, the user gets the turn and performs an unavailable filler inform, which the system must react to. This is an instance of the second deficiency of the KR approach, where 183 Table 6: Sample IDTB dialogue in Combined User condition; Cost=6 Sys Usr Spkr Utt 1 low mid U: inform burger b1 2 h-est low S: inform burger have b3 3 mid low S: inform side have s1 4 mid h-est U: inform burger b3 5 mid high U: inform drink d1 6 l-est h-est U: inform side s1 7 high mid S: bye Table 7: Sample KR dialogue in Combined User condition; Cost=7 Agent Utt Turn-Action 1 U: inform burger b1 Release 2 S: inform burger have b3 Release 3 U: inform side s1 Keep 4 U: inform drink d1 Keep 5 U: inform burger b3 Release 6 S: inform side have s2 Release 7 U: inform side s2 Release 8 S: bye the speaking system should not have released the turn. The user has the same belief in both scenarios, but the negotiative nature of IDTB enables a shorter dialogues. In short, the IDTB system can win the turn when it should have it, but the KR system cannot. A lesser cause of longer dialogues is an instance of the first deficiency of the KR systems; the listening user cannot get the turn when it should have it. Usually, this situation presents itself when the user releases the turn, having randomly chosen the weaker of the two unfilled slots. The system then has the turn for more than one utterance, informing the available fillers for two slots. However, the user already had a strong belief and available top filler for one of those slots, and the system has increased the dialogue length unnecessarily. In the combined condition, the KR system produces 0.06 unnecessary informs per dialogue, whereas the IDTB system produces 0.045 per dialogue. The novice and intermediate conditions mirror this (IDTB: 0.009, 0.076 ; KR: 0.019, 0.096 respectfully), but the expert condition does not (IDTB: 0.011, KR: 0.0014). In this case, the IDTB system wins the turn initially using a low bid and informs one of the strong slots, whereas the expert user initiates the dialogue for the KR environment and unnecessary informs are rarer. In general, however, the KR approach has more unnecessary informs since the KR system can only infer that one of the user’s beliefs was probably weak, otherwise the user would not have released the turn. The IDTB system handles this situation by using a high bid, allowing the user to outbid the system as its contribution is more important. In other words, the IDTB user can win the turn when it should have it, but the KR user cannot. 7 Conclusion This paper presented the Importance-Driven TurnBidding model of turn-taking. The IDTB model is motivated by turn-conflict studies showing that the interest in holding the turn influences conversant turn-cues. A computational prototype using Reinforcement Learning to choose appropriate turnbids performs better than the standard KR and SU approaches in an artificial collaborative dialogue domain. In short, the Importance-Driven TurnBidding model provides a negotiative turn-taking framework that supports mixed-initiative interactions. In the previous section, we showed that the KR approach is deficient for two reasons: the speaking system might not keep the turn when it should have, and might release the turn when it should not have. This is driven by KR’s speaker-centric nature; the speaker has no way of judging the potential contribution of the listener. The IDTB approach however, due to its negotiative quality, does not have this problem. Our performance differences arise from situations when the system is the speaker and the user is the listener. The IDTB model also excels in the opposite situation, when the system is the listener and the user is the speaker, though our domain is not sophisticated enough for this situation to occur. In the future we hope to develop a domain with more realistic speech acts and a more difficult dialogue task that will, among other things, highlight this situation. We also plan on implementing a fully functional IDTB system, using an incremental processing architecture that not only detects, but generates, a wide array of turn-cues. Acknowledgments We gratefully acknowledge funding from the National Science Foundation under grant IIS0713698. 184 References J.E Allen, C.I. Guinn, and Horvitz E. 1999. Mixedinitiative interaction. IEEE Intelligent Systems, 14(5):14–23. Jennifer Chu-Carroll and Sandra Carberry. 1995. Response generation in collaborative negotiation. In Proceedings of the 33rd annual meeting on Association for Computational Linguistics, pages 136– 143, Morristown, NJ, USA. Association for Computational Linguistics. David DeVault, Kenji Sagae, and David Traum. 2009. Can i finish? learning when to respond to incremental interpretation results in interactive dialogue. In Proceedings of the SIGDIAL 2009 Conference, pages 11–20, London, UK, September. Association for Computational Linguistics. S.J. Duncan and G. Niederehe. 1974. On signalling that it’s your turn to speak. Journal of Experimental Social Psychology, 10:234–247. S.J. Duncan. 1972. Some signals and rules for taking speaking turns in conversations. Journal of Personality and Social Psychology, 23:283–292. M. English and Peter A. Heeman. 2005. Learning mixed initiative dialog strategies by using reinforcement learning on both conversants. In Proceedings of HLT/EMNLP, pages 1011–1018. G. Ferguson, J. Allen, and B. Miller. 1996. TRAINS95: Towards a mixed-initiative planning assistant. In Proceedings of the Third Conference on Artificial Intelligence Planning Systems (AIPS-96), pages 70– 77. A. Gravano and J. Hirschberg. 2009. Turn-yielding cues in task-oriented dialogue. In Proceedings of the SIGDIAL 2009 Conference: The 10th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 253–261. Association for Computational Linguistics. C.I. Guinn. 1996. Mechanisms for mixed-initiative human-computer collaborative discourse. In Proceedings of the 34th annual meeting on Association for Computational Linguistics, pages 278–285. Association for Computational Linguistics. P.A. Heeman. 2007. Combining reinforcement learning with information-state update rules. In Proceedings of the Annual Conference of the North American Association for Computational Linguistics, pages 268–275, Rochester, NY. Gudny Ragna Jonsdottir, Kristinn R. Thorisson, and Eric Nivel. 2008. Learning smooth, human-like turntaking in realtime dialogue. In IVA ’08: Proceedings of the 8th international conference on Intelligent Virtual Agents, pages 162–175, Berlin, Heidelberg. Springer-Verlag. S. Larsson and D. Traum. 2000. Information state and dialogue managment in the trindi dialogue move engine toolkit. Natural Language Engineering, 6:323– 340. E. Levin, R. Pieraccini, and W. Eckert. 2000. A stochastic model of human-machine interaction for learning dialog strategies. IEEE Transactions on Speech and Audio Processing, 8(1):11 – 23. A. Raux and M. Eskenazi. 2009. A finite-state turntaking model for spoken dialog systems. In Proceedings of HLT/NAACL, pages 629–637. Association for Computational Linguistics. H. Sacks, E.A. Schegloff, and G. Jefferson. 1974. A simplest systematics for the organization of turntaking for conversation. Language, 50(4):696–735. R. Sato, R. Higashinaka, M. Tamoto, M. Nakano, and K. Aikawa. 2002. Learning decision trees to determine turn-taking by spoken dialogue systems. In ICSLP, pages 861–864, Denver, CO. E.A. Schegloff. 2000). Overlapping talk and the organization of turn-taking for conversation. Language in Society, 29:1 – 63. E. O. Selfridge and Peter A. Heeman. 2009. A bidding approach to turn-taking. In 1st International Workshop on Spoken Dialogue Systems. G. Skantze and D. Schlangen. 2009. Incremental dialogue processing in a micro-domain. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics, pages 745–753. Association for Computational Linguistics. N. Str¨om and S. Seneff. 2000. Intelligent barge-in in conversational systems. In Sixth International Conference on Spoken Language Processing. Citeseer. R. Sutton and A. Barto. 1998. Reinforcement Learning. MIT Press. S. Sutton, D. Novick, R. Cole, P. Vermeulen, J. de Villiers, J. Schalkwyk, and M. Fanty. 1996. Building 10,000 spoken-dialogue systems. In ICSLP, Philadelphia, Oct. M. Walker and S. Whittaker. 1990. Mixed initiative in dialoge: an investigation into discourse segmentation. In Proceedings of the 28th Annual Meeting of the Association for Computational Linguistics, pages 70–76. M. Walker, D. Hindle, J. Fromer, G.D. Fabbrizio, and C. Mestel. 1997. Evaluating competing agent strategies for a voice email agent. In Fifth European Conference on Speech Communication and Technology. Fan Yang and Peter A. Heeman. 2010. Initiative conflicts in task-oriented dialogue”. Computer Speech Language, 24(2):175 – 189. 185
|
2010
|
19
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 12–20, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Dependency Parsing and Projection Based on Word-Pair Classification Wenbin Jiang and Qun Liu Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100190, China {jiangwenbin, liuqun}@ict.ac.cn Abstract In this paper we describe an intuitionistic method for dependency parsing, where a classifier is used to determine whether a pair of words forms a dependency edge. And we also propose an effective strategy for dependency projection, where the dependency relationships of the word pairs in the source language are projected to the word pairs of the target language, leading to a set of classification instances rather than a complete tree. Experiments show that, the classifier trained on the projected classification instances significantly outperforms previous projected dependency parsers. More importantly, when this classifier is integrated into a maximum spanning tree (MST) dependency parser, obvious improvement is obtained over the MST baseline. 1 Introduction Supervised dependency parsing achieves the stateof-the-art in recent years (McDonald et al., 2005a; McDonald and Pereira, 2006; Nivre et al., 2006). Since it is costly and difficult to build humanannotated treebanks, a lot of works have also been devoted to the utilization of unannotated text. For example, the unsupervised dependency parsing (Klein and Manning, 2004) which is totally based on unannotated data, and the semisupervised dependency parsing (Koo et al., 2008) which is based on both annotated and unannotated data. Considering the higher complexity and lower performance in unsupervised parsing, and the need of reliable priori knowledge in semisupervised parsing, it is a promising strategy to project the dependency structures from a resource-rich language to a resource-scarce one across a bilingual corpus (Hwa et al., 2002; Hwa et al., 2005; Ganchev et al., 2009; Smith and Eisner, 2009; Jiang et al., 2009). For dependency projection, the relationship between words in the parsed sentences can be simply projected across the word alignment to words in the unparsed sentences, according to the DCA assumption (Hwa et al., 2005). Such a projection procedure suffers much from the word alignment errors and syntactic isomerism between languages, which usually lead to relationship projection conflict and incomplete projected dependency structures. To tackle this problem, Hwa et al. (2005) use some filtering rules to reduce noise, and some hand-designed rules to handle language heterogeneity. Smith and Eisner (2009) perform dependency projection and annotation adaptation with quasi-synchronous grammar features. Jiang and Liu (2009) resort to a dynamic programming procedure to search for a completed projected tree. However, these strategies are all confined to the same category that dependency projection must produce completed projected trees. Because of the free translation, the syntactic isomerism between languages and word alignment errors, it would be strained to completely project the dependency structure from one language to another. We propose an effective method for dependency projection, which does not have to produce complete projected trees. Given a wordaligned bilingual corpus with source language sentences parsed, the dependency relationships of the word pairs in the source language are projected to the word pairs of the target language. A dependency relationship is a boolean value that represents whether this word pair forms a dependency edge. Thus a set of classification instances are obtained. Meanwhile, we propose an intuitionistic model for dependency parsing, which uses a classifier to determine whether a pair of words form a dependency edge. The classifier can then be trained on the projected classification instance set, so as to build a projected dependency parser without the need of complete projected trees. 12 i j j i Figure 1: Illegal (a) and incomplete (b) dependency tree produced by the simple-collection method. Experimental results show that, the classifier trained on the projected classification instances significantly outperforms the projected dependency parsers in previous works. The classifier trained on the Chinese projected classification instances achieves a precision of 58.59% on the CTB standard test set. More importantly, when this classifier is integrated into a 2nd-ordered maximum spanning tree (MST) dependency parser (McDonald and Pereira, 2006) in a weighted average manner, significant improvement is obtained over the MST baselines. For the 2nd-order MST parser trained on Penn Chinese Treebank (CTB) 5.0, the classifier give an precision increment of 0.5 points. Especially for the parser trained on the smaller CTB 1.0, more than 1 points precision increment is obtained. In the rest of this paper, we first describe the word-pair classification model for dependency parsing (section 2) and the generation method of projected classification instances (section 3). Then we describe an application of the projected parser: boosting a state-of-the-art 2nd-ordered MST parser (section 4). After the comparisons with previous works on dependency parsing and projection, we finally five the experimental results. 2 Word-Pair Classification Model 2.1 Model Definition Following (McDonald et al., 2005a), x is used to denote the sentence to be parsed, and xi to denote the i-th word in the sentence. y denotes the dependency tree for sentence x, and (i, j) ∈y represents a dependency edge from word xi to word xj, where xi is the parent of xj. The task of the word-pair classification model is to determine whether any candidate word pair, xi and xj s.t. 1 ≤i, j ≤|x| and i ̸= j, forms a dependency edge. The classification result C(i, j) can be a boolean value: C(i, j) = p p ∈{0, 1} (1) as produced by a support vector machine (SVM) classifier (Vapnik, 1998). p = 1 indicates that the classifier supports the candidate edge (i, j), and p = 0 the contrary. C(i, j) can also be a realvalued probability: C(i, j) = p 0 ≤p ≤1 (2) as produced by an maximum entropy (ME) classifier (Berger et al., 1996). p is a probability which indicates the degree the classifier support the candidate edge (i, j). Ideally, given the classification results for all candidate word pairs, the dependency parse tree can be composed of the candidate edges with higher score (1 for the boolean-valued classifier, and large p for the real-valued classifier). However, more robust strategies should be investigated since the ambiguity of the language syntax and the classification errors usually lead to illegal or incomplete parsing result, as shown in Figure 1. Follow the edge based factorization method (Eisner, 1996), we factorize the score of a dependency tree s(x, y) into its dependency edges, and design a dynamic programming algorithm to search for the candidate parse with maximum score. This strategy alleviate the classification errors to some degree and ensure a valid, complete dependency parsing tree. If a boolean-valued classifier is used, the search algorithm can be formalized as: ˜y = argmax y s(x, y) = argmax y X (i,j)∈y C(i, j) (3) And if a probability-valued classifier is used instead, we replace the accumulation with cumula13 Type Features Unigram wordi ◦posi wordi posi wordj ◦posj wordj posj Bigram wordi ◦posi ◦wordj ◦posj posi ◦wordj ◦posj wordi ◦wordj ◦posj wordi ◦posi ◦posj wordi ◦posi ◦wordj wordi ◦wordj posi ◦posj wordi ◦posj posi ◦wordj Surrounding posi ◦posi+1 ◦posj−1 ◦posj posi−1 ◦posi ◦posj−1 ◦posj posi ◦posi+1 ◦posj ◦posj+1 posi−1 ◦posi ◦posj ◦posj+1 posi−1 ◦posi ◦posj−1 posi−1 ◦posi ◦posj+1 posi ◦posi+1 ◦posj−1 posi ◦posi+1 ◦posj+1 posi−1 ◦posj−1 ◦posj posi−1 ◦posj ◦posj+1 posi+1 ◦posj−1 ◦posj posi+1 ◦posj ◦posj+1 posi ◦posj−1 ◦posj posi ◦posj ◦posj+1 posi−1 ◦posi ◦posj posi ◦posi+1 ◦posj Table 1: Feature templates for the word-pair classification model. tive product: ˜y = argmax y s(x, y) = argmax y Y (i,j)∈y C(i, j) (4) Where y is searched from the set of well-formed dependency trees. In our work we choose a real-valued ME classifier. Here we give the calculation of dependency probability C(i, j). We use w to denote the parameter vector of the ME model, and f(i, j, r) to denote the feature vector for the assumption that the word pair i and j has a dependency relationship r. The symbol r indicates the supposed classification result, where r = + means we suppose it as a dependency edge and r = −means the contrary. A feature fk(i, j, r) ∈f(i, j, r) equals 1 if it is activated by the assumption and equals 0 otherwise. The dependency probability can then be defined as: C(i, j) = exp(w · f(i, j, +)) P r exp(w · f(i, j, r)) = exp(P k wk × fk(i, j, +)) P r exp(P k wk × fk(i, j, r)) (5) 2.2 Features for Classification The feature templates for the classifier are similar to those of 1st-ordered MST model (McDonald et al., 2005a). 1 Each feature is composed of some words and POS tags surrounded word i and/or word j, as well as an optional distance representations between this two words. Table shows the feature templates we use. Previous graph-based dependency models usually use the index distance of word i and word j 1We exclude the in between features of McDonald et al. (2005a) since preliminary experiments show that these features bring no improvement to the word-pair classification model. to enrich the features with word distance information. However, in order to utilize some syntax information between the pair of words, we adopt the syntactic distance representation of (Collins, 1996), named Collins distance for convenience. A Collins distance comprises the answers of 6 questions: • Does word i precede or follow word j? • Are word i and word j adjacent? • Is there a verb between word i and word j? • Are there 0, 1, 2 or more than 2 commas between word i and word j? • Is there a comma immediately following the first of word i and word j? • Is there a comma immediately preceding the second of word i and word j? Besides the original features generated according to the templates in Table 1, the enhanced features with Collins distance as postfixes are also used in training and decoding of the word-pair classifier. 2.3 Parsing Algorithm We adopt logarithmic dependency probabilities in decoding, therefore the cumulative product of probabilities in formula 6 can be replaced by accumulation of logarithmic probabilities: ˜y = argmax y s(x, y) = argmax y Y (i,j)∈y C(i, j) = argmax y X (i,j)∈y log(C(i, j)) (6) Thus, the decoding algorithm for 1st-ordered MST model, such as the Chu-Liu-Edmonds algorithm 14 Algorithm 1 Dependency Parsing Algorithm. 1: Input: sentence x to be parsed 2: for ⟨i, j⟩⊆⟨1, |x|⟩in topological order do 3: buf ←∅ 4: for k ←i..j −1 do ⊲all partitions 5: for l ∈V[i, k] and r ∈V[k + 1, j] do 6: insert DERIV(l, r) into buf 7: insert DERIV(r, l) into buf 8: V[i, j] ←top K derivations of buf 9: Output: the best derivation of V[1, |x|] 10: function DERIV(p, c) 11: d ←p ∪c ∪{(p · root, c · root)} ⊲new derivation 12: d · evl ←EVAL(d) ⊲evaluation function 13: return d used in McDonald et al. (2005b), is also applicable here. In this work, however, we still adopt the more general, bottom-up dynamic programming algorithm Algorithm 1 in order to facilitate the possible expansions. Here, V[i, j] contains the candidate parsing segments of the span [i, j], and the function EVAL(d) accumulates the scores of all the edges in dependency segment d. In practice, the cube-pruning strategy (Huang and Chiang, 2005) is used to speed up the enumeration of derivations (loops started by line 4 and 5). 3 Projected Classification Instance After the introduction of the word-pair classification model, we now describe the extraction of projected dependency instances. In order to alleviate the effect of word alignment errors, we base the projection on the alignment matrix, a compact representation of multiple GIZA++ (Och and Ney, 2000) results, rather than a single word alignment in previous dependency projection works. Figure 2 shows an example. Suppose a bilingual sentence pair, composed of a source sentence e and its target translation f. ye is the parse tree of the source sentence. A is the alignment matrix between them, and each element Ai,j denotes the degree of the alignment between word ei and word fj. We define a boolean-valued function δ(y, i, j, r) to investigate the dependency relationship of word i and word j in parse tree y: δ(y, i, j, r) = 1 (i, j) ∈y and r = + or (i, j) /∈y and r = − 0 otherwise (7) Then the score that word i and word j in the target sentence y forms a projected dependency edge, Figure 2: The word alignment matrix between a Chinese sentence and its English translation. Note that probabilities need not to be normalized across rows or columns. s+(i, j), can be defined as: s+(i, j) = X i′,j′ Ai,i′ × Aj,j′ × δ(ye, i′, j′, +) (8) The score that they do not form a projected dependency edge can be defined similarly: s−(i, j) = X i′,j′ Ai,i′ × Aj,j′ × δ(ye, i′, j′, −) (9) Note that for simplicity, the condition factors ye and A are omitted from these two formulas. We finally define the probability of the supposed projected dependency edge as: Cp(i, j) = exp(s+(i, j)) exp(s+(i, j)) + exp(s−(i, j)) (10) The probability Cp(i, j) is a real value between 0 and 1. Obviously, Cp(i, j) = 0.5 indicates the most ambiguous case, where we can not distinguish between positive and negative at all. On the other hand, there are as many as 2|f|(|f|−1) candidate projected dependency instances for the target sentence f. Therefore, we need choose a threshold b for Cp(i, j) to filter out the ambiguous instances: the instances with Cp(i, j) > b are selected as the positive, and the instances with Cp(i, j) < 1 −b are selected as the negative. 4 Boosting an MST Parser The classifier can be used to boost a existing parser trained on human-annotated trees. We first establish a unified framework for the enhanced parser. For a sentence to be parsed, x, the enhanced parser selects the best parse ˜y according to both the baseline model B and the projected classifier C. ˜y = argmax y [sB(x, y) + λsC(x, y)] (11) 15 Here, sB and sC denote the evaluation functions of the baseline model and the projected classifier, respectively. The parameter λ is the relative weight of the projected classifier against the baseline model. There are several strategies to integrate the two evaluation functions. For example, they can be integrated deeply at each decoding step (Carreras et al., 2008; Zhang and Clark, 2008; Huang, 2008), or can be integrated shallowly in a reranking manner (Collins, 2000; Charniak and Johnson, 2005). As described previously, the score of a dependency tree given by a word-pair classifier can be factored into each candidate dependency edge in this tree. Therefore, the projected classifier can be integrated with a baseline model deeply at each dependency edge, if the evaluation score given by the baseline model can also be factored into dependency edges. We choose the 2nd-ordered MST model (McDonald and Pereira, 2006) as the baseline. Especially, the effect of the Collins distance in the baseline model is also investigated. The relative weight λ is adjusted to maximize the performance on the development set, using an algorithm similar to minimum error-rate training (Och, 2003). 5 Related Works 5.1 Dependency Parsing Both the graph-based (McDonald et al., 2005a; McDonald and Pereira, 2006; Carreras et al., 2006) and the transition-based (Yamada and Matsumoto, 2003; Nivre et al., 2006) parsing algorithms are related to our word-pair classification model. Similar to the graph-based method, our model is factored on dependency edges, and its decoding procedure also aims to find a maximum spanning tree in a fully connected directed graph. From this point, our model can be classified into the graph-based category. On the training method, however, our model obviously differs from other graph-based models, that we only need a set of word-pair dependency instances rather than a regular dependency treebank. Therefore, our model is more suitable for the partially bracketed or noisy training corpus. The most apparent similarity between our model and the transition-based category is that they all need a classifier to perform classification conditioned on a certain configuration. However, they differ from each other in the classification results. The classifier in our model predicates a dependency probability for each pair of words, while the classifier in a transition-based model gives a possible next transition operation such as shift or reduce. Another difference lies in the factorization strategy. For our method, the evaluation score of a candidate parse is factorized into each dependency edge, while for the transition-based models, the score is factorized into each transition operation. Thanks to the reminding of the third reviewer of our paper, we find that the pairwise classification schema has also been used in Japanese dependency parsing (Uchimoto et al., 1999; Kudo and Matsumoto, 2000). However, our work shows more advantage in feature engineering, model training and decoding algorithm. 5.2 Dependency Projection Many works try to learn parsing knowledge from bilingual corpora. L¨u et al. (2002) aims to obtain Chinese bracketing knowledge via ITG (Wu, 1997) alignment. Hwa et al. (2005) and Ganchev et al. (2009) induce dependency grammar via projection from aligned bilingual corpora, and use some thresholds to filter out noise and some hand-written rules to handle heterogeneity. Smith and Eisner (2009) perform dependency projection and annotation adaptation with Quasi-Synchronous Grammar features. Jiang and Liu (2009) refer to alignment matrix and a dynamic programming search algorithm to obtain better projected dependency trees. All previous works for dependency projection (Hwa et al., 2005; Ganchev et al., 2009; Smith and Eisner, 2009; Jiang and Liu, 2009) need complete projected trees to train the projected parsers. Because of the free translation, the word alignment errors, and the heterogeneity between two languages, it is reluctant and less effective to project the dependency tree completely to the target language sentence. On the contrary, our dependency projection strategy prefer to extract a set of dependency instances, which coincides our model’s demand for training corpus. An obvious advantage of this strategy is that, we can select an appropriate filtering threshold to obtain dependency instances of good quality. In addition, our word-pair classification model can be integrated deeply into a state-of-the-art MST dependency model. Since both of them are 16 Corpus Train Dev Test WSJ (section) 2-21 22 23 CTB 5.0 (chapter) others 301-325 271-300 Table 2: The corpus partition for WSJ and CTB 5.0. factorized into dependency edges, the integration can be conducted at each dependency edge, by weightedly averaging their evaluation scores for this dependency edge. This strategy makes better use of the projected parser while with faster decoding, compared with the cascaded approach of Jiang and Liu (2009). 6 Experiments In this section, we first validate the word-pair classification model by experimenting on humanannotated treebanks. Then we investigate the effectiveness of the dependency projection by evaluating the projected classifiers trained on the projected classification instances. Finally, we report the performance of the integrated dependency parser which integrates the projected classifier and the 2nd-ordered MST dependency parser. We evaluate the parsing accuracy by the precision of lexical heads, which is the percentage of the words that have found their correct parents. 6.1 Word-Pair Classification Model We experiment on two popular treebanks, the Wall Street Journal (WSJ) portion of the Penn English Treebank (Marcus et al., 1993), and the Penn Chinese Treebank (CTB) 5.0 (Xue et al., 2005). The constituent trees in the two treebanks are transformed to dependency trees according to the headfinding rules of Yamada and Matsumoto (2003). For English, we use the automatically-assigned POS tags produced by an implementation of the POS tagger of Collins (2002). While for Chinese, we just use the gold-standard POS tags following the tradition. Each treebank is splitted into three partitions, for training, development and testing, respectively, as shown in Table 2. For a dependency tree with n words, only n − 1 positive dependency instances can be extracted. They account for only a small proportion of all the dependency instances. As we know, it is important to balance the proportions of the positive and the negative instances for a batched-trained classifier. We define a new parameter r to denote the ratio of the negative instances relative to the positive ones. 84 84.5 85 85.5 86 86.5 87 1 1.5 2 2.5 3 Dependency Precision (%) Ratio r (#negative/#positive) WSJ CTB 5.0 Figure 3: Performance curves of the word-pair classification model on the development sets of WSJ and CTB 5.0, with respect to a series of ratio r. Corpus System P % WSJ Yamada and Matsumoto (2003) 90.3 Nivre and Scholz (2004) 87.3 1st-ordered MST 90.7 2nd-ordered MST 91.5 our model 86.8 CTB 5.0 1st-ordered MST 86.53 2nd-ordered MST 87.15 our model 82.06 Table 3: Performance of the word-pair classification model on WSJ and CTB 5.0, compared with the current state-of-the-art models. For example, r = 2 means we reserve negative instances two times as many as the positive ones. The MaxEnt toolkit by Zhang 2 is adopted to train the ME classifier on extracted instances. We set the gaussian prior as 1.0 and the iteration limit as 100, leaving other parameters as default values. We first investigate the impact of the ratio r on the performance of the classifier. Curves in Figure 3 show the performance of the English and Chinese parsers, each of which is trained on an instance set corresponding to a certain r. We find that for both English and Chinese, maximum performance is achieved at about r = 2.5. 3 The English and Chinese classifiers trained on the instance sets with r = 2.5 are used in the final evaluation phase. Table 3 shows the performances on the test sets of WSJ and CTB 5.0. We also compare them with previous works on the same test sets. On both English and Chinese, the word-pair classification model falls behind of the state-of-the-art. We think that it is probably 2http://homepages.inf.ed.ac.uk/s0450736/ maxent toolkit.html. 3We did not investigate more fine-grained ratios, since the performance curves show no dramatic fluctuation along with the alteration of r. 17 54 54.5 55 55.5 56 0.65 0.7 0.75 0.8 0.85 0.9 0.95 Dependency Precision (%) Threshold b Figure 4: The performance curve of the wordpair classification model on the development set of CTB 5.0, with respect to a series of threshold b. due to the local optimization of the training procedure. Given complete trees as training data, it is easy for previous models to utilize structural, global and linguistical information in order to obtain more powerful parameters. The main advantage of our model is that it doesn’t need complete trees to tune its parameters. Therefore, if trained on instances extracted from human-annotated treebanks, the word-pair classification model would not demonstrate its advantage over existed stateof-the-art dependency parsing methods. 6.2 Dependency Projection In this work we focus on the dependency projection from English to Chinese. We use the FBIS Chinese-English bitext as the bilingual corpus for dependency projection. It contains 239K sentence pairs with about 6.9M/8.9M words in Chinese/English. Both English and Chinese sentences are tagged by the implementations of the POS tagger of Collins (2002), which trained on WSJ and CTB 5.0 respectively. The English sentences are then parsed by an implementation of 2nd-ordered MST model of McDonald and Pereira (2006), which is trained on dependency trees extracted from WSJ. The alignment matrixes for sentence pairs are generated according to (Liu et al., 2009). Similar to the ratio r, the threshold b need also be assigned an appropriate value to achieve a better performance. Larger thresholds result in better but less classification instances, the lower coverage of the instances would hurt the performance of the classifier. On the other hand, smaller thresholds lead to worse but more instances, and too much noisy instances will bring down the classifier’s discriminating power. We extract a series of classification instance sets Corpus System P % CTB 2.0 Hwa et al. (2005) 53.9 our model 56.9 CTB 5.0 Jiang and Liu (2009) 53.28 our model 58.59 Table 4: The performance of the projected classifier on the test sets of CTB 2.0 and CTB 5.0, compared with the performance of previous works on the corresponding test sets. Corpus Baseline P% Integrated P% CTB 1.0 82.23 83.70 CTB 5.0 87.15 87.65 Table 5: Performance improvement brought by the projected classifier to the baseline 2nd-ordered MST parsers trained on CTB 1.0 and CTB 5.0, respectively. with different thresholds. Then, on each instance set we train a classifier and test it on the development set of CTB 5.0. Figure 4 presents the experimental results. The curve shows that the maximum performance is achieved at the threshold of about 0.85. The classifier corresponding to this threshold is evaluated on the test set of CTB 5.0, and the test set of CTB 2.0 determined by Hwa et al. (2005). Table 4 shows the performance of the projected classifier, as well as the performance of previous works on the corresponding test sets. The projected classifier significantly outperforms previous works on both test sets, which demonstrates that the word-pair classification model, although falling behind of the state-of-the-art on humanannotated treebanks, performs well in projected dependency parsing. We give the credit to its good collaboration with the word-pair classification instance extraction for dependency projection. 6.3 Integrated Dependency Parser We integrate the word-pair classification model into the state-of-the-art 2nd-ordered MST model. First, we implement a chart-based dynamic programming parser for the 2nd-ordered MST model, and develop a training procedure based on the perceptron algorithm with averaged parameters (Collins, 2002). On the WSJ corpus, this parser achieves the same performance as that of McDonald and Pereira (2006). Then, at each derivation step of this 2nd-ordered MST parser, we weightedly add the evaluation score given by the projected classifier to the original MST evaluation score. Such a weighted summation of two eval18 uation scores provides better evaluation for candidate parses. The weight parameter λ is tuned by a minimum error-rate training algorithm (Och, 2003). Given a 2nd-ordered MST parser trained on CTB 5.0 as the baseline, the projected classifier brings an accuracy improvement of about 0.5 points. For the baseline trained on the smaller CTB 1.0, whose training set is chapters 1-270 of CTB 5.0, the accuracy improvement is much significant, about 1.5 points over the baseline. It indicates that, the smaller the human-annotated treebank we have, the more significant improvement we can achieve by integrating the projecting classifier. This provides a promising strategy for boosting the parsing performance of resourcescarce languages. Table 5 summarizes the experimental results. 7 Conclusion and Future Works In this paper, we first describe an intuitionistic method for dependency parsing, which resorts to a classifier to determine whether a word pair forms a dependency edge, and then propose an effective strategy for dependency projection, which produces a set of projected classification instances rather than complete projected trees. Although this parsing method falls behind of previous models, it can collaborate well with the word-pair classification instance extraction strategy for dependency projection, and achieves the state-of-the-art in projected dependency parsing. In addition, when integrated into a 2nd-ordered MST parser, the projected parser brings significant improvement to the baseline, especially for the baseline trained on smaller treebanks. This provides a new strategy for resource-scarce languages to train high-precision dependency parsers. However, considering its lower performance on human-annotated treebanks, the dependency parsing method itself still need a lot of investigations, especially on the training method of the classifier. Acknowledgement This project was supported by National Natural Science Foundation of China, Contract 60736014, and 863 State Key Project No. 2006AA010108. We are grateful to the anonymous reviewers for their thorough reviewing and valuable suggestions. We show special thanks to Dr. Rebecca Hwa for generous help of sharing the experimental data. We also thank Dr. Yang Liu for sharing the codes of alignment matrix generation, and Dr. Liang Huang for helpful discussions. References Adam L. Berger, Stephen A. Della Pietra, and Vincent J. Della Pietra. 1996. A maximum entropy approach to natural language processing. Computational Linguistics. Xavier Carreras, Mihai Surdeanu, and Lluis Marquez. 2006. Projective dependency parsing with perceptron. In Proceedings of the CoNLL. Xavier Carreras, Michael Collins, and Terry Koo. 2008. Tag, dynamic programming, and the perceptron for efficient, feature-rich parsing. In Proceedings of the CoNLL. Eugene Charniak and Mark Johnson. 2005. Coarseto-fine-grained n-best parsing and discriminative reranking. In Proceedings of the ACL. Michael Collins. 1996. A new statistical parser based on bigram lexical dependencies. In Proceedings of ACL. Michael Collins. 2000. Discriminative reranking for natural language parsing. In Proceedings of the ICML, pages 175–182. Michael Collins. 2002. Discriminative training methods for hidden markov models: Theory and experiments with perceptron algorithms. In Proceedings of the EMNLP, pages 1–8, Philadelphia, USA. Jason M. Eisner. 1996. Three new probabilistic models for dependency parsing: An exploration. In Proceedings of COLING, pages 340–345. Kuzman Ganchev, Jennifer Gillenwater, and Ben Taskar. 2009. Dependency grammar induction via bitext projection constraints. In Proceedings of the 47th ACL. Liang Huang and David Chiang. 2005. Better k-best parsing. In Proceedings of the IWPT, pages 53–64. Liang Huang. 2008. Forest reranking: Discriminative parsing with non-local features. In Proceedings of the ACL. Rebecca Hwa, Philip Resnik, Amy Weinberg, and Okan Kolak. 2002. Evaluating translational correspondence using annotation projection. In Proceedings of the ACL. Rebecca Hwa, Philip Resnik, Amy Weinberg, Clara Cabezas, and Okan Kolak. 2005. Bootstrapping parsers via syntactic projection across parallel texts. In Natural Language Engineering, volume 11, pages 311–325. 19 Wenbin Jiang and Qun Liu. 2009. Automatic adaptation of annotation standards for dependency parsing using projected treebank as source corpus. In Proceedings of IWPT. Wenbin Jiang, Liang Huang, and Qun Liu. 2009. Automatic adaptation of annotation standards: Chinese word segmentation and pos tagging–a case study. In Proceedings of the 47th ACL. Dan Klein and Christopher D. Manning. 2004. Corpusbased induction of syntactic structure: Models of dependency and constituency. In Proceedings of the ACL. Terry Koo, Xavier Carreras, and Michael Collins. 2008. Simple semi-supervised dependency parsing. In Proceedings of the ACL. Taku Kudo and Yuji Matsumoto. 2000. Japanese dependency structure analysis based on support vector machines. In Proceedings of the EMNLP. Yang Liu, Tian Xia, Xinyan Xiao, and Qun Liu. 2009. Weighted alignment matrices for statistical machine translation. In Proceedings of the EMNLP. Yajuan L¨u, Sheng Li, Tiejun Zhao, and Muyun Yang. 2002. Learning chinese bracketing knowledge based on a bilingual language model. In Proceedings of the COLING. Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of english: The penn treebank. In Computational Linguistics. Ryan McDonald and Fernando Pereira. 2006. Online learning of approximate dependency parsing algorithms. In Proceedings of EACL, pages 81–88. Ryan McDonald, Koby Crammer, and Fernando Pereira. 2005a. Online large-margin training of dependency parsers. In Proceedings of ACL, pages 91– 98. Ryan McDonald, Fernando Pereira, Kiril Ribarov, and Jan Haji˘c. 2005b. Non-projective dependency parsing using spanning tree algorithms. In Proceedings of HLT-EMNLP. J. Nivre and M. Scholz. 2004. Deterministic dependency parsing of english text. In Proceedings of the COLING. Joakim Nivre, Johan Hall, Jens Nilsson, Gulsen Eryigit, and Svetoslav Marinov. 2006. Labeled pseudoprojective dependency parsing with support vector machines. In Proceedings of CoNLL, pages 221–225. Franz J. Och and Hermann Ney. 2000. Improved statistical alignment models. In Proceedings of the ACL. Franz Joseph Och. 2003. Minimum error rate training in statistical machine translation. In Proceedings of the ACL, pages 160–167. David Smith and Jason Eisner. 2009. Parser adaptation and projection with quasi-synchronous grammar features. In Proceedings of EMNLP. Kiyotaka Uchimoto, Satoshi Sekine, and Hitoshi Isahara. 1999. Japanese dependency structure analysis based on maximum entropy models. In Proceedings of the EACL. Vladimir N. Vapnik. 1998. Statistical learning theory. In A Wiley-Interscience Publication. Dekai Wu. 1997. Stochastic inversion transduction grammars and bilingual parsing of parallel corpora. Computational Linguistics. Nianwen Xue, Fei Xia, Fu-Dong Chiou, and Martha Palmer. 2005. The penn chinese treebank: Phrase structure annotation of a large corpus. In Natural Language Engineering. H Yamada and Y Matsumoto. 2003. Statistical dependency analysis using support vector machines. In Proceedings of IWPT. Yue Zhang and Stephen Clark. 2008. Joint word segmentation and pos tagging using a single perceptron. In Proceedings of the ACL. 20
|
2010
|
2
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 186–195, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Entity-based local coherence modelling using topological fields Jackie Chi Kit Cheung and Gerald Penn Department of Computer Science University of Toronto Toronto, ON, M5S 3G4, Canada {jcheung,gpenn}@cs.toronto.edu Abstract One goal of natural language generation is to produce coherent text that presents information in a logical order. In this paper, we show that topological fields, which model high-level clausal structure, are an important component of local coherence in German. First, we show in a sentence ordering experiment that topological field information improves the entity grid model of Barzilay and Lapata (2008) more than grammatical role and simple clausal order information do, particularly when manual annotations of this information are not available. Then, we incorporate the model enhanced with topological fields into a natural language generation system that generates constituent orders for German text, and show that the added coherence component improves performance slightly, though not statistically significantly. 1 Introduction One type of coherence modelling that has captured recent research interest is local coherence modelling, which measures the coherence of a document by examining the similarity between neighbouring text spans. The entity-based approach, in particular, considers the occurrences of noun phrase entities in a document (Barzilay and Lapata, 2008). Local coherence modelling has been shown to be useful for tasks like natural language generation and summarization, (Barzilay and Lee, 2004) and genre classification (Barzilay and Lapata, 2008). Previous work on English, a language with relatively fixed word order, has identified factors that contribute to local coherence, such as the grammatical roles associated with the entities. There is good reason to believe that the importance of these factors vary across languages. For instance, freerword-order languages exhibit word order patterns which are dependent on discourse factors relating to information structure, in addition to the grammatical roles of nominal arguments of the main verb. We thus expect word order information to be particularly important in these languages in discourse analysis, which includes coherence modelling. For example, Strube and Hahn (1999) introduce Functional Centering, a variant of Centering Theory which utilizes information status distinctions between hearer-old and hearer-new entities. They apply their model to pronominal anaphora resolution, identifying potential antecedents of subsequent anaphora by considering syntactic and word order information, classifying constituents by their familiarity to the reader. They find that their approach correctly resolves more pronominal anaphora than a grammatical role-based approach which ignores word order, and the difference between the two approaches is larger in German corpora than in English ones. Unfortunately, their criteria for ranking potential antecedents require complex syntactic information in order to classify whether proper names are known to the hearer, which makes their algorithm hard to automate. Indeed, all evaluation is done manually. We instead use topological fields, a model of clausal structure which is indicative of information structure in German, but shallow enough to be automatically parsed at high accuracy. We test the hypothesis that they would provide a good complement or alternative to grammatical roles in local coherence modelling. We show that they are superior to grammatical roles in a sentence ordering experiment, and in fact outperforms simple word-order information as well. We further show that these differences are particularly large when manual syntactic and grammatical role an186 Millionen von Mark verschwendet der Senat jeden Monat, weil er sparen will. LK MF VC VF LK MF S NF S “The senate wastes millions of marks each month, because it wants to save.” Figure 1: The clausal and topological field structure of a German sentence. Notice that the subordinate clause receives its own topology. notations are not available. We then embed these topological field annotations into a natural language generation system to show the utility of local coherence information in an applied setting. We add contextual features using topological field transitions to the model of Filippova and Strube (2007b) and achieve a slight improvement over their model in a constituent ordering task, though not statistically significantly. We conclude by discussing possible reasons for the utility of topological fields in local coherence modelling. 2 Background and Related Work 2.1 German Topological Field Parsing Topological fields are sequences of one or more contiguous phrases found in an enclosing syntactic region, which is the clause in the case of the German topological field model (H¨ohle, 1983). These fields may have constraints on the number of words or phrases they contain, and do not necessarily form a semantically coherent constituent. In German, the topology serves to identify all of the components of the verbal head of a clause, as well as clause-level structure such as complementizers and subordinating conjunctions. Topological fields are a useful abstraction of word order, because while Germanic word order is relatively free with respect to grammatical functions, the order of the topological fields is strict and unvarying. A German clause can be considered to be anchored by two “brackets” which contain modals, verbs and complementizers. The left bracket (linke Klammer, LK) may contain a complementizer, subordinating conjunction, or a finite verb, depending on the clause type, and the right bracket contains the verbal complex (VC). The other topological fields are defined in relation to these two brackets, and contain all other parts of the clause such as verbal arguments, adjuncts, and discourse cues. The VF (Vorfeld or “pre-field”) is so-named because it occurs before the left bracket. As the first constituent of most matrix clauses in declarative sentences, it has special significance for the coherence of a passage, which we will further discuss below. The MF (Mittelfeld or “middle field”) is the field bounded by the two brackets. Most verb arguments, adverbs, and prepositional phrases are found here, unless they have been fronted and put in the VF, or are prosodically heavy and postposed to the NF field. The NF (Nachfeld or “post-field”) contains prosodically heavy elements such as postposed prepositional phrases or relative clauses, and occasionally postposed noun phrases. 2.2 The Role of the Vorfeld One of the reasons that we use topological fields for local coherence modelling is the role that the VF plays in signalling the information structure of German clauses, as it often contains the topic of the sentence. In fact, its role is much more complex than being simply the topic position. Dipper and Zinsmeister (2009) distinguish multiple uses of the VF depending on whether it contains an element related to the surrounding discourse. They find that 45.1% of VFs are clearly related to the previous context by a reference or discourse relation, and a further 21.9% are deictic and refer to the situation described in the passage in a corpus study. They also run a sentence insertion experiment where subjects are asked to place an extracted sentence in its original location in a passage. The authors remark that extracted sentences with VFs that are referentially related to previous context (e.g., they contain a coreferential noun phrase or a discourse relation like “therefore”) are reinserted at higher accuracies. 187 a) # Original Sentence and Translation 1 Einen Zufluchtsort f¨ur Frauen, die von ihren M¨annern mißhandelt werden, gibt es nunmehr auch in Treptow. “There is now a sanctuary for women who are mistreated by their husbands in Treptow as well.” 2 Das Bezirksamt bietet Frauen (auch mit Kindern) in derartigen Notsituationen vor¨ubergehend eine Unterkunft. “The district office offers women (even with children) in this type of emergency temporary accommodation.” 3 Zugleich werden die Betroffenen der Regelung des Unterhalts, bei Beh¨ordeng¨angen und auch bei der Wohnungssuche unterst¨utzt. “At the same time, the affected are supported with provisions of necessities, in dealing with authorities, and also in the search for new accommodations.” b) DE Zufluchtsort Frauen M¨annern Treptow Kindern EN sanctuary women husbands Treptow children 1 acc oth oth oth − 2 − oth − − oth 3 − nom − − − c) −− −nom −acc −oth nom − nom nom nom acc nom oth 0.3 0.0 0.0 0.1 0.0 0.0 0.0 0.0 acc − acc nom acc acc acc oth oth − oth nom oth acc oth oth 0.1 0.0 0.0 0.0 0.3 0.1 0.0 0.1 Table 1: a) An example of a document from T¨uBa-D/Z, b) an abbreviated entity grid representation of it, and c) the feature vector representation of the abbreviated entity grid for transitions of length two. Mentions of the entity Frauen are underlined. nom: nominative, acc: accusative, oth: dative, oblique, and other arguments Filippova and Strube (2007c) also examine the role of the VF in local coherence and natural language generation, focusing on the correlation between VFs and sentential topics. They follow Jacobs (2001) in distinguishing the topic of addressation, which is the constituent for which the proposition holds, and frame-setting topics, which is the domain in which the proposition holds, such as a temporal expression. They show in a user study that frame-setting topics are preferred to topics of addressation in the VF, except when a constituent needs to be established as the topic of addressation. 2.3 Using Entity Grids to Model Local Coherence Barzilay and Lapata (2008) introduce the entity grid as a method of representing the coherence of a document. Entity grids indicate the location of the occurrences of an entity in a document, which is important for coherence modelling because mentions of an entity tend to appear in clusters of neighbouring or nearby sentences in coherent documents. This last assumption is adapted from Centering Theory approaches to discourse modelling. In Barzilay and Lapata (2008), an entity grid is constructed for each document, and is represented as a matrix in which each row represents a sentence, and each column represents an entity. Thus, a cell in the matrix contains information about an entity in a sentence. The cell is marked by the presence or absence of the entity, and can also be augmented with other information about the entity in this sentence, such as the grammatical role of the noun phrase representing that entity in that sentence, or the topological field in which the noun phrase appears. Consider the document in Table 1. An entity grid representation which incorporates the syntactic role of the noun phrase in which the entity ap188 pears is also shown (not all entities are listed for brevity). We tabulate the transitions of entities between different syntactic positions (or their nonoccurrence) in sentences, and convert the frequencies of transitions into a feature vector representation of transition probabilities in the document. To calculate transition probabilities, we divide the frequency of a particular transition by the total number of transitions of that length. This model of local coherence was investigated for German by Filippova and Strube (2007a). The main focus of that work, however, was to adapt the model for use in a low-resource situation when perfect coreference information is not available. This is particularly useful in natural language understanding tasks. They employ a semantic clustering model to relate entities. In contrast, our work focuses on improving performance by annotating entities with additional linguistic information, such as topological fields, and is geared towards natural language generation systems where perfect information is available. Similar models of local coherence include various Centering Theory accounts of local coherence ((Kibble and Power, 2004; Poesio et al., 2004) inter alia). The model of Elsner and Charniak (2007) uses syntactic cues to model the discoursenewness of noun phrases. There are also more global content models of topic shifts between sentences like Barzilay and Lee (2004). 3 Sentence Ordering Experiments 3.1 Method We test a version of the entity grid representation augmented with topological fields in a sentence ordering experiment corresponding to Experiment 1 of Barzilay and Lapata (2008). The task is a binary classification task to identify the original version of a document from another version which contains the sentences in a randomly permuted order, which is taken to be incoherent. We solve this problem in a supervised machine learning setting, where the input is the feature vector representations of the two versions of the document, and the output is a binary value indicating the document with the original sentence ordering. We use SVMlight’s ranking module for classification (Joachims, 2002). The corpus in our experiments consists of the last 480 documents of T¨uBa-D/Z version 4 (Telljohann et al., 2004), which contains manual coreference, grammatical role and topological field information. This set is larger than the set that was used in Experiment 1 of Barzilay and Lapata (2008), which consists of 400 documents in two English subcorpora on earthquakes and accidents respectively. The average document length in the T¨uBaD/Z subcorpus is also greater, at 19.2 sentences compared to about 11 for the two subcorpora. Up to 20 random permutations of sentences were generated from each document, with duplicates removed. There are 216 documents and 4126 originalpermutation pairs in the training set, and 24 documents and 465 pairs in the development set. The remaining 240 documents are in the final test set (4243 pairs). The entity-based model is parameterized as follows. Transition length – the maximum length of the transitions used in the feature vector representation of a document. Representation – when marking the presence of an entity in a sentence, what information about the entity is marked (topological field, grammatical role, or none). We will describe the representations that we try in section 3.2. Salience – whether to set a threshold for the frequency of occurrence of entities. If this is set, all entities below a certain frequency are treated separately from those reaching this frequency threshold when calculating transition probabilities. In the example in Table 1, with a salience threshold of 2, Frauen would be treated separately from M¨annern or Kindern. Transition length, salience, and a regularization parameter are tuned on the development set. We only report results using the setting of transition length ≤4, and no salience threshold, because they give the best performance on the development set. This is in contrast to the findings of Barzilay and Lapata (2008), who report that transition length ≤3 and a salience threshold of 2 perform best on their data. 3.2 Entity Representations The main goal of this study is to compare word order, grammatical role and topological field information, which is encoded into the entity grid at each occurrence of an entity. Here, we describe the variants of the entity representations that we compare. 189 Baseline Representations We implement several baseline representations against which we test our topological field-enhanced model. The simplest baseline representation marks the mere appearance of an entity without any additional information, which we refer to as default. Another class of baseline representations mark the order in which entities appear in the clause. The correlation between word order and information structure is well known, and has formed the basis of some theories of syntax such as the Prague School’s (Sgall et al., 1986). The two versions of clausal order we tried are order 1/2/3+, which marks a noun phrase as the first, the second, or the third or later to appear in a clause, and order 1/2+, which marks a noun phrase as the first, or the second or later to appear in a clause. Since noun phrases can be embedded in other noun phrases, overlaps can occur. In this case, the dominating noun phrase takes the smallest order number among its dominated noun phrases. The third class of baseline representations we employ mark an entity by its grammatical role in the clause. Barzilay and Lapata (2008) found that grammatical role improves performance in this task for an English corpus. Because German distinguishes more grammatical roles morphologically than English, we experiment with various granularities of role labelling. In particular, subj/obj distinguishes the subject position, the object position, and another category for all other positions. cases distinguishes five types of entities corresponding to the four morphological cases of German in addition to another category for noun phrases which are not complements of the main verb. Topological Field-Based These representations mark the topological field in which an entity appears. Some versions mark entities which are prepositional objects separately. We try versions which distinguish VF from non-VF, as well as more general versions that make use of a greater set of topological fields. vf marks the noun phrase as belonging to a VF (and not in a PP) or not. vfpp is the same as above, but allows prepositional objects inside the VF to be marked as VF. topf/pp distinguishes entities in the topological fields VF, MF, and NF, contains a separate category for PP, and a category for all other noun phrases. topf distinguishes between VF, MF, and NF, on the one hand, and everything else on the other. Prepositional objects are treated the same as other noun phrases here. Combined We tried a representation which combines grammatical role and topological field into a single representation, subj/obj×vf, which takes the Cartesian product of subj/obj and vf above. Topological fields do not map directly to topicfocus distinctions. For example, besides the topic of the sentence, the Vorfeld may contain discourse cues, expletive pronouns, or the informational or contrastive focus. Furthermore, there are additional constraints on constituent order related to pronominalization. Thus, we devised additional entity representations to account for these aspects of German. topic attempts to identify the sentential topic of a clause. A noun phrase is marked as TOPIC if it is in VF as in vfpp, or if it is the first noun phrase in MF and also the first NP in the clause. Other noun phrases in MF are marked as NONTOPIC. Categories for NF and miscellaneous noun phrases also exist. While this representation may appear to be very similar to simply distinguishing the first entity in a clause as for order 1/2+ in that TOPIC would correspond to the first entity in the clause, they are in fact distinct. Due to issues related to coordination, appositive constructions, and fragments which do not receive a topology of fields, the first entity in a clause is labelled the TOPIC only 80.8% of the time in the corpus. This representation also distinguishes NFs, which clausal order does not. topic+pron refines the above by taking into account a word order restriction in German that pronouns appear before full noun phrases in the MF field. The following set of decisions represents how a noun phrase is marked: If the first NP in the clause is a pronoun in an MF field and is the subject, we mark it as TOPIC. If it is not the subject, we mark it as NONTOPIC. For other NPs, we follow the topic representation. 3.3 Automatic annotations While it is reasonable to assume perfect annotations of topological fields and grammatical roles in many NLG contexts, this assumption may be less appropriate in other applications involving text-totext generation where the input to the system is text such as paraphrasing or machine translation. Thus, we test the robustness of the entity repre190 Representation Manual Automatic topf/pp 94.44 94.89 topic 94.13 94.53 topic+pron 94.08 94.51 topf 93.87 93.11 subj/obj 93.831 91.7++ cases 93.312 90.93++ order 1/2+ 92.51++ 92.1+ subj/obj×vf 92.32++ 90.74++ default 91.42++ 91.42++ vfpp 91.37++ 91.68++ vf 91.21++ 91.16++ order 1/2/3+ 91.16++ 90.71++ Table 2: Accuracy (%) of the permutation detection experiment with various entity representations using manual and automatic annotations of topological fields and grammatical roles. The baseline without any additional annotation is underlined. Two-tailed sign tests were calculated for each result against the best performing model in each column (1: p = 0.101; 2: p = 0.053; +: statistically significant, p < 0.05; ++: very statistically significant, p < 0.01 ). sentations to automatic extraction in the absence of manual annotations. We employ the following two systems for extracting topological fields and grammatical roles. To parse topological fields, we use the Berkeley parser of Petrov and Klein (2007), which has been shown to perform well at this task (Cheung and Penn, 2009). The parser is trained on sections of T¨uBa-D/Z which do not overlap with the section from which the documents for this experiment were drawn, and obtains an overall parsing performance of 93.35% F1 on topological fields and clausal nodes without gold POS tags on the section of T¨uBa-D/Z it was tested on. We tried two methods to obtain grammatical roles. First, we tried extracting grammatical roles from the parse trees which we obtained from the Berkeley parser, as this information is present in the edge labels that can be recovered from the parse. However, we found that we achieved better accuracy by using RFTagger (Schmid and Laws, 2008), which tags nouns with their morphological case. Morphological case is distinct from grammatical role, as noun phrases can function as adjuncts in possessive constructions and preposiAnnotation Accuracy (%) Grammatical role 83.6 Topological field (+PP) 93.8 Topological field (−PP) 95.7 Clausal order 90.8 Table 3: Accuracy of automatic annotations of noun phrases with coreferents. +PP means that prepositional objects are treated as a separate category from topological fields. −PP means they are treated as other noun phrases. tional phrases. However, we can approximate the grammatical role of an entity using the morphological case. We follow the annotation conventions of T¨uBa-D/Z in not assigning a grammatical role when the noun phrase is a prepositional object. We also do not assign a grammatical role when the noun phrase is in the genitive case, as genitive objects are very rare in German and are far outnumbered by the possessive genitive construction. 3.4 Results Table 2 shows the results of the sentence ordering permutation detection experiment. The top four performing entity representations are all topological field-based, and they outperform grammatical role-based and simple clausal order-based models. These results indicate that the information that topological fields provide about clause structure, appositives, right dislocation, etc. which is not captured by simple clausal order is important for coherence modelling. The representations incorporating linguistics-based heuristics do not outperform purely topological field-based models. Surprisingly, the VF-based models fare quite poorly, performing worse than not adding any annotations, despite the fact that topological fieldbased models in general perform well. This result may be a result of the heterogeneous uses of the VF. The automatic topological field annotations are more accurate than the automatic grammatical role annotations (Table 3), which may partly explain why grammatical role-based models suffer more when using automatic annotations. Note, however, that the models based on automatic topological field annotations outperform even the grammatical role-based models using manual annotation (at marginal significance, p < 0.1). The topo191 logical field annotations are accurate enough that automatic annotations produce no decrease in performance. These results show the upper bound of entitybased local coherence modelling with perfect coreference information. The results we obtain are higher than the results for the English corpora of Barzilay and Lapata (2008) (87.2% on the Earthquakes corpus and 90.4% on the Accidents corpus), but this is probably due to corpus differences as well as the availability of perfect coreference information in our experiments1. Due to the high performance we obtained, we calculated Kendall tau coefficients (Lapata, 2006) over the sentence orderings of the cases in which our best performing model is incorrect, to determine whether the remaining errors are instances where the permuted ordering is nearly identical to the original ordering. We obtained a τ of 0.0456 in these cases, compared to a τ of −0.0084 for all the pairs, indicating that this is not the case. To facilitate comparison to the results of Filippova and Strube (2007a), we rerun this experiment on the same subsections of the corpus as in that work for training and testing. The first 100 articles of T¨uBa-D/Z are used for testing, while the next 200 are used for training and development. Unlike the previous experiments, we do not do parameter tuning on this set of data. Instead, we follow Filippova and Strube (2007a) in using transition lengths of up to three. We do not put in a salience threshold. We see that our results are much better than the ones reported in that work, even for the default representation. The main reason for this discrepancy is probably the way that entities are created from the corpus. In our experiments, we create an entity for every single noun phrase node that we encounter, then merge the entities that are linked by coreference. Filippova and Strube (2007a) convert the annotations of T¨uBa-D/Z into a dependency format, then extract entities from the noun phrases found there. They may thus annotate fewer entities, as there 1Barzilay and Lapata (2008) use the coreference system of Ng and Cardie (2002) to obtain coreference annotations. We are not aware of similarly well-tested, publicly available coreference resolution systems that handle all types of anaphora for German. We considered adapting the BART coreference resolution toolkit (Versley et al., 2008) to German, but a number of language-dependent decisions regarding preprocessing, feature engineering, and the learning paradigm would need to be made in order to achieve reasonable performance comparable to state-of-the-art English coreference resolution systems. Representation Accuracy (%) topf/pp 93.83 topic 93.31 topic+pron 93.31 topf 92.49 subj/obj 88.99 order 1/2+ 88.89 order 1/2/3+ 88.84 cases 88.63 vf 87.60 vfpp 88.17 default 87.55 subj/obj×vf 87.71 (Filippova and Strube, 2007) 75 Table 4: Accuracy (%) of permutation detection experiment with various entity representations using manual and automatic annotations of topological fields and grammatical roles on subset of corpus used by Filippova and Strube (2007a). may be nested NP nodes in the original corpus. There may also be noise in the dependency conversion process. The relative rankings of different entity representations in this experiment are similar to the rankings of the previous experiment, with topological field-based models outperforming grammatical role and clausal order models. 4 Local Coherence for Natural Language Generation One of the motivations of the entity grid-based model is to improve surface realization decisions in NLG systems. A typical experimental design would pass the contents of the test section of a corpus as input to the NLG system with the ordering information stripped away. The task is then to regenerate the ordering of the information found in the original corpus. Various coherence models have been tested in corpus-based NLG settings. For example, Karamanis et al. (2009) compare several versions of Centering Theory-based metrics of coherence on corpora by examining how highly the original ordering found in the corpus is ranked compared to other possible orderings of propositions. A metric performs well if it ranks the original ordering better than the alternative orderings. In our next experiment, we incorporate local co192 herence information into the system of Filippova and Strube (2007b). We embed entity topological field transitions into their probabilistic model, and show that the added coherence component slightly improves the performance of the baseline NLG system in generating constituent orderings in a German corpus, though not to a statistically significant degree. 4.1 Method We use the WikiBiography corpus2 for our experiments. The corpus consists of more than 1100 biographies taken from the German Wikipedia, and contains automatic annotations of morphological, syntactic, and semantic information. Each article also contains the coreference chain of the subject of the biography (the biographee). The first 100 articles are used for testing, the next 200 for development, and the rest for training. The baseline generation system already incorporates topological field information into the constituent ordering process. The system operates in two steps. First, in main clauses, one constituent is selected as the Vorfeld (VF). This is done using a maximum entropy model (call it MAXENT). Then, the remaining constituents are ordered using a second maximum entropy model (MAXENT2). Significantly, Filippova and Strube (2007b) found that selecting the VF first, and then ordering the remaining constituents results in a 9% absolute improvement over the corresponding model where the selection is performed in one step by the sorting algorithm alone. The maximum entropy model for both steps rely on the following features: • features on the voice, valency, and identity of the main verb of the clause • features on the morphological and syntactic status of the constituent to be ordered • whether the constituent occurs in the preceding sentence • features for whether the constituent contains a determiner, an anaphoric pronoun, or a relative clause • the size of the constituent in number of modifiers, in depth, and in number of words 2http://www.eml-research.de/english/ research/nlp/download/wikibiography.php • the semantic class of the constituent (person, temporal, location, etc.) The biographee, in particular, is marked by its own semantic class. In the first VF selection step, MAXENT simply produces a probability of each constituent being a VF, and the constituent with the highest probability is selected. In the second step, MAXENT2 takes the featural representation of two constituents, and produces an output probability of the first constituent preceding the second constituent. The final ordering is achieved by first randomizing the order of the constituents in a clause (besides the first one, which is selected to be the VF), then sorting them according to the precedence probabilities. Specifically, a constituent A is put before a constituent B if MAXENT2(A,B) > 0.5. Because this precedence relation is not antisymmetric (i.e., MAXENT2(A,B) > 0.5 and MAXENT2(B,A) > 0.5 may be simultaneously true or simultaneously false), different initializations of the order produce different sorted results. In our experiments, we correct this by defining the precedence relation to be A precedes B iff MAXENT2(A,B) > MAXENT2(B,A). This change does not greatly impact the performance, and removes the randomized element of the algorithm. The baseline system does not directly model the context when ordering constituents. All of the features but one in the original maximum entropy models rely on local properties of the clause. We incorporate local coherence information into the model by adding entity transition features which we found to be useful in the sentence ordering experiment in Section 3 above. Specifically, we add features indicating the topological fields in which entities occur in the previous sentences. We found that looking back up to two sentences produces the best results (by tuning on the development set). Because this corpus does not come with general coreference information except for the coreference chain of the biographee, we use the semantic classes instead. So, all constituents in the same semantic class are treated as one coreference chain. An example of a feature may be biog-last2, which takes on a value such as ‘v−’, meaning that this constituent refers to the biographee, and the biographee occurs in the VF two clauses ago (v), but does not appear in the previous clause (−). For a constituent which is not the biographee, this feature would be marked 193 Method VF Acc (%) Acc (%) Tau Baseline 68.7 60.9 0.72 +Coherence 69.2 61.5 0.72 Table 5: Results of adding coherence features into a natural language generation system. VF Acc% is the accuracy of selecting the first constituent in main clauses. Acc % is the percentage of perfectly ordered clauses, tau is Kendall’s τ on the constituent ordering. The test set contains 2246 clauses, of which 1662 are main clauses. ‘na’ (not applicable). 4.2 Results Table 5 shows the results of adding these contextual features into the maximum entropy models. We see that we obtain a small improvement in the accuracy of VF selection, and in the accuracy of correctly ordering the entire clause. These improvements are not statistically significant by McNemar’s test. We suggest that the lack of coreference information for all entities in the article may have reduced the benefit of the coherence component. Also, the topline of performance is substantially lower than 100%, as multiple orderings are possible and equally valid. Human judgements on information structuring for both interand intra-sentential units are known to have low agreement (Barzilay et al., 2002; Filippova and Strube, 2007c; Lapata, 2003; Chen et al., 2007). Thus, the relative error reduction is higher than the absolute reduction might suggest. 5 Conclusions We have shown that topological fields are a useful source of information for local coherence modelling. In a sentence-order permutation detection task, models which use topological field information outperform both grammatical role-based models and models based on simple clausal order, with the best performing model achieving a relative error reduction of 40.4% over the original baseline without any additional annotation. Applying our local coherence model in another setting, we have embedded topological field transitions of entities into an NLG system which orders constituents in German clauses. We find that the coherence-enhanced model slightly outperforms the baseline system, but this was not statistically significant. We suggest that the utility of topological fields in local coherence modelling comes from the interaction between word order and information structure in freer-word-order languages. Crucially, topological fields take into account issues such as coordination, appositives, sentential fragments and differences in clause types, which word order alone does not. They are also shallow enough to be accurately parsed automatically for use in resource-poor applications. Further refinement of the topological field annotations to take advantage of the fact that they do not correspond neatly to any single information status such as topic or focus could provide additional performance gains. The model also shows promise for other discourse-related tasks such as coreference resolution and discourse parsing. Acknowledgements We are grateful to Katja Filippova for providing us with source code for the experiments in Section 4 and for answering related questions, and to Timothy Fowler for useful discussions and comments on a draft of the paper. This work is supported in part by the Natural Sciences and Engineering Research Council of Canada. References R. Barzilay and M. Lapata. 2008. Modeling local coherence: An entity-based approach. Computational Linguistics, 34(1):1–34. R. Barzilay and L. Lee. 2004. Catching the drift: Probabilistic content models, with applications to generation and summarization. In Proc. HLT-NAACL 2004, pages 113–120. R. Barzilay, N. Elhadad, and K. McKeown. 2002. Inferring strategies for sentence ordering in multidocument news summarization. Journal of Artificial Intelligence Research, 17:35–55. E. Chen, B. Snyder, and R. Barzilay. 2007. Incremental text structuring with online hierarchical ranking. In Proceedings of EMNLP, pages 83–91. J.C.K. Cheung and G. Penn. 2009. Topological Field Parsing of German. In Proc. 47th ACL and 4th IJCNLP, pages 64–72. Association for Computational Linguistics. S. Dipper and H. Zinsmeister. 2009. The Role of the German Vorfeld for Local Coherence: A Pilot Study. In Proceedings of the Conference of the German Society for Computational Linguistics and Language Technology (GSCL), pages 69–79. Gunter Narr. 194 M. Elsner and E. Charniak. 2007. A generative discourse-new model for text coherence. Technical report, Technical Report CS-07-04, Brown University. K. Filippova and M. Strube. 2007a. Extending the entity-grid coherence model to semantically related entities. In Proceedings of the Eleventh European Workshop on Natural Language Generation, pages 139–142. Association for Computational Linguistics. K. Filippova and M. Strube. 2007b. Generating constituent order in German clauses. In Proc. 45th ACL, pages 320–327. K. Filippova and M. Strube. 2007c. The German Vorfeld and Local Coherence. Journal of Logic, Language and Information, 16(4):465–485. T.N. H¨ohle. 1983. Topologische Felder. Ph.D. thesis, K¨oln. J. Jacobs. 2001. The dimensions of topiccomment. Linguistics, 39(4):641–681. T. Joachims. 2002. Learning to Classify Text Using Support Vector Machines. Kluwer. N. Karamanis, C. Mellish, M. Poesio, and J. Oberlander. 2009. Evaluating centering for information ordering using corpora. Computational Linguistics, 35(1):29–46. R. Kibble and R. Power. 2004. Optimizing referential coherence in text generation. Computational Linguistics, 30(4):401–416. M. Lapata. 2003. Probabilistic text structuring: Experiments with sentence ordering. In Proc. 41st ACL, pages 545–552. M. Lapata. 2006. Automatic evaluation of information ordering: Kendall’s tau. Computational Linguistics, 32(4):471–484. V. Ng and C. Cardie. 2002. Improving machine learning approaches to coreference resolution. In Proc. 40th ACL, pages 104–111. S. Petrov and D. Klein. 2007. Improved inference for unlexicalized parsing. In Proceedings of NAACL HLT 2007, pages 404–411. M. Poesio, R. Stevenson, B.D. Eugenio, and J. Hitzeman. 2004. Centering: A parametric theory and its instantiations. Computational Linguistics, 30(3):309–363. H. Schmid and F. Laws. 2008. Estimation of conditional probabilities with decision trees and an application to fine-grained POS tagging. In Proc. 22nd COLING, pages 777–784. Association for Computational Linguistics. P. Sgall, E. Hajiˇcov´a, J. Panevov´a, and J. Mey. 1986. The meaning of the sentence in its semantic and pragmatic aspects. Springer. M. Strube and U. Hahn. 1999. Functional centering: Grounding referential coherence in information structure. Computational Linguistics, 25(3):309– 344. H. Telljohann, E. Hinrichs, and S. Kubler. 2004. The T¨uBa-D/Z treebank: Annotating German with a context-free backbone. In Proc. Fourth International Conference on Language Resources and Evaluation (LREC 2004), pages 2229–2235. Y. Versley, S.P. Ponzetto, M. Poesio, V. Eidelman, A. Jern, J. Smith, X. Yang, and A. Moschitti. 2008. BART: A modular toolkit for coreference resolution. In Proc. 46th ACL-HLT Demo Session, pages 9–12. Association for Computational Linguistics. 195
|
2010
|
20
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 196–206, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Syntactic and Semantic Factors in Processing Difficulty: An Integrated Measure Jeff Mitchell, Mirella Lapata, Vera Demberg and Frank Keller University of Edinburgh Edinburgh, United Kingdom [email protected], [email protected], [email protected], [email protected] Abstract The analysis of reading times can provide insights into the processes that underlie language comprehension, with longer reading times indicating greater cognitive load. There is evidence that the language processor is highly predictive, such that prior context allows upcoming linguistic material to be anticipated. Previous work has investigated the contributions of semantic and syntactic contexts in isolation, essentially treating them as independent factors. In this paper we analyze reading times in terms of a single predictive measure which integrates a model of semantic composition with an incremental parser and a language model. 1 Introduction Psycholinguists have long realized that language comprehension is highly incremental, with readers and listeners continuously extracting the meaning of utterances on a word-by-word basis. As soon as they encounter a word in a sentence, they integrate it as fully as possible into a representation of the sentence thus far (Marslen-Wilson 1973; Konieczny 2000; Tanenhaus et al. 1995; Sturt and Lombardo 2005). Recent research suggests that language comprehension can also be highly predictive, i.e., comprehenders are able to anticipate upcoming linguistic material. This is beneficial as it gives them more time to keep up with the input, and predictions can be used to compensate for problems with noise or ambiguity. Two types of prediction have been observed in the literature. The first type is semantic prediction, as evidenced in semantic priming: a word that is preceded by a semantically related prime or a semantically congruous sentence fragment is processed faster (Stanovich and West 1981; van Berkum et al. 1999; Clifton et al. 2007). Another example is argument prediction: listeners are able to launch eye-movements to the predicted argument of a verb before having encountered it, e.g., they will fixate an edible object as soon as they hear the word eat (Altmann and Kamide 1999). The second type of prediction is syntactic prediction. Comprehenders are faster at naming words that are syntactically compatible with prior context, even when they bear no semantic relationship to the context (Wright and Garrett 1984). Another instance of syntactic prediction has been reported by Staub and Clifton (2006): following the word either, readers predict or and the complement that follows it, and process it faster compared to a control condition without either. Thus, human language processing takes advantage of the constraints imposed by the preceding semantic and syntactic context to derive expectations about the upcoming input. Much recent work has focused on developing computational measures of these constraints and expectations. Again, the literature is split into syntactic and semantic models. Probably the best known measure of syntactic expectation is surprisal (Hale 2001) which can be coarsely defined as the negative log probability of word wt given the preceding words, typically computed using a probabilistic context-free grammar. Modeling work on semantic constraint focuses on the degree to which a word is related to its preceding context. Pynte et al. (2008) use Latent Semantic Analysis (LSA, Landauer and Dumais 1997) to assess the degree of contextual constraint exerted on a word by its context. In this framework, word meanings are represented as vectors in a high dimensional space and distance in this space is interpreted as an index of processing difficulty. Other work (McDonald and Brew 2004) models contextual constraint in information theoretic terms. The assumption is that words carry prior semantic expectations which are updated upon seeing the next word. Expectations are represented by a vector of probabilities which reflects the likely location in semantic space of the upcoming word. The measures discussed above are typically computed automatically on real-language corpora using data-driven methods and their predictions are verified through analysis of eye-movements that people make while reading. Ample evidence 196 (Rayner 1998) demonstrates that eye-movements are related to the moment-to-moment cognitive activities of readers. They also provide an accurate temporal record of the on-line processing of natural language, and through the analysis of eyemovement measurements (e.g., the amount of time spent looking at a word) can give insight into the processing difficulty involved in reading. In this paper, we investigate a model of prediction that is incremental and takes into account syntactic as well as semantic constraint. The model essentially integrates the predictions of an incremental parser (Roark 2001) together with those of a semantic space model (Mitchell and Lapata 2009). The latter creates meaning representations compositionally, and therefore builds semantic expectations for word sequences (e.g., phrases, sentences, even documents) rather than isolated words. Some existing models of sentence processing integrate semantic information into a probabilistic parser (Narayanan and Jurafsky 2002; Pad´o et al. 2009); however, the semantic component of these models is limited to semantic role information, rather than attempting to build a full semantic representation for a sentence. Furthermore, the models of Narayanan and Jurafsky (2002) and Pad´o et al. (2009) do not explicitly model prediction, but rather focus on accounting for garden path effects. The proposed model simultaneously captures semantic and syntactic effects in a single measure which we empirically show is predictive of processing difficulty as manifested in eyemovements. 2 Models of Processing Difficulty As described in Section 1, reading times provide an insight into the various cognitive activities that contribute to the overall processing difficulty involved in comprehending a written text. To quantify and understand the overall cognitive load associated with processing a word in context, we will break that load down into a sum of terms representing distinct computational costs (semantic and syntactic). For example, surprisal can be thought of as measuring the cost of dealing with unexpected input. When a word conforms to the language processor’s expectations, surprisal is low, and the cognitive load associated with processing that input will also be low. In contrast, unexpected words will have a high surprisal and a high cognitive cost. However, high-level syntactic and semantic factors are only one source of cognitive costs. A sizable proportion of the variance in reading times is accounted for by costs associated with low-level features of the stimuli, e.g.. relating to orthography and eye-movement control (Rayner 1998). In addition, there may also be costs associated with the integration of new input into an incremental representation. Dependency Locality Theory (DLT, Gibson 2000) is essentially a distance-based measure of the amount of processing effort required when the head of a phrase is integrated with its syntactic dependents. We do not consider integration costs here (as they have not been shown to correlate reliably with reading times; see Demberg and Keller 2008 for details) and instead focus on the costs associated with semantic and syntactic constraint and low-level features, which appear to make the most substantial contributions. In the following subsections we describe the various features which contribute to the processing costs of a word in context. We begin by looking at the low-level costs and move on to consider the costs associated with syntactic and semantic constraint. For readers unfamiliar with the methodology involved in modeling eye-tracking data, we note that regression analysis (or the more general mixed effects models) is typically used to study the relationship between dependent and independent variables. The independent variables are the various costs of processing effort and the dependent variables are measurements of eyemovements, three of which are routinely used in the literature: first fixation duration (the duration of the first fixation on a word regardless of whether it is the first fixation on a word or the first of multiple fixations on the same word), first pass duration, also known as gaze duration, (the sum of all fixations made on a word prior to looking at another word), and total reading time (the sum of all fixations on a word including refixations after moving on to other words). 2.1 Low-level Costs Low-level features include word frequency (more frequent words are read faster), word length (shorter words are read faster), and the position of the word in the sentence (later words are read faster). Oculomotor variables have also been found to influence reading times. These include previous fixation (indicating whether or not the previous word has been fixated), launch distance (how many characters intervene between the current fixation and the previous fixation), and landing position (which letter in the word the fixation landed on). Information about the sequential context of a word can also influence reading times. Mc197 Donald and Shillcock (2003) show that forward and backward transitional probabilities are predictive of first fixation and first pass durations: the higher the transitional probability, the shorter the fixation time. Backward transitional probability is essentially the conditional probability of a word given its immediately preceding word, P(wk|wk−1). Analogously, forward probability is the conditional probability of the current word given the next word, P(wk|wk+1). 2.2 Syntactic Constraint As mentioned earlier, surprisal (Hale 2001; Levy 2008) is one of the best known models of processing difficulty associated with syntactic constraint, and has been previously applied to the modeling of reading times (Demberg and Keller 2008; Ferrara Boston et al. 2008; Roark et al. 2009; Frank 2009). The basic idea is that the processing costs relating to the expectations of the language processor can be expressed in terms of the probabilities assigned by some form of language model to the input. These processing costs are assumed to arise from the change in the expectations of the language processor as new input arrives. If we express these expectations in terms of a distribution over all possible continuations of the input seen so far, then we can measure the magnitude of this change in terms of the Kullback-Leibler divergence of the old distribution to the updated distribution. This measure of processing cost for an input word, wk+1, given the previous context, w1 ...wk, can be expressed straightforwardly in terms of its conditional probability as: S = −logP(wk+1|w1 ...wk) (1) That is, the processing cost for a word decreases as its probability increases, with zero processing cost incurred for words which must appear in a given context, as these do not result in any change in the expectations of the language processor. The original formulation of surprisal (Hale 2001) used a probabilistic parser to calculate these probabilities, as the emphasis was on the processing costs incurred when parsing structurally ambiguous garden path sentences.1 Several variants of calculating surprisal have been developed in the literature since using different parsing strategies 1While hearing a sentence like The horse raced past the barn fell (Bever 1970), English speakers are inclined to interpreted horse as the subject of raced expecting the sentence to end at the word barn. So upon hearing the word fell they are forced to revise their analysis of the sentence thus far and adopt a reduced relative reading. (e.g., left-to-right vs. top-down, PCFGs vs dependency parsing) and different degrees of lexicalization (see Roark et al. 2009 for an overview) . For instance, unlexicalized surprisal can be easily derived by substituting the words in Equation (1) with parts of speech (Demberg and Keller 2008). Surprisal could be also defined using a vanilla language model that does not take any structural or grammatical information into account (Frank 2009). 2.3 Semantic Constraint Distributional models of meaning have been commonly used to quantify the semantic relation between a word and its context in computational studies of lexical processing. These models are based on the idea that words with similar meanings will be found in similar contexts. In putting this idea into practice, the meaning of a word is then represented as a vector in a high dimensional space, with the vector components relating to the strength on occurrence of that word in various types of context. Semantic similarities are then modeled in terms of geometric similarities within the space. To give a concrete example, Latent Semantic Analysis (LSA, Landauer and Dumais 1997) creates a meaning representation for words by constructing a word-document co-occurrence matrix from a large collection of documents. Each row in the matrix represents a word, each column a document, and each entry the frequency with which the word appeared within that document. Because this matrix tends to be quite large it is often transformed via a singular value decomposition (Berry et al. 1995) into three component matrices: a matrix of word vectors, a matrix of document vectors, and a diagonal matrix containing singular values. Re-multiplying these matrices together using only the initial portions of each (corresponding to the use of a lower dimensional spatial representation) produces a tractable approximation to the original matrix. In this framework, the similarity between two words can be easily quantified, e.g., by measuring the cosine of the angle of the vectors representing them. As LSA is one the best known semantic space models it comes as no surprise that it has been used to analyze semantic constraint. Pynte et al. (2008) measure the similarity between the next word and its preceding context under the assumption that high similarity indicates high semantic constraint (i.e., the word was expected) and analogously low similarity indicates low semantic constraint (i.e., the word was unexpected). They oper198 ationalize preceding contexts in two ways, either as the word immediately preceding the next word as the sentence fragment preceding it. Sentence fragments are represented as the average of the words they contain independently of their order. The model takes into account only content words, function words are of little interest here as they can be found in any context. Pynte et al. (2008) analyze reading times on the French part of the Dundee corpus (Kennedy and Pynte 2005) and find that word-level LSA similarities are predictive of first fixation and first pass durations, whereas sentence-level LSA is only predictive of first pass duration (i.e., for a measure that includes refixation). This latter finding is somewhat counterintuitive, one would expect longer contexts to have an immediate effect as they are presumably more constraining. One reason why sentence-level influences are only visible on first pass duration may be due to LSA itself, which is syntax-blind. Another reason relates to the way sentential context was modeled as vector addition (or averaging). The idea of averaging is not very attractive from a linguistic perspective as it blends the meanings of individual words together. Ideally, the combination of simple elements onto more complex ones must allow the construction of novel meanings which go beyond those of the individual elements (Pinker 1994). The only other model of semantic constraint we are aware of is Incremental Contextual Distinctiveness (ICD, McDonald 2000; McDonald and Brew 2004). ICD assumes that words carry prior semantic expectations which are updated upon seeing the next word. Context is represented by a vector of probabilities which reflects the likely location in semantic space of the upcoming word. When the latter is observed, the prior expectation is updated using a Bayesian inference mechanism to reflect the newly arrived information. Like LSA, ICD is based on word co-occurrence vectors, however it does not employ singular value decomposition, and constructs a word-word rather than a word-document co-occurrence matrix. Although this model has been shown to successfully simulate single- and multiple-word priming (McDonald and Brew 2004), it failed to predict processing costs in the Embra eye-tracking corpus (McDonald and Shillcock 2003). In this work we model semantic constraint using the representational framework put forward in Mitchell and Lapata (2008). Their aim is not so much to model processing difficulty, but to construct vector-based meaning representations that go beyond individual words. They introduce a general framework for studying vector composition, which they formulate as a function f of two vectors u and v: h = f(u,v) (2) where h denotes the composition of u and v. Different composition models arise, depending on how f is chosen. Assuming that h is a linear function of the Cartesian product of u and v allows to specify additive models which are by far the most common method of vector combination in the literature: hi = ui +vi (3) Alternatively, we can assume that h is a linear function of the tensor product of u and v, and thus derive models based on multiplication: hi = ui ·vi (4) Mitchell and Lapata (2008) show that several additive and multiplicative models can be formulated under this framework, including the wellknown tensor products (Smolensky 1990) and circular convolution (Plate 1995). Importantly, composition models are not defined with a specific semantic space in mind, they could easily be adapted to LSA, or simple co-occurrence vectors, or more sophisticated semantic representations (e.g., Griffiths et al. 2007), although admittedly some composition functions may be better suited for particular semantic spaces. Composition models can be straightforwardly used as predictors of processing difficulty, again via measuring the cosine of the angle between a vector w representing the upcoming word and a vector h representing the words preceding it: sim(w,h) = w·h |w||h| (5) where h is created compositionally, via some (additive or multiplicative) function f. In this paper we evaluate additive and compositional models in their ability to capture semantic prediction. We also examine the influence of the underlying meaning representations by comparing a simple semantic space similar to McDonald (2000) against Latent Dirichlet Allocation (Blei et al. 2003; Griffiths et al. 2007). Specifically, the simpler space is based on word cooccurrence counts; it constructs the vector representing a given target word, t, by identifying all the tokens of t in a corpus and recording the counts of context words, ci (within a specific window). The context words, ci, are limited to a set of the n most 199 common content words and each vector component is given by the ratio of the probability of a ci given t to the overall probability of ci. vi = p(ci|t) p(ci) (6) Despite its simplicity, the above semantic space (and variants thereof) has been used to successfully simulate lexical priming (e.g., McDonald 2000), human judgments of semantic similarity (Bullinaria and Levy 2007), and synonymy tests (Pad´o and Lapata 2007) such as those included in the Test of English as Foreign Language (TOEFL). LDA is a probabilistic topic model offering an alternative to spatial semantic representations. It is similar in spirit to LSA, it also operates on a word-document co-occurrence matrix and derives a reduced dimensionality description of words and documents. Whereas in LSA words are represented as points in a multi-dimensional space, LDA represents words using topics. Specifically, each document in a corpus is modeled as a distribution over K topics, which are themselves characterized as distribution over words. The individual words in a document are generated by repeatedly sampling a topic according to the topic distribution and then sampling a single word from the chosen topic. Under this framework, word meaning is represented as a probability distribution over a set of latent topics, essentially a vector whose dimensions correspond to topics and values to the probability of the word given these topics. Topic models have been recently gaining ground as a more structured representation of word meaning (Griffiths et al. 2007; Steyvers and Griffiths 2007). In contrast to more standard semantic space models where word senses are conflated into a single representation, topics have an intuitive correspondence to coarse-grained sense distinctions. 3 Integrating Semantic Constraint into Surprisal The treatment of semantic and syntactic constraint in models of processing difficulty has been somewhat inconsistent. While surprisal is a theoretically well-motivated measure, formalizing the idea of linguistic processing being highly predictive in terms of probabilistic language models, the measurement of semantic constraint in terms of vector similarities lacks a clear motivation. Moreover, the two approaches, surprisal and similarity, produce mathematically different types of measures. Formally, it would be preferable to have a single approach to capturing constraint and the obvious solution is to derive some form of semantic surprisal rather than sticking with similarity. This can be achieved by turning a vector model of semantic similarity into a probabilistic language model. There are in fact a number of approaches to deriving language models from distributional models of semantics (e.g., Bellegarda 2000; Coccaro and Jurafsky 1998; Gildea and Hofmann 1999). We focus here on the model of Mitchell and Lapata (2009) which tackles the issue of the composition of semantic vectors and also integrates the output of an incremental parser. The core of their model is based on the product of a trigram model p(wn|wn−1 n−2) and a semantic component ∆(wn,h) which determines the factor by which this probability should be scaled up or down given the prior semantic context h: p(wn) = p(wn|wn−1 n−2)·∆(wn,h) (7) The factor ∆(wn,h) is essentially based on a comparison between the vector representing the current word wn and the vector representing the prior history h. Varying the method for constructing word vectors (e.g., using LDA or a simpler semantic space model) and for combining them into a representation of the prior context h (e.g., using additive or multiplicative functions) produces distinct models of semantic composition. The calculation of ∆is then based on a weighted dot product of the vector representing the upcoming word w, with the vector representing the prior context h: ∆(w,h) = ∑ i wihip(ci) (8) As shown in Equation (7) this semantic factor then modulates the trigram probabilities, to take account of the effect of the semantic content outside the n-gram window. Mitchell and Lapata (2009) show that a combined semantic-trigram language model derived from this approach and trained on the Wall Street Journal outperforms a baseline trigram model in terms of perplexity on a held out set. They also linearly interpolate this semantic language model with the output of an incremental parser, which computes the following probability: p(w|h) = λp1(w|h)+(1−λ)p2(w|h) (9) where p1(w|h) is computed as in Equation (7) and p2(w|h) is computed by the parser. Their implementation uses Roark’s (2001) top-down incremental parser which estimates the probability of 200 the next word based upon the previous words of the sentence. These prefix probabilities are calculated from a grammar, by considering the likelihood of seeing the next word given the possible grammatical relations representing the prior context. Equation (9) essentially defines a language model which combines semantic, syntactic and n-gram structure, and Mitchell and Lapata (2009) demonstrate that it improves further upon a semantic language model in terms of perplexity. We argue that the probabilities from this model give us a means to model the incrementally and predictivity of the language processor in a manner that integrates both syntactic and semantic constraints. Converting these probabilities to surprisal should result in a single measure of the processing cost associated with semantic and syntactic expectations. 4 Method Data The models discussed in the previous section were evaluated against an eye-tracking corpus. Specifically, we used the English portion of the Dundee Corpus (Kennedy and Pynte 2005) which contains 20 texts taken from The Independent newspaper. The corpus consists of 51,502 tokens and 9,776 types in total. It is annotated with the eye-movement records of 10 English native speakers, who each read the whole corpus. The eye-tracking data was preprocessed following the methodology described in Demberg and Keller (2008). From this data, we computed total reading time for each word in the corpus. Our statistical analyses were based on actual reading times, and so we only included words that were not skipped. We also excluded words for which the previous word had been skipped, and words on which the normal left-to-right movement of gaze had been interrupted, i.e., by blinks, regressions, etc. Finally, because our focus is the influence of semantic context, we selected only content words whose prior sentential context contained at least two further content words. The resulting data set consisted of 53,704 data points, which is about 10% of the theoretically possible total.2 2The total of all words read by all subjects is 515,020. The pre-processing recommended by Demberg and Keller’s (2008) results in a data sets containing 436,000 data points. Removing non-content words leaves 205,922 data points. It only makes sense to consider words that were actually fixated (the eye-tracking measures used are not defined on skipped words), which leaves 162,129 data points. Following Pynte et al. (2008), we require that the previous word was fixated, with 70,051 data points remaining. We exclude words on which the normal left to right movement of gaze had been interrupted, e.g., by blinks and regressions, which results in the final total to 53,704 data points. Model Implementation All elements of our model were trained on the BLLIP corpus, a collection of texts from the Wall Street Journal (years 1987–89). The training corpus consisted of 38,521,346 words. We used a development corpus of 50,006 words and a test corpus of similar size. All words were converted to lowercase and numbers were replaced with the symbol ⟨num⟩. A vocabulary of 20,000 words was chosen and the remaining tokens were replaced with ⟨unk⟩. Following Mitchell and Lapata (2009), we constructed a simple semantic space based on cooccurrence statistics from the BLLIP training set. We used the 2,000 most frequent word types as contexts and a symmetric five word window. Vector components were defined as in Equation (6). We also trained the LDA model on BLLIP, using the Gibb’s sampling procedure discussed in Griffiths et al. (2007). We experimented with different numbers of topics on the development set (from 10 to 1,000) and report results on the test set with 100 topics. In our experiments, the hyperparameter α was initialized to .5, and the β word probabilities were initialized randomly. We integrated our compositional models with a trigram model which we also trained on BLLIP. The model was built using the SRILM toolkit (Stolcke 2002) with backoff and Kneser-Ney smoothing. As our incremental parser we used Roark’s (2001) parser trained on sections 2–21 of the Penn Treebank containing 936,017 words. The parser produces prefix probabilities for each word of a sentence which we converted to conditional probabilities by dividing each current probability by the previous one. Statistical Analysis The statistical analyses in this paper were carried out using linear mixed effects models (LME, Pinheiro and Bates 2000). The latter can be thought of as generalization of linear regression that allows the inclusion of random factors (such as participants or items) as well as fixed factors (e.g., word frequency). In our analyses, we treat participant as a random factor, which means that our models contain an intercept term for each participant, representing the individual differences in the rates at which they read.3 We evaluated the effect of adding a factor to a model by comparing the likelihoods of the models with and without that factor. If a χ2 test on the 3Other random factors that are appropriate for our analyses are word and sentence; however, due to the large number of instances for these factors (given that the Dundee corpus contains 51,502 tokens), we were not able to include them: the model fitting algorithm we used (implemented in the R package lme4) does not converge for such large models. 201 Factor Coefficient Intercept −.011 Word Length .264 Launch Distance .109 Landing Position .612 Word Frequency −.010 Reading Time of Last Word .151 Table 1: Coefficients of the baseline LME model for total reading time likelihood ratio is significant, then this indicates that the new factor significantly improves model fit. We also experimented with adding random slopes for participant to the model (in addition to the random intercept); however, this either led to non-convergence of the model fitting procedure, or failed to result in an increase in model fit according to the likelihood ratio test. Therefore, all models reported in the rest of this paper contain random intercept of participants as the sole random factor. Rather than model raw reading times, we model times on the log scale. This is desirable for a number of reasons. Firstly, the raw reading times tend to have a skew distribution and taking logs produces something closer to normal, which is preferable for modeling. Secondly, the regression equation makes more sense on the log scale as the contribution of each term to raw reading time is multiplicative rather than additive. That is, log(t) = ∑i βixi implies t = ∏i eβixi. In particular, the intercept term for each participant now represents a multiplicative factor by which that participant is slower or faster. 5 Results We computed separate mixed effects models for three dependent variables, namely first fixation duration, first pass duration, and total reading time. We report results for total times throughout, as the results of the other two dependent variables are broadly similar. Our strategy was to first construct a baseline model of low-level factors influencing reading time, and then to take the residuals from that model as the dependent variable in subsequent analyses. In this way we removed the effects of low-level factors before investigating the factors associated with syntactic and semantic constraint. This avoids problems with collinearity between low-level factors and the factors we are interested in (e.g., trigram probability is highly correlated with word frequency). The baseline model contained the factors word length, word freModel Composition Coefficient SSS Additive −.03820∗∗∗ Multiplicative −.00895∗∗∗ LDA Additive −.02500∗∗∗ Multiplicative −.00262∗∗∗ Table 2: Coefficients of LME models including simple semantic space (SSS) or Latent Dirichlet Allocation (LDA) as factors; ∗∗∗p < .001 quency, launch distance, landing position, and the reading time for the last fixated word, and its parameter estimates are given in Table 1. To further reduce collinearity, we also centered all fixed factors, both in the baseline model, and in the models fitted on the residuals that we report in the following. Note that some intercorrelations remain between the factors, which we will discuss at the end of Section 5. Before investigating whether an integrated model of semantic and syntactic constraint improves the goodness of fit over the baseline, we examined the influence of semantic constraint alone. This was necessary as compositional models have not been previously used to model processing difficulty. Besides, replicating Pynte et al.’s (2008) finding, we were also interested in assessing whether the underlying semantic representation (simple semantic space or LDA) and composition function (additive versus multiplicative) modulate reading times differentially. We built an LME model that predicted the residual reading times of the baseline model using the similarity scores from our composition models as factors. We then carried out a χ2 test on the likelihood ratio of a model only containing the random factor and the intercept, and a model also containing the semantic factor (cosine similarity). The addition of the semantic factor significantly improves model fit for both the simple semantic space and LDA. This result is observed for both additive and multiplicative composition functions. Our results are summarized in Table 2 which reports the coefficients of the four LME models fitted against the residuals of the baseline model, together with the p-values of the χ2 test. Before evaluating our integrated surprisal measure, we evaluated its components individually in order to tease their contributions apart. For example, it may be the case that syntactic surprisal is an overwhelmingly better predictor of reading time than semantic surprisal, however we would not be able to detect this by simply adding a factor based on Equation (9) to the baseline model. The 202 Factor SSS Coef LDA Coef −log(p) .00760∗∗∗ .00760∗∗∗ Add −log(∆) .03810∗∗∗ .00622∗∗∗ log(λ+(1−λ) p2 p1 ) .00953∗∗∗ .00943∗∗∗ Mult −log(∆) .01110∗∗∗−.00033 log(λ+(1−λ) p2 p1 ) .00882∗∗∗ .00133 Table 3: Coefficients of nested LME models with the components of SSS or LDA surprisal as factors; only the coefficient of the additional factor at each step are shown integrated surprisal measure can be written as: S = −log(λp1 +(1−λ)p2) (10) Where p2 is the incremental parser probability and p1 is the product of the semantic component, ∆, and the trigram probability, p. This can be broken down into the sum of two terms: S = −log(p1)−log(λ+(1−λ) p2 p1 ) (11) Since the first term, −log(p1) is itself a product it can also be broken down further: S = −log(p)−log(∆)−log(λ+(1−λ) p2 p1 ) (12) Thus, to evaluate the contribution of the three components to the integrated surprisal measure we fitted nested LME models, i.e., we entered these terms one at a time into a mixed effects model and tested the significance of the improvement in model fit for each additional term. We again start with an LME model that only contains the random factor and the intercept, with the residuals of the baseline models as the dependent variable. Considering the trigram model first, we find that adding this factor to the model gives a significant improvement in fit. Also adding the semantic component (−log(∆)) improves fit further, both for additive and multiplicative composition functions using a simple semantic space. Finally, the addition of the parser probabilities (log(λ + (1−λ) p2 p1 )) again improves model fit significantly. As far as LDA is concerned, the additive model significantly improves model fit, whereas the multiplicative one does not. These results mirror the findings of Mitchell and Lapata (2009), who report that a multiplicative composition function produced the lowest perplexity for the simple semantic space model, whereas an additive function gave the best perplexity for the LDA space. Table 3 lists the coefficients for the nested models for Model Composition Coefficient SSS Additive .00804∗∗∗ Multiplicative .00819∗∗∗ LDA Additive .00817∗∗∗ Multiplicative .00640∗∗∗ Table 4: Coefficients of LME models with integrated surprisal measure (based on SSS or LDA) as factor all four variants of our semantic constraint measure. Finally, we built a separate LME model where we added the integrated surprisal measure (see Equation (9)) to the model only containing the random factor and the intercept (see Table 4). We did this separately for all four versions of the integrated surprisal measure (SSS, LDA; additive, multiplicative). We find that model fit improved significantly all versions of integrated surprisal. One technical issue that remains to be discussed is collinearity, i.e., intercorrelations between the factors in a model. The presence of collinearity is problematic, as it can render the model fitting procedure unstable; it can also affect the significance of individual factors. As mentioned in Section 4 we used two techniques to reduce collinearity: residualizing and centering. Table 5 gives an overview of the correlation coefficients for all pairs of factors. It becomes clear that collinearity has mostly been removed; there is a remaining relationship between word length and word frequency, which is expected as shorter words tend to be more frequent. This correlation is not a problem for our analysis, as it is confined to the baseline model. Furthermore, word frequency and trigram probability are highly correlated. Again this is expected, given that the frequencies of unigrams and higher-level n-grams tend to be related. This correlation is taken care of by residualizing, which isolates the two factors: word frequency is part of the baseline model, while trigram probability is part of the separate models that we fit on the residuals. All other correlations are small (with coefficients of .27 or less), with one exception: there is a high correlation between the −log(∆) term and the log(λ + (1 −λ) p2 p1 ) term in the multiplicative LDA model. This collinearity issue may explain the absence of a significant improvement in model fit when these two terms are added to the baseline (see Table 3). 203 Factor Len Freq −l(p) −l(∆) Frequency −.310 −log(p) .230 −.700 SSS Add −log(∆) .016 −.120 .025 log(λ+(1−λ) p2 p1 ) .024 .036 −.270 .065 SSS Mult −log(∆) −.015 −.110 .035 log(λ+(1−λ) p2 p1 ) .020 .028 −.260 .160 LDA Add −log(∆) −.024 −.130 .046 log(λ+(1−λ) p2 p1 ) .005 .014 −.250 .030 LDA Mult −log(∆) −.120 .006 −.046 log(λ+(1−λ) p2 p1 )−.089 −.005 −.180 .740 Table 5: Intercorrelations between model factors 6 Discussion In this paper we investigated the contributions of syntactic and semantic constraint in modeling processing difficulty. Our work departs from previous approaches in that we propose a single measure which integrates syntactic and semantic factors. Evaluation on an eye-tracking corpus shows that our measure predicts reading time better than a baseline model that captures low-level factors in reading (word length, landing position, etc.). Crucially, we were able to show that the semantic component of our measure improves reading time predictions over and above a model that includes syntactic measures (based on a trigram model and incremental parser). This means that semantic costs are a significant predictor of reading time in addition to the well-known syntactic surprisal. An open issue is whether a single, integrated measure (as evaluated in Table 4) fits the eyemovement data significantly better than separate measures for trigram, syntactic, and semantic surprisal (as evaluated in Table 3. However, we are not able to investigate this hypothesis: our approach to testing the significance of factors requires nested models; the log-likelihood test (see Section 4) is only able to establish whether adding a factor to a model improves its fit; it cannot compare models with disjunct sets of factors (such as a model containing the integrated surprisal measure and one containing the three separate ones). However, we would argue that a single, integrated measure that captures human predictive processing is preferable over a collection of separate measures. It is conceptually simpler (as it is more parsimonious), and is also easier to use in applications (such as readability prediction). Finally, an integrated measure requires less parameters; our definition of surprisal in 12 is simply the sum of the trigram, syntactic, and semantic components. An LME model containing separate factors, on the other hand, requires a coefficient for each of them, and thus has more parameters. In evaluating our model, we adopted a broad coverage approach using the reading time data from a naturalistic corpus rather than artificially constructed experimental materials. In doing so, we were able to compare different syntactic and semantic costs on the same footing. Previous analyses of semantic constraint have been conducted on different eye-tracking corpora (Dundee and Embra Corpus) and on different languages (English, French). Moreover, comparisons of the individual contributions of syntactic and semantic factors were generally absent from the literature. Our analysis showed that both of these factors can be captured by our integrated surprisal measure which is uniformly probabilistic and thus preferable to modeling semantic and syntactic costs disjointly using a mixture of probabilistic and nonprobabilistic measures. An interesting question is which aspects of semantics our model is able to capture, i.e., why does the combination of LSA or LDA representations with an incremental parser yield a better fit of the behavioral data. In the psycholinguistic literature, various types of semantic information have been investigated: lexical semantics (word senses, selectional restrictions, thematic roles), sentential semantics (scope, binding), and discourse semantics (coreference and coherence); see Keller (2010) of a detailed discussion. We conjecture that our model is mainly capturing lexical semantics (through the vector space representation of words) and sentential semantics (through the multiplication or addition of words). However, discourse coreference effects (such as the ones reported by Altmann and Steedman (1988) and much subsequent work) are probably not amenable to a treatment in terms of vector space semantics; an explicit representation of discourse entities and coreference relations is required (see Dubey 2010 for a model of human sentence processing that can handle coreference). A key objective for future work will be to investigate models that integrate semantic constraint with syntactic predictions more tightly. For example, we could envisage a parser that uses semantic representations to guide its search, e.g., by pruning syntactic analyses that have a low semantic probability. At the same time, the semantic model should have access to syntactic information, i.e., the composition of word representations should take their syntactic relationships into account, rather than just linear order. 204 References ACL. 2010. Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics. Uppsala. Altmann, Gerry T. M. and Yuki Kamide. 1999. Incremental interpretation at verbs: Restricting the domain of subsequent reference. Cognition 73:247–264. Altmann, Gerry T. M. and Mark J. Steedman. 1988. Interaction with context during human sentence processing. Cognition 30(3):191–238. Bellegarda, Jerome R. 2000. Exploiting latent semantic information in statistical language modeling. Proceedings of the IEEE 88(8):1279– 1296. Berry, Michael W., Susan T. Dumais, and Gavin W. O’Brien. 1995. Using linear algebra for intelligent information retrieval. SIAM review 37(4):573–595. Bever, Thomas G. 1970. The cognitive basis for linguistic strutures. In J. R. Hayes, editor, Cognition and the Development of Language, Wiley, New York, pages 279–362. Blei, David M., Andrew Y. Ng, and Michael I. Jordan. 2003. Latent Dirichlet allocation. Journal of Machine Learning Research 3:993–1022. Bullinaria, John A. and Joseph P. Levy. 2007. Extracting semantic representations from word cooccurrence statistics: A computational study. Behavior Research Methods 39:510–526. Clifton, Charles, Adrian Staub, and Keith Rayner. 2007. Eye movement in reading words and sentences. In R V Gompel, M Fisher, W Murray, and R L Hill, editors, Eye Movements: A Window in Mind and Brain, Elsevier, pages 341– 372. Coccaro, Noah and Daniel Jurafsky. 1998. Towards better integration of semantic predictors in satistical language modeling. In Proceedings of the 5th International Conference on Spoken Language Processing. Sydney, Australia, pages 2403–2406. Demberg, Vera and Frank Keller. 2008. Data from eye-tracking corpora as evidence for theories of syntactic processing complexity. Cognition 101(2):193–210. Dubey, Amit. 2010. The influence of discourse on syntax: A psycholinguistic model of sentence processing. In ACL. Ferrara Boston, Marisa, John Hale, Reinhold Kliegl, Umesh Patil, and Shravan Vasishth. 2008. Parsing costs as predictors of reading difficulty: An evaluation using the Potsdam Sentence Corpus. Journal of Eye Movement Research 2(1):1–12. Frank, Stefan L. 2009. Surprisal-based comparison between a symbolic and a connectionist model of sentence processing. In Proceedings of the 31st Annual Conference of the Cognitive Science Society. Austin, TX, pages 139–1144. Gibson, Edward. 2000. Dependency locality theory: A distance-dased theory of linguistic complexity. In Alec Marantz, Yasushi Miyashita, and Wayne O’Neil, editors, Image, Language, Brain: Papers from the First Mind Articulation Project Symposium, MIT Press, Cambridge, MA, pages 95–126. Gildea, Daniel and Thomas Hofmann. 1999. Topic-based language models using EM. In Proceedings of the 6th European Conference on Speech Communiation and Technology. Budapest, Hungary, pages 2167–2170. Griffiths, Thomas L., Mark Steyvers, and Joshua B. Tenenbaum. 2007. Topics in semantic representation. Psychological Review 114(2):211–244. Hale, John. 2001. A probabilistic Earley parser as a psycholinguistic model. In Proceedings of the 2nd Conference of the North American Chapter of the Association. Association for Computational Linguistics, Pittsburgh, PA, volume 2, pages 159–166. Keller, Frank. 2010. Cognitively plausible models of human language processing. In ACL. Kennedy, Alan and Joel Pynte. 2005. Parafovealon-foveal effects in normal reading. Vision Research 45:153–168. Konieczny, Lars. 2000. Locality and parsing complexity. Journal of Psycholinguistic Research 29(6):627–645. Landauer, Thomas K. and Susan T. Dumais. 1997. A solution to Plato’s problem: the latent semantic analysis theory of acquisition, induction and representation of knowledge. Psychological Review 104(2):211–240. Levy, Roger. 2008. Expectation-based syntactic comprehension. Cognition 106(3):1126–1177. Marslen-Wilson, William D. 1973. Linguistic structure and speech shadowing at very short latencies. Nature 244:522–523. McDonald, Scott. 2000. Environmental Determinants of Lexical Processing Effort. Ph.D. thesis, University of Edinburgh. 205 McDonald, Scott and Chris Brew. 2004. A distributional model of semantic context effects in lexical processing. In Proceedings of the 42th Annual Meeting of the Association for Computational Linguistics. Barcelona, Spain, pages 17–24. McDonald, Scott A. and Richard C. Shillcock. 2003. Low-level predictive inference in reading: The influence of transitional probabilities on eye movements. Vision Research 43:1735– 1751. Mitchell, Jeff and Mirella Lapata. 2008. Vectorbased models of semantic composition. In Proceedings of ACL-08: HLT. Columbus, OH, pages 236–244. Mitchell, Jeff and Mirella Lapata. 2009. Language models based on semantic composition. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing. Singapore, pages 430–439. Narayanan, Srini and Daniel Jurafsky. 2002. A Bayesian model predicts human parse preference and reading time in sentence processing. In Thomas G. Dietterich, Sue Becker, and Zoubin Ghahramani, editors, Advances in Neural Information Processing Systems 14. MIT Press, Cambridge, MA, pages 59–65. Pad´o, Sebastian and Mirella Lapata. 2007. Dependency-based construction of semantic space models. Computational Linguistics 33(2):161–199. Pad´o, Ulrike, Matthew W. Crocker, and Frank Keller. 2009. A probabilistic model of semantic plausibility in sentence processing. Cognitive Science 33(5):794–838. Pinheiro, Jose C. and Douglas M. Bates. 2000. Mixed-effects Models in S and S-PLUS. Springer, New York. Pinker, Steven. 1994. The Language Instinct: How the Mind Creates Language. HarperCollins, New York. Plate, Tony A. 1995. Holographic reduced representations. IEEE Transactions on Neural Networks 6(3):623–641. Pynte, Joel, Boris New, and Alan Kennedy. 2008. On-line contextual influences during reading normal text: A multiple-regression analysis. Vision Research 48:2172–2183. Rayner, Keith. 1998. Eye movements in reading and information processing: 20 years of research. Psychological Bulletin 124(3):372–422. Roark, Brian. 2001. Probabilistic top-down parsing and language modeling. Computational Linguistics 27(2):249–276. Roark, Brian, Asaf Bachrach, Carlos Cardenas, and Christophe Pallier. 2009. Deriving lexical and syntactic expectation-based measures for psycholinguistic modeling via incremental top-down parsing. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, Singapore, pages 324–333. Smolensky, Paul. 1990. Tensor product variable binding and the representation of symbolic structures in connectionist systems. Artificial Intelligence 46:159–216. Stanovich, Kieth E. and Richard F. West. 1981. The effect of sentence context on ongoing word recognition: Tests of a two-pricess theory. Journal of Experimental Psychology: Human Perception and Performance 7:658–672. Staub, Adrian and Charles Clifton. 2006. Syntactic prediction in language comprehension: Evidence from either ...or. Journal of Experimental Psychology: Learning, Memory, and Cognition 32:425–436. Steyvers, Mark and Tom Griffiths. 2007. Probabilistic topic models. In T. Landauer, D. McNamara, S Dennis, and W Kintsch, editors, A Handbook of Latent Semantic Analysis, Psychology Press. Stolcke, Andreas. 2002. Srilm - an extensible language modeling toolkit. In Proceedings of the Internatinal Conference on Spoken Language Processing. Denver, Colorado. Sturt, Patrick and Vincenzo Lombardo. 2005. Processing coordinated structures: Incrementality and connectedness. Cognitive Science 29(2):291–305. Tanenhaus, Michael K., Michael J. SpiveyKnowlton, Kathleen M. Eberhard, and Julie C. Sedivy. 1995. Integration of visual and linguistic information in spoken language comprehension. Science 268:1632–1634. van Berkum, Jos J. A., Colin M. Brown, and Peter Hagoort. 1999. Early referential context effects in sentence processing: Evidence from eventrelated brain potentials. Journal of Memory and Language 41:147–182. Wright, Barton and Merrill F. Garrett. 1984. Lexical decision in sentences: Effects of syntactic structure. Memory and Cognition 12:31–45. 206
|
2010
|
21
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 207–215, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Rebanking CCGbank for improved NP interpretation Matthew Honnibal and James R. Curran School of Information Technologies University of Sydney NSW 2006, Australia {mhonn,james}@it.usyd.edu.au Johan Bos University of Groningen The Netherlands [email protected] Abstract Once released, treebanks tend to remain unchanged despite any shortcomings in their depth of linguistic analysis or coverage of specific phenomena. Instead, separate resources are created to address such problems. In this paper we show how to improve the quality of a treebank, by integrating resources and implementing improved analyses for specific constructions. We demonstrate this rebanking process by creating an updated version of CCGbank that includes the predicate-argument structure of both verbs and nouns, baseNP brackets, verb-particle constructions, and restrictive and non-restrictive nominal modifiers; and evaluate the impact of these changes on a statistical parser. 1 Introduction Progress in natural language processing relies on direct comparison on shared data, discouraging improvements to the evaluation data. This means that we often spend years competing to reproduce partially incorrect annotations. It also encourages us to approach related problems as discrete tasks, when a new data set that adds deeper information establishes a new incompatible evaluation. Direct comparison has been central to progress in statistical parsing, but it has also caused problems. Treebanking is a difficult engineering task: coverage, cost, consistency and granularity are all competing concerns that must be balanced against each other when the annotation scheme is developed. The difficulty of the task means that we ought to view treebanking as an ongoing process akin to grammar development, such as the many years of work on the ERG (Flickinger, 2000). This paper demonstrates how a treebank can be rebanked to incorporate novel analyses and information from existing resources. We chose to work on CCGbank (Hockenmaier and Steedman, 2007), a Combinatory Categorial Grammar (Steedman, 2000) treebank acquired from the Penn Treebank (Marcus et al., 1993). This work is equally applicable to the corpora described by Miyao et al. (2004), Shen et al. (2008) or Cahill et al. (2008). Our first changes integrate four previously suggested improvements to CCGbank. We then describe a novel CCG analysis of NP predicateargument structure, which we implement using NomBank (Meyers et al., 2004). Our analysis allows the distinction between core and peripheral arguments to be represented for predicate nouns. With this distinction, an entailment recognition system could recognise that Google’s acquisition of YouTube entailed Google acquired YouTube, because equivalent predicate-argument structures are built for both. Our analysis also recovers nonlocal dependencies mediated by nominal predicates; for instance, Google is the agent of acquire in Google’s decision to acquire YouTube. The rebanked corpus extends CCGbank with: 1. NP brackets from Vadas and Curran (2008); 2. Restored and normalised punctuation; 3. Propbank-derived verb subcategorisation; 4. Verb particle structure drawn from Propbank; 5. Restrictive and non-restrictive adnominals; 6. Reanalyses to promote better head-finding; 7. Nombank-derived noun subcategorisation. Together, these changes modify 30% of the labelled dependencies in CCGbank, demonstrating how multiple resources can be brought together in a single, richly annotated corpus. We then train and evaluate a parser for these changes, to investigate their impact on the accuracy of a state-of-theart statistical CCG parser. 207 2 Background and motivation Formalisms like HPSG (Pollard and Sag, 1994), LFG (Kaplan and Bresnan, 1982), and CCG (Steedman, 2000) are linguistically motivated in the sense that they attempt to explain and predict the limited variation found in the grammars of natural languages. They also attempt to specify how grammars construct semantic representations from surface strings, which is why they are sometimes referred to as deep grammars. Analyses produced by these formalisms can be more detailed than those produced by skeletal phrasestructure parsers, because they produce fully specified predicate-argument structures. Unfortunately, statistical parsers do not take advantage of this potential detail. Statistical parsers induce their grammars from corpora, and the corpora for linguistically motivated formalisms currently do not contain high quality predicateargument annotation, because they were derived from the Penn Treebank (PTB Marcus et al., 1993). Manually written grammars for these formalisms, such as the ERG HPSG grammar (Flickinger, 2000) and the XLE LFG grammar (Butt et al., 2006) produce far more detailed and linguistically correct analyses than any English statistical parser, due to the comparatively coarse-grained annotation schemes of the corpora statistical parsers are trained on. While rule-based parsers use grammars that are carefully engineered (e.g. Oepen et al., 2004), and can be updated to reflect the best linguistic analyses, statistical parsers have so far had to take what they are given. What we suggest in this paper is that a treebank’s grammar need not last its lifetime. For a start, there have been many annotations of the PTB that add much of the extra information needed to produce very high quality analyses for a linguistically motivated grammar. There are also other transformations which can be made with no additional information. That is, sometimes the existing trees allow transformation rules to be written that improve the quality of the grammar. Linguistic theories are constantly changing, which means that there is a substantial lag between what we (think we) understand of grammar and the annotations in our corpora. The grammar engineering process we describe, which we dub rebanking, is intended to reduce this gap, tightening the feedback loop between formal and computational linguistics. 2.1 Combinatory Categorial Grammar Combinatory Categorial Grammar (CCG; Steedman, 2000) is a lexicalised grammar, which means that all grammatical dependencies are specified in the lexical entries and that the production of derivations is governed by a small set of rules. Lexical categories are either atomic (S, NP, PP, N ), or a functor consisting of a result, directional slash, and argument. For instance, in might head a PP-typed constituent with one NP-typed argument, written as PP/NP. A category can have a functor as its result, so that a word can have a complex valency structure. For instance, a verb phrase is represented by the category S\NP: it is a function from a leftward NP (a subject) to a sentence. A transitive verb requires an object to become a verb phrase, producing the category (S\NP)/NP. A CCG grammar consists of a small number of schematic rules, called combinators. CCG extends the basic application rules of pure categorial grammar with (generalised) composition rules and type raising. The most common rules are: X /Y Y ⇒ X (>) Y X \Y ⇒ X (<) X /Y Y /Z ⇒X /Z (>B) Y \Z X \Y ⇒X \Z (<B) Y /Z X \Y ⇒X /Z (<B×) CCGbank (Hockenmaier and Steedman, 2007) extends this compact set of combinatory rules with a set of type-changing rules, designed to strike a better balance between sparsity in the category set and ambiguity in the grammar. We mark typechanging rules TC in our derivations. In wide-coverage descriptions, categories are generally modelled as typed-feature structures (Shieber, 1986), rather than atomic symbols. This allows the grammar to include a notion of headedness, and to unify under-specified features. We occasionally must refer to these additional details, for which we employ the following notation. Features are annotated in square-brackets, e.g. S[dcl]. Head-finding indices are annotated on categories in subscripts, e.g. (NPy\NPy)/NPz. The index of the word the category is assigned to is left implicit. We will sometimes also annotate derivations with the heads of categories as they are being built, to help the reader keep track of what lexemes have been bound to which categories. 208 3 Combining CCGbank corrections There have been a few papers describing corrections to CCGbank. We bring these corrections together for the first time, before building on them with our further changes. 3.1 Compound noun brackets Compound noun phrases can nest inside each other, creating bracketing ambiguities: (1) (crude oil) prices (2) crude (oil prices) The structure of such compound noun phrases is left underspecified in the Penn Treebank (PTB), because the annotation procedure involved stitching together partial parses produced by the Fidditch parser (Hindle, 1983), which produced flat brackets for these constructions. The bracketing decision was also a source of annotator disagreement (Bies et al., 1995). When Hockenmaier and Steedman (2002) went to acquire a CCG treebank from the PTB, this posed a problem. There is no equivalent way to leave these structures under-specified in CCG, because derivations must be binary branching. They therefore employed a simple heuristic: assume all such structures branch to the right. Under this analysis, crude oil is not a constituent, producing an incorrect analysis as in (1). Vadas and Curran (2007) addressed this by manually annotating all of the ambiguous noun phrases in the PTB, and went on to use this information to correct 20,409 dependencies (1.95%) in CCGbank (Vadas and Curran, 2008). Our changes build on this corrected corpus. 3.2 Punctuation corrections The syntactic analysis of punctuation is notoriously difficult, and punctuation is not always treated consistently in the Penn Treebank (Bies et al., 1995). Hockenmaier (2003) determined that quotation marks were particularly problematic, and therefore removed them from CCGbank altogether. We use the process described by Tse and Curran (2008) to restore the quotation marks and shift commas so that they always attach to the constituent to their left. This allows a grammar rule to be removed, preventing a great deal of spurious ambiguity and improving the speed of the C&C parser (Clark and Curran, 2007) by 37%. 3.3 Verb predicate-argument corrections Semantic role descriptions generally recognise a distinction between core arguments, whose role comes from a set specific to the predicate, and peripheral arguments, who have a role drawn from a small, generic set. This distinction is represented in the surface syntax in CCG, because the category of a verb must specify its argument structure. In (3) as a director is annotated as a complement; in (4) it is an adjunct: (3) He NP joined (S\NP)/PP as a director PP (4) He NP joined S\NP as a director (S\NP)\(S\NP) CCGbank contains noisy complement and adjunct distinctions, because they were drawn from PTB function labels which imperfectly represent the distinction. In our previous work we used Propbank (Palmer et al., 2005) to convert 1,543 complements to adjuncts and 13,256 adjuncts to complements (Honnibal and Curran, 2007). If a constituent such as as a director received an adjunct category, but was labelled as a core argument in Propbank, we changed it to a complement, using its head’s part-of-speech tag to infer its constituent type. We performed the equivalent transformation to ensure all peripheral arguments of verbs were analysed as adjuncts. 3.4 Verb-particle constructions Propbank also offers reliable annotation of verbparticle constructions. This was not available in the PTB, so Hockenmaier and Steedman (2007) annotated all intransitive prepositions as adjuncts: (5) He NP woke S\NP up (S\NP)\(S\NP) We follow Constable and Curran (2009) in exploiting the Propbank annotations to add verbparticle distinctions to CCGbank, by introducing a new atomic category PT for particles, and changing their status from adjuncts to complements: (6) He NP woke (S\NP)/PT up PT This analysis could be improved by adding extra head-finding logic to the verbal category, to recognise the multi-word expression as the head. 209 Rome ′s gift of peace to Europe NP (NP/(N /PP))\NP (N /PP)/PP)/PP PP/NP NP PP/NP NP < > > N /(N /PP) PP PP > (N /PP)/PP > N /PP > NP Figure 1: Deverbal noun predicate with agent, patient and beneficiary arguments. 4 Noun predicate-argument structure Many common nouns in English can receive optional complements and adjuncts, realised by prepositional phrases, genitive determiners, compound nouns, relative clauses, and for some nouns, complementised clauses. For example, deverbal nouns generally have argument structures similar to the verbs they are derived from: (7) Rome’s destruction of Carthage (8) Rome destroyed Carthage The semantic roles of Rome and Carthage are the same in (7) and (8), but the noun cannot casemark them directly, so of and the genitive clitic are pressed into service. The semantic role depends on both the predicate and subcategorisation frame: (9) Carthage’sp destructionPred. (10) Rome’sa destructionPred. of Carthagep (11) Rome’sa giftPred. (12) Rome’sa giftPred. of peacep to Europeb In (9), the genitive introduces the patient, but when the patient is supplied by the PP, it instead introduces the agent. The mapping differs for gift, where the genitive introduces the agent. Peripheral arguments, which supply generically available modifiers of time, place, cause, quality etc, can be realised by pre- and post-modifiers: (13) The portrait in the Louvre (14) The fine portrait (15) The Louvre’s portraits These are distinct from core arguments because their interpretation does not depend on the predicate. The ambiguity can be seen in an NP such as The nobleman’s portrait, where the genitive could mark possession (peripheral), or it could introduce the patient (core). The distinction between core and peripheral arguments is particularly difficult for compound nouns, as pre-modification is very productive in English. 4.1 CCG analysis We designed our analysis for transparency between the syntax and the predicate-argument structure, by stipulating that all and only the core arguments should be syntactic arguments of the predicate’s category. This is fairly straightforward for arguments introduced by prepositions: destruction of Carthage N /PPy PPy/NPy NP > PPCarthage > Ndestruction In our analysis, the head of of Carthage is Carthage, as of is assumed to be a semantically transparent case-marker. We apply this analysis to prepositional phrases that provide arguments to verbs as well — a departure from CCGbank. Prepositional phrases that introduce peripheral arguments are analysed as syntactic adjuncts: The war in 149 B.C. NPy/Ny N (Ny\Ny)/NPz NP > (Ny\Ny)in < Nwar > NPwar Adjunct prepositional phrases remain headed by the preposition, as it is the preposition’s semantics that determines whether they function as temporal, causal, spatial etc. arguments. We follow Hockenmaier and Steedman (2007) in our analysis of genitives which realise peripheral arguments, such as the literal possessive: Rome ′s aqueducts NP (NPy/Ny)\NPz N < (NPy/Ny)′s > NPaqueducts Arguments introduced by possessives are a little trickier, because the genitive also functions as a determiner. We achieve this by having the noun subcategorise for the argument, which we type PP, and having the possessive subcategorise for the unsaturated noun to ultimately produce an NP: 210 Google ′s decision to buy YouTube NP (NPy/(Ny/PPz)y)\NPz (N /PPy)/(S[to]z\NPy)z (S[to]y\NPz)y/(S[b]y\NPz)y (S[b]\NPy)/NPz NP < > NPy/(Ny/PPGoogle)y S[b]\NPy >B > NPdecision/(S[to]y\NPGoogle)y S[to]buy\NPy > NP Figure 2: The coindexing on decision’s category allows the hard-to-reach agent of buy to be recovered. A non-normal form derivation is shown so that instantiated variables can be seen. Carthage ′s destruction NP (NPy/(Ny/PPz)y)\NPz N /PPy < (NPy/(Ny/PPCarthage)y)′s > NPdestruction In this analysis, we regard the genitive clitic as a case-marker that performs a movement operation roughly analogous to WH-extraction. Its category is therefore similar to the one used in object extraction, (N \N )/(S/NP). Figure 1 shows an example with multiple core arguments. This analysis allows recovery of verbal arguments of nominalised raising and control verbs, a construction which both Gildea and Hockenmaier (2003) and Boxwell and White (2008) identify as a problem case when aligning Propbank and CCGbank. Our analysis accommodates this construction effortlessly, as shown in Figure 2. The category assigned to decision can coindex the missing NP argument of buy with its own PP argument. When that argument is supplied by the genitive, it is also supplied to the verb, buy, filling its dependency with its agent, Google. This argument would be quite difficult to recover using a shallow syntactic analysis, as the path would be quite long. There are 494 such verb arguments mediated by nominal predicates in Sections 02-21. These analyses allow us to draw complement/adjunct distinctions for nominal predicates, so that the surface syntax takes us very close to a full predicate-argument analysis. The only information we are not specifying in the syntactic analysis are the role labels assigned to each of the syntactic arguments. We could go further and express these labels in the syntax, producing categories like (N /PP{0}y)/PP{1}z and (N /PP{1}y)/PP{0}z, but we expect that this would cause sparse data problems given the limited size of the corpus. This experiment would be an interesting subject of future work. The only local core arguments that we do not annotate as syntactic complements are compound nouns, such as decision makers. We avoided these arguments because of the productivity of nounnoun compounding in English, which makes these argument structures very difficult to recover. We currently do not have an analysis that allows support verbs to supply noun arguments, so we do not recover any of the long-range dependency structures described by Meyers et al. (2004). 4.2 Implementation and statistics Our analysis requires semantic role labels for each argument of the nominal predicates in the Penn Treebank — precisely what NomBank (Meyers et al., 2004) provides. We can therefore draw our distinctions using the process described in our previous work, Honnibal and Curran (2007). NomBank follows the same format as Propbank, so the procedure is exactly the same. First, we align CCGbank and the Penn Treebank, and produce a version of NomBank that refers to CCGbank nodes. We then assume that any prepositional phrase or genitive determiner annotated as a core argument in NomBank should be analysed as a complement, while peripheral arguments and adnominals that receive no semantic role label at all are analysed as adjuncts. We converted 34,345 adnominal prepositional phrases to complements, leaving 18,919 as adjuncts. The most common preposition converted was of, which was labelled as a core argument 99.1% of the 19,283 times it occurred as an adnominal. The most common adjunct preposition was in, which realised a peripheral argument in 59.1% of its 7,725 occurrences. The frequent prepositions were more skewed towards core arguments. 73% of the occurrences of the 5 most frequent prepositions (of, in, for, on and to) realised peripheral arguments, compared with 53% for other prepositions. Core arguments were also more common than peripheral arguments for possessives. There are 20,250 possessives in the corpus, of which 75% were converted to complements. The percentage was similar for both personal pronouns (such as his) and genitive phrases (such as the boy’s). 211 5 Adding restrictivity distinctions Adnominals can have either a restrictive or a nonrestrictive (appositional) interpretation, determining the potential reference of the noun phrase it modifies. This ambiguity manifests itself in whether prepositional phrases, relative clauses and other adnominals are analysed as modifiers of either N or NP, yielding a restrictive or nonrestrictive interpretation respectively. In CCGbank, all adnominals attach to NPs, producing non-restrictive interpretations. We therefore move restrictive adnominals to N nodes: All staff on casual contracts NP/N N (N \N )/NP N /N N > N TC NP > N \N < N > NP This corrects the previous interpretation, which stated that there were no permanent staff. 5.1 Implementation and statistics The Wall Street Journal’s style guide mandates that this attachment ambiguity be managed by bracketing non-restrictive relatives with commas (Martin, 2002, p. 82), as in casual staff, who have no health insurance, support it. We thus use punctuation to make the attachment decision. All NP\NP modifiers that are not preceded by punctuation were moved to the lowest N node possible and relabelled N \N . We select the lowest (i.e. closest to leaf) N node because some adjectives, such as present or former, require scope over the qualified noun, making it safer to attach the adnominal first. Some adnominals in CCGbank are created by the S\NP →NP\NP unary type-changing rule, which transforms reduced relative clauses. We introduce a S\NP →N \N in its place, and add a binary rule cued by punctuation to handle the relatively rare non-restrictive reduced relative clauses. The rebanked corpus contains 34,134 N \N restrictive modifiers, and 9,784 non-restrictive modifiers. Most (61%) of the non-restrictive modifiers were relative clauses. 6 Reanalysing partitive constructions True partitive constructions consist of a quantifier (16), a cardinal (17) or demonstrative (18) applied to an NP via of. There are similar constructions headed by common nouns, as in (19): (16) Some of us (17) Four of our members (18) Those of us who smoke (19) A glass of wine We regard the common noun partitives as headed by the initial noun, such as glass, because this noun usually controls the number agreement. We therefore analyse these cases as nouns with prepositional arguments. In (19), glass would be assigned the category N /PP. True partitive constructions are different, however: they are always headed by the head of the NP supplied by of. The construction is quite common, because it provides a way to quantify or apply two different determiners. Partitive constructions are not given special treatment in the PTB, and were analysed as noun phrases with a PP modifier in CCGbank: Four of our members NP (NPy\NPy)/NPz NPy/Ny N > NPmembers > (NPy\NPy)of < NPFour This analysis does not yield the correct semantics, and may even hurt parser performance, because the head of the phrase is incorrectly assigned. We correct this with the following analysis, which takes the head from the NP argument of the PP: Four of our members NPy/PPy PPy/NPy NPy/Ny N > NPmembers > PPmembers > NPmembers The cardinal is given the category NP/PP, in analogy with the standard determiner category which is a function from a noun to a noun phrase (NP/N ). 212 Corpus L. DEPS U. DEPS CATS +NP brackets 97.2 97.7 98.5 +Quotes 97.2 97.7 98.5 +Propbank 93.0 94.9 96.7 +Particles 92.5 94.8 96.2 +Restrictivity 79.5 94.4 90.6 +Part. Gen. 76.1 90.1 90.4 +NP Pred-Arg 70.6 83.3 84.8 Table 1: Effect of the changes on CCGbank, by percentage of dependencies and categories left unchanged in Section 00. 6.1 Implementation and Statistics We detect this construction by identifying NPs post-modified by an of PP. The NP’s head must either have the POS tag CD, or be one of the following words, determined through manual inspection of Sections 02-21: all, another, average, both, each, another, any, anything, both, certain, each, either, enough, few, little, most, much, neither, nothing, other, part, plenty, several, some, something, that, those. Having identified the construction, we simply relabel the NP to NP/PP, and the NP\NP adnominal to PP. We identified and reanalysed 3,010 partitive genitives in CCGbank. 7 Similarity to CCGbank Table 1 shows the percentage of labelled dependencies (L. Deps), unlabelled dependencies (U. Deps) and lexical categories (Cats) that remained the same after each set of changes. A labelled dependency is a 4-tuple consisting of the head, the argument, the lexical category of the head, and the argument slot that the dependency fills. For instance, the subject fills slot 1 and the object fills slot 2 on the transitive verb category (S\NP)/NP. There are more changes to labelled dependencies than lexical categories because one lexical category change alters all of the dependencies headed by a predicate, as they all depend on its lexical category. Unlabelled dependencies consist of only the head and argument. The biggest changes were those described in Sections 4 and 5. After the addition of nominal predicate-argument structure, over 50% of the labelled dependencies were changed. Many of these changes involved changing an adjunct to a complement, which affects the unlabelled dependencies because the head and argument are inverted. 8 Lexicon statistics Our changes make the grammar sensitive to new distinctions, which increases the number of lexical categories required. Table 2 shows the number Corpus CATS Cats ≥10 CATS/WORD CCGbank 1286 425 8.6 +NP brackets 1298 429 8.9 +Quotes 1300 431 8.8 +Propbank 1342 433 8.9 +Particles 1405 458 9.1 +Restrictivity 1447 471 9.3 +Part. Gen. 1455 474 9.5 +NP Pred-Arg 1574 511 10.1 Table 2: Effect of the changes on the size of the lexicon. of lexical categories (Cats), the number of lexical categories that occur at least 10 times in Sections 02-21 (Cats ≥10), and the average number of categories available for assignment to each token in Section 00 (Cats/Word). We followed Clark and Curran’s (2007) process to determine the set of categories a word could receive, which includes a part-of-speech back-off for infrequent words. The lexicon steadily grew with each set of changes, because each added information to the corpus. The addition of quotes only added two categories (LQU and RQU ), and the addition of the quote tokens slightly decreased the average categories per word. The Propbank and verb-particle changes both introduced rare categories for complicated, infrequent argument structures. The NP predicate-argument structure modifications added the most information. Head nouns were previously guaranteed the category N in CCGbank; possessive clitics always received the category (NP/N )\NP; and possessive personal pronouns were always NP/N . Our changes introduce new categories for these frequent tokens, which meant a substantial increase in the number of possible categories per word. 9 Parsing Evaluation Some of the changes we have made correct problems that have caused the performance of a statistical CCG parser to be over-estimated. Other changes introduce new distinctions, which a parser may or may not find difficult to reproduce. To investigate these issues, we trained and evaluated the C&C CCG parser on our rebanked corpora. The experiments were set up as follows. We used the highest scoring configuration described by Clark and Curran (2007), the hybrid dependency model, using gold-standard POS tags. We followed Clark and Curran in excluding sentences that could not be parsed from the evaluation. All models obtained similar coverage, between 99.0 and 99.3%. The parser was evaluated using depen213 WSJ 00 WSJ 23 Corpus LF UF CAT LF UF CAT CCGbank 87.2 92.9 94.1 87.7 93.0 94.4 +NP brackets 86.9 92.8 93.8 87.3 92.8 93.9 +Quotes 86.8 92.7 93.9 87.1 92.6 94.0 +Propbank 86.7 92.6 94.0 87.0 92.6 94.0 +Particles 86.4 92.5 93.8 86.8 92.6 93.8 All Rebanking 84.2 91.2 91.9 84.7 91.3 92.2 Table 3: Parser evaluation on the rebanked corpora. Corpus Rebanked CCGbank LF UF LF UF +NP brackets 86.45 92.36 86.52 92.35 +Quotes 86.57 92.40 86.52 92.35 +Propbank 87.76 92.96 87.74 92.99 +Particles 87.50 92.77 87.67 92.93 All Rebanking 87.23 92.71 88.02 93.51 Table 4: Comparison of parsers trained on CCGbank and the rebanked corpora, using dependencies that occur in both. dencies generated from the gold-standard derivations (Boxwell, p.c., 2010). Table 3 shows the accuracy of the parser on Sections 00 and 23. The parser scored slightly lower as the NP brackets, Quotes, Propbank and Particles corrections were added. This apparent decline in performance is at least partially an artefact of the evaluation. CCGbank contains some dependencies that are trivial to recover, because Hockenmaier and Steedman (2007) was forced to adopt a strictly right-branching analysis for NP brackets. There was a larger drop in accuracy on the fully rebanked corpus, which included our analyses of restrictivity, partitive constructions and noun predicate-argument structure. This might also be explained by the evaluation, as the rebanked corpus includes much more fine-grained distinctions. The labelled dependencies evaluation is particularly sensitive to this, as a single category change affects multiple dependencies. This can be seen in the smaller gap in category accuracy. We investigated whether the differences in performance were due to the different evaluation data by comparing the parsers’ performance against the original parser on the dependencies they agreed upon, to allow direct comparison. To do this, we extracted the CCGbank intersection of each corpus’s Section 00 dependencies. Table 4 compares the labelled and unlabelled recall of the rebanked parsers we trained against the CCGbank parser on these intersections. Note that each row refers to a different intersection, so results are not comparable between rows. This comparison shows that the declines in accuracy seen in Table 3 were largely confined to the corrected dependencies. The parser’s performance remained fairly stable on the dependencies left unchanged. The rebanked parser performed 0.8% worse than the CCGbank parser on the intersection dependencies, suggesting that the fine-grained distinctions we introduced did cause some sparse data problems. However, we did not change any of the parser’s maximum entropy features or hyperparameters, which are tuned for CCGbank. 10 Conclusion Research in natural language understanding is driven by the datasets that we have available. The most cited computational linguistics work to date is the Penn Treebank (Marcus et al., 1993)1. Propbank (Palmer et al., 2005) has also been very influential since its release, and NomBank has been used for semantic dependency parsing in the CoNLL 2008 and 2009 shared tasks. This paper has described how these resources can be jointly exploited using a linguistically motivated theory of syntax and semantics. The semantic annotations provided by Propbank and NomBank allowed us to build a corpus that takes much greater advantage of the semantic transparency of a deep grammar, using careful analyses and phenomenon-specific conversion rules. The major areas of CCGbank’s grammar left to be improved are the analysis of comparatives, and the analysis of named entities. English comparatives are diverse and difficult to analyse. Even the XTAG grammar (Doran et al., 1994), which deals with the major constructions of English in enviable detail, does not offer a full analysis of these phenomena. Named entities are also difficult to analyse, as many entity types obey their own specific grammars. This is another example of a phenomenon that could be analysed much better in CCGbank using an existing resource, the BBN named entity corpus. Our rebanking has substantially improved CCGbank, by increasing the granularity and linguistic fidelity of its analyses. We achieved this by exploiting existing resources and crafting novel analyses. The process we have demonstrated can be used to train a parser that returns dependencies that abstract away as much surface syntactic variation as possible — including, now, even whether the predicate and arguments are expressed in a noun phrase or a full clause. 1http://clair.si.umich.edu/clair/anthology/rankings.cgi 214 Acknowledgments James Curran was supported by Australian Research Council Discovery grant DP1097291 and the Capital Markets Cooperative Research Centre. The parsing evaluation for this paper would have been much more difficult without the assistance of Stephen Boxwell, who helped generate the gold-standard dependencies with his software. We are also grateful to the members of the CCG technicians mailing list for their help crafting the analyses, particularly Michael White, Mark Steedman and Dennis Mehay. References Ann Bies, Mark Ferguson, Karen Katz, and Robert MacIntyre. 1995. Bracketing guidelines for Treebank II style Penn Treebank project. Technical report, MS-CIS-95-06, University of Pennsylvania, Philadelphia, PA, USA. Stephen Boxwell and Michael White. 2008. Projecting propbank roles onto the CCGbank. In Proceedings of the Sixth International Language Resources and Evaluation (LREC’08), pages 3112–3117. European Language Resources Association (ELRA), Marrakech, Morocco. Miriam Butt, Mary Dalrymple, and Tracy H. King, editors. 2006. Lexical Semantics in LFG. CSLI Publications, Stanford, CA. Aoife Cahill, Michael Burke, Ruth O’Donovan, Stefan Riezler, Josef van Genabith, and Andy Way. 2008. Widecoverage deep statistical parsing using automatic dependency structure annotation. Computational Linguistics, 34(1):81–124. Stephen Clark and James R. Curran. 2007. Wide-coverage efficient statistical parsing with CCG and log-linear models. Computational Linguistics, 33(4):493–552. James Constable and James Curran. 2009. Integrating verbparticle constructions into CCG parsing. In Proceedings of the Australasian Language Technology Association Workshop 2009, pages 114–118. Sydney, Australia. Christy Doran, Dania Egedi, Beth Ann Hockey, B. Srinivas, and Martin Zaidel. 1994. Xtag system: a wide coverage grammar for english. In Proceedings of the 15th conference on Computational linguistics, pages 922–928. ACL, Morristown, NJ, USA. Dan Flickinger. 2000. On building a more efficient grammar by exploiting types. Natural Language Engineering, 6(1):15–28. Daniel Gildea and Julia Hockenmaier. 2003. Identifying semantic roles using combinatory categorial grammar. In Proceedings of the 2003 conference on Empirical methods in natural language processing, pages 57–64. ACL, Morristown, NJ, USA. Donald Hindle. 1983. User manual for fidditch, a deterministic parser. Technical Memorandum 7590-142, Naval Research Laboratory. Julia Hockenmaier. 2003. Data and Models for Statistical Parsing with Combinatory Categorial Grammar. Ph.D. thesis, University of Edinburgh, Edinburgh, UK. Julia Hockenmaier and Mark Steedman. 2002. Acquiring compact lexicalized grammars from a cleaner treebank. In Proceedings of the Third Conference on Language Resources and Evaluation Conference, pages 1974–1981. Las Palmas, Spain. Julia Hockenmaier and Mark Steedman. 2007. CCGbank: a corpus of CCG derivations and dependency structures extracted from the Penn Treebank. Computational Linguistics, 33(3):355–396. Matthew Honnibal and James R. Curran. 2007. Improving the complement/adjunct distinction in CCGBank. In Proceedings of the Conference of the Pacific Association for Computational Linguistics, pages 210–217. Melbourne, Australia. Ronald M. Kaplan and Joan Bresnan. 1982. LexicalFunctional Grammar: A formal system for grammatical representation. In Joan Bresnan, editor, The mental representation of grammatical relations, pages 173–281. MIT Press, Cambridge, MA, USA. Mitchell Marcus, Beatrice Santorini, and Mary Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19(2):313–330. Paul Martin. 2002. The Wall Street Journal Guide to Business Style and Usage. Free Press, New York. Adam Meyers, Ruth Reeves, Catherine Macleod, Rachel Szekely, Veronika Zielinska, Brian Young, and Ralph Grishman. 2004. The NomBank project: An interim report. In Frontiers in Corpus Annotation: Proceedings of the Workshop, pages 24–31. Boston, MA, USA. Yusuke Miyao, Takashi Ninomiya, and Jun’ichi Tsujii. 2004. Corpus-oriented grammar development for acquiring a head-driven phrase structure grammar from the Penn Treebank. In Proceedings of the First International Joint Conference on Natural Language Processing (IJCNLP-04), pages 684–693. Hainan Island, China. Stepan Oepen, Daniel Flickenger, Kristina Toutanova, and Christopher D. Manning. 2004. LinGO Redwoods. a rich and dynamic treebank for HPSG. Research on Language and Computation, 2(4):575–596. Martha Palmer, Daniel Gildea, and Paul Kingsbury. 2005. The proposition bank: An annotated corpus of semantic roles. Computational Linguistics, 31(1):71–106. Carl Pollard and Ivan Sag. 1994. Head-Driven Phrase Structure Grammar. The University of Chicago Press, Chicago. Libin Shen, Lucas Champollion, and Aravind K. Joshi. 2008. LTAG-spinal and the treebank: A new resource for incremental, dependency and semantic parsing. Language Resources and Evaluation, 42(1):1–19. Stuart M. Shieber. 1986. An Introduction to UnificationBased Approaches to Grammar, volume 4 of CSLI Lecture Notes. CSLI Publications, Stanford, CA. Mark Steedman. 2000. The Syntactic Process. The MIT Press, Cambridge, MA, USA. Daniel Tse and James R. Curran. 2008. Punctuation normalisation for cleaner treebanks and parsers. In Proceedings of the Australian Language Technology Workshop, volume 6, pages 151–159. ALTW, Hobart, Australia. David Vadas and James Curran. 2007. Adding noun phrase structure to the Penn Treebank. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 240–247. ACL, Prague, Czech Republic. David Vadas and James R. Curran. 2008. Parsing noun phrase structure with CCG. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics, pages 335–343. ACL, Columbus, Ohio, USA. 215
|
2010
|
22
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 216–225, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics BabelNet: Building a Very Large Multilingual Semantic Network Roberto Navigli Dipartimento di Informatica Sapienza Universit`a di Roma [email protected] Simone Paolo Ponzetto Department of Computational Linguistics Heidelberg University [email protected] Abstract In this paper we present BabelNet – a very large, wide-coverage multilingual semantic network. The resource is automatically constructed by means of a methodology that integrates lexicographic and encyclopedic knowledge from WordNet and Wikipedia. In addition Machine Translation is also applied to enrich the resource with lexical information for all languages. We conduct experiments on new and existing gold-standard datasets to show the high quality and coverage of the resource. 1 Introduction In many research areas of Natural Language Processing (NLP) lexical knowledge is exploited to perform tasks effectively. These include, among others, text summarization (Nastase, 2008), Named Entity Recognition (Bunescu and Pas¸ca, 2006), Question Answering (Harabagiu et al., 2000) and text categorization (Gabrilovich and Markovitch, 2006). Recent studies in the difficult task of Word Sense Disambiguation (Navigli, 2009b, WSD) have shown the impact of the amount and quality of lexical knowledge (Cuadros and Rigau, 2006): richer knowledge sources can be of great benefit to both knowledge-lean systems (Navigli and Lapata, 2010) and supervised classifiers (Ng and Lee, 1996; Yarowsky and Florian, 2002). Various projects have been undertaken to make lexical knowledge available in a machine readable format. A pioneering endeavor was WordNet (Fellbaum, 1998), a computational lexicon of English based on psycholinguistic theories. Subsequent projects have also tackled the significant problem of multilinguality. These include EuroWordNet (Vossen, 1998), MultiWordNet (Pianta et al., 2002), the Multilingual Central Repository (Atserias et al., 2004), and many others. However, manual construction methods inherently suffer from a number of drawbacks. First, maintaining and updating lexical knowledge resources is expensive and time-consuming. Second, such resources are typically lexicographic, and thus contain mainly concepts and only a few named entities. Third, resources for non-English languages often have a much poorer coverage since the construction effort must be repeated for every language of interest. As a result, an obvious bias exists towards conducting research in resource-rich languages, such as English. A solution to these issues is to draw upon a large-scale collaborative resource, namely Wikipedia1. Wikipedia represents the perfect complement to WordNet, as it provides multilingual lexical knowledge of a mostly encyclopedic nature. While the contribution of any individual user might be imprecise or inaccurate, the continual intervention of expert contributors in all domains results in a resource of the highest quality (Giles, 2005). But while a great deal of work has been recently devoted to the automatic extraction of structured information from Wikipedia (Wu and Weld, 2007; Ponzetto and Strube, 2007; Suchanek et al., 2008; Medelyan et al., 2009, inter alia), the knowledge extracted is organized in a looser way than in a computational lexicon such as WordNet. In this paper, we make a major step towards the vision of a wide-coverage multilingual knowledge resource. We present a novel methodology that produces a very large multilingual semantic network: BabelNet. This resource is created by linking Wikipedia to WordNet via an automatic mapping and by integrating lexical gaps in resource1http://download.wikipedia.org. We use the English Wikipedia database dump from November 3, 2009, which includes 3,083,466 articles. Throughout this paper, we use Sans Serif for words, SMALL CAPS for Wikipedia pages and CAPITALS for Wikipedia categories. 216 high wind blow gas gasbag wind hot-air balloon gas cluster ballooning Montgolfier brothers Fermi gas is-a has-part is-a is-a Wikipedia WordNet balloon BABEL SYNSET balloonEN, BallonDE, aerostatoES, globusCA, pallone aerostaticoIT, ballonFR, montgolfi`ereFR WIKIPEDIA SENTENCES ...world’s first hydrogen balloon flight. ...an interim balloon altitude record... ...from a British balloon near B´ecourt... + SEMCOR SENTENCES ...look at the balloon and the... ...suspended like a huge balloon, in... ...the balloon would go up... Machine Translation system Figure 1: An illustrative overview of BabelNet. poor languages with the aid of Machine Translation. The result is an “encyclopedic dictionary”, that provides concepts and named entities lexicalized in many languages and connected with large amounts of semantic relations. 2 BabelNet We encode knowledge as a labeled directed graph G = (V, E) where V is the set of vertices – i.e. concepts2 such as balloon – and E ⊆V ×R×V is the set of edges connecting pairs of concepts. Each edge is labeled with a semantic relation from R, e.g. {is-a, part-of, . . . , ϵ}, where ϵ denotes an unspecified semantic relation. Importantly, each vertex v ∈V contains a set of lexicalizations of the concept for different languages, e.g. { balloonEN, BallonDE, aerostatoES, . . . , montgolfi`ereFR }. Concepts and relations in BabelNet are harvested from the largest available semantic lexicon of English, WordNet, and a wide-coverage collaboratively edited encyclopedia, the English Wikipedia (Section 3.1). We collect (a) from WordNet, all available word senses (as concepts) and all the semantic pointers between synsets (as relations); (b) from Wikipedia, all encyclopedic entries (i.e. pages, as concepts) and semantically unspecified relations from hyperlinked text. In order to provide a unified resource, we merge the intersection of these two knowledge sources (i.e. their concepts in common) by establishing a mapping between Wikipedia pages and WordNet senses (Section 3.2). This avoids duplicate concepts and allows their inventories of concepts to complement each other. Finally, to enable multilinguality, we collect the lexical realizations of the available concepts in different languages by 2Throughout the paper, unless otherwise stated, we use the general term concept to denote either a concept or a named entity. using (a) the human-generated translations provided in Wikipedia (the so-called inter-language links), as well as (b) a machine translation system to translate occurrences of the concepts within sense-tagged corpora, namely SemCor (Miller et al., 1993) – a corpus annotated with WordNet senses – and Wikipedia itself (Section 3.3). We call the resulting set of multilingual lexicalizations of a given concept a babel synset. An overview of BabelNet is given in Figure 1 (we label vertices with English lexicalizations): unlabeled edges are obtained from links in the Wikipedia pages (e.g. BALLOON (AIRCRAFT) links to WIND), whereas labeled ones from WordNet3 (e.g. balloon1 n haspart gasbag1 n). In this paper we restrict ourselves to concepts lexicalized as nouns. Nonetheless, our methodology can be applied to all parts of speech, but in that case Wikipedia cannot be exploited, since it mainly contains nominal entities. 3 Methodology 3.1 Knowledge Resources WordNet. The most popular lexical knowledge resource in the field of NLP is certainly WordNet, a computational lexicon of the English language. A concept in WordNet is represented as a synonym set (called synset), i.e. the set of words that share the same meaning. For instance, the concept wind is expressed by the following synset: { wind1 n, air current1 n, current of air1 n }, where each word’s subscripts and superscripts indicate their parts of speech (e.g. n stands for noun) 3We use in the following WordNet version 3.0. We denote with wi p the i-th sense of a word w with part of speech p. We use word senses to unambiguously denote the corresponding synsets (e.g. plane1 n for { airplane1 n, aeroplane1 n, plane1 n }). Hereafter, we use word sense and synset interchangeably. 217 and sense number, respectively. For each synset, WordNet provides a textual definition, or gloss. For example, the gloss of the above synset is: “air moving from an area of high pressure to an area of low pressure”. Wikipedia. Our second resource, Wikipedia, is a Web-based collaborative encyclopedia. A Wikipedia page (henceforth, Wikipage) presents the knowledge about a specific concept (e.g. BALLOON (AIRCRAFT)) or named entity (e.g. MONTGOLFIER BROTHERS). The page typically contains hypertext linked to other relevant Wikipages. For instance, BALLOON (AIRCRAFT) is linked to WIND, GAS, and so on. The title of a Wikipage (e.g. BALLOON (AIRCRAFT)) is composed of the lemma of the concept defined (e.g. balloon) plus an optional label in parentheses which specifies its meaning if the lemma is ambiguous (e.g. AIRCRAFT vs. TOY). Wikipages also provide inter-language links to their counterparts in other languages (e.g. BALLOON (AIRCRAFT) links to the Spanish page AEROSTATO). Finally, some Wikipages are redirections to other pages, e.g. the Spanish BAL ´ON AEROST ´ATICO redirects to AEROSTATO. 3.2 Mapping Wikipedia to WordNet The first phase of our methodology aims to establish links between Wikipages and WordNet senses. We aim to acquire a mapping µ such that, for each Wikipage w, we have: µ(w) = s ∈SensesWN(w) if a link can be established, ϵ otherwise, where SensesWN(w) is the set of senses of the lemma of w in WordNet. For example, if our mapping methodology linked BALLOON (AIRCRAFT) to the corresponding WordNet sense balloon1 n, we would have µ(BALLOON (AIRCRAFT)) = balloon1 n. In order to establish a mapping between the two resources, we first identify the disambiguation contexts for Wikipages (Section 3.2.1) and WordNet senses (Section 3.2.2). Next, we intersect these contexts to perform the mapping (see Section 3.2.3). 3.2.1 Disambiguation Context of a Wikipage Given a Wikipage w, we use the following information as disambiguation context: • Sense labels: e.g. given the page BALLOON (AIRCRAFT), the word aircraft is added to the disambiguation context. • Links: the titles’ lemmas of the pages linked from the target Wikipage (i.e., outgoing links). For instance, the links in the Wikipage BALLOON (AIRCRAFT) include wind, gas, etc. • Categories: Wikipages are typically classified according to one or more categories. For example, the Wikipage BALLOON (AIRCRAFT) is categorized as BALLOONS, BALLOONING, etc. While many categories are very specific and do not appear in WordNet (e.g., SWEDISH WRITERS or SCIENTISTS WHO COMMITTED SUICIDE), we use their syntactic heads as disambiguation context (i.e. writer and scientist, respectively). Given a Wikipage w, we define its disambiguation context Ctx(w) as the set of words obtained from all of the three sources above. 3.2.2 Disambiguation Context of a WordNet Sense Given a WordNet sense s and its synset S, we collect the following information: • Synonymy: all synonyms of s in S. For instance, given the sense airplane1 n and its corresponding synset { airplane1 n, aeroplane1 n, plane1 n }, the words contained therein are included in the context. • Hypernymy/Hyponymy: all synonyms in the synsets H such that H is either a hypernym (i.e., a generalization) or a hyponym (i.e., a specialization) of S. For example, given balloon1 n, we include the words from its hypernym { lighter-than-air craft1 n } and all its hyponyms (e.g. { hot-air balloon1 n }). • Sisterhood: words from the sisters of S. A sister synset S′ is such that S and S′ have a common direct hypernym. For example, given balloon1 n, it can be found that { balloon1 n } and { airship1 n, dirigible1 n } are sisters. Thus airship and dirigible are included in the disambiguation context of s. • Gloss: the set of lemmas of the content words occurring within the WordNet gloss of S. We thus define the disambiguation context Ctx(s) of sense s as the set of words obtained from all of the four sources above. 218 3.2.3 Mapping Algorithm In order to link each Wikipedia page to a WordNet sense, we perform the following steps: • Initially, our mapping µ is empty, i.e. it links each Wikipage w to ϵ. • For each Wikipage w whose lemma is monosemous both in Wikipedia and WordNet we map w to its only WordNet sense. • For each remaining Wikipage w for which no mapping was previously found (i.e., µ(w) = ϵ), we assign the most likely sense to w based on the maximization of the conditional probabilities p(s|w) over the senses s ∈SensesWN(w) (no mapping is established if a tie occurs). To find the mapping of a Wikipage w, we need to compute the conditional probability p(s|w) of selecting the WordNet sense s given w. The sense s which maximizes this probability is determined as follows: µ(w) = argmax s∈SensesWN(w) p(s|w) = argmax s p(s, w) p(w) = argmax s p(s, w) The latter formula is obtained by observing that p(w) does not influence our maximization, as it is a constant independent of s. As a result, determining the most appropriate sense s consists of finding the sense s that maximizes the joint probability p(s, w). We estimate p(s, w) as: p(s, w) = score(s, w) X s′∈SensesWN(w), w′∈SensesWiki(w) score(s′, w′) , where score(s, w) = |Ctx(s) ∩Ctx(w)| + 1 (we add 1 as a smoothing factor). Thus, in our algorithm we determine the best sense s by computing the intersection of the disambiguation contexts of s and w, and normalizing by the scores summed over all senses of w in Wikipedia and WordNet. More details on the mapping algorithm can be found in Ponzetto and Navigli (2010). 3.3 Translating Babel Synsets So far we have linked English Wikipages to WordNet senses. Given a Wikipage w, and provided it is mapped to a sense s (i.e., µ(w) = s), we create a babel synset S ∪W, where S is the WordNet synset to which sense s belongs, and W includes: (i) w; (ii) all its inter-language links (that is, translations of the Wikipage to other languages); (iii) the redirections to the inter-language links found in the Wikipedia of the target language. For instance, given that µ(BALLOON) = balloon1 n, the corresponding babel synset is { balloonEN, BallonDE, aerostatoES, bal´on aerost´aticoES, ..., pallone aerostaticoIT }. However, two issues arise: first, a concept might be covered only in one of the two resources (either WordNet or Wikipedia), meaning that no link can be established (e.g., FERMI GAS or gasbag1 n in Figure 1); second, even if covered in both resources, the Wikipage for the concept might not provide any translation for the language of interest (e.g., the Catalan for BALLOON is missing in Wikipedia). In order to address the above issues and thus guarantee high coverage for all languages we developed a methodology for translating senses in the babel synset to missing languages. Given a WordNet word sense in our babel synset of interest (e.g. balloon1 n) we collect its occurrences in SemCor (Miller et al., 1993), a corpus of more than 200,000 words annotated with WordNet senses. We do the same for Wikipages by retrieving sentences in Wikipedia with links to the Wikipage of interest. By repeating this step for each English lexicalization in a babel synset, we obtain a collection of sentences for the babel synset (see left part of Figure 1). Next, we apply state-of-the-art Machine Translation4 and translate the set of sentences in all the languages of interest. Given a specific term in the initial babel synset, we collect the set of its translations. We then identify the most frequent translation in each language and add it to the babel synset. Note that translations are sensespecific, as the context in which a term occurs is provided to the translation system. 3.4 Example We now illustrate the execution of our methodology by way of an example. Let us focus on the Wikipage BALLOON (AIRCRAFT). The word is polysemous both in Wikipedia and WordNet. In the first phase of our methodology we aim to find a mapping µ(BALLOON (AIRCRAFT)) to an appropriate WordNet sense of the word. To 4We use the Google Translate API. An initial prototype used a statistical machine translation system based on Moses (Koehn et al., 2007) and trained on Europarl (Koehn, 2005). However, we found such system unable to cope with many technical names, such as in the domains of sciences, literature, history, etc. 219 this end we construct the disambiguation context for the Wikipage by including words from its label, links and categories (cf. Section 3.2.1). The context thus includes, among others, the following words: aircraft, wind, airship, lighter-thanair. We now construct the disambiguation context for the two WordNet senses of balloon (cf. Section 3.2.2), namely the aircraft (#1) and the toy (#2) senses. To do so, we include words from their synsets, hypernyms, hyponyms, sisters, and glosses. The context for balloon1 n includes: aircraft, craft, airship, lighter-than-air. The context for balloon2 n contains: toy, doll, hobby. The sense with the largest intersection is #1, so the following mapping is established: µ(BALLOON (AIRCRAFT)) = balloon1 n. After the first phase, our babel synset includes the following English words from WordNet plus the Wikipedia interlanguage links to other languages (we report German, Spanish and Italian): { balloonEN, BallonDE, aerostatoES, bal´on aerost´aticoES, pallone aerostaticoIT }. In the second phase (see Section 3.3), we collect all the sentences in SemCor and Wikipedia in which the above English word sense occurs. We translate these sentences with the Google Translate API and select the most frequent translation in each language. As a result, we can enrich the initial babel synset with the following words: mongolfi`ereFR, globusCA, globoES, mongolfieraIT. Note that we had no translation for Catalan and French in the first phase, because the inter-language link was not available, and we also obtain new lexicalizations for the Spanish and Italian languages. 4 Experiment 1: Mapping Evaluation Experimental setting. We first performed an evaluation of the quality of our mapping from Wikipedia to WordNet. To create a gold standard for evaluation we considered all lemmas whose senses are contained both in WordNet and Wikipedia: the intersection between the two resources contains 80,295 lemmas which correspond to 105,797 WordNet senses and 199,735 Wikipedia pages. The average polysemy is 1.3 and 2.5 for WordNet senses and Wikipages, respectively (2.8 and 4.7 when excluding monosemous words). We then selected a random sample of 1,000 Wikipages and asked an annotator with previous experience in lexicographic annotaP R F1 A Mapping algorithm 81.9 77.5 79.6 84.4 MFS BL 24.3 47.8 32.2 24.3 Random BL 23.8 46.8 31.6 23.9 Table 1: Performance of the mapping algorithm. tion to provide the correct WordNet sense for each page (an empty sense label was given, if no correct mapping was possible). The gold-standard dataset includes 505 non-empty mappings, i.e. Wikipages with a corresponding WordNet sense. In order to quantify the quality of the annotations and the difficulty of the task, a second annotator sense tagged a subset of 200 pages from the original sample. Our annotators achieved a κ inter-annotator agreement (Carletta, 1996) of 0.9, indicating almost perfect agreement. Results and discussion. Table 1 summarizes the performance of our mapping algorithm against the manually annotated dataset. Evaluation is performed in terms of standard measures of precision, recall, and F1-measure. In addition we calculate accuracy, which also takes into account empty sense labels. As baselines we use the most frequent WordNet sense (MFS), and a random sense assignment. The results show that our method achieves almost 80% F1 and it improves over the baselines by a large margin. The final mapping contains 81,533 pairs of Wikipages and word senses they map to, covering 55.7% of the noun senses in WordNet. As for the baselines, the most frequent sense is just 0.6% and 0.4% above the random baseline in terms of F1 and accuracy, respectively. A χ2 test reveals in fact no statistical significant difference at p < 0.05. This is related to the random distribution of senses in our dataset and the Wikipedia unbiased coverage of WordNet senses. So selecting the first WordNet sense rather than any other sense for each target page represents a choice as arbitrary as picking a sense at random. 5 Experiment 2: Translation Evaluation We perform a second set of experiments concerning the quality of the acquired concepts. This is assessed in terms of coverage against gold-standard resources (Section 5.1) and against a manuallyvalidated dataset of translations (Section 5.2). 220 Language Word senses Synsets German 15,762 9,877 Spanish 83,114 55,365 Catalan 64,171 40,466 Italian 57,255 32,156 French 44,265 31,742 Table 2: Size of the gold-standard wordnets. 5.1 Automatic Evaluation Datasets. We compare BabelNet against goldstandard resources for 5 languages, namely: the subset of GermaNet (Lemnitzer and Kunze, 2002) included in EuroWordNet for German, MultiWordNet (Pianta et al., 2002) for Italian, the Multilingual Central Repository for Spanish and Catalan (Atserias et al., 2004), and WOrdnet Libre du Franc¸ais (Benoˆıt and Fiˇser, 2008, WOLF) for French. In Table 2 we report the number of synsets and word senses available in the gold-standard resources for the 5 languages. Measures. Let B be BabelNet, F our goldstandard non-English wordnet (e.g. GermaNet), and let E be the English WordNet. All the goldstandard non-English resources, as well as BabelNet, are linked to the English WordNet: given a synset SF ∈F, we denote its corresponding babel synset as SB and its synset in the English WordNet as SE. We assess the coverage of BabelNet against our gold-standard wordnets both in terms of synsets and word senses. For synsets, we calculate coverage as follows: SynsetCov(B, F) = P SF∈F δ(SB, SF) |{SF ∈F}| , where δ(SB, SF) = 1 if the two synsets SB and SF have a synonym in common, 0 otherwise. That is, synset coverage is determined as the percentage of synsets of F that share a term with the corresponding babel synsets. For word senses we calculate a similar measure of coverage: WordCov(B, F) = P SF∈F P sF∈SF δ′(sF, SB) |{sF ∈SF : SF ∈F}| , where sF is a word sense in synset SF and δ′(sF, SB) = 1 if sF ∈SB, 0 otherwise. That is we calculate the ratio of word senses in our gold-standard resource F that also occur in the corresponding synset SB to the overall number of senses in F. However, our gold-standard resources cover only a portion of the English WordNet, whereas the overall coverage of BabelNet is much higher. We calculate extra coverage for synsets as follows: SynsetExtraCov(B, F) = P SE∈E\F δ(SB, SE) |{SF ∈F}| . Similarly, we calculate extra coverage for word senses in BabelNet corresponding to WordNet synsets not covered by the reference resource F. Results and discussion. We evaluate the coverage and extra coverage of word senses and synsets at different stages: (a) using only the interlanguage links from Wikipedia (WIKI Links); (b) and (c) using only the automatic translations of the sentences from Wikipedia (WIKI Transl.) or SemCor (WN Transl.); (d) using all available translations, i.e. BABELNET. Coverage results are reported in Table 3. The percentage of word senses covered by BabelNet ranges from 52.9% (Italian) to 66.4 (Spanish) and 86.0% (French). Synset coverage ranges from 73.3% (Catalan) to 76.6% (Spanish) and 92.9% (French). As expected, synset coverage is higher, because a synset in the reference resource is considered to be covered if it shares at least one word with the corresponding synset in BabelNet. Numbers for the extra coverage, which provides information about the percentage of word senses and synsets in BabelNet but not in the goldstandard resources, are given in Figure 2. The results show that we provide for all languages a high extra coverage for both word senses – between 340.1% (Catalan) and 2,298% (German) – and synsets – between 102.8% (Spanish) and 902.6% (German). Table 3 and Figure 2 show that the best results are obtained when combining all available translations, i.e. both from Wikipedia and the machine translation system. The performance figures suffer from the errors of the mapping phase (see Section 4). Nonetheless, the results are generally high, with a peak for French, since WOLF has been created semi-automatically by combining several resources, including Wikipedia. The relatively low word sense coverage for Italian (55.4%) is, instead, due to the lack of many common words in the Italian gold-standard synsets. Examples include whipEN translated as staffileIT but not as the more common frustaIT, playboyEN translated as vitaioloIT but not gigol`oIT, etc. 221 0%
500%
1000%
1500%
2000%
2500%
German
Spanish
Catalan
Italian
French
Wiki
Links
Wiki
Transl.
WN
Transl.
BabelNet
(a) word senses 0%
100%
200%
300%
400%
500%
600%
700%
800%
900%
1000%
German
Spanish
Catalan
Italian
French
Wiki
Links
Wiki
Transl.
WN
Transl.
BabelNet
(b) synsets Figure 2: Extra coverage against gold-standard wordnets: word senses (a) and synsets (b). Resource Method SENSES SYNSETS German WIKI n Links 39.6 50.7 Transl. 42.6 58.2 WN Transl. 21.0 28.6 BABELNET All 57.6 73.4 Spanish WIKI n Links 34.4 40.7 Transl. 47.9 56.1 WN Transl. 25.2 30.0 BABELNET All 66.4 76.6 Catalan WIKI n Links 20.3 25.2 Transl. 46.9 54.1 WN Transl. 25.0 29.6 BABELNET All 64.0 73.3 Italian WIKI n Links 28.1 40.0 Transl. 39.9 58.0 WN Transl. 19.7 28.7 BABELNET All 52.9 73.7 French WIKI n Links 70.0 72.4 Transl. 69.6 79.6 WN Transl. 16.3 19.4 BABELNET All 86.0 92.9 Table 3: Coverage against gold-standard wordnets (we report percentages). 5.2 Manual Evaluation Experimental setup. The automatic evaluation quantifies how much of the gold-standard resources is covered by BabelNet. However, it does not say anything about the precision of the additional lexicalizations provided by BabelNet. Given that our resource has displayed a remarkably high extra coverage – ranging from 340% to 2,298% of the national wordnets (see Figure 2) – we performed a second evaluation to assess its precision. For each of our 5 languages, we selected a random set of 600 babel synsets composed as follows: 200 synsets whose senses exist in WordNet only, 200 synsets in the intersection between WordNet and Wikipedia (i.e. those mapped with our method illustrated in Section 3.2), 200 synsets whose lexicalizations exist in Wikipedia only. Therefore, our dataset included 600 × 5 = 3,000 babel synsets. None of the synsets was covered by any of the five reference wordnets. The babel synsets were manually validated by expert annotators who decided which senses (i.e. lexicalizations) were appropriate given the corresponding WordNet gloss and/or Wikipage. Results and discussion. We report the results in Table 4. For each language (rows) and for each of the three regions of BabelNet (columns), we report precision (i.e. the percentage of synonyms deemed correct) and, in parentheses, the overall number of synonyms evaluated. The results show that the different regions of BabelNet contain translations of different quality: while on average translations for WordNet-only synsets have a precision around 72%, when Wikipedia comes into play the performance increases considerably (around 80% in the intersection and 95% with Wikipedia-only translations). As can be seen from the figures in parentheses, the number of translations available in the presence of Wikipedia is higher. This quantitative difference is due to our method collecting many translations from the redirections in the Wikipedia of the target language (Section 3.3), as well as to the paucity of examples in SemCor for many synsets. In addition, some of the synsets in WordNet with no Wikipedia counterpart are very difficult to translate. Examples include terms like stammel, crape fern, baseball clinic, and many others for which we could 222 Language WN WN ∩Wiki Wiki German 73.76 (282) 78.37 (777) 97.74 (709) Spanish 69.45 (275) 78.53 (643) 92.46 (703) Catalan 75.58 (258) 82.98 (517) 92.71 (398) Italian 72.32 (271) 80.83 (574) 99.09 (552) French 67.16 (268) 77.43 (709) 96.44 (758) Table 4: Precision of BabelNet on synonyms in WordNet (WN), Wikipedia (Wiki) and their intersection (WN ∩Wiki): percentage and total number of words (in parentheses) are reported. not find translations in major editions of bilingual dictionaries. In contrast, good translations were produced using our machine translation method when enough sentences were available. Examples are: chaudr´ee de poissonFR for fish chowderEN, grano de caf´eES for coffee beanEN, etc. 6 Related Work Previous attempts to manually build multilingual resources have led to the creation of a multitude of wordnets such as EuroWordNet (Vossen, 1998), MultiWordNet (Pianta et al., 2002), BalkaNet (Tufis¸ et al., 2004), Arabic WordNet (Black et al., 2006), the Multilingual Central Repository (Atserias et al., 2004), bilingual electronic dictionaries such as EDR (Yokoi, 1995), and fullyfledged frameworks for the development of multilingual lexicons (Lenci et al., 2000). As it is often the case with manually assembled resources, these lexical knowledge repositories are hindered by high development costs and an insufficient coverage. This barrier has led to proposals that acquire multilingual lexicons from either parallel text (Gale and Church, 1993; Fung, 1995, inter alia) or monolingual corpora (Sammer and Soderland, 2007; Haghighi et al., 2008). The disambiguation of bilingual dictionary glosses has also been proposed to create a bilingual semantic network from a machine readable dictionary (Navigli, 2009a). Recently, Etzioni et al. (2007) and Mausam et al. (2009) presented methods to produce massive multilingual translation dictionaries from Web resources such as online lexicons and Wiktionaries. However, while providing lexical resources on a very large scale for hundreds of thousands of language pairs, these do not encode semantic relations between concepts denoted by their lexical entries. The research closest to ours is presented by de Melo and Weikum (2009), who developed a Universal WordNet (UWN) by automatically acquiring a semantic network for languages other than English. UWN is bootstrapped from WordNet and is built by collecting evidence extracted from existing wordnets, translation dictionaries, and parallel corpora. The result is a graph containing 800,000 words from over 200 languages in a hierarchically structured semantic network with over 1.5 million links from words to word senses. Our work goes one step further by (1) developing an even larger multilingual resource including both lexical semantic and encyclopedic knowledge, (2) enriching the structure of the ‘core’ semantic network (i.e. the semantic pointers from WordNet) with topical, semantically unspecified relations from the link structure of Wikipedia. This result is essentially achieved by complementing WordNet with Wikipedia, as well as by leveraging the multilingual structure of the latter. Previous attempts at linking the two resources have been proposed. These include associating Wikipedia pages with the most frequent WordNet sense (Suchanek et al., 2008), extracting domain information from Wikipedia and providing a manual mapping to WordNet concepts (Auer et al., 2007), a model based on vector spaces (Ruiz-Casado et al., 2005), a supervised approach using keyword extraction (Reiter et al., 2008), as well as automatically linking Wikipedia categories to WordNet based on structural information (Ponzetto and Navigli, 2009). In contrast to previous work, BabelNet is the first proposal that integrates the relational structure of WordNet with the semi-structured information from Wikipedia into a unified, widecoverage, multilingual semantic network. 7 Conclusions In this paper we have presented a novel methodology for the automatic construction of a large multilingual lexical knowledge resource. Key to our approach is the establishment of a mapping between a multilingual encyclopedic knowledge repository (Wikipedia) and a computational lexicon of English (WordNet). This integration process has several advantages. Firstly, the two resources contribute different kinds of lexical knowledge, one is concerned mostly with named entities, the other with concepts. Secondly, while Wikipedia is less structured than WordNet, it provides large 223 amounts of semantic relations and can be leveraged to enable multilinguality. Thus, even when they overlap, the two resources provide complementary information about the same named entities or concepts. Further, we contribute a large set of sense occurrences harvested from Wikipedia and SemCor, a corpus that we input to a state-ofthe-art machine translation system to fill in the gap between resource-rich languages – such as English – and resource-poorer ones. Our hope is that the availability of such a language-rich resource5 will enable many non-English and multilingual NLP applications to be developed. Our experiments show that our fully-automated approach produces a large-scale lexical resource with high accuracy. The resource includes millions of semantic relations, mainly from Wikipedia (however, WordNet relations are labeled), and contains almost 3 million concepts (6.7 labels per concept on average). As pointed out in Section 5, such coverage is much wider than that of existing wordnets in non-English languages. While BabelNet currently includes 6 languages, links to freely-available wordnets6 can immediately be established by utilizing the English WordNet as an interlanguage index. Indeed, BabelNet can be extended to virtually any language of interest. In fact, our translation method allows it to cope with any resource-poor language. As future work, we plan to apply our method to other languages, including Eastern European, Arabic, and Asian languages. We also intend to link missing concepts in WordNet, by establishing their most likely hypernyms – e.g., `a la Snow et al. (2006). We will perform a semi-automatic validation of BabelNet, e.g. by exploiting Amazon’s Mechanical Turk (Callison-Burch, 2009) or designing a collaborative game (von Ahn, 2006) to validate low-ranking mappings and translations. Finally, we aim to apply BabelNet to a variety of applications which are known to benefit from a wide-coverage knowledge resource. We have already shown that the English-only subset of BabelNet allows simple knowledge-based algorithms to compete with supervised systems in standard coarse-grained and domain-specific WSD settings (Ponzetto and Navigli, 2010). We plan in the near future to apply BabelNet to the challenging task of cross-lingual WSD (Lefever and Hoste, 2009). 5BabelNet can be freely downloaded for research purposes at http://lcl.uniroma1.it/babelnet. 6http://www.globalwordnet.org. References Jordi Atserias, Luis Villarejo, German Rigau, Eneko Agirre, John Carroll, Bernardo Magnini, and Piek Vossen. 2004. The MEANING multilingual central repository. In Proc. of GWC-04, pages 80–210. S¨oren Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyganiak, and Zachary Ive. 2007. Dbpedia: A nucleus for a web of open data. In Proceedings of 6th International Semantic Web Conference joint with 2nd Asian Semantic Web Conference (ISWC+ASWC 2007), pages 722–735. Sagot Benoˆıt and Darja Fiˇser. 2008. Building a free French WordNet from multilingual resources. In Proceedings of the Ontolex 2008 Workshop. William Black, Sabri Elkateb Horacio Rodriguez, Musa Alkhalifa, Piek Vossen, and Adam Pease. 2006. Introducing the Arabic WordNet project. In Proc. of GWC-06, pages 295–299. Razvan Bunescu and Marius Pas¸ca. 2006. Using encyclopedic knowledge for named entity disambiguation. In Proc. of EACL-06, pages 9–16. Chris Callison-Burch. 2009. Fast, cheap, and creative: Evaluating translation quality using Amazon’s Mechanical Turk. In Proc. of EMNLP-09, pages 286– 295. Jean Carletta. 1996. Assessing agreement on classification tasks: The kappa statistic. Computational Linguistics, 22(2):249–254. Montse Cuadros and German Rigau. 2006. Quality assessment of large scale knowledge resources. In Proc. of EMNLP-06, pages 534–541. Gerard de Melo and Gerhard Weikum. 2009. Towards a universal wordnet by learning from combined evidence. In Proc. of CIKM-09, pages 513–522. Oren Etzioni, Kobi Reiter, Stephen Soderland, and Marcus Sammer. 2007. Lexical translation with application to image search on the Web. In Proceedings of Machine Translation Summit XI. Christiane Fellbaum, editor. 1998. WordNet: An Electronic Database. MIT Press, Cambridge, MA. Pascale Fung. 1995. A pattern matching method for finding noun and proper noun translations from noisy parallel corpora. In Proc. of ACL-95, pages 236–243. Evgeniy Gabrilovich and Shaul Markovitch. 2006. Overcoming the brittleness bottleneck using Wikipedia: Enhancing text categorization with encyclopedic knowledge. In Proc. of AAAI-06, pages 1301–1306. William A. Gale and Kenneth W. Church. 1993. A program for aligning sentences in bilingual corpora. Computational Linguistics, 19(1):75–102. Jim Giles. 2005. Internet encyclopedias go head to head. Nature, 438:900–901. Aria Haghighi, Percy Liang, Taylor Berg-Kirkpatrick, and Dan Klein. 2008. Learning bilingual lexicons from monolingual corpora. In Proc. of ACL-08, pages 771–779. 224 Sanda M. Harabagiu, Dan Moldovan, Marius Pas¸ca, Rada Mihalcea, Mihai Surdeanu, Razvan Bunescu, Roxana Girju, Vasile Rus, and Paul Morarescu. 2000. FALCON: Boosting knowledge for answer engines. In Proc. of TREC-9, pages 479–488. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondˇrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: open source toolkit for statistical machine translation. In Comp. Vol. to Proc. of ACL-07, pages 177–180. Philipp Koehn. 2005. Europarl: A parallel corpus for statistical machine translation. In Proceedings of Machine Translation Summit X. Els Lefever and Veronique Hoste. 2009. Semeval2010 task 3: Cross-lingual Word Sense Disambiguation. In Proc. of the Workshop on Semantic Evaluations: Recent Achievements and Future Directions (SEW-2009), pages 82–87, Boulder, Colorado. Lothar Lemnitzer and Claudia Kunze. 2002. GermaNet – representation, visualization, application. In Proc. of LREC ’02, pages 1485–1491. Alessandro Lenci, Nuria Bel, Federica Busa, Nicoletta Calzolari, Elisabetta Gola, Monica Monachini, Antoine Ogonowski, Ivonne Peters, Wim Peters, Nilda Ruimy, Marta Villegas, and Antonio Zampolli. 2000. SIMPLE: A general framework for the development of multilingual lexicons. International Journal of Lexicography, 13(4):249–263. Mausam, Stephen Soderland, Oren Etzioni, Daniel Weld, Michael Skinner, and Jeff Bilmes. 2009. Compiling a massive, multilingual dictionary via probabilistic inference. In Proc. of ACL-IJCNLP09, pages 262–270. Olena Medelyan, David Milne, Catherine Legg, and Ian H. Witten. 2009. Mining meaning from Wikipedia. Int. J. Hum.-Comput. Stud., 67(9):716– 754. George A. Miller, Claudia Leacock, Randee Tengi, and Ross Bunker. 1993. A semantic concordance. In Proceedings of the 3rd DARPA Workshop on Human Language Technology, pages 303–308, Plainsboro, N.J. Vivi Nastase. 2008. Topic-driven multi-document summarization with encyclopedic knowledge and activation spreading. In Proc. of EMNLP-08, pages 763–772. Roberto Navigli and Mirella Lapata. 2010. An experimental study on graph connectivity for unsupervised Word Sense Disambiguation. IEEE Transactions on Pattern Anaylsis and Machine Intelligence, 32(4):678–692. Roberto Navigli. 2009a. Using cycles and quasicycles to disambiguate dictionary glosses. In Proc. of EACL-09, pages 594–602. Roberto Navigli. 2009b. Word Sense Disambiguation: A survey. ACM Computing Surveys, 41(2):1–69. Hwee Tou Ng and Hian Beng Lee. 1996. Integrating multiple knowledge sources to disambiguate word senses: An exemplar-based approach. In Proc. of ACL-96, pages 40–47. Emanuele Pianta, Luisa Bentivogli, and Christian Girardi. 2002. MultiWordNet: Developing an aligned multilingual database. In Proc. of GWC-02, pages 21–25. Simone Paolo Ponzetto and Roberto Navigli. 2009. Large-scale taxonomy mapping for restructuring and integrating Wikipedia. In Proc. of IJCAI-09, pages 2083–2088. Simone Paolo Ponzetto and Roberto Navigli. 2010. Knowledge-rich Word Sense Disambiguation rivaling supervised system. In Proc. of ACL-10. Simone Paolo Ponzetto and Michael Strube. 2007. Deriving a large scale taxonomy from Wikipedia. In Proc. of AAAI-07, pages 1440–1445. Nils Reiter, Matthias Hartung, and Anette Frank. 2008. A resource-poor approach for linking ontology classes to Wikipedia articles. In Johan Bos and Rodolfo Delmonte, editors, Semantics in Text Processing, volume 1 of Research in Computational Semantics, pages 381–387. College Publications, London, England. Maria Ruiz-Casado, Enrique Alfonseca, and Pablo Castells. 2005. Automatic assignment of Wikipedia encyclopedic entries to WordNet synsets. In Advances in Web Intelligence, volume 3528 of Lecture Notes in Computer Science. Springer Verlag. Marcus Sammer and Stephen Soderland. 2007. Building a sense-distinguished multilingual lexicon from monolingual corpora and bilingual lexicons. In Proceedings of Machine Translation Summit XI. Rion Snow, Dan Jurafsky, and Andrew Ng. 2006. Semantic taxonomy induction from heterogeneous evidence. In Proc. of COLING-ACL-06, pages 801– 808. Fabian M. Suchanek, Gjergji Kasneci, and Gerhard Weikum. 2008. Yago: A large ontology from Wikipedia and WordNet. Journal of Web Semantics, 6(3):203–217. Dan Tufis¸, Dan Cristea, and Sofia Stamou. 2004. BalkaNet: Aims, methods, results and perspectives. a general overview. Romanian Journal on Science and Technology of Information, 7(1-2):9–43. Luis von Ahn. 2006. Games with a purpose. IEEE Computer, 6(39):92–94. Piek Vossen, editor. 1998. EuroWordNet: A Multilingual Database with Lexical Semantic Networks. Kluwer, Dordrecht, The Netherlands. Fei Wu and Daniel Weld. 2007. Automatically semantifying Wikipedia. In Proc. of CIKM-07, pages 41–50. David Yarowsky and Radu Florian. 2002. Evaluating sense disambiguation across diverse parameter spaces. Natural Language Engineering, 9(4):293– 310. Toshio Yokoi. 1995. The EDR electronic dictionary. Communications of the ACM, 38(11):42–44. 225
|
2010
|
23
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 226–236, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Fully Unsupervised Core-Adjunct Argument Classification Omri Abend∗ Institute of Computer Science The Hebrew University [email protected] Ari Rappoport Institute of Computer Science The Hebrew University [email protected] Abstract The core-adjunct argument distinction is a basic one in the theory of argument structure. The task of distinguishing between the two has strong relations to various basic NLP tasks such as syntactic parsing, semantic role labeling and subcategorization acquisition. This paper presents a novel unsupervised algorithm for the task that uses no supervised models, utilizing instead state-of-the-art syntactic induction algorithms. This is the first work to tackle this task in a fully unsupervised scenario. 1 Introduction The distinction between core arguments (henceforth, cores) and adjuncts is included in most theories on argument structure (Dowty, 2000). The distinction can be viewed syntactically, as one between obligatory and optional arguments, or semantically, as one between arguments whose meanings are predicate dependent and independent. The latter (cores) are those whose function in the described event is to a large extent determined by the predicate, and are obligatory. Adjuncts are optional arguments which, like adverbs, modify the meaning of the described event in a predictable or predicate-independent manner. Consider the following examples: 1. The surgeon operated [on his colleague]. 2. Ron will drop by [after lunch]. 3. Yuri played football [in the park]. The marked argument is a core in 1 and an adjunct in 2 and 3. Adjuncts form an independent semantic unit and their semantic role can often be inferred independently of the predicate (e.g., [after lunch] is usually a temporal modifier). Core ∗Omri Abend is grateful to the Azrieli Foundation for the award of an Azrieli Fellowship. roles are more predicate-specific, e.g., [on his colleague] has a different meaning with the verbs ‘operate’ and ‘count’. Sometimes the same argument plays a different role in different sentences. In (3), [in the park] places a well-defined situation (Yuri playing football) in a certain location. However, in “The troops are based [in the park]”, the same argument is obligatory, since being based requires a place to be based in. Distinguishing between the two argument types has been discussed extensively in various formulations in the NLP literature, notably in PP attachment, semantic role labeling (SRL) and subcategorization acquisition. However, no work has tackled it yet in a fully unsupervised scenario. Unsupervised models reduce reliance on the costly and error prone manual multi-layer annotation (POS tagging, parsing, core-adjunct tagging) commonly used for this task. They also allow to examine the nature of the distinction and to what extent it is accounted for in real data in a theory-independent manner. In this paper we present a fully unsupervised algorithm for core-adjunct classification. We utilize leading fully unsupervised grammar induction and POS induction algorithms. We focus on prepositional arguments, since non-prepositional ones are generally cores. The algorithm uses three measures based on different characterizations of the core-adjunct distinction, and combines them using an ensemble method followed by self-training. The measures used are based on selectional preference, predicate-slot collocation and argument-slot collocation. We evaluate against PropBank (Palmer et al., 2005), obtaining roughly 70% accuracy when evaluated on the prepositional arguments and more than 80% for the entire argument set. These results are substantially better than those obtained by a non-trivial baseline. 226 Section 2 discusses the core-adjunct distinction. Section 3 describes the algorithm. Sections 4 and 5 present our experimental setup and results. 2 Core-Adjunct in Previous Work PropBank. PropBank (PB) (Palmer et al., 2005) is a widely used corpus, providing SRL annotation for the entire WSJ Penn Treebank. Its core labels are predicate specific, while adjunct (or modifiers under their terminology) labels are shared across predicates. The adjuncts are subcategorized into several classes, the most frequent of which are locative, temporal and manner1. The organization of PropBank is based on the notion of diathesis alternations, which are (roughly) defined to be alternations between two subcategorization frames that preserve meaning or change it systematically. The frames in which each verb appears were collected and sets of alternating frames were defined. Each such set was assumed to have a unique set of roles, named ‘roleset’. These roles include all roles appearing in any of the frames, except of those defined as adjuncts. Adjuncts are defined to be optional arguments appearing with a wide variety of verbs and frames. They can be viewed as fixed points with respect to alternations, i.e., as arguments that do not change their place or slot when the frame undergoes an alternation. This follows the notions of optionality and compositionality that define adjuncts. Detecting diathesis alternations automatically is difficult (McCarthy, 2001), requiring an initial acquisition of a subcategorization lexicon. This alone is a challenging task tackled in the past using supervised parsers (see below). FrameNet. FrameNet (FN) (Baker et al., 1998) is a large-scale lexicon based on frame semantics. It takes a different approach from PB to semantic roles. Like PB, it distinguishes between core and non-core arguments, but it does so for each and every frame separately. It does not commit that a semantic role is consistently tagged as a core or a non-core across frames. For example, the semantic role ‘path’ is considered core in the ‘Self Motion’ frame, but as non-core in the ‘Placing’ frame. Another difference is that FN does not allow any type of non-core argument to attach to a given frame. For instance, while the ‘Getting’ 1PropBank annotates modals and negation words as modifiers. Since these are not arguments in the common usage of the term, we exclude them from the discussion in this paper. frame allows a ‘Duration’ non-core argument, the ‘Active Perception’ frame does not. PB and FN tend to agree in clear (prototypical) cases, but to differ in others. For instance, both schemes would tag “Yuri played football [in the park]” as an adjunct and “The commander placed a guard [in the park]” as a core. However, in “He walked [into his office]”, the marked argument is tagged as a directional adjunct in PB but as a ‘Direction’ core in FN. Under both schemes, non-cores are usually confined to a few specific semantic domains, notably time, place and manner, in contrast to cores that are not restricted in their scope of applicability. This approach is quite common, e.g., the COBUILD English grammar (Willis, 2004) categorizes adjuncts to be of manner, aspect, opinion, place, time, frequency, duration, degree, extent, emphasis, focus and probability. Semantic Role Labeling. Work in SRL does not tackle the core-adjunct task separately but as part of general argument classification. Supervised approaches obtain an almost perfect score in distinguishing between the two in an in-domain scenario. For instance, the confusion matrix in (Toutanova et al., 2008) indicates that their model scores 99.5% accuracy on this task. However, adaptation results are lower, with the best two models in the CoNLL 2005 shared task (Carreras and M`arquez, 2005) achieving 95.3% (Pradhan et al., 2008) and 95.6% (Punyakanok et al., 2008) accuracy in an adaptation between the relatively similar corpora WSJ and Brown. Despite the high performance in supervised scenarios, tackling the task in an unsupervised manner is not easy. The success of supervised methods stems from the fact that the predicate-slot combination (slot is represented in this paper by its preposition) strongly determines whether a given argument is an adjunct or a core (see Section 3.4). Supervised models are provided with an annotated corpus from which they can easily learn the mapping between predicate-slot pairs and their core/adjunct label. However, induction of the mapping in an unsupervised manner must be based on inherent core-adjunct properties. In addition, supervised models utilize supervised parsers and POS taggers, while the current state-of-the-art in unsupervised parsing and POS tagging is considerably worse than their supervised counterparts. This challenge has some resemblance to un227 supervised detection of multiword expressions (MWEs). An important MWE sub-class is that of phrasal verbs, which are also characterized by verb-preposition pairs (Li et al., 2003; Sporleder and Li, 2009) (see also (Boukobza and Rappoport, 2009)). Both tasks aim to determine semantic compositionality, which is a highly challenging task. Few works addressed unsupervised SRL-related tasks. The setup of (Grenager and Manning, 2006), who presented a Bayesian Network model for argument classification, is perhaps closest to ours. Their work relied on a supervised parser and a rule-based argument identification (both during training and testing). Swier and Stevenson (2004, 2005), while addressing an unsupervised SRL task, greatly differ from us as their algorithm uses the VerbNet (Kipper et al., 2000) verb lexicon, in addition to supervised parses. Finally, Abend et al. (2009) tackled the argument identification task alone and did not perform argument classification of any sort. PP attachment. PP attachment is the task of determining whether a prepositional phrase which immediately follows a noun phrase attaches to the latter or to the preceding verb. This task’s relation to the core-adjunct distinction was addressed in several works. For instance, the results of (Hindle and Rooth, 1993) indicate that their PP attachment system works better for cores than for adjuncts. Merlo and Esteve Ferrer (2006) suggest a system that jointly tackles the PP attachment and the core-adjunct distinction tasks. Unlike in this work, their classifier requires extensive supervision including WordNet, language-specific features and a supervised parser. Their features are generally motivated by common linguistic considerations. Features found adaptable to a completely unsupervised scenario are used in this work as well. Syntactic Parsing. The core-adjunct distinction is included in many syntactic annotation schemes. Although the Penn Treebank does not explicitly annotate adjuncts and cores, a few works suggested mapping its annotation (including function tags) to core-adjunct labels. Such a mapping was presented in (Collins, 1999). In his Model 2, Collins modifies his parser to provide a coreadjunct prediction, thereby improving its performance. The Combinatory Categorial Grammar (CCG) formulation models the core-adjunct distinction explicitly. Therefore, any CCG parser can be used as a core-adjunct classifier (Hockenmaier, 2003). Subcategorization Acquisition. This task specifies for each predicate the number, type and order of obligatory arguments. Determining the allowable subcategorization frames for a given predicate necessarily involves separating its cores from its allowable adjuncts (which are not framed). Notable works in the field include (Briscoe and Carroll, 1997; Sarkar and Zeman, 2000; Korhonen, 2002). All these works used a parsed corpus in order to collect, for each predicate, a set of hypothesized subcategorization frames, to be filtered by hypothesis testing methods. This line of work differs from ours in a few aspects. First, all works use manual or supervised syntactic annotations, usually including a POS tagger. Second, the common approach to the task focuses on syntax and tries to identify the entire frame, rather than to tag each argument separately. Finally, most works address the task at the verb type level, trying to detect the allowable frames for each type. Consequently, the common evaluation focuses on the quality of the allowable frames acquired for each verb type, and not on the classification of specific arguments in a given corpus. Such a token level evaluation was conducted in a few works (Briscoe and Carroll, 1997; Sarkar and Zeman, 2000), but often with a small number of verbs or a small number of frames. A discussion of the differences between type and token level evaluation can be found in (Reichart et al., 2010). The core-adjunct distinction task was tackled in the context of child language acquisition. Villavicencio (2002) developed a classifier based on preposition selection and frequency information for modeling the distinction for locative prepositional phrases. Her approach is not entirely corpus based, as it assumes the input sentences are given in a basic logical form. The study of prepositions is a vibrant research area in NLP. A special issue of Computational Linguistics, which includes an extensive survey of related work, was recently devoted to the field (Baldwin et al., 2009). 228 3 Algorithm We are given a (predicate, argument) pair in a test sentence, and we need to determine whether the argument is a core or an adjunct. Test arguments are assumed to be correctly bracketed. We are allowed to utilize a training corpus of raw text. 3.1 Overview Our algorithm utilizes statistics based on the (predicate, slot, argument head) (PSH) joint distribution (a slot is represented by its preposition). To estimate this joint distribution, PSH samples are extracted from the training corpus using unsupervised POS taggers (Clark, 2003; Abend et al., 2010) and an unsupervised parser (Seginer, 2007). As current performance of unsupervised parsers for long sentences is low, we use only short sentences (up to 10 words, excluding punctuation). The length of test sentences is not bounded. Our results will show that the training data accounts well for the argument realization phenomena in the test set, despite the length bound on its sentences. The sample extraction process is detailed in Section 3.2. Our approach makes use of both aspects of the distinction – obligatoriness and compositionality. We define three measures, one quantifying the obligatoriness of the slot, another quantifying the selectional preference of the verb to the argument and a third that quantifies the association between the head word and the slot irrespective of the predicate (Section 3.3). The measures’ predictions are expected to coincide in clear cases, but may be less successful in others. Therefore, an ensemble-based method is used to combine the three measures into a single classifier. This results in a high accuracy classifier with relatively low coverage. A self-training step is now performed to increase coverage with only a minor deterioration in accuracy (Section 3.4). We focus on prepositional arguments. Nonprepositional arguments in English tend to be cores (e.g., in more than 85% of the cases in PB sections 2–21), while prepositional arguments tend to be equally divided between cores and adjuncts. The difficulty of the task thus lies in the classification of prepositional arguments. 3.2 Data Collection The statistical measures used by our classifier are based on the (predicate, slot, argument head) (PSH) joint distribution. This section details the process of extracting samples from this joint distribution given a raw text corpus. We start by parsing the corpus using the Seginer parser (Seginer, 2007). This parser is unique in its ability to induce a bracketing (unlabeled parsing) from raw text (without even using POS tags) with strong results. Its high speed (thousands of words per second) allows us to use millions of sentences, a prohibitive number for other parsers. We continue by tagging the corpus using Clark’s unsupervised POS tagger (Clark, 2003) and the unsupervised Prototype Tagger (Abend et al., 2010)2. The classes corresponding to prepositions and to verbs are manually selected from the induced clusters3. A preposition is defined to be any word which is the first word of an argument and belongs to a prepositions cluster. A verb is any word belonging to a verb cluster. This manual selection requires only a minute, since the number of classes is very small (34 in our experiments). In addition, knowing what is considered a preposition is part of the task definition itself. Argument identification is hard even for supervised models and is considerably more so for unsupervised ones (Abend et al., 2009). We therefore confine ourselves to sentences of length not greater than 10 (excluding punctuation) which contain a single verb. A sequence of words will be marked as an argument of the verb if it is a constituent that does not contain the verb (according to the unsupervised parse tree), whose parent is an ancestor of the verb. This follows the pruning heuristic of (Xue and Palmer, 2004) often used by SRL algorithms. The corpus is now tagged using an unsupervised POS tagger. Since the sentences in question are short, we consider every word which does not belong to a closed class cluster as a head word (an argument can have several head words). A closed class is a class of function words with relatively few word types, each of which is very frequent. Typical examples include determiners, prepositions and conjunctions. A class which is not closed is open. In this paper, we define closed classes to be clusters in which the ratio between the number of word tokens and the number of word types ex2Clark’s tagger was replaced by the Prototype Tagger where the latter gave a significant improvement. See Section 4. 3We also explore a scenario in which they are identified by a supervised tagger. See Section 4. 229 ceeds a threshold T 4. Using these annotation layers, we traverse the corpus and extract every (predicate, slot, argument head) triplet. In case an argument has several head words, each of them is considered as an independent sample. We denote the number of times that a triplet occurred in the training corpus by N(p, s, h). 3.3 Collocation Measures In this section we present the three types of measures used by the algorithm and the rationale behind each of them. These measures are all based on the PSH joint distribution. Given a (predicate, prepositional argument) pair from the test set, we first tag and parse the argument using the unsupervised tools above5. Each word in the argument is now represented by its word form (without lemmatization), its unsupervised POS tag and its depth in the parse tree of the argument. The last two will be used to determine which are the head words of the argument (see below). The head words themselves, once chosen, are represented by the lemma. We now compute the following measures. Selectional Preference (SP). Since the semantics of cores is more predicate dependent than the semantics of adjuncts, we expect arguments for which the predicate has a strong preference (in a specific slot) to be cores. Selectional preference induction is a wellestablished task in NLP. It aims to quantify the likelihood that a certain argument appears in a certain slot of a predicate. Several methods have been suggested (Resnik, 1996; Li and Abe, 1998; Schulte im Walde et al., 2008). We use the paradigm of (Erk, 2007). For a given predicate slot pair (p, s), we define its preference to the argument head h to be: SP(p, s, h) = X h′∈Heads Pr(h′|p, s) · sim(h, h′) Pr(h|p, s) = N(p, s, h) Σh′N(p, s, h′) sim(h, h′) is a similarity measure between argument heads. Heads is the set of all head words. 4We use sections 2–21 of the PTB WSJ for these counts, containing 0.95M words. Our T was set to 50. 5Note that while current unsupervised parsers have low performance on long sentences, arguments, even in long sentences, are usually still short enough for them to operate well. Their average length in the test set is 5.1 words. This is a natural extension of the naive (and sparse) maximum likelihood estimator Pr(h|p, s), which is obtained by taking sim(h, h′) to be 1 if h = h′ and 0 otherwise. The similarity measure we use is based on the slot distributions of the arguments. That is, two arguments are considered similar if they tend to appear in the same slots. Each head word h is assigned a vector where each coordinate corresponds to a slot s. The value of the coordinate is the number of times h appeared in s, i.e. Σp′N(p′, s, h) (p′ is summed over all predicates). The similarity measure between two head words is then defined as the cosine measure of their vectors. Since arguments in the test set can be quite long, not every open class word in the argument is taken to be a head word. Instead, only those appearing in the top level (depth = 1) of the argument under its unsupervised parse tree are taken. In case there are no such open class words, we take those appearing in depth 2. The selectional preference of the whole argument is then defined to be the arithmetic mean of this measure over all of its head words. If the argument has no head words under this definition or if none of the head words appeared in the training corpus, the selectional preference is undefined. Predicate-Slot Collocation. Since cores are obligatory, when a predicate persistently appears with an argument in a certain slot, the arguments in this slot tends to be cores. This notion can be captured by the (predicate, slot) joint distribution. We use the Pointwise Mutual Information measure (PMI) to capture the slot and the predicate’s collocation tendency. Let p be a predicate and s a slot, then: PS(p, s) = PMI(p, s) = log Pr(p, s) Pr(s) · Pr(p) = = log N(p, s)Σp′,s′N(p′, s′) Σs′N(p, s′)Σp′N(p′, s) Since there is only a meager number of possible slots (that is, of prepositions), estimating the (predicate, slot) distribution can be made by the maximum likelihood estimator with manageable sparsity. In order not to bias the counts towards predicates which tend to take more arguments, we define here N(p, s) to be the number of times the (p, s) pair occurred in the training corpus, irrespective of the number of head words the argument had (and not e.g., ΣhN(p, s, h)). Argu230 ments with no prepositions are included in these counts as well (with s = NULL), so not to bias against predicates which tend to have less nonprepositional arguments. Argument-Slot Collocation. Adjuncts tend to belong to one of a few specific semantic domains (see Section 2). Therefore, if an argument tends to appear in a certain slot in many of its instances, it is an indication that this argument tends to have a consistent semantic flavor in most of its instances. In this case, the argument and the preposition can be viewed as forming a unit on their own, independent of the predicate with which they appear. We therefore expect such arguments to be adjuncts. We formalize this notion using the following measure. Let p, s, h be a predicate, a slot and a head word respectively. We then use6: AS(s, h) = 1 −Pr(s|h) = 1 − Σp′N(p′, s, h) Σp′,s′N(p′, s′, h) We select the head words of the argument as we did with the selectional preference measure. Again, the AS of the whole argument is defined to be the arithmetic mean of the measure over all of its head words. Thresholding. In order to turn these measures into classifiers, we set a threshold below which arguments are marked as adjuncts and above which as cores. In order to avoid tuning a parameter for each of the measures, we set the threshold as the median value of this measure in the test set. That is, we find the threshold which tags half of the arguments as cores and half as adjuncts. This relies on the prior knowledge that prepositional arguments are roughly equally divided between cores and adjuncts7. 3.4 Combination Model The algorithm proceeds to integrate the predictions of the weak classifiers into a single classifier. We use an ensemble method (Breiman, 1996). Each of the classifiers may either classify an argument as an adjunct, classify it as a core, or abstain. In order to obtain a high accuracy classifier, to be used for self-training below, the ensemble classifier only tags arguments for which none of 6The conditional probability is subtracted from 1 so that higher values correspond to cores, as with the other measures. 7In case the test data is small, we can use the median value on the training data instead. the classifiers abstained, i.e., when sufficient information was available to make all three predictions. The prediction is determined by the majority vote. The ensemble classifier has high precision but low coverage. In order to increase its coverage, a self-training step is performed. We observe that a predicate and a slot generally determine whether the argument is a core or an adjunct. For instance, in our development data, a classifier which assigns all arguments that share a predicate and a slot their most common label, yields 94.3% accuracy on the pairs appearing at least 5 times. This property of the core-adjunct distinction greatly simplifies the task for supervised algorithms (see Section 2). We therefore apply the following procedure: (1) tag the training data with the ensemble classifier; (2) for each test sample x, if more than a ratio of α of the training samples sharing the same predicate and slot with x are labeled as cores, tag x as core. Otherwise, tag x as adjunct. Test samples which do not share a predicate and a slot with any training sample are considered out of coverage. The parameter α is chosen so half of the arguments are tagged as cores and half as adjuncts. In our experiments α was about 0.25. 4 Experimental Setup Experiments were conducted in two scenarios. In the ‘SID’ (supervised identification of prepositions and verbs) scenario, a gold standard list of prepositions was provided. The list was generated by taking every word tagged by the preposition tag (‘IN’) in at least one of its instances under the gold standard annotation of the WSJ sections 2– 21. Verbs were identified using MXPOST (Ratnaparkhi, 1996). Words tagged with any of the verb tags, except of the auxiliary verbs (‘have’, ‘be’ and ‘do’) were considered predicates. This scenario decouples the accuracy of the algorithm from the quality of the unsupervised POS tagging. In the ‘Fully Unsupervised’ scenario, prepositions and verbs were identified using Clark’s tagger (Clark, 2003). It was asked to produce a tagging into 34 classes. The classes corresponding to prepositions and to verbs were manually identified. Prepositions in the test set were detected with 84.2% precision and 91.6% recall. The prediction of whether a word belongs to an open class or a closed was based on the output of the Prototype tagger (Abend et al., 2010). The Prototype tagger provided significantly more ac231 curate predictions in this context than Clark’s. The 39832 sentences of PropBank’s sections 2– 21 were used as a test set without bounding their lengths8. Cores were defined to be any argument bearing the labels ‘A0’ – ‘A5’, ‘C-A0’ – ‘C-A5’ or ‘R-A0’ – ‘R-A5’. Adjuncts were defined to be arguments bearing the labels ‘AM’, ‘C-AM’ or ‘R-AM’. Modals (‘AM-MOD’) and negation modifiers (‘AM-NEG’) were omitted since they do not represent adjuncts. The test set includes 213473 arguments, 45939 (21.5%) are prepositional. Of the latter, 22442 (48.9%) are cores and 23497 (51.1%) are adjuncts. The non-prepositional arguments include 145767 (87%) cores and 21767 (13%) adjuncts. The average number of words per argument is 5.1. The NANC (Graff, 1995) corpus was used as a training set. Only sentences of length not greater than 10 excluding punctuation were used (see Section 3.2), totaling 4955181 sentences. 7673878 (5635810) arguments were identified in the ‘SID’ (‘Fully Unsupervised’) scenario. The average number of words per argument is 1.6 (1.7). Since this is the first work to tackle this task using neither manual nor supervised syntactic annotation, there is no previous work to compare to. However, we do compare against a non-trivial baseline, which closely follows the rationale of cores as obligatory arguments. Our Window Baseline tags a corpus using MXPOST and computes, for each predicate and preposition, the ratio between the number of times that the preposition appeared in a window of W words after the verb and the total number of times that the verb appeared. If this number exceeds a certain threshold β, all arguments having that predicate and preposition are tagged as cores. Otherwise, they are tagged as adjuncts. We used 18.7M sentences from NANC of unbounded length for this baseline. W and β were fine-tuned against the test set9. We also report results for partial versions of the algorithm, starting with the three measures used (selectional preference, predicate-slot collocation and argument-slot collocation). Results for the ensemble classifier (prior to the bootstrapping stage) are presented in two variants: one 8The first 15K arguments were used for the algorithm’s development and therefore excluded from the evaluation. 9Their optimal value was found to be W=2, β=0.03. The low optimal value of β is an indication of the noisiness of this technique. in which the ensemble is used to tag arguments for which all three measures give a prediction (the ‘Ensemble(Intersection)’ classifier) and one in which the ensemble tags all arguments for which at least one classifier gives a prediction (the ‘Ensemble(Union)’ classifier). For the latter, a tie is broken in favor of the core label. The ‘Ensemble(Union)’ classifier is not a part of our model and is evaluated only as a reference. In order to provide a broader perspective on the task, we compare the measures in the basis of our algorithm to simplified or alternative measures. We experiment with the following measures: 1. Simple SP – a selectional preference measure defined to be Pr(head|slot, predicate). 2. Vast Corpus SP – similar to ‘Simple SP’ but with a much larger corpus. It uses roughly 100M arguments which were extracted from the web-crawling based corpus of (Gabrilovich and Markovitch, 2005) and the British National Corpus (Burnard, 2000). 3. Thesaurus SP – a selectional preference measure which follows the paradigm of (Erk, 2007) (Section 3.3) and defines the similarity between two heads to be the Jaccard affinity between their two entries in Lin’s automatically compiled thesaurus (Lin, 1998)10. 4. Pr(slot|predicate) – an alternative to the used predicate-slot collocation measure. 5. PMI(slot, head) – an alternative to the used argument-slot collocation measure. 6. Head Dependence – the entropy of the predicate distribution given the slot and the head (following (Merlo and Esteve Ferrer, 2006)): HD(s, h) = −ΣpPr(p|s, h) · log(Pr(p|s, h)) Low entropy implies a core. For each of the scenarios and the algorithms, we report accuracy, coverage and effective accuracy. Effective accuracy is defined to be the accuracy obtained when all out of coverage arguments are tagged as adjuncts. This procedure always yields a classifier with 100% coverage and therefore provides an even ground for comparing the algorithms’ performance. We see accuracy as important on its own right since increasing coverage is often straightforward given easily obtainable larger training corpora. 10Since we aim for a minimally supervised scenario, we used the proximity-based version of his thesaurus which does not require parsing as pre-processing. http://webdocs.cs.ualberta.ca/∼lindek/Downloads/sims.lsp.gz 232 Collocation Measures Ensemble + Cov. Sel. Preference Pred-Slot Arg-Slot Ensemble(I) Ensemble(U) E(I) + ST SID Scenario Accuracy 65.6 64.5 72.4 74.1 68.7 70.6 Coverage 35.6 77.8 44.7 33.2 88.1 74.2 Eff. Acc. 56.7 64.8 58.8 58.8 67.8 68.4 Fully Unsupervised Accuracy 62.6 61.1 69.4 70.6 64.8 68.8 Scenario Coverage 24.8 59.0 38.7 22.8 74.2 56.9 Eff. Acc. 52.6 57.5 55.8 53.8 61.0 61.4 Table 1: Results for the various models. Accuracy, coverage and effective accuracy are presented in percents. Effective accuracy is defined to be the accuracy resulting from labeling each out of coverage argument with an adjunct label. The rows represent the following models (left to right): selectional preference, predicate-slot collocation, argument-slot collocation, ‘Ensemble(Intersection)’, ‘Ensemble(Union)’ and the ‘Ensemble(Intersection)’ followed by self-training (see Section 3.4). ‘Ensemble(Intersection)’ obtains the highest accuracy. The ensemble + self-training obtains the highest effective accuracy. Selectional Preference Measures Pred-Slot Measures Arg-Slot Measures SP∗ S. SP V.C. SP Lin SP PS∗ Pr(s|p) Window AS∗ PMI(s, h) HD Acc. 65.6 41.6 44.8 49.9 64.5 58.9 64.1 72.4 67.5 67.4 Cov. 35.6 36.9 45.3 36.7 77.8 77.8 92.6 44.7 44.7 44.7 Eff. Acc. 56.7 48.2 47.7 51.3 64.8 60.5 65.0 58.8 56.6 56.6 Table 2: Comparison of the measures used by our model to alternative measures in the ‘SID’ scenario. Results are in percents. The sections of the table are (from left to right): selectional preference measures, predicate-slot measures, argument-slot measures and head dependence. The measures are (left to right): SP∗, Simple SP, Vast Corpus SP, Lin SP, PS∗, Pr(slot|predicate), Window Baseline, AS∗, PMI(slot, head) and Head Dependence. The measures marked with ∗are the ones used by our model. See Section 4. Another reason is that a high accuracy classifier may provide training data to be used by subsequent supervised algorithms. For completeness, we also provide results for the entire set of arguments. The great majority of non-prepositional arguments are cores (87% in the test set). We therefore tag all non-prepositional as cores and tag prepositional arguments using our model. In order to minimize supervision, we distinguish between the prepositional and the nonprepositional arguments using Clark’s tagger. Finally, we experiment on a scenario where even argument identification on the test set is not provided, but performed by the algorithm of (Abend et al., 2009), which uses neither syntactic nor SRL annotation but does utilize a supervised POS tagger. We therefore run it in the ‘SID’ scenario. We apply it to the sentences of length at most 10 contained in sections 2–21 of PB (11586 arguments in 6007 sentences). Non-prepositional arguments are invariably tagged as cores and out of coverage prepositional arguments as adjuncts. We report labeled and unlabeled recall, precision and F-scores for this experiment. An unlabeled match is defined to be an argument that agrees in its boundaries with a gold standard argument and a labeled match requires in addition that the arguments agree in their core/adjunct label. We also report labeling accuracy which is the ratio between the number of labeled matches and the number of unlabeled matches11. 5 Results Table 1 presents the results of our main experiments. In both scenarios, the most accurate of the three basic classifiers was the argument-slot collocation classifier. This is an indication that the collocation between the argument and the preposition is more indicative of the core/adjunct label than the obligatoriness of the slot (as expressed by the predicate-slot collocation). Indeed, we can find examples where adjuncts, although optional, appear very often with a certain verb. An example is ‘meet’, which often takes a temporal adjunct, as in ‘Let’s meet [in July]’. This is a semantic property of ‘meet’, whose syntactic expression is not obligatory. All measures suffered from a comparable deterioration of accuracy when moving from the ‘SID’ to the ‘Fully Unsupervised’ scenario. The deterioration in coverage, however, was considerably lower for the argument-slot collocation. The ‘Ensemble(Intersection)’ model in both cases is more accurate than each of the basic classifiers alone. This is to be expected as it combines the predictions of all three. The self-training step significantly increases the ensemble model’s cov11Note that the reported unlabeled scores are slightly lower than those reported in the 2009 paper, due to the exclusion of the modals and negation modifiers. 233 Precision Recall F-score lAcc. Unlabeled 50.7 66.3 57.5 – Labeled 42.4 55.4 48.0 83.6 Table 3: Unlabeled and labeled scores for the experiments using the unsupervised argument identification system of (Abend et al., 2009). Precision, recall, F-score and labeling accuracy are given in percents. erage (with some loss in accuracy), thus obtaining the highest effective accuracy. It is also more accurate than the simpler classifier ‘Ensemble(Union)’ (although the latter’s coverage is higher). Table 2 presents results for the comparison to simpler or alternative measures. Results indicate that the three measures used by our algorithm (leftmost column in each section) obtain superior results. The only case in which performance is comparable is the window baseline compared to the Pred-Slot measure. However, the baseline’s score was obtained by using a much larger corpus and a careful hand-tuning of the parameters12. The poor performance of Simple SP can be ascribed to sparsity. This is demonstrated by the median value of 0, which this measure obtained on the test set. Accuracy is only somewhat better with a much larger corpus (Vast Corpus SP). The Thesaurus SP most probably failed due to insufficient coverage, despite its applicability in a similar supervised task (Zapirain et al., 2009). The Head Dependence measure achieves a relatively high accuracy of 67.4%. We therefore attempted to incorporate it into our model, but failed to achieve a significant improvement to the overall result. We expect a further study of the relations between the measures will suggest better ways of combining their predictions. The obtained effective accuracy for the entire set of arguments, where the prepositional arguments are automatically identified, was 81.6%. Table 3 presents results of our experiments with the unsupervised argument identification model of (Abend et al., 2009). The unlabeled scores reflect performance on argument identification alone, while the labeled scores reflect the joint performance of both the 2009 and our algorithms. These results, albeit low, are potentially beneficial for unsupervised subcategorization acquisition. The accuracy of our model on the entire set (prepositional argument subset) of correctly identified arguments was 83.6% (71.7%). This is 12We tried about 150 parameter pairs for the baseline. The average of the five best effective accuracies was 64.3%. somewhat higher than the score on the entire test set (‘SID’ scenario), which was 83.0% (68.4%), probably due to the bounded length of the test sentences in this case. 6 Conclusion We presented a fully unsupervised algorithm for the classification of arguments into cores and adjuncts. Since most non-prepositional arguments are cores, we focused on prepositional arguments, which are roughly equally divided between cores and adjuncts. The algorithm computes three statistical measures and utilizes ensemble-based and self-training methods to combine their predictions. The algorithm applies state-of-the-art unsupervised parser and POS tagger to collect statistics from a large raw text corpus. It obtains an accuracy of roughly 70%. We also show that (somewhat surprisingly) an argument-slot collocation measure gives more accurate predictions than a predicate-slot collocation measure on this task. We speculate the reason is that the head word disambiguates the preposition and that this disambiguation generally determines whether a prepositional argument is a core or an adjunct (somewhat independently of the predicate). This calls for a future study into the semantics of prepositions and their relation to the core-adjunct distinction. In this context two recent projects, The Preposition Project (Litkowski and Hargraves, 2005) and PrepNet (Saint-Dizier, 2006), which attempt to characterize and categorize the complex syntactic and semantic behavior of prepositions, may be of relevance. It is our hope that this work will provide a better understanding of core-adjunct phenomena. Current supervised SRL models tend to perform worse on adjuncts than on cores (Pradhan et al., 2008; Toutanova et al., 2008). We believe a better understanding of the differences between cores and adjuncts may contribute to the development of better SRL techniques, in both its supervised and unsupervised variants. References Omri Abend, Roi Reichart and Ari Rappoport, 2009. Unsupervised Argument Identification for Semantic Role Labeling. ACL ’09. Omri Abend, Roi Reichart and Ari Rappoport, 2010. Improved Unsupervised POS Induction through Prototype Discovery. ACL ’10. 234 Collin F. Baker, Charles J. Fillmore and John B. Lowe, 1998. The Berkeley FrameNet Project. ACLCOLING ’98. Timothy Baldwin, Valia Kordoni and Aline Villavicencio, 2009. Prepositions in Applications: A Survey and Introduction to the Special Issue. Computational Linguistics, 35(2):119–147. Ram Boukobza and Ari Rappoport, 2009. MultiWord Expression Identification Using Sentence Surface Features. EMNLP ’09. Leo Breiman, 1996. Bagging Predictors. Machine Learning, 24(2):123–140. Ted Briscoe and John Carroll, 1997. Automatic Extraction of Subcategorization from Corpora. Applied NLP ’97. Lou Burnard, 2000. User Reference Guide for the British National Corpus. Technical report, Oxford University. Xavier Carreras and Llu`ıs M`arquez, 2005. Introduction to the CoNLL–2005 Shared Task: Semantic Role Labeling. CoNLL ’05. Alexander Clark, 2003. Combining Distributional and Morphological Information for Part of Speech Induction. EACL ’03. Michael Collins, 1999. Head-driven statistical models for natural language parsing. Ph.D. thesis, University of Pennsylvania. David Dowty, 2000. The Dual Analysis of Adjuncts and Complements in Categorial Grammar. Modifying Adjuncts, ed. Lang, Maienborn and Fabricius– Hansen, de Gruyter, 2003. Katrin Erk, 2007. A Simple, Similarity-based Model for Selectional Preferences. ACL ’07. Evgeniy Gabrilovich and Shaul Markovitch, 2005. Feature Generation for Text Categorization using World Knowledge. IJCAI ’05. David Graff, 1995. North American News Text Corpus. Linguistic Data Consortium. LDC95T21. Trond Grenager and Christopher D. Manning, 2006. Unsupervised Discovery of a Statistical Verb Lexicon. EMNLP ’06. Donald Hindle and Mats Rooth, 1993. Structural Ambiguity and Lexical Relations. Computational Linguistics, 19(1):103–120. Julia Hockenmaier, 2003. Data and Models for Statistical Parsing with Combinatory Categorial Grammar. Ph.D. thesis, University of Edinburgh. Karin Kipper, Hoa Trang Dang and Martha Palmer, 2000. Class-Based Construction of a Verb Lexicon. AAAI ’00. Anna Korhonen, 2002. Subcategorization Acquisition. Ph.D. thesis, University of Cambridge. Hang Li and Naoki Abe, 1998. Generalizing Case Frames using a Thesaurus and the MDL Principle. Computational Linguistics, 24(2):217–244. Wei Li, Xiuhong Zhang, Cheng Niu, Yuankai Jiang and Rohini Srihari, 2003. An Expert Lexicon Approach to Identifying English Phrasal Verbs. ACL ’03. Dekang Lin, 1998. Automatic Retrieval and Clustering of Similar Words. COLING–ACL ’98. Ken Litkowski and Orin Hargraves, 2005. The Preposition Project. ACL-SIGSEM Workshop on “The Linguistic Dimensions of Prepositions and Their Use in Computational Linguistic Formalisms and Applications”. Diana McCarthy, 2001. Lexical Acquisition at the Syntax-Semantics Interface: Diathesis Alternations, Subcategorization Frames and Selectional Preferences. Ph.D. thesis, University of Sussex. Paula Merlo and Eva Esteve Ferrer, 2006. The Notion of Argument in Prepositional Phrase Attachment. Computational Linguistics, 32(3):341–377. Martha Palmer, Daniel Gildea and Paul Kingsbury, 2005. The Proposition Bank: A Corpus Annotated with Semantic Roles. Computational Linguistics, 31(1):71–106. Sameer Pradhan, Wayne Ward and James H. Martin, 2008. Towards Robust Semantic Role Labeling. Computational Linguistics, 34(2):289–310. Vasin Punyakanok, Dan Roth and Wen-tau Yih, 2008. The Importance of Syntactic Parsing and Inference in Semantic Role Labeling. Computational Linguistics, 34(2):257–287. Adwait Ratnaparkhi, 1996. Maximum Entropy PartOf-Speech Tagger. EMNLP ’96. Roi Reichart, Omri Abend and Ari Rappoport, 2010. Type Level Clustering Evaluation: New Measures and a POS Induction Case Study. CoNLL ’10. Philip Resnik, 1996. Selectional constraints: An information-theoretic model and its computational realization. Cognition, 61:127–159. Patrick Saint-Dizier, 2006. PrepNet: A Multilingual Lexical Description of Prepositions. LREC ’06. Anoop Sarkar and Daniel Zeman, 2000. Automatic Extraction of Subcategorization Frames for Czech. COLING ’00. Sabine Schulte im Walde, Christian Hying, Christian Scheible and Helmut Schmid, 2008. Combining EM Training and the MDL Principle for an Automatic Verb Classification Incorporating Selectional Preferences. ACL ’08. 235 Yoav Seginer, 2007. Fast Unsupervised Incremental Parsing. ACL ’07. Caroline Sporleder and Linlin Li, 2009. Unsupervised Recognition of Literal and Non-Literal Use of Idiomatic Expressions. EACL ’09. Robert S. Swier and Suzanne Stevenson, 2004. Unsupervised Semantic Role Labeling. EMNLP ’04. Robert S. Swier and Suzanne Stevenson, 2005. Exploiting a Verb Lexicon in Automatic Semantic Role Labelling. EMNLP ’05. Kristina Toutanova, Aria Haghighi and Christopher D. Manning, 2008. A Global Joint Model for Semantic Role Labeling. Computational Linguistics, 34(2):161–191. Aline Villavicencio, 2002. Learning to Distinguish PP Arguments from Adjuncts. CoNLL ’02. Dave Willis, 2004. Collins Cobuild Intermedia English Grammar, Second Edition. HarperCollins Publishers. Nianwen Xue and Martha Palmer, 2004. Calibrating Features for Semantic Role Labeling. EMNLP ’04. Be˜nat Zapirain, Eneko Agirre and Llu´ıs M`arquez, 2009. Generalizing over Lexical Features: Selectional Preferences for Semantic Role Classification. ACL ’09, short paper. 236
|
2010
|
24
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 237–246, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Towards Open-Domain Semantic Role Labeling Danilo Croce, Cristina Giannone, Paolo Annesi, Roberto Basili {croce,giannone,annesi,basili}@info.uniroma2.it Department of Computer Science, Systems and Production University of Roma, Tor Vergata Abstract Current Semantic Role Labeling technologies are based on inductive algorithms trained over large scale repositories of annotated examples. Frame-based systems currently make use of the FrameNet database but fail to show suitable generalization capabilities in out-of-domain scenarios. In this paper, a state-of-art system for frame-based SRL is extended through the encapsulation of a distributional model of semantic similarity. The resulting argument classification model promotes a simpler feature space that limits the potential overfitting effects. The large scale empirical study here discussed confirms that state-of-art accuracy can be obtained for out-of-domain evaluations. 1 Introduction The availability of large scale semantic lexicons, such as FrameNet (Baker et al., 1998), allowed the adoption of a wide family of learning paradigms in the automation of semantic parsing. Building upon the so called frame semantic model (Fillmore, 1985), the Berkeley FrameNet project has developed a semantic lexicon for the core vocabulary of English, since 1997. A frame is evoked in texts through the occurrence of its lexical units (LU), i.e. predicate words such verbs, nouns, or adjectives, and specifies the participants and properties of the situation it describes, the so called frame elements (FEs). Semantic Role Labeling (SRL) is the task of automatic recognition of individual predicates together with their major roles (e.g. frame elements) as they are grammatically realized in input sentences. It has been a popular task since the availability of the PropBank and FrameNet annotated corpora (Palmer et al., 2005), the seminal work of (Gildea and Jurafsky, 2002) and the successful CoNLL evaluation campaigns (Carreras and M`arquez, 2005). Statistical machine learning methods, ranging from joint probabilistic models to support vector machines, have been successfully adopted to provide very accurate semantic labeling, e.g. (Carreras and M`arquez, 2005). SRL based on FrameNet is thus not a novel task, although very few systems are known capable of completing a general frame-based annotation process over raw texts, noticeable exceptions being discussed for example in (Erk and Pado, 2006), (Johansson and Nugues, 2008b) and (Coppola et al., 2009). Some critical limitations have been outlined in literature, some of them independent from the underlying semantic paradigm. Parsing Accuracy. Most of the employed learning algorithms are based on complex sets of syntagmatic features, as deeply investigated in (Johansson and Nugues, 2008b). The resulting recognition is thus highly dependent on the accuracy of the underlying parser, whereas wrong structures returned by the parser usually imply large misclassification errors. Annotation costs. Statistical learning approaches applied to SRL are very demanding with respect to the amount and quality of the training material. The complex SRL architectures proposed (usually combining local and global, i.e. joint, models of argument classification, e.g. (Toutanova et al., 2008)) require a large number of annotated examples. The amount and quality of the training data required to reach a significant accuracy is a serious limitation to the exploitation of SRL in many NLP applications. Limited Linguistic Generalization. Several studies showed that even when large training sets exist the corresponding learning exhibits poor generalization power. Most of the CoNLL 2005 systems show a significant performance drop when the tested corpus, i.e. Brown, differs from 237 the training one (i.e. Wall Street Journal), e.g. (Toutanova et al., 2008). More recently, the stateof-art frame-based semantic role labeling system discussed in (Johansson and Nugues, 2008b) reports a 19% drop in accuracy for the argument classification task when a different test domain is targeted (i.e. NTI corpus). Out-of-domain tests seem to suggest the models trained on BNC do not generalize well to novel grammatical and lexical phenomena. As also suggested in (Pradhan et al., 2008), the major drawback is the poor generalization power affecting lexical features. Notice how this is also a general problem of statistical learning processes, as large fine grain feature sets are more exposed to the risks of overfitting. The above problems are particularly critical for frame-based shallow semantic parsing where, as opposed to more syntactic-oriented semantic labeling schemes (as Propbank (Palmer et al., 2005)), a significant mismatch exists between the semantic descriptors and the underlying syntactic annotation level. In (Johansson and Nugues, 2008b) an upper bound of about 83.9% for the accuracy of the argument identification task is reported, it is due to the complexity in projecting frame element boundaries out from the dependency graph: more than 16% of the roles in the annotated material lack of a clear grammatical status. The limited level of linguistic generalization outlined above is still an open research problem. Existing solutions have been proposed in literature along different lines. Learning from richer linguistic descriptions of more complex structures is proposed in (Toutanova et al., 2008). Limiting the cost required for developing large domainspecific training data sets has been also studied, e.g., (F¨urstenau and Lapata, 2009). Finally, the application of semi-supervised learning is attempted to increase the lexical expressiveness of the model, e.g. (Goldberg and Elhadad, 2009). In this paper, this last direction is pursued. A semi-supervised statistical model exploiting useful lexical information from unlabeled corpora is proposed. The model adopts a simple feature space by relying on a limited set of grammatical properties, thus reducing its learning capacity. Moreover, it generalizes lexical information about the annotated examples by applying a geometrical model, in a Latent Semantic Analysis style, inspired by a distributional paradigm (Pado and Lapata, 2007). As we will see, the accuracy reachable through a restricted feature space is still quite close to the state-of-art, but interestingly the performance drops in out-of-domain tests are avoided. In the following, after discussing existing approaches to SRL (Section 2), a distributional approach is defined in Section 3. Section 3.2 discusses the proposed HMM-based treatment of joint inferences in argument classification. The large scale experiments described in Section 4 will allow to draw the conclusions of Section 5. 2 Related Work State-of-art approaches to frame-based SRL are based on Support Vector Machines, trained over linear models of syntactic features, e.g. (Johansson and Nugues, 2008b), or tree-kernels, e.g. (Coppola et al., 2009). SRL proceeds through two main steps: the localization of arguments in a sentence, called boundary detection (BD), and the assignment of the proper role to the detected constituents, that is the argument classification, (AC) step. In (Toutanova et al., 2008) a SRL model over Propbank that effectively exploits the semantic argument frame as a joint structure, is presented. It incorporates strong dependencies within a comprehensive statistical joint model with a rich set of features over multiple argument phrases. This approach effectively introduces a new step in SRL, also called Joint Re-ranking, (RR), e.g. (Toutanova et al., 2008) or (Moschitti et al., 2008). First local models are applied to produce role labels over individual arguments, then the joint model is used to decide the entire argument sequence among the set of the n-best competing solutions. While these approaches increase the expressive power of the models to capture more general linguistic properties, they rely on complex feature sets, are more demanding about the amount of training information and increase the overall exposure to overfitting effects. In (Johansson and Nugues, 2008b) the impact of different grammatical representations on the task of frame-based shallow semantic parsing is studied and the poor lexical generalization problem is outlined. An argument classification accuracy of 89.9% over the FrameNet (i.e. BNC) dataset is shown to decrease to 71.1% when a different test domain is evaluated (i.e. the Nuclear Threat Initiative corpus). The argument classification 238 component is thus shown to be heavily domaindependent whereas the inclusion of grammatical function features is just able to mitigate this sensitivity. In line with (Pradhan et al., 2008), it is suggested that lexical features are domain specific and their suitable generalization is not achieved. The lack of suitable lexical information is also discussed in (F¨urstenau and Lapata, 2009) through an approach aiming to support the creation of novel annotated resources. Accordingly a semisupervised approach for reducing the costs of the manual annotation effort is proposed. Through a graph alignment algorithm triggered by annotated resources, the method acquires training instances from an unlabeled corpus also for verbs not listed as existing FrameNet predicates. 2.1 The role of Lexical Semantic Information It is widely accepted that lexical information (as features directly derived from word forms) is crucial for training accurate systems in a number of NLP tasks. Indeed, all the best systems in the CoNLL shared task competitions (e.g. Chunking (Tjong Kim Sang and Buchholz, 2000)) make extensive use of lexical information. Also lexical features are beneficial in SRL usually either for systems on Propbank as well as for FrameNetbased annotation. In (Goldberg and Elhadad, 2009), a different strategy to incorporate lexical features into classification models is proposed. A more expressive training algorithm (i.e. anchored SVM) coupled with an aggressive feature pruning strategy is shown to achieve high accuracy over a chunking and named entity recognition task. The suggested perspective here is that effective semantic knowledge can be collected from sources external to the annotated corpora (very large unannotated corpora or on manually constructed lexical resources) rather than learned from the raw lexical counts of the annotated corpus. Notice how this is also the strategy pursued in recent work on deep learning approaches to NLP tasks. In (Collobert and Weston, 2008) a unified architecture for Natural Language Processing that learns features relevant to the tasks at hand given very limited prior knowledge is presented. It embodies the idea that a multitask learning architecture coupled with semi-supervised learning can be effectively applied even to complex linguistic tasks such as SRL. In particular, (Collobert and Weston, 2008) proposes an embedding of lexical information using Wikipedia as source, and exploiting the resulting language model within the multitask learning process. The idea of (Collobert and Weston, 2008) to obtain an embedding of lexical information by acquiring a language model from unlabeled data is an interesting approach to the problem of performance degradation in out-of-domain tests, as already pursued by (Deschacht and Moens, 2009). The extensive use of unlabeled texts allows to achieve a significant level of lexical generalization that seems better capitalize the smaller annotated data sets. 3 A Distributional Model for Argument Classification High quality lexical information is crucial for robust open-domain SRL, as semantic generalization highly depends on lexical information. For example, the following two sentences evoke the STATEMENT frame, through the LUs say and state, where the FEs, SPEAKER and MEDIUM, are shown. [President Kennedy] SPEAKER said to an astronaut, ”Man is still the most extraordinary computer of all.” (1) [The report] MEDIUM stated, that some problems needed to be solved. (2) In sentence (1), for example, President Kennedy is the grammatical subject of the verb say and this justifies its role of SPEAKER. However, syntax does not entirely characterize argument semantics. In (1) and (2), the same syntactic relation is observed. It is the semantics of the grammatical heads, i.e. report and Kennedy, the main responsible for the difference between the two resulting proto-agentive roles, SPEAKER and MEDIUM. In this work we explore two different aspects. First, we propose a model that does not depend on complex syntactic information in order to minimize the risk of overfitting. Second, we improve the lexical semantic information available to the learning algorithm. The proposed ”minimalistic” approach will consider only two independent features: • the semantic head (h) of a role, as it can be observed in the grammatical structure. In sentence (2), for example, the MEDIUM FE is realized as the logical subject, whose head is report. 239 • the dependency relation (r) connecting the semantic head to the predicate words. In (2), the semantic head report is connected to the LU stated through the subject (SBJ) relation. In the rest of the section the distributional model for the argument classification step is presented. A lexicalized model for individual semantic roles is first defined in order to achieve robust semantic classification local to each argument. Then a Hidden Markov Model is introduced in order to exploit the local probability estimators, sensitive to lexical similarity, as well as the global information on the entire argument sequence. 3.1 Distributional Local Models As the classification of semantic roles is strictly related to the lexical meaning of argument heads, we adopt a distributional perspective, where the meaning is described by the set of textual contexts in which words appear. In distributional models, words are thus represented through vectors built over these observable contexts: similar vectors suggest semantic relatedness as a function of the distance between two words, capturing paradigmatic (e.g. synonymy) or syntagmatic relations (Pado, 2007). Vectors −→h are described by an adjacency matrix M, whose rows describe target words (h) and whose columns describe their corpus contexts. Latent Semantic Analysis (LSA) (Landauer and Dumais, 1997), is then applied to M to acquire meaningful representations −→h . LSA exploits the linear transformation called Singular Value Decomposition (SVD) and produces an approximation of the original matrix M, capturing (semantic) dependencies between context vectors. M is replaced by a lower dimensional matrix Ml, capturing the same statistical information in a new l-dimensional space, where each dimension is a linear combination of some of the original features (i.e. contexts). These derived features may be thought as artificial concepts, each one representing an emerging meaning component, as the linear combination of many different words. In the argument classification task, the similarity between two argument heads h1 and h2 observed in FrameNet can be computed over −→ h1 and −→ h2. The model for a given frame element FEk is built around the semantic heads h observed in the role FEk: they form a set denoted by HFEk. These LSA vectors −→h express the individual annotated examples as they are immerse in the LSA Role, FEk Clusters of semantic heads MEDIUM c1: {article, report, statement} c2: {constitution, decree, rule} SPEAKER c3: {brother, father, mother, sister } c4: {biographer, philosopher, ....} c5: {he, she, we, you} c6: {friend} TOPIC c7: {privilege, unresponsiveness} c8: {pattern} Table 1: Clusters of semantic heads in the Subj position for the frame STATEMENT with σ = 0.5 space acquired from the unlabeled texts. Moreover, given FEk, a model for each individual syntactic relation r (i.e. that links h labeled as FEk to their corresponding predicates) is a partition of the set HFEk called HFEk r , i.e. the subset of HFEk produced by examples of the relation r (e.g. Subj). Given the annotated sentence (2), we have that report ∈HMEDIUM SBJ . As the LSA vectors −→h are available for the semantic heads h, a vector representation −−→ FEk for the role FEk can be obtained from the annotated data. However, one single vector is a too simplistic representation given the rich nature of semantic roles FEk. In order to better represent FEk, multiple regions in the semantic space are used. They are obtained by a clustering process applied to the set HFEk r according to the Quality Threshold (QT) algorithm (Heyer et al., 1999). QT is a generalization of k-mean where a variable number of clusters can be obtained. This number depends on the minimal value of intra-cluster similarity accepted by the algorithm and controlled by a parameter, σ: lower values of σ correspond to more heterogeneous (i.e. larger grain) clusters, while values close to 1 characterize stricter policies and more fine-grained results. Given a syntactic relation r, CFEk r denotes the clusters derived by QT clustering over HFEk r . Each cluster c ∈CFEk r is represented by a vector −→c , computed as the geometric centroid of its semantic heads h ∈c. For a frame F, clusters define a geometric model of every frame elements FEk: it consists of centroids −→c with c ⊆HFEk r . Each c represents FEk through a set of similar heads, as role fillers observed in FrameNet. Table 1 represents clusters for the heads HFEk Subj of the STATEMENT frame. In argument classification we assume that the evoking predicate word for the frame F in an input sentence s is known. A sentence s can be seen as a sequence of role-relation pairs: 240 s = {(r1, h1), ..., (rn, hn)} where the heads hi are in the syntactic relation ri with the underlying lexical unit of F. For every head h in s, the vector −→h can be then used to estimate its similarity with the different candidate roles FEk. Given the syntactic relation r, the clusters c ∈CFEk r whose centroid vector ⃗c is closer to ⃗h are selected. Dr,h is the set of the representations semantically related to h: Dr,h = [ k {ckj ∈CFEk r |sim(h, ckj) ≥τ} (3) where the similarity between the j-th cluster for the FEk, i.e. ckj ∈CFEk r , and h is the usual cosine similarity: simcos(h, ckj) = − → h ·− →c kj ∥− → h ∥∥− →c kj∥ Then, through a k-nearest neighbours (k-NN) strategy within Dr,h, the m clusters ckj most similar to h are retained in the set D(m) r,h . A probabilistic preference for the role FEk is estimated for h through a cluster-based voting scheme, prob(FEk|r, h) = |CFEk r ∩D(m) r,h | |D(m) r,h | (4) or, alternatively, an instance-based one over D(m) r,h : prob(FEk|r, h) = P c∈CF Ek r ∩D(m) r,h |c| P c∈D(m) r,h |c| (5) In Fig. 1 the preference estimation for the incoming head h = professor connected to a LU by the Subj relation is shown. Clusters for the heads in Table 1 are also reported. First, in the set of clusters whose similarity with professor is higher than a threshold τ the m = 5 most similar clusters are selected. Accordingly, the preferences given by Eq. 4 are prob(SPEAKER|SBJ, h) = 3/5, prob(MEDIUM|SBJ, h) = 2/5 and prob(TOPIC|SBJ, h) = 0. The strategy modeled by Eq. 5 amplifies the role of larger clusters, e.g. prob(SPEAKER|SBJ, h) = 9/14 and prob(MEDIUM|SBJ, h) = 5/14. We call Distributional, the model that applies Eq. 5 to the source (r, h) arguments, by rejecting cases only when no information about the head h is available from the unlabeled corpus or no example of relation r for the role FEk is available from the annotated corpus. Eq. 4 and 5 in fact do not cover all possible cases. Often the incoming head h or the relation r may be unavailable: 1. If the head h has never been met in the unlabeled corpus or the high grammatical ambiguity of the sentence does not allow to locate it reliably, Eq. 4 (or 5) should be backed off to a purely syntactic model, that is prob(FEk|r) 2. If the relation r can not be properly located in s, h is also unknown: the prior probability of individual arguments, i.e. prob(FEk), is here employed. Both prob(FEk|r) and prob(FEk) can be estimated from the training set and smoothing can be also applied1. A more robust argument preference function for all arguments (ri, hi) ∈s of the frame F is thus given by: prob(FEk|ri, hi) = λ1prob(FEk|ri, hi) + λ2prob(FEk|ri) + λ3prob(FEk) (6) where weights λ1, λ2, λ3 can be heuristically assigned or estimated from the training set2. The resulting model is hereafter called Backoff model: although simply based on a single feature (i.e. the syntactic relation r), it accounts for information at different reliability degrees. 3.2 A Joint Model for Argument Classification Eq. 6 defines roles preferences local to individual arguments (ri, hi). However, an argument frame is a joint structure, with strong dependencies between arguments. We thus propose to model the reranking phase (RR) as a HMM sequence labeling task. It defines a stochastic inference over multiple (locally justified) alternative sequences through a Hidden Markov Model (HMM). It infers the best sequence FE(k1,...,kn) over all the possible hidden state sequences (i.e. made by the target FEki) given the observable emissions, i.e. the arguments (ri, hi). Viterbi inference is applied to build the best (role) interpretation for the input sentence. Once Eq. 6 is available, the best frame element sequence FE(θ(1),...,θ(n)) for the entire sentence s can be selected by defining the function θ(·) that maps arguments (ri, hi) ∈s to frame elements FEk: θ(i) = k s.t. FEk ∈F (7) 1Lindstone smoothing was applied with δ = 1. 2In each test discussed hereafter, λ1, λ2, λ3 were assigned to .9,.09 and .01, in order to impose a strict priority to the model contributions. 241 report statement article survey review constitution decree rule translator archaeologist philosopher biographer friend pattern president king sister mother brother father we she he you MEDIUM SPEAKER TOPIC target head professor manifesto privilege unresponsiveness Figure 1: A k-NN approach to the role classification for hi = professor Notice that different transfer functions θ(·) are usually possible. By computing their probability we can solve the SRL task by selecting the most likely interpretation, bθ(·), via argmaxθ P θ(·) | s , as follows: bθ(·) = argmax θ P s|θ(·) P θ(·) (8) In Eq. 8, the emission probability P s|θ(·) and the transition probability P θ(·) are explicit. Notice that the emission probability corresponds to an argument interpretation (e.g. Eq. 5) and it is assigned independently from the rest of the sentence. On the other hand, transition probabilities model role sequences and support the expectations about argument frames of a sentence. The emission probability is approximated as: P s | θ(1) . . . θ(n) ≈ n Y i=1 P(ri, hi | FEθ(i)) (9) as it is made independent from previous states in a Viterbi path. Again the emission probability can be rewritten as: P(ri, hi|FEθ(i)) = P(FEθ(i)|ri, hi) P(ri, hi) P(FEθ(i)) (10) Since P(ri, hi) does not depend on the role labeling, maximizing Eq. 10 corresponds to maximize: P(FEθ(i)|ri, hi) P(FEθ(i)) (11) whereas P(FEθ(i)|ri, hi) is thus estimated through Eq. 6. The transition probability, estimated through P θ(1) . . . θ(n) ≈ n Y i=1 P FEθ(i)|FEθ(i−1), FEθ(i−2) (12) accounts FEs sequence via a 3-gram model3 . 4 Empirical Analysis The aim of the evaluation is to measure the reachable accuracy of the simple model proposed and to compare its impact over in-domain and out-ofdomain semantic role labeling tasks. In particular, we will evaluate the argument classification (AC) task in Section 4.2. Experimental Set-Up. The in-domain test has been run over the FrameNet annotated corpus, derived from the British National Corpus (BNC). The splitting between train and test set is 90%10% according to the same data set of (Johansson and Nugues, 2008b). In all experiments, the FrameNet 1.3 version and the dependencybased system using the LTH parser (Johansson and Nugues, 2008a) have been employed. Outof-domain tests are run over the two training corpora as made available by the Semeval 2007 Task 194 (Baker et al., 2007): the Nuclear Threat Initiative (NTI) and the American National Corpus 3Two empty states are added at the beginning of any sequence. Moreover, Laplace smoothing was also applied to each estimator. 4The NTI and ANC annotated collections are downloadable at: nlp.cs.swarthmore.edu/semeval/tasks/task19/data/train.tar.gz 242 Corpus Predicates Arguments training FN-BNC 134,697 271,560 test in-domain FN-BNC 14,952 30,173 out-of-domain NTI 8,208 14,422 ANC 760 1,389 Table 2: Training and Testing data sets (ANC)5. Table 2 shows the predicates and arguments in each data set. All null-instantiated arguments were removed from the training and test sets. Vectors ⃗h representing semantic heads have been computed according to the ”dependencybased” vector space discussed in (Pado and Lapata, 2007). The entire BNC corpus has been parsed and the dependency graphs derived from individual sentences provided the basic observable contexts: every co-occurrence is thus syntactically justified by a dependency arc. The most frequent 30,000 basic features, i.e. (syntactic relation,lemma) pairs, have been used to build the matrix M, vector components corresponding to point-wise mutual information scores. Finally, the final space is obtained by applying the SVD reduction over M, with a dimensionality cut of l = 250. In the evaluation of the AC task, accuracy is computed over the nodes of the dependency graph, in line with (Johansson and Nugues, 2008b) or (Coppola et al., 2009). Accordingly, also recall, precision and F-measure are reported on a per node basis, against the binary BD task or for the full BD + AC chain. 4.1 The Role of Lexical Clustering The first study aims at detecting the impact of different clustering policies on the resulting AC accuracy. Clustering, as discussed in Section 3.1, allows to generalize lexical information: similar heads within the latent semantic space are built from the annotated examples and they allow to predict the behavior of new unseen words as found in the test sentences. The system performances have been here measured under different clustering conditions, i.e. grains at which the clustering of annotated examples is applied. This grain is determined by the parameter σ of the applied Quality Threshold algorithm (Heyer et al., 1999). Notice that small values of σ imply large clusters, while if 5Sentences whose arguments were not represented in the FrameNet training material were removed from all tests. Frames with a number of annotated examples Eq. - σ >0 >100 >500 >1K >3K >5K (5) - .85 86.3 86.5 87.2 88.3 85.9 82.0 (4) - .5 85.1 85.5 85.8 87.2 83.5 79.4 (4) - .1 84.5 84.8 85.1 86.5 83.0 78.7 Table 3: Accuracy on Arg classification tasks wrt different clustering policies σ ≈1 then many singleton clusters are promoted (i.e. one cluster for each example). By varying the threshold σ we thus account for prototype-based as well exemplar-based strategies, as discussed in (Erk, 2009). We measured the performance on the argument classification tasks of different models obtained by combing different choices of σ with Eq. (4) or (5). Results are reported in Table 3. The leftmost column reports the different clustering settings, while in the remaining columns we see performances over test sentences related to different frames: we selected frames for which an increasing number of annotated examples are available: from all frames (for more than 0 examples) to the only frame (i.e. SELF MOTION) that has more than 5,000 examples in our training data set. The reported accuracies suggest that Eq. (5), promoting an example driven strategy, better captures the role preference, as it always outperforms alternative settings (i.e. more prototype oriented methods). It limits overgeneralization and promotes fine grained clusters. An interesting result is that a per-node accuracy of 86.3 (i.e. only 3 points under the state of-the art on the same data set, (Johansson and Nugues, 2008b)) is achieved. All the remaining tests have been run with the clustering configuration characterized by Eq. (5) and σ = 0.85. 4.2 Argument Classification Accuracy In these experiments we evaluate the quality of the argument classification step against the lexical knowledge acquired from unlabeled texts and the reranking step. The accuracy reachable on the gold standard argument boundaries has been compared across several experimental settings. Two baseline systems have been obtained. The Local Prior model outputs the sequence that maximizes the prior probability locally to individual arguments. The Global Prior model is obtained by applying re-ranking (Section 3.2) to the best n = 10 candidates provided by the Local Prior model. Fi243 Model FN-BNC NTI ANC Local Prior 43.9 50.9 50.4 Global Prior 67.7 (+54.2%) 75.9 (+49.0%) 68.8 (+36.4%) Distributional 81.1 (+19.8%) 82.3 (+8.4%) 69.7 (+1.3%) Backoff 84.6 (+4.3%) 87.2 (+6.0%) 76.2 (+9.3%) Backoff+HMMRR 86.3 (+2.0%) 90.5 (+3.8%) 79.9 (+5.0%) (Johansson&Nugues, 2008) 89.9 71.1 Table 4: Accuracy of the Argument Classification task over the different corpora. In parenthesis the relative increment with respect to the immediately simpler model, previous row nally, the application of the backoff strategies (as in Eq. 6) and the HMM-based reranking characterize the final two configurations. Table 4 reports the accuracy results obtained over the three corpora (defined as in Table 2): the accuracy scores are averaged over different values of m in Eq. 5, ranging from 3 to 30. In the in-domain scenario, i.e. the FN-BNC dataset reported in column 2, it is worth noticing that the proposed model, with backoff and global reranking, is quite effective with respect to the state-of-the-art. Although results on the FN-BNC do not outperform the state-of-the-art for the FrameNet corpus, we still need to study the generalization capability of our SRL model in out-of-domain conditions. In a further experiment, we applied the same system, as trained over the FN-BNC data, to the other corpora, i.e. NTI and ANC, used entirely as test sets. Results, reported in column 3 and 4 of Table 4 and shown in Figure 2, confirm that no major drop in performance is observed. Notice how the positive impact of the backoff models and the HMM reranking policy is similarly reflected by all the collections. Moreover, the results on the NTI corpus are even better than those obtained on the BNC, with a resulting 90.5% accuracy on the AC task. 86,3% 90,5% 79,9% 40,0% 50,0% 60,0% 70,0% 80,0% 90,0% 100,0% Local Prior Global Prior Distributional Backoff Backoff +HMMRR FN-BNC NTI ANC Figure 2: Accuracy of the AC task over different corpora 4.3 Discussion The above empirical findings are relevant if compared with the outcome of a similar test on the NTI collection, discussed in (Johansson and Nugues, 2008b)6. There, under the same training conditions, a performance drop of about -19% is reported (from 89.9 to 71.1%) over gold standard argument boundaries. The model proposed in this paper exhibits no such drop in any collection (NTI and ANC). This seems to confirm the hypothesis that the model is able to properly generalize the required lexical information across different domains. It is interesting to outline that the individual stages of the proposed model play different roles in the different domains, as Table 4 suggests. Although the positive contributions of the individual processing stages are uniformly confirmed, some differences can be outlined: • The beneficial impact of the lexical information (i.e. the distributional model) applies differently across the different domains. The ANC domain seems not to significantly benefit when the distributional model (Eq. 5) is applied. Notice how Eq. 5 depends both from the evidence gathered in the corpus about lexical heads h as well as about the relation r. In ANC the percentage of times that the Eq. 5 is backed off against test instances (as h or r are not available from the training data) is twice as high as in the BNC-FN or in the NTI domain (i.e. 15.5 vs. 7.2 or 8.7, respectively). The different syntactic style of ANC seems thus the main responsible of the poor impact of distributional information, as it is often unapplicable to ANC test cases. • The complexity of the three test sets is different, as the three plots show. The NTI col6Notice that in this paper only the training portion of the NTI data set is employed as reported in Table 2 and results are not directly comparable to (Johansson and Nugues, 2008b). 244 lections seems characterized by a lower level of complexity (see for example the accuracy of the Local prior model, that is about 51% as for the ANC). It then gets benefits from all the analysis stages, in particular the final HMM reranking. The BNC-FN test collection seems the most complex one, and the impact of the lexical information brought by the distributional model is here maximal. This is mainly due to the coherence between the distributions of lexical and grammatical phenomena in the test and training data. • The role of HMM reranking is an effective way to compensate errors in the local argument classifications for all the three domains. However, it is particularly effective for the outside domain cases, while, in the BNC corpus, it produces just a small improvement instead (i.e. +2%, as shown in Table 4 ). It is worth noticing that the average length of the sentences in the BNC test collection is about 23 words per sentence, while it is higher for the NTI and ANC data sets (i.e. 34 and 31, respectively). It seems that the HMM model well captures some information on the global semantic structure of a sentence: this is helpful in cases where errors in the grammatical recognition (of individual arguments or at sentence level) are more frequent and afflict the local distributional model. The more complex is the syntax of a corpus (e.g. in the NTI and ANC data sets), the higher seems the impact of the reranking phase. The significant performance of the AC model here presented suggest to test it when integrated within a full SRL architecture. Table 5 reports the results of the processing cascade over three collections. Results on the Boundary Detection BD task are obtained by training an SVM model on the same feature set presented in (Johansson and Nugues, 2008b) and are slightly below the stateof-the art BD accuracy reported in (Coppola et al., 2009). However, the accuracy of the complete BD + AC + RR chain (i.e. 68%) improves the corresponding results of (Coppola et al., 2009). Given the relatively simple feature set adopted here, this result is very significant as for its resulting efficiency. The overall BD recognition process is, on a standard architecture, performed at about 6.74 sentences per second, that is basically Corpus Eval. Setting Recall Precision F1 BNC BD 72.6 85.1 78.4 BD+AC+RR 62.6 74.5 68.0 NTI BD 63.9 80.0 71.0 BD+AC+RR 56.7 72.1 63.5 ANC BD 64.0 81.5 71.7 BD+AC+RR 47.4 62.5 53.9 Table 5: Accuracy of the full cascade of the SRL system over three domain the same as the time needed for applying the entire BD + AC + RR chain, i.e. 6.21 sentence per second. 5 Conclusions In this paper, a distributional approach for acquiring a semi-supervised model of argument classification (AC) preferences has been proposed. It aims at improving the generalization capability of the inductive SRL approach by reducing the complexity of the employed grammatical features and through a distributional representation of lexical features. The obtained results are close to the state-of-art in FrameNet semantic parsing. State of the art accuracy is obtained instead in out-ofdomain experiments. The model seems to capitalize from simple methods of lexical modeling (i.e. the estimation of lexico-grammatical preferences through distributional analysis over unlabeled data), estimation (through syntactic or lexical back-off where necessary) and reranking. The result is an accurate and highly portable SRL cascade. Experiments on the integrated SRL architecture (i.e. BD + AC + RR chain) show that state-of-art accuracy (i.e. 68%) can be obtained on raw texts. This result is also very significant as for the achieved efficiency. The system is able to apply the entire BD + AC + RR chain at a speed of 6.21 sentences per second. This significant efficiency confirms the applicability of the SRL approach proposed here in large scale NLP applications. Future work will study the application of the flexible SRL method proposed to other languages, for which less resources are available and worst training conditions are the norm. Moreover, dimensionality reduction methods alternative to LSA, as currently studied on semisupervised spectral learning (Johnson and Zhang, 2008), will be experimented. 245 References Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998. The Berkeley FrameNet project. In Proc. of COLING-ACL, Montreal, Canada. Collin Baker, Michael Ellsworth, and Katrin Erk. 2007. Semeval-2007 task 19: Frame semantic structure extraction. In Proceedings of SemEval-2007, pages 99–104, Prague, Czech Republic, June. Association for Computational Linguistics. Xavier Carreras and Llu´ıs M`arquez. 2005. Introduction to the CoNLL-2005 Shared Task: Semantic Role Labeling. In Proc. of CoNLL-2005, pages 152–164, Ann Arbor, Michigan, June. Ronan Collobert and Jason Weston. 2008. A unified architecture for natural language processing: deep neural networks with multitask learning. In In Proceedings of ICML ’08, pages 160–167, New York, NY, USA. ACM. Bonaventura Coppola, Alessandro Moschitti, and Giuseppe Riccardi. 2009. Shallow semantic parsing for spoken language understanding. In Proceedings of NAACL ’09, pages 85–88, Morristown, NJ, USA. Koen Deschacht and Marie-Francine Moens. 2009. Semi-supervised semantic role labeling using the latent words language model. In EMNLP ’09: Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 21–29, Morristown, NJ, USA. Association for Computational Linguistics. Katrin Erk and Sebastian Pado. 2006. Shalmaneser a flexible toolbox for semantic role assignment. In Proceedings of LREC 2006, Genoa, Italy. Katrin Erk. 2009. Representing words as regions in vector space. In In Proceedings of CoNLL ’09, pages 57–65, Morristown, NJ, USA. Association for Computational Linguistics. Charles J. Fillmore. 1985. Frames and the semantics of understanding. Quaderni di Semantica, 4(2):222– 254. Hagen F¨urstenau and Mirella Lapata. 2009. Graph alignment for semi-supervised semantic role labeling. In In Proceedings of EMNLP ’09, pages 11–20, Morristown, NJ, USA. Daniel Gildea and Daniel Jurafsky. 2002. Automatic Labeling of Semantic Roles. Computational Linguistics, 28(3):245–288. Yoav Goldberg and Michael Elhadad. 2009. On the role of lexical features in sequence labeling. In In Proceedings of EMNLP ’09, pages 1142–1151, Singapore, August. Association for Computational Linguistics. L.J. Heyer, S. Kruglyak, and S. Yooseph. 1999. Exploring expression data: Identification and analysis of coexpressed genes. Genome Research, (9):1106– 1115. Richard Johansson and Pierre Nugues. 2008a. Dependency-based syntactic-semantic analysis with propbank and nombank. In Proceedings of CoNLL2008, Manchester, UK, August 16-17. Richard Johansson and Pierre Nugues. 2008b. The effect of syntactic representation on semantic role labeling. In Proceedings of COLING, Manchester, UK, August 18-22. Rie Johnson and Tong Zhang. 2008. Graph-based semi-supervised learning and spectral kernel design. IEEE Transactions on Information Theory, 54(1):275–288. Tom Landauer and Sue Dumais. 1997. A solution to plato’s problem: The latent semantic analysis theory of acquisition, induction and representation of knowledge. Psychological Review, 104. A. Moschitti, D. Pighin, and R. Basili. 2008. Tree kernels for semantic role labeling. Computational Linguistics, 34. Sebastian Pado and Mirella Lapata. 2007. Dependency-based construction of semantic space models. Computational Linguistics, 33(2). Sebastian Pado. 2007. Cross-Lingual Annotation Projection Models for Role-Semantic Information. Ph.D. thesis, Saarland University. Martha Palmer, Dan Gildea, and Paul Kingsbury. 2005. The proposition bank: A corpus annotated with semantic roles. Computational Linguistics, 31(1), March. Sameer S. Pradhan, Wayne Ward, and James H. Martin. 2008. Towards robust semantic role labeling. Comput. Linguist., 34(2):289–310. Erik F. Tjong Kim Sang and Sabine Buchholz. 2000. Introduction to the conll-2000 shared task: chunking. In Proceedings of the 2nd workshop on Learning language in logic and the 4th conference on Computational natural language learning, pages 127–132, Morristown, NJ, USA. Association for Computational Linguistics. Kristina Toutanova, Aria Haghighi, and Christopher D. Manning. 2008. A global joint model for semantic role labeling. Comput. Linguist., 34(2):161–191. 246
|
2010
|
25
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 247–256, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics A Bayesian Method for Robust Estimation of Distributional Similarities Jun’ichi Kazama Stijn De Saeger Kow Kuroda Masaki Murata† Kentaro Torisawa Language Infrastructure Group, MASTAR Project National Institute of Information and Communications Technology (NICT) 3-5 Hikaridai, Seika-cho, Soraku-gun, Kyoto, 619-0289 Japan {kazama, stijn, kuroda, torisawa}@nict.go.jp †Department of Information and Knowledge Engineering Faculty/Graduate School of Engineering, Tottori University 4-101 Koyama-Minami, Tottori, 680-8550 Japan∗ [email protected] Abstract Existing word similarity measures are not robust to data sparseness since they rely only on the point estimation of words’ context profiles obtained from a limited amount of data. This paper proposes a Bayesian method for robust distributional word similarities. The method uses a distribution of context profiles obtained by Bayesian estimation and takes the expectation of a base similarity measure under that distribution. When the context profiles are multinomial distributions, the priors are Dirichlet, and the base measure is the Bhattacharyya coefficient, we can derive an analytical form that allows efficient calculation. For the task of word similarity estimation using a large amount of Web data in Japanese, we show that the proposed measure gives better accuracies than other well-known similarity measures. 1 Introduction The semantic similarity of words is a longstanding topic in computational linguistics because it is theoretically intriguing and has many applications in the field. Many researchers have conducted studies based on the distributional hypothesis (Harris, 1954), which states that words that occur in the same contexts tend to have similar meanings. A number of semantic similarity measures have been proposed based on this hypothesis (Hindle, 1990; Grefenstette, 1994; Dagan et al., 1994; Dagan et al., 1995; Lin, 1998; Dagan et al., 1999). ∗The work was done while the author was at NICT. In general, most semantic similarity measures have the following form: sim(w1, w2) = g(v(w1), v(w2)), (1) where v(wi) is a vector that represents the contexts in which wi appears, which we call a context profile of wi. The function g is a function on these context profiles that is expected to produce good similarities. Each dimension of the vector corresponds to a context, fk, which is typically a neighboring word or a word having dependency relations with wi in a corpus. Its value, vk(wi), is typically a co-occurrence frequency c(wi, fk), a conditional probability p(fk|wi), or point-wise mutual information (PMI) between wi and fk, which are all calculated from a corpus. For g, various works have used the cosine, the Jaccard coefficient, or the Jensen-Shannon divergence is utilized, to name only a few measures. Previous studies have focused on how to devise good contexts and a good function g for semantic similarities. On the other hand, our approach in this paper is to estimate context profiles (v(wi)) robustly and thus to estimate the similarity robustly. The problem here is that v(wi) is computed from a corpus of limited size, and thus inevitably contains uncertainty and sparseness. The guiding intuition behind our method is as follows. All other things being equal, the similarity with a more frequent word should be larger, since it would be more reliable. For example, if p(fk|w1) and p(fk|w2) for two given words w1 and w2 are equal, but w1 is more frequent, we would expect that sim(w0, w1) > sim(w0, w2). In the NLP field, data sparseness has been recognized as a serious problem and tackled in the context of language modeling and supervised machine learning. However, to our knowledge, there 247 has been no study that seriously dealt with data sparseness in the context of semantic similarity calculation. The data sparseness problem is usually solved by smoothing, regularization, margin maximization and so on (Chen and Goodman, 1998; Chen and Rosenfeld, 2000; Cortes and Vapnik, 1995). Recently, the Bayesian approach has emerged and achieved promising results with a clearer formulation (Teh, 2006; Mochihashi et al., 2009). In this paper, we apply the Bayesian framework to the calculation of distributional similarity. The method is straightforward: Instead of using the point estimation of v(wi), we first estimate the distribution of the context profile, p(v(wi)), by Bayesian estimation and then take the expectation of the original similarity under this distribution as follows: simb(w1, w2) (2) = E[sim(w1, w2)]{p(v(w1)),p(v(w2))} = E[g(v(w1), v(w2))]{p(v(w1)),p(v(w2))}. The uncertainty due to data sparseness is represented by p(v(wi)), and taking the expectation enables us to take this into account. The Bayesian estimation usually gives diverging distributions for infrequent observations and thus decreases the expectation value as expected. The Bayesian estimation and the expectation calculation in Eq. 2 are generally difficult and usually require computationally expensive procedures. Since our motivation for this research is to calculate good semantic similarities for a large set of words (e.g., one million nouns) and apply them to a wide range of NLP tasks, such costs must be minimized. Our technical contribution in this paper is to show that in the case where the context profiles are multinomial distributions, the priors are Dirichlet, and the base similarity measure is the Bhattacharyya coefficient (Bhattacharyya, 1943), we can derive an analytical form for Eq. 2, that enables efficient calculation (with some implementation tricks). In experiments, we estimate semantic similarities using a large amount of Web data in Japanese and show that the proposed measure gives better word similarities than a non-Bayesian Bhattacharyya coefficient or other well-known similarity measures such as Jensen-Shannon divergence and the cosine with PMI weights. The rest of the paper is organized as follows. In Section 2, we briefly introduce the Bayesian estimation and the Bhattacharyya coefficient. Section 3 proposes our new Bayesian Bhattacharyya coefficient for robust similarity calculation. Section 4 mentions some implementation issues and the solutions. Then, Section 5 reports the experimental results. 2 Background 2.1 Bayesian estimation with Dirichlet prior Assume that we estimate a probabilistic model for the observed data D, p(D|φ), which is parameterized with parameters φ. In the maximum likelihood estimation (MLE), we find the point estimation φ∗= argmaxφp(D|φ). For example, we estimate p(fk|wi) as follows with MLE: p(fk|wi) = c(wi, fk)/ X k c(wi, fk). (3) On the other hand, the objective of the Bayesian estimation is to find the distribution of φ given the observed data D, i.e., p(φ|D), and use it in later processes. Using Bayes’ rule, this can also be viewed as: p(φ|D) = p(D|φ)pprior(φ) p(D) . (4) pprior(φ) is a prior distribution that represents the plausibility of each φ based on the prior knowledge. In this paper, we consider the case where φ is a multinomial distribution, i.e., ∑ k φk = 1, that models the process of choosing one out of K choices. Estimating a conditional probability distribution φk = p(fk|wi) as a context profile for each wi falls into this case. In this paper, we also assume that the prior is the Dirichlet distribution, Dir(α). The Dirichlet distribution is defined as follows. Dir(φ|α) = Γ(PK k=1 αk) QK k=1 Γ(αk) K Y k=1 φαk−1 k . (5) Γ(.) is the Gamma function. The Dirichlet distribution is parametrized by hyperparameters αk(> 0). It is known that p(φ|D) is also a Dirichlet distribution for this simplest case, and it can be analytically calculated as follows. p(φ|D) = Dir(φ|{αk + c(k)}), (6) where c(k) is the frequency of choice k in data D. For example, c(k) = c(wi, fk) in the estimation of p(fk|wi). This is very simple: we just need to add the observed counts to the hyperparameters. 248 2.2 Bhattacharyya coefficient When the context profiles are probability distributions, we usually utilize the measures on probability distributions such as the Jensen-Shannon (JS) divergence to calculate similarities (Dagan et al., 1994; Dagan et al., 1997). The JS divergence is defined as follows. JS(p1||p2) = 1 2(KL(p1||pavg) + KL(p2||pavg)), where pavg = p1+p2 2 is a point-wise average of p1 and p2 and KL(.) is the Kullback-Leibler divergence. Although we found that the JS divergence is a good measure, it is difficult to derive an efficient calculation of Eq. 2, even in the Dirichlet prior case.1 In this study, we employ the Bhattacharyya coefficient (Bhattacharyya, 1943) (BC for short), which is defined as follows: BC(p1, p2) = K X k=1 √p1k × p2k. The BC is also a similarity measure on probability distributions and is suitable for our purposes as we describe in the next section. Although BC has not been explored well in the literature on distributional word similarities, it is also a good similarity measure as the experiments show. 3 Method In this section, we show that if our base similarity measure is BC and the distributions under which we take the expectation are Dirichlet distributions, then Eq. 2 also has an analytical form, allowing efficient calculation. Here, we calculate the following value given two Dirichlet distributions: BCb(p1, p2) = E[BC(p1, p2)]{Dir(p1|α′ ),Dir(p2|β′ )} = ZZ △×△ Dir(p1|α ′)Dir(p2|β ′)BC(p1, p2)dp1dp2. After several derivation steps (see Appendix A), we obtain the following analytical solution for the above: 1A naive but general way might be to draw samples of v(wi) from p(v(wi)) and approximate the expectation using these samples. However, such a method will be slow. = Γ(α ′ 0)Γ(β ′ 0) Γ(α ′ 0 + 1 2)Γ(β ′ 0 + 1 2) K X k=1 Γ(α ′ k + 1 2)Γ(β ′ k + 1 2) Γ(α ′ k)Γ(β ′ k) , (7) where α ′ 0 = ∑ k α ′ k and β ′ 0 = ∑ k β ′ k. Note that with the Dirichlet prior, α ′ k = αk + c(w1, fk) and β ′ k = βk + c(w2, fk), where αk and βk are the hyperparameters of the priors of w1 and w2, respectively. To put it all together, we can obtain a new Bayesian similarity measure on words, which can be calculated only from the hyperparameters for the Dirichlet prior, α and β, and the observed counts c(wi, fk). It is written as follows. BCb(w1, w2) = (8) Γ(α0 + a0)Γ(β0 + b0) Γ(α0 + a0 + 1 2)Γ(β0 + b0 + 1 2) × K X k=1 Γ(αk + c(w1, fk) + 1 2)Γ(βk + c(w2, fk) + 1 2) Γ(αk + c(w1, fk))Γ(βk + c(w2, fk)) , where a0 = ∑ k c(w1, fk) and b0 = ∑ k c(w2, fk). We call this new measure the Bayesian Bhattacharyya coefficient (BCb for short). For simplicity, we assume αk = βk = α in this paper. We can see that BCb actually encodes our guiding intuition. Consider four words, w0, w1, w2, and w4, for which we have c(w0, f1) = 10, c(w1, f1) = 2, c(w2, f1) = 10, and c(w3, f1) = 20. They have counts only for the first dimension, i.e., they have the same context profile: p(f1|wi) = 1.0, when we employ MLE. When K = 10, 000 and αk = 1.0, the Bayesian similarity between these words is calculated as BCb(w0, w1) = 0.785368 BCb(w0, w2) = 0.785421 BCb(w0, w3) = 0.785463 We can see that similarities are different according to the number of observations, as expected. Note that the non-Bayesian BC will return the same value, 1.0, for all cases. Note also that BCb(w0, w0) = 0.78542 if we use Eq. 8, meaning that the self-similarity might not be the maximum. This is conceptually strange, although not a serious problem since we hardly use sim(wi, wi) in practice. If we want to fix this, we can use the special definition: BCb(wi, wi) ≡ 1. This is equivalent to using simb(wi, wi) = E[sim(wi, wi)]{p(v(wi))} = 1 only for this case. 249 4 Implementation Issues Although we have derived the analytical form (Eq. 8), there are several problems in implementing robust and efficient calculations. First, the Gamma function in Eq. 8 overflows when the argument is larger than 170. In such cases, a commonly used way is to work in the logarithmic space. In this study, we utilize the “log Gamma” function: lnΓ(x), which returns the logarithm of the Gamma function directly without the overflow problem.2 Second, the calculation of the log Gamma function is heavier than operations such as simple multiplication, which is used in existing measures. In fact, the log Gamma function is implemented using an iterative algorithm such as the Lanczos method. In addition, according to Eq. 8, it seems that we have to sum up the values for all k, because even if c(wi, fk) is zero the value inside the summation will not be zero. In the existing measures, it is often the case that we only need to sum up for k where c(wi, fk) > 0. Because c(wi, fk) is usually sparse, that technique speeds up the calculation of the existing measures drastically and makes it practical. In this study, the above problem is solved by pre-computing the required log Gamma values, assuming that we calculate similarities for a large set of words, and pre-computing default values for cases where c(wi, fk) = 0. The following values are pre-computed once at the start-up time. For each word: (A) lnΓ(α0 + a0) −lnΓ(α0 + a0 + 1 2) (B) lnΓ(αk+c(wi, fk))−lnΓ(αk+c(wi, fk)+ 1 2) for all k where c(wi, fk) > 0 (C) −exp(2(lnΓ(αk + 1 2) −lnΓ(αk)))) + exp(lnΓ(αk + c(wi, fk)) −lnΓ(αk + c(wi, fk) + 1 2) + lnΓ(αk + 1 2) −lnΓ(αk)) for all k where c(wi, fk) > 0; For each k: (D): exp(2(lnΓ(αk + 1 2)). In the calculation of BCb(w1, w2), we first assume that all c(wi, fk) = 0 and set the output variable to the default value. Then, we iterate over the sparse vectors c(w1, fk) and c(w2, fk). If 2We used the GNU Scientific Library (GSL) (www.gnu.org/software/gsl/), which implements this function. c(w1, fk) > 0 and c(w2, fk) = 0 (and vice versa), we update the output variable just by adding (C). If c(w1, fk) > 0 and c(w2, fk) > 0, we update the output value using (B), (D) and one additional exp(.) operation. With this implementation, we can make the computation of BCb practically as fast as using other measures. 5 Experiments 5.1 Evaluation setting We evaluated our method in the calculation of similarities between nouns in Japanese. Because human evaluation of word similarities is very difficult and costly, we conducted automatic evaluation in the set expansion setting, following previous studies such as Pantel et al. (2009). Given a word set, which is expected to contain similar words, we assume that a good similarity measure should output, for each word in the set, the other words in the set as similar words. For given word sets, we can construct input-andanswers pairs, where the answers for each word are the other words in the set the word appears in. We output a ranked list of 500 similar words for each word using a given similarity measure and checked whether they are included in the answers. This setting could be seen as document retrieval, and we can use an evaluation measure such as the mean of the precision at top T (MP@T) or the mean average precision (MAP). For each input word, P@T (precision at top T) and AP (average precision) are defined as follows. P@T = 1 T T X i=1 δ(wi ∈ans), AP = 1 R N X i=1 δ(wi ∈ans)P@i. δ(wi ∈ans) returns 1 if the output word wi is in the answers, and 0 otherwise. N is the number of outputs and R is the number of the answers. MP@T and MAP are the averages of these values over all input words. 5.2 Collecting context profiles Dependency relations are used as context profiles as in Kazama and Torisawa (2008) and Kazama et al. (2009). From a large corpus of Japanese Web documents (Shinzato et al., 2008) (100 million 250 documents), where each sentence has a dependency parse, we extracted noun-verb and nounnoun dependencies with relation types and then calculated their frequencies in the corpus. If a noun, n, depends on a word, w, with a relation, r, we collect a dependency pair, (n, 〈w, r〉). That is, a context fk, is 〈w, r〉here. For noun-verb dependencies, postpositions in Japanese represent relation types. For example, we extract a dependency relation (ワイン, 〈買う, を〉) from the sentence below, where a postposition “を(wo)” is used to mark the verb object. ワイン(wine) を(wo) 買う(buy) (≈buy a wine) Note that we leave various auxiliary verb suffixes, such as “れる(reru),” which is for passivization, as a part of w, since these greatly change the type of n in the dependent position. As for noun-noun dependencies, we considered expressions of type “n1 のn2” (≈“n2 of n1”) as dependencies (n1, 〈n2, の〉). We extracted about 470 million unique dependencies from the corpus, containing 31 million unique nouns (including compound nouns as determined by our filters) and 22 million unique contexts, fk. We sorted the nouns according to the number of unique co-occurring contexts and the contexts according to the number of unique cooccurring nouns, and then we selected the top one million nouns and 100,000 contexts. We used only 260 million dependency pairs that contained both the selected nouns and the selected contexts. 5.3 Test sets We prepared three test sets as follows. Set “A” and “B”: Thesaurus siblings We considered that words having a common hypernym (i.e., siblings) in a manually constructed thesaurus could constitute a similar word set. We extracted such sets from a Japanese dictionary, EDR (V3.0) (CRL, 2002), which contains concept hierarchies and the mapping between words and concepts. The dictionary contains 304,884 nouns. In all, 6,703 noun sibling sets were extracted with the average set size of 45.96. We randomly chose 200 sets each for sets “A” and “B.” Set “A” is a development set to tune the value of the hyperparameters and “B” is for the validation of the parameter tuning. Set “C”: Closed sets Murata et al. (2004) constructed a dataset that contains several closed word sets such as the names of countries, rivers, sumo wrestlers, etc. We used all of the 45 sets that are marked as “complete” in the data, containing 12,827 unique words in total. Note that we do not deal with ambiguities in the construction of these sets as well as in the calculation of similarities. That is, a word can be contained in several sets, and the answers for such a word is the union of the words in the sets it belongs to (excluding the word itself). In addition, note that the words in these test sets are different from those of our one-million-word vocabulary. We filtered out the words that are not included in our vocabulary and removed the sets with size less than 2 after the filtering. Set “A” contained 3,740 words that are actually evaluated, with about 115 answers on average, and “B” contained 3,657 words with about 65 answers on average. Set “C” contained 8,853 words with about 1,700 answers on average. 5.4 Compared similarity measures We compared our Bayesian Bhattacharyya similarity measure, BCb, with the following similarity measures. JS Jensen-Shannon divergence between p(fk|w1) and p(fk|w2) (Dagan et al., 1994; Dagan et al., 1999). PMI-cos The cosine of the context profile vectors, where the k-th dimension is the pointwise mutual information (PMI) between wi and fk defined as: PMI(wi, fk) = log p(wi,fk) p(wi)p(fk) (Pantel and Lin, 2002; Pantel et al., 2009).3 Cls-JS Kazama et al. (2009) proposed using the Jensen-Shannon divergence between hidden class distributions, p(c|w1) and p(c|w2), which are obtained by using an EM-based clustering of dependency relations with a model p(wi, fk) = ∑ c p(wi|c)p(fk|c)p(c) (Kazama and Torisawa, 2008). In order to 3We did not use the discounting of the PMI values described in Pantel and Lin (2002). 251 alleviate the effect of local minima of the EM clustering, they proposed averaging the similarities by several different clustering results, which can be obtained by using different initial parameters. In this study, we combined two clustering results (denoted as “s1+s2” in the results), each of which (“s1” and “s2”) has 2,000 hidden classes.4 We included this method since clustering can be regarded as another way of treating data sparseness. BC The Bhattacharyya coefficient (Bhattacharyya, 1943) between p(fk|w1) and p(fk|w2). This is the baseline for BCb. BCa The Bhattacharyya coefficient with absolute discounting. In calculating p(fk|wi), we subtract the discounting value, α, from c(wi, fk) and equally distribute the residual probability mass to the contexts whose frequency is zero. This is included as an example of naive smoothing methods. Since it is very costly to calculate the similarities with all of the other words (one million in our case), we used the following approximation method that exploits the sparseness of c(wi, fk). Similar methods were used in Pantel and Lin (2002), Kazama et al. (2009), and Pantel et al. (2009) as well. For a given word, wi, we sort the contexts in descending order according to c(wi, fk) and retrieve the top-L contexts.5 For each selected context, we sort the words in descending order according to c(wi, fk) and retrieve the top-M words (L = M = 1600).6 We merge all of the words above as candidate words and calculate the similarity only for the candidate words. Finally, the top 500 similar words are output. Note also that we used modified counts, log(c(wi, fk)) + 1, instead of raw counts, c(wi, fk), with the intention of alleviating the effect of strangely frequent dependencies, which can be found in the Web data. In preliminary experiments, we observed that this modification improves the quality of the top 500 similar words as reported in Terada et al. (2004) and Kazama et al. (2009). 4In the case of EM clustering, the number of unique contexts, fk, was also set to one million instead of 100,000, following Kazama et al. (2009). 5It is possible that the number of contexts with non-zero counts is less than L. In that case, all of the contexts with non-zero counts are used. 6Sorting is performed only once in the initialization step. Table 1: Performance on siblings (Set A). Measure MAP MP @1 @5 @10 @20 JS 0.0299 0.197 0.122 0.0990 0.0792 PMI-cos 0.0332 0.195 0.124 0.0993 0.0798 Cls-JS (s1) 0.0319 0.195 0.122 0.0988 0.0796 Cls-JS (s2) 0.0295 0.198 0.122 0.0981 0.0786 Cls-JS (s1+s2) 0.0333 0.206 0.129 0.103 0.0841 BC 0.0334 0.211 0.131 0.106 0.0854 BCb (0.0002) 0.0345 0.223 0.138 0.109 0.0873 BCb (0.0016) 0.0356 0.242 0.148 0.119 0.0955 BCb (0.0032) 0.0325 0.223 0.137 0.111 0.0895 BCa (0.0016) 0.0337 0.212 0.133 0.107 0.0863 BCa (0.0362) 0.0345 0.221 0.136 0.110 0.0890 BCa (0.1) 0.0324 0.214 0.128 0.101 0.0825 without log(c(wi, fk)) + 1 modification JS 0.0294 0.197 0.116 0.0912 0.0712 PMI-cos 0.0342 0.197 0.125 0.0987 0.0793 BC 0.0296 0.201 0.118 0.0915 0.0721 As for BCb, we assumed that all of the hyperparameters had the same value, i.e., αk = α. It is apparent that an excessively large α is not appropriate because it means ignoring observations. Therefore, α must be tuned. The discounting value of BCa is also tuned. 5.5 Results Table 1 shows the results for Set A. The MAP and the MPs at the top 1, 5, 10, and 20 are shown for each similarity measure. As for BCb and BCa, the results for the tuned and several other values for α are shown. Figure 1 shows the parameter tuning for BCb with MAP as the y-axis (results for BCa are shown as well). Figure 2 shows the same results with MPs as the y-axis. The MAP and MPs showed a correlation here. From these results, we can see that BCb surely improves upon BC, with 6.6% improvement in MAP and 14.7% improvement in MP@1 when α = 0.0016. BCb achieved the best performance among the compared measures with this setting. The absolute discounting, BCa, improved upon BC as well, but the improvement was smaller than with BCb. Table 1 also shows the results for the case where we did not use the log-modified counts. We can see that this modification gives improvements (though slight or unclear for PMI-cos). Because tuning hyperparameters involves the possibility of overfitting, its robustness should be assessed. We checked whether the tuned α with Set A works well for Set B. The results are shown in Table 2. We can see that the best α (= 0.0016) found for Set A works well for Set B as well. That is, the tuning of α as above is not unrealistic in 252 0.02 0.022 0.024 0.026 0.028 0.03 0.032 0.034 0.036 1e-06 1e-05 0.0001 0.001 0.01 0.1 1 MAP α (log-scale) Bayes Absolute Discounting Figure 1: Tuning of α (MAP). The dashed horizontal line indicates the score of BC. 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24 0.26 1e-06 1e-05 0.0001 0.001 0.01 MP α (log-scale) MP@1 MP@5 MP@10 MP@20 MP@30 MP@40 Figure 2: Tuning of α (MP). practice because it seems that we can tune it robustly using a small subset of the vocabulary as shown by this experiment. Next, we evaluated the measures on Set C, i.e., the closed set data. The results are shown in Table 3. For this set, we observed a tendency that is different from Sets A and B. Cls-JS showed a particularly good performance. BCb surely improves upon BC. For example, the improvement was 7.5% for MP@1. However, the improvement in MAP was slight, and MAP did not correlate well with MPs, unlike in the case of Sets A and B. We thought one possible reason is that the number of outputs, 500, for each word was not large enough to assess MAP values correctly because the average number of answers is 1,700 for this dataset. In fact, we could output more than 500 words if we ignored the cost of storage. Therefore, we also included the results for the case where L = M = 3600 and N = 2, 000. Even with this setting, however, MAP did not correlate well with MPs. Although Cls-JS showed very good performance for Set C, note that the EM clustering is very time-consuming (Kazama and Torisawa, 2008), and it took about one week with 24 CPU cores to get one clustering result in our computing environment. On the other hand, the preparation Table 2: Performance on siblings (Set B). Measure MAP MP @1 @5 @10 @20 JS 0.0265 0.208 0.116 0.0855 0.0627 PMI-cos 0.0283 0.203 0.116 0.0871 0.0660 Cls-JS (s1+s2) 0.0274 0.194 0.115 0.0859 0.0643 BC 0.0295 0.223 0.124 0.0922 0.0693 BCb (0.0002) 0.0301 0.225 0.128 0.0958 0.0718 BCb (0.0016) 0.0313 0.246 0.135 0.103 0.0758 BCb (0.0032) 0.0279 0.228 0.127 0.0938 0.0698 BCa (0.0016) 0.0297 0.223 0.125 0.0934 0.0700 BCa (0.0362) 0.0298 0.223 0.125 0.0934 0.0705 BCa (0.01) 0.0300 0.224 0.126 0.0949 0.0710 Table 3: Performance on closed-sets (Set C). Measure MAP MP @1 @5 @10 @20 JS 0.127 0.607 0.582 0.566 0.544 PMI-cos 0.124 0.531 0.519 0.508 0.493 Cls-JS (s1) 0.125 0.589 0.566 0.548 0.525 Cls-JS (s2) 0.137 0.608 0.592 0.576 0.554 Cls-JS (s1+s2) 0.152 0.638 0.617 0.603 0.583 BC 0.131 0.602 0.579 0.565 0.545 BCb (0.0004) 0.133 0.636 0.605 0.587 0.563 BCb (0.0008) 0.131 0.647 0.615 0.594 0.568 BCb (0.0016) 0.126 0.644 0.615 0.593 0.564 BCb (0.0032) 0.107 0.573 0.556 0.529 0.496 L = M = 3200 and N = 2000 JS 0.165 0.605 0.580 0.564 0.543 PMI-cos 0.165 0.530 0.517 0.507 0.492 Cls-JS (s1+s2) 0.209 0.639 0.618 0.603 0.584 BC 0.168 0.600 0.577 0.562 0.542 BCb (0.0004) 0.170 0.635 0.604 0.586 0.562 BCb (0.0008) 0.168 0.647 0.615 0.594 0.568 BCb (0.0016) 0.161 0.644 0.615 0.593 0.564 BCb (0.0032) 0.140 0.573 0.556 0.529 0.496 for our method requires just an hour with a single core. 6 Discussion We should note that the improvement by using our method is just “on average,” as in many other NLP tasks, and observing clear qualitative change is relatively difficult, for example, by just showing examples of similar word lists here. Comparing the results of BCb and BC, Table 4 lists the numbers of improved, unchanged, and degraded words in terms of MP@20 for each evaluation set. As can be seen, there are a number of degraded words, although they are fewer than the improved words. Next, Figure 3 shows the averaged differences of MP@20 in each 40,000 word-ID range.7 We can observe that the advantage of BCb is lessened es7Word IDs are assigned in ascending order when we chose the top one million words as described in Section 5.2, and they roughly correlate with frequencies. So, frequent words tend to have low-IDs. 253 Table 4: The numbers of improved, unchanged, and degraded words in terms of MP@20 for each evaluation set. # improved # unchanged # degraded Set A 755 2,585 400 Set B 643 2,610 404 Set C 3,153 3,962 1,738 -0.01 0 0.01 0.02 0.03 0.04 0.05 0.06 0 500000 1e+06 Avg. Diff. of MP@20 ID range -0.01 0 0.01 0.02 0.03 0.04 0.05 0.06 0 500000 1e+06 Avg. Diff. of MP@20 ID range -0.01 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0 500000 1e+06 Avg. Diff. of MP@20 ID range Figure 3: Averaged Differences of MP@20 between BCb (0.0016) and BC within each 40,000 ID range (Left: Set A. Right: Set B. Bottom: Set C). pecially for low-ID words (as expected) with onaverage degradation.8 The improvement is “on average” in this sense as well. One might suspect that the answer words tended to be low-ID words, and the proposed method is simply biased towards low-ID words because of its nature. Then, the observed improvement is a trivial consequence. Table 5 lists some interesting statistics about the IDs. We can see that BCb surely outputs more low-ID words than BC, and BC more than Cls-JS and JS.9 However, the average ID of the outputs of BC is already lower than the average ID of the answer words. Therefore, even if BCb preferred lower-ID words than BC, it should not have the effect of improving the accuracy. That is, the improvement by BCb is not superficial. From BC/BCb, we can also see that the IDs of the correct outputs did not become smaller compared to the IDs of the system outputs. Clearly, we need more analysis on what caused the improvement by the proposed method and how that affects the efficacy in real applications of similarity measures. The proposed Bayesian similarity measure outperformed the baseline Bhattacharyya coefficient 8This suggests the use of different αs depending on ID ranges (e.g., smaller α for low-ID words) in practice. 9The outputs of Cls-JS are well-balanced in the ID space. Table 5: Statistics on IDs. (A): Avg. ID of answers. (B): Avg. ID of system outputs. (C): Avg. ID of correct system outputs. Set A Set C (A) 238,483 255,248 (B) (C) (B) (C) Cls-JS (s1+s2) 282,098 176,706 273,768 232,796 JS 183,054 11,3442 211,671 201,214 BC 162,758 98,433 193,508 189,345 BCb(0.0016) 55,915 54,786 90,472 127,877 BC/BCb 2.91 1.80 2.14 1.48 and other well-known similarity measures. As a smoothing method, it also outperformed a naive absolute discounting. Of course, we cannot say that the proposed method is better than any other sophisticated smoothing method at this point. However, as noted above, there has been no serious attempt to assess the effect of smoothing in the context of word similarity calculation. Recent studies have pointed out that the Bayesian framework derives state-of-the-art smoothing methods such as Kneser-Ney smoothing as a special case (Teh, 2006; Mochihashi et al., 2009). Consequently, it is reasonable to resort to the Bayesian framework. Conceptually, our method is equivalent to modifying p(fk|wi) as p(fk|wi) = {Γ(α0+a0)Γ(αk+c(wi,fk)+ 1 2 ) Γ(α0+a0+ 1 2 )Γ(αk+c(wi,fk)) }2 and taking the Bhattacharyya coefficient. However, the implication of this form has not yet been investigated, and so we leave it for future research. Our method is the simplest one as a Bayesian method. We did not employ any numerical optimization or sampling iterations, as in a more complete use of the Bayesian framework (Teh, 2006; Mochihashi et al., 2009). Instead, we used the obtained analytical form directly with the assumption that αk = α and α can be tuned directly by using a simple grid search with a small subset of the vocabulary as the development set. If substantial additional costs are allowed, we can fine-tune each αk using more complete Bayesian methods. We also leave this for future research. In terms of calculation procedure, BCb has the same form as other similarity measures, which is basically the same as the inner product of sparse vectors. Thus, it can be as fast as other similarity measures with some effort as we described in Section 4 when our aim is to calculate similarities between words in a fixed large vocabulary. For example, BCb took about 100 hours to calculate the 254 top 500 similar nouns for all of the one million nouns (using 16 CPU cores), while JS took about 57 hours. We think this is an acceptable additional cost. The limitation of our method is that it cannot be used efficiently with similarity measures other than the Bhattacharyya coefficient, although that choice seems good as shown in the experiments. For example, it seems difficult to use the Jensen-Shannon divergence as the base similarity because the analytical form cannot be derived. One way we are considering to give more flexibility to our method is to adjust αk depending on external knowledge such as the importance of a context (e.g., PMIs). In another direction, we will be able to use a “weighted” Bhattacharyya coefficient: ∑ k µ(w1, fk)µ(w2, fk)√p1k × p2k, where the weights, µ(wi, fk), do not depend on pik, as the base similarity measure. The analytical form for it will be a weighted version of BCb. BCb can also be generalized to the case where the base similarity is BCd(p1, p2) = ∑K k=1 pd 1k × pd 2k, where d > 0. The Bayesian analytical form becomes as follows. BCd b (w1, w2) = Γ(α0 + a0)Γ(β0 + b0) Γ(α0 + a0 + d)Γ(β0 + b0 + d) × K X k=1 Γ(αk + c(w1, fk) + d)Γ(βk + c(w2, fk) + d) Γ(αk + c(w1, fk))Γ(βk + c(w2, fk)) . See Appendix A for the derivation. However, we restricted ourselves to the case of d = 1 2 in this study. Finally, note that our BCb is different from the Bhattacharyya distance measure on Dirichlet distributions of the following form described in Rauber et al. (2008) in its motivation and analytical form: p Γ(α ′ 0)Γ(β ′ 0) qQ k Γ(α ′ k) qQ k Γ(β ′ k) × Q k Γ((α ′ k + β ′ k)/2) Γ( 1 2 PK k (α ′ k + β ′ k)) . (9) Empirical and theoretical comparisons with this measure also form one of the future directions.10 7 Conclusion We proposed a Bayesian method for robust distributional word similarities. Our method uses a distribution of context profiles obtained by Bayesian 10Our preliminary experiments show that calculating similarity using Eq. 9 for the Dirichlet distributions obtained by Eq. 6 does not produce meaningful similarity (i.e., the accuracy is very low). estimation and takes the expectation of a base similarity measure under that distribution. We showed that, in the case where the context profiles are multinomial distributions, the priors are Dirichlet, and the base measure is the Bhattacharyya coefficient, we can derive an analytical form, permitting efficient calculation. Experimental results show that the proposed measure gives better word similarities than a non-Bayesian Bhattacharyya coefficient, other well-known similarity measures such as Jensen-Shannon divergence and the cosine with PMI weights, and the Bhattacharyya coefficient with absolute discounting. Appendix A Here, we give the analytical form for the generalized case (BCd b ) in Section 6. Recall the following relation, which is used to derive the normalization factor of the Dirichlet distribution: Z △ Y k φ α ′ k−1 k dφ = Q k Γ(α ′ k) Γ(α ′ 0) = Z(α ′)−1. (10) Then, BCd b (w1, w2) = ZZ △×△ Dir(φ1|α ′)Dir(φ2|β ′) X k φd 1kφd 2k dφ1 dφ2 = Z(α ′)Z(β ′) × ZZ △×△ Y l φ α ′ l−1 1l Y m φ β ′ m−1 2m X k φd 1kφd 2k dφ1 dφ2 | {z } A . Using Eq. 10, A in the above can be calculated as follows: = Z △ Y m φ β ′ m−1 2m 2 4X k φd 2k Z △ φ α ′ k+d−1 1k Y l̸=k φ α ′ l−1 1l dφ1 3 5 dφ2 = Z △ Y m φ β ′ m−1 2m "X k φd 2k Γ(α ′ k + d) Q l̸=k Γ(α ′ l) Γ(α ′ 0 + d) # dφ2 = X k Γ(α ′ k + d) Q l̸=k Γ(α ′ l) Γ(α ′ 0 + d) Z △ φ β ′ k+d−1 2k Y m̸=k φ β ′ m−1 2m dφ2 = X k Γ(α ′ k + d) Q l̸=k Γ(α ′ l) Γ(α ′ 0 + d) Γ(β ′ k + d) Q m̸=k Γ(β ′ m) Γ(β ′ 0 + d) = Q Γ(α ′ l) Q Γ(β ′ m) Γ(α ′ 0 + d)Γ(β ′ 0 + d) X k Γ(α ′ k + d) Γ(α ′ k) Γ(β ′ k + d) Γ(β ′ k) . This will give: BCd b (w1, w2) = Γ(α ′ 0)Γ(β ′ 0) Γ(α ′ 0 + d)Γ(β ′ 0 + d) K X k=1 Γ(α ′ k + d)Γ(β ′ k + d) Γ(α ′ k)Γ(β ′ k) . 255 References A. Bhattacharyya. 1943. On a measure of divergence between two statistical populations defined by their probability distributions. Bull. Calcutta Math. Soc., 49:214–224. Stanley F. Chen and Joshua Goodman. 1998. An empirical study of smoothing techniques for language modeling. TR-10-98, Computer Science Group, Harvard University. Stanley F. Chen and Ronald Rosenfeld. 2000. A survey of smoothing techniques for ME models. IEEE Transactions on Speech and Audio Processing, 8(1):37–50. Corinna Cortes and Vladimir Vapnik. 1995. Support vector networks. Machine Learning, 20:273–297. CRL. 2002. EDR electronic dictionary version 2.0 technical guide. Communications Research Laboratory (CRL). Ido Dagan, Fernando Pereira, and Lillian Lee. 1994. Similarity-based estimation of word cooccurrence probabilities. In Proceedings of ACL 94. Ido Dagan, Shaul Marcus, and Shaul Markovitch. 1995. Contextual word similarity and estimation from sparse data. Computer, Speech and Language, 9:123–152. Ido Dagan, Lillian Lee, and Fernando Pereira. 1997. Similarity-based methods for word sense disambiguation. In Proceedings of ACL 97. Ido Dagan, Lillian Lee, and Fernando Pereira. 1999. Similarity-based models of word cooccurrence probabilities. Machine Learning, 34(1-3):43–69. Gregory Grefenstette. 1994. Explorations In Automatic Thesaurus Discovery. Kluwer Academic Publishers. Zellig Harris. 1954. Distributional structure. Word, pages 146–142. Donald Hindle. 1990. Noun classification from predicate-argument structures. In Proceedings of ACL-90, pages 268–275. Jun’ichi Kazama and Kentaro Torisawa. 2008. Inducing gazetteers for named entity recognition by large-scale clustering of dependency relations. In Proceedings of ACL-08: HLT. Jun’ichi Kazama, Stijn De Saeger, Kentaro Torisawa, and Masaki Murata. 2009. Generating a large-scale analogy list using a probabilistic clustering based on noun-verb dependency profiles. In Proceedings of 15th Annual Meeting of The Association for Natural Language Processing (in Japanese). Dekang Lin. 1998. Automatic retrieval and clustering of similar words. In Proceedings of COLING/ACL98, pages 768–774. Daichi Mochihashi, Takeshi Yamada, and Naonori Ueda. 2009. Bayesian unsupervised word segmentation with nested Pitman-Yor language modeling. In Proceedings of ACL-IJCNLP 2009, pages 100– 108. Masaki Murata, Qing Ma, Tamotsu Shirado, and Hitoshi Isahara. 2004. Database for evaluating extracted terms and tool for visualizing the terms. In Proceedings of LREC 2004 Workshop: Computational and Computer-Assisted Terminology, pages 6–9. Patrick Pantel and Dekang Lin. 2002. Discovering word senses from text. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 613–619. Patrick Pantel, Eric Crestan, Arkady Borkovsky, AnaMaria Popescu, and Vishnu Vyas. 2009. Web-scale distributional similarity and entity set expansion. In Proceedings of EMNLP 2009, pages 938–947. T. W. Rauber, T. Braun, and K. Berns. 2008. Probabilistic distance measures of the Dirichlet and Beta distributions. Pattern Recognition, 41:637–645. Keiji Shinzato, Tomohide Shibata, Daisuke Kawahara, Chikara Hashimoto, and Sadao Kurohashi. 2008. Tsubaki: An open search engine infrastructure for developing new information access. In Proceedings of IJCNLP 2008. Yee Whye Teh. 2006. A hierarchical Bayesian language model based on Pitman-Yor processes. In Proceedings of COLING-ACL 2006, pages 985–992. Akira Terada, Minoru Yoshida, and Hiroshi Nakagawa. 2004. A tool for constructing a synonym dictionary using context information. In IPSJ SIG Technical Report (in Japanese), pages 87–94. 256
|
2010
|
26
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 257–265, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Recommendation in Internet Forums and Blogs Jia Wang Southwestern Univ. of Finance & Economics China [email protected] Qing Li Southwestern Univ. of Finance & Economics China liq [email protected] Yuanzhu Peter Chen Memorial Univ. of Newfoundland Canada [email protected] Zhangxi Lin Texas Tech Univ. USA zhangxi.lin @ttu.edu Abstract The variety of engaging interactions among users in social medial distinguishes it from traditional Web media. Such a feature should be utilized while attempting to provide intelligent services to social media participants. In this article, we present a framework to recommend relevant information in Internet forums and blogs using user comments, one of the most representative of user behaviors in online discussion. When incorporating user comments, we consider structural, semantic, and authority information carried by them. One of the most important observation from this work is that semantic contents of user comments can play a fairly different role in a different form of social media. When designing a recommendation system for this purpose, such a difference must be considered with caution. 1 Introduction In the past twenty years, the Web has evolved from a framework of information dissemination to a social interaction facilitator for its users. From the initial dominance of static pages or sites, with addition of dynamic content generation and provision of client-side computation and event handling, Web applications have become a prevalent framework for distributed GUI applications. Such technological advancement has fertilized vibrant creation, sharing, and collaboration among the users (Ahn et al., 2007). As a result, the role of Computer Science is not as much of designing or implementing certain data communication techniques, but more of enabling a variety of creative uses of the Web. In a more general context, Web is one of the most important carriers for “social media”, e.g. Internet forums, blogs, wikis, podcasts, instant messaging, and social networking. Various engaging interactions among users in social media differentiate it from traditional Web sites. Such characteristics should be utilized in attempt to provide intelligent services to social media users. One form of such interactions of particular interest here is user comments. In self-publication, or customer-generated media, a user can publish an article or post news to share with others. Other users can read and comment on the posting and these comments can, in turn, be read and commented on. Digg (www.digg.com), Yahoo!Buzz (buzz.yahoo.com) and various kinds of blogs are commercial examples of self-publication. Therefore, reader responses to earlier discussion provide a valuable source of information for effective recommendation. Currently, self-publishing media are becoming increasingly popular. For instance, at this point of writing, Technorati is indexing over 133 million blogs, and about 900,000 new blogs are created worldwide daily1. With such a large scale, information in the blogosphere follows a Long Tail Distribution (Agarwal et al., 2010). That is, in aggregate, the not-so-well-known blogs can have more valuable information than the popular ones. This gives us an incentive to develop a recommender to provide a set of relevant articles, which are expected to be of interest to the current reader. The user experience with the system can be immensely enhanced with the recommended articles. In this work, we focus on recommendation in Internet forums and blogs with discussion threads. Here, a fundamental challenge is to account for topic divergence, i.e. the change of gist during the process of discussion. In a discussion thread, the original posting is typically followed by other readers’ opinions, in the form of comments. Inten1http://technorati.com/ 257 tion and concerns of active users may change as the discussion goes on. Therefore, recommendation, if it were only based on the original posting, can not benefit the potentially evolving interests of the users. Apparently, there is a need to consider topic evolution in adaptive content-based recommendation and this requires novel techniques in order to capture topic evolution precisely and to prevent drastic topic shifting which returns completely irrelevant articles to users. In this work, we present a framework to recommend relevant information in Internet forums and blogs using user comments, one of the most representative recordings of user behaviors in these forms of social media. It has the following contributions. ∙ The relevant information is recommended based on a balanced perspective of both the authors and readers. ∙ We model the relationship among comments and that relative to the original posting using graphs in order to evaluate their combined impact. In addition, the weight of a comment is further enhanced with its content and with the authority of its poster. 2 Related Work In a broader context, a related problem is contentbased information recommendation (or filtering). Most information recommender systems select articles based on the contents of the original postings. For instance, Chiang and Chen (Chiang and Chen, 2004) study a few classifiers for agent-based news recommendations. The relevant news selections of these work are determined by the textual similarity between the recommended news and the original news posting. A number of later proposals incorporate additional metadata, such as user behaviors and timestamps. For example, Claypool et al. (Claypool et al., 1999) combine the news content with numerical user ratings. Del Corso, Gull´ı, and Romani (Del Corso et al., 2005) use timestamps to favor more recent news. Cantador, Bellogin, and Castells (Cantador et al., 2008) utilize domain ontology. Lee and Park (Lee and Park, 2007) consider matching between news article attributes and user preferences. Anh et al. (Ahn et al., 2007) and Lai, Liang, and Ku (Lai et al., 2003) construct explicit user profiles, respectively. Lavrenko et al. (Lavrenko et al., 2000) propose the e-Analyst system which combines news stories with trends in financial time series. Some go even further by ignoring the news contents and only using browsing behaviors of the readers with similar interests (Das et al., 2007). Another related problem is topic detection and tracking (TDT), i.e. automated categorization of news stories by their themes. TDT consists of breaking the stream of news into individual news stories, monitoring the stories for events that have not been seen before, and categorizing them (Lavrenko and Croft, 2001). A topic is modeled with a language profile deduced by the news. Most existing TDT schemes calculate the similarity between a piece of news and a topic profile to determine its topic relevance (Lavrenko and Croft, 2001) (Yang et al., 1999). Qiu (Qiu et al., 2009) apply TDT techniques to group news for collaborative news recommendation. Some work on TDT takes one step further in that they update the topic profiles as part of the learning process during its operation (Allan et al., 2002) (Leek et al., 2002). Most recent researches on information recommendation in social media focus on the blogosphere. Various types of user interactions in the blogosphere have been observed. A prominent feature of the blogosphere is the collective wisdom (Agarwal et al., 2010). That is, the knowledge in the blogosphere is enriched by such engaging interactions among bloggers and readers as posting, commenting and tagging. Prior to this work, the linking structure and user tagging mechanisms in the blogosphere are the most widely adopted ones to model such collective wisdom. For example, Esmaili et al. (Esmaili et al., 2006) focus on the linking structure among blogs. Hayes, Avesani, and Bojars (Hayes et al., 2007) explore measures based on blog authorship and reader tagging to improve recommendation. Li and Chen further integrate trust, social relation and semantic analysis (Li and Chen, 2009). These approaches attempt to capture accurate similarities between postings without using reader comments. Due to the interactions between bloggers and readers, blog recommendation should not limit its input to only blog postings themselves but also incorporate feedbacks from the readers. The rest of this article is organized as follows. We first describe the design of our recommendation framework in Section 3. We then evaluate the performance of such a recommender using two 258
! " # $ $ " # $ " # % & Figure 1: Design scheme different social media corpora (Section 4). This paper is concluded with speculation on how the current prototype can be further improved in Section 5. 3 System Design In this section, we present a mechanism for recommendation in Internet forums and blogs. The framework is sketched in Figure 1. Essentially, it builds a topic profile for each original posting along with the comments from readers, and uses this profile to retrieve relevant articles. In particular, we first extract structural, semantic, and authority information carried by the comments. Then, with such collective wisdom, we use a graph to model the relationship among comments and that relative to the original posting in order to evaluate the impact of each comment. The graph is weighted with postings’ contents and the authors’ authority. This information along with the original posting and its comments are fed into a synthesizer. The synthesizer balances views from both authors and readers to construct a topic profile to retrieve relevant articles. 3.1 Incorporating Comments In a discussion thread, comments made at different levels reflect the variation of focus of readers. Therefore, recommended articles should reflect their concerns to complement the author’s opinion. The degree of contribution from each comment, however, is different. In the extreme case, some of them are even advertisements which are completely irrelevant to the discussion topics. In this work, we use a graph model to differentiate the importance of each comment. That is, we model the authority, semantic, structural relations of comments to determine their combined impact. 3.1.1 Authority Scoring Comments Intuitively, each comment may have a different degree of authority determined by the status of its author (Hu et al., 2007). Assume we have 푛users in a forum, denoted by 푈 = { 푢1 ,푢2 ,...,푢푛} . We calculate the authority 푎푖for user 푢푖. To do that, we employ a variant of the PageRank algorithm (Brin and Page, 1998). We consider the cases that a user replies to a previous posting and that a user quotes a previous posting separately. For user 푢푗, we use 푙푟(푖,푗) to denote the number of times that 푢푗has replied to user 푢푖. Similarly, we use 푙푞(푖,푗) to denote the number of times that 푢푗has quoted user 푢푖. We combine them linearly: 푙′(푖,푗) = 훽1 푙푟(푖,푗) + 훽2 푙푞(푖,푗). Further, we normalize the above quantity to record how frequently a user refers to another: 푙(푖,푗) = 푙′(푖,푗) ∑ 푛 푘= 1 푙′(푖,푘) + 휖. Inline with the PageRank algorithm, we define the authority of user 푢푖as 푎푖= 휆 푛+ (1 − 휆) × 푛∑ 푘= 1 (푙(푘,푖) × 푎푘) . 3.1.2 Differentiating comments with Semantic and Structural relations Next, we construct a similar model in terms of the comments themselves. In this model, we treat the original posting and the comments each as a text node. This model considers both the content similarity between text nodes and the logic relationship among them. On the one hand, the semantic similarity between two nodes can be measured with any commonly adopted metric, such as cosine similarity and Jaccard coefficient (Baeza-Yates and RibeiroNeto, 1999). On the other hand, the structural relation between a pair of nodes takes two forms as we have discussed earlier. First, a comment can be made in response to the original posting or at most one earlier comment. In graph theoretic terms, the hierarchy can be represented as a tree 퐺푇= (푉,퐸푇), where 푉is the set of all text nodes and 퐸푇is the edge set. In particular, the original posting is the root and all the comments are ordinary nodes. There is an arc (directed edge) 푒푇∈ 퐸푇from node 푣to node 푢, denoted (푣,푢), if the corresponding comment 푢is made in response to comment (or original posting) 푣. Second, a comment can quote from one or more earlier comments. From this perspective, the hierarchy can be modeled using a directed acyclic graph (DAG), 259 M 0.8 0.5 0.8 0 0 0 0 0.5 0 0 0 0 0 0 0 0 0 1 0 0 0 1 0 0 0 0 0.5 0 1 0 0.8 C MT MD 2 1 3 2 1 3 Semantic Relation Quotation Relation Reply Relation M 0 0 0 1.5 0.8 Figure 2: Multi-relation graph of comments based on the structural and semantic information denoted 퐺퐷= (푉,퐸퐷). There is an arc 푒퐷∈ 퐸퐷 from node 푣to node 푢, denoted (푣,푢), if the corresponding comment 푢quotes comment (or original posting) 푣. As shown in Figure 2, for either graph 퐺푇or 퐺퐷, we can use a ∣푉∣× ∣푉∣adjacency matrix, denoted 푀 푇and 푀 퐷, respectively, to record them. Similarly, we can also use a ∣푉∣× ∣푉∣matrix defined on [0 ,1 ] to record the content similarity between nodes and denote it by 푀 퐶. Thus, we combine these three aspects linearly: 푀 = 훾1 × 푀 퐶+ 훾2 × 푀 푇+ 훾3 × 푀 퐷. The importance of a text node can be quantized by the times it has been referred to. Considering the semantic similarity between nodes, we use another variant of the PageRank algorithm to calculate the weight of comment 푗: 푠′ 푗= 휆 ∣푉∣+ (1 − 휆) × ∣푉∣ ∑ 푘= 1 푟푘,푗× 푠′ 푘, where 휆is a damping factor, and 푟푘,푗is the normalized weight of comment 푘referring to 푗defined as 푟(푘,푗) = 푀 푘,푗 ∑ 푗 푀 푘,푗+ 휖, where 푀 푘,푗is an entry in the graph adjacency matrix M and 휖is a constant to avoid division by zero. In some social networking media, a user may have a subset of other users as “friends”. This can be captured by a ∣푈∣× ∣푈∣matrix of { 0 ,1 } , whose entries are denoted by 푓푖,푗. Thus, with this information and assuming poster 푖has made a comment k for user 푗’s posting, the final weight of this comment is defined as 푠푘= 푠′ 푘× ( 푎푖+ 푓푖,푗 2 ) . 3.2 Topic Profile Construction Once the weight of comments on one posting is quantified by our models, this information along with the entire discussion thread is fed into a synthesizer to construct a topic profile. As such, the perspectives of both authors and readers are balanced for recommendation. The profile is a weight vector of terms to model the language used in the discussion thread. Consider a posting 푑0 and its comment sequence { 푑1 ,푑2 ,⋅⋅⋅,푑푚} . For each term 푡, a compound weight 푊(푡) = (1 − 훼) × 푊 1 (푡) + 훼× 푊 2 (푡) is calculated. It is a linear combination of the contribution by the posting itself, 푊 1 (푡), and that by the comments, 푊 2 (푡). We assume that each term is associated with an “inverted document frequency”, denoted by 퐼(푡) = lo g 푁 푛(푡) , where 푁 is the corpus size and 푛(푡) is the number of documents in corpus containing term 푡. We use a function 푓(푡,푑) to denote the number of occurrences of term 푡in document 푑, i.e. “term frequency”. Thus, when the original posting and comments are each considered as a document, this term frequency can be calculated for any term in any document. We thus define the weight of term 푡in document 푑, be the posting itself or a comment, using the standard TF/IDF definition (Baeza-Yates and Ribeiro-Neto, 1999): 푤(푡,푑) = ( 0 .5 + 0 .5 × 푓(푡,푑) m a x 푡′ 푓(푡′,푑) ) × 퐼(푡). The weight contributed by the posting itself, 푑0 , is thus: 푊 1 (푡) = 푤(푡,푑0 ) m a x 푡′ 푤(푡′,푑0 ) . The weight contribution from the comments { 푑1 ,푑2 ,⋅⋅⋅,푑푚} incorporates not only the language features of these documents but also their importance in the discussion thread. That is, the contribution of comment score is incorporated into weight calculation of the words in a comment. 푊 2 (푡) = 푚∑ 푖= 1 ( 푤(푡,푑푖) m a x 푡′ 푤(푡′,푑푖) ) × ( 푠(푖) m a x 푖′ 푠(푖′) ) . Such a treatment of compounded weight 푊(푡) is essentially to recognize that readers’ impact on selecting relevant articles and the difference of their influence. For each profile, we select the top푛highest weighted words to represent the topic. 260 With the topic profile thus constructed, the retriever returns an ordered list of articles with decreasing relevance to the topic. Note that our approach to differentiate the importance of each comment can be easily incorporated into any generic retrieval model. In this work, our retriever is adopted from (Lavrenko et al., 2000). 3.3 Interpretation of Recommendation Since interpreting recommended items enhances users’ trusting beliefs (Wang and Benbasat, 2007), we design a creative approach to generate hints to indicate the relationship (generalization, specialization and duplication) between the recommended articles and the original posting based on our previous work (Candan et al., 2009). Article 퐴being more general than 퐵can be interpreted as 퐴being less constrained than 퐵 by the keywords they contain. Let us consider two articles, 퐴and 퐵, where 퐴contains keywords, 푘1 and 푘2 , and 퐵only contains 푘1 . ∙ If 퐴is said to be more general than 퐵, then the additional keyword, 푘2 , of article 퐴must render 퐴less constrained than 퐵. Therefore, the content of 퐴can be interpreted as 푘1 ∪푘2 . ∙ If, on the other hand, 퐴is said to be more specific than 퐵, then the additional keyword, 푘2 , must render 퐴more constrained than 퐵. Therefore, the content of 퐴can be interpreted as 푘1 ∩ 푘2 . Note that, in the two-keyword space ⟨푘1 ,푘2 ⟩, 퐴 can be denoted by a vector ⟨푎퐴,푏퐴⟩and 퐵can be denoted by ⟨푎퐵,0 ⟩. The origin 푂 = ⟨0 ,0 ⟩corresponds to the case where an article does contain neither 푘1 nor 푘2 . That is, 푂 corresponds to an article which can be interpreted as ¬ 푘1 ∩ ¬ 푘2 ≡ ¬ (푘1 ∪ 푘2 ). Therefore, if 퐴 is said to be more general than 퐵, Δ 퐴= 푑(퐴,푂) should be greater than Δ 퐵 = 푑(퐵,푂). This allows us to measure the degrees of generalization and specialization of two articles. Given two articles, 퐴and 퐵, of the same topic, they will have a common keyword base, while both articles will also have their own content, different from their common base. Let us denote the common part of 퐴by 퐴푐and common part of 퐵by 퐵푐. Note that Δ 퐴퐶and Δ 퐵퐶 are usually unequal because the same words in the common part have different term weights in article 퐴and 퐵respectively. Given these and the generalization concept introduced above for two similar articles 퐴and 퐵, we can define the degree of generalization (퐺퐴퐵) and specialization (푆퐴퐵) of 퐵 with respect to 퐴as 퐺퐴퐵= Δ 퐴/ Δ 퐵푐,푆퐴퐵= Δ 퐵/ Δ 퐴푐. To alleviate the effect of document length, we revise the definition as 퐺퐴퐵= Δ 퐴/ lo g (Δ 퐴) Δ 퐵푐/ lo g (Δ 퐴+ Δ 퐵) , 푆퐴퐵= Δ 퐵/ lo g (Δ 퐵) Δ 퐴푐/ lo g (Δ 퐴+ Δ 퐵) . The relative specialization and generalization values can be used to reveal the relationships between recommended articles and the original posting. Given original posting 퐴and recommended article 퐵, if 퐺퐴퐵> Θ 푔, for a given generalization threshold Θ 푔, then B is marked as a generalization. When this is not the case, if 푆퐴퐵> Θ 푠, for a given specialization threshold, Θ 푠, then 퐵is marked as a specialization. If neither of these cases is true, then 퐵is duplicate of 퐴. Such an interpretation provides a control on delivering recommended articles. In particular, we can filter the duplicate articles to avoid recommending the same information. 4 Experimental Evaluation To evaluate the effectiveness of our proposed recommendation mechanism, we carry out a series of experiments on two synthetic data sets, collected from Internet forums and blogs, respectively. The first data set is called Forum. This data set is constructed by randomly selecting 20 news articles with corresponding reader comments from the Digg Web site and 16,718 news articles from the Reuters news Web site. This simulates the scenario of recommending relevant news from traditional media to social media users for their further reading. The second one is the Blog data set containing 15 blog articles with user comments and 15,110 articles obtained from the Myhome Web site 2. Details of these two data sets are shown in Table 1. For evaluation purposes, we adopt the traditional pooling strategy (Zobel, 1998) and apply to the TREC data set to mark the relevant articles for each topic. 2http://blogs.myhome.ie 261 Table 1: Evaluation data set Synthetic Data Set Forum Blog Topics No. of postings 20 15 Ave. length of postings 676 236 No. of comments per posting 81.4 46 Ave. length of comments 45 150 Target No. of articles 16718 15110 Ave. length of articles 583 317 The recommendation engine may return a set of essentially the same articles re-posted at different sites. Therefore, we introduce a metric of novelty to measure the topic diversity of returned suggestions. In our experiments, we define precision and novelty metrics as 푃@ 푁 = ∣퐶∩ 푅∣ ∣푅∣ and 퐷@ 푁 = ∣퐸∩ 푅∣ ∣푅∣ , where 푅is the subset of the top-푛articles returned by the recommender, 퐶 is the set of manually tagged relevant articles, and 퐸is the set of manually tagged relevant articles excluding duplicate ones to the original posting. We select the top 10 articles for evaluation assuming most readers only browse up to 10 recommended articles (Karypis, 2001). Meanwhile, we also utilize mean average precision (MAP) and mean average novelty (MAN) to evaluate the entire set of returned article. We test our proposal in four aspects. First, we compare our work to two baseline works. We then present results for some preliminary tests to find out the optimal values for two critical parameters. Next, we study the effect of user authority and its integration to comment weighting. Fourth, we evaluate the performance gain obtained from interpreting recommendation. In addition, we provide a significance test to show that the observed differences in effectiveness for different approaches are not incidental. In particular, we use the 푡-test here, which is commonly used for significance tests in information retrieval experiments (Hull, 1993). 4.1 Overall Performance As baseline proposals, we also implement two well-known content-based recommendation methods (Bogers and Bosch, 2007). The first method, Okapi, is commonly applied as a representative of the classic probabilistic model for relevant information retrieval (Robertson and Walker, 1994). The second one, LM, is based on statistical language models for relevant information retrieval (Ponte and Croft, 1998). It builds a probaTable 2: Overall performance Precision Novelty Data Method 푃@ 1 0 푀퐴푃 퐷@ 1 0 푀퐴푁 Forum Okapi 0.827 0.833 0.807 0.751 LM 0.804 0.833 0.807 0.731 Our 0.967 0.967 0.9 0.85 Blog Okapi 0.733 0.651 0.667 0.466 LM 0.767 0.718 0.70 0.524 Our 0.933 0.894 0.867 0.756 bilistic language model for each article, and ranks them on query likelihood, i.e. the probability of the model generating the query. Following the strategy of Bogers and Bosch, relevant articles are selected based on the title and the first 10 sentences of the original postings. This is because articles are organized in the so-called inverted pyramid style, meaning that the most important information is usually placed at the beginning. Trimming the rest of an article would usually remove relatively less crucial information, which speeds up the recommendation process. A paired 푡-test shows that using 푃@ 1 0 and 퐷@ 1 0 as performance measures, our approach performs significantly better than the baseline methods for both Forum and Blog data sets as shown in Table 2. In addition, we conduct 푡-tests using MAP and MAN as performance measures, respectively, and the 푝-values of these tests are all less than 0.05, meaning that the results of experiments are statistically significant. We believe that such gains are introduced by the additional information from the collective wisdom, i.e. user authority and comments. Note that the retrieval precision for Blog of two baseline methods is not as good as that for Forum. Our explanation is that blog articles may not be organized in the inverted pyramid style as strictly as news forum articles. 4.2 Parameters of Topic Profile There are two important parameters to be considered to construct topic profiles for recommendation. 1) the number of the most weighted words to represent the topic, and 2) combination coefficient 훼to determine the contribution of original posting and comments in selecting relevant articles.We conduct a series of experiments and find out that the optimal performance is obtained when the number of words is between 50 and 70, and 훼is between 0.65 and 0.75. When 훼is set to 0, the recommended articles only reflect the author’s opinion. When 훼= 1 , the suggested articles represent the concerns of readers exclusively. In the 262 Table 3: Performance of four runs Precision Novelty Method 푃@ 1 0 푀퐴푃 퐷@ 1 0 푀퐴푁 Forum RUN1 0.88 0.869 0.853 0.794 RUN2 0.933 0.911 0.9 0.814 RUN3 0.94 0.932 0.9 0.848 RUN4 0.967 0.967 0.9 0.85 Blog RUN1 0.767 0.758 0.7 0.574 RUN2 0.867 0.828 0.833 0.739 RUN3 0.9 0.858 0.833 0.728 RUN4 0.933 0.894 0.867 0.756 following experiments, we set topic word number to 60 and combination coefficient 훼to 0.7. 4.3 Effect of Authority and Comments In this part, we explore the contribution of user authority and comments in social media recommender. In particular, we study the following scenarios with increasing system capabilities. Note that, lacking friend information (Section 3.1.2) in the Forum data set, 푓푖,푗is set to zero. ∙ RUN 1 (Posting): the topic profile is constructed only based on the original posting itself. This is analogous to traditional recommenders which only consider the focus of authors for suggesting further readings. ∙ RUN 2 (Posting+Authority): the topic profile is constructed based on the original posting and participant authority. ∙ RUN 3 (Posting+Comment): the topic profile is constructed based on the original posting and its comments. ∙ RUN 4 (All): the topic profile is constructed based on the original posting, user authority, and its comments. Here, we set 훾1 = 훾2 = 훾3 = 1 . Our 푡-test shows that using 푃@ 1 0 and 퐷@ 1 0 as performance measures, RUN4 performs best in both Forum and Blog data sets as shown in Table 3. There is a stepwise performance improvement while integrating user authority, comments and both. With the assistance of user authority and comments, the recommendation precision is improved up to 9.8% and 21.6% for Forum and Blog, respectively. The opinion of readers is an effective complementarity to the authors’ view in suggesting relevant information for further reading. Moreover, we investigate the effect of the semantic and structural relations among comments, i.e. semantic similarity, reply, and quotation. For this purpose, we carry out a series of experiments based on different combinations of these relations. CR RR QR CQR CRR QRR All MAP 0.6 0.7 0.8 0.9 1.0 Forum Data Set Blog Data Set Figure 3: Effect of content, quotation and reply relation ∙ Content Relation (CR): only the content relation matrix is used in scoring the comments. ∙ Quotation Relation (QR): only the quotation relation matrix is used in scoring the comments. ∙ Reply Relation (RR): only the reply relation matrix is used in scoring the comments. ∙ Content+Quotation Relation (CQR): both the content and quotation relation matrices is used in scoring the comments. ∙ Content+Reply Relation(CRR): both the content and reply relation matrices are used in scoring the comments. ∙ Quotation+Reply Relation (QRR): both the quotation and reply relation matrices are used in scoring the comments. ∙ All: all three matrices are used. The MAP yielded by these combinations for both data sets is plotted in Figure 3. For the case of Forum, we observe that incorporating content information adversely affects recommendation precision. This concurs with what we saw in our previous work (Wang et al., 2010). On the other hand, when we test the Blog data set, the trend is the opposite, i.e. content similarity does contribute to retrieval performance positively. This is attributed by the text characteristics of these two forms of social media. Specifically, comments in news forums usually carry much richer structural information than blogs where comments are usually “flat” among themselves. 4.4 Recommendation Interpretation To evaluate the precision of interpreting the relationship between recommended articles and the 263 original posting, the evaluation metric of success rate 푆is defined as 푆= 푚∑ 푖= 1 (1 − 푒푖)/ 푚, where 푚 is the number of recommended articles, 푒푖is the error weight of recommended article 푖. Here, the error weight is set to one if the result interpretation is mis-labelled. From our studies, we observe that the success rate at top-10 is around 89.3% and 87.5% for the Forum and Blog data sets, respectively. Note that these rates include the errors introduced by the irrelevant articles returned by the retrieval module. To estimate optimal thresholds of generalization and specialization, we calculate the success rate at different threshold values and find that neither too small nor too large a value is appropriate for interpretation. In our experiments, we set generalization threshold Θ 푔to 3.2 and specialization threshold Θ 푠to 1.8 for the Forum data set, and Θ 푔to 3.5 and Θ 푠to 2.0 for Blog. Ideally, threshold values would need to be set through a machine learning process, which identifies proper values based on a given training sample. 5 Conclusion and Future Work The Web has become a platform for social networking, in addition to information dissemination at its earlier stage. Many of its applications are also being extended in this fashion. Traditional recommendation is essentially a push service to provide information according to the profile of individual or groups of users. Its niche at the Web 2.0 era lies in its ability to enable online discussion by serving up relevant references to the participants. In this work, we present a framework for information recommendation in such social media as Internet forums and blogs. This model incorporates information of user status and comment semantics and structures within the entire discussion thread. This framework models the logic connections among readers and the innovativeness of comments. By combining such information with traditional statistical language models, it is capable of suggesting relevant articles that meet the dynamic nature of a discussion in social media. One important discovery from this work is that, when integrating comment contents, the structural information among comments, and reader relationship, it is crucial to distinguish the characteristics of various forms of social media. The reason is that the role that the semantic content of a comment plays can differ from one form to another. This study can be extended in a few interesting ways. For example, we can also evaluate its effectiveness and costs during the operation of a discussion forum, where the discussion thread is continually updated by new comments and votes. Indeed, its power is yet to be further improved and investigated. Acknowledgments Li’s research is supported by National Natural Science Foundation of China (Grant No.60803106), the Scientific Research Foundation for the Returned Overseas Chinese Scholars, State Education Ministry, and the Fok Ying-Tong Education Foundation for Young Teachers in the Higher Education Institutions of China. Research of Chen is supported by Natural Science and Engineering Council (NSERC) of Canada. References Nitin Agarwal, Magdiel Galan, Huan Liu, and Shankar Subramanya. 2010. Wiscoll: Collective wisdom based blog clustering. Information Sciences, 180(1):39–61. Jae-wook Ahn, Peter Brusilovsky, Jonathan Grady, Daqing He, and Sue Yeon Syn. 2007. Open user profiles for adaptive news systems: help or harm? In Proceedings of the 16th International Conference on World Wide Web (WWW), pages 11–20. James Allan, Victor Lavrenko, and Russell Swan. 2002. Explorations within topic tracking and detection. Topic detection and tracking: event-based information organization Kluwer Academic Publishers, pages 197–224. Ricardo Baeza-Yates and Berthier Ribeiro-Neto. 1999. Modern information retrieval. Addison Wesley Longman Publisher. Toine Bogers and Antal Bosch. 2007. Comparing and evaluating information retrieval algorithms for news recommendation. In Proceedings of 2007 ACM conference on Recommender Systems, pages 141–144. Sergey Brin and Lawrence Page. 1998. The anatomy of a large-scale hypertextual web search engine. Computer networks and ISDN systems, 30(1-7):107–117. K. Selc¸uk Candan, Mehmet E. D¨onderler, Terri Hedgpeth, Jong Wook Kim, Qing Li, and Maria Luisa Sapino. 2009. SEA: Segment-enrich-annotate paradigm for adapting dialog-based content for improved accessibility. ACM Transactions on Information Systems (TOIS), 27(3):1–45. 264 Ivan Cantador, Alejandro Bellogin, and Pablo Castells. 2008. Ontology-based personalized and contextaware recommendations of news items. In Proceedings of IEEE/WIC/ACM international Conference on Web Intelligence and Intelligent Agent Technology (WI), pages 562–565. Jung-Hsien Chiang and Yan-Cheng Chen. 2004. An intelligent news recommender agent for filtering and categorizing large volumes of text corpus. International Journal of Intelligent Systems, 19(3):201– 216. Mark Claypool, Anuja Gokhale, Tim Miranda, Pavel Murnikov, Dmitry Netes, and Matthew Sartin. 1999. Combining content-based and collaborative filters in an online newspaper. In Proceedings of the ACM SIGIR Workshop on Recommender Systems. Abhinandan S. Das, Mayur Datar, Ashutosh Garg, and Shyam Rajaram. 2007. Google news personalization: scalable online collaborative filtering. In Proceedings of the 16th International Conference on World Wide Web (WWW), pages 271–280. Gianna M. Del Corso, Antonio Gull´ı, and Francesco Romani. 2005. Ranking a stream of news. In Proceedings of the 14th International Conference on World Wide Web(WWW), pages 97–106. Kyumars Sheykh Esmaili, Mahmood Neshati, Mohsen Jamali, Hassan Abolhassani, and Jafar Habibi. 2006. Comparing performance of recommendation techniques in the blogsphere. In ECAI 2006 Workshop on Recommender Systems. Conor Hayes, Paolo Avesani, and Uldis Bojars. 2007. An analysis of bloggers, topics and tags for a blog recommender system. In Workshop on Web Mining (WebMine), pages 1–20. Meishan Hu, Aixin Sun, and Ee-Peng Lim. 2007. Comments-oriented blog summarization by sentence extraction. In Proceedings of the sixteenth ACM Conference on Conference on Information and Knowledge Management(CIKM), pages 901–904. David Hull. 1993. Using statistical testing in the evaluation of retrieval experiments. In Proceedings of the 16th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 329–338. George Karypis. 2001. Evaluation of item-based TopN recommendation algorithms. In Proceedings of the 10th International Conference on Information and Knowledge Management (CIKM), pages 247– 254. Hung-Jen Lai, Ting-Peng Liang, and Yi Cheng Ku. 2003. Customized internet news services based on customer profiles. In Proceedings of the 5th International Conference on Electronic commerce (ICEC), pages 225–229. Victor Lavrenko and W. Bruce Croft. 2001. Relevance based language models. In Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 120–127. Victor Lavrenko, Matt Schmill, Dawn Lawrie, Paul Ogilvie, David Jensen, and James Allan. 2000. Language models for financial news recommendation. In Proceedings of the 9th International Conference on Information and Knowledge Management (CIKM), pages 389–396. Hong Joo Lee and Sung Joo Park. 2007. MONERS: A news recommender for the mobile web. Expert Systems with Applications, 32(1):143–150. Tim Leek, Richard Schwartz, and Srinivasa Sista. 2002. Probabilistic approaches to topic detection and tracking. Topic detection and tracking: eventbased information organization, pages 67–83. Yung-Ming Li and Ching-Wen Chen. 2009. A synthetical approach for blog recommendation: Combining trust, social relation, and semantic analysis. Expert Systems with Applications, 36(3):6536 – 6547. Jay Michael Ponte and William Bruce Croft. 1998. A language modeling approach to information retrieval. In Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 275–281. Jing Qiu, Lejian Liao, and Peng Li. 2009. News recommender system based on topic detection and tracking. In Proceedings of the 4th Rough Sets and Knowledge Technology. Stephen E. Robertson and Stephen G Walker. 1994. Some simple effective approximations to the 2poisson model for probabilistic weighted retrieval. In Proceedings of the 17th ACM SIGIR conference on Research and Development in Information Retrieval, pages 232–241. Weiquan Wang and Izak Benbasat. 2007. Recommendation agents for electronic commerce: Effects of explanation facilities on trusting beliefs. Journal of Management Information Systems, 23(4):217–246. Jia Wang, Qing Li, and Yuanzhu Peter Chen. 2010. User comments for news recommendation in social media. In Proceedings of the 33rd ACM SIGIR Conference on Research and Development in Information Retrieval, pages 295–296. Yiming Yang, Jaime Guillermo Carbonell, Ralf D. Brown, Thomas Pierce, Brian T. Archibald, and Xin Liu. 1999. Learning approaches for detecting and tracking news events. IEEE Intelligent Systems, 14(4):32–43. Justin Zobel. 1998. How reliable are the results of large-scale information retrieval experiments? In Proceedings of the 21st International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 307–314. 265
|
2010
|
27
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 266–274, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Learning Phrase-Based Spelling Error Models from Clickthrough Data Xu Sun∗ Dept. of Mathematical Informatics University of Tokyo, Tokyo, Japan [email protected] Jianfeng Gao Microsoft Research Redmond, WA, USA [email protected] Daniel Micol Microsoft Corporation Munich, Germany [email protected] Chris Quirk Microsoft Research Redmond, WA, USA [email protected] ∗ The work was done when Xu Sun was visiting Microsoft Research Redmond. Abstract This paper explores the use of clickthrough data for query spelling correction. First, large amounts of query-correction pairs are derived by analyzing users' query reformulation behavior encoded in the clickthrough data. Then, a phrase-based error model that accounts for the transformation probability between multi-term phrases is trained and integrated into a query speller system. Experiments are carried out on a human-labeled data set. Results show that the system using the phrase-based error model outperforms significantly its baseline systems. 1 Introduction Search queries present a particular challenge for traditional spelling correction methods for three main reasons (Ahmad and Kondrak, 2004). First, spelling errors are more common in search queries than in regular written text: roughly 10-15% of queries contain misspelled terms (Cucerzan and Brill, 2004). Second, most search queries consist of a few key words rather than grammatical sentences, making a grammar-based approach inappropriate. Most importantly, many queries contain search terms, such as proper nouns and names, which are not well established in the language. For example, Chen et al. (2007) reported that 16.5% of valid search terms do not occur in their 200K-entry spelling lexicon. Therefore, recent research has focused on the use of Web corpora and query logs, rather than human-compiled lexicons, to infer knowledge about misspellings and word usage in search queries (e.g., Whitelaw et al., 2009). Another important data source that would be useful for this purpose is clickthrough data. Although it is well-known that clickthrough data contain rich information about users' search behavior, e.g., how a user (re-) formulates a query in order to find the relevant document, there has been little research on exploiting the data for the development of a query speller system. In this paper we present a novel method of extracting large amounts of query-correction pairs from the clickthrough data. These pairs, implicitly judged by millions of users, are used to train a set of spelling error models. Among these models, the most effective one is a phrase-based error model that captures the probability of transforming one multi-term phrase into another multi-term phrase. Comparing to traditional error models that account for transformation probabilities between single characters (Kernighan et al., 1990) or sub-word strings (Brill and Moore, 2000), the phrase-based model is more powerful in that it captures some contextual information by retaining inter-term dependencies. We show that this information is crucial to detect the correction of a query term, because unlike in regular written text, any query word can be a valid search term and in many cases the only way for a speller system to make the judgment is to explore its usage according to the contextual information. We conduct a set of experiments on a large data set, consisting of human-labeled 266 query-correction pairs. Results show that the error models learned from clickthrough data lead to significant improvements on the task of query spelling correction. In particular, the speller system incorporating a phrase-based error model significantly outperforms its baseline systems. To the best of our knowledge, this is the first extensive study of learning phase-based error models from clickthrough data for query spelling correction. The rest of the paper is structured as follows. Section 2 reviews related work. Section 3 presents the way query-correction pairs are extracted from the clickthrough data. Section 4 presents the baseline speller system used in this study. Section 5 describes in detail the phrase- based error model. Section 6 presents the experiments. Section 7 concludes the paper. 2 Related Work Spelling correction for regular written text is a long standing research topic. Previous researches can be roughly grouped into two categories: correcting non-word errors and real-word errors. In non-word error spelling correction, any word that is not found in a pre-compiled lexicon is considered to be misspelled. Then, a list of lexical words that are similar to the misspelled word are proposed as candidate spelling corrections. Most traditional systems use a manually tuned similarity function (e.g., edit distance function) to rank the candidates, as reviewed by Kukich (1992). During the last two decades, statistical error models learned on training data (i.e., query-correction pairs) have become increasingly popular, and have proven more effective (Kernighan et al., 1990; Brill and Moore, 2000; Toutanova and Moore, 2002; Okazaki et al., 2008). Real-word spelling correction is also referred to as context sensitive spelling correction (CSSC). It tries to detect incorrect usages of a valid word based on its context, such as "peace" and "piece" in the context "a _ of cake". A common strategy in CSSC is as follows. First, a pre-defined confusion set is used to generate candidate corrections, then a scoring model, such as a trigram language model or naïve Bayes classifier, is used to rank the candidates according to their context (e.g., Golding and Roth, 1996; Mangu and Brill, 1997; Church et al., 2007). When designed to handle regular written text, both CSSC and non-word error speller systems rely on a pre-defined vocabulary (i.e., either a lexicon or a confusion set). However, in query spelling correction, it is impossible to compile such a vocabulary, and the boundary between the non-word and real-word errors is quite vague. Therefore, recent research on query spelling correction has focused on exploiting noisy Web data and query logs to infer knowledge about misspellings and word usage in search queries. Cucerzan and Brill (2004) discuss in detail the challenges of query spelling correction, and suggest the use of query logs. Ahmad and Kondrak (2005) propose a method of estimating an error model from query logs using the EM algorithm. Li et al. (2006) extend the error model by capturing word-level similarities learned from query logs. Chen et al. (2007) suggest using web search results to improve spelling correction. Whitelaw et al. (2009) present a query speller system in which both the error model and the language model are trained using Web data. Compared to Web corpora and query logs, clickthrough data contain much richer information about users’ search behavior. Although there has been a lot of research on using clickthrough data to improve Web document retrieval (e.g., Joachims, 2002; Agichtein et al., 2006; Gao et al., 2009), the data have not been fully explored for query spelling correction. This study tries to learn error models from clickthrough data. To our knowledge, this is the first such attempt using clickthrough data. Most of the speller systems reviewed above are based on the framework of the source channel model. Typically, a language model (source model) is used to capture contextual information, while an error model (channel model) is considered to be context free in that it does not take into account any contextual information in modeling word transformation probabilities. In this study we argue that it is beneficial to capture contextual information in the error model. To this end, inspired by the phrase-based statistical machine translation (SMT) systems (Koehn et al., 2003; Och and Ney, 2004), we propose a phrase-based error model where we assume that query spelling correction is performed at the phrase level. In what follows, before presenting the phrase- based error model, we will first describe the clickthrough data and the query speller system we used in this study. 3 Clickthrough Data and Spelling Correction This section describes the way the query-correction pairs are extracted from click267 through data. Two types of clickthrough data are explored in our experiment. The clickthrough data of the first type has been widely used in previous research and proved to be useful for Web search (Joachims, 2002; Agichtein et al., 2006; Gao et al., 2009) and query reformulation (Wang and Zhai, 2008; Suzuki et al., 2009). We start with this same data with the hope of achieving similar improvements in our task. The data consist of a set of query sessions that were extracted from one year of log files from a commercial Web search engine. A query session contains a query issued by a user and a ranked list of links (i.e., URLs) returned to that same user along with records of which URLs were clicked. Following Suzuki et al. (2009), we extract query-correction pairs as follows. First, we extract pairs of queries Q1 and Q2 such that (1) they are issued by the same user; (2) Q2 was issued within 3 minutes of Q1; and (3) Q2 contained at least one clicked URL in the result page while Q1 did not result in any clicks. We then scored each query pair (Q1, Q2) using the edit distance between Q1 and Q2, and retained those with an edit distance score lower than a pre-set threshold as query correction pairs. Unfortunately, we found in our experiments that the pairs extracted using the method are too noisy for reliable error model training, even with a very tight threshold, and we did not see any significant improvement. Therefore, in Section 6 we will not report results using this dataset. The clickthrough data of the second type consists of a set of query reformulation sessions extracted from 3 months of log files from a commercial Web browser. A query reformulation session contains a list of URLs that record user behaviors that relate to the query reformulation functions, provided by a Web search engine. For example, almost all commercial search engines offer the "did you mean" function, suggesting a possible alternate interpretation or spelling of a user-issued query. Figure 1 shows a sample of the query reformulation sessions that record the "did you mean" sessions from three of the most popular search engines. These sessions encode the same user behavior: A user first queries for "harrypotter sheme park", and then clicks on the resulting spelling suggestion "harry potter theme park". In our experiments, we "reverse-engineer" the parameters from the URLs of these sessions, and deduce how each search engine encodes both a query and the fact that a user arrived at a URL by clicking on the spelling suggestion of the query – an important indication that the spelling suggestion is desired. From these three months of query reformulation sessions, we extracted about 3 million query-correction pairs. Compared to the pairs extracted from the clickthrough data of the first type (query sessions), this data set is much cleaner because all these spelling corrections are actually clicked, and thus judged implicitly, by many users. In addition to the "did you mean" function, recently some search engines have introduced two new spelling suggestion functions. One is the "auto-correction" function, where the search engine is confident enough to automatically apply the spelling correction to the query and execute it to produce search results for the user. The other is the "split pane" result page, where one half portion of the search results are produced using the original query, while the other half, usually visually separate portion of results are produced using the auto-corrected query. In neither of these functions does the user ever receive an opportunity to approve or disapprove of the correction. Since our extraction approach focuses on user-approved spelling suggestions, Google: http://www.google.com/search? hl=en&source=hp& q=harrypotter+sheme+park&aq=f&oq=&aqi= http://www.google.com/search? hl=en&ei=rnNAS8-oKsWe_AaB2eHlCA& sa=X&oi=spell&resnum=0&ct= result&cd=1&ved=0CA4QBSgA& q=harry+potter+theme+park&spell=1 Yahoo: http://search.yahoo.com/search; _ylt=A0geu6ywckBL_XIBSDtXNyoA? p=harrypotter+sheme+park& fr2=sb-top&fr=yfp-t-701&sao=1 http://search.yahoo.com/search? ei=UTF-8&fr=yfp-t-701& p=harry+potter+theme+park &SpellState=n-2672070758_q-tsI55N6srhZa. qORA0MuawAAAA%40%40&fr2=sp-top Bing: http://www.bing.com/search? q=harrypotter+sheme+park&form=QBRE&qs=n http://www.bing.com/search? q=harry+potter+theme+park&FORM=SSRE Figure 1. A sample of query reformulation sessions from three popular search engines. These sessions show that a user first issues the query "harrypotter sheme park", and then clicks on the resulting spell suggestion "harry potter theme park". 268 we ignore the query reformulation sessions recording either of the two functions. Although by doing so we could miss some basic, obvious spelling corrections, our experiments show that the negative impact on error model training is negligible. One possible reason is that our baseline system, which does not use any error model learned from the clickthrough data, is already able to correct these basic, obvious spelling mistakes. Thus, including these data for training is unlikely to bring any further improvement. We found that the error models trained using the data directly extracted from the query reformulation sessions suffer from the problem of underestimating the self-transformation probability of a query P(Q2=Q1|Q1), because we only included in the training data the pairs where the query is different from the correction. To deal with this problem, we augmented the training data by including correctly spelled queries, i.e., the pairs (Q1, Q2) where Q1 = Q2. First, we extracted a set of queries from the sessions where no spell suggestion is presented or clicked on. Second, we removed from the set those queries that were recognized as being auto-corrected by a search engine. We do so by running a sanity check of the queries against our baseline spelling correction system, which will be described in Section 6. If the system thinks an input query is misspelled, we assumed it was an obvious misspelling, and removed it. The remaining queries were assumed to be correctly spelled and were added to the training data. 4 The Baseline Speller System The spelling correction problem is typically formulated under the framework of the source channel model. Given an input query ܳൌ ݍଵ. . . ݍூ, we want to find the best spelling correction ܥൌܿଵ. . . ܿ among all candidate spelling corrections: ܥכ ൌargmax ܲሺܥ|ܳሻ (1) Applying Bayes' Rule and dropping the constant denominator, we have ܥכ ൌargmax ܲሺܳ|ܥሻܲሺܥሻ (2) where the error model ܲሺܳ|ܥሻ models the transformation probability from C to Q, and the language model ܲሺܥሻ models how likely C is a correctly spelled query. The speller system used in our experiments is based on a ranking model (or ranker), which can be viewed as a generalization of the source channel model. The system consists of two components: (1) a candidate generator, and (2) a ranker. In candidate generation, an input query is first tokenized into a sequence of terms. Then we scan the query from left to right, and each query term q is looked up in lexicon to generate a list of spelling suggestions c whose edit distance from q is lower than a preset threshold. The lexicon we used contains around 430,000 entries; these are high frequency query terms collected from one year of search query logs. The lexicon is stored using a trie-based data structure that allows efficient search for all terms within a maximum edit distance. The set of all the generated spelling suggestions is stored using a lattice data structure, which is a compact representation of exponentially many possible candidate spelling corrections. We then use a decoder to identify the top twenty candidates from the lattice according to the source channel model of Equation (2). The language model (the second factor) is a backoff bigram model trained on the tokenized form of one year of query logs, using maximum likelihood estimation with absolute discounting smoothing. The error model (the first factor) is approximated by the edit distance function as െlogܲሺܳ|ܥሻן EditDistሺܳ, ܥሻ (3) The decoder uses a standard two-pass algorithm to generate 20-best candidates. The first pass uses the Viterbi algorithm to find the best C according to the model of Equations (2) and (3). In the second pass, the A-Star algorithm is used to find the 20-best corrections, using the Viterbi scores computed at each state in the first pass as heuristics. Notice that we always include the input query Q in the 20-best candidate list. The core of the second component of the speller system is a ranker, which re-ranks the 20-best candidate spelling corrections. If the top C after re-ranking is different than the original query Q, the system returns C as the correction. Let f be a feature vector extracted from a query and candidate spelling correction pair (Q, C). The ranker maps f to a real value y that indicates how likely C is a desired correction of Q. For example, a linear ranker simply maps f to y with a learned weight vector w such as ݕൌܟ· , where w is optimized w.r.t. accuracy on a set of hu269 man-labeled (Q, C) pairs. The features in f are arbitrary functions that map (Q, C) to a real value. Since we define the logarithm of the probabilities of the language model and the error model (i.e., the edit distance function) as features, the ranker can be viewed as a more general framework, subsuming the source channel model as a special case. In our experiments we used 96 features and a non-linear model, implemented as a two-layer neural net, though the details of the ranker and the features are beyond the scope of this paper. 5 A Phrase-Based Error Model The goal of the phrase-based error model is to transform a correctly spelled query C into a misspelled query Q. Rather than replacing single words in isolation, this model replaces sequences of words with sequences of words, thus incorporating contextual information. For instance, we might learn that “theme part” can be replaced by “theme park” with relatively high probability, even though “part” is not a misspelled word. We assume the following generative story: first the correctly spelled query C is broken into K non-empty word sequences c1, …, ck, then each is replaced with a new non-empty word sequence q1, …, qk, and finally these phrases are permuted and concatenated to form the misspelled Q. Here, c and q denote consecutive sequences of words. To formalize this generative process, let S denote the segmentation of C into K phrases c1…cK, and let T denote the K replacement phrases q1…qK – we refer to these (ci, qi) pairs as bi-phrases. Finally, let M denote a permutation of K elements representing the final reordering step. Figure 2 demonstrates the generative procedure. Next let us place a probability distribution over rewrite pairs. Let B(C, Q) denote the set of S, T, M triples that transform C into Q. If we assume a uniform probability over segmentations, then the phrase-based probability can be defined as: ܲሺܳ|ܥሻן ܲሺܶ|ܥ, ܵሻڄ ܲሺܯ|ܥ, ܵ, ܶሻ ሺௌ,்,ெሻא ሺ,ொሻ (4) As is common practice in SMT, we use the maximum approximation to the sum: ܲሺܳ|ܥሻൎ max ሺௌ,்,ெሻא ሺ,ொሻ ܲሺܶ|ܥ, ܵሻڄ ܲሺܯ|ܥ, ܵ, ܶሻ (5) 5.1 Forced Alignments Although we have defined a generative model for transforming queries, our goal is not to propose new queries, but rather to provide scores over existing Q and C pairs which act as features for the ranker. Furthermore, the word-level alignments between Q and C can most often be identified with little ambiguity. Thus we restrict our attention to those phrase transformations consistent with a good word-level alignment. Let J be the length of Q, L be the length of C, and A = a1, …, aJ be a hidden variable representing the word alignment. Each ai takes on a value ranging from 1 to L indicating its corresponding word position in C, or 0 if the ith word in Q is unaligned. The cost of assigning k to ai is equal to the Levenshtein edit distance (Levenshtein, 1966) between the ith word in Q and the kth word in C, and the cost of assigning 0 to ai is equal to the length of the ith word in Q. We can determine the least cost alignment A* between Q and C efficiently using the A-star algorithm. When scoring a given candidate pair, we further restrict our attention to those S, T, M triples that are consistent with the word alignment, which we denote as B(C, Q, A*). Here, consistency requires that if two words are aligned in A*, then they must appear in the same bi-phrase (ci, qi). Once the word alignment is fixed, the final permutation is uniquely determined, so we can safely discard that factor. Thus we have: ܲሺܳ|ܥሻൎ max ሺௌ,்,ெሻא ࣜሺ,ொ,څሻ ܲሺܶ|ܥ, ܵሻ (6) For the sole remaining factor P(T|C, S), we make the assumption that a segmented query T = q1… qK is generated from left to right by transforming each phrase c1…cK independently: C: “disney theme park” correct query S: [“disney”, “theme park”] segmentation T: [“disnee”, “theme part”] translation M: (1 Æ 2, 2Æ 1) permutation Q: “theme part disnee” misspelled query Figure 2: Example demonstrating the generative procedure behind the phrase-based error model. 270 ܲሺܶ|ܥ, ܵሻൌ∏ ܲሺܙ|܋ሻ ୀଵ , (7) where ܲሺܙ|܋ሻ is a phrase transformation probability, the estimation of which will be described in Section 5.2. To find the maximum probability assignment efficiently, we can use a dynamic programming approach, somewhat similar to the monotone decoding algorithm described in Och (2002). Here, though, both the input and the output word sequences are specified as the input to the algorithm, as is the word alignment. We define the quantity ߙ to be the probability of the most likely sequence of bi-phrases that produce the first j terms of Q and are consistent with the word alignment and C. It can be calculated using the following algorithm: 1. Initialization: ߙൌ1 (8) 2. Induction: ߙൌ max ′ழ,ܙୀ ೕ′శభ …ೕ ቄߙ′ܲ൫ܙห܋ܙ൯ቅ (9) 3. Total: ܲሺܳ|ܥሻൌߙ (10) The pseudo-code of the above algorithm is shown in Figure 3. After generating Q from left to right according to Equations (8) to (10), we record at each possible bi-phrase boundary its maximum probability, and we obtain the total probability at the end-position of Q. Then, by back-tracking the most probable bi-phrase boundaries, we obtain B*. The algorithm takes a complexity of O(KL2), where K is the total number of word alignments in A* which does not contain empty words, and L is the maximum length of a bi-phrase, which is a hyper-parameter of the algorithm. Notice that when we set L=1, the phrase-based error model is reduced to a word-based error model which assumes that words are transformed independently from C to Q, without taking into account any contextual information. 5.2 Model Estimation We follow a method commonly used in SMT (Koehn et al., 2003) to extract bi-phrases and estimate their replacement probabilities. From each query-correction pair with its word alignment (Q, C, A*), all bi-phrases consistent with the word alignment are identified. Consistency here implies two things. First, there must be at least one aligned word pair in the bi-phrase. Second, there must not be any word alignments from words inside the bi-phrase to words outside the bi-phrase. That is, we do not extract a phrase pair if there is an alignment from within the phrase pair to outside the phrase pair. The toy example shown in Figure 4 illustrates the bilingual phrases we can generate by this process. After gathering all such bi-phrases from the full training data, we can estimate conditional relative frequency estimates without smoothing. For example, the phrase transformation probability ܲሺܙ|܋ሻ in Equation (7) can be estimated approximately as Input: biPhraseLattice “PL” with length = K & height = L; Initialization: biPhrase.maxProb = 0; for (x = 0; x <= K – 1; x++) for (y = 1; y <= L; y++) for (yPre = 1; yPre <= L; yPre++) { xPre = x – y; biPhrasePre = PL.get(xPre, yPre); biPhrase = PL.get(x, y); if (!biPhrasePre || !biPhrase) continue; probIncrs = PL.getProbIncrease(biPhrasePre, biPhrase); maxProbPre = biPhrasePre.maxProb; totalProb = probIncrs + maxProbPre; if (totalProb > biPhrase.maxProb) { biPhrase.maxProb = totalProb; biPhrase.yPre = yPre; } } Result: record at each bi-phrase boundary its maximum probability (biPhrase.maxProb) and optimal back-tracking biPhrases (biPhrase.yPre). Figure 3: The dynamic programming algorithm for Viterbi bi-phrase segmentation. A B C D E F a A a # adc ABCD d # d D c # dc CD f # dcf CDEF c C f F Figure 4: Toy example of (left) a word alignment between two strings "adcf" and "ABCDEF"; and (right) the bi-phrases containing up to four words that are consistent with the word alignment. 271 ܲሺܙ|܋ሻൌ ܰሺ܋, ܙሻ ∑ ܰሺ܋, ܙᇱሻ ܙᇲ (11) where ܰሺ܋, ܙሻ is the number of times that c is aligned to q in training data. These estimates are useful for contextual lexical selection with sufficient training data, but can be subject to data sparsity issues. An alternate translation probability estimate not subject to data sparsity issues is the so-called lexical weight estimate (Koehn et al., 2003). Assume we have a word translation distribution ݐሺݍ|ܿሻ (defined over individual words, not phrases), and a word alignment A between q and c; here, the word alignment contains (i, j) pairs, where ݅א 1. . || and ݆א 0. . |܋|, with 0 indicating an inserted word. Then we can use the following estimate: ܲ௪ሺܙ|܋, ܣሻൌෑ 1 |ሼ݆|ሺ݆, ݅ሻא ܣሽ| ݐሺݍ|ܿሻ ሺ,ሻא |ܙ| ୀଵ (12) We assume that for every position in q, there is either a single alignment to 0, or multiple alignments to non-zero positions in c. In effect, this computes a product of per-word translation scores; the per-word scores are averages of all the translations for the alignment links of that word. We estimate the word translation probabilities using counts from the word aligned corpus: ݐሺݍ|ܿሻൌ ேሺ,ሻ ∑ ே൫,′൯ ′ . Here ܰሺܿ, ݍሻ is the number of times that the words (not phrases as in Equation 11) c and q are aligned in the training data. These word based scores of bi-phrases, though not as effective in contextual selection, are more robust to noise and sparsity. Throughout this section, we have approached this model in a noisy channel approach, finding probabilities of the misspelled query given the corrected query. However, the method can be run in both directions, and in practice SMT systems benefit from also including the direct probability of the corrected query given this misspelled query (Och, 2002). 5.3 Phrase-Based Error Model Features To use the phrase-based error model for spelling correction, we derive five features and integrate them into the ranker-based query speller system, described in Section 4. These features are as follows. • Two phrase transformation features: These are the phrase transformation scores based on relative frequency estimates in two directions. In the correction-to-query direction, we define the feature as ்݂ሺܳ, ܥ, ܣሻൌ log ܲሺܳ|ܥሻ, where ܲሺܳ|ܥሻ is computed by Equations (8) to (10), and ܲ൫ܙห܋ܙ൯ is the relative frequency estimate of Equation (11). • Two lexical weight features: These are the phrase transformation scores based on the lexical weighting models in two directions. For example, in the correction-to-query direction, we define the feature as݂ௐሺܳ, ܥ, ܣሻൌlog ܲሺܳ|ܥሻ, where ܲሺܳ|ܥሻ is computed by Equations (8) to (10), and the phrase transformation probability is the computed as lexical weight according to Equation (12). • Unaligned word penalty feature: the feature is defined as the ratio between the number of unaligned query words and the total number of query words. 6 Experiments We evaluate the spelling error models on a large scale real world data set containing 24,172 queries sampled from one year’s worth of query logs from a commercial search engine. The spelling of each query is judged and corrected by four annotators. We divided the data set into training and test data sets. The two data sets do not overlap. The training data contains 8,515 query-correction pairs, among which 1,743 queries are misspelled (i.e., in these pairs, the corrections are different from the queries). The test data contains 15,657 query-correction pairs, among which 2,960 queries are misspelled. The average length of queries in the training and test data is 2.7 words. The speller systems we developed in this study are evaluated using the following three metrics. • Accuracy: The number of correct outputs generated by the system divided by the total number of queries in the test set. • Precision: The number of correct spelling corrections for misspelled queries generated by the system divided by the total number of corrections generated by the system. • Recall: The number of correct spelling corrections for misspelled queries generated by the system divided by the total number of misspelled queries in the test set. We also perform a significance test, i.e., a t-test with a significance level of 0.05. A significant difference should be read as significant at the 95% level. 272 In our experiments, all the speller systems are ranker-based. In most cases, other than the baseline system (a linear neural net), the ranker is a two-layer neural net with 5 hidden nodes. The free parameters of the neural net are trained to optimize accuracy on the training data using the back propagation algorithm, running for 500 iterations with a very small learning rate (0.1) to avoid overfitting. We did not adjust the neural net structure (e.g., the number of hidden nodes) or any training parameters for different speller systems. Neither did we try to seek the best tradeoff between precision and recall. Since all the systems are optimized for accuracy, we use accuracy as the primary metric for comparison. Table 1 summarizes the main spelling correction results. Row 1 is the baseline speller system where the source-channel model of Equations (2) and (3) is used. In our implementation, we use a linear ranker with only two features, derived respectively from the language model and the error model models. The error model is based on the edit distance function. Row 2 is the ranker-based spelling system that uses all 96 ranking features, as described in Section 4. Note that the system uses the features derived from two error models. One is the edit distance model used for candidate generation. The other is a phonetic model that measures the edit distance between the metaphones (Philips, 1990) of a query word and its aligned correction word. Row 3 is the same system as Row 2, with an additional set of features derived from a word-based error model. This model is a special case of the phrase-based error model described in Section 5 with the maximum phrase length set to one. Row 4 is the system that uses the additional 5 features derived from the phrase-based error models with a maximum bi-phrase length of 3. In phrase based error model, L is the maximum length of a bi-phrase (Figure 3). This value is important for the spelling performance. We perform experiments to study the impact of L; the results are displayed in Table 2. Moreover, since we proposed to use clickthrough data for spelling correction, it is interesting to study the impact on spelling performance from the size of clickthrough data used for training. We varied the size of clickthrough data and the experimental results are presented in Table 3. The results show first and foremost that the ranker-based system significantly outperforms the spelling system based solely on the source-channel model, largely due to the richer set of features used (Row 1 vs. Row 2). Second, the error model learned from clickthrough data leads to significant improvements (Rows 3 and 4 vs. Row 2). The phrase-based error model, due to its capability of capturing contextual information, outperforms the word-based model with a small but statistically significant margin (Row 4 vs. Row 3), though using phrases longer (L > 3) does not lead to further significant improvement (Rows 6 and 7 vs. Rows 8 and 9). Finally, using more clickthrough data leads to significant improvement (Row 13 vs. Rows 10 to 12). The benefit does not appear to have peaked – further improvements are likely given a larger data set. 7 Conclusions Unlike conventional textual documents, most search queries consist of a sequence of key words, many of which are valid search terms but are not stored in any compiled lexicon. This presents a challenge to any speller system that is based on a dictionary. This paper extends the recent research on using Web data and query logs for query spelling correction in two aspects. First, we show that a large amount of training data (i.e. query-correction pairs) can be extracted from clickthrough data, focusing on query reformulation sessions. The resulting data are very clean and effective for error model training. Second, we argue that it is critical to capture contextual information for query spelling correction. To this end, we propose # System Accuracy Precision Recall 1 Source-channel 0.8526 0.7213 0.3586 2 Ranker-based 0.8904 0.7414 0.4964 3 Word model 0.8994 0.7709 0.5413 4 Phrase model (L=3) 0.9043 0.7814 0.5732 Table 1. Summary of spelling correction results. # System Accuracy Precision Recall 5 Phrase model (L=1) 0.8994 0.7709 0.5413 6 Phrase model (L=2) 0.9014 0.7795 0.5605 7 Phrase model (L=3) 0.9043 0.7814 0.5732 8 Phrase model (L=5) 0.9035 0.7834 0.5698 9 Phrase model (L=8) 0.9033 0.7821 0.5713 Table 2. Variations of spelling performance as a function of phrase length. # System Accuracy Precision Recall 10 L=3; 0 month data 0.8904 0.7414 0.4964 11 L=3; 0.5 month data 0.8959 0.7701 0.5234 12 L=3; 1.5 month data 0.9023 0.7787 0.5667 13 L=3; 3 month data 0.9043 0.7814 0.5732 Table 3. Variations of spelling performance as a function of the size of clickthrough data used for training. 273 a new phrase-based error model, which leads to significant improvement in our spelling correction experiments. There is additional potentially useful information that can be exploited in this type of model. For example, in future work we plan to investigate the combination of the clickthrough data collected from a Web browser with the noisy but large query sessions collected from a commercial search engine. Acknowledgments The authors would like to thank Andreas Bode, Mei Li, Chenyu Yan and Galen Andrew for the very helpful discussions and collaboration. References Agichtein, E., Brill, E. and Dumais, S. 2006. Improving web search ranking by incorporating user behavior information. In SIGIR, pp. 19-26. Ahmad, F., and Kondrak, G. 2005. Learning a spelling error model from search query logs. In HLT-EMNLP, pp 955-962. Brill, E., and Moore, R. C. 2000. An improved error model for noisy channel spelling correction. In ACL, pp. 286-293. Chen, Q., Li, M., and Zhou, M. 2007. Improving query spelling correction using web search results. In EMNLP-CoNLL, pp. 181-189. Church, K., Hard, T., and Gao, J. 2007. Compressing trigram language models with Golomb coding. In EMNLP-CoNLL, pp. 199-207. Cucerzan, S., and Brill, E. 2004. Spelling correction as an iterative process that exploits the collective knowledge of web users. In EMNLP, pp. 293-300. Gao, J., Yuan, W., Li, X., Deng, K., and Nie, J-Y. 2009. Smoothing clickthrough data for web search ranking. In SIGIR. Golding, A. R., and Roth, D. 1996. Applying winnow to context-sensitive spelling correction. In ICML, pp. 182-190. Joachims, T. 2002. Optimizing search engines using clickthrough data. In SIGKDD, pp. 133-142. Kernighan, M. D., Church, K. W., and Gale, W. A. 1990. A spelling correction program based on a noisy channel model. In COLING, pp. 205-210. Koehn, P., Och, F., and Marcu, D. 2003. Statistical phrase-based translation. In HLT/NAACL, pp. 127-133. Kukich, K. 1992. Techniques for automatically correcting words in text. ACM Computing Surveys. 24(4): 377-439. Levenshtein, V. I. 1966. Binary codes capable of correcting deletions, insertions and reversals. Soviet Physics Doklady, 10(8):707-710. Li, M., Zhu, M., Zhang, Y., and Zhou, M. 2006. Exploring distributional similarity based models for query spelling correction. In ACL, pp. 1025-1032. Mangu, L., and Brill, E. 1997. Automatic rule acquisition for spelling correction. In ICML, pp. 187-194. Och, F. 2002. Statistical machine translation: from single-word models to alignment templates. PhD thesis, RWTH Aachen. Och, F., and Ney, H. 2004. The alignment template approach to statistical machine translation. Computational Linguistics, 30(4): 417-449. Okazaki, N., Tsuruoka, Y., Ananiadou, S., and Tsujii, J. 2008. A discriminative candidate generator for string transformations. In EMNLP, pp. 447-456. Philips, L. 1990. Hanging on the metaphone. Computer Language Magazine, 7(12):38-44. Suzuki, H., Li, X., and Gao, J. 2009. Discovery of term variation in Japanese web search queries. In EMNLP. Toutanova, K., and Moore, R. 2002. Pronunciation modeling for improved spelling correction. In ACL, pp. 144-151. Wang, X., and Zhai, C. 2008. Mining term association patterns from search logs for effective query reformulation. In CIKM, pp. 479-488. Whitelaw, C., Hutchinson, B., Chung, G. Y., and Ellis, G. 2009. Using the web for language independent spellchecking and autocorrection. In EMNLP, pp. 890-899. 274
|
2010
|
28
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 275–285, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Inducing Domain-specific Semantic Class Taggers from (Almost) Nothing Ruihong Huang and Ellen Riloff School of Computing University of Utah Salt Lake City, UT 84112 {huangrh,riloff}@cs.utah.edu Abstract This research explores the idea of inducing domain-specific semantic class taggers using only a domain-specific text collection and seed words. The learning process begins by inducing a classifier that only has access to contextual features, forcing it to generalize beyond the seeds. The contextual classifier then labels new instances, to expand and diversify the training set. Next, a cross-category bootstrapping process simultaneously trains a suite of classifiers for multiple semantic classes. The positive instances for one class are used as negative instances for the others in an iterative bootstrapping cycle. We also explore a one-semantic-class-per-discourse heuristic, and use the classifiers to dynamically create semantic features. We evaluate our approach by inducing six semantic taggers from a collection of veterinary medicine message board posts. 1 Introduction The goal of our research is to create semantic class taggers that can assign a semantic class label to every noun phrase in a sentence. For example, consider the sentence: “The lab mix was diagnosed with parvo and given abx”. A semantic tagger should identify the “the lab mix” as an ANIMAL, “parvo” as a DISEASE, and “abx” (antibiotics) as a DRUG. Accurate semantic tagging could be beneficial for many NLP tasks, including coreference resolution and word sense disambiguation, and many NLP applications, such as event extraction systems and question answering technology. Semantic class tagging has been the subject of previous research, primarily under the guises of named entity recognition (NER) and mention detection. Named entity recognizers perform semantic tagging on proper name noun phrases, and sometimes temporal and numeric expressions as well. The mention detection task was introduced in recent ACE evaluations (e.g., (ACE, 2007; ACE, 2008)) and requires systems to identify all noun phrases (proper names, nominals, and pronouns) that correspond to 5-7 semantic classes. Despite widespread interest in semantic tagging, nearly all semantic taggers for comprehensive NP tagging still rely on supervised learning, which requires annotated data for training. A few annotated corpora exist, but they are relatively small and most were developed for broadcoverage NLP. Many domains, however, are replete with specialized terminology and jargon that cannot be adequately handled by general-purpose systems. Domains such as biology, medicine, and law are teeming with specialized vocabulary that necessitates training on domain-specific corpora. Our research explores the idea of inducing domain-specific semantic taggers using a small set of seed words as the only form of human supervision. Given an (unannotated) collection of domain-specific text, we automatically generate training instances by labelling every instance of a seed word with its designated semantic class. We then train a classifier to do semantic tagging using these seed-based annotations, using bootstrapping to iteratively improve performance. On the surface, this approach appears to be a contradiction. The classifier must learn how to assign different semantic tags to different instances of the same word based on context (e.g., “lab” may refer to an animal in one context but a laboratory in another). And yet, we plan to train the classifier using stand-alone seed words, making the assumption that every instance of the seed belongs to the same semantic class. We resolve this apparent contradiction by using semantically unambiguous seeds and by introducing an initial context-only training phase before bootstrapping begins. First, we train a strictly contextual classifier that only 275 has access to contextual features and cannot see the seed. Then we apply the classifier to the corpus to automatically label new instances, and combine these new instances with the seed-based instances. This process expands and diversifies the training set to fuel subsequent bootstrapping. Another challenge is that we want to use a small set of seeds to minimize the amount of human effort, and then use bootstrapping to fully exploit the domain-specific corpus. Iterative self-training, however, often has difficulty sustaining momentum or it succumbs to semantic drift (Komachi et al., 2008; McIntosh and Curran, 2009). To address these issues, we simultaneously induce a suite of classifiers for multiple semantic categories, using the positive instances of one semantic category as negative instances for the others. As bootstrapping progresses, the classifiers gradually improve themselves, and each other, over many iterations. We also explore a onesemantic-class-per-discourse (OSCPD) heuristic that infuses the learning process with fresh training instances, which may be substantially different from the ones seen previously, and we use the labels produced by the classifiers to dynamically create semantic features. We evaluate our approach by creating six semantic taggers using a collection of message board posts in the domain of veterinary medicine. Our results show this approach produces high-quality semantic taggers after a sustained bootstrapping cycle that maintains good precision while steadily increasing recall over many iterations. 2 Related Work Semantic class tagging is most closely related to named entity recognition (NER), mention detection, and semantic lexicon induction. NER systems (e.g., (Bikel et al., 1997; Collins and Singer, 1999; Cucerzan and Yarowsky, 1999; Fleischman and Hovy, 2002) identify proper named entities, such as people, organizations, and locations. Several bootstrapping methods for NER have been previously developed (e.g., (Collins and Singer, 1999; Niu et al., 2003)). NER systems, however, do not identify nominal NP instances (e.g., “a software manufacturer” or “the beach”), or handle semantic classes that are not associated with proper named entities (e.g., symptoms).1 ACE 1Some NER systems also handle specialized constructs such as dates and monetary amounts. mention detection systems (e.g., see (ACE, 2005; ACE, 2007; ACE, 2008)) require tagging of NPs that correspond to 5-7 general semantic classes. These systems are typically trained with supervised learning using annotated corpora, although techniques have been developed to use resources for one language to train systems for different languages (e.g., (Zitouni and Florian, 2009)). Another line of relevant work is semantic class induction (e.g., (Riloff and Shepherd, 1997; Roark and Charniak, 1998; Thelen and Riloff, 2002; Ng, 2007; McIntosh and Curran, 2009), where the goal is to induce a stand-alone dictionary of words with semantic class labels. These techniques are often designed to learn specialized terminology from unannotated domain-specific texts via bootstrapping. Our work, however, focuses on classification of NP instances in context, so the same phrase may be assigned to different semantic classes in different contexts. Consequently, our classifier can also assign semantic class labels to pronouns. There has also been work on extracting semantically related terms or category members from the Web (e.g., (Pas¸ca, 2004; Etzioni et al., 2005; Kozareva et al., 2008; Carlson et al., 2009)). These techniques harvest broad-coverage semantic information from the Web using patterns and statistics, typically for the purpose of knowledge acquisition. Importantly, our goal is to classify instances in context, rather than generate lists of terms. In addition, the goal of our research is to learn specialized terms and jargon that may not be common on the Web, as well as domain-specific usages that may differ from the norm (e.g., “mix” and “lab” are usually ANIMALS in our domain). The idea of simulataneously learning multiple semantic categories to prevent semantic drift has been explored for other tasks, such as semantic lexicon induction (Thelen and Riloff, 2002; McIntosh and Curran, 2009) and pattern learning (Yangarber, 2003). Our bootstrapping model can be viewed as a form of self-training (e.g., (Ng and Cardie, 2003; Mihalcea, 2004; McClosky et al., 2006)), and cross-category training is similar in spirit to co-training (e.g., (Blum and Mitchell, 1998; Collins and Singer, 1999; Riloff and Jones, 1999; Mueller et al., 2002; Phillips and Riloff, 2002)). But, importantly, our classifiers all use the same feature set so they do not represent independent views of the data. They do, however, offer slightly different perspectives because each is at276 tempting to recognize a different semantic class. 3 Bootstrapping an Instance-based Semantic Class Tagger from Seeds 3.1 Motivation Our goal is to create a bootstrapping model that can rapidly create semantic class taggers using just a small set of seed words and an unannotated domain-specific corpus. Our motivation comes from specialized domains that cannot be adequately handled by general-purpose NLP systems. As an example of such a domain, we have been working with a collection of message board posts in the field of veterinary medicine. Given a document, we want a semantic class tagger to label every NP with a semantic category, for example: [A 14yo doxy]ANIMAL owned by [a reputable breeder]HUMAN is being treated for [IBD]DISEASE with [pred]DRUG. When we began working with these texts, we were immediately confronted by a dizzying array of non-standard words and word uses. In addition to formal veterinary vocabulary (e.g., animal diseases), veterinarians often use informal, shorthand terms when posting on-line. For example, they frequently refer to breeds using “nicknames” or shortened terms (e.g., gshep for German shepherd, doxy for dachsund, bxr for boxer, labx for labrador cross). Often, they refer to animals based solely on their physical characteristics, for example “a dlh” (domestic long hair), “a m/n” (male, neutered), or “a 2yo” (2 year old). This phenomenon occurs with other semantic categories as well, such as drugs and medical tests (e.g., pred for prednisone, and rads for radiographs). Nearly all semantic class taggers are trained using supervised learning with manually annotated data. However, annotated data is rarely available for specialized domains, and it is expensive to obtain because domain experts must do the annotation work. So we set out to create a bootstrapping model that can rapidly create domain-specific semantic taggers using just a few seed words and a domain-specific text collection. Our bootstrapping model consists of two distinct phases. First, we train strictly contextual classifiers from the seed annotations. We then apply the classifiers to the unlabeled data to generate new annotated instances that are added to the training set. Second, we employ a cross-category bootstrapping process that simultaneously trains a suite of classifiers for multiple semantic categories, using the positive instances for one semantic class as negative instances for the others. This cross-category training process gives the learner sustained momentum over many bootstrapping iterations. Finally, we explore two additional enhancements: (1) a one-semantic-classper-discourse heuristic to automatically generate new training instances, and (2) dynamically created semantic features produced by the classifiers themselves. In the following sections, we explain each of these steps in detail. 3.2 Phase 1: Inducing a Contextual Classifier The main challenge that we faced was how to train an instance-based classifier using seed words as the only form of human supervision. First, the user must provide a small set of seed words that are relatively unambiguous (e.g., “dog” will nearly always refer to an animal in our domain). But even so, training a traditional classifier from seedbased instances would likely produce a classifier that learns to recognize the seeds but is unable to classify new examples. We needed to force the classifier to generalize beyond the seed words. Our solution was to introduce an initial training step that induces a strictly contextual classifier. First, we generate training instances by automatically labeling each instance of a seed word with its designated semantic class. However, when we create feature vectors for the classifier, the seeds themselves are hidden and only contextual features are used to represent each training instance. By essentially “masking” the seed words so the classifier can only see the contexts around them, we force the classifier to generalize. We create a suite of strictly contextual classifiers, one for each semantic category. Each classifier makes a binary decision as to whether a noun phrase belongs to its semantic category. We use the seed words for category Ck to generate positive training instances for the Ck classifier, and the seed words for all other categories to generate the negative training instances for Ck. We use an in-house sentence segmenter and NP chunker to identify the base NPs in each sentence and create feature vectors that represent each constituent in the sentence as either an NP or an individual word. For each seed word, the feature 277 vector captures a context window of 3 constituents (word or NP) to its left and 3 constituents (word or NP) to its right. Each constituent is represented with a lexical feature: for NPs, we use its head noun; for individual words, we use the word itself. The seed word, however, is discarded so that the classifier is essentially blind-folded and cannot see the seed that produced the training instance. We also create a feature for every modifier that precedes the head noun in the target NP, except for articles which are discarded. As an example, consider the following sentence: Fluffy was diagnosed with FELV after a blood test showed that he tested positive. Suppose that “FELV” is a seed for the DISEASE category and “test” is a seed for the TEST category. Two training instances would be created, with feature vectors that look like this, where M represents a modifier inside the target NP: was−3 diagnosed−2 with−1 after1 test2 showed3 ⇒DISEASE with−3 FELV−2 after−1 bloodM showed1 that2 he3 ⇒TEST The contextual classifiers are then applied to the corpus to automatically label new instances. We use a confidence score to label only the instances that the classifiers are most certain about. We compute a confidence score for instance i with respect to semantic class Ck by considering both the score of the Ck classifier as well as the scores of the competing classifiers. Intuitively, we have confidence in labeling an instance as category Ck if the Ck classifier gave it a positive score, and its score is much higher than the score of any other classifier. We use the following scoring function: Confidence(i,Ck) = score(i,Ck) - max(∀j̸=k score(i,Cj)) We employ support vector machines (SVMs) (Joachims, 1999) with a linear kernel as our classifiers, using the SVMlin software (Keerthi and DeCoste, 2005). We use the value produced by the decision function (essentially the distance from the hyperplane) as the score for a classifier. We specify a threshold θcf and only assign a semantic tag Ck to an instance i if Confidence(i,Ck) ≥θcf. All instances that pass the confidence threshold are labeled and added to the training set. This process greatly enhances the diversity of the training data. In this initial learning step, the strictly contextual classifiers substantially increase the number of training instances for each semantic category, producing a more diverse mix of seed-generated instances and context-generated instances. 3.3 Phase 2: Cross-Category Bootstrapping The next phase of the learning process is an iterative bootstrapping procedure. The key challenge was to design a bootstrapping model that would not succumb to semantic drift and would have sustained momentum to continue learning over many iterations. Figure 1 shows the design of our cross-category bootstrapping model.2 We simultaneously train a suite of binary classifiers, one for each semantic category, C1 . . . Cn. After each training cycle, all of the classifiers are applied to the remaining unlabeled instances and each classifier labels the (positive) instances that it is most confident about (i.e., the instances that it classifies with a confidence score ≥θcf). The set of instances positively labeled by classifier Ck are shown as C+ k in Figure 1. All of the new instances produced by classifier Ck are then added to the set of positive training instances for Ck and to the set of negative training instances for all of the other classifiers. One potential problem with this scheme is that some categories are more prolific than others, plus we are collecting negative instances from a set of competing classifiers. Consequently, this approach can produce highly imbalanced training sets. Therefore we enforced a 3:1 ratio of negatives to positives by randomly selecting a subset of the possible negative instances. We discuss this issue further in Section 4.4. labeled C+ 1 C2 C+ 2 C+ n Cn unlabeled seeds C1 i=1 C+ _ ( ) C+ 1 (+) (+) C+ 2 i=2 C+ _ ( ) _ ( ) i=n C+ C+ n (+) Figure 1: Cross-Category Bootstrapping 2For simplicity, this picture does not depict the initial contextual training step, but that can be viewed as the first iteration in this general framework. 278 Cross-category training has two advantages over independent self-training. First, as others have shown for pattern learning and lexicon induction (Thelen and Riloff, 2002; Yangarber, 2003; McIntosh and Curran, 2009), simultaneously training classifiers for multiple categories reduces semantic drift because each classifier is deterred from encroaching on another one’s territory (i.e., claiming the instances from a competing class as its own). Second, similar in spirit to co-training3, this approach allows each classifier to obtain new training instances from an outside source that has a slightly different perspective. While independent self-training can quickly run out of steam, cross-category training supplies each classifier with a constant stream of new (negative) instances produced by competing classifiers. In Section 4, we will show that cross-category bootstrapping performs substantially better than an independent self-training model, where each classifier is bootstrapped separately. The feature set for these classifiers is exactly the same as described in Section 3.2, except that we add a new lexical feature that represents the head noun of the target NP (i.e., the NP that needs to be tagged). This allows the classifiers to consider the local context as well as the target word itself when making decisions. 3.4 One Semantic Class Per Discourse We also explored the idea of using a one semantic class per discourse (OSCPD) heuristic to generate additional training instances during bootstrapping. Inspired by Yarowsky’s one sense per discourse heuristic for word sense disambiguation (Yarowsky, 1995), we make the assumption that multiple instances of a word in the same discourse will nearly always correspond to the same semantic class. Since our data set consists of message board posts organized as threads, we consider all posts in the same thread to be a single discourse. After each training step, we apply the classifiers to the unlabeled data to label some new instances. For each newly labeled instance, the OSCPD heuristic collects all instances with the same head noun in the same discourse (thread) and unilaterally labels them with the same semantic class. This heuristic serves as meta-knowledge to label instances that (potentially) occur in very different 3But technically this is not co-training because our feature sets are all the same. contexts, thereby infusing the bootstrapping process with “fresh” training examples. In early experiments, we found that OSCPD can be aggressive, pulling in many new instances. If the classifier labels a word incorrectly, however, then the OSCPD heuristic will compound the error and mislabel even more instances incorrectly. Therefore we only apply this heuristic to instances that are labeled with extremely high confidence (θcf ≥2.5) and that pass a global sanity check, gsc(w) ≥0.2, which ensures that a relatively high proportion of labeled instances with the same head noun have been assigned to the same semantic class. Specifically, gsc(w) = 0.1∗ wl/c wl +0.9∗ wu/c wu where wl and wu are the # of labeled and unlabeled instances, respectively, wl/c is the # of instances labeled as c, and wu/c is the # of unlabeled instances that receive a positive confidence score for c when given to the classifier. The intuition behind the second term is that most instances are initially unlabeled and we want to make sure that many of the unlabeled instances are likely to belong to the same semantic class (even though the classifier isn’t ready to commit to them yet). 3.5 Dynamic Semantic Features For many NLP tasks, classifiers use semantic features to represent the semantic class of words. These features are typically obtained from external resources such as Wordnet (Miller, 1990). Our bootstrapping model incrementally trains semantic class taggers, so we explored the idea of using the labels assigned by the classifiers to create enhanced feature vectors by dynamically adding semantic features. This process allows later stages of bootstrapping to directly benefit from earlier stages. For example, consider the sentence: He started the doxy on Vetsulin today. If “Vetsulin” was labeled as a DRUG in a previous bootstrapping iteration, then the feature vector representing the context around “doxy” can be enhanced to include an additional semantic feature identifying Vetsulin as a DRUG, which would look like this: He−2 started−1 on1 V etsulin2 DRUG2 today3 Intuitively, the semantic features should help the classifier identify more general contextual patterns, such as “started <X> on DRUG”. To create semantic features, we use the semantic tags that 279 have been assigned to the current set of labeled instances. When a feature vector is created for a target NP, we check every noun instance in its context window to see if it has been assigned a semantic tag, and if so, then we add a semantic feature. In the early stages of bootstrapping, however, relatively few nouns will be assigned semantic tags, so these features are often missing. 3.6 Thresholds and Stopping Criterion When new instances are automatically labeled during bootstrapping, it is critically important that most of the labels are correct or performance rapidly deteriorates. This suggests that we should only label instances in which the classifier has high confidence. On the other hand, a high threshold often yields few new instances, which can cause the bootstrapping process to sputter and halt. To balance these competing demands, we used a sliding threshold that begins conservatively but gradually loosens the reins as bootstrapping progresses. Initially, we set θcf = 2.0, which only labels instances that the classifier is highly confident about. When fewer than min new instances can be labeled, we automatically decrease θcf by 0.2, allowing another batch of new instances to be labeled, albeit with slightly less confidence. We continue decreasing the threshold, as needed, until θcf < 1.0, when we end the bootstrapping process. In Section 4, we show that this sliding threshold outperforms fixed threshold values. 4 Evaluation 4.1 Data Our data set consists of message board posts from the Veterinary Information Network (VIN), which is a web site (www.vin.com) for professionals in veterinary medicine. Among other things, VIN hosts forums where veterinarians engage in discussions about medical issues, cases in their practices, etc. Over half of the small animal veterinarians in the U.S. and Canada use VIN. Analysis of veterinary data could not only improve pet health care, but also provide early warning signs of infectious disease outbreaks, emerging zoonotic diseases, exposures to environmental toxins, and contamination in the food chain. We obtained over 15,000 VIN message board threads representing three topics: cardiology, endocrinology, and feline internal medicine. We did basic cleaning, removing html tags and tokenizing numbers. For training, we used 4,629 threads, consisting of 25,944 individual posts. We developed classifiers to identify six semantic categories: ANIMAL, DISEASE/SYMPTOM4, DRUG, HUMAN, TEST, and OTHER. The message board posts contain an abundance of veterinary terminology and jargon, so two domain experts5 from VIN created a test set (answer key) for our evaluation. We defined annotation guidelines6 for each semantic category and conducted an inter-annotator agreement study to measure the consistency of the two domain experts on 30 message board posts, which contained 1,473 noun phrases. The annotators achieved a relatively high κ score of .80. Each annotator then labeled an additional 35 documents, which gave us a test set containing 100 manually annotated message board posts. The table below shows the distribution of semantic classes in the test set. Animal Dis/Sym Drug Test Human Other 612 900 369 404 818 1723 To select seed words, we used the procedure proposed by Roark and Charniak (1998), ranking all of the head nouns in the training corpus by frequency and manually selecting the first 10 nouns that unambiguously belong to each category.7 This process is fast, relatively objective, and guaranteed to yield high-frequency terms, which is important for bootstrapping. We used the Stanford part-ofspeech tagger (Toutanova et al., 2003) to identify nouns, and our own simple rule-based NP chunker. 4.2 Baselines To assess the difficulty of our data set and task, we evaluated several baselines. The first baseline searches for each head noun in WordNet and labels the noun as category Ck if it has a hypernym synset corresponding to that category. We manually identified the WordNet synsets that, to the best of our ability, seem to most closely correspond 4We used a single category for diseases and symptoms because our domain experts had difficulty distinguishing between them. A veterinary consultant explained that the same term (e.g., diabetes) may be considered a symptom in one context if it is secondary to another condition (e.g., pancreatitis) and a disease in a different context if it is the primary diagnosis. 5One annotator is a veterinarian and the other is a veterinary technician. 6The annotators were also allowed to label an NP as POS Error if it was clearly misparsed. These cases were not used in the evaluation. 7We used 20 seeds for DIS/SYM because we merged two categories and for OTHER because it is a broad catch-all class. 280 Method Animal Dis/Sym Drug Test Human Other Avg BASELINES WordNet 32/80/46 21/81/34 25/35/29 NA 62/66/64 NA 35/66/45.8 Seeds 38/100/55 14/99/25 21/97/35 29/94/45 80/99/88 18/93/30 37/98/53.1 Supervised 67/94/78 20/88/33 24/96/39 34/90/49 79/99/88 31/91/46 45/94/60.7 Ind. Self-Train I.13 61/84/71 39/80/52 53/77/62 55/70/61 81/96/88 30/82/44 58/81/67.4 CROSS-CATEGORY BOOTSTRAPPED CLASSIFIERS Contextual I.1 59/77/67 33/84/47 42/80/55 49/77/59 82/93/87 33/80/47 53/82/64.3 XCategory I.45 86/71/78 57/82/67 70/78/74 73/65/69 85/92/89 46/82/59 75/78/76.1 XCat+OSCPD I.40 86/69/77 59/81/68 72/70/71 72/69/71 86/92/89 50/81/62 75/76/75.6 XCat+OSCPD+SF I.39 86/70/77 60/81/69 69/81/75 73/69/71 86/91/89 50/81/62 75/78/76.6 Table 1: Experimental results, reported as Recall/Precision/F score to each semantic class. We do not report WordNet results for TEST because there did not seem be an appropriate synset, or for the OTHER category because that is a catch-all class. The first row of Table 1 shows the results, which are reported as Recall/Precision/F score8. The WordNet baseline yields low recall (21-32%) for every category except HUMAN, which confirms that many veterinary terms are not present in WordNet. The surprisingly low precision for some categories is due to atypical word uses (e.g., patient, boy, and girl are HUMAN in WordNet but nearly always ANIMALS in our domain), and overgeneralities (e.g., WordNet lists calcium as a DRUG). The second baseline simply labels every instance of a seed with its designated semantic class. All non-seed instances remain unlabeled. As expected, the seeds produce high precision but low recall. The exception is HUMAN, where 80% of the instances match a seed word, undoubtedly because five of the ten HUMAN seeds are 1st and 2nd person pronouns, which are extremely common. A third baseline trains semantic classifiers using supervised learning by performing 10-fold crossvalidation on the test set. The feature set and classifier settings are exactly the same as with our bootstrapped classifiers.9 Supervised learning achieves good precision but low recall for all categories except ANIMAL and HUMAN. In the next section, we present the experimental results for our bootstrapped classifiers. 4.3 Results for Bootstrapped Classifiers The bottom section of Table 1 displays the results for our bootstrapped classifiers. The Contextual I.1 row shows results after just the first iteration, 8We use an F(1) score, where recall and precision are equally weighted. 9For all of our classifiers, supervised and bootstrapped, we label all instances of the seed words first and then apply the classifiers to the unlabeled (non-seed) instances. which trains only the strictly contextual classifiers. The average F score improved from 53.1 for the seeds alone to 64.3 with the contextual classifiers. The next row, XCategory I.45, shows the results after cross-category bootstrapping, which ended after 45 iterations. (We indicate the number of iterations until bootstrapping ended using the notation I.#.) With cross-category bootstrapping, the average F score increased from 64.3 to 76.1. A closer inspection reveals that all of the semantic categories except HUMAN achieved large recall gains. And importantly, these recall gains were obtained with relatively little loss of precision, with the exception of TEST. Next, we measured the impact of the onesemantic-class-per-discourse heuristic, shown as XCat+OSCPD I.40. From Table 1, it appears that OSCPD produced mixed results: recall increased by 1-4 points for DIS/SYM, DRUG, HUMAN, and OTHER, but precision was inconsistent, improving by +4 for TEST but dropping by -8 for DRUG. However, this single snapshot in time does not tell the full story. Figure 2 shows the performance of the classifiers during the course of bootstrapping. The OSCPD heuristic produced a steeper learning curve, and consistently improved performance until the last few iterations when its performance dipped. This is probably due to the fact that noise gradually increases during bootstrapping, so incorrect labels are more likely and OSCPD will compound any mistakes by the classifier. A good future strategy might be to use the OSCPD heuristic only during the early stages of bootstrapping when the classifier’s decisions are most reliable. We also evaluated the effect of dynamically created semantic features. When added to the basic XCategory system, they had almost no effect. We suspect this is because the semantic features are sparse during most of the bootstrapping process. However, the semantic features did im281 0 5 10 15 20 25 30 35 40 45 64 66 68 70 72 74 76 78 F measure (%) # of iterations independent self−training cross−category bootstrapping +OSCPD +OSCPD+SemFeat Figure 2: Average F scores after each iteration prove performance when coupled with the OSCPD heuristic, presumably because the OSCPD heuristic aggressively labels more instances in the earlier stages of bootstrapping, increasing the prevalence of semantic class tags. The XCat+OSCPD+SF I.39 row in Table 1 shows that the semantic features coupled with OSCPD dramatically increased the precision for DRUG, yielding the best overall F score of 76.6. We conducted one additional experiment to assess the benefits of cross-category bootstrapping. We created an analogous suite of classifiers using self-training, where each classifier independently labels the instances that it is most confident about, adds them only to its own training set, and then retrains itself. The Ind. Self-Train I.13 row in Table 1 shows that these classifiers achieved only 58% recall (compared to 75% for XCategory) and an average F score of 67.4 (compared to 76.1 for XCategory). One reason for the disparity is that the self-training model ended after just 13 bootstrapping cycles (I.13), given the same threshold values. To see if we could push it further, we lowered the confidence threshold to 0 and it continued learning through 35 iterations. Even so, its final score was 65% recall with 79% precision, which is still well below the 75% recall with 78% precision produced by the XCategory model. These results support our claim that cross-category bootstrapping is more effective than independently selftrained models. Figure 3 tracks the recall and precision scores of the XCat+OSCPD+SF system as bootstrapping progresses. This graph shows the sustained momentum of cross-category bootstrapping: re0 5 10 15 20 25 30 35 40 50 55 60 65 70 75 80 85 # of iterations Precision Recall Figure 3: Recall and Precision scores during cross-category bootstrapping call steadily improves while precision stays consistently high with only a slight dropoff at the end. 4.4 Analysis To assess the impact of corpus size, we generated a learning curve with randomly selected subsets of the training texts. Figure 4 shows the average F score of our best system using 1 16, 1 8, 1 4, 1 2, 3 4, and all of the data. With just 1 16th of the training set, the system has about 1,600 message board posts to use for training, which yields a similar F score (roughly 61%) as the supervised baseline that used 100 manually annotated posts via 10-fold crossvalidation. So with 16 times more text, seed-based bootstrapping achieves roughly the same results as supervised learning. This result reflects the natural trade-off between supervised learning and seedbased bootstrapping. Supervised learning exploits manually annotated data, but must make do with a relatively small amount of training text because manual annotations are expensive. In contrast, seed-based bootstrapping exploits a small number of human-provided seeds, but needs a larger set of (unannotated) texts for training because the seeds produce relatively sparse annotations of the texts. An additional advantage of seed-based bootstrapping methods is that they can easily exploit unlimited amounts of training text. For many domains, large text collections are readily available. Figure 4 shows a steady improvement in performance as the amount of training text grows. Overall, the F score improves from roughly 61% to nearly 77% simply by giving the system access to more unannotated text during bootstrapping. We also evaluated the effectiveness of our sliding confidence threshold (Section 3.6). The table below shows the results using fixed thresholds 282 0 1/16 1/8 1/4 1/2 3/4 1 0 20 40 60 65 70 75 80 ration of data F measure (%) Figure 4: Learning Curve of 1.0, 1.5, 2.0, as well as the sliding threshold (which begins at 2.0 and ends at 1.0 decreasing by 0.2 when the number of newly labeled instances falls below 3000 (i.e., < 500 per category, on average). This table depicts the expected trade-off between recall and precision for the fixed thresholds, with higher thresholds producing higher precision but lower recall. The sliding threshold produces the best F score, achieving the best balance of high recall and precision. θcf R/P/F 1.0 71/77/74.1 1.5 69/81/74.7 2.0 65/82/72.4 Sliding 75/78/76.6 As mentioned in Section 3.3, we used 3 times as many negative instances as positive instances for every semantic category during bootstrapping. This ratio was based on early experiments where we needed to limit the number of negative instances per category because the crosscategory framework naturally produces an extremely skewed negative/positive training set. We revisited this issue to empirically assess the impact of the negative/positive ratio on performance. The table below shows recall, precision, and F score results when we vary the ratio from 1:1 to 5:1. A 1:1 ratio seems to be too conservative, improving precision a bit but lowering recall. However the difference in performance between the other ratios is small. Our conclusion is that a 1:1 ratio is too restrictive but, in general, the cross-category bootstrapping process is relatively insensitive to the specific negative/positive ratio used. Our observation from preliminary experiments, however, is that the negative/positive ratio does need to be controlled, or else the dominant categories overwhelm the less frequent categories with negative instances. Neg:Pos R/P/F 1:1 72/79/75.2 2:1 74/78/76.1 3:1 75/78/76.6 4:1 75/77/76.0 5:1 76/77/76.4 Finally, we examined performance on gendered pronouns (he/she/him/her), which can refer to either animals or people in the veterinary domain. 84% (220/261) of the gendered pronouns were annotated as ANIMAL in the test set. Our classifier achieved 95% recall (209/220) and 90% precision (209/232) for ANIMAL and 15% recall (6/41) and 100% precision (6/6) for HUMAN. So while it failed to recognize most of the (relatively few) gendered pronouns that refer to a person, it was highly effective at identifying the ANIMAL references and it was always correct when it did assign a HUMAN tag to a pronoun. 5 Conclusions We presented a novel technique for inducing domain-specific semantic class taggers from a handful of seed words and an unannotated text collection. Our results showed that the induced taggers achieve good performance on six semantic categories associated with the domain of veterinary medicine. Our technique allows semantic class taggers to be rapidly created for specialized domains with minimal human effort. In future work, we plan to investigate whether these semantic taggers can be used to improve other tasks. Acknowledgments We are very grateful to the people at the Veterinary Information Network for providing us access to their resources. Special thanks to Paul Pion, DVM and Nicky Mastin, DVM for making their data available to us, and to Sherri Lofing and Becky Lundgren, DVM for their time and expertise in creating the gold standard annotations. This research was supported in part by Department of Homeland Security Grant N0014-07-1-0152 and Air Force Contract FA8750-09-C-0172 under the DARPA Machine Reading Program. References ACE. 2005. NIST ACE evaluation website. In http://www.nist.gov/speech/tests/ace/2005. 283 ACE. 2007. NIST ACE evaluation website. In http://www.nist.gov/speech/tests/ace/2007. ACE. 2008. NIST ACE evaluation website. In http://www.nist.gov/speech/tests/ace/2008. Daniel M. Bikel, Scott Miller, Richard Schwartz, and Ralph Weischedel. 1997. Nymble: a highperformance learning name-finder. In Proceedings of ANLP-97, pages 194–201. A. Blum and T. Mitchell. 1998. Combining Labeled and Unlabeled Data with Co-Training. In Proceedings of the 11th Annual Conference on Computational Learning Theory (COLT-98). Andrew Carlson, Justin Betteridge, Estevam R. Hruschka Jr., and Tom M. Mitchell. 2009. Coupling semi-supervised learning of categories and relations. In HLT-NAACL 2009 Workshop on Semi-Supervised Learning for NLP. M. Collins and Y. Singer. 1999. Unsupervised Models for Named Entity Classification. In Proceedings of the Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora (EMNLP/VLC-99). S. Cucerzan and D. Yarowsky. 1999. Language Independent Named Entity Recognition Combining Morphologi cal and Contextual Evidence. In Proceedings of the Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora (EMNLP/VLC-99). O. Etzioni, M. Cafarella, D. Downey, A. Popescu, T. Shaked, S. Soderland, D. Weld, and A. Yates. 2005. Unsupervised named-entity extraction from the web: an experimental study. Artificial Intelligence, 165(1):91–134, June. M.B. Fleischman and E.H. Hovy. 2002. Fine grained classification of named entities. In Proceedings of the COLING conference, August. T. Joachims. 1999. Making Large-Scale Support Vector Machine Learning Practical. In A. Smola B. Sch¨olkopf, C. Burges, editor, Advances in Kernel Methods: Support Vector Machines. MIT Press, Cambridge, MA. S. Keerthi and D. DeCoste. 2005. A Modified Finite Newton Method for Fast Solution of Large Scale Linear SVMs. Journal of Machine Learning Research. Mamoru Komachi, Taku Kudo, Masashi Shimbo, and Yuji Matsumoto. 2008. Graph-based analysis of semantic drift in espresso-like bootstrapping algorithms. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing. Z. Kozareva, E. Riloff, and E. Hovy. 2008. Semantic Class Learning from the Web with Hyponym Pattern Linkage Graphs. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (ACL-08). D. McClosky, E. Charniak, and M Johnson. 2006. Effective self-training for parsing. In HLT-NAACL2006. T. McIntosh and J. Curran. 2009. Reducing Semantic Drift with Bagging and Distributional Similarity. In Proceedings of the 47th Annual Meeting of the Association for Computational Linguistics. R. Mihalcea. 2004. Co-training and Self-training for Word Sense Disambiguation. In CoNLL-2004. G. Miller. 1990. Wordnet: An On-line Lexical Database. International Journal of Lexicography, 3(4). C. Mueller, S. Rapp, and M. Strube. 2002. Applying co-training to reference resolution. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. V. Ng and C. Cardie. 2003. Weakly supervised natural language learning without redundant views. In HLTNAACL-2003. V. Ng. 2007. Semantic Class Induction and Coreference Resolution. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics. Cheng Niu, Wei Li, Jihong Ding, and Rohini K. Srihari. 2003. A bootstrapping approach to named entity classification using successive learners. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics (ACL-03), pages 335–342. M. Pas¸ca. 2004. Acquisition of categorized named entities for web search. In Proc. of the Thirteenth ACM International Conference on Information and Knowledge Management, pages 137–145. W. Phillips and E. Riloff. 2002. Exploiting Strong Syntactic Heuristics and Co-Training to Learn Semantic Lexicons. In Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing, pages 125–132. E. Riloff and R. Jones. 1999. Learning Dictionaries for Information Extraction by Multi-Level Bootstrapping. In Proceedings of the Sixteenth National Conference on Artificial Intelligence. E. Riloff and J. Shepherd. 1997. A Corpus-Based Approach for Building Semantic Lexicons. In Proceedings of the Second Conference on Empirical Methods in Natural Language Processing, pages 117– 124. B. Roark and E. Charniak. 1998. Noun-phrase Cooccurrence Statistics for Semi-automatic Semantic Lexicon Construction. In Proceedings of the 36th Annual Meeting of the Association for Computational Linguistics, pages 1110–1116. 284 M. Thelen and E. Riloff. 2002. A Bootstrapping Method for Learning Semantic Lexicons Using Extraction Pa ttern Contexts. In Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing, pages 214–221. K. Toutanova, D. Klein, C. Manning, and Y. Singer. 2003. Feature-Rich Part-of-Speech Tagging with a Cyclic Dependency Network. In Proceedings of HLT-NAACL 2003. R. Yangarber. 2003. Counter-training in the discovery of semantic patterns. In Proceedings of the 41th Annual Meeting of the Association for Computational Linguistics. D. Yarowsky. 1995. Unsupervised Word Sense Disambiguation Rivaling Supervised Methods. In Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics. Imed Zitouni and Radu Florian. 2009. Cross-language information propagation for arabic mention detection. ACM Transactions on Asian Language Information Processing (TALIP), 8(4):1–21. 285
|
2010
|
29
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 21–29, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Bitext Dependency Parsing with Bilingual Subtree Constraints Wenliang Chen, Jun’ichi Kazama and Kentaro Torisawa Language Infrastructure Group, MASTAR Project National Institute of Information and Communications Technology 3-5 Hikari-dai, Seika-cho, Soraku-gun, Kyoto, Japan, 619-0289 {chenwl, kazama, torisawa}@nict.go.jp Abstract This paper proposes a dependency parsing method that uses bilingual constraints to improve the accuracy of parsing bilingual texts (bitexts). In our method, a targetside tree fragment that corresponds to a source-side tree fragment is identified via word alignment and mapping rules that are automatically learned. Then it is verified by checking the subtree list that is collected from large scale automatically parsed data on the target side. Our method, thus, requires gold standard trees only on the source side of a bilingual corpus in the training phase, unlike the joint parsing model, which requires gold standard trees on the both sides. Compared to the reordering constraint model, which requires the same training data as ours, our method achieved higher accuracy because of richer bilingual constraints. Experiments on the translated portion of the Chinese Treebank show that our system outperforms monolingual parsers by 2.93 points for Chinese and 1.64 points for English. 1 Introduction Parsing bilingual texts (bitexts) is crucial for training machine translation systems that rely on syntactic structures on either the source side or the target side, or the both (Ding and Palmer, 2005; Nakazawa et al., 2006). Bitexts could provide more information, which is useful in parsing, than a usual monolingual texts that can be called “bilingual constraints”, and we expect to obtain more accurate parsing results that can be effectively used in the training of MT systems. With this motivation, there are several studies aiming at highly accurate bitext parsing (Smith and Smith, 2004; Burkett and Klein, 2008; Huang et al., 2009). This paper proposes a dependency parsing method, which uses the bilingual constraints that we call bilingual subtree constraints and statistics concerning the constraints estimated from large unlabeled monolingual corpora. Basically, a (candidate) dependency subtree in a source-language sentence is mapped to a subtree in the corresponding target-language sentence by using word alignment and mapping rules that are automatically learned. The target subtree is verified by checking the subtree list that is collected from unlabeled sentences in the target language parsed by a usual monolingual parser. The result is used as additional features for the source side dependency parser. In this paper, our task is to improve the source side parser with the help of the translations on the target side. Many researchers have investigated the use of bilingual constraints for parsing (Burkett and Klein, 2008; Zhao et al., 2009; Huang et al., 2009). For example, Burkett and Klein (2008) show that parsing with joint models on bitexts improves performance on either or both sides. However, their methods require that the training data have tree structures on both sides, which are hard to obtain. Our method only requires dependency annotation on the source side and is much simpler and faster. Huang et al. (2009) proposes a method, bilingual-constrained monolingual parsing, in which a source-language parser is extended to use the re-ordering of words between two sides’ sentences as additional information. The input of their method is the source trees with their translation on the target side as ours, which is much easier to obtain than trees on both sides. However, their method does not use any tree structures on 21 the target side that might be useful for ambiguity resolution. Our method achieves much greater improvement because it uses the richer subtree constraints. Our approach takes the same input as Huang et al. (2009) and exploits the subtree structure on the target side to provide the bilingual constraints. The subtrees are extracted from large-scale autoparsed monolingual data on the target side. The main problem to be addressed is mapping words on the source side to the target subtree because there are many to many mappings and reordering problems that often occur in translation (Koehn et al., 2003). We use an automatic way for generating mapping rules to solve the problems. Based on the mapping rules, we design a set of features for parsing models. The basic idea is as follows: if the words form a subtree on one side, their corresponding words on the another side will also probably form a subtree. Experiments on the translated portion of the Chinese Treebank (Xue et al., 2002; Bies et al., 2007) show that our system outperforms state-ofthe-art monolingual parsers by 2.93 points for Chinese and 1.64 points for English. The results also show that our system provides higher accuracies than the parser of Huang et al. (2009). The rest of the paper is organized as follows: Section 2 introduces the motivation of our idea. Section 3 introduces the background of dependency parsing. Section 4 proposes an approach of constructing bilingual subtree constraints. Section 5 explains the experimental results. Finally, in Section 6 we draw conclusions and discuss future work. 2 Motivation In this section, we use an example to show the idea of using the bilingual subtree constraints to improve parsing performance. Suppose that we have an input sentence pair as shown in Figure 1, where the source sentence is in English, the target is in Chinese, the dashed undirected links are word alignment links, and the directed links between words indicate that they have a (candidate) dependency relation. In the English side, it is difficult for a parser to determine the head of word “with” because there is a PP-attachment problem. However, in Chinese it is unambiguous. Therefore, we can use the information on the Chinese side to help disambiguaHe ate the meat with a fork . Ԇ(He) ⭘(use) ৹ᆀ(fork) ਲ਼(eat) 㚹(meat) DŽ(.) Figure 1: Example for disambiguation tion. There are two candidates “ate” and “meat” to be the head of “with” as the dashed directed links in Figure 1 show. By adding “fork”, we have two possible dependency relations, “meat-with-fork” and “ate-with-fork”, to be verified. First, we check the possible relation of “meat”, “with”, and “fork”. We obtain their corresponding words “肉(meat)”, “用(use)”, and “叉子(fork)” in Chinese via the word alignment links. We verify that the corresponding words form a subtree by looking up a subtree list in Chinese (described in Section 4.1). But we can not find a subtree for them. Next, we check the possible relation of “ate”, “with”, and “fork”. We obtain their corresponding words “吃(ate)”, “用(use)”, and “叉子(fork)”. Then we verify that the words form a subtree by looking up the subtree list. This time we can find the subtree as shown in Figure 2. ⭘(use) ৹ᆀ(fork) ਲ਼(eat) Figure 2: Example for a searched subtree Finally, the parser may assign “ate” to be the head of “with” based on the verification results. This simple example shows how to use the subtree information on the target side. 3 Dependency parsing For dependency parsing, there are two main types of parsing models (Nivre and McDonald, 2008; Nivre and Kubler, 2006): transition-based (Nivre, 2003; Yamada and Matsumoto, 2003) and graphbased (McDonald et al., 2005; Carreras, 2007). Our approach can be applied to both parsing models. In this paper, we employ the graph-based MST parsing model proposed by McDonald and Pereira 22 (2006), which is an extension of the projective parsing algorithm of Eisner (1996). To use richer second-order information, we also implement parent-child-grandchild features (Carreras, 2007) in the MST parsing algorithm. 3.1 Parsing with monolingual features Figure 3 shows an example of dependency parsing. In the graph-based parsing model, features are represented for all the possible relations on single edges (two words) or adjacent edges (three words). The parsing algorithm chooses the tree with the highest score in a bottom-up fashion. ROOT He ate the meat with a fork . Figure 3: Example of dependency tree In our systems, the monolingual features include the first- and second- order features presented in (McDonald et al., 2005; McDonald and Pereira, 2006) and the parent-child-grandchild features used in (Carreras, 2007). We call the parser with the monolingual features monolingual parser. 3.2 Parsing with bilingual features In this paper, we parse source sentences with the help of their translations. A set of bilingual features are designed for the parsing model. 3.2.1 Bilingual subtree features We design bilingual subtree features, as described in Section 4, based on the constraints between the source subtrees and the target subtrees that are verified by the subtree list on the target side. The source subtrees are from the possible dependency relations. 3.2.2 Bilingual reordering feature Huang et al. (2009) propose features based on reordering between languages for a shift-reduce parser. They define the features based on wordalignment information to verify that the corresponding words form a contiguous span for resolving shift-reduce conflicts. We also implement similar features in our system. 4 Bilingual subtree constraints In this section, we propose an approach that uses the bilingual subtree constraints to help parse source sentences that have translations on the target side. We use large-scale auto-parsed data to obtain subtrees on the target side. Then we generate the mapping rules to map the source subtrees onto the extracted target subtrees. Finally, we design the bilingual subtree features based on the mapping rules for the parsing model. These features indicate the information of the constraints between bilingual subtrees, that are called bilingual subtree constraints. 4.1 Subtree extraction Chen et al. (2009) propose a simple method to extract subtrees from large-scale monolingual data and use them as features to improve monolingual parsing. Following their method, we parse large unannotated data with a monolingual parser and obtain a set of subtrees (STt) in the target language. We encode the subtrees into string format that is expressed as st = w : hid(−w : hid)+1, where w refers to a word in the subtree and hid refers to the word ID of the word’s head (hid=0 means that this word is the root of a subtree). Here, word ID refers to the ID (starting from 1) of a word in the subtree (words are ordered based on the positions of the original sentence). For example, “He” and “ate” have a left dependency arc in the sentence shown in Figure 3. The subtree is encoded as “He:2ate:0”. There is also a parent-child-grandchild relation among “ate”, “with”, and “fork”. So the subtree is encoded as “ate:0-with:1-fork:2”. If a subtree contains two nodes, we call it a bigramsubtree. If a subtree contains three nodes, we call it a trigram-subtree. From the dependency tree of Figure 3, we obtain the subtrees, as shown in Figure 4 and Figure 5. Figure 4 shows the extracted bigram-subtrees and Figure 5 shows the extracted trigram-subtrees. After extraction, we obtain a set of subtrees. We remove the subtrees occurring only once in the data. Following Chen et al. (2009), we also group the subtrees into different sets based on their frequencies. 1+ refers to matching the preceding element one or more times and is the same as a regular expression in Perl. 23 ate He He:1:2-ate:2:0 ate meat ate:1:0-meat:2:1 ate with ate:1:0-with:2:1 meat the the:1:2-meat:2:0 with fork with:1:0-fork:2:1 fork a a:1:2-fork:2:0 Figure 4: Examples of bigram-subtrees ate meat with ate:1:0-meat:2:1-with:3:1 ate with . ate:1:0-with:2:1-.:3:1 (a) He:1:3-NULL:2:3-ate:3:0 ate He NULL ate NULL meat ate:1:0-NULL:2:1-meat:3:1 the:1:3-NULL:2:3-meat:3:0 a:1:3-NULL:2:3-fork:3:0 with:1:0-NULL:2:1-fork:3:1 ate:1:0-the:2:3-meat:3:1 ate:1:0-with:2:1-fork:3:2 with:1:0-a:2:3-fork:3:1 NULL:1:2-He:2:3-ate:3:0 He:1:3-NULL:2:1-ate:3:0 ate:1:0-meat:2:1-NULL:3:2 ate:1:0-NULL:2:3-with:3:1 with:1:0-fork:2:1-NULL:3:2 NULL:1:2-a:2:3-fork:3:0 a:1:3-NULL:2:1-fork:3:0 ate:1:0-NULL:2:3-.:3:1 ate:1:0-.:2:1-NULL:3:2 (b) NULL:1:2-the:2:3-meat:3:0 the:1:3-NULL:2:1-meat:3:0 Figure 5: Examples of trigram-subtrees 4.2 Mapping rules To provide bilingual subtree constraints, we need to find the characteristics of subtree mapping for the two given languages. However, subtree mapping is not easy. There are two main problems: MtoN (words) mapping and reordering, which often occur in translation. MtoN (words) mapping means that a source subtree with M words is mapped onto a target subtree with N words. For example, 2to3 means that a source bigram-subtree is mapped onto a target trigram-subtree. Due to the limitations of the parsing algorithm (McDonald and Pereira, 2006; Carreras, 2007), we only use bigram- and trigram-subtrees in our approach. We generate the mapping rules for the 2to2, 2to3, 3to3, and 3to2 cases. For trigram-subtrees, we only consider the parentchild-grandchild type. As for the use of other types of trigram-subtrees, we leave it for future work. We first show the MtoN and reordering problems by using an example in Chinese-English translation. Then we propose a method to automatically generate mapping rules. 4.2.1 Reordering and MtoN mapping in translation Both Chinese and English are classified as SVO languages because verbs precede objects in simple sentences. However, Chinese has many characteristics of such SOV languages as Japanese. The typical cases are listed below: 1) Prepositional phrases modifying a verb precede the verb. Figure 6 shows an example. In English the prepositional phrase “at the ceremony” follows the verb “said”, while its corresponding prepositional phrase “在(NULL) 仪式(ceremony) 上(at)” precedes the verb “说(say)” in Chinese. ൘ Ԛᔿ к 䈤 Said at the ceremony Figure 6: Example for prepositional phrases modifying a verb 2) Relative clauses precede head noun. Figure 7 shows an example. In Chinese the relative clause “今天(today) 签字(signed)” precedes the head noun “项目(project)”, while its corresponding clause “signed today” follows the head noun “projects” in English. ӺཙㆮᆇⲴй њ 亩ⴞ The 3 projects signed today Figure 7: Example for relative clauses preceding the head noun 3) Genitive constructions precede head noun. For example, “汽车(car) 轮子(wheel)” can be translated as “the wheel of the car”. 4) Postposition in many constructions rather than prepositions. For example, “桌子(table) 上(on)” can be translated as “on the table”. 24 We can find the MtoN mapping problem occurring in the above cases. For example, in Figure 6, trigram-subtree “在(NULL):3-上(at):1-说(say):0” is mapped onto bigram-subtree “said:0-at:1”. Since asking linguists to define the mapping rules is very expensive, we propose a simple method to easily obtain the mapping rules. 4.2.2 Bilingual subtree mapping To solve the mapping problems, we use a bilingual corpus, which includes sentence pairs, to automatically generate the mapping rules. First, the sentence pairs are parsed by monolingual parsers on both sides. Then we perform word alignment using a word-level aligner (Liang et al., 2006; DeNero and Klein, 2007). Figure 8 shows an example of a processed sentence pair that has tree structures on both sides and word alignment links. ROOT ԆԜ ༴Ҿ ⽮Պ 䗩㕈 DŽ ROOT They are on the fringes of society . Figure 8: Example of auto-parsed bilingual sentence pair From these sentence pairs, we obtain subtree pairs. First, we extract a subtree (sts) from a source sentence. Then through word alignment links, we obtain the corresponding words of the words of sts. Because of the MtoN problem, some words lack of corresponding words in the target sentence. Here, our approach requires that at least two words of sts have corresponding words and nouns and verbs need corresponding words. If not, it fails to find a subtree pair for sts. If the corresponding words form a subtree (stt) in the target sentence, sts and stt are a subtree pair. We also keep the word alignment information in the target subtree. For example, we extract subtree “社 会(society):2-边缘(fringe):0” on the Chinese side and get its corresponding subtree “fringes(W 2):0of:1-society(W 1):2” on the English side, where W 1 means that the target word is aligned to the first word of the source subtree, and W 2 means that the target word is aligned to the second word of the source subtree. That is, we have a subtree pair: “社会(society):2-边缘(fringe):0” and “fringe(W 2):0-of:1-society(W 1):2”. The extracted subtree pairs indicate the translation characteristics between Chinese and English. For example, the pair “社会(society):2边缘(fringe):0” and “fringes:0-of:1-society:2” is a case where “Genitive constructions precede/follow the head noun”. 4.2.3 Generalized mapping rules To increase the mapping coverage, we generalize the mapping rules from the extracted subtree pairs by using the following procedure. The rules are divided by “=>” into two parts: source (left) and target (right). The source part is from the source subtree and the target part is from the target subtree. For the source part, we replace nouns and verbs using their POS tags (coarse grained tags). For the target part, we use the word alignment information to represent the target words that have corresponding source words. For example, we have the subtree pair: “社会(society):2-边缘(fringe):0” and “fringes(W 2):0-of:1-society(W 1):2”, where “of” does not have a corresponding word, the POS tag of “社会(society)” is N, and the POS tag of “边缘(fringe)” is N. The source part of the rule becomes “N:2-N:0” and the target part becomes “W 2:0-of:1-W 1:2”. Table 1 shows the top five mapping rules of all four types ordered by their frequencies, where W 1 means that the target word is aligned to the first word of the source subtree, W 2 means that the target word is aligned to the second word, and W 3 means that the target word is aligned to the third word. We remove the rules that occur less than three times. Finally, we obtain 9,134 rules for 2to2, 5,335 for 2to3, 7,450 for 3to3, and 1,244 for 3to2 from our data. After experiments with different threshold settings on the development data sets, we use the top 20 rules for each type in our experiments. The generalized mapping rules might generate incorrect target subtrees. However, as described in Section 4.3.1, the generated subtrees are verified by looking up list STt before they are used in the parsing models. 4.3 Bilingual subtree features Informally, if the words form a subtree on the source side, then the corresponding words on the target side will also probably form a subtree. For 25 # rules freq 2to2 mapping 1 N:2 N:0 => W 1:2 W 2:0 92776 2 V:0 N:1 => W 1:0 W 2:1 62437 3 V:0 V:1 => W 1:0 W 2:1 49633 4 N:2 V:0 => W 1:2 W 2:0 43999 5 的:2 N:0 => W 2:0 W 1:2 25301 2to3 mapping 1 N:2-N:0 => W 2:0-of:1-W 1:2 10361 2 V:0-N:1 => W 1:0-of:1-W 2:2 4521 3 V:0-N:1 => W 1:0-to:1-W 2:2 2917 4 N:2-V:0 => W 2:0-of:1-W 1:2 2578 5 N:2-N:0 => W 1:2-’:3-W 2:0 2316 3to2 mapping 1 V:2-的/DEC:3-N:0 => W 1:0-W 3:1 873 2 V:2-的/DEC:3-N:0 => W 3:2-W 1:0 634 3 N:2-的/DEG:3-N:0 => W 1:0-W 3:1 319 4 N:2-的/DEG:3-N:0 => W 3:2-W 1:0 301 5 V:0-的/DEG:3-N:1 => W 3:0-W 1:1 247 3to3 mapping 1 V:0-V:1-N:2 => W 1:0-W 2:1-W 3:2 9580 2 N:2-的/DEG:3-N:0 => W 3:0-W 2:1-W 1:2 7010 3 V:0-N:3-N:1 => W 1:0-W 2:3-W 3:1 5642 4 V:0-V:1-V:2 => W 1:0-W 2:1-W 3:2 4563 5 N:2-N:3-N:0 => W 1:2-W 2:3-W 3:0 3570 Table 1: Top five mapping rules of 2to3 and 3to2 example, in Figure 8, words “他们(they)” and “处于(be on)” form a subtree , which is mapped onto the words “they” and “are” on the target side. These two target words form a subtree. We now develop this idea as bilingual subtree features. In the parsing process, we build relations for two or three words on the source side. The conditions of generating bilingual subtree features are that at least two of these source words must have corresponding words on the target side and nouns and verbs must have corresponding words. At first, we have a possible dependency relation (represented as a source subtree) of words to be verified. Then we obtain the corresponding target subtree based on the mapping rules. Finally, we verify that the target subtree is included in STt. If yes, we activate a positive feature to encourage the dependency relation. 䘉 ᱟӺཙ ㆮᆇ Ⲵй њ 亩ⴞ 䘉 ᱟӺཙ ㆮᆇ Ⲵй њ 亩ⴞ Those are the 3 projects signed today Those are the 3 projects signed today Figure 9: Example of features for parsing We consider four types of features based on 2to2, 3to3, 3to2, and 2to3 mappings. In the 2to2, 3to3, and 3to2 cases, the target subtrees do not add new words. We represent features in a direct way. For the 2to3 case, we represent features using a different strategy. 4.3.1 Features for 2to2, 3to3, and 3to2 We design the features based on the mapping rules of 2to2, 3to3, and 3to2. For example, we design features for a 3to2 case from Figure 9. The possible relation to be verified forms source subtree “签字(signed)/VV:2-的(NULL)/DEC:3项目(project)/NN:0” in which “项目(project)” is aligned to “projects” and “签字(signed)” is aligned to “signed” as shown in Figure 9. The procedure of generating the features is shown in Figure 10. We explain Steps (1), (2), (3), and (4) as follows: ㆮᆇ/VV:2-Ⲵ/DEC:3-亩ⴞ/NN:0 projects(W_3) signed(W_1) (1) V:2-Ⲵ/DEC:3-N:0 W_3:0-W_1:1 W 3:2 W 1:0 (2) W_3:2-W_1:0 (3) projects:0-signed:1 projects:2-signed:0 STt (4) 3to2:YES (4) Figure 10: Example of feature generation for 3to2 case (1) Generate source part from the source subtree. We obtain “V:2-的/DEC:3-N:0” from “签 字(signed)/VV:2-的(NULL)/DEC:3-项 目(project)/NN:0”. (2) Obtain target parts based on the matched mapping rules, whose source parts equal “V:2-的/DEC:3-N:0”. The matched rules are “V:2-的/DEC:3-N:0 =>W 3:0-W 1:1” and “V:2-的/DEC:3-N:0 => W 3:2-W 1:0”. Thus, we have two target parts “W 3:0-W 1:1” and “W 3:2-W 1:0”. (3) Generate possible subtrees by consider26 ing the dependency relation indicated in the target parts. We generate a possible subtree “projects:0-signed:1” from the target part “W 3:0W 1:1”, where “projects” is aligned to “项 目(project)(W 3)” and “signed” is aligned to “签 字(signed)(W 1)”. We also generate another possible subtree “projects:2-signed:0” from “W 3:2W 1:0”. (4) Verify that at least one of the generated possible subtrees is a target subtree, which is included in STt. If yes, we activate this feature. In the figure, “projects:0-signed:1” is a target subtree in STt. So we activate the feature “3to2:YES” to encourage dependency relations among “签 字(signed)”, “的(NULL)”, and “项目(project)”. 4.3.2 Features for 2to3 In the 2to3 case, a new word is added on the target side. The first two steps are identical as those in the previous section. For example, a source part “N:2-N:0” is generated from “汽车(car)/NN:2-轮 子(wheel)/NN:0”. Then we obtain target parts such as “W 2:0-of/IN:1-W 1:2”, “W 2:0-in/IN:1W 1:2”, and so on, according to the matched mapping rules. The third step is different. In the target parts, there is an added word. We first check if the added word is in the span of the corresponding words, which can be obtained through word alignment links. We can find that “of” is in the span “wheel of the car”, which is the span of the corresponding words of “汽车(car)/NN:2-轮子(wheel)/NN:0”. Then we choose the target part “W 2:0-of/IN:1W 1:2” to generate a possible subtree. Finally, we verify that the subtree is a target subtree included in STt. If yes, we say feature “2to3:YES” to encourage a dependency relation between “汽 车(car)” and “轮子(wheel)”. 4.4 Source subtree features Chen et al. (2009) shows that the source subtree features (Fsrc−st) significantly improve performance. The subtrees are obtained from the auto-parsed data on the source side. Then they are used to verify the possible dependency relations among source words. In our approach, we also use the same source subtree features described in Chen et al. (2009). So the possible dependency relations are verified by the source and target subtrees. Combining two types of features together provides strong discrimination power. If both types of features are active, building relations is very likely among source words. If both are inactive, this is a strong negative signal for their relations. 5 Experiments All the bilingual data were taken from the translated portion of the Chinese Treebank (CTB) (Xue et al., 2002; Bies et al., 2007), articles 1-325 of CTB, which have English translations with gold-standard parse trees. We used the tool “Penn2Malt”2 to convert the data into dependency structures. Following the study of Huang et al. (2009), we used the same split of this data: 1-270 for training, 301-325 for development, and 271300 for test. Note that some sentence pairs were removed because they are not one-to-one aligned at the sentence level (Burkett and Klein, 2008; Huang et al., 2009). Word alignments were generated from the Berkeley Aligner (Liang et al., 2006; DeNero and Klein, 2007) trained on a bilingual corpus having approximately 0.8M sentence pairs. We removed notoriously bad links in {a, an, the}×{的(DE), 了(LE)} following the work of Huang et al. (2009). For Chinese unannotated data, we used the XIN CMN portion of Chinese Gigaword Version 2.0 (LDC2009T14) (Huang, 2009), which has approximately 311 million words whose segmentation and POS tags are given. To avoid unfair comparison, we excluded the sentences of the CTB data from the Gigaword data. We discarded the annotations because there are differences in annotation policy between CTB and this corpus. We used the MMA system (Kruengkrai et al., 2009) trained on the training data to perform word segmentation and POS tagging and used the Baseline Parser to parse all the sentences in the data. For English unannotated data, we used the BLLIP corpus that contains about 43 million words of WSJ text. The POS tags were assigned by the MXPOST tagger trained on training data. Then we used the Baseline Parser to parse all the sentences in the data. We reported the parser quality by the unlabeled attachment score (UAS), i.e., the percentage of tokens (excluding all punctuation tokens) with correct HEADs. 5.1 Main results The results on the Chinese-source side are shown in Table 2, where “Baseline” refers to the systems 2http://w3.msi.vxu.se/˜nivre/research/Penn2Malt.html 27 with monolingual features, “Baseline2” refers to adding the reordering features to the Baseline, “FBI” refers to adding all the bilingual subtree features to “Baseline2”, “Fsrc−st” refers to the monolingual parsing systems with source subtree features, “Order-1” refers to the first-order models, and “Order-2” refers to the second-order models. The results showed that the reordering features yielded an improvement of 0.53 and 0.58 points (UAS) for the first- and second-order models respectively. Then we added four types of bilingual constraint features one by one to “Baseline2”. Note that the features based on 3to2 and 3to3 can not be applied to the first-order models, because they only consider single dependencies (bigram). That is, in the first model, FBI only includes the features based on 2to2 and 2to3. The results showed that the systems performed better and better. In total, we obtained an absolute improvement of 0.88 points (UAS) for the first-order model and 1.36 points for the second-order model by adding all the bilingual subtree features. Finally, the system with all the features (OURS) outperformed the Baseline by an absolute improvement of 3.12 points for the first-order model and 2.93 points for the second-order model. The improvements of the final systems (OURS) were significant in McNemar’s Test (p < 10−4). Order-1 Order-2 Baseline 84.35 87.20 Baseline2 84.88 87.78 +2to2 85.08 88.07 +2to3 85.23 88.14 +3to3 – 88.29 +3to2 – 88.56 FBI 85.23(+0.88) 88.56(+1.36) Fsrc−st 86.54(+2.19) 89.49(+2.29) OURS 87.47(+3.12) 90.13(+2.93) Table 2: Dependency parsing results of Chinesesource case We also conducted experiments on the Englishsource side. Table 3 shows the results, where abbreviations are the same as in Table 2. As in the Chinese experiments, the parsers with bilingual subtree features outperformed the Baselines. Finally, the systems (OURS) with all the features outperformed the Baselines by 1.30 points for the first-order model and 1.64 for the second-order model. The improvements of the final systems (OURS) were significant in McNemar’s Test (p < 10−3). Order-1 Order-2 Baseline 86.41 87.37 Baseline2 86.86 87.66 +2to2 87.23 87.87 +2to3 87.35 87.96 +3to3 – 88.25 +3to2 – 88.37 FBI 87.35(+0.94) 88.37(+1.00) Fsrc−st 87.25(+0.84) 88.57(+1.20) OURS 87.71(+1.30) 89.01(+1.64) Table 3: Dependency parsing results of Englishsource case 5.2 Comparative results Table 4 shows the performance of the system we compared, where Huang2009 refers to the result of Huang et al. (2009). The results showed that our system performed better than Huang2009. Compared with the approach of Huang et al. (2009), our approach used additional large-scale autoparsed data. We did not compare our system with the joint model of Burkett and Klein (2008) because they reported the results on phrase structures. Chinese English Huang2009 86.3 87.5 Baseline 87.20 87.37 OURS 90.13 89.01 Table 4: Comparative results 6 Conclusion We presented an approach using large automatically parsed monolingual data to provide bilingual subtree constraints to improve bitexts parsing. Our approach remains the efficiency of monolingual parsing and exploits the subtree structure on the target side. The experimental results show that the proposed approach is simple yet still provides significant improvements over the baselines in parsing accuracy. The results also show that our systems outperform the system of previous work on the same data. There are many ways in which this research could be continued. First, we may attempt to apply the bilingual subtree constraints to transition28 based parsing models (Nivre, 2003; Yamada and Matsumoto, 2003). Here, we may design new features for the models. Second, we may apply the proposed method for other language pairs such as Japanese-English and Chinese-Japanese. Third, larger unannotated data can be used to improve the performance further. References Ann Bies, Martha Palmer, Justin Mott, and Colin Warner. 2007. English Chinese translation treebank v 1.0. In LDC2007T02. David Burkett and Dan Klein. 2008. Two languages are better than one (for syntactic parsing). In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 877– 886, Honolulu, Hawaii, October. Association for Computational Linguistics. X. Carreras. 2007. Experiments with a higher-order projective dependency parser. In Proceedings of the CoNLL Shared Task Session of EMNLP-CoNLL 2007, pages 957–961. WL. Chen, J. Kazama, K. Uchimoto, and K. Torisawa. 2009. Improving dependency parsing with subtrees from auto-parsed data. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 570–579, Singapore, August. Association for Computational Linguistics. John DeNero and Dan Klein. 2007. Tailoring word alignments to syntactic machine translation. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 17–24, Prague, Czech Republic, June. Association for Computational Linguistics. Yuan Ding and Martha Palmer. 2005. Machine translation using probabilistic synchronous dependency insertion grammars. In ACL ’05: Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, pages 541–548, Morristown, NJ, USA. Association for Computational Linguistics. J. Eisner. 1996. Three new probabilistic models for dependency parsing: An exploration. In Proc. of the 16th Intern. Conf. on Computational Linguistics (COLING), pages 340–345. Liang Huang, Wenbin Jiang, and Qun Liu. 2009. Bilingually-constrained (monolingual) shift-reduce parsing. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1222–1231, Singapore, August. Association for Computational Linguistics. Chu-Ren Huang. 2009. Tagged Chinese Gigaword Version 2.0, LDC2009T14. Linguistic Data Consortium. P. Koehn, F.J. Och, and D. Marcu. 2003. Statistical phrase-based translation. In Proceedings of NAACL, page 54. Association for Computational Linguistics. Canasai Kruengkrai, Kiyotaka Uchimoto, Jun’ichi Kazama, Yiou Wang, Kentaro Torisawa, and Hitoshi Isahara. 2009. An error-driven word-character hybrid model for joint Chinese word segmentation and POS tagging. In Proceedings of ACL-IJCNLP2009, pages 513–521, Suntec, Singapore, August. Association for Computational Linguistics. Percy Liang, Ben Taskar, and Dan Klein. 2006. Alignment by agreement. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 104–111, New York City, USA, June. Association for Computational Linguistics. R. McDonald and F. Pereira. 2006. Online learning of approximate dependency parsing algorithms. In Proc. of EACL2006. R. McDonald, K. Crammer, and F. Pereira. 2005. Online large-margin training of dependency parsers. In Proc. of ACL 2005. T. Nakazawa, K. Yu, D. Kawahara, and S. Kurohashi. 2006. Example-based machine translation based on deeper nlp. In Proceedings of IWSLT 2006, pages 64–70, Kyoto, Japan. J. Nivre and S. Kubler. 2006. Dependency parsing: Tutorial at Coling-ACL 2006. In CoLING-ACL. J. Nivre and R. McDonald. 2008. Integrating graphbased and transition-based dependency parsers. In Proceedings of ACL-08: HLT, Columbus, Ohio, June. J. Nivre. 2003. An efficient algorithm for projective dependency parsing. In Proceedings of IWPT2003, pages 149–160. David A. Smith and Noah A. Smith. 2004. Bilingual parsing with factored estimation: Using English to parse Korean. In Proceedings of EMNLP. Nianwen Xue, Fu-Dong Chiou, and Martha Palmer. 2002. Building a large-scale annotated Chinese corpus. In Coling. H. Yamada and Y. Matsumoto. 2003. Statistical dependency analysis with support vector machines. In Proceedings of IWPT2003, pages 195–206. Hai Zhao, Yan Song, Chunyu Kit, and Guodong Zhou. 2009. Cross language dependency parsing using a bilingual lexicon. In Proceedings of ACLIJCNLP2009, pages 55–63, Suntec, Singapore, August. Association for Computational Linguistics. 29
|
2010
|
3
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 286–295, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Learning 5000 Relational Extractors Raphael Hoffmann, Congle Zhang, Daniel S. Weld Computer Science & Engineering University of Washington Seattle, WA-98195, USA {raphaelh,clzhang,weld}@cs.washington.edu Abstract Many researchers are trying to use information extraction (IE) to create large-scale knowledge bases from natural language text on the Web. However, the primary approach (supervised learning of relation-specific extractors) requires manually-labeled training data for each relation and doesn’t scale to the thousands of relations encoded in Web text. This paper presents LUCHS, a self-supervised, relation-specific IE system which learns 5025 relations — more than an order of magnitude greater than any previous approach — with an average F1 score of 61%. Crucial to LUCHS’s performance is an automated system for dynamic lexicon learning, which allows it to learn accurately from heuristically-generated training data, which is often noisy and sparse. 1 Introduction Information extraction (IE), the process of generating relational data from natural-language text, has gained popularity for its potential applications in Web search, question answering and other tasks. Two main approaches have been attempted: • Supervised learning of relation-specific extractors (e.g., (Freitag, 1998)), and • “Open” IE — self-supervised learning of unlexicalized, relation-independent extractors (e.g., Textrunner (Banko et al., 2007)). Unfortunately, both methods have problems. Supervised approaches require manually-labeled training data for each relation and hence can’t scale to handle the thousands of relations encoded in Web text. Open extraction is more scalable, but has lower precision and recall. Furthermore, open extraction doesn’t canonicalize relations, so any application using the output must deal with homonymy and synonymy. A third approach, sometimes refered to as weak supervision, is to heuristically match values from a database to text, thus generating a set of training data for self-supervised learning of relationspecific extractors (Craven and Kumlien, 1999). With the Kylin system (Wu and Weld, 2007) applied this idea to Wikipedia by matching values of an article’s infobox1 attributes to corresponding sentences in the article, and suggested that their approach could extract thousands of relations (Wu et al., 2008). Unfortunately, however, they never tested the idea on more than a dozen relations. Indeed, no one has demonstrated a practical way to extract more than about one hundred relations. We note that Wikipedia’s infobox ‘ontology’ is a particularly interesting target for extraction. As a by-product of thousands of contributors, it is broad in coverage and growing quickly. Unfortunately, the schemata are surprisingly noisy and most are sparsely populated; challenging conditions for extraction. This paper presents LUCHS, an autonomous, self-supervised system, which learns 5025 relational extractors — an order of magnitude greater than any previous effort. Like Kylin, LUCHS creates training data by matching Wikipedia attribute values with corresponding sentences, but by itself, this method was insufficient for accurate extraction of most relations. Thus, LUCHS introduces a new technique, dynamic lexicon features, which dramatically improves performance when learning from sparse data and that way enables scalability. 1.1 Dynamic Lexicon Features Figure 1 summarizes the architecture of LUCHS. At the highest level, LUCHS’s offline training process resembles that of Kylin. Wikipedia pages 1A sizable fraction of Wikipedia articles have associated infoboxes — relational summaries of the key aspects of the subject of the article. For example, the infobox for Alan Turing’s Wikipedia page lists the values of 10 attributes, including his birthdate, nationality and doctoral advisor. 286 Matcher Harvester CRF Learner Filtered Lists WWW Lexicon Learner Classifier Learner Training Data Extractor Training Data Lexicons Tuples Pages Article Classifier Extractor Extractor Classified Pages Extraction Learning Figure 1: Architecture of LUCHS. In order to handle sparsity in its heuristically-generated training data, LUCHS generates custom lexicon features when learning each relational extractor. containing infoboxes are used to train a classifier that can predict the appropriate schema for pages missing infoboxes. Additionally, the values of infobox attributes are compared with article sentences to heuristically generate training data. LUCHS’s major innovation is a feature-generation process, which starts by harvesting HTML lists from a 5B document Web crawl, discarding 98% to create a set of 49M semantically-relevant lists. When learning an extractor for relation R, LUCHS extracts seed phrases from R’s training data and uses a semi-supervised learning algorithm to create several relation-specific lexicons at different points on a precision-recall spectrum. These lexicons form Boolean features which, along with lexical and dependency parser-based features, are used to produce a CRF extractor for each relation — one which performs much better than lexiconfree extraction on sparse training data. At runtime, LUCHS feeds pages to the article classfier, which predicts which infobox schema is most appropriate for extraction. Then a small set of relation-specific extractors are applied to each sentence, outputting tuples. Our experiments demonstrate a high F1 score, 61%, across the 5025 relational extractors learned. 1.2 Summary This paper makes several contributions: • We present LUCHS, a self-supervised IE system capable of learning more than an order of magnitude more relation-specific extractors than previous systems. • We describe the construction and use of dynamic lexicon features, a novel technique, that enables hyper-lexicalized extractors which cope effectively with sparse training data. • We evaluate the overall end-to-end performance of LUCHS, showing an F1 score of 61% when extracting relations from randomly selected Wikipedia pages. • We present a comprehensive set of additional experiments, evaluating LUCHS’s individual components, measuring the effect of dynamic lexicon features, testing sensitivity to varying amounts of training data, and categorizing the types of relations LUCHS can extract. 2 Heuristic Generation of Training Data Wikipedia is an ideal starting point for our longterm goal of creating a massive knowledge base of extracted facts for two reasons. First, it is comprehensive, containing a diverse body of content with significant depth. Perhaps more importantly, Wikipedia’s structure facilitates self-supervised extraction. Infoboxes are short, manually-created tabular summaries of many articles’ key facts — effectively defining a relational schema for that class of entity. Since the same facts are often expressed in both article and ontology, matching values of the ontology to the article can deliver valuable, though noisy, training data. For example, the Wikipedia article on “Jerry Seinfeld” contains the sentence “Seinfeld was born in Brooklyn, New York.” and the article’s infobox contains the attribute “birth place = Brooklyn”. By matching the attribute’s value “Brooklyn” to the sentence, we can heuristically generate training data for a birth place extractor. This data is noisy; some attributes will not find matches, while others will find many co-incidental matches. 3 Learning Extractors We first assume that each Wikipedia infobox attribute corresponds to a unique relation (but see Section 5.6) for which we would like to learn a specific extractor. A major challenge with such an approach is scalability. Running a relationspecific extractor for each of Wikipedia’s 34,000 unique infobox attributes on each of Wikipedia’s 50 million sentences would require 1.7 trillion extractor executions. We therefore choose a hierarchical approach that combines both article classifiers and relation extractors. For each infobox schema, LUCHS trains a classifier that predicts if an article is likely to contain that schema. Only when an article 287 is likely to contain a schema, does LUCHS run that schema’s relation extractors. To extract infobox attributes from all of Wikipedia, LUCHS now needs orders of magnitude fewer executions. While this approach does not propagate information from extractors back to article classifiers, experiments confirm that our article classifiers nonetheless deliver accurate results (Section 5.2), reducing the potential benefit of joint inference. In addition, our approach reduces the need for extractors to keep track of the larger context, thus simplifying the extraction problem. We briefly summarize article classification: We use a linear, multi-class classifier with six kinds of features: words in the article title, words in the first sentence, words in the first sentence which are direct objects to the verb ‘to be’, article section headers, Wikipedia categories, and their ancestor categories. We use the voted perceptron algorithm (Freund and Schapire, 1999) for training. More challenging are the attribute extractors, which we wish to be simple, fast, and able to well capture local dependencies. We use a linear-chain conditional random field (CRF) — an undirected graphical model connecting a sequence of input and output random variables, x = (x0, . . . , xT ) and y = (y0, . . . , yT ) (Lafferty et al., 2001). Input variables are assigned words w. The states of output variables represent discrete labels l, e.g. Argi-of-Relj and Other. In our case, variables are connected in a chain, following the first-order Markov assumption. We train to maximize conditional likelihood of output variables given an input probability distribution. The CRF models p(y|x) are represented with a log-linear distribution p(y|x) = 1 Z(x) exp T X t=1 K X k=1 λkfk(yt−1, yt, x, t) where feature functions, f, encode sufficient statistics of (x, y), T is the length of the sequence, K is the number of feature functions, and λk are parameters representing feature weights, which we learn during training. Z(x) is a partition function used to normalize the probabilities to 1. Feature functions allow complex, overlapping global features with lookahead. Common techniques for learning the weights λk include numeric optimization algorithms such as stochastic gradient descent or L-BFGS. In our experiments, we again use the simpler and more efficient voted-perceptron algorithm (Collins, 2002). The linear-chain layout enables efficient interence using the dynamic programming-based Viterbi algorithm (Lafferty et al., 2001). We evaluate nine kinds of Boolean features: Words For each input word w we introduce feature fw w (yt−1, yt, x, t) := 1[xt=w]. State Transitions For each transition between output labels li, lj we add feature ftran li,lj (yt−1, yt, x, t) := 1[yt−1=li∧yt=lj]. Word Contextualization For parameters p and s we add features fprev w (yt−1, yt, x, t) := 1[w∈{xt−p,...,xt−1}] and fsub w (yt−1, yt, x, t) := 1[w∈{xt+1,...,xt+s}] which capture a window of words appearing before and after each position t. Capitalization We add feature fcap(yt−1, yt, x, t) := 1[xtis capitalized]. Digits We add feature fdig(yt−1, yt, x, t) := 1[xtis digits]. Dependencies We set fdep(yt−1, yt, x, t) to the lemmatized sequence of words from xt to the root of the dependency tree, computed using the Stanford parser (Marneffe et al., 2006). First Sentence We set ffs(yt−1, yt, x, t) := 1[xtin first sentence of article]. Gaussians For numeric attributes, we fit a Gaussian (µ, σ) and add feature fgau i (yt−1, yt, x, t) := 1[|xt−µ|<iσ] for parameters i. Lexicons For non-numeric attributes, and for a lexicon l, i.e. a set of related words, we add feature flex l (yt−1, yt, x, t) := 1[xt∈l]. Lexicons are explained in the following section. 4 Extraction with Lexicons It is often possible to group words that are likely to be assigned similar labels, even if many of these words do not appear in our training set. The obtained lexicons then provide an elegant way to improve the generalization ability of an extractor, especially when only little training data is available. However, there is a danger of overfitting, which we discuss in Section 4.2.4. The next section explains how we mine the Web to obtain a large corpus of quality lists. Then Section 4.2 presents our semi-supervised algorithm for learning semantic lexicons from these lists. 288 4.1 Harvesting Lists from the Web Domain-independence requires access to an extremely large number of lists, but our tight integration of lexicon acquisition and CRF learning requires that relevant lists be accessed instantaneously. Approaches using search engines or wrappers at query time (Etzioni et al., 2004; Wang and Cohen, 2008) are too slow; we must extract and index lists prior to learning. We begin with a 5 billion page Web crawl. LUCHS can be combined with any list harvesting technique, but we choose a simple approach, extracting lists defined by HTML <ul> or <ol> tags. The set of lists obtained in this way is extremely noisy — many lists comprise navigation bars, tag sets, spam links, or a series of long text paragraphs. This is consistent with the observation that less than 2% of Web tables are relational (Cafarella et al., 2008). We therefore apply a series of filtering steps. We remove lists of only one or two items, lists containing long phrases, and duplicate lists from the same host. After filtering we obtain 49 million lists, containing 56 million unique phrases. 4.2 Semi-Supervised Learning of Lexicons While training a CRF extractor for a given relation, LUCHS uses its corpus of lists to automatically generate a set of semantic lexicons — specific to that relation. The technique proceeds in three steps, which have been engineered to run extremely quickly: 1. Seed phrases are extracted from the labeled training set. 2. A learning algorithm expands the seed phrases into a set of lexicons. 3. The semantic lexicons are added as features to the CRF learning algorithm. 4.2.1 Extracting Seed Phrases For each training sentence LUCHS first identifies subsequences of labeled words, and for each such labeled subsequence, LUCHS creates one or more seed phrases p. Typically, a set of seeds consists precisely of the labeled subsequences. However, if the labeled subsequences are long and have substructure, e.g., ‘San Remo, Italy’, our system splits at the separator token, and creates additional seed sets from prefixes and postfixes. 4.2.2 From Seeds to Lexicons To expand a set of seeds into a lexicon, LUCHS must identify relevant lists in the corpus. Relevancy can be computed by defining a similarity between lists using the vector-space model. Specifically, let L denote the corpus of lists, and P be the set of unique phrases from L. Each list l0 ∈L can be represented as a vector of weighted phrases p ∈ P appearing on the list, l0 = (l0 p1l0 p2 . . . l0 p|P|). Following the notion of inverse document frequency, a phrase’s weight is inversely proportional to the number of lists containing the phrase. Popular phrases which appear on many lists thus receive a small weight, whereas rare phrases are weighted higher: l0 pi = 1 |{l ∈L|p ∈l}| Unlike the vector space model for documents, we ignore term frequency, since the vast majority of lists in our corpus don’t contain duplicates. This vector representation supports the simple cosine definition of list similarity, which for lists l0, l1 ∈ L is defined as simcos := l0 · l1 ∥l0∥∥l1∥. Intuitively, two lists are similar if they have many overlapping phrases, the phrases are not too common, and the lists don’t contain many other phrases. By representing the seed set as another vector, we can find similar lists, hopefully containing related phrases. We then create a semantic lexicon by collecting phrases from a range of related lists. For example, one lexicon may be created as the union of all phrases on lists that have non-zero similarity to the seed list. Unfortunately, due to the noisy nature of the Web lists such a lexicon may be very large and may contain many irrelevant phrases. We expect that lists with higher similarity are more likely to contain phrases which are related to our seeds; hence, by varying the similarity threshold one may produce lexicons representing different compromises between lexicon precision and recall. Not knowing which lexicon will be most useful to the extractors, LUCHS generates several and lets the extractors learn appropriate weights. However, since list similarities vary depending on the seeds, fixed thresholds are not an option. If #similarlists denotes the number of lists that have non-zero similarity to the seed list and #lexicons 289 the total number of lexicons we want to generate, LUCHS sets lexicon i ∈{0, . . . , #lexicons −1} to be the union of prases on the #similarlistsi/#lexicons most similar lists.2 4.2.3 Efficiently Creating Lexicons We create lexicons from lists that are similar to our seed vector, so we only consider lists that have at least one phrase in common. Importantly, our index structures allow LUCHS to select the relevant lists efficiently. For each seed, LUCHS retrieves the set of containing lists as a sorted sequence of list identifiers. These sequences are then merged yielding a sequence of list identifiers with associated seed-hit counts. Precomputed list lengths and inverse document frequencies are also retrieved from indices, allowing efficient computation of similarity. The worst case complexity is O(log(S)SK) where S is the number of seeds and K the maximum number of lists to consider per seed. 4.2.4 Preventing Lexicon Overfitting Finally, we integrate the acquired semantic lexicons as features into the CRF. Although Section 3 discussed how to use lexicons as CRF features, there are some subtleties. Recall that the lexicons were created from seeds extracted from the training set. If we now train the CRF on the same examples that generated the lexicon features, then the CRF will likely overfit, and weight the lexicon features too highly! Before training, we therefore split the training set into k partitions. For each example in a partition we assign features based on lexicons generated from only the k−1 remaining partitions. This avoids overfitting and ensures that we will not perform much worse than without lexicon features. When we apply the CRF to our test set, we use the lexicons based on all k partitions. We refer to this technique as cross-training. 5 Experiments We start by evaluating end-to-end performance of LUCHS when applied to Wikipedia text, then analyze the characteristics of its components. Our experiments use the 10/2008 English Wikipedia dump. 2For practical reasons, we exclude the case i = #lexicons in our experiments. 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 recall precision Figure 2: Precision / recall curve for end-to-end system performance on 100 random articles. 5.1 Overall Extraction Performance To evaluate the end-to-end performance of LUCHS, we test the pipeline which first classifies incoming pages, activating a small set of extractors on the text. To ensure adequate training and test data, we limit ourselves to infobox classes with at least ten instances; there exist 1,583 such classes, together comprising 981,387 articles. We only consider the first ten sentences for each article, and we only consider 5025 attributes.3 We create a test set by sampling 100 articles randomly; these articles are not used to train article classifiers or extractors. Each test article is then automatically classified, and a random attribute of the predicted schema is selected for extraction. Gold labels for the selected attribute and article are created manually by a human judge and compared to the token-level predictions from the extractors which are trainined on the remaining articles with heuristic matches. Overall, LUCHS reaches a precision of .55 at a recall of .68, giving an F1-score of .61 (Figure 2). Analyzing the errors in more detail, we find that in 11 of 100 cases an article was incorrectly classified. We note that in at least two of these cases the predicted class could also be considered correct. For example, instead of Infobox Minor Planet the extractor predicted Infobox Planet. On five of the selected attributes the extractor failed because the attributes could be considered unlearnable: The flexibility of Wikipedia’s infobox system allows contributors to introduce attributes for formatting, for example defining el3Attributes were selected to have at least 10 heuristic matches, to have 10% of values covered by matches, and 10% of articles with attribute in infobox covered by matches. 290 ement order. In the future we wish to train LUCHS to ignore this type of attribute. We also compared the heuristic matches contained in the selected 100 articles to the gold standard: The matches reach a precision of .90 at a recall of .33, giving an F1-score of .48. So while most heuristic matches hit mentions of attribute values, many other mentions go unmatched. Manual analysis shows that these values are often missing from an infobox, are formatted differently, or are inconsistent to what is stated in the article. So why did the low recall of the heuristic matches not adversely affect recall of our extractors? For most articles, an attribute can be assigned a single unique value. When training an attribute extractor, only articles that contained a heuristic match for that attribute were considered, thus avoiding many cases of unmatched mentions. Subsequent experiments evaluate the performance of LUCHS components in more detail. 5.2 Article Classification The first step in LUCHS’s run-time pipeline is determining which infobox schemata are most likely to be found in a given article. To test this, we randomly split our 981,387 articles into 4/5 for training and 1/5 for testing, and train a single multiclass classifier. For this experiment, we use the original infobox class of an article as its gold label. We compute the accuracy of the prediction at .92. Since some classes can be considered interchangeable, this number represents a lower bound on performance. 5.3 Factors Affecting Extraction Accuracy We now evaluate attribute extraction assuming perfect article classification. To keep training time manageable, we sample 100 articles for training and 100 articles for testing4 for each of 100 random attributes. We again only consider the first ten sentences of each article, and we only consider articles that have heuristic matches with the attribute. We measure F1-score at a token-level, taking the heuristic matches as ground-truth. We first test the performance of extractors trained using our basic features (Section 3)5, not including lexicons and Gaussians. We begin using word features and obtain a token-level F1score of .311 for text and .311 for numeric attributes. Adding any of our additional features 4These numbers are smaller for attributes with less training data available, but the same split is maintained. 5For contextualization features we choose p, s = 5. Features F1-Score Text attributes Baseline .491 Baseline + Lexicons w/o CT .367 Baseline + Lexicons .545 Numeric attributes Baseline .586 Baseline + Gaussians w/o CT .623 Baseline + Gaussians .627 Table 1: Impact of Lexicon and Gaussian features. Cross-Training (CT) is essential to improve performance. improves these scores, but the relative improvements vary: For both text and numeric attributes, contextualization and dependency features deliver the largest improvement. We then iteratively add the feature with largest improvement until no further improvement is observed. We finally obtain an F1-score of .491 for text and .586 for numeric attributes. For text attributes the extractor uses word, contextualization, first sentence, capitalization, and digit features; for numeric attributes the extractor uses word, contextualization, digit, first sentence, and dependency features. We use these extractors as a baseline to evaluate our lexicon and Gaussian features. Varying the size of the training sets affects results: Taking more articles raises the F1-score, but taking more sentences per article reduces it. This is because Wikipedia articles often summarize a topic in the first few paragraphs and later discuss related topics, necessitating reference resolution which we plan to add in future work. 5.4 Lexicon and Gaussian Features We next study how our distribution features6 impact the quality of the baseline extractors (Table 1). Without cross-training we observe a reduction in performance, due to overfitting. Cross-training avoids this, and substantially improves results over the baseline. While cross-training is particularly critical for lexicon features, it is less needed for Gaussians where only two parameters, mean and deviation, are fitted to the training set. The relative improvements depend on the number of available training examples (Table 2). Lexicon and Gaussian features especially benefit extractors for sparse attributes. Here we can also see that the improvements are mainly due to increases in recall. 6We set the number of lexicon and Gaussian features to 4. 291 # Train F1-B F1-LUCHS ∆F1 ∆Pr ∆Re Text attributes 10 .379 .439 +16% +10% +20% 25 .447 .504 +13% +7% +20% 100 .491 .545 +11% +5% +17% Numeric attributes 10 .484 .531 +10% +4% +13% 25 .552 .596 +8% +4% +10% 100 .586 .627 +7% +5% +8% Table 2: Lexicon and Gaussian features greatly expand F1 score (F1-LUCHS) over the baseline (F1B), in particular for attributes with few training examples. Gains are mainly due to increased recall. 5.5 Scaling to All of Wikipedia Finally, we take our best extractors and run them on all 5025 attributes, again assuming perfect article classification and using heuristic matches as gold-standard. Figure 3 shows the distribution of obtained F1 scores. 810 text attributes and 328 numeric attributes reach a score of 0.80 or higher. The performance depends on the number of available training examples, and that number is governed by a long-tailed distribution. For example, 61% of the attributes in our set have 50 or fewer examples, 36% have 20 or fewer. Interestingly, the number of training examples had a smaller effect on performance than expected. Figure 4 shows the correlation between these variables. Lexicon and Gaussian features enables acceptable performance even for sparse attributes. Averaging across all attributes we obtain F1 scores of 0.56 and 0.60 for textual and numeric values respectively. We note that these scores assume that all attributes are equally important, weighting rare attributes just like common ones. If we weight scores by the number of attribute instances, we obtain F1 scores of 0.64 (textual) and 0.78 (numeric). In each case, precision is slightly higher than recall. 5.6 Towards an Attribute Ontology The true promise of relation-specific extractors comes when an ontology ties the system together. By learning a probabilistic model of selectional preferences, one can use joint inference to improve extraction accuracy. One can also answer scientific questions, such as “How many of the learned Wikipedia attributes are distinct?” It is clear that many duplicates exist due to collaborative sloppiness, but semantic similarity is a matter of opinion and an exact answer is impossible. 0% 20% 40% 60% 80% 100% 0.0 0.2 0.4 0.6 0.8 1.0 Text attr. (3962) Numeric attr. (1063) # Attributes F1 Score Figure 3: F1 scores among attributes, ranked by score. 810 text attributes (20%) and 328 numeric attributes (31%) had an F1-score of .80 or higher. 0 20 40 60 80 100 0.0 0.2 0.4 0.6 0.8 Text attr. Numeric attr. # Training Examples Average F1 Score Figure 4: Average F1 score by number of training examples. While more training data helps, even sparse attributes reach acceptable performance. Nevertheless, we clustered the textual attributes in several ways. First, we cleaned the attribute names heuristically and performed spell check. The “distance” between two attributes was calculated with a combination of edit distance and IR metrics with Wordnet synonyms; then hierarchical agglomerative clustering was performed. We manually assigned names to the clusters and cleaned them, splitting and joining as needed. The result is too crude to be called an ontology, but we continue its elaboration. There are a total of 3962 attributes grouped in about 1282 clusters (not yet counting attributes with numerical values); the largest cluster, location, has 115 similar attributes. Figure 5 shows the confusion matrix between attributes in the biggest clusters; the shade of the i, jth pixel indicates the F1 score achieved by training on instances of attribute i and testing on attribute j. 292 location birthplace p title country full name city nationality nationality birth name date of birth date of death date states Figure 5: Confusion matrix for extractor accuracy training on one attribute then testing on another. Note the extraction similarity between title and full-name, as well as between dates of birth and death. Space constraints allow us to show only 1000 of LUCHS’s 5025 extracted attributes, those in the largest clusters. 6 Related Work Large-scale extraction A popular approach to IE is supervised learning of relation-specific extractors (Freitag, 1998). Open IE, self-supervised learning of unlexicalized, relation-independent extractors (Banko et al., 2007), is a more scalable approach, but suffers from lower precision and recall, and doesn’t canonicalize the relations. A third approach, weak supervision, performs selfsupervised learning of relation-specific extractors from noisy training data, heuristically generated by matching database values to text. (Craven and Kumlien, 1999; Hirschman et al., 2002) apply this technique to the biological domain, and (Mintz et al., 2009) apply it to 102 relations from Freebase. LUCHS differs from these approaches in that its “database” – the set of infobox values – itself is noisy, contains many more relations, and has few instances per relation. Whereas the existing approaches focus on syntactic extraction patterns, LUCHS focuses on lexical information enhanced by dynamic lexicon learning. Extraction from Wikipedia Wikipedia has become an interesting target for extraction. (Suchanek et al., 2008) build a knowledgebase from Wikipedia’s semi-structured data. (Wang et al., 2007) propose a semisupervised positive-only learning technique. Although that extracts from text, its reliance on hyperlinks and other semistructured data limits extraction. (Wu and Weld, 2007; Wu et al., 2008)’s systems generate training data similar to LUCHS, but were only on a few infobox classes. In contrast, LUCHS shows that the idea scales to more than 5000 relations, but that additional techniques, such as dynamic lexicon learning, are necessary to deal with sparsity. Extraction with lexicons While lexicons have been commonly used for IE (Cohen and Sarawagi, 2004; Agichtein and Ganti, 2004; Bellare and McCallum, 2007), many approaches assume that lexicons are clean and are supplied by a user before training. Other approaches (Talukdar et al., 2006; Miller et al., 2004; Riloff, 1993) learn lexicons automatically from distributional patterns in text. (Wang et al., 2009) learns lexicons from Web lists for query tagging. LUCHS differs from these approaches in that it is not limited to a small set of well-defined relations. Rather than creating large lexicons of common entities, LUCHS attempts to efficiently instantiate a series of lexicons from a small set of seeds to bias extractors of sparse attributes. Crucual to LUCHS’s different setting is also the need to avoid overfitting. Set expansion A large amount of work has looked at automatically generating sets of related items. Starting with a set of seed terms, (Etzioni et al., 2004) extract lists by learning wrappers for Web pages containing those terms. (Wang and Cohen, 2007; Wang and Cohen, 2008) extend the idea, computing term relatedness through a random walk algorithm that takes into account seeds, documents, wrappers and mentions. Other approaches include Bayesian methods (Ghahramani and Heller, 2005) and graph label propagation algorithms (Talukdar et al., 2008; Bengio et al., 2006). The goal of set expansion techniques is to generate high precision sets of related items; hence, these techniques are evaluated based on lexicon precision and recall. For LUCHS, which is evaluated based on the quality of an extractor using the lexicons, lexicon precision is not important – as long as it does not confuse the extractor. 7 Future Work We envision a Web-scale machine reading system which simultaneously learns ontologies and extractors, and we believe that LUCHS’s approach of leveraging noisy semi-structured information (such as lists or formatting templates) is a key towards this goal. For future work, we plan to enhance LUCHS in two major ways. First, we note that a big weakness is that the system currently only works for Wikipedia pages. 293 For example, LUCHS assumes that each page corresponds to exactly one schema and that the subject of relations on a page are the same. Also, LUCHS makes predictions on a token basis, thus sometimes failing to recognize larger segments. To remove these limitations we plan to add a deeper linguistic analysis, making better use of parse and dependency information and including coreference resolution. We also plan to employ relation-independent Open extraction techniques, e.g. as suggested in (Wu and Weld, 2008) (retraining). Second, we note that LUCHS’s performance may benefit substantially from an attribute ontology. As we showed in Section 5.6, LUCHS’s current extractors can also greatly facilitate learning a full attribute ontology. We therefore plan to interleave extractor learning and ontology inference, hence jointly learning ontology and extractors. 8 Conclusion Many researchers are trying to use IE to create large-scale knowledge bases from natural language text on the Web, but existing relationspecific techniques do not scale to the thousands of relations encoded in Web text – while relationindependent techniques suffer from lower precision and recall, and do not canonicalize the relations. This paper shows that – with new techniques – self-supervised learning of relation-specific extractors from Wikipedia infoboxes does scale. In particular, we present LUCHS, a selfsupervised IE system capable of learning more than an order of magnitude more relation-specific extractors than previous systems. LUCHS uses dynamic lexicon features that enable hyperlexicalized extractors which cope effectively with sparse training data. We show an overall performance of 61% F1 score, and present experiments evaluating LUCHS’s individual components. Datasets generated in this work are available to the community7. Acknowledgments We thank Jesse Davis, Oren Etzioni, Andrey Kolobov, Mausam, Fei Wu, and the anonymous reviewers for helpful comments and suggestions. This material is based upon work supported by a WRF / TJ Cable Professorship, a gift from Google and by the Air Force Research Laboratory (AFRL) under prime contract no. FA8750-09-C-0181. Any opinions, findings, and conclusion or recommendations expressed in this material are those of 7http://www.cs.washington.edu/ai/iwp the author(s) and do not necessarily reflect the view of the Air Force Research Laboratory (AFRL). References Eugene Agichtein and Venkatesh Ganti. 2004. Mining reference tables for automatic text segmentation. In Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD-2004), pages 20–29. S¨oren Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyganiak, and Zachary G. Ives. 2007. Dbpedia: A nucleus for a web of open data. In Proceedings of the 6th International Semantic Web Conference and 2nd Asian Semantic Web Conference (ISWC/ASWC2007), pages 722–735. Michele Banko, Michael J. Cafarella, Stephen Soderland, Matthew Broadhead, and Oren Etzioni. 2007. Open information extraction from the web. In Proceedings of the 20th International Joint Conference on Artificial Intelligence (IJCAI-2007), pages 2670–2676. Kedar Bellare and Andrew McCallum. 2007. Learning extractors from unlabeled text using relevant databases. In Sixth International Workshop on Information Integration on the Web. Yoshua Bengio, Olivier Delalleau, and Nicolas Le Roux. 2006. Label propagation and quadratic criterion. In Olivier Chapelle, Bernhard Sch¨olkopf, and Alexander Zien, editors, Semi-Supervised Learning, pages 193–216. MIT Press. Michael J. Cafarella, Alon Y. Halevy, Daisy Zhe Wang, Eugene Wu, and Yang Zhang. 2008. Webtables: exploring the power of tables on the web. Proceedings of the International Conference on Very Large Databases (VLDB2008), 1(1):538–549. Andrew Carlson, Justin Betteridge, Estevam R. Hruschka Jr., and Tom M. Mitchell. 2009a. Coupling semi-supervised learning of categories and relations. In NAACL HLT 2009 Workskop on Semi-supervised Learning for Natural Language Processing. Andrew Carlson, Scott Gaffney, and Flavian Vasile. 2009b. Learning a named entity tagger from gazetteers with the partial perceptron. In AAAI Spring Symposium on Learning by Reading and Learning to Read. William W. Cohen and Sunita Sarawagi. 2004. Exploiting dictionaries in named entity extraction: combining semimarkov extraction processes and data integration methods. In Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD-2004), pages 89–98. Michael Collins. 2002. Discriminative training methods for hidden markov models: Theory and experiments with perceptron algorithms. In Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing (EMNLP-2002). Mark Craven and Johan Kumlien. 1999. Constructing biological knowledge bases by extracting information from text sources. In Proceedings of the Seventh International Conference on Intelligent Systems for Molecular Biology (ISMB-1999), pages 77–86. 294 Benjamin Van Durme and Marius Pasca. 2008. Finding cars, goddesses and enzymes: Parametrizable acquisition of labeled instances for open-domain information extraction. In Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence (AAAI-2008), pages 1243–1248. Oren Etzioni, Michael J. Cafarella, Doug Downey, AnaMaria Popescu, Tal Shaked, Stephen Soderland, Daniel S. Weld, and Alexander Yates. 2004. Methods for domainindependent information extraction from the web: An experimental comparison. In Proceedings of the Nineteenth National Conference on Artificial Intelligence (AAAI2004), pages 391–398. Dayne Freitag. 1998. Toward general-purpose learning for information extraction. In Proceedings of the 17th international conference on Computational linguistics, pages 404–408. Association for Computational Linguistics. Yoav Freund and Robert E. Schapire. 1999. Large margin classification using the perceptron algorithm. Machine Learning, 37(3):277–296. Zoubin Ghahramani and Katherine A. Heller. 2005. Bayesian sets. In Neural Information Processing Systems (NIPS-2005). Lynette Hirschman, Alexander A. Morgan, and Alexander S. Yeh. 2002. Rutabaga by any other name: extracting biological names. Journal of Biomedical Informatics, 35(4):247–259. John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning (ICML-2001), pages 282–289. Marie-Catherine De Marneffe, Bill Maccartney, and Christopher D. Manning. 2006. Generating typed dependency parses from phrase structure parses. In Proceedings of the fifth international conference on Language Resources and Evaluation (LREC-2006). Scott Miller, Jethran Guinness, and Alex Zamanian. 2004. Name tagging with word clusters and discriminative training. In Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics (HLT-NAACL-2004). Mike Mintz, Steven Bills, Rion Snow, and Daniel Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In The Annual Meeting of the Association for Computational Linguistics (ACL-2009). Marius Pasca. 2009. Outclassing wikipedia in open-domain information extraction: Weakly-supervised acquisition of attributes over conceptual hierarchies. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics (EACL-2009), pages 639–647. Ellen Riloff. 1993. Automatically constructing a dictionary for information extraction tasks. In Proceedings of the 11th National Conference on Artificial Intelligence (AAAI1993), pages 811–816. Fabian M. Suchanek, Gjergji Kasneci, and Gerhard Weikum. 2008. Yago: A large ontology from wikipedia and wordnet. Elsevier Journal of Web Semantics, 6(3):203–217. Fabian M. Suchanek, Mauro Sozio, and Gerhard Weikum. 2009. Sofie: A self-organizing framework for information extraction. In Proceedings of the 18th International Conference on World Wide Web (WWW-2009). Partha Pratim Talukdar, Thorsten Brants, Mark Liberman, and Fernando Pereira. 2006. A context pattern induction method for named entity extraction. In The Tenth Conference on Natural Language Learning (CoNLL-X-2006). Partha Pratim Talukdar, Joseph Reisinger, Marius Pasca, Deepak Ravichandran, Rahul Bhagat, and Fernando Pereira. 2008. Weakly-supervised acquisition of labeled class instances using graph random walks. In EMNLP, pages 582–590. Richard C. Wang and William W. Cohen. 2007. Languageindependent set expansion of named entities using the web. In Proceedings of the 7th IEEE International Conference on Data Mining (ICDM-2007), pages 342–350. Richard C. Wang and William W. Cohen. 2008. Iterative set expansion of named entities using the web. In Proceedings of the 8th IEEE International Conference on Data Mining (ICDM-2008). Gang Wang, Yong Yu, and Haiping Zhu. 2007. Pore: Positive-only relation extraction from wikipedia text. In Proceedings of the 6th International Semantic Web Conference and 2nd Asian Semantic Web Conference (ISWC/ASWC-2007), pages 580–594. Ye-Yi Wang, Raphael Hoffmann, Xiao Li, and Alex Acero. 2009. Semi-supervised acquisition of semantic classes – from the web and for the web. In International Conference on Information and Knowledge Management (CIKM2009), pages 37–46. Fei Wu and Daniel S. Weld. 2007. Autonomously semantifying wikipedia. In Proceedings of the International Conference on Information and Knowledge Management (CIKM-2007), pages 41–50. Fei Wu and Daniel S. Weld. 2008. Automatically refining the wikipedia infobox ontology. In Proceedings of the 17th International Conference on World Wide Web (WWW-2008), pages 635–644. Fei Wu and Daniel S. Weld. 2010. Open information extraction using wikipedia. In The Annual Meeting of the Association for Computational Linguistics (ACL-2010). Fei Wu, Raphael Hoffmann, and Daniel S. Weld. 2008. Information extraction from wikipedia: moving down the long tail. In Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD-2008), pages 731–739. 295
|
2010
|
30
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 296–305, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Unsupervised Ontology Induction from Text Hoifung Poon and Pedro Domingos Department of Computer Science & Engineering University of Washington hoifung,[email protected] Abstract Extracting knowledge from unstructured text is a long-standing goal of NLP. Although learning approaches to many of its subtasks have been developed (e.g., parsing, taxonomy induction, information extraction), all end-to-end solutions to date require heavy supervision and/or manual engineering, limiting their scope and scalability. We present OntoUSP, a system that induces and populates a probabilistic ontology using only dependency-parsed text as input. OntoUSP builds on the USP unsupervised semantic parser by jointly forming ISA and IS-PART hierarchies of lambda-form clusters. The ISA hierarchy allows more general knowledge to be learned, and the use of smoothing for parameter estimation. We evaluate OntoUSP by using it to extract a knowledge base from biomedical abstracts and answer questions. OntoUSP improves on the recall of USP by 47% and greatly outperforms previous state-of-the-art approaches. 1 Introduction Knowledge acquisition has been a major goal of NLP since its early days. We would like computers to be able to read text and express the knowledge it contains in a formal representation, suitable for answering questions and solving problems. However, progress has been difficult. The earliest approaches were manual, but the sheer amount of coding and knowledge engineering needed makes them very costly and limits them to well-circumscribed domains. More recently, machine learning approaches to a number of key subproblems have been developed (e.g., Snow et al. (2006)), but to date there is no sufficiently automatic end-to-end solution. Most saliently, supervised learning requires labeled data, which itself is costly and infeasible for large-scale, open-domain knowledge acquisition. Ideally, we would like to have an end-to-end unsupervised (or lightly supervised) solution to the problem of knowledge acquisition from text. The TextRunner system (Banko et al., 2007) can extract a large number of ground atoms from the Web using only a small number of seed patterns as guidance, but it is unable to extract non-atomic formulas, and the mass of facts it extracts is unstructured and very noisy. The USP system (Poon and Domingos, 2009) can extract formulas and appears to be fairly robust to noise. However, it is still limited to extractions for which there is substantial evidence in the corpus, and in most corpora most pieces of knowledge are stated only once or a few times, making them very difficult to extract without supervision. Also, the knowledge extracted is simply a large set of formulas without ontological structure, and the latter is essential for compact representation and efficient reasoning (Staab and Studer, 2004). We propose OntoUSP (Ontological USP), a system that learns an ISA hierarchy over clusters of logical expressions, and populates it by translating sentences to logical form. OntoUSP is encoded in a few formulas of higher-order Markov logic (Domingos and Lowd, 2009), and can be viewed as extending USP with the capability to perform hierarchical (as opposed to flat) clustering. This clustering is then used to perform hierarchical smoothing (a.k.a. shrinkage), greatly increasing the system’s capability to generalize from 296 sparse data. We begin by reviewing the necessary background. We then present the OntoUSP Markov logic network and the inference and learning algorithms used with it. Finally, experiments on a biomedical knowledge acquisition and question answering task show that OntoUSP can greatly outperform USP and previous systems. 2 Background 2.1 Ontology Learning In general, ontology induction (constructing an ontology) and ontology population (mapping textual expressions to concepts and relations in the ontology) remain difficult open problems (Staab and Studer, 2004). Recently, ontology learning has attracted increasing interest in both NLP and semantic Web communities (Cimiano, 2006; Maedche, 2002), and a number of machine learning approaches have been developed (e.g., Snow et al. (2006), Cimiano (2006), Suchanek et al. (2008,2009), Wu & Weld (2008)). However, they are still limited in several aspects. Most approaches induce and populate a deterministic ontology, which does not capture the inherent uncertainty among the entities and relations. Besides, many of them either bootstrap from heuristic patterns (e.g., Hearst patterns (Hearst, 1992)) or build on existing structured or semi-structured knowledge bases (e.g., WordNet (Fellbaum, 1998) and Wikipedia1), thus are limited in coverage. Moreover, they often focus on inducing ontology over individual words rather than arbitrarily large meaning units (e.g., idioms, phrasal verbs, etc.). Most importantly, existing approaches typically separate ontology induction from population and knowledge extraction, and pursue each task in a standalone fashion. While computationally efficient, this is suboptimal. The resulted ontology is disconnected from text and requires additional effort to map between the two (Tsujii, 2004). In addition, this fails to leverage the intimate connections between the three tasks for joint inference and mutual disambiguiation. Our approach differs from existing ones in two main aspects: we induce a probabilistic ontology from text, and we do so by jointly conducting ontology induction, population, and knowledge extraction. Probabilistic modeling handles uncertainty and noise. A joint approach propagates in1http : //www.wikipedia.org formation among the three tasks, uncovers more implicit information from text, and can potentially work well even in domains not well covered by existing resources like WordNet and Wikipedia. Furthermore, we leverage the ontology for hierarchical smoothing and incorporate this smoothing into the induction process. This facilitates more accurate parameter estimation and better generalization. Our approach can also leverage existing ontologies and knowledge bases to conduct semisupervised ontology induction (e.g., by incorporating existing structures as hard constraints or penalizing deviation from them). 2.2 Markov Logic Combining uncertainty handling and joint inference is the hallmark of the emerging field of statistical relational learning (a.k.a. structured prediction), where a plethora of approaches have been developed (Getoor and Taskar, 2007; Bakir et al., 2007). In this paper, we use Markov logic (Domingos and Lowd, 2009), which is the leading unifying framework, but other approaches can be used as well. Markov logic is a probabilistic extension of first-order logic and can compactly specify probability distributions over complex relational domains. It has been successfully applied to unsupervised learning for various NLP tasks such as coreference resolution (Poon and Domingos, 2008) and semantic parsing (Poon and Domingos, 2009). A Markov logic network (MLN) is a set of weighted first-order clauses. Together with a set of constants, it defines a Markov network with one node per ground atom and one feature per ground clause. The weight of a feature is the weight of the first-order clause that originated it. The probability of a state x in such a network is given by the log-linear model P(x) = 1 Z exp (P i wini(x)), where Z is a normalization constant, wi is the weight of the ith formula, and ni is the number of satisfied groundings. 2.3 Unsupervised Semantic Parsing Semantic parsing aims to obtain a complete canonical meaning representation for input sentences. It can be viewed as a structured prediction problem, where a semantic parse is formed by partitioning the input sentence (or a syntactic analysis such as a dependency tree) into meaning units and assigning each unit to the logical form representing an entity or relation (Figure 1). In effect, a semantic 297 induces protein CD11b nsubj dobj IL-4 nn induces protein CD11b nsubj dobj IL-4 nn INDUCE INDUCER INDUCED IL-4 CD11B INDUCE(e1) INDUCER(e1,e2) INDUCED(e1,e3) IL-4(e2) CD11B(e3) IL-4 protein induces CD11b Structured prediction: Partition + Assignment Figure 1: An example of semantic parsing. Top: semantic parsing converts an input sentence into logical form in Davidsonian semantics. Bottom: a semantic parse consists of a partition of the dependency tree and an assignment of its parts. parser extracts knowledge from input text and converts them into logical form (the semantic parse), which can then be used in logical and probabilistic inference and support end tasks such as question answering. A major challenge to semantic parsing is syntactic and lexical variations of the same meaning, which abound in natural languages. For example, the fact that IL-4 protein induces CD11b can be expressed in a variety of ways, such as, “Interleukin-4 enhances the expression of CD11b”, “CD11b is upregulated by IL-4”, etc. Past approaches either manually construct a grammar or require example sentences with meaning annotation, and do not scale beyond restricted domains. Recently, we developed the USP system (Poon and Domingos, 2009), the first unsupervised approach for semantic parsing.2 USP inputs dependency trees of sentences and first transforms them into quasi-logical forms (QLFs) by converting each node to a unary atom and each dependency edge to a binary atom (e.g., the node for “induces” becomes induces(e1) and the subject dependency becomes nsubj(e1, e2), where ei’s are Skolem constants indexed by the nodes.).3 For each sentence, a semantic parse comprises of a partition of its QLF into subexpressions, each of which has a naturally corresponding lambda 2In this paper, we use a slightly different formulation of USP and its MLN to facilitate the exposition of OntoUSP. 3We call these QLFs because they are not true logical form (the ambiguities are not yet resolved). This is related to but not identical with the definition in Alshawi (1990). Object Cluster: INDUCE induces 0.1 enhances 0.4 … …… Property Cluster: INDUCER 0.5 0.4 … IL-4 0.2 IL-8 0.1 … None 0.1 One 0.8 … nsubj agent Core Form Figure 2: An example of object/property clusters: INDUCE contains the core-form property cluster and others, such as the agent argument INDUCER. form,4 and an assignment of each subexpression to a lambda-form cluster. The lambda-form clusters naturally form an ISPART hierarchy (Figure 2). An object cluster corresponds to semantic concepts or relations such as INDUCE, and contains a variable number of property clusters. A special property cluster of core forms maintains a distribution over variations in lambda forms for expressing this concept or relation. Other property clusters correspond to modifiers or arguments such as INDUCER (the agent argument of INDUCE), each of which in turn contains three subclusters of property values: the argument-object subcluster maintains a distribution over object clusters that may occur in this argument (e.g., IL −4), the argument-form subcluster maintains a distribution over lambda forms that corresponds to syntactic variations for this argument (e.g., nsubj in active voice and agent in passive voice), and the argument-number subcluster maintains a distribution over total numbers of this argument that may occur in a sentence (e.g., zero if the argument is not mentioned). Effectively, USP simultaneously discovers the lambda-form clusters and an IS-PART hierarchy among them. It does so by recursively combining subexpressions that are composed with or by similar subexpressions. The partition breaks a sentence into subexpressions that are meaning units, and the clustering abstracts away syntactic and lexical variations for the same meaning. This novel form of relational clustering is governed by a joint probability distribution P(T, L) defined in higher-order5 Markov logic, where T are the input dependency trees, and L the semantic parses. The 4The lambda form is derived by replacing every Skolem constant ei that does not appear in any unary atom in the subexpression with a lambda variable xi that is uniquely indexed by the corresponding node i. For example, the lambda form for nsubj(e1, e2) is λx1λx2.nsubj(x1, x2). 5Variables can range over arbitrary lambda forms. 298 main predicates are: e ∈c: expression e is assigned to cluster c; SubExpr(s, e): s is a subexpression of e; HasValue(s, v): s is of value v; IsPart(c, i, p): p is the property cluster in object cluster c uniquely indexed by i. In USP, property clusters in different object clusters use distinct index i’s. As we will see later, in OntoUSP, property clusters with ISA relation share the same index i, which corresponds to a generic semantic frame such as agent and patient. The probability model of USP can be captured by two formulas: x ∈+p ∧HasValue(x, +v) e ∈c ∧SubExpr(x, e) ∧x ∈p ⇒∃1i.IsPart(c, i, p). All free variables are implicitly universally quantified. The “+” notation signifies that the MLN contains an instance of the formula, with a separate weight, for each value combination of the variables with a plus sign. The first formula is the core of the model and represents the mixture of property values given the cluster. The second formula ensures that a property cluster must be a part in the corresponding object cluster; it is a hard constraint, as signified by the period at the end. To encourage clustering, USP imposes an exponential prior over the number of parameters. To parse a new sentence, USP starts by partitioning the QLF into atomic forms, and then hillclimbs on the probability using a search operator based on lambda reduction until it finds the maximum a posteriori (MAP) parse. During learning, USP starts with clusters of atomic forms, maintains the optimal semantic parses according to current parameters, and hill-climbs on the loglikelihood of observed QLFs using two search operators: MERGE(c1, c2) merges clusters c1, c2 into a larger cluster c by merging the core-form clusters and argument clusters of c1, c2, respectively. E.g., c1 = {“induce”}, c2 = {“enhance”}, and c = {“induce”, “enhance”}. COMPOSE(c1, c2) creates a new lambda-form cluster c formed by composing the lambda forms in c1, c2 into larger ones. E.g., c1 = {“amino”}, c2 = {“acid”}, and c = {“amino acid”}. Each time, USP executes the highest-scored operator and reparses affected sentences using the new parameters. The output contains the optimal lambda-form clusters and parameters, as well as the MAP semantic parses of input sentences. 3 Unsupervised Ontology Induction with Markov Logic A major limitation of USP is that it either merges two object clusters into one, or leaves them separate. This is suboptimal, because different object clusters may still possess substantial commonalities. Modeling these can help extract more general knowledge and answer many more questions. The best way to capture such commonalities is by forming an ISA hierarchy among the clusters. For example, INDUCE and INHIBIT are both subconcepts of REGULATE. Learning these ISA relations helps answer questions like “What regulates CD11b?”, when the text states that “IL-4 induces CD11b” or “AP-1 suppresses CD11b”. For parameter learning, this is also undesirable. Without the hierarchical structure, each cluster estimates its parameters solely based on its own observations, which can be extremely sparse. The better solution is to leverage the hierarchical structure for smoothing (a.k.a. shrinkage (McCallum et al., 1998; Gelman and Hill, 2006)). For example, if we learn that “super-induce” is a verb and that in general verbs have active and passive voices, then even though “super-induce” only shows up once in the corpus as in “AP-1 is super-induced by IL4”, by smoothing we can still infer that this probably means the same as “IL-4 super-induces AP-1”, which in turn helps answer questions like “What super-induces AP-1”. OntoUSP overcomes the limitations of USP by replacing the flat clustering process with a hierarchical clustering one, and learns an ISA hierarchy of lambda-form clusters in addition to the IS-PART one. The output of OntoUSP consists of an ontology, a semantic parser, and the MAP parses. In effect, OntoUSP conducts ontology induction, population, and knowledge extraction in a single integrated process. Specifically, given clusters c1, c2, in addition to merge vs. separate, OntoUSP evaluates a third option called abstraction, in which a new object cluster c is created, and ISA links are added from ci to c; the argument clusters in c are formed by merging that of ci’s. In the remainder of the section, we describe the 299 details of OntoUSP. We start by presenting the OntoUSP MLN. We then describe our inference algorithm and how to parse a new sentence using OntoUSP. Finally, we describe the learning algorithm and how OntoUSP induces the ontology while learning the semantic parser. 3.1 The OntoUSP MLN The OntoUSP MLN can be obtained by modifying the USP MLN with three simple changes. First, we introduce a new predicate IsA(c1, c2), which is true if cluster c1 is a subconcept of c2. For convenience, we stipulate that IsA is reflexive (i.e., IsA(c, c) is true for any c). Second, we add two formulas to the MLN: IsA(c1, c2) ∧IsA(c2, c3) ⇒IsA(c1, c3). IsPart(c1, i1, p1) ∧IsPart(c2, i2, p2) ∧IsA(c1, c2) ⇒(i1 = i2 ⇔IsA(p1, p2)). The first formula simply enforces the transitivity of ISA relation. The second formula states that if the ISA relation holds for a pair of object clusters, it also holds between their corresponding property clusters. Both are hard constraints. Third, we introduce hierarchical smoothing into the model by replacing the USP mixture formula x ∈+p ∧HasValue(x, +v) with a new formula ISA(p1, +p2) ∧x ∈p1 ∧HasValue(x, +v) Intuitively, for each p2, the weight corresponds to the delta in log-probability of v comparing to the prediction according to all ancestors of p2. The effect of this change is that now the value v of a subexpression x is not solely determined by its property cluster p1, but is also smoothed by statistics of all p2 that are super clusters of p1. Shrinkage takes place via interaction among the weights of the ISA mixture formula. In particular, if the weights for some property cluster p are all zero, it means that values in p are completely predicted by p’s ancestors. In effect, p is backed off to its parent. 3.2 Inference Given the dependency tree T of a sentence, the conditional probability of a semantic parse L is given by Pr(L|T) ∝ exp (P i wini(T, L)). The MAP semantic parse is simply Algorithm 1 OntoUSP-Parse(MLN, T) Initialize semantic parse L with individual atoms in the QLF of T repeat for all subexpressions e in L do Evaluate all semantic parses that are lambda-reducible from e end for L ←the new semantic parse with the highest gain in probability until none of these improve the probability return L arg maxL P i wini(T, L). Directly enumerating all L’s is intractable. OntoUSP uses the same inference algorithm as USP by hill-climbing on the probability of L; in each step, OntoUSP evaluates the alternative semantic parses that can be formed by lambda-reducing a current subexpression with one of its arguments. The only difference is that OntoUSP uses a different MLN and so the probabilities and resulting semantic parses may be different. Algorithm 1 gives pseudo-code for OntoUSP’s inference algorithm. 3.3 Learning OntoUSP uses the same learning objective as USP, i.e., to find parameters θ that maximizes the loglikelihood of observing the dependency trees T, summing out the unobserved semantic parses L: Lθ(T) = log Pθ(L) = log P L Pθ(T, L) However, the learning problem in OntoUSP is distinct in two important aspects. First, OntoUSP learns in addition an ISA hierarchy among the lambda-form clusters. Second and more importantly, OntoUSP leverages this hierarchy during learning to smooth the parameter estimation of individual clusters, as embodied by the new ISA mixture formula in the OntoUSP MLN. OntoUSP faces several new challenges unseen in previous hierarchical-smoothing approaches. The ISA hierarchy in OntoUSP is not known in advance, but needs to be learned as well. Similarly, OntoUSP has no known examples of populated facts and rules in the ontology, but has to infer that in the same joint learning process. Finally, OntoUSP does not start from well-formed structured input like relational tuples, but rather directly from raw text. In sum, OntoUSP tackles a 300 Algorithm 2 OntoUSP-Learn(MLN, T’s) Initialize with a flat ontology, along with clusters and semantic parses Merge clusters with the same core form Agenda ←∅ repeat for all candidate operations O do Score O by log-likelihood improvement if score is above a threshold then Add O to agenda end if end for Execute the highest scoring operation O∗in the agenda Regenerate MAP parses for affected trees and update agenda and candidate operations until agenda is empty return the learned ontology and MLN, and the semantic parses very hard problem with exceedingly little aid from user supervision. To combat these challenges, OntoUSP adopts a novel form of hierarchical smoothing by integrating it with the search process for identifying the hierarchy. Algorithm 2 gives pseudocode for OntoUSP’s learning algorithm. Like USP, OntoUSP approximates the sum over all semantic parses with the most probable parse, and searches for both θ and the MAP semantic parses L that maximize Pθ(T, L). In addition to MERGE and COMPOSE, OntoUSP uses a new operator ABSTRACT(c1, c2), which does the following: 1. Create an abstract cluster c; 2. Create ISA links from c1, c2 to c; 3. Align property clusters of c1 and c2; for each aligned pair p1 and p2, either merge them into a single property cluster, or create an abstract property cluster p in c and create ISA links from pi to p, so as to maximize loglikelihood. Intuitively, c corresponds to a more abstract concept that summarizes similar properties in ci’s. To add a child cluster c2 to an existing abstract cluster c1, OntoUSP also uses an operator ADDCHILD(c1, c2) that does the following: 1. Create an ISA link from c2 to c1; 2. For each property cluster of c2, maximize the log-likelihood by doing one of the following: merge it with a property cluster in an existing child of c1; create ISA link from it to an abstract property cluster in c; leave it unchanged. For efficiency, in both operators, the best option is chosen greedily for each property cluster in c2, in descending order of cluster size. Notice that once an abstract cluster is created, it could be merged with an existing cluster using MERGE. Thus with the new operators, OntoUSP is capable of inducing any ISA hierarchy among abstract and existing clusters. (Of course, the ISA hierarchy it actually induces depends on the data.) Learning the shrinkage weights has been approached in a variety of ways; examples include EM and cross-validation (McCallum et al., 1998), hierarchical Bayesian methods (Gelman and Hill, 2006), and maximum entropy with L1 priors (Dudik et al., 2007). The past methods either only learn parameters with one or two levels (e.g., in hierarchical Bayes), or requires significant amount of computation (e.g., in EM and in L1-regularized maxent), while also typically assuming a given hierarchy. In contrast, OntoUSP has to both induce the hierarchy and populate it, with potentially many levels in the induced hierarchy, starting from raw text with little user supervision. Therefore, OntoUSP simplifies the weight learning problem by adopting standard mestimation for smoothing. Namely, the weights for cluster c are set by counting its observations plus m fractional samples from its parent distribution. When c has few observations, its unreliable statistics can be significantly augmented via the smoothing by its parent (and in turn to a gradually smaller degree by its ancestors). m is a hyperparameter that can be used to trade off bias towards statistics for parent vs oneself. OntoUSP also needs to balance between two conflicting aspects during learning. On one hand, it should encourage creating abstract clusters to summarize intrinsic commonalities among the children. On the other hand, this needs to be heavily regularized to avoid mistaking noise for the signal. OntoUSP does this by a combination of priors and thresholding. To encourage the induction of higher-level nodes and inheritance, OntoUSP imposes an exponential prior β on the number of parameter slots. Each slot corresponds to a distinct property value. A child cluster inherits its parent’s slots (and thus avoids the penalty on them). On301 toUSP also stipulates that, in an ABSTRACT operation, a new property cluster can be created either as a concrete cluster with full parameterization, or as an abstract cluster that merely serves for smoothing purposes. To discourage overproposing clusters and ISA links, OntoUSP imposes a large exponential prior γ on the number of concrete clusters created by ABSTRACT. For abstract cluster, it sets a cut-off tp and only allows storing a probability value no less than tp. Like USP, it also rejects MERGE and COMPOSE operations that improve loglikelihood by less than to. These priors and cut-off values can be tuned to control the granularity of the induced ontology and clusters. Concretely, given semantic parses L, OntoUSP computes the optimal parameters and evaluates the regularized log-likelihood as follows. Let wp2,v denote the weight of the ISA mixture formula ISA(p1, +p2) ∧x ∈p1 ∧HasValue(x, +v). For convenience, for each pair of property cluster c and value v, OntoUSP instead computes and stores w′ c,v = P ISA(c, a) wa,v, which sums over all weights for c and its ancestors. (Thus wc,v = w′ c,v −w′ p,v, where p is the parent of c.) Like USP, OntoUSP imposes local normalization constraints that enable closed-form estimation of the optimal parameters and likelihood. Specifically, using m-estimation, the optimal w′ c,v is log((m·ew′ p,v +nc,v)/(m+nc)), where p is the parent of c and n is the count. The log-likelihood is P c,v w′ c,v ·nc,v, which is then augmented by the priors. 4 Experiments 4.1 Methodology Evaluating unsupervised ontology induction is difficult, because there is no gold ontology for comparison. Moreover, our ultimate goal is to aid knowledge acquisition, rather than just inducing an ontology for its own sake. Therefore, we used the same methodology and dataset as the USP paper to evaluate OntoUSP on its capability in knowledge acquisition. Specifically, we applied OntoUSP to extract knowledge from the GENIA dataset (Kim et al., 2003) and answer questions, and we evaluated it on the number of extracted answers and accuracy. GENIA contains 1999 PubMed abstracts.6 The question set con6http://www-tsujii.is.s.u-tokyo.ac.jp/GENIA/home/wiki.cgi. tains 2000 questions which were created by sampling verbs and entities according to their frequencies in GENIA. Sample questions include “What regulates MIP-1alpha?”, “What does anti-STAT 1 inhibit?”. These simple question types were used to focus the evaluation on the knowledge extraction aspect, rather than engineering for handling special question types and/or reasoning. 4.2 Systems OntoUSP is the first unsupervised approach that synergistically conducts ontology induction, population, and knowledge extraction. The system closest in aim and capability is USP. We thus compared OntoUSP with USP and all other systems evaluated in the USP paper (Poon and Domingos, 2009). Below is a brief description of the systems. (For more details, see Poon & Domingos (2009).) Keyword is a baseline system based on keyword matching. It directly matches the question substring containing the verb and the available argument with the input text, ignoring case and morphology. Given a match, two ways to derive the answer were considered: KW simply returns the rest of sentence on the other side of the verb, whereas KW-SYN is informed by syntax and extracts the answer from the subject or object of the verb, depending on the question (if the expected argument is absent, the sentence is ignored). TextRunner (Banko et al., 2007) is the state-ofthe-art system for open-domain information extraction. It inputs text and outputs relational triples in the form (R, A1, A2), where R is the relation string, and A1, A2 the argument strings. To answer questions, each triple-question pair is considered in turn by first matching their relation strings, and then the available argument strings. If both match, the remaining argument string in the triple is returned as an answer. Results were reported when exact match is used (TR-EXACT), or when the triple strings may contain the question ones as substrings (TR-SUB). RESOLVER (Yates and Etzioni, 2009) inputs TextRunner triples and collectively resolves coreferent relation and argument strings. To answer questions, the only difference from TextRunner is that a question string can match any string in its cluster. As in TextRunner, results were reported for both exact match (RS-EXACT) and substring (RS-SUB). DIRT (Lin and Pantel, 2001) resolves binary rela302 Table 1: Comparison of question answering results on the GENIA dataset. Results for systems other than OntoUSP are from Poon & Domingos (2009). # Total # Correct Accuracy KW 150 67 45% KW-SYN 87 67 77% TR-EXACT 29 23 79% TR-SUB 152 81 53% RS-EXACT 53 24 45% RS-SUB 196 81 41% DIRT 159 94 59% USP 334 295 88% OntoUSP 480 435 91% tions by inputting a dependency path that signifies the relation and returns a set of similar paths. To use DIRT in question answering, it was queried to obtain similar paths for the relation of the question, which were then used to match sentences. USP (Poon and Domingos, 2009) parses the input text using the Stanford dependency parser (Klein and Manning, 2003; de Marneffe et al., 2006), learns an MLN for semantic parsing from the dependency trees, and outputs this MLN and the MAP semantic parses of the input sentences. These MAP parses formed the knowledge base (KB). To answer questions, USP first parses the questions (with the question slot replaced by a dummy word), and then matches the question parse to parses in the KB by testing subsumption. OntoUSP uses a similar procedure as USP for extracting knowledge and answering questions, except for two changes. First, USP’s learning and parsing algorithms are replaced with OntoUSPLearn and OntoUSP-Parse, respectively. Second, when OntoUSP matches a question to its KB, it not only considers the lambda-form cluster of the question relation, but also all its sub-clusters.7 4.3 Results Table 1 shows the results comparing OntoUSP with other systems. While USP already greatly outperformed other systems in both precision and recall, OntoUSP further substantially improved on the recall of USP, without any loss in precision. In particular, OntoUSP extracted 140 more correct answers than USP, for a gain of 47% in absolute 7Additional details are available at http : //alchemy.cs.washington.edu/papers/poon10. ISA ISA INHIBIT induce, enhance, trigger, augment, up-regulate INDUCE inhibit, block, suppress, prevent, abolish, abrogate, down-regulate activate regulate, control, govern, modulate ISA ACTIVATE REGULATE Figure 3: A fragment of the induced ISA hierarchy, showing the core forms for each cluster (the cluster labels are added by the authors for illustration purpose). recall. Compared to TextRunner (TR-SUB), OntoUSP gained on precision by 38 points and extracted more than five times of correct answers. Manual inspection shows that the induced ISA hierarchy is the key for the recall gain. Like USP, OntoUSP discovered the following clusters (in core forms) that represent some of the core concepts in biomedical research: {regulate, control, govern, modulate} {induce, enhance, trigger, augment, upregulate} {inhibit, block, suppress, prevent, abolish, abrogate, down-regulate} However, USP formed these as separate clusters, whereas OntoUSP in addition induces ISA relations from the INDUCE and INHIBIT clusters to the REGULATE cluster (Figure 3). This allows OntoUSP to answer many more questions that are asked about general regulation events, even though the text states them with specific regulation directions like “induce” or “inhibit”. Below is an example question-answer pair output by OntoUSP; neither USP nor any other system were able to extract the necessary knowledge. Q: What does IL-2 control? A: The DEX-mediated IkappaBalpha induction. Sentence: Interestingly, the DEX-mediated IkappaBalpha induction was completely inhibited by IL-2, but not IL-4, in Th1 cells, while the reverse profile was seen in Th2 cells. OntoUSP also discovered other interesting commonalities among the clusters. For example, both USP and OntoUSP formed a singleton cluster with core form “activate”. Although this cluster may appear similar to the INDUCE cluster, the data in GENIA does not support merging the two. However, OntoUSP discovered that 303 the ACTIVATE cluster, while not completely resolvent with INDUCE, shared very similar distributions in their agent arguments. In fact, they are so similar that OntoUSP merges them into a single property cluster. It found that the patient arguments of INDUCE and INHIBIT are very similar and merged them. In turn, OntoUSP formed ISA links from these three object clusters to REGULATE, as well as among their property clusters. Intuitively, this makes sense. The positive- and negative-regulation events, as signified by INDUCE and INHIBIT, often target similar object entities or processes. However, their agents tend to differ since in one case they are inducers, and in the other they are inhibitors. On the other hand, ACTIVATE and INDUCE share similar agents since they both signify positive regulation. However, “activate” tends to be used more often when the patient argument is a concrete entity (e.g., cells, genes, proteins), whereas “induce” and others are also used with processes and events (e.g., expressions, inhibition, pathways). USP was able to resolve common syntactic differences such as active vs. passive voice. However, it does so on the basis of individual verbs, and there is no generalization beyond their clusters. OntoUSP, on the other hand, formed a highlevel cluster with two abstract property clusters, corresponding to general agent argument and patient argument. The active-passive alternation is captured in these clusters, and is inherited by all descendant clusters, including many rare verbs like “super-induce” which only occur once in GENIA and for which there is no way that USP could have learned about their active-passive alternations. This illustrates the importance of discovering ISA relations and performing hierarchical smoothing. 4.4 Discussion OntoUSP is a first step towards joint ontology induction and knowledge extraction. The experimental results demonstrate the promise in this direction. However, we also notice some limitations in the current system. While OntoUSP induced meaningful ISA relations among relation clusters like REGULATE, INDUCE, etc., it was less successful in inducing ISA relations among entity clusters such as specific genes and proteins. This is probably due to the fact that our model only considers local features such as the parent and arguments. A relation is often manifested as verbs and has several arguments, whereas an entity typically appears as an argument of others and has few arguments of its own. As a result, in average, there is less information available for entities than relations. Presumably, we can address this limitation by modeling longer-ranged dependencies such as grandparents, siblings, etc. This is straightforward to do in Markov logic. OntoUSP also uses a rather elaborate scheme for regularization. We hypothesize that this can be much simplified and improved by adopting a principled framework such as Dudik et al. (2007). 5 Conclusion This paper introduced OntoUSP, the first unsupervised end-to-end system for ontology induction and knowledge extraction from text. OntoUSP builds on the USP semantic parser by adding the capability to form hierarchical clusterings of logical expressions, linked by ISA relations, and using them for hierarchical smoothing. OntoUSP greatly outperformed USP and other state-of-theart systems in a biomedical knowledge acquisition task. Directions for future work include: exploiting the ontological structure for principled handling of antonyms and (more generally) expressions with opposite meanings; developing and testing alternate methods for hierarchical modeling in OntoUSP; scaling up learning and inference to larger corpora; investigating the theoretical properties of OntoUSP’s learning approach and generalizing it to other tasks; answering questions that require inference over multiple extractions; etc. 6 Acknowledgements We give warm thanks to the anonymous reviewers for their comments. This research was partly funded by ARO grant W911NF-08-1-0242, AFRL contract FA8750-09-C0181, DARPA contracts FA8750-05-2-0283, FA8750-07-D0185, HR0011-06-C-0025, HR0011-07-C-0060 and NBCHD030010, NSF grants IIS-0534881 and IIS-0803481, and ONR grant N00014-08-1-0670. The views and conclusions contained in this document are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ARO, DARPA, NSF, ONR, or the United States Government. References Hiyan Alshawi. 1990. Resolving quasi logical forms. Computational Linguistics, 16:133–144. G. Bakir, T. Hofmann, B. B. Sch¨olkopf, A. Smola, B. Taskar, 304 S. Vishwanathan, and (eds.). 2007. Predicting Structured Data. MIT Press, Cambridge, MA. Michele Banko, Michael J. Cafarella, Stephen Soderland, Matt Broadhead, and Oren Etzioni. 2007. Open information extraction from the web. In Proceedings of the Twentieth International Joint Conference on Artificial Intelligence, pages 2670–2676, Hyderabad, India. AAAI Press. Philipp Cimiano. 2006. Ontology learning and population from text. Springer. Marie-Catherine de Marneffe, Bill MacCartney, and Christopher D. Manning. 2006. Generating typed dependency parses from phrase structure parses. In Proceedings of the Fifth International Conference on Language Resources and Evaluation, pages 449–454, Genoa, Italy. ELRA. Pedro Domingos and Daniel Lowd. 2009. Markov Logic: An Interface Layer for Artificial Intelligence. Morgan & Claypool, San Rafael, CA. Miroslav Dudik, David Blei, and Robert Schapire. 2007. Hierarchical maximum entropy density estimation. In Proceedings of the Twenty Fourth International Conference on Machine Learning. Christiane Fellbaum, editor. 1998. WordNet: An Electronic Lexical Database. MIT Press, Cambridge, MA. Andrew Gelman and Jennifer Hill. 2006. Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge University Press. Lise Getoor and Ben Taskar, editors. 2007. Introduction to Statistical Relational Learning. MIT Press, Cambridge, MA. Marti Hearst. 1992. Automatic acquisition of hyponyms from large text corpora. In Proceedings of the 14th International Conference on Computational Linguistics. Jin-Dong Kim, Tomoko Ohta, Yuka Tateisi, and Jun’ichi Tsujii. 2003. GENIA corpus - a semantically annotated corpus for bio-textmining. Bioinformatics, 19:180–82. Dan Klein and Christopher D. Manning. 2003. Accurate unlexicalized parsing. In Proceedings of the Forty First Annual Meeting of the Association for Computational Linguistics, pages 423–430. Dekang Lin and Patrick Pantel. 2001. DIRT - discovery of inference rules from text. In Proceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 323–328, San Francisco, CA. ACM Press. Alexander Maedche. 2002. Ontology learning for the semantic Web. Kluwer Academic Publishers, Boston, Massachusetts. Andrew McCallum, Ronald Rosenfeld, Tom Mitchell, and Andrew Ng. 1998. Improving text classification by shrinkage in a hierarchy of classes. In Proceedings of the Fifteenth International Conference on Machine Learning. Hoifung Poon and Pedro Domingos. 2008. Joint unsupervised coreference resolution with Markov logic. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 649–658, Honolulu, HI. ACL. Hoifung Poon and Pedro Domingos. 2009. Unsupervised semantic parsing. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1–10, Singapore. ACL. Rion Snow, Daniel Jurafsky, and Andrew Ng. 2006. Semantic taxonomy induction from heterogenous evidence. In Proceedings of COLING/ACL 2006. S. Staab and R. Studer. 2004. Handbook on ontologies. Springer. Fabian Suchanek, Gjergji Kasneci, and Gerhard Weikum. 2008. Yago - a large ontology from Wikipedia and WordNet. Journal of Web Semantics. Fabian Suchanek, Mauro Sozio, and Gerhard Weikum. 2009. Sofie: A self-organizing framework for information extraction. In Proceedings of the Eighteenth International Conference on World Wide Web. Jun-ichi Tsujii. 2004. Thesaurus or logical ontology, which do we need for mining text? In Proceedings of the Language Resources and Evaluation Conference. Fei Wu and Daniel S. Weld. 2008. Automatically refining the wikipedia infobox ontology. In Proceedings of the Seventeenth International Conference on World Wide Web, Beijing, China. Alexander Yates and Oren Etzioni. 2009. Unsupervised methods for determining object and relation synonyms on the web. Journal of Artificial Intelligence Research, 34:255–296. 305
|
2010
|
31
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 306–315, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Exploring Syntactic Structural Features for Sub-Tree Alignment using Bilingual Tree Kernels Jun Sun1,2 Min Zhang1 Chew Lim Tan2 1 Institute for Infocomm Research 2School of Computing, National University of Singapore [email protected] [email protected] [email protected] Abstract We propose Bilingual Tree Kernels (BTKs) to capture the structural similarities across a pair of syntactic translational equivalences and apply BTKs to sub-tree alignment along with some plain features. Our study reveals that the structural features embedded in a bilingual parse tree pair are very effective for sub-tree alignment and the bilingual tree kernels can well capture such features. The experimental results show that our approach achieves a significant improvement on both gold standard tree bank and automatically parsed tree pairs against a heuristic similarity based method. We further apply the sub-tree alignment in machine translation with two methods. It is suggested that the subtree alignment benefits both phrase and syntax based systems by relaxing the constraint of the word alignment. 1 Introduction Syntax based Statistical Machine Translation (SMT) systems allow the translation process to be more grammatically performed, which provides decent reordering capability. However, most of the syntax based systems construct the syntactic translation rules based on word alignment, which not only suffers from the pipeline errors, but also fails to effectively utilize the syntactic structural features. To address those deficiencies, Tinsley et al. (2007) attempt to directly capture the syntactic translational equivalences by automatically conducting sub-tree alignment, which can be defined as follows: A sub-tree alignment process pairs up sub-tree pairs across bilingual parse trees whose contexts are semantically translational equivalent. According to Tinsley et al. (2007), a sub-tree aligned parse tree pair follows the following criteria: (i) a node can only be linked once; (ii) descendants of a source linked node may only link to descendants of its target linked counterpart; (iii) ancestors of a source linked node may only link to ancestors of its target linked counterpart. By sub-tree alignment, translational equivalent sub-tree pairs are coupled as aligned counterparts. Each pair consists of both the lexical constituents and their maximum tree structures generated over the lexical sequences in the original parse trees. Due to the 1-to-1 mapping between sub-trees and tree nodes, sub-tree alignment can also be considered as node alignment by conducting multiple links across the internal nodes as shown in Fig. 1. Previous studies conduct sub-tree alignments by either using a rule based method or conducting some similarity measurement only based on lexical features. Groves et al. (2004) conduct sub-tree alignment by using some heuristic rules, lack of extensibility and generality. Tinsley et al. (2007) Figure 1: Sub-tree alignment as referred to Node alignment 306 and Imamura (2001) propose some score functions based on the lexical similarity and co-occurrence. These works fail to utilize the structural features, rendering the syntactic rich task of sub-tree alignment less convincing and attractive. This may be due to the fact that the syntactic structures in a parse tree pair are hard to describe using plain features. In addition, explicitly utilizing syntactic tree fragments results in exponentially high dimensional feature vectors, which is hard to compute. Alternatively, convolution parse tree kernels (Collins and Duffy, 2001), which implicitly explore the tree structure information, have been successfully applied in many NLP tasks, such as Semantic parsing (Moschitti, 2004) and Relation Extraction (Zhang et al. 2006). However, all those studies are carried out in monolingual tasks. In multilingual tasks such as machine translation, tree kernels are seldom applied. In this paper, we propose Bilingual Tree Kernels (BTKs) to model the bilingual translational equivalences, in our case, to conduct sub-tree alignment. This is motivated by the decent effectiveness of tree kernels in expressing the similarity between tree structures. We propose two kinds of BTKs named dependent Bilingual Tree Kernel (dBTK), which takes the sub-tree pair as a whole and independent Bilingual Tree Kernel (iBTK), which individually models the source and the target sub-trees. Both kernels can be utilized within different feature spaces using various representations of the sub-structures. Along with BTKs, various lexical and syntactic structural features are proposed to capture the correspondence between bilingual sub-trees using a polynomial kernel. We then attempt to combine the polynomial kernel and BTKs to construct a composite kernel. The sub-tree alignment task is considered as a binary classification problem. We employ a kernel based classifier with the composite kernel to classify each candidate of sub-tree pair as aligned or unaligned. Then a greedy search algorithm is performed according to the three criteria of sub-tree alignment within the space of candidates classified as aligned. We evaluate the sub-tree alignment on both the gold standard tree bank and an automatically parsed corpus. Experimental results show that the proposed BTKs benefit sub-tree alignment on both corpora, along with the lexical features and the plain structural features. Further experiments in machine translation also suggest that the obtained sub-tree alignment can improve the performance of both phrase and syntax based SMT systems. 2 Bilingual Tree Kernels In this section, we propose the two BTKs and study their capability and complexity in modeling the bilingual structural similarity. Before elaborating the concepts of BTKs, we first illustrate some notations to facilitate further understanding. Each sub-tree pair ሺܵ· ܶሻ can be explicitly decomposed into multiple sub-structures which belong to the given sub-structure spaces. ࣭ூൌ ሼݏଵ, … , ݏ, … , ݏூሽ refers to the source tree substructure space; while ࣮ൌ൛ݐଵ, … , ݐ, … , ݐൟ refers to the target sub-structure space. A sub-structure pair ሺݏ, ݐሻ refers to an element in the set of the Cartesian product of the two sub-structure spaces: ሺݏ, ݐሻא ࣭ூൈ࣮. 2.1 Independent Bilingual Tree Kernel (iBTK) Given the sub-structure spaces ࣭ூ and ࣮, we construct two vectors using the integer counts of the source and target sub-structures: ߶ሺܵሻൌ൫#ሺݏଵሻ, … , #ሺݏሻ, … , #ሺݏ|࣭|ሻ൯ ߶ሺܶሻൌቀ#ሺݐଵሻ, … , #ሺݐሻ, … , #ሺݐห࣮หሻቁ where #ሺݏሻ and #ሺݐሻ are the numbers of occurrences of the sub-structures ݏ and ݐ. In order to compute the dot product of the feature vectors in the exponentially high dimensional feature space, we introduce the tree kernel functions as follows: ்ࣥሺܵ· ܶ, ܵᇱ· ܶᇱሻൌࣥሺܵ, ܵᇱሻࣥሺܶ, ܶᇱሻ The iBTK is defined as a composite kernel consisting of a source tree kernel and a target tree kernel which measures the source and the target structural similarity respectively. Therefore, the composite kernel can be computed using the ordinary monolingual tree kernels (Collins and Duffy, 2001). ࣥሺܵ, ܵᇱሻൌ൏߶ሺܵሻ, ߶ሺܵᇱሻ ൌ∑ ቀ∑ ܫሺ݊௦ሻ ೞאேೄ · ∑ ܫሺ݊௦ ᇱሻ ೞᇲאேೄ ᇲ ቁ |࣭| ୀଵ ൌ∑ ∑ ∆ሺ݊௦, ݊௦ ᇱሻ ೞᇲאேೄ ᇲ ೞאேೄ where ܰௌ and ܰௌ ᇱ refer to the node sets of the source sub-tree ܵ and ܵᇱ respectvely. ܫሺ݊௦ሻ is an indicator function which equals to 1 iff the substructure ݏ is rooted at the node ݊௦ and 0 otherwise.∆ሺ݊௦, ݊௦ ᇱሻൌ∑ ൫ܫሺ݊௦ሻ· ܫሺ݊௦ ᇱሻ൯ |࣭| ୀଵ is the number of identical sub-structures rooted at ݊௦ and ݊௦ᇱ. Then we compute the ∆ሺ݊௦, ݊௦ ᇱሻ function as follows: 307 (1) If the production rule at ݊௦ and ݊௦ᇱ are different, ∆ሺ݊௦, ݊௦ ᇱሻൌ0; (2)else if both ݊௦and ݊௦ᇱare POS tags, ∆ሺ݊௦, ݊௦ ᇱሻൌߣ; (3)else, ∆ሺ݊௦, ݊௦ ᇱሻൌߣ∏ ቀ1 ∆൫ܿሺ݊௦, ݈ሻ, ܿሺ݊௦ ᇱ, ݈ሻ൯ቁ ሺೞሻ ୀଵ . where ݊ܿሺ݊௦ሻ is the child number of ݊௦, ܿሺ݊௦, ݈ሻ is the lth child of ݊௦, ߣ is the decay factor used to make the kernel value less variable with respect to the number of sub-structures. Similarly, we can decompose the target kernel as ࣥሺܶ, ܶᇱሻൌ∑ ∑ ∆ሺ݊௧, ݊௧ ᇱሻ ᇲאே ᇲ אே and run the algorithm above as well. The disadvantage of the iBTK is that it fails to capture the correspondence across the substructure pairs. However, the composite style of constructing the iBTK helps keep the computational complexity comparable to the monolingual tree kernel, which is ܱሺ|ܰௌ| · |ܰௌ ᇱ| |்ܰ| · |்ܰ ᇱ|ሻ. 2.2 Dependent Bilingual Tree Kernel (dBTK) The iBTK explores the structural similarity of the source and the target sub-trees respectively. As an alternative, we further define a kernel to capture the relationship across the counterparts without increasing the computational complexity. As a result, we propose the dependent Bilingual Tree kernel (dBTK) to jointly evaluate the similarity across sub-tree pairs by enlarging the feature space to the Cartesian product of the two substructure sets. A dBTK takes the source and the target substructure pair as a whole and recursively calculate over the joint sub-structures of the given sub-tree pair. We define the dBTK as follows: Given the sub-structure space ࣭ூൈ࣮, we construct a vector using the integer counts of the substructure pairs to represent a sub-tree pair: ߶ሺܵ· ܶሻൌቆ #ሺݏଵ, ݐଵሻ, … , #ሺݏଵ, ݐห࣮หሻ, #ሺݏଶ, ݐଵሻ, … , #ሺݏ|࣭|, ݐଵሻ, … , #ሺݏ|࣭|, ݐห࣮หሻቇ where #൫ݏ, ݐ൯ is the number of occurrences of the sub-structure pair ൫ݏ, ݐ൯. ࣥௗ்ሺܵ· ܶ, ܵᇱ· ܶᇱሻ ൌ൏߶ሺܵ· ܶሻ, ߶ሺܵᇱ· ܶᇱሻ ൌ∑ ቆ ∑ ∑ ܫሺ݊௦, ݊௧ሻ אே ೞאேೄ · ∑ ∑ ܫሺ݊௦ ᇱ, ݊௧ ᇱሻ ᇲאே ᇲ ೞᇲאேೄ ᇲ ቇ |࣭ൈ࣮| ୀଵ ൌ∑ ∑ ∑ ∑ ∆൬ሺ݊௦, ݊௧ሻ, ሺ݊௦ ᇱ, ݊௧ ᇱሻ൰ ᇲאே ᇲ ೞᇲאேೄ ᇲ אே ೞאேೄ (1) ൌ∑ ∑ ∑ ∑ ൬∆ሺ݊௦, ݊௦ ᇱሻ· ∆ሺ݊௧, ݊௧ ᇱሻ൰ ᇲאே ᇲ ೞᇲאேೄ ᇲ אே ೞאேೄ (2) ൌ∑ ∑ ∆ሺ݊௦, ݊௦ ᇱሻ ೞᇲאேೄ ᇲ ೞאேೄ ∑ ∑ ∆ሺ݊௧, ݊௧ ᇱሻ ᇲאே ᇲ אே ൌࣥሺܵ, ܵᇱሻ· ࣥሺܶ, ܶᇱሻ It is infeasible to explicitly compute the kernel function by expressing the sub-trees as feature vectors. In order to achieve convenient computation, we deduce the kernel function as the above. The deduction from (1) to (2) is derived according to the fact that the number of identical substructure pairs rooted in the node pairs ሺ݊௦, ݊௧ሻ and ሺ݊௦ ᇱ, ݊௧ ᇱሻ equals to the product of the respective counts. As a result, the dBTK can be evaluated as a product of two monolingual tree kernels. Here we verify the correctness of the kernel by directly constructing the feature space for the inner product. Alternatively, Cristianini and Shawe-Taylor (2000) prove the positive semi-definite characteristic of the tensor product of two kernels. The decomposition benefits the efficient computation to use the algorithm for the monolingual tree kernel in Section 2.1. The computational complexity of the dBTK is still ܱሺ|ܰௌ| · |ܰௌ ᇱ| |்ܰ| · |்ܰ ᇱ|ሻ. 3 Sub-structure Spaces for BTKs The syntactic translational equivalences under BTKs are evaluated with respective to the substructures factorized from the candidate sub-tree pairs. In this section, we propose different substructures to facilitate the measurement of syntactic similarity for sub-tree alignment. Since the proposed BTKs can be computed by individually evaluating the source and target monolingual tree kernels, the definition of the sub-structure can be simplified to base only on monolingual sub-trees. 3.1 Subset Tree Motivated from Collins and Duffy (2002) in monolingual tree kernels, the Subset Tree (SST) can be employed as sub-structures. An SST is any subgraph, which includes more than one non-terminal node, with the constraint that the entire rule productions are included. Fig. 2 shows an example of the SSTs decomposed from the source sub-tree rooted at VP*. 3.2 Root directed Subset Tree Monolingual Tree kernels achieve decent performance using the SSTs due to the rich exploration of syntactic information. However, the sub-tree alignment task requires strong capability of discriminating the sub-trees with their roots across adjacent generations, because those candidates share many identical SSTs. As illustrated in Fig 2, the source sub-tree rooted at VP*, which should be aligned to the target sub-tree rooted at NP*, may be likely aligned to the sub-tree rooted at PP*, 308 which shares quite a similar context with NP*. It is also easy to show that the latter shares all the SSTs that the former obtains. In consequence, the values of the SST based kernel function are quite similar between the candidate sub-tree pair rooted at (VP*,NP*) and (VP*,PP*). In order to effectively differentiate the candidates like the above, we propose the Root directed Subset Tree (RdSST) by encapsulating each SST with the root of the given sub-tree. As shown in Fig 2, a sub-structure is considered identical to the given examples, when the SST is identical and the root tag of the given sub-tree is NP. As a result, the kernel function in Section 2.1 is re-defined as: ࣥሺܵ, ܵᇱሻൌ∑ ∑ ∆ሺ݊௦, ݊௦ ᇱሻܫሺݎ௦, ݎ௦ ᇱሻ ೞᇲאேೄ ᇲ ೞאேೄ ൌܫሺݎ௦, ݎ௦ᇱሻ∑ ∑ ∆ሺ݊௦, ݊௦ᇱሻ ೞᇲאேೄ ᇲ ೞאேೄ where ݎ௦ and ݎ௦ ᇱ are the root nodes of the subtree ܵ and ܵᇱrespectively. The indicator function ܫሺݎ௦, ݎ௦ ᇱሻ equals to 1 if ݎ௦ and ݎ௦ ᇱ are identical, and 0 otherwise. Although defined for individual SST, the indicator function can be evaluated outside the summation, without increasing the computational complexity of the kernel function. 3.3 Root generated Subset Tree Some grammatical tags (NP/VP) may have identical tags as their parents or children which may make RdSST less effective. Consequently, we step further to propose the sub-structure of Root generated Subset Tree (RgSST). An RgSST requires the root node of the given sub-tree to be part of the sub-structure. In other words, all sub-structures should be generated from the root of the given sub-tree as presented in Fig. 2. Therefore the kernel function can be simplified to only capture the sub-structure rooted at the root of the sub-tree. ࣥሺܵ, ܵᇱሻൌ∆ሺݎ௦, ݎ௦ ᇱሻ where ݎ௦ and ݎ௦ ᇱ are the root nodes of the subtree ܵ and ܵᇱ respectively. The time complexity is reduced to ܱሺ|ܰௌ| |ܰௌ ᇱ| |்ܰ| |்ܰ ᇱ|ሻ. 3.4 Root only More aggressively, we can simplify the kernel to only measure the common root node without considering the complex tree structures. Therefore the kernel function is simplified to be a binary function with time complexity ܱሺ1ሻ. ࣥሺܵ, ܵᇱሻൌܫሺݎ௦, ݎ௦ ᇱሻ 4 Plain features Besides BTKs, we introduce various plain lexical features and structural features which can be expressed as feature functions. The lexical features with directions are defined as conditional feature functions based on the conditional lexical translation probabilities. The plain syntactic structural features can deal with the structural divergence of bilingual parse trees in a more general perspective. 4.1 Lexical and Word Alignment Features In this section, we define seven lexical features to measure semantic similarity of a given sub-tree pair. Internal Lexical Features: We define two lexical features with respective to the internal span of the sub-tree pair. ߶ଵሺܵ|ܶሻൌ൫∏ ∑ ܲሺݑ|ݒሻ ௨אሺௌሻ ௩אሺ்ሻ ൯ భ |ሺሻ| Figure 2: Illustration of SST, RdSST and RgSST 309 ߶ଶሺܶ|ܵሻൌ൫∏ ∑ ܲሺݒ|ݑሻ ௩אሺ்ሻ ௨אሺௌሻ ൯ భ |ሺೄሻ| where ܲሺݒ|ݑሻ refers to the lexical translation probability from the source word ݑ to the target word ݒ within the sub-tree spans, while ܲሺݑ|ݒሻ refers to that from target to source; ݅݊ሺܵሻ refers to the word set for the internal span of the source sub-tree ܵ, while ݅݊ሺܶሻ refers to that of the target sub-tree ܶ. Internal-External Lexical Features: These features are motivated by the fact that lexical translation probabilities within the translational equivalence tend to be high, and that of the nonequivalent counterparts tend to be low. ߶ଷሺܵ|ܶሻൌ൫∏ ∑ ܲሺݑ|ݒሻ ௨א௨௧ሺௌሻ ௩אሺ்ሻ ൯ భ |ሺሻ| ߶ସሺܶ|ܵሻൌ൫∏ ∑ ܲሺݒ|ݑሻ ௩א௨௧ሺ்ሻ ௨אሺௌሻ ൯ భ |ሺೄሻ| where ݑݐሺܵሻ refers to the word set for the external span of the source sub-tree ܵ, while ݑݐሺܶሻ refers to that of the target sub-tree ܶ. Internal Word Alignment Features: The word alignment links account much for the cooccurrence of the aligned terms. We define the internal word alignment features as follows: ߶ହሺܵ, ܶሻൌ ∑ ∑ ߜሺݑ, ݒሻ· ൫ܲሺݑ|ݒሻ· ܲሺݒ|ݑሻ൯ ଵ ଶ ௨אሺௌሻ ௩אሺ்ሻ ሺ|݅݊ሺܵሻ| · |݅݊ሺܶሻ|ሻ ଵ ଶ where ߜሺݑ, ݒሻൌቄ1 if ሺݑ, ݒሻ is aligned 0 otherwise The binary function ߜሺݑ, ݒሻ is employed to trigger the computation only when a word aligned link exists for the two words ሺݑ, ݒሻ within the subtree span. Internal-External Word Alignment Features: Similar to the lexical features, we also introduce the internal-external word alignment features as follows: ߶ሺܵ, ܶሻൌ ∑ ∑ ߜሺݑ, ݒሻ· ൫ܲሺݑ|ݒሻ· ܲሺݒ|ݑሻ൯ ଵ ଶ ௨א௨௧ሺௌሻ ௩אሺ்ሻ ሺ|ݑݐሺܵሻ| · |݅݊ሺܶሻ|ሻ ଵ ଶ ߶ሺܵ, ܶሻൌ ∑ ∑ ߜሺݑ, ݒሻ· ൫ܲሺݑ|ݒሻ· ܲሺݒ|ݑሻ൯ ଵ ଶ ௨אሺௌሻ ௩א௨௧ሺ்ሻ ሺ|݅݊ሺܵሻ| · |ݑݐሺܶሻ|ሻ ଵ ଶ where ߜሺݑ, ݒሻൌቄ1 if ሺݑ, ݒሻ is aligned 0 otherwise 4.2 Online Structural Features In addition to the lexical correspondence, we also capture the structural divergence by introducing the following tree structural features. Span difference: Translational equivalent subtree pairs tend to share similar length of spans. Thus the model will penalize the candidate subtree pairs with largely different length of spans. ߮ଵሺܵ, ܶሻൌቚ |ሺௌሻ| |ሺ܁ሻ| െ |ሺ்ሻ| |ሺ܂ሻ| ቚ ܁ and ܂ refer to the entire source and target parse trees respectively. Therefore, |݅݊ሺ܁ሻ| and |݅݊ሺ܂ሻ| are the respective span length of the parse tree used for normalization. Number of Descendants: Similarly, the number of the root’s descendants of the aligned subtrees should also correspond. ߮ଶሺܵ, ܶሻൌቚ |ோሺௌሻ| |ோሺ܁ሻ| െ |ோሺ்ሻ| |ோሺ܂ሻ|ቚ where ܴሺ. ሻ refers to the descendant set of the root to a sub-tree. Tree Depth difference: Intuitively, translational equivalent sub-tree pairs tend to have similar depth from the root of the parse tree. We allow the model to penalize the candidate sub-tree pairs with quite different distance of path from the root of the parse tree to the root of the sub-tree. ߮ଷሺܵ, ܶሻൌቚ ௧ሺௌሻ ு௧ሺ܁ሻെ ௧ሺ்ሻ ு௧ሺ܂ሻቚ 5 Alignment Model Given feature spaces defined in the last two sections, we propose a 2-phase sub-tree alignment model as follows: In the 1st phase, a kernel based classifier, SVM in our study, is employed to classify each candidate sub-tree pair as aligned or unaligned. The feature vector of the classifier is computed using a composite kernel: ࣥሺܵ· ܶ, ܵᇱ· ܶᇱሻൌ ߠࣥሺܵ· ܶ, ܵᇱ· ܶᇱሻ∑ ߠ୧்ࣥ ሺܵ· ܶ, ܵᇱ· ܶᇱሻ K ୧ୀଵ ࣥሺ·,·ሻ is the normalized form of the polynomial kennel ࣥሺ·,·ሻ, which is a polynomial kernel with the degree of 2, utilizing the plain features. ்ࣥ ሺ·,·ሻ is the normalized form of the BTK ்ࣥ ሺ·,·ሻ, exploring the corresponding substructure space. The composite kernel can be constructed using the polynomial kernel for plain features and various BTKs for tree structure by linear combination with coefficient ߠ୧, where ∑ ߠ୧ K ୧ୀ ൌ1. In the 2nd phase, we adopt a greedy search with respect to the alignment probabilities. Since SVM is a large margin based discriminative classifier rather than a probabilistic model, we introduce a sigmoid function to convert the distance against the hyperplane to a posterior alignment probability as follows: 310 ܲሺܽା|ܵ, ܶሻൌ 1 1 ݁ିశ ܲሺܽି|ܵ, ܶሻൌ 1 1 ݁ିష where ܦା is the distance for the instances classified as aligned and ܦି is that for the unaligned. We use ܲሺܽା|ܵ, ܶሻ as the confidence to conduct the sure links for those classified as aligned. On this perspective, the alignment probability is suitable as a searching metric. The search space is reduced to that of the candidates classified as aligned after the 1st phase. 6 Experiments on Sub-Tree Alignments In order to evaluate the effectiveness of the alignment model and its capability in the applications requiring syntactic translational equivalences, we employ two corpora to carry out the sub-tree alignment evaluation. The first is HIT gold standard English Chinese parallel tree bank referred as HIT corpus1. The other is the automatically parsed bilingual tree pairs selected from FBIS corpus (allowing minor parsing errors) with human annotated sub-tree alignment. 6.1 Data preparation HIT corpus, which is collected from English learning text books in China as well as example sentences in dictionaries, is used for the gold standard corpus evaluation. The word segmentation, tokenization and parse-tree in the corpus are manually constructed or checked. The corpus is constructed with manually annotated sub-tree alignment. The annotation strictly reserves the semantic equivalence of the aligned sub-tree pair. Only sure links are conducted in the internal node level, without considering possible links adopted in word alignment. A different annotation criterion of the Chinese parse tree, designed by the annotator, is employed. Compared with the widely used Penn TreeBank annotation, the new criterion utilizes some different grammar tags and is able to effectively describe some rare language phenomena in Chinese. The annotator still uses Penn TreeBank annotation on the English side. The statistics of HIT corpus used in our experiment is shown in Table 1. We use 5000 sentences for experiment and divide them into three parts, with 3k for training, 1k for testing and 1k for tuning the parameters of kernels and thresholds of pruning the negative instances. 1HIT corpus is designed and constructed by HIT-MITLAB. http://mitlab.hit.edu.cn/index.php/resources.html . Most linguistically motivated syntax based SMT systems require an automatic parser to perform the rule induction. Thus, it is important to evaluate the sub-tree alignment on the automatically parsed corpus with parsing errors. In addition, HIT corpus is not applicable for MT experiment due to the problems of domain divergence, annotation discrepancy (Chinese parse tree employs a different grammar from Penn Treebank annotations) and degree of tolerance for parsing errors. Due to the above issues, we annotate a new data set to apply the sub-tree alignment in machine translation. We randomly select 300 bilingual sentence pairs from the Chinese-English FBIS corpus with the length 30 in both the source and target sides. The selected plain sentence pairs are further parsed by Stanford parser (Klein and Manning, 2003) on both the English and Chinese sides. We manually annotate the sub-tree alignment for the automatically parsed tree pairs according to the definition in Section 1. To be fully consistent with the definition, we strictly reserve the semantic equivalence for the aligned sub-trees to keep a high precision. In other words, we do not conduct any doubtful links. The corpus is further divided into 200 aligned tree pairs for training and 100 for testing as shown in Table 2. 6.2 Baseline approach We implement the work in Tinsley et al. (2007) as our baseline methodology. Given a tree pair ൏܁, ܂, the baseline approach first takes all the links between the sub-tree pairs as alignment hypotheses, i.e., the Cartesian product of the two sub-tree sets: ሼܵଵ, … , ܵ, … , ܵூሽൈ൛ܶଵ, … , ܶ, … , ܶൟ By using the lexical translation probabilities, each hypothesis is assigned an alignment score. All hypotheses with zero score are pruned out. Chinese English # of Sentence pair 300 Avg. Sentence Length 16.94 20.81 Avg. # of sub-tree 28.97 34.39 Avg. # of alignment 17.07 Table 2. Statistics of FBIS selected Corpus Chinese English # of Sentence pair 5000 Avg. Sentence Length 12.93 12.92 Avg. # of sub-tree 21.40 23.58 Avg. # of alignment 11.60 Table 1. Corpus Statistics for HIT corpus 311 Then the algorithm iteratively selects the link of the sub-tree pairs with the maximum score as a sure link, and blocks all hypotheses that contradict with this link and itself, until no non-blocked hypotheses remain. The baseline system uses many heuristics in searching the optimal solutions with alternative score functions. Heuristic skip1 skips the tied hypotheses with the same score, until it finds the highest-scoring hypothesis with no competitors of the same score. Heuristic skip2 deals with the same problem. Initially, it skips over the tied hypotheses. When a hypothesis sub-tree pair ൫ܵ, ܶ൯ without any competitor of the same score is found, where neither ܵ nor ܶ has been skipped over, the hypothesis is chosen as a sure link. Heuristic span1 postpones the selection of the hypotheses on the POS level. Since the highest-scoring hypotheses tend to appear on the leaf nodes, it may introduce ambiguity when conducting the alignment for a POS node whose child word appears twice in a sentence. The baseline method proposes two score functions based on the lexical translation probability. They also compute the score function by splitting the tree into the internal and external components. Tinsley et al. (2007) adopt the lexical translation probabilities dumped by GIZA++ (Och and Ney, 2003) to compute the span based scores for each pair of sub-trees. Although all of their heuristics combinations are re-implemented in our study, we only present the best result among them with the highest Recall and F-value as our baseline, denoted as skip2_s1_span12. 2 s1 denotes score function 1 in Tinsley et al. (2007), skip2_s1_span1 denotes the utilization of heuristics skip2 and span1 while using score function 1 6.3 Experimental settings We use SVM with binary classes as the classifier. In case of the implementation, we modify the Tree Kernel tool (Moschitti, 2004) and SVMLight (Joachims, 1999). The coefficient ߠ୧ for the composite kernel are tuned with respect to F-measure (F) on the development set of HIT corpus. We empirically set C=2.4 for SVM and use ߙൌ0.23, the default parameter ߣൌ0.4 for BTKs. Since the negative training instances largely overwhelm the positive instances, we prune the negative instances using the thresholds according to the lexical feature functions (߶ଵ, ߶ଶ, ߶ଷ, ߶ସ) and online structural feature functions ( ߮ଵ, ߮ଶ, ߮ଷ). Those thresholds are also tuned on the development set of HIT corpus with respect to F-measure. To learn the lexical and word alignment features for both the proposed model and the baseline method, we train GIZA++ on the entire FBIS bilingual corpus (240k). The evaluation is conducted by means of Precision (P), Recall (R) and Fmeasure (F). 6.4 Experimental results In Tables 3 and 4, we incrementally enlarge the feature spaces in certain order for both corpora and examine the feature contribution to the alignment results. In detail, the iBTKs and dBTKs are firstly combined with the polynomial kernel for plain features individually, then the best iBTK and dBTK are chosen to construct a more complex composite kernel along with the polynomial kernel for both corpora. The experimental results show that: • All the settings with structural features of the proposed approach achieve better performance than the baseline method. This is because the Feature Space P R F Lex 73.48 71.66 72.56 Lex +Online Str 77.02 73.63 75.28 Plain +dBTK-STT 81.44 74.42 77.77 Plain +dBTK-RdSTT 81.40 69.29 74.86 Plain +dBTK-RgSTT 81.90 67.32 73.90 Plain +dBTK-Root 78.60 80.90 79.73 Plain +iBTK-STT 82.94 79.44 81.15 Plain +iBTK-RdSTT 83.14 80 81.54 Plain +iBTK-RgSTT 83.09 79.72 81.37 Plain +iBTK-Root 78.61 79.49 79.05 Plain +dBTK-Root +iBTK-RdSTT 82.70 82.70 82.70 Baseline 70.48 78.70 74.36 Table 4. Structure feature contribution for FBIS test set Feature Space P R F Lex 61.62 58.33 59.93 Lex +Online Str 70.08 69.02 69.54 Plain +dBTK-STT 80.36 78.08 79.20 Plain +dBTK-RdSTT 87.52 74.13 80.27 Plain +dBTK-RgSTT 88.54 70.18 78.30 Plain +dBTK-Root 81.05 84.38 82.68 Plain +iBTK-STT 81.57 73.51 77.33 Plain +iBTK-RdSTT 82.27 77.85 80.00 Plain +iBTK-RgSTT 82.92 78.77 80.80 Plain +iBTK-Root 76.37 76.81 76.59 Plain +dBTK-Root +iBTK-RgSTT 85.53 85.12 85.32 Baseline 64.14 66.99 65.53 Table 3. Structure feature contribution for HIT test set *Plain= Lex +Online Str 312 baseline only assesses semantic similarity using the lexical features. The improvement suggests that the proposed framework with syntactic structural features is more effective in modeling the bilingual syntactic correspondence. • By introducing BTKs to construct a composite kernel, the performance in both corpora is significantly improved against only using the polynomial kernel for plain features. This suggests that the structural features captured by BTKs are quite useful for the sub-tree alignment task. We also try to use BTKs alone without the polynomial kernel for plain features; however, the performance is rather low. This suggests that the structure correspondence cannot be used to measure the semantically equivalent tree structures alone, since the same syntactic structure tends to be reused in the same parse tree and lose the ability of disambiguation to some extent. In other words, to capture the semantic similarity, structure features requires lexical features to cooperate. • After comparing iBTKs with the corresponding dBTKs, we find that for FBIS corpus, iBTK greatly outperforms dBTK in any feature space except the Root space. However, when it comes the HIT corpus, the gaps between the corresponding iBTKs and dBTKs are much closer, while on the Root space, dBTK outperforms iBTK to a large amount. This finding can be explained by the relationship between the amount of training data and the high dimensional feature space. Since dBTKs are constructed in a joint manner which obtains a much larger high dimensional feature space than those of iBTKs, dBTKs require more training data to excel its capability, otherwise it will suffer from the data sparseness problem. The reason that dBTK outperforms iBTK in the feature space of Root in FBIS corpus is that although it is a joint feature space, the Root node pairs can be constructed from a close set of grammar tags and to form a relatively low dimensional space. As a result, when applying to FBIS corpus, which only contains limited amount of training data, dBTKs will suffer more from the data sparseness problem, and therefore, a relatively low performance. When enlarging the amount of training corpus to the HIT corpus, the ability of dBTKs excels and the benefit from data increasing of dBTKs is more significant than iBTKs. • We also find that the introduction of BTKs gains more improvement in HIT gold standard corpus than in FBIS corpus. Other than the factor of the amount of training data, this is also because the plain features in Table 3 are not as effective as those in Table 4, since they are trained on FBIS corpus which facilitates Table 4 more with respect to the domains. On the other hand, the grammatical tags and syntactic tree structures are more accurate in HIT corpus, which facilitates the performance of BTKs in Table 3. • On the comparison across the different feature spaces of BTKs, we find that STT, RdSTT and TgSTT are rather selective, since Recalls of those feature spaces are relatively low, exp. for HIT corpus. However, the Root sub-structure obtains a satisfactory Recall for both corpora. That’s why we attempt to construct a more complex composite kernel in adoption of the kernel of dBTK-Root as below. • To gain an extra performance boosting, we further construct a composite kernel which includes the best iBTK and the best dBTK for each corpus along with the polynomial kernel for plain features. In the HIT corpus, we use dBTK in the Root space and iBTK in the RgSST space; while for FBIS corpus, we use dBTK in the Root space and iBTK in the RdSST space. The experimental results suggest that by combining iBTK and dBTK together, we can achieve more improvement. 7 Experiments on Machine Translation In addition to the intrinsic alignment evaluation, we further conduct the extrinsic MT evaluation. We explore the effectiveness of sub-tree alignment for both phrase based and linguistically motivated syntax based SMT systems. 7.1 Experimental configuration In the experiments, we train the translation model on FBIS corpus (7.2M (Chinese) + 9.2M (English) words in 240,000 sentence pairs) and train a 4gram language model on the Xinhua portion of the English Gigaword corpus (181M words) using the SRILM Toolkits (Stolcke, 2002). We use these sentences with less than 50 characters from the NIST MT-2002 test set as the development set (to speed up tuning for syntax based system) and the NIST MT-2005 test set as our test set. We use the Stanford parser (Klein and Manning, 2003) to parse bilingual sentences on the training set and Chinese sentences on the development and test set. The evaluation metric is case-sensitive BLEU-4. 313 For the phrase based system, we use Moses (Koehn et al., 2007) with its default settings. For the syntax based system, since sub-tree alignment can directly benefit Tree-2-Tree based systems, we apply the sub-tree alignment in a syntax system based on Synchronous Tree Substitution Grammar (STSG) (Zhang et al., 2007). The STSG based decoder uses a pair of elementary tree3 as a basic translation unit. Recent research on tree based systems shows that relaxing the restriction from tree structure to tree sequence structure (Synchronous Tree Sequence Substitution Grammar: STSSG) significantly improves the translation performance (Zhang et al., 2008). We implement the STSG/STSSG based model in the Pisces decoder with the identical features and settings in Sun et al. (2009). In the Pisces decoder, the STSSG based decoder translates each span iteratively in a bottom up manner which guarantees that when translating a source span, any of its subspans is already translated. The STSG based decoding can be easily performed with the STSSG decoder by restricting the translation rule set to be elementary tree pairs only. As for the alignment setting, we use the word alignment trained on the entire FBIS (240k) corpus by GIZA++ with heuristic grow-diag-final for both Moses and the syntax system. For sub-treealignment, we use the above word alignment to learn lexical/word alignment feature, and train with the FBIS training corpus (200) using the composite kernel of Plain+dBTK-Root+iBTKRdSTT. 7.2 Experimental results Compared with the adoption of word alignment, translational equivalences generated from structural alignment tend to be more grammatically 3 An elementary tree is a fragment whose leaf nodes can be either non-terminal symbols or terminal symbols. aware and syntactically meaningful. However, utilizing syntactic translational equivalences alone for machine translation loses the capability of modeling non-syntactic phrases (Koehn et al., 2003). Consequently, instead of using phrases constraint by sub-tree alignment alone, we attempt to combine word alignment and sub-tree alignment and deploy the capability of both with two methods. • Directly Concatenate (DirC) is operated by directly concatenating the rule set genereted from sub-tree alignment and the original rule set generated from word alignment (Tinsley et al., 2009). As shown in Table 5, we gain minor improvement in the Bleu score for all configurations. • Alternatively, we proposed a new approach to generate the rule set from the scratch. We constrain the bilingual phrases to be consistent with Either Word alignment or Sub-tree alignment (EWoS) instead of being originally consistent with the word alignment only. The method helps tailoring the rule set decently without redundant counts for syntactic rules. The performance is further improved compared to DirC in all systems. The findings suggest that with the modeling of non-syntactic phrases maintained, more emphasis on syntactic phrases can benefit both the phrase and syntax based SMT systems. 8 Conclusion In this paper, we explore syntactic structure features by means of Bilingual Tree Kernels and apply them to bilingual sub-tree alignment along with various lexical and plain structural features. We use both gold standard tree bank and the automatically parsed corpus for the sub-tree alignment evaluation. Experimental results show that our model significantly outperforms the baseline method and the proposed Bilingual Tree Kernels are very effective in capturing the cross-lingual structural similarity. Further experiment shows that the obtained sub-tree alignment benefits both phrase and syntax based MT systems by delivering more weight on syntactic phrases. Acknowledgments We thank MITLAB4 in Harbin Institute of Technology for licensing us their sub-tree alignment corpus for our research. 4 http://mitlab.hit.edu.cn/ . System Model BLEU Moses BP* 23.86 DirC 23.98 EWoS 24.48 Syntax STSG STSG 24.71 DirC 25.16 EWoS 25.38 Syntax STSSG 25.92 STSSG DirC 25.95 EWoS 26.45 Table 5. MT evaluation on various systems *BP denotes bilingual phrases 314 References David Burkett and Dan Klein. 2008. Two languages are better than one (for syntactic parsing). In Proceedings of EMNLP-08. 877-886. Nello Cristianini and John Shawe-Taylor. 2000. An introduction to support vector machines and other kernelbased learning methods. Cambridge: Cambridge University Press. Michael Collins and Nigel Duffy. 2001. Convolution Kernels for Natural Language. In Proceedings of NIPS-01. Declan Groves, Mary Hearne and Andy Way. 2004. Robust sub-sentential alignment of phrase-structure trees. In Proceedings of COLING-04, pages 10721078. Kenji Imamura. 2001. Hierarchical Phrase Alignment Harmonized with Parsing. In Proceedings of NLPRS-01, Tokyo. 377-384. Thorsten Joachims. 1999. Making large-scale SVM learning practical. In B. SchÄolkopf, C. Burges, and A. Smola, editors, Advances in Kernel Methods - Support Vector Learning, MIT press. Dan Klein and Christopher D. Manning. 2003. Accurate Unlexicalized Parsing. In Proceedings of ACL03. 423-430. Philipp Koehn, Franz Josef Och and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of HLT-NAACL-03. 48-54. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin and Evan Herbst. 2007. Moses: Open Source Toolkit for Statistical Machine Translation. In Proceedings of ACL-07. 177-180. Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1):19-51, March. Alessandro Moschitti. 2004. A Study on Convolution Kernels for Shallow Semantic Parsing. In Proceedings of ACL-04. Andreas Stolcke. 2002. SRILM - an extensible language modeling toolkit. In Proceedings of ICSLP-02. 901-904. Jun Sun, Min Zhang and Chew Lim Tan. 2009. A noncontiguous Tree Sequence Alignment-based Model for Statistical Machine Translation. In Proceedings of ACL-IJCNLP-09. 914-922. John Tinsley, Ventsislav Zhechev, Mary Hearne and Andy Way. 2007. Robust language pair-independent sub-tree alignment. In Proceedings of MT Summit XI -07. John Tinsley, Mary Hearne and Andy Way. 2009. Parallel treebanks in phrase-based statistical machine translation. In Proceedings of CICLING-09. Min Zhang, Jie Zhang, Jian Su and Guodong Zhou. 2006. A Composite Kernel to Extract Relations between Entities with both Flat and Structured Features. In Proceedings of ACL-COLING-06. 825832. Min Zhang, Hongfei Jiang, AiTi Aw, Jun Sun, Sheng Li and Chew Lim Tan. 2007. A tree-to-tree alignment-based model for statistical machine translation. In Proceedings of MT Summit XI -07. 535-542. Min Zhang, Hongfei Jiang, AiTi Aw, Haizhou Li, Chew Lim Tan and Sheng Li. 2008. A tree sequence alignment-based tree-to-tree translation model. In Proceedings of ACL-08. 559-567. 315
|
2010
|
32
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 316–324, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Discriminative Pruning for Discriminative ITG Alignment Shujie Liu†, Chi-Ho Li‡ and Ming Zhou‡ †School of Computer Science and Technology Harbin Institute of Technology, Harbin, China [email protected] ‡Microsoft Research Asia, Beijing, China {chl, mingzhou}@microsoft.com Abstract While Inversion Transduction Grammar (ITG) has regained more and more attention in recent years, it still suffers from the major obstacle of speed. We propose a discriminative ITG pruning framework using Minimum Error Rate Training and various features from previous work on ITG alignment. Experiment results show that it is superior to all existing heuristics in ITG pruning. On top of the pruning framework, we also propose a discriminative ITG alignment model using hierarchical phrase pairs, which improves both F-score and Bleu score over the baseline alignment system of GIZA++. 1 Introduction Inversion transduction grammar (ITG) (Wu, 1997) is an adaptation of SCFG to bilingual parsing. It does synchronous parsing of two languages with phrasal and word-level alignment as by-product. For this reason ITG has gained more and more attention recently in the word alignment community (Zhang and Gildea, 2005; Cherry and Lin, 2006; Haghighi et al., 2009). A major obstacle in ITG alignment is speed. The original (unsupervised) ITG algorithm has complexity of O(n6). When extended to supervised/discriminative framework, ITG runs even more slowly. Therefore all attempts to ITG alignment come with some pruning method. For example, Haghighi et al. (2009) do pruning based on the probabilities of links from a simpler alignment model (viz. HMM); Zhang and Gildea (2005) propose Tic-tac-toe pruning, which is based on the Model 1 probabilities of word pairs inside and outside a pair of spans. As all the principles behind these techniques have certain contribution in making good pruning decision, it is tempting to incorporate all these features in ITG pruning. In this paper, we propose a novel discriminative pruning framework for discriminative ITG. The pruning model uses no more training data than the discriminative ITG parser itself, and it uses a log-linear model to integrate all features that help identify the correct span pair (like Model 1 probability and HMM posterior). On top of the discriminative pruning method, we also propose a discriminative ITG alignment system using hierarchical phrase pairs. In the following, some basic details on the ITG formalism and ITG parsing are first reviewed (Sections 2 and 3), followed by the definition of pruning in ITG (Section 4). The “Discriminative Pruning for Discriminative ITG” model (DPDI) and our discriminative ITG (DITG) parsers will be elaborated in Sections 5 and 6 respectively. The merits of DPDI and DITG are illustrated with the experiments described in Section 7. 2 Basics of ITG The simplest formulation of ITG contains three types of rules: terminal unary rules 𝑋→𝑒/𝑓, where 𝑒 and 𝑓 represent words (possibly a null word, ε) in the English and foreign language respectively, and the binary rules 𝑋→ 𝑋, 𝑋 and 𝑋→ 𝑋, 𝑋 , which refer to that the component English and foreign phrases are combined in the same and inverted order respectively. From the viewpoint of word alignment, the terminal unary rules provide the links of word pairs, whereas the binary rules represent the reordering factor. One of the merits of ITG is that it is less biased towards short-distance reordering. Such a formulation has two drawbacks. First of all, it imposes a 1-to-1 constraint in word alignment. That is, a word is not allowed to align to more than one word. This is a strong limitation as no idiom or multi-word expression is allowed to align to a single word on the other side. In fact there have been various attempts in relaxing the 1-to-1 constraint. Both ITG alignment 316 approaches with and without this constraint will be elaborated in Section 6. Secondly, the simple ITG leads to redundancy if word alignment is the sole purpose of applying ITG. For instance, there are two parses for three consecutive word pairs, viz. [𝑎/𝑎’ [𝑏/𝑏’ 𝑐/ 𝑐’] ] and [[𝑎/𝑎’ 𝑏/𝑏’] 𝑐/𝑐’] . The problem of redundancy is fixed by adopting ITG normal form. In fact, normal form is the very first key to speeding up ITG. The ITG normal form grammar as used in this paper is described in Appendix A. 3 Basics of ITG Parsing Based on the rules in normal form, ITG word alignment is done in a similar way to chart parsing (Wu, 1997). The base step applies all relevant terminal unary rules to establish the links of word pairs. The word pairs are then combined into span pairs in all possible ways. Larger and larger span pairs are recursively built until the sentence pair is built. Figure 1(a) shows one possible derivation for a toy example sentence pair with three words in each sentence. Each node (rectangle) represents a pair, marked with certain phrase category, of foreign span (F-span) and English span (E-span) (the upper half of the rectangle) and the associated alignment hypothesis (the lower half). Each graph like Figure 1(a) shows only one derivation and also only one alignment hypothesis. The various derivations in ITG parsing can be compactly represented in hypergraph (Klein and Manning, 2001) like Figure 1(b). Each hypernode (rectangle) comprises both a span pair (upper half) and the list of possible alignment hypotheses (lower half) for that span pair. The hyperedges show how larger span pairs are derived from smaller span pairs. Note that a hypernode may have more than one alignment hypothesis, since a hypernode may be derived through more than one hyperedge (e.g. the topmost hypernode in Figure 1(b)). Due to the use of normal form, the hypotheses of a span pair are different from each other. 4 Pruning in ITG Parsing The ITG parsing framework has three levels of pruning: 1) To discard some unpromising span pairs; 2) To discard some unpromising F-spans and/or E-spans; 3) To discard some unpromising alignment hypotheses for a particular span pair. The second type of pruning (used in Zhang et. al. (2008)) is very radical as it implies discarding too many span pairs. It is empirically found to be highly harmful to alignment performance and therefore not adopted in this paper. The third type of pruning is equivalent to minimizing the beam size of alignment hypotheses in each hypernode. It is found to be well handled by the K-Best parsing method in Huang and Chiang (2005). That is, during the bottom-up construction of the span pair repertoire, each span pair keeps only the best alignment hypothesis. Once the complete parse tree is built, the k-best list of the topmost span is obtained by minimally expanding the list of alignment hypotheses of minimal number of span pairs. The first type of pruning is equivalent to minimizing the number of hypernodes in a hypergraph. The task of ITG pruning is defined in this paper as the first type of pruning; i.e. the search for, given an F-span, the minimal number of Espans which are the most likely counterpart of that F-span.1 The pruning method should maintain a balance between efficiency (run as quickly as possible) and performance (keep as many correct span pairs as possible). 1 Alternatively it can be defined as the search of the minimal number of E-spans per F-span. That is simply an arbitrary decision on how the data are organized in the ITG parser. B:[e1,e2]/[f1,f2] {e1/f2,e2/f1} C:[e1,e1]/[f2,f2] {e1/f2} C:[e2,e2]/[f1,f1] {e2/f1} C:[e3,e3]/[f3,f3] {e3/f3} A:[e1,e3]/[f1,f3] {e1/f2,e2/f1,e3/f3} (a) C:[e2,e2]/[f2,f2] {e2/f2} C:[e1,e1]/[f1,f1] {e1/f1} C:[e3,e3]/[f3,f3] {e3/f3} C:[e2,e2]/[f1,f1] {e2/f1} C:[e1,e1]/[f2,f2] {e1/f2} B:[e1,e2]/[f1,f2] {e1/f2} A:[e1,e2]/[f1,f2] {e2/f2} A:[e1,e3]/[f1,f3] {e1/f2,e2/f1,e3/f3} , {e1/f1,e2/f2,e3,f3} (b) B→<C,C> A→[C,C] A→[A,C] A→[B,C] Figure 1: Example ITG parses in graph (a) and hypergraph (b). 317 A naïve approach is that the required pruning method outputs a score given a span pair. This score is used to rank all E-spans for a particular F-span, and the score of the correct E-span should be in general higher than most of the incorrect ones. 5 The DPDI Framework DPDI, the discriminative pruning model proposed in this paper, assigns score to a span pair 𝑓 , 𝑒 as probability from a log-linear model: 𝑃 𝑒 𝑓 = 𝑒𝑥𝑝( 𝜆𝑖𝛹𝑖 𝑓 , 𝑒 𝑖 ) 𝑒𝑥𝑝( 𝜆𝑖𝛹𝑖(𝑓 , 𝑒 ′)) 𝑖 𝑒 ′∈𝐸 (1) where each 𝛹𝑖(𝑓 ,𝑒 ) is some feature about the span pair, and each 𝜆 is the weight of the corresponding feature. There are three major questions to this model: 1) How to acquire training samples? (Section 5.1) 2) How to train the parameters 𝜆 ? (Section 5.2) 3) What are the features? (Section 5.3) 5.1 Training Samples Discriminative approaches to word alignment use manually annotated alignment for sentence pairs. Discriminative pruning, however, handles not only a sentence pair but every possible span pair. The required training samples consist of various F-spans and their corresponding E-spans. Rather than recruiting annotators for marking span pairs, we modify the parsing algorithm in Section 3 so as to produce span pair annotation out of sentence-level annotation. In the base step, only the word pairs listed in sentence-level annotation are inserted in the hypergraph, and the recursive steps are just the same as usual. If the sentence-level annotation satisfies the alignment constraints of ITG, then each F-span will have only one E-span in the parse tree. However, in reality there are often the cases where a foreign word aligns to more than one English word. In such cases the F-span covering that foreign word has more than one corresponding Espans. Consider the example in Figure 2, where the golden links in the alignment annotation are 𝑒1/𝑓1, 𝑒2/𝑓1, and 𝑒3/𝑓2; i.e. the foreign word 𝑓1 aligns to both the English words 𝑒1 and 𝑒2. Therefore the F-span 𝑓1, 𝑓1 aligns to the Espan 𝑒1, 𝑒1 in one hypernode and to the E-span 𝑒2, 𝑒2 in another hypernode. When such situation happens, we calculate the product of the inside and outside probability of each alignment hypothesis of the span pair, based on the probabilities of the links from some simpler alignment model2. The E-span with the most probable hypothesis is selected as the alignment of the F-span. A→[C,C] Cw: [e1,e1]/[f1,f1] {e1/f1} Ce: [e1]/ε Cw: [e2,e2]/[f1,f1] Ce: [e2]/ε Cw: [e3,e3]/[f2,f2] C: [e1,e2]/[f1,f1] {e2/f1} C: [e2,e3]/[f2,f2] {e3/f2} A: [e1,e3]/[f1,f2] {e1/f1,e3/f2},{e2/f1,e3/f2} C→ [Ce,Cw] A→[C,C] C→ [Ce,Cw] {e1/f1} {e1/f1} (a) (b) [f1,f1] [e1,e1] [e1,e2] [e2,e2] [f2,f2] [e2,e3] [e3,e3] [f1,f2] [e1,e3] Figure 2: Training sample collection. Table (b) lists, for the hypergraph in (a), the candidate E-spans for each F-span. It should be noted that this automatic span pair annotation may violate some of the links in the original sentence-level alignment annotation. We have already seen how the 1-to-1 constraint in ITG leads to the violation. Another situation is the „inside-out‟ alignment pattern (c.f. Figure 3). The ITG reordering constraint cannot be satisfied unless one of the links in this pattern is removed. f1 f2 f3 f4 e1 e2 e3 e4 Figure 3: An example of inside-out alignment The training samples thus obtained are positive training samples. If we apply some classifier for parameter training, then negative samples are also needed. Fortunately, our parameter training does not rely on any negative samples. 5.2 MERT for Pruning Parameter training of DPDI is based on Minimum Error Rate Training (MERT) (Och, 2003), a widely used method in SMT. MERT for SMT estimates model parameters with the objective of minimizing certain measure of translation errors (or maximizing certain performance measure of translation quality) for a development corpus. Given an SMT system which produces, with 2 The formulae of the inside and outside probability of a span pair will be elaborated in Section 5.3. The simpler alignment model we used is HMM. 318 model parameters 𝜆1 𝑀, the K-best candidate translations 𝑒 (𝑓𝑠; 𝜆1 𝑀) for a source sentence 𝑓𝑠, and an error measure 𝐸(𝑟𝑠, 𝑒𝑠,𝑘) of a particular candidate 𝑒𝑠,𝑘 with respect to the reference translation 𝑟𝑠, the optimal parameter values will be: 𝜆 1 𝑀= 𝑎𝑟𝑔𝑚𝑖𝑛 𝜆1𝑀 𝐸 𝑟𝑠, 𝑒 𝑓𝑠; 𝜆1 𝑀 𝑆 𝑠=1 = 𝑎𝑟𝑔𝑚𝑖𝑛 𝜆1𝑀 𝐸 𝑟𝑠, 𝑒𝑠,𝑘 𝛿(𝑒 𝑓𝑠; 𝜆1 𝑀 , 𝑒𝑠,𝑘) 𝐾 𝑘=1 𝑆 𝑠=1 DPDI applies the same equation for parameter tuning, with different interpretation of the components in the equation. Instead of a development corpus with reference translations, we have a collection of training samples, each of which is a pair of F-span (𝑓𝑠) and its corresponding E-span (𝑟𝑠). These samples are acquired from some manually aligned dataset by the method elaborated in Section 5.1. The ITG parser outputs for each fs a K-best list of E-spans 𝑒 𝑓𝑠; 𝜆1 𝑀 based on the current parameter values 𝜆1 𝑀. The error function is based on the presence and the rank of the correct E-span in the K-best list: 𝐸 𝑟𝑠, 𝑒 𝑓𝑠; 𝜆1 𝑀 = −𝑟𝑎𝑛𝑘 𝑟𝑠 𝑖𝑓 𝑟𝑠∈𝑒 𝑓𝑠; 𝜆1 𝑀 𝑝𝑒𝑛𝑎𝑙𝑡𝑦 𝑜𝑡𝑒𝑟𝑤𝑖𝑠𝑒 (2) where 𝑟𝑎𝑛𝑘 𝑟𝑠 is the (0-based) rank of the correct E-span 𝑟𝑠 in the K-best list 𝑒 𝑓𝑠; 𝜆1 𝑀 . If 𝑟𝑠 is not in the K-best list at all, then the error is defined to be 𝑝𝑒𝑛𝑎𝑙𝑡𝑦, which is set as -100000 in our experiments. The rationale underlying this error function is to keep as many correct E-spans as possible in the K-best lists of E-spans, and push the correct E-spans upward as much as possible in the K-best lists. This new error measure leads to a change in details of the training algorithm. In MERT for SMT, the interval boundaries at which the performance or error measure changes are defined by the upper envelope (illustrated by the dash line in Figure 4(a)), since the performance/error measure depends on the best candidate translation. In MERT for DPDI, however, the error measure depends on the correct E-span rather than the E-span leading to the highest system score. Thus the interval boundaries are the intersections between the correct E-span and all other candidate E-spans (as shown in Figure 4(b)). The rank of the correct E-span in each interval can then be figured out as shown in Figure 4(c). Finally, the error measure in each interval can be calculated by Equation (2) (as shown in Figure 4(d)). All other steps in MERT for DPDI are the same as that for SMT. Σλmfm -index loss λk -8 -9 -10 -8 -9 -100,000 gold Σλmfm λk (a) (b) (c) (d) λk λk Figure 4: MERT for DPDI Part (a) shows how intervals are defined for SMT and part (b) for DPDI. Part (c) obtains the rank of correct E-spans in each interval and part (d) the error measure. Note that the beam size (max number of E-spans) for each F-span is 10. 5.3 Features The features used in DPDI are divided into three categories: 1) Model 1-based probabilities. Zhang and Gildea (2005) show that Model 1 (Brown et al., 1993; Och and Ney., 2000) probabilities of the word pairs inside and outside a span pair ( 𝑒𝑖1, 𝑒𝑖2 /[𝑓𝑗1, 𝑓𝑗2]) are useful. Hence these two features: a) Inside probability (i.e. probability of word pairs within the span pair): 𝑝𝑖𝑛𝑐 𝑒𝑖1,𝑖2 𝑓𝑗1,𝑗2 = 1 𝑗2 −𝑗1 𝑝𝑀1 𝑒𝑖 𝑓𝑗 𝑗∈ 𝑗1,𝑗2 𝑖∈ 𝑖1,𝑖2 b) Outside probability (i.e. probability of the word pairs outside the span pair): 𝑝𝑜𝑢𝑡 𝑒𝑖1,𝑖2 𝑓𝑗1,𝑗2 = 1 𝐽−𝑗2 + 𝑗1 𝑝𝑀1 𝑒𝑖 𝑓𝑗 𝑗∉ 𝑗1,𝑗2 𝑖∉ 𝑖1,𝑖2 where 𝐽 is the length of the foreign sentence. 2) Heuristics. There are four features in this category. The features are explained with the 319 example of Figure 5, in which the span pair in interest is 𝑒2, 𝑒3 /[𝑓1, 𝑓2]. The four links are produced by some simpler alignment model like HMM. The word pair 𝑒2/𝑓1 is the only link in the span pair. The links 𝑒4/𝑓2 and 𝑒3/𝑓3 are inconsistent with the span pair.3 f1 f2 f3 f4 e1 e2 e3 e4 Figure 5: Example for heuristic features a) Link ratio: 2×#𝑙𝑖𝑛𝑘𝑠 𝑓𝑙𝑒𝑛+𝑒𝑙𝑒𝑛 where #𝑙𝑖𝑛𝑘𝑠 is the number of links in the span pair, and 𝑓𝑙𝑒𝑛 and 𝑒𝑙𝑒𝑛 are the length of the foreign and English spans respectively. The feature value of the example span pair is (2*1)/(2+2)=0.5. b) inconsistent link ratio: 2×#𝑙𝑖𝑛𝑘𝑠𝑖𝑛𝑐𝑜𝑛 𝑓𝑙𝑒𝑛+𝑒𝑙𝑒𝑛 where #𝑙𝑖𝑛𝑘𝑠𝑖𝑛𝑐𝑜𝑛 is the number of links which are inconsistent with the phrase pair according to some simpler alignment model (e.g. HMM). The feature value of the example is (2*2)/(2+2) =1.0. c) Length ratio: 𝑓𝑙𝑒𝑛 𝑒𝑙𝑒𝑛−𝑟𝑎𝑡𝑖𝑜𝑎𝑣𝑔 where 𝑟𝑎𝑡𝑖𝑜𝑎𝑣𝑔 is defined as the average ratio of foreign sentence length to English sentence length, and it is estimated to be around 1.15 in our training dataset. The rationale underlying this feature is that the ratio of span length should not be too deviated from the average ratio of sentence length. The feature value for the example is |2/2-1.15|=0.15. d) Position Deviation: 𝑝𝑜𝑠𝑓 −𝑝𝑜𝑠𝑒 where 𝑝𝑜𝑠𝑓 refers to the position of the F-span in the entire foreign sentence, and it is defined as 1 2𝐽 𝑠𝑡𝑎𝑟𝑡𝑓 + 𝑒𝑛𝑑𝑓 , 𝑠𝑡𝑎𝑟𝑡𝑓 /𝑒𝑛𝑑𝑓 being the position of the first/last word of the F-span in the foreign sentence. 𝑝𝑜𝑠𝑒 is defined similarly. The rationale behind this feature is the monotonic assumption, i.e. a phrase of the foreign sentence usually occupies roughly the same position of the equivalent English phrase. The feature value for 3 An inconsistent link connects a word within the phrase pair to some word outside the phrase pair. C.f. Deng et al. (2008) the example is |(1+2)/(2*4)-(2+3)/(2*4)| =0.25. 3) HMM-based probabilities. Haghighi et al. (2009) show that posterior probabilities from the HMM alignment model is useful for pruning. Therefore, we design two new features by replacing the link count in link ratio and inconsistent link ratio with the sum of the link‟s posterior probability. 6 The DITG Models The discriminative ITG alignment can be conceived as a two-staged process. In the first stage DPDI selects good span pairs. In the second stage good alignment hypotheses are assigned to the span pairs selected by DPDI. Two discriminative ITG (DITG) models are investigated. One is word-to-word DITG (henceforth W-DITG), which observes the 1-to-1 constraint on alignment. Another is DITG with hierarchical phrase pairs (henceforth HP-DITG), which relaxes the 1to-1 constraint by adopting hierarchical phrase pairs in Chiang (2007). Each model selects the best alignment hypotheses of each span pair, given a set of features. The contributions of these features are integrated through a log linear model (similar to Liu et al., 2005; Moore, 2005) like Equation (1). The discriminative training of the feature weights is again MERT (Och, 2003). The MERT module for DITG takes alignment F-score of a sentence pair as the performance measure. Given an input sentence pair and the reference annotated alignment, MERT aims to maximize the F-score of DITG-produced alignment. Like SMT (and unlike DPDI), it is the upper envelope which defines the intervals where the performance measure changes. 6.1 Word-to-word DITG The following features about alignment link are used in W-DITG: 1) Word pair translation probabilities trained from HMM model (Vogel, et.al., 1996) and IBM model 4 (Brown et.al., 1993; Och and Ney, 2000). 2) Conditional link probability (Moore, 2005). 3) Association score rank features (Moore et al., 2006). 4) Distortion features: counts of inversion and concatenation. 5) Difference between the relative positions of the words. The relative position of a word in a sentence is defined as the posi320 tion of the word divided by sentence length. 6) Boolean features like whether a word in the word pair is a stop word. 6.2 DITG with Hierarchical Phrase Pairs The 1-to-1 assumption in ITG is a serious limitation as in reality there are always segmentation or tokenization errors as well as idiomatic expressions. Wu (1997) proposes a bilingual segmentation grammar extending the terminal rules by including phrase pairs. Cherry and Lin (2007) incorporate phrase pairs in phrase-based SMT into ITG, and Haghighi et al. (2009) introduce Block ITG (BITG), which adds 1-to-many or many-to-1 terminal unary rules. It is interesting to see if DPDI can benefit the parsing of a more realistic ITG. HP-DITG extends Cherry and Lin‟s approach by not only employing simple phrase pairs but also hierarchical phrase pairs (Chiang, 2007). The grammar is enriched with rules of the format: 𝑋𝑒 𝑖/𝑓 𝑖 where 𝑒 𝑖 and 𝑓 𝑖 refer to the English and foreign side of the i-th (simple/hierarchical) phrase pair respectively. As example, if there is a simple phrase pair 𝑋 𝑁𝑜𝑟𝑡 𝐾𝑜𝑟𝑒𝑎, 北 朝鲜 , then it is transformed into the ITG rule 𝐶"North Korea"/ "北 朝鲜". During parsing, each span pair does not only examine all possible combinations of sub-span pairs using binary rules, but also checks if the yield of that span pair is exactly the same as that phrase pair. If so, then the alignment links within the phrase pair (which are obtained in standard phrase pair extraction procedure) are taken as an alternative alignment hypothesis of that span pair. For a hierarchical phrase pair like 𝑋 𝑋1 𝑜𝑓 𝑋2, 𝑋2 的 𝑋1 , it is transformed into the ITG rule 𝐶"𝑋1 𝑜𝑓 𝑋2"/"𝑋2 的 𝑋1" during parsing, each span pair checks if it contains the lexical anchors "of" and "的", and if the remaining words in its yield can form two sub-span pairs which fit the reordering constraint among 𝑋1 and 𝑋2. (Note that span pairs of any category in the ITG normal form grammar can substitute for 𝑋1or 𝑋2.) If both conditions hold, then the span pair is assigned an alignment hypothesis which combines the alignment links among the lexical anchors (𝑙𝑖𝑘𝑒 𝑜𝑓/的) and those links among the sub-span pairs. HP-ITG acquires the rules from HMM-based word-aligned corpus using standard phrase pair extraction as stated in Chiang (2007). The rule probabilities and lexical weights in both Englishto-foreign and foreign-to-English directions are estimated and taken as features, in addition to those features in W-DITG, in the discriminative model of alignment hypothesis selection. 7 Evaluation DPDI is evaluated against the baselines of Tictac-toe (TTT) pruning (Zhang and Gildea, 2005) and Dynamic Program (DP) pruning (Haghighi et al., 2009; DeNero et al., 2009) with respect to Chinese-to-English alignment and translation. Based on DPDI, HP-DITG is evaluated against the alignment systems GIZA++ and BITG. 7.1 Evaluation Criteria Four evaluation criteria are used in addition to the time spent on ITG parsing. We will first evaluate pruning regarding the pruning decisions themselves. That is, the first evaluation metric, pruning error rate (henceforth PER), measures how many correct E-spans are discarded. The major drawback of PER is that not all decisions in pruning would impact on alignment quality, since certain F-spans are of little use to the entire ITG parse tree. An alternative criterion is the upper bound on alignment F-score, which essentially measures how many links in annotated alignment can be kept in ITG parse. The calculation of F-score upper bound is done in a bottom-up way like ITG parsing. All leaf hypernodes which contain a correct link are assigned a score (known as hit) of 1. The hit of a non-leaf hypernode is based on the sum of hits of its daughter hypernodes. The maximal sum among all hyperedges of a hypernode is assigned to that hypernode. Formally, 𝑖𝑡 𝑋 𝑓 , 𝑒 = 𝑚𝑎𝑥 𝑌,𝑍,𝑓 1,𝑒 1,𝑓 2,𝑒 2 (𝑖𝑡 𝑌 𝑓 1, 𝑒 1 + 𝑖𝑡[𝑓 2, 𝑒 2]) 𝑖𝑡 𝐶𝑤 𝑢, 𝑣 = 1 𝑖𝑓 𝑢, 𝑣 ∈𝑅 0 𝑜𝑡𝑒𝑟𝑤𝑖𝑠𝑒 𝑖𝑡 𝐶𝑒 = 0; 𝑖𝑡 𝐶𝑓 = 0 where 𝑋, 𝑌, 𝑍 are variables for the categories in ITG grammar, and 𝑅 comprises the golden links in annotated alignment. 𝐶𝑤, 𝐶𝑒, 𝐶𝑓 are defined in Appendix A. Figure 6 illustrates the calculation of the hit score for the example in Section 5.1/Figure 2. The upper bound of recall is the hit score divided by the total number of golden links. The upper 321 ID pruning beam size pruning/total time cost PER F-UB F-score 1 DPDI 10 72‟‟/3‟03‟‟ 4.9% 88.5% 82.5% 2 TTT 10 58’’/2’38’’ 8.6% 87.5% 81.1% 3 TTT 20 53‟‟/6‟55‟‟ 5.2% 88.6% 82.4% 4 DP -- 11‟‟/6‟01‟‟ 12.1% 86.1% 80.5% Table 1: Evaluation of DPDI against TTT (Tic-tac-toe) and DP (Dynamic Program) for W-DITG ID pruning beam size pruning/total time cost PER F-UB F-score 1 DPDI 10 72‟‟/5‟18‟‟ 4.9% 93.9% 87.0% 2 TTT 10 58’’/4’51’’ 8.6% 93.0% 84.8% 3 TTT 20 53‟‟/12‟5‟‟ 5.2% 94.0% 86.5% 4 DP -- 11‟‟/15‟39‟‟ 12.1% 91.4% 83.6% Table 2: Evaluation of DPDI against TTT (Tic-tac-toe) and DP (Dynamic Program) for HP-DITG. bound of precision, which should be defined as the hit score divided by the number of links produced by the system, is almost always 1.0 in practice. The upper bound of alignment F-score can thus be calculated as well. A→[C,C] Cw: [e1,e1]/[f1,f1] hit=1 Ce: [e1]/ε Cw: [e2,e2]/[f1,f1] Ce: [e2]/ε Cw: [e3,e3]/[f2,f2] C: [e1,e2]/[f1,f1] hit=max{0+1}=1 C: [e2,e3]/[f2,f2] hit=max{0+1}=1 A: [e1,e3]/[f1,f2] hit=max{1+1,1+1}=2 C→ [Ce,Cw] A→[C,C] C→ [Ce,Cw] hit=1 hit=1 hit=0 hit=0 Figure 6: Recall Upper Bound Calculation Finally, we also do end-to-end evaluation using both F-score in alignment and Bleu score in translation. We use our implementation of hierarchical phrase-based SMT (Chiang, 2007), with standard features, for the SMT experiments. 7.2 Experiment Data Both discriminative pruning and alignment need training data and test data. We use the manually aligned Chinese-English dataset as used in Haghighi et al. (2009). The 491 sentence pairs in this dataset are adapted to our own Chinese word segmentation standard. 250 sentence pairs are used as training data and the other 241 are test data. The corresponding numbers of F-spans in training and test data are 4590 and 3951 respectively. In SMT experiments, the bilingual training dataset is the NIST training set excluding the Hong Kong Law and Hong Kong Hansard, and our 5gram language model is trained from the Xinhua section of the Gigaword corpus. The NIST‟03 test set is used as our development corpus and the NIST‟05 and NIST‟08 test sets are our test sets. 7.3 Small-scale Evaluation The first set of experiments evaluates the performance of the three pruning methods using the small 241-sentence set. Each pruning method is plugged in both W-DITG and HP-DITG. IBM Model 1 and HMM alignment model are reimplemented as they are required by the three ITG pruning methods. The results for W-DITG are listed in Table 1. Tests 1 and 2 show that with the same beam size (i.e. number of E-spans per F-span), although DPDI spends a bit more time (due to the more complicated model), DPDI makes far less incorrect pruning decisions than the TTT. In terms of F-score upper bound, DPDI is 1 percent higher. DPDI achieves even larger improvement in actual F-score. To enable TTT achieving similar F-score or Fscore upper bound, the beam size has to be doubled and the time cost is more than twice the original (c.f. Tests 1 and 3 in Table 1) . The DP pruning in Haghighi et.al. (2009) performs much poorer than the other two pruning methods. In fact, we fail to enable DP achieve the same F-score upper bound as the other two methods before DP leads to intolerable memory consumption. This may be due to the use of different HMM model implementations between our work and Haghighi et.al. (2009). Table 2 lists the results for HP-DITG. Roughly the same observation as in W-DITG can be made. In addition to the superiority of DPDI, it can also be noted that HP-DITG achieves much higher Fscore and F-score upper bound. This shows that 322 hierarchical phrase is a powerful tool in rectifying the 1-to-1 constraint in ITG. Note also that while TTT in Test 3 gets roughly the same F-score upper bound as DPDI in Test 1, the corresponding F-score is slightly worse. A possible explanation is that better pruning not only speeds up the parsing/alignment process but also guides the search process to focus on the most promising region of the search space. 7.4 Large-scale End-to-End Experiment ID Pruning beam size time cost Bleu05 Bleu08 1 DPDI 10 1092h 38.57 28.31 2 TTT 10 972h 37.96 27.37 3 TTT 20 2376h 38.13 27.58 4 DP -- 2068h 37.43 27.12 Table 3: Evaluation of DPDI against TTT and DP for HP-DITG ID WAModel F-Score Bleu-05 Bleu-08 1 HMM 80.1% 36.91 26.86 2 Giza++ 84.2% 37.70 27.33 3 BITG 85.9% 37.92 27.85 4 HP-DITG 87.0% 38.57 28.31 Table 4: Evaluation of DPDI against HMM, Giza++ and BITG Table 3 lists the word alignment time cost and SMT performance of different pruning methods. HP-DITG using DPDI achieves the best Bleu score with acceptable time cost. Table 4 compares HP-DITG to HMM (Vogel, et al., 1996), GIZA++ (Och and Ney, 2000) and BITG (Haghighi et al., 2009). It shows that HP-DITG (with DPDI) is better than the three baselines both in alignment F-score and Bleu score. Note that the Bleu score differences between HP-DITG and the three baselines are statistically significant (Koehn, 2004). An explanation of the better performance by HP-DITG is the better phrase pair extraction due to DPDI. On the one hand, a good phrase pair often fails to be extracted due to a link inconsistent with the pair. On the other hand, ITG pruning can be considered as phrase pair selection, and good ITG pruning like DPDI guides the subsequent ITG alignment process so that less links inconsistent to good phrase pairs are produced. This also explains (in Tables 2 and 3) why DPDI with beam size 10 leads to higher Bleu than TTT with beam size 20, even though both pruning methods lead to roughly the same alignment F-score. 8 Conclusion and Future Work This paper reviews word alignment through ITG parsing, and clarifies the problem of ITG pruning. A discriminative pruning model and two discriminative ITG alignments systems are proposed. The pruning model is shown to be superior to all existing ITG pruning methods, and the HP-DITG alignment system is shown to improve state-ofthe-art alignment and translation quality. The current DPDI model employs a very limited set of features. Many features are related only to probabilities of word pairs. As the success of HP-DITG illustrates the merit of hierarchical phrase pair, in future we should investigate more features on the relationship between span pair and hierarchical phrase pair. Appendix A. The Normal Form Grammar Table 5 lists the ITG rules in normal form as used in this paper, which extend the normal form in Wu (1997) so as to handle the case of alignment to null. 1 𝑆 →𝐴|𝐵|𝐶 2 𝐴 → 𝐴 𝐵 | 𝐴 𝐶 | 𝐵 𝐵 | 𝐵𝐶 | 𝐶 𝐵 | 𝐶 𝐶 3 𝐵 → 𝐴 𝐴 | 𝐴 𝐶 | 𝐵 𝐴 | 𝐵 𝐶 𝐵 → 𝐶 𝐴 | 𝐶 𝐶 4 𝐶 →𝐶𝑤|𝐶𝑓𝑤|𝐶𝑒𝑤 5 𝐶 → 𝐶𝑒𝑤 𝐶𝑓𝑤 6 𝐶𝑤 →𝑢/𝑣 7 𝐶𝑒 →𝜀/𝑣; 𝐶𝑓→𝑢/𝜀 8 𝐶𝑒𝑚→𝐶𝑒| 𝐶𝑒𝑚 𝐶𝑒 ; 𝐶𝑓𝑚→𝐶𝑓| 𝐶𝑓𝑚 𝐶𝑓 9 𝐶𝑒𝑤→ 𝐶𝑒𝑚 𝐶𝑤 ; 𝐶𝑓𝑤→ 𝐶𝑓𝑚 𝐶𝑤 Table 5: ITG Rules in Normal Form In these rules, 𝑆 is the Start symbol; 𝐴 is the category for concatenating combination whereas 𝐵 for inverted combination. Rules (2) and (3) are inherited from Wu (1997). Rules (4) divide the terminal category 𝐶 into subcategories. Rule schema (6) subsumes all terminal unary rules for some English word 𝑢 and foreign word 𝑣, and rule schemas (7) are unary rules for alignment to null. Rules (8) ensure all words linked to null are combined in left branching manner, while rules (9) ensure those words linked to null combine with some following, rather than preceding, word pair. (Note: Accordingly, all sentences must be ended by a special token 𝑒𝑛𝑑 , otherwise the last word(s) of a sentence cannot be linked to null.) If there are both English and foreign words linked to null, rule (5) ensures that those English 323 words linked to null precede those foreign words linked to null. References Peter F. Brown, Stephen A. Della Pietra, Vincent J. Della Peitra, Robert L. Mercer. 1993. The Mathematics of Statistical Machine Translation: Parameter Estimation. Computational Linguistics, 19(2):263-311. Colin Cherry and Dekang Lin. 2006. Soft Syntactic Constraints for Word Alignment through Discriminative Training. In Proceedings of ACLCOLING. Colin Cherry and Dekang Lin. 2007. Inversion Transduction Grammar for Joint Phrasal Translation Modeling. In Proceedings of SSST, NAACL-HLT, Pages:17-24. David Chiang. 2007. Hierarchical Phrase-based Translation. Computational Linguistics, 33(2). John DeNero, Mohit Bansal, Adam Pauls, and Dan Klein. 2009. Efficient Parsing for Transducer Grammars. In Proceedings of NAACL, Pages:227-235. Alexander Fraser and Daniel Marcu. 2006. SemiSupervised Training for StatisticalWord Alignment. In Proceedings of ACL, Pages:769776. Aria Haghighi, John Blitzer, John DeNero, and Dan Klein. 2009. Better Word Alignments with Supervised ITG Models. In Proceedings of ACL, Pages: 923-931. Liang Huang and David Chiang. 2005. Better k-best Parsing. In Proceedings of IWPT 2005, Pages:173-180. Franz Josef Och and Hermann Ney. 2000. Improved statistical alignment models. In Proceedings of ACL. Pages: 440-447 Franz Josef Och. 2003. Minimum error rate training in statistical machine translation. In Proceedings of ACL, Pages:160-167. Dan Klein and Christopher D. Manning. 2001. Parsing and Hypergraphs. In Proceedings of IWPT, Pages:17-19 Philipp Koehn. 2004. Statistical Significance Tests for Machine Translation Evaluation. In Proceedings of EMNLP, Pages: 388-395. Yang Liu, Qun Liu and Shouxun Lin. 2005. Loglinear models for word alignment. In Proceedings of ACL, Pages: 81-88. Robert Moore. 2005. A Discriminative Framework for Bilingual Word Alignment. In Proceedings of EMNLP 2005, Pages: 81-88. Robert Moore, Wen-tau Yih, and Andreas Bode. 2006. Improved Discriminative Bilingual Word Alignment. In Proceedings of ACL, Pages: 513520. Stephan Vogel, Hermann Ney, and Christoph Tillmann. 1996. HMM-based word alignment in statistical translation. In Proceedings of COLING, Pages: 836-841. Stephan Vogel. 2005. PESA: Phrase Pair Extraction as Sentence Splitting. In Proceedings of MT Summit. Dekai Wu. 1997. Stochastic Inversion Transduction Grammars and Bilingual Parsing of Parallel Corpora. Computational Linguistics, 23(3). Hao Zhang and Daniel Gildea. 2005. Stochastic Lexicalized Inversion Transduction Grammar for Alignment. In Proceedings of ACL. Hao Zhang, Chris Quirk, Robert Moore, and Daniel Gildea. 2008. Bayesian learning of noncompositional phrases with synchronous parsing. In Proceedings of ACL, Pages: 314-323. 324
|
2010
|
33
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 325–334, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Fine-grained Tree-to-String Translation Rule Extraction Xianchao Wu† Takuya Matsuzaki† Jun’ichi Tsujii†‡∗ †Department of Computer Science, The University of Tokyo 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-0033, Japan ‡School of Computer Science, University of Manchester ∗National Centre for Text Mining (NaCTeM) Manchester Interdisciplinary Biocentre, 131 Princess Street, Manchester M1 7DN, UK {wxc, matuzaki, tsujii}@is.s.u-tokyo.ac.jp Abstract Tree-to-string translation rules are widely used in linguistically syntax-based statistical machine translation systems. In this paper, we propose to use deep syntactic information for obtaining fine-grained translation rules. A head-driven phrase structure grammar (HPSG) parser is used to obtain the deep syntactic information, which includes a fine-grained description of the syntactic property and a semantic representation of a sentence. We extract fine-grained rules from aligned HPSG tree/forest-string pairs and use them in our tree-to-string and string-to-tree systems. Extensive experiments on largescale bidirectional Japanese-English translations testified the effectiveness of our approach. 1 Introduction Tree-to-string translation rules are generic and applicable to numerous linguistically syntax-based Statistical Machine Translation (SMT) systems, such as string-to-tree translation (Galley et al., 2004; Galley et al., 2006; Chiang et al., 2009), tree-to-string translation (Liu et al., 2006; Huang et al., 2006), and forest-to-string translation (Mi et al., 2008; Mi and Huang, 2008). The algorithms proposed by Galley et al. (2004; 2006) are frequently used for extracting minimal and composed rules from aligned 1-best tree-string pairs. Dealing with the parse error problem and rule sparseness problem, Mi and Huang (2008) replaced the 1-best parse tree with a packed forest which compactly encodes exponentially many parses for treeto-string rule extraction. However, current tree-to-string rules only make use of Probabilistic Context-Free Grammar tree fragments, in which part-of-speech (POS) or koroshita korosareta (active) (passive) VBN(killed) 6 (6/10,6/6) 4 (4/10,4/4) VBN(killed:active) 5 (5/6,5/6) 1 (1/6,1/4) VBN(killed:passive) 1 (1/4,1/6) 3 (3/4,3/4) Table 1: Bidirectional translation probabilities of rules, denoted in the brackets, change when voice is attached to “killed”. phrasal tags are used as the tree node labels. As will be testified by our experiments, we argue that the simple POS/phrasal tags are too coarse to reflect the accurate translation probabilities of the translation rules. For example, as shown in Table 1, suppose a simple tree fragment “VBN(killed)” appears 6 times with “koroshita”, which is a Japanese translation of an active form of “killed”, and 4 times with “korosareta”, which is a Japanese translation of a passive form of “killed”. Then, without larger tree fragments, we will more frequently translate “VBN(killed)” into “koroshita” (with a probability of 0.6). But, “VBN(killed)” is indeed separable into two finegrained tree fragments of “VBN(killed:active)” and “VBN(killed:passive)”1. Consequently, “VBN(killed:active)” appears 5 times with “koroshita” and 1 time with “korosareta”; and “VBN(killed:passive)” appears 1 time with “koroshita” and 3 times with “korosareta”. Now, by attaching the voice information to “killed”, we are gaining a rule set that is more appropriate to reflect the real translation situations. This motivates our proposal of using deep syntactic information to obtain a fine-grained translation rule set. We name the information such as the voice of a verb in a tree fragment as deep syntactic information. We use a head-driven phrase structure grammar (HPSG) parser to obtain the 1For example, “John has killed Mary.” versus “John was killed by Mary.” 325 deep syntactic information of an English sentence, which includes a fine-grained description of the syntactic property and a semantic representation of the sentence. We extract fine-grained translation rules from aligned HPSG tree/forest-string pairs. We localize an HPSG tree/forest to make it segmentable at any nodes to fit the extraction algorithms described in (Galley et al., 2006; Mi and Huang, 2008). We also propose a linear-time algorithm for extracting composed rules guided by predicate-argument structures. The effectiveness of the rules are testified in our tree-to-string and string-to-tree systems, taking bidirectional Japanese-English translations as our test cases. This paper is organized as follows. In Section 2, we briefly review the tree-to-string and string-totree translation frameworks, tree-to-string rule extraction algorithms, and rich syntactic information previously used for SMT. The HPSG grammar and our proposal of fine-grained rule extraction algorithms are described in Section 3. Section 4 gives the experiments for applying fine-grained translation rules to large-scale Japanese-English translation tasks. Finally, we conclude in Section 5. 2 Related Work 2.1 Tree-to-string and string-to-tree translations Tree-to-string translation (Liu et al., 2006; Huang et al., 2006) first uses a parser to parse a source sentence into a 1-best tree and then searches for the best derivation that segments and converts the tree into a target string. In contrast, string-to-tree translation (Galley et al., 2004; Galley et al., 2006; Chiang et al., 2009) is like bilingual parsing. That is, giving a (bilingual) translation grammar and a source sentence, we are trying to construct a parse forest in the target language. Consequently, the translation results can be collected from the leaves of the parse forest. Figure 1 illustrates the training and decoding processes of bidirectional Japanese-English translations. The English sentence is “John killed Mary” and the Japanese sentence is “jyon ha mari wo koroshita”, in which the function words “ha” and “wo” are not aligned with any English word. 2.2 Tree/forest-based rule extraction Galley et al. (2004) proposed the GHKM algorithm for extracting (minimal) tree-to-string translation rules from a tuple of ⟨F, Et, A⟩, where F = x0 は x1 x0 x1 x1 をx0 NP John ジョン V killed 殺した NP Mary マリー NP V NP VP S John killed Mary ジョン は マリー を 殺した NP VP S V NP VP x0 x1 Training Aligned tree-string pair: Extract rules John killed Mary ジョン は マリー を 殺した CKY decoding Testing NP V NP VP S John killed Mary NP VP V NP Apply rules …… jyon ha mari wo koroshita parsing Bottom-up decoding tree-to-string string-to-tree Figure 1: Illustration of the training and decoding processes for tree-to-string and string-to-tree translations. fJ 1 is a sentence of a foreign language other than English, Et is a 1-best parse tree of an English sentence E = eI 1, and A = {(j, i)} is an alignment between the words in F and E. The basic idea of GHKM algorithm is to decompose Et into a series of tree fragments, each of which will form a rule with its corresponding translation in the foreign language. A is used as a constraint to guide the segmentation procedure, so that the root node of every tree fragment of Et exactly corresponds to a contiguous span on the foreign language side. Based on this consideration, a frontier set (fs) is defined to be a set of nodes n in Et that satisfies the following constraint: fs = {n|span(n) ∩comp span(n) = ϕ}. (1) Here, span(n) is defined by the indices of the first and last word in F that are reachable from a node n, and comp span(n) is defined to be the complement set of span(n), i.e., the union of the spans of all nodes n′ in Et that are neither descendants nor ancestors of n. span(n) and comp span(n) of each n can be computed by first a bottom-up exploration and then a top-down traversal of Et. By restricting each fragment so that it only takes 326 John CAT N POS NNP BASE john LEXENTRY [D< N.3sg>]_lxm PRED noun_arg0 t0 HEAD t0 SEM_HEAD t0 CAT NX XCAT c2 killed CAT V POS VBD BASE kill LEXENTRY [NP.nom <V.bse> NP.acc] _lxm-past_verb_rule PRED verb_arg12 TENSE past ASPECT none VOICE active AUX minus ARG1 c1 ARG2 c5 t1 HEAD t1 SEM_HEAD t1 CAT VX XCAT c4 HEAD c6 SEM_HEAD c6 CAT NP XCAT SCHEMA empty_spec_head c5 HEAD t2 SEM_HEAD t2 CAT NX XCAT c6 HEAD c3 SEM_HEAD c3 CAT S XCAT SCHEMA subj_head c0 HEAD c2 SEM_HEAD c2 CAT NP XCAT SCHEMA empty_spec_head c1 HEAD c4 SEM_HEAD c4 CAT VP XCAT SCHEMA head_comp c3 Mary CAT N POS NNP BASE mary LEXENTRY [D<N.3sg>]_lxm PRED noun_arg0 t2 ジョン は マリー を 殺した 1. c0(x0:c1, x1:c3) x0 は x1 2. c1(x0:c2) x0 3. c2(t0) ジョン 4. c3(x0:c4, x1:c5) x1 を x0 5. c4(t1) 殺した 6. c5(x0:c6) x0 7. c6(t2) マリー c0 c1 c3 c4 c5 t1 minimum covering tree x0 は x1 を 殺した An HPSG-tree based minimal rule set A PAS-based composed rule John killed Mary HEAD c8 SEM_HEAD c8 CAT S XCAT SCHEMA head_mod c7 HEAD c9 SEM_HEAD c9 CAT S XCAT SCHEMA subj_head c8 killed CAT V POS VBD BASE kill LEXENTRY [NP.nom <V.bse>]_lxmpast_verb_rule PRED verb_arg1 TENSE past ASPECT none VOICE active AUX minus ARG1 c1 t3 HEAD t3 SEM_HEAD t3 CAT VP XCAT c9 HEAD c11 SEM_HEAD c11 CAT NP XCAT SCHEMA empty_spec_head c10 HEAD t4 SEM_HEAD t4 CAT NX XCAT c11 Mary CAT N POS NNP BASE mary LEXENTRY V[D<N.3sg>] PRED noun_arg0 t4 2.77 4.52 0.81 2.25 0 0.00 -3.47 -0.03 0 -2.82 -0.07 -0.001 Figure 2: Illustration of an aligned HPSG forest-string pair. The forest includes two parse trees by taking “Mary” as a modifier (t3, t4) or an argument (t1, t2) of “killed”. Arrows with broken lines denote the PAS dependencies from the terminal node t1 to its argument nodes (c1 and c5). The scores of the hyperedges are attached to the forest as well. the nodes in fs as the root and leaf nodes, a wellformed fragmentation of Et is generated. With fs computed, rules are extracted through a depthfirst traversal of Et: we cut Et at all nodes in fs to form tree fragments and extract a rule for each fragment. These extracted rules are called minimal rules (Galley et al., 2004). For example, the 1best tree (with gray nodes) in Figure 2 is cut into 7 pieces, each of which corresponds to the tree fragment in a rule (bottom-left corner of the figure). In order to include richer context information and account for multiple interpretations of unaligned words of foreign language, minimal rules which share adjacent tree fragments are connected together to form composed rules (Galley et al., 2006). For each aligned tree-string pair, Galley et al. (2006) constructed a derivation-forest, in which composed rules were generated, unaligned words of foreign language were consistently attached, and the translation probabilities of rules were estimated by using ExpectationMaximization (EM) (Dempster et al., 1977) training. For example, by combining the minimal rules of 1, 4, and 5, we obtain a composed rule, as shown in the bottom-right corner of Figure 2. Considering the parse error problem in the 1-best or k-best parse trees, Mi and Huang (2008) extracted tree-to-string translation rules from aligned packed forest-string pairs. A forest compactly encodes exponentially many trees 327 rather than the 1-best tree used by Galley et al. (2004; 2006). Two problems were managed to be tackled during extracting rules from an aligned forest-string pair: where to cut and how to cut. Equation 1 was used again to compute a frontier node set to determine where to cut the packed forest into a number of tree-fragments. The difference with tree-based rule extraction is that the nodes in a packed forest (which is a hypergraph) now are hypernodes, which can take a set of incoming hyperedges. Then, by limiting each fragment to be a tree and whose root/leaf hypernodes all appearing in the frontier set, the packed forest can be segmented properly into a set of tree fragments, each of which can be used to generate a tree-to-string translation rule. 2.3 Rich syntactic information for SMT Before describing our approaches of applying deep syntactic information yielded by an HPSG parser for fine-grained rule extraction, we would like to briefly review what kinds of deep syntactic information have been employed for SMT. Two kinds of supertags, from Lexicalized TreeAdjoining Grammar and Combinatory Categorial Grammar (CCG), have been used as lexical syntactic descriptions (Hassan et al., 2007) for phrasebased SMT (Koehn et al., 2007). By introducing supertags into the target language side, i.e., the target language model and the target side of the phrase table, significant improvement was achieved for Arabic-to-English translation. Birch et al. (2007) also reported a significant improvement for Dutch-English translation by applying CCG supertags at a word level to a factorized SMT system (Koehn et al., 2007). In this paper, we also make use of supertags on the English language side. In an HPSG parse tree, these lexical syntactic descriptions are included in the LEXENTRY feature (refer to Table 2) of a lexical node (Matsuzaki et al., 2007). For example, the LEXENTRY feature of “t1:killed” takes the value of [NP.nom<V.bse>NP.acc]_lxm-past _verb_rule in Figure 2. In which, [NP.nom<V.bse>NP.acc] is an HPSG style supertag, which tells us that the base form of “killed” needs a nominative NP in the left hand side and an accessorial NP in the right hand side. The major differences are that, we use a larger feature set (Table 2) including the supertags for fine-grained tree-to-string rule extraction, rather than string-to-string translation (Hassan et al., 2007; Birch et al., 2007). The Logon project2 (Oepen et al., 2007) for Norwegian-English translation integrates in-depth grammatical analysis of Norwegian (using lexical functional grammar, similar to (Riezler and Maxwell, 2006)) with semantic representations in the minimal recursion semantics framework, and fully grammar-based generation for English using HPSG. A hybrid (of rule-based and data-driven) architecture with a semantic transfer backbone is taken as the vantage point of this project. In contrast, the fine-grained tree-to-string translation rule extraction approaches in this paper are totally data-driven, and easily applicable to numerous language pairs by taking English as the source or target language. 3 Fine-grained rule extraction We now introduce the deep syntactic information generated by an HPSG parser and then describe our approaches for fine-grained tree-tostring rule extraction. Especially, we localize an HPSG tree/forest to fit the extraction algorithms described in (Galley et al., 2006; Mi and Huang, 2008). Also, we propose a linear-time composed rule extraction algorithm by making use of predicate-argument structures. 3.1 Deep syntactic information by HPSG parsing Head-driven phrase structure grammar (HPSG) is a lexicalist grammar framework. In HPSG, linguistic entities such as words and phrases are represented by a data structure called a sign. A sign gives a factored representation of the syntactic features of a word/phrase, as well as a representation of their semantic content. Phrases and words represented by signs are composed into larger phrases by applications of schemata. The semantic representation of the new phrase is calculated at the same time. As such, an HPSG parse tree/forest can be considered as a tree/forest of signs (c.f. the HPSG forest in Figure 2). An HPSG parse tree/forest has two attractive properties as a representation of an English sentence in syntax-based SMT. First, we can carefully control the condition of the application of a translation rule by exploiting the fine-grained syntactic 2http://www.emmtee.net/ 328 Feature Description CAT phrasal category XCAT fine-grained phrasal category SCHEMA name of the schema applied in the node HEAD pointer to the head daughter SEM HEAD pointer to the semantic head daughter CAT syntactic category POS Penn Treebank-style part-of-speech tag BASE base form TENSE tense of a verb (past, present, untensed) ASPECT aspect of a verb (none, perfect, progressive, perfect-progressive) VOICE voice of a verb (passive, active) AUX auxiliary verb or not (minus, modal, have, be, do, to, copular) LEXENTRY lexical entry, with supertags embedded PRED type of a predicate ARG⟨x⟩ pointer to semantic arguments, x = 1..4 Table 2: Syntactic/semantic features extracted from HPSG signs that are included in the output of Enju. Features in phrasal nodes (top) and lexical nodes (bottom) are listed separately. description in the English parse tree/forest, as well as those in the translation rules. Second, we can identify sub-trees in a parse tree/forest that correspond to basic units of the semantics, namely sub-trees covering a predicate and its arguments, by using the semantic representation given in the signs. We expect that extraction of translation rules based on such semantically-connected subtrees will give a compact and effective set of translation rules. A sign in the HPSG tree/forest is represented by a typed feature structure (TFS) (Carpenter, 1992). A TFS is a directed-acyclic graph (DAG) wherein the edges are labeled with feature names and the nodes (feature values) are typed. In the original HPSG formalism, the types are defined in a hierarchy and the DAG can have arbitrary shape (e.g., it can be of any depth). We however use a simplified form of TFS, for simplicity of the algorithms. In the simplified form, a TFS is converted to a (flat) set of pairs of feature names and their values. Table 2 lists the features used in this paper, which are a subset of those in the original output from an HPSG parser, Enju3. The HPSG forest shown in Figure 2 is in this simplified format. An important detail is that we allow a feature value to be a pointer to another (simplified) TFS. Such pointervalued features are necessary for denoting the semantics, as explained shortly. In the Enju English HPSG grammar (Miyao et 3http://www-tsujii.is.s.u-tokyo.ac.jp/enju/index.html She ignore fact want I dispute ARG1 ARG2 ARG1 ARG1 ARG2 ARG2 John kill Mary ARG2 ARG1 Figure 3: Predicate argument structures for the sentences of “John killed Mary” and “She ignored the fact that I wanted to dispute”. al., 2003) used in this paper, the semantic content of a sentence/phrase is represented by a predicateargument structure (PAS). Figure 3 shows the PAS of the example sentence in Figure 2, “John killed Mary”, and a more complex PAS for another sentence, “She ignored the fact that I wanted to dispute”, which is adopted from (Miyao et al., 2003). In an HPSG tree/forest, each leaf node generally introduces a predicate, which is represented by the pair of LEXENTRY (lexical entry) feature and PRED (predicate type) feature. The arguments of a predicate are designated by the pointers from the ARG⟨x⟩features in a leaf node to non-terminal nodes. 3.2 Localize HPSG forest Our fine-grained translation rule extraction algorithm is sketched in Algorithm 1. Considering that a parse tree is a trivial packed forest, we only use the term forest to expand our discussion, hereafter. Recall that there are pointer-valued features in the TFSs (Table 2) which prevent arbitrary segmentation of a packed forest. Hence, we have to localize an HPSG forest. For example, there are ARG pointers from t1 to c1 and c5 in the HPSG forest of Figure 2. However, the three nodes are not included in one (minimal) translation rule. This problem is caused by not considering the predicate argument dependency among t1, c1, and c5 while performing the GHKM algorithm. We can combine several minimal rules (Galley et al., 2006) together to address this dependency. Yet we have a faster way to tackle PASs, as will be described in the next subsection. Even if we omit ARG, there are still two kinds of pointer-valued features in TFSs, HEAD and SEM HEAD. Localizing these pointer-valued features is straightforward, since during parsing, the HEAD and SEM HEAD of a node are automatically transferred to its mother node. That is, the syntactic and semantic head of a node only take 329 Algorithm 1 Fine-grained rule extraction Input: HPSG tree/forest Ef, foreign sentence F, and alignment A Output: a PAS-based rule set R1 and/or a tree-rule set R2 1: if Ef is an HPSG tree then 2: E ′ f = localize Tree(Ef) 3: R1 = PASR extraction(E ′ f, F, A) ◃Algorithm 2 4: E ′′ f = ignore PAS(E ′ f) 5: R2 = TR extraction(E ′′ f , F, A) ◃composed rule extraction algorithm in (Galley et al., 2006) 6: else if Ef is an HPSG forest then 7: E ′ f = localize Forest(Ef); 8: R2 = forest based rule extraction(E ′ f, F, A) ◃Algorithm 1 in (Mi and Huang, 2008) 9: end if the identifier of the daughter node as the values. For example, HEAD and SEM HEAD of node c0 take the identical value to be c3 in Figure 2. To extract tree-to-string rules from the tree structures of an HPSG forest, our solution is to pre-process an HPSG forest in the following way: • for a phrasal hypernode, replace its HEAD and SEM HEAD value with L, R, or S, which respectively represent left daughter, right daughter, or single daughter (line 2 and 7); and, • for a lexical node, ARG⟨x⟩and PRED features are ignored (line 4). A pure syntactic-based HPSG forest without any pointer-valued features can be yielded through this pre-processing for the consequent execution of the extraction algorithms (Galley et al., 2006; Mi and Huang, 2008). 3.3 Predicate-argument structures In order to extract translation rules from PASs, we want to localize a predicate word and its arguments into one tree fragment. For example, in Figure 2, we can use a tree fragment which takes c0 as its root node and c1, t1, and c5 on its yield (= leaf nodes of a tree fragment) to cover “killed” and its subject and direct object arguments. We define this kind of tree fragment to be a minimum covering tree. For example, the minimum covering tree of {t1, c1, c5} is shown in the bottom-right corner of Figure 2. The definition supplies us a linear-time algorithm to directly find the tree fragment that covers a PAS during both rule extracting and rule matching when decoding an HPSG tree. Algorithm 2 PASR extraction Input: HPSG tree Et, foreign sentence F, and alignment A Output: a PAS-based rule set R 1: R = {} 2: for node n ∈Leaves(Et) do 3: if Open(n.ARG) then 4: Tc = MinimumCoveringTree(Et, n, n.ARGs) 5: if root and leaf nodes of Tc are in fs then 6: generate a rule r using fragment Tc 7: R.append(r) 8: end if 9: end if 10: end for See (Wu, 2010) for more examples of minimum covering trees. Taking a minimum covering tree as the tree fragment, we can easily build a tree-to-string translation rule that reflects the semantic dependency of a PAS. The algorithm of PAS-based rule (PASR) extraction is sketched in Algorithm 2. Suppose we are given a tuple of ⟨F, Et, A⟩. Et is pre-processed by replacing HEAD and SEM HEAD to be L, R, or S, and computing the span and comp span of each node. We extract PAS-based rules through one-time traversal of the leaf nodes in Et (line 2). For each leaf node n, we extract a minimum covering tree Tc if n contains at least one argument. That is, at least one ARG⟨x⟩takes the value of some node identifier, where x ranges 1 over 4 (line 3). Then, we require the root and yield nodes of Tc being in the frontier set of Et (line 5). Based on Tc, we can easily build a tree-to-string translation rule by further completing the right-hand-side string by sorting the spans of Tc’s leaf nodes, lexicalizing the terminal node’s span(s), and assigning a variable to each non-terminal node’s span. Maximum likelihood estimation is used to calculate the translation probabilities of each rule. An example of PAS-based rule is shown in the bottom-right corner of Figure 2. In the rule, the subject and direct-object of “killed” are generalized into two variables, x0 and x1. 4 Experiments 4.1 Translation models We use a tree-to-string model and a string-to-tree model for bidirectional Japanese-English translations. Both models use a phrase translation table (PTT), an HPSG tree-based rule set (TRS), and a PAS-based rule set (PRS). Since the three rule sets are independently extracted and estimated, we 330 use Minimum Error Rate Training (MERT) (Och, 2003) to tune the weights of the features from the three rule sets on the development set. Given a 1-best (localized) HPSG tree Et, the tree-to-string decoder searches for the optimal derivation d∗that transforms Et into a Japanese string among the set of all possible derivations D: d∗= arg max d∈D {λ1 log pLM(τ(d)) + λ2|τ(d)| + log s(d|Et)}. (2) Here, the first item is the language model (LM) probability where τ(d) is the target string of derivation d; the second item is the translation length penalty; and the third item is the translation score, which is decomposed into a product of feature values of rules: s(d|Et) = ∏ r∈d f(r∈PTT )f(r∈TRS)f(r∈PRS). This equation reflects that the translation rules in one d come from three sets. Inspired by (Liu et al., 2009b), it is appealing to combine these rule sets together in one decoder because PTT provides excellent rule coverages while TRS and PRS offer linguistically motivated phrase selections and nonlocal reorderings. Each f(r) is in turn a product of five features: f(r) = p(s|t)λ3 · p(t|s)λ4 · l(s|t)λ5 · l(t|s)λ6 · eλ7. Here, s/t represent the source/target part of a rule in PTT, TRS, or PRS; p(·|·) and l(·|·) are translation probabilities and lexical weights of rules from PTT, TRS, and PRS. The derivation length penalty is controlled by λ7. In our string-to-tree model, for efficient decoding with integrated n-gram LM, we follow (Zhang et al., 2006) and inversely binarize all translation rules into Chomsky Normal Forms that contain at most two variables and can be incrementally scored by LM. In order to make use of the binarized rules in the CKY decoding, we add two kinds of glues rules: S → Xm(1), Xm(1); S → S(1)Xm(2), S(1)Xm(2). Here Xm ranges over the nonterminals appearing in a binarized rule set. These glue rules can be seen as an extension from X to {Xm}of the two glue rules described in (Chiang, 2007). The string-to-tree decoder searches for the optimal derivation d∗that parses a Japanese string F into a packed forest of the set of all possible derivations D: d∗= arg max d∈D {λ1 log pLM(τ(d)) + λ2|τ(d)| + λ3g(d) + log s(d|F)}. (3) This formula differs from Equation 2 by replacing Et with F in s(d|·) and adding g(d), which is the number of glue rules used in d. Further definitions of s(d|F) and f(r) are identical with those used in Equation 2. 4.2 Decoding algorithms In our translation models, we have made use of three kinds of translation rule sets which are trained separately. We perform derivation-level combination as described in (Liu et al., 2009b) for mixing different types of translation rules within one derivation. For tree-to-string translation, we use a bottomup beam search algorithm (Liu et al., 2006) for decoding an HPSG tree Et. We keep at most 10 best derivations with distinct τ(d)s at each node. Recall the definition of minimum covering tree, which supports a faster way to retrieve available rules from PRS without generating all the subtrees. That is, when node n fortunately to be the root of some minimum covering tree(s), we use the tree(s) to seek available PAS-based rules in PRS. We keep a hash-table with the key to be the node identifier of n and the value to be a priority queue of available PAS-based rules. The hash-table is easy to be filled by one-time traversal of the terminal nodes in Et. At each terminal node, we seek its minimum covering tree, retrieve PRS, and update the hash-table. For example, suppose we are decoding an HPSG tree (with gray nodes) shown in Figure 2. At t1, we can extract its minimum covering tree with the root node to be c0, then take this tree fragment as the key to retrieve PRS, and consequently put c0 and the available rules in the hash-table. When decoding at c0, we can directly access the hash-table looking for available PASbased rules. In contrast, we use a CKY-style algorithm with beam-pruning and cube-pruning (Chiang, 2007) to decode Japanese sentences. For each Japanese sentence F, the output of the chart-parsing algorithm is expressed as a hypergraph representing a set of derivations. Given such a hypergraph, we 331 Train Dev. Test # of sentences 994K 2K 2K # of Jp words 28.2M 57.4K 57.1K # of En words 24.7M 50.3K 49.9K Table 3: Statistics of the JST corpus. use the Algorithm 3 described in (Huang and Chiang, 2005) to extract its k-best (k = 500 in our experiments) derivations. Since different derivations may lead to the same target language string, we further adopt Algorithm 3’s modification, i.e., keep a hash-table to maintain the unique target sentences (Huang et al., 2006), to efficiently generate the unique k-best translations. 4.3 Setups The JST Japanese-English paper abstract corpus4, which consists of one million parallel sentences, was used for training and testing. This corpus was constructed from a Japanese-English paper abstract corpus by using the method of Utiyama and Isahara (2007). Table 3 shows the statistics of this corpus. Making use of Enju 2.3.1, we successfully parsed 987,401 English sentences in the training set, with a parse rate of 99.3%. We modified this parser to output a packed forest for each English sentence. We executed GIZA++ (Och and Ney, 2003) and grow-diag-final-and balancing strategy (Koehn et al., 2007) on the training set to obtain a phrasealigned parallel corpus, from which bidirectional phrase translation tables were estimated. SRI Language Modeling Toolkit (Stolcke, 2002) was employed to train 5-gram English and Japanese LMs on the training set. We evaluated the translation quality using the case-insensitive BLEU-4 metric (Papineni et al., 2002). The MERT toolkit we used is Z-mert5 (Zaidan, 2009). The baseline system for comparison is Joshua (Li et al., 2009), a freely available decoder for hierarchical phrase-based SMT (Chiang, 2005). We respectively extracted 4.5M and 5.3M translation rules from the training set for the 4K English and Japanese sentences in the development and test sets. We used the default configuration of Joshua, expect setting the maximum number of items/rules and the k of k-best outputs to be the identical 4http://www.jst.go.jp. The corpus can be conditionally obtained from NTCIR-7 patent translation workshop homepage: http://research.nii.ac.jp/ntcir/permission/ntcir-7/permen-PATMT.html. 5http://www.cs.jhu.edu/ ozaidan/zmert/ PRS CS 3 C3 F S F tree nodes TFS POS TFS POS TFS # rules 0.9 62.1 83.9 92.5 103.7 # tree types 0.4 23.5 34.7 40.6 45.2 extract time 3.5 98.6 121.2 Table 4: Statistics of several kinds of tree-to-string rules. Here, the number is in million level and the time is in hour. 200 for English-to-Japanese translation and 500 for Japanese-to-English translation. We used four dual core Xeon machines (4×3.0GHz×2CPU, 4×64GB memory) to run all the experiments. 4.4 Results Table 4 illustrates the statistics of several translation rule sets, which are classified by: • using TFSs or simple POS/phrasal tags (annotated by a superscript S) to represent tree nodes; • composed rules (PRS) extracted from the PAS of 1-best HPSG trees; • composed rules (C3), extracted from the tree structures of 1-best HPSG trees, and 3 is the maximum number of internal nodes in the tree fragments; and • forest-based rules (F), where the packed forests are pre-pruned by the marginal probability-based inside-outside algorithm used in (Mi and Huang, 2008). Table 5 reports the BLEU-4 scores achieved by decoding the test set making use of Joshua and our systems (t2s = tree-to-string and s2t = string-totree) under numerous rule sets. We analyze this table in terms of several aspects to prove the effectiveness of deep syntactic information for SMT. Let’s first look at the performance of TFSs. We take CS 3 and F S as approximations of CFG-based translation rules. Comparing the BLEU-4 scores of PTT+CS 3 and PTT+C3, we gained 0.56 (t2s) and 0.57 (s2t) BLEU-4 points which are significant improvements (p < 0.05). Furthermore, we gained 0.50 (t2s) and 0.62 (s2t) BLEU-4 points from PTT+F S to PTT+F, which are also significant improvements (p < 0.05). The rich features included in TFSs contribute to these improvements. 332 Systems BLEU-t2s Decoding BLEU-s2t Joshua 21.79 0.486 19.73 PTT 18.40 0.013 17.21 PTT+PRS 22.12 0.031 19.33 PTT+CS 3 23.56 2.686 20.59 PTT+C3 24.12 2.753 21.16 PTT+C3+PRS 24.13 2.930 21.20 PTT+F S 24.25 3.241 22.05 PTT+F 24.75 3.470 22.67 Table 5: BLEU-4 scores (%) achieved by Joshua and our systems under numerous rule configurations. The decoding time (seconds per sentence) of tree-to-string translation is listed as well. Also, BLEU-4 scores were inspiringly increased 3.72 (t2s) and 2.12 (s2t) points by appending PRS to PTT, comparing PTT with PTT+PRS. Furthermore, in Table 5, the decoding time (seconds per sentence) of tree-to-string translation by using PTT+PRS is more than 86 times faster than using the other tree-to-string rule sets. This suggests that the direct generation of minimum covering trees for rule matching is extremely faster than generating all subtrees of a tree node. Note that PTT performed extremely bad compared with all other systems or tree-based rule sets. The major reason is that we did not perform any reordering or distorting during decoding with PTT. However, in both t2s and s2t systems, the BLEU-4 score benefits of PRS were covered by the composed rules: both PTT+CS 3 and PTT+C3 performed significant better (p < 0.01) than PTT+PRS, and there are no significant differences when appending PRS to PTT+C3. The reason is obvious: PRS is only a small subset of the composed rules, and the probabilities of rules in PRS were estimated by maximum likelihood, which is fast but biased compared with EM based estimation (Galley et al., 2006). Finally, by using PTT+F, our systems achieved the best BLEU-4 scores of 24.75% (t2s) and 22.67% (s2t), both are significantly better (p < 0.01) than that achieved by Joshua. 5 Conclusion We have proposed approaches of using deep syntactic information for extracting fine-grained treeto-string translation rules from aligned HPSG forest-string pairs. The main contributions are the applications of GHKM-related algorithms (Galley et al., 2006; Mi and Huang, 2008) to HPSG forests and a linear-time algorithm for extracting composed rules from predicate-argument structures. We applied our fine-grained translation rules to a tree-to-string system and an Hiero-style string-totree system. Extensive experiments on large-scale bidirectional Japanese-English translations testified the significant improvements on BLEU score. We argue the fine-grained translation rules are generic and applicable to many syntax-based SMT frameworks such as the forest-to-string model (Mi et al., 2008). Furthermore, it will be interesting to extract fine-grained tree-to-tree translation rules by integrating deep syntactic information in the source and/or target language side(s). These treeto-tree rules are applicable for forest-to-tree translation models (Liu et al., 2009a). Acknowledgments This work was partially supported by Grant-inAid for Specially Promoted Research (MEXT, Japan) and Japanese/Chinese Machine Translation Project in Special Coordination Funds for Promoting Science and Technology (MEXT, Japan), and Microsoft Research Asia Machine Translation Theme. The first author thanks Naoaki Okazaki and Yusuke Miyao for their help and the anonymous reviewers for improving the earlier version. References Alexandra Birch, Miles Osborne, and Philipp Koehn. 2007. Ccg supertags in factored statistical machine translation. In Proceedings of the Second Workshop on Statistical Machine Translation, pages 9– 16, June. Bob Carpenter. 1992. The Logic of Typed Feature Structures. Cambridge University Press. David Chiang, Kevin Knight, and Wei Wang. 2009. 11,001 new features for statistical machine translation. In Proceedings of HLT-NAACL, pages 218– 226, June. David Chiang. 2005. A hierarchical phrase-based model for statistical machine translation. In Proceedings of ACL, pages 263–270, Ann Arbor, MI. David Chiang. 2007. Hierarchical phrase-based translation. Computational Lingustics, 33(2):201–228. A. P. Dempster, N. M. Laird, and D. B. Rubin. 1977. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society, 39:1–38. Michel Galley, Mark Hopkins, Kevin Knight, and Daniel Marcu. 2004. What’s in a translation rule? In Proceedings of HLT-NAACL. 333 Michel Galley, Jonathan Graehl, Kevin Knight, Daniel Marcu, Steve DeNeefe, Wei Wang, and Ignacio Thayer. 2006. Scalable inference and training of context-rich syntactic translation models. In Proceedings of COLING-ACL, pages 961–968, Sydney. Hany Hassan, Khalil Sima’an, and Andy Way. 2007. Supertagged phrase-based statistical machine translation. In Proceedings of ACL, pages 288–295, June. Liang Huang and David Chiang. 2005. Better k-best parsing. In Proceedings of IWPT. Liang Huang, Kevin Knight, and Aravind Joshi. 2006. Statistical syntax-directed translation with extended domain of locality. In Proceedings of 7th AMTA. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondˇrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the ACL 2007 Demo and Poster Sessions, pages 177–180. Zhifei Li, Chris Callison-Burch, Chris Dyery, Juri Ganitkevitch, Sanjeev Khudanpur, Lane Schwartz, Wren N. G. Thornton, Jonathan Weese, and Omar F. Zaidan. 2009. Demonstration of joshua: An open source toolkit for parsing-based machine translation. In Proceedings of the ACL-IJCNLP 2009 Software Demonstrations, pages 25–28, August. Yang Liu, Qun Liu, and Shouxun Lin. 2006. Treeto-string alignment templates for statistical machine transaltion. In Proceedings of COLING-ACL, pages 609–616, Sydney, Australia. Yang Liu, Yajuan L¨u, and Qun Liu. 2009a. Improving tree-to-tree translation with packed forests. In Proceedings of ACL-IJCNLP, pages 558–566, August. Yang Liu, Haitao Mi, Yang Feng, and Qun Liu. 2009b. Joint decoding with multiple translation models. In Proceedings of ACL-IJCNLP, pages 576–584, August. Takuya Matsuzaki, Yusuke Miyao, and Jun’ichi Tsujii. 2007. Efficient hpsg parsing with supertagging and cfg-filtering. In Proceedings of IJCAI, pages 1671– 1676, January. Haitao Mi and Liang Huang. 2008. Forest-based translation rule extraction. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 206–214, October. Haitao Mi, Liang Huang, and Qun Liu. 2008. Forestbased translation. In Proceedings of ACL-08:HLT, pages 192–199, Columbus, Ohio. Yusuke Miyao, Takashi Ninomiya, and Jun’ichi Tsujii. 2003. Probabilistic modeling of argument structures including non-local dependencies. In Proceedings of the International Conference on Recent Advances in Natural Language Processing, pages 285– 291, Borovets. Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1):19–51. Franz Josef Och. 2003. Minimum error rate training in statistical machine translation. In Proceedings of ACL, pages 160–167. Stephan Oepen, Erik Velldal, Jan Tore Lønning, Paul Meurer, and Victoria Ros´en. 2007. Towards hybrid quality-oriented machine translation - on linguistics and probabilities in mt. In Proceedings of the 11th International Conference on Theoretical and Methodological Issues in Machine Translation (TMI-07), September. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of ACL, pages 311–318. Stefan Riezler and John T. Maxwell, III. 2006. Grammatical machine translation. In Proceedings of HLTNAACL, pages 248–255. Andreas Stolcke. 2002. Srilm-an extensible language modeling toolkit. In Proceedings of International Conference on Spoken Language Processing, pages 901–904. Masao Utiyama and Hitoshi Isahara. 2007. A japanese-english patent parallel corpus. In Proceedings of MT Summit XI, pages 475–482, Copenhagen. Xianchao Wu. 2010. Statistical Machine Translation Using Large-Scale Lexicon and Deep Syntactic Structures. Ph.D. dissertation. Department of Computer Science, The University of Tokyo. Omar F. Zaidan. 2009. Z-MERT: A fully configurable open source tool for minimum error rate training of machine translation systems. The Prague Bulletin of Mathematical Linguistics, 91:79–88. Hao Zhang, Liang Huang, Daniel Gildea, and Kevin Knight. 2006. Synchronous binarization for machine translation. In Proceedings of HLT-NAACL, pages 256–263, June. 334
|
2010
|
34
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 335–344, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Accurate Context-Free Parsing with Combinatory Categorial Grammar Timothy A. D. Fowler and Gerald Penn Department of Computer Science, University of Toronto Toronto, ON, M5S 3G4, Canada {tfowler, gpenn}@cs.toronto.edu Abstract The definition of combinatory categorial grammar (CCG) in the literature varies quite a bit from author to author. However, the differences between the definitions are important in terms of the language classes of each CCG. We prove that a wide range of CCGs are strongly context-free, including the CCG of CCGbank and of the parser of Clark and Curran (2007). In light of these new results, we train the PCFG parser of Petrov and Klein (2007) on CCGbank and achieve state of the art results in supertagging accuracy, PARSEVAL measures and dependency accuracy. 1 Introduction Combinatory categorial grammar (CCG) is a variant of categorial grammar which has attracted interest for both theoretical and practical reasons. On the theoretical side, we know that it is mildly context-sensitive (Vijay-Shanker and Weir, 1994) and that it can elegantly analyze a wide range of linguistic phenomena (Steedman, 2000). On the practical side, we have corpora with CCG derivations for each sentence (Hockenmaier and Steedman, 2007), a wide-coverage parser trained on that corpus (Clark and Curran, 2007) and a system for converting CCG derivations into semantic representations (Bos et al., 2004). However, despite being treated as a single unified grammar formalism, each of these authors use variations of CCG which differ primarily on which combinators are included in the grammar and the restrictions that are put on them. These differences are important because they affect whether the mild context-sensitivity proof of Vijay-Shanker and Weir (1994) applies. We will provide a generalized framework for CCG within which the full variation of CCG seen in the literature can be defined. Then, we prove that for a wide range of CCGs there is a context-free grammar (CFG) that has exactly the same derivations. Included in this class of strongly context-free CCGs are a grammar including all the derivations in CCGbank and the grammar used in the Clark and Curran parser. Due to this insight, we investigate the potential of using tools from the probabilistic CFG community to improve CCG parsing results. The Petrov parser (Petrov and Klein, 2007) uses latent variables to refine the grammar extracted from a corpus to improve accuracy, originally used to improve parsing results on the Penn treebank (PTB). We train the Petrov parser on CCGbank and achieve the best results to date on sentences from section 23 in terms of supertagging accuracy, PARSEVAL measures and dependency accuracy. These results should not be interpreted as proof that grammars extracted from the Penn treebank and from CCGbank are equivalent. Bos’s system for building semantic representations from CCG derivations is only possible due to the categorial nature of CCG. Furthermore, the long distance dependencies involved in extraction and coordination phenomena have a more natural representation in CCG. 2 The Language Classes of Combinatory Categorial Grammars A categorial grammar is a grammatical system consisting of a finite set of words, a set of categories, a finite set of sentential categories, a finite lexicon mapping words to categories and a rule system dictating how the categories can be combined. The set of categories are constructed from a finite set of atoms A (e.g. A = {S, NP, N, PP}) and a finite set of binary connectives B (e.g. B = {/, \}) to build an infinite set of categories C(A, B) (e.g. C(A, B) = {S, S\NP, (S\NP)/ NP, . . .}). For a category C, its size |C| is the 335 number of atom occurrences it contains. When not specified, connectives are left associative. According to the literature, combinatory categorial grammar has been defined to have a variety of rule systems. These rule systems vary from a small rule set, motivated theoretically (VijayShanker and Weir, 1994), to a larger rule set, motivated linguistically, (Steedman, 2000) to a very large rule set, motivated by practical coverage (Hockenmaier and Steedman, 2007; Clark and Curran, 2007). We provide a definition general enough to incorporate these four main variants of CCG, as well as others. A combinatory categorial grammar (CCG) is a categorial grammar whose rule system consists of rule schemata where the left side is a sequence of categories and the right side is a single category where the categories may include variables over both categories and connectives. In addition, rule schemata may specify a sequence of categories and connectives using the . . . convention1. When . . . appears in a rule, it matches any sequence of categories and connectives according to the connectives adjacent to the . . .. For example, the rule schema for forward composition is: X/Y, Y/Z →X/Z and the rule schema for generalized forward crossed composition is: X/Y, Y |1Z1|2 . . . |nZn →X|1Z1|2 . . . |nZn where X, Y and Zi for 1 ≤i ≤n are variables over categories and |i for 1 ≤i ≤n are variables over connectives. Figure 1 shows a CCG derivation from CCGbank. A well-known categorial grammar which is not a CCG is Lambek categorial grammar (Lambek, 1958) whose introduction rules cannot be characterized as combinatory rules (Zielonka, 1981). 2.1 Classes for defining CCG We define a number of schema classes general enough that the important variants of CCG can be defined by selecting some subset of the classes. In addition to the schema classes, we also define two restriction classes which define ways in which the rule schemata from the schema classes can be restricted. We define the following schema classes: 1The . . . convention (Vijay-Shanker and Weir, 1994) is essentially identical to the $ convention of Steedman (2000). (1) Application • X/Y, Y →X • Y, X\Y →X (2) Composition • X/Y, Y/Z →X/Z • Y \Z, X\Y →X\Z (3) Crossed Composition • X/Y, Y \Z →X\Z • Y/Z, X\Y →X/Z (4) Generalized Composition • X/Y, Y/Z1/ . . . /Zn →X/Z1/ . . . /Zn • Y \Z1\ . . . \Zn, X\Y →X\Z1\ . . . \Zn (5) Generalized Crossed Composition • X/Y, Y |1Z1|2 . . . |nZn →X|1Z1|2 . . . |nZn • Y |1Z1|2 . . . |nZn, X\Y →X|1Z1|2 . . . |nZn (6) Reducing Generalized Crossed Composition Generalized Composition or Generalized Crossed Composition where |X| ≤|Y |. (7) Substitution • (X/Y )|1Z, Y |1Z →X|1Z • Y |1Z, (X\Y )|1Z →X|1Z (8) D Combinator2 • X/(Y |1Z), Y |2W →X|2(W|1Z) • Y |2W, X\(Y |1Z) →X|2(W|1Z) (9) Type-Raising • X →T/(T\X) • X →T\(T/X) (10) Finitely Restricted Type-Raising • X →T/(T\X) where ⟨X, T⟩∈S for finite S • X →T\(T/X) where ⟨X, T⟩∈S for finite S (11) Finite Unrestricted Variable-Free Rules • ⃗X →Y where ⟨⃗X, Y ⟩∈S for finite S 2Hoyt and Baldridge (2008) argue for the inclusion of the D Combinator in CCG. 336 Mr. Vinken is chairman of Elsevier N.V. , the Dutch publishing group . N/N N S[dcl]\NP/NP N NP\NP/NP N/N N , NP[nb]/N N/N N/N N . N N NP NP[conj] N NP NP NP\NP NP NP S[dcl]\NP N NP S[dcl] S[dcl] Figure 1: A CCG derivation from section 00 of CCGbank. We define the following restriction classes: (A) Rule Restriction to a Finite Set The rule schemata in the schema classes of a CCG are limited to a finite number of instantiations. (B) Rule Restrictions to Certain Categories 3 The rule schemata in the schema classes of a CCG are limited to a finite number of instantiations although variables are allowed in the instantiations. Vijay-Shanker and Weir (1994) define CCG to be schema class (4) with restriction class (B). Steedman (2000) defines CCG to be schema classes (1-5), (6), (10) with restriction class (B). 2.2 Strongly Context-Free CCGs Proposition 1. The set of atoms in any derivation of any CCG consisting of a subset of the schema classes (1-8) and (10-11) is finite. Proof. A finite lexicon can introduce only a finite number of atoms in lexical categories. Any rule corresponding to a schema in the schema classes (1-8) has only those atoms on the right that occur somewhere on the left. Rules in classes (10-11) can each introduce a finite number of atoms, but there can be only a finite number of 3Baldridge (2002) introduced a variant of CCG where modalities are added to the connectives / and \ along with variants of the combinatory rules based on these modalities. Our proofs about restriction class (B) are essentially identical to proofs regarding the multi-modal variant. such rules, limiting the new atoms to a finite number. Definition 1. The subcategories for a category c are c1 and c2 if c = c1 • c2 for • ∈B and c if c is atomic. Its second subcategories are the subcategories of its subcategories. Proposition 2. Any CCG consisting of a subset of the rule schemata (1-3), (6-8) and (10-11) has derivations consisting of only a finite number of categories. Proof. We first prove the proposition excluding schema class (8). We will use structural induction on the derivations to prove that there is a bound on the size of the subcategories of any category in the derivation. The base case is the assignment of a lexical category to a word and the inductive step is the use of a rule from schema classes (1-4), (6-7) and (10-11). Given that the lexicon is finite, there is a bound k on the size of the subcategories of lexical categories. Furthermore, there is a bound l on the size of the subcategories of categories on the right side of any rule in (10) and (11). Let m = max(k, l). For rules from schema class (1), the category on the right is a subcategory of the first category on the left, so the subcategories on the right are bound by m. For rules from schema classes (2-3), the category on the right has subcategories X and Z each of which is bound in size by m since they occur as subcategories of categories on the left. For rules from schema class (6), since reducing generalized composition is a special case of re337 ducing generalized crossing composition, we need only consider the latter. The category on the right has subcategories X|1Z1|2 . . . |n−1|Zn−1 and Zn. Zn is bound in size by m because it occurs as a subcategory of the second category on the left. Then, the size of Y |1Z1|2 . . . |n−1|Zn−1 must be bound by m and since |X| ≤|Y |, the size of X|1Z1|2 . . . |n−1|Zn−1 must also be bound by m. For rules from schema class (7), the category on the right has subcategories X and Z. The size of Z is bound by m because it is a subcategory of a category on the left. The size of X is bound by m because it is a second subcategory of a category on the left. Finally, the use of rules in schema classes (1011) have categories on the right that are bounded by l, which is, in turn, bounded by m. Then, by proposition 1, there must only be a finite number of categories in any derivation in a CCG consisting of a subset of rule schemata (1-3), (6-7) and (1011). The proof including schema class (8) is essentially identical except that k must be defined in terms of the size of the second subcategories. Definition 2. A grammar is strongly context-free if there exists a CFG such that the derivations of the two grammars are identical. Proposition 3. Any CCG consisting of a subset of the schema classes (1-3), (6-8) and (10-11) is strongly context-free. Proof. Since the CCG generates derivations whose categories are finite in number let C be that set of categories. Let S(C, X) be the subset of C matching category X (which may have variables). Then, for each rule schema C1, C2 →C3 in (1-3) and (6-8), we construct a context-free rule C′ 3 → C′ 1, C′ 2 for each C′ i in S(C, Ci) for 1 ≤i ≤3. Similarly, for each rule schema C1 →C2 in (10), we construct a context-free rule C′ 2 →C′ 1 which results in a finite number of such rules. Finally, for each rule schema ⃗X →Z in (11) we construct a context-free rule Z →⃗X. Then, for each entry in the lexicon w →C, we construct a context-free rule C →w. The constructed CFG has precisely the same rules as the CCG restricted to the categories in C except that the left and right sides have been reversed. Thus, by proposition 2, the CFG has exactly the same derivations as the CCG. Proposition 4. Any CCG consisting of a subset of the schema classes (1-3), (6-8) and (10-11) along with restriction class (B) is strongly context-free. Proof. If a CCG is allowed to restrict the use of its rules to certain categories as in schema class (B), then when we construct the context-free rules by enumerating only those categories in the set C allowed by the restriction. Proposition 5. Any CCG that includes restriction class (A) is strongly context-free. Proof. We construct a context-free grammar with exactly those rules in the finite set of instantiations of the CCG rule schemata along with contextfree rules corresponding to the lexicon. This CFG generates exactly the same derivations as the CCG. We have thus proved that of a wide range of the rule schemata used to define CCGs are contextfree. 2.3 Combinatory Categorial Grammars in Practice CCGbank (Hockenmaier and Steedman, 2007) is a corpus of CCG derivations that was semiautomatically converted from the Wall Street Journal section of the Penn treebank. Figure 2 shows a categorization of the rules used in CCGbank according to the schema classes defined in the preceding section where a rule is placed into the least general class to which it belongs. In addition to having no generalized composition other than the reducing variant, it should also be noted that in all generalized composition rules, X = Y implying that the reducing class of generalized composition is a very natural schema class for CCGbank. If we assume that type-raising is restricted to those instances occurring in CCGbank4, then a CCG consisting of schema classes (1-3), (6-7) and (10-11) can generate all the derivations in CCGbank. By proposition 3, such a CCG is strongly context-free. One could also observe that since CCGbank is finite, its grammar is not only a context-free grammar but can produce only a finite number of derivations. However, our statement is much stronger because this CCG can generate all of the derivations in CCGbank given only the lexicon, the finite set of unrestricted rules and the finite number of type-raising rules. 4Without such an assumption, parsing is intractable. 338 Schema Class Rules Instances Application 519 902176 Composition 102 7189 Crossed Composition 64 14114 Reducing Generalized 50 612 Crossed Composition Generalized Composition 0 0 Generalized Crossed 0 0 Composition Substitution 3 4 Type-Raising 27 3996 Unrestricted Rules 642 335011 Total 1407 1263102 Figure 2: The rules of CCGbank by schema class. The Clark and Curran CCG Parser (Clark and Curran, 2007) is a CCG parser which uses CCGbank as a training corpus. Despite the fact that there is a strongly context-free CCG which generates all of the derivations in CCGbank, it is still possible that the grammar learned by the Clark and Curran parser is not a context-free grammar. However, in addition to rule schemata (1-6) and (10-11) they also include restriction class (A) by restricting rules to only those found in the training data5. Thus, by proposition 5, the Clark and Curran parser is a context-free parser. 3 A Latent Variable CCG Parser The context-freeness of a number of CCGs should not be considered evidence that there is no advantage to CCG as a grammar formalism. Unlike the context-free grammars extracted from the Penn treebank, these allow for the categorial semantics that accompanies any categorial parse and for a more elegant analysis of linguistic structures such as extraction and coordination. However, because we now know that the CCG defined by CCGbank is strongly context-free, we can use tools from the CFG parsing community to improve CCG parsing. To illustrate this point, we train the Petrov parser (Petrov and Klein, 2007) on CCGbank. The Petrov parser uses latent variables to refine a coarse-grained grammar extracted from a training corpus to a grammar which makes much more fine-grained syntactic distinctions. For example, 5The Clark and Curran parser has an option, which is disabled by default, for not restricting the rules to those that appear in the training data. However, they find that this restriction is “detrimental to neither parser accuracy or coverage” (Clark and Curran, 2007). in Petrov’s experiments on the Penn treebank, the syntactic category NP was refined to the more fine-grained NP 1 and NP 2 roughly corresponding to NPs in subject and object positions. Rather than requiring such distinctions to be made in the corpus, the Petrov parser hypothesizes these splits automatically. The Petrov parser operates by performing a fixed number of iterations of splitting, merging and smoothing. The splitting process is done by performing Expectation-Maximization to determine a likely potential split for each syntactic category. Then, during the merging process some of the splits are undone to reduce grammar size and avoid overfitting according to the likelihood of the split against the training data. The Petrov parser was chosen for our experiments because it refines the grammar in a mathematically principled way without altering the nature of the derivations that are output. This is important because the input to the semantic backend and the system that converts CCG derivations to dependencies requires CCG derivations as they appear in CCGbank. 3.1 Experiments Our experiments use CCGbank as the corpus and we use sections 02-21 for training (39603 sentences), 00 for development (1913 sentences) and 23 for testing (2407 sentences). CCGbank, in addition to the basic atoms S, N, NP and PP, also differentiates both the S and NP atoms with features allowing more subtle distinctions. For example, declarative sentences are S[dcl], wh-questions are S[wq] and sentence fragments are S[frg] (Hockenmaier and Steedman, 2007). These features allow finer control of the use of combinatory rules in the resulting grammars. However, this fine-grained control is exactly what the Petrov parser does automatically. Therefore, we trained the Petrov parser twice, once on the original version of CCGbank (denoted “Petrov”) and once on a version of CCGbank without these features (denoted “Petrov no feats”). Furthermore, we will evaluate the parsers obtained after 0, 4, 5 and 6 training iterations (denoted I-0, I-4, I-5 and I-6). When we evaluate on sets of sentences for which not all parsers return an analysis, we report the coverage (denoted “Cover”). We use the evalb package for PARSEVAL evaluation and a modified version of Clark and 339 Parser Accuracy % No feats % C&C Normal Form 92.92 93.38 C&C Hybrid 93.06 93.52 Petrov I-5 93.18 93.73 Petrov no feats I-6 93.74 Figure 3: Supertagging accuracy on the sentences in section 00 that receive derivations from the four parsers shown. Parser Accuracy % No feats % C&C Hybrid 92.98 93.43 Petrov I-5 93.10 93.59 Petrov no feats I-6 93.62 Figure 4: Supertagging accuracy on the sentences in section 23 that receive derivations from the three parsers shown. Curran’s evaluate script for dependency evaluation. To determine statistical significance, we obtain p-values from Bikel’s randomized parsing evaluation comparator6, modified for use with tagging accuracy, F-score and dependency accuracy. 3.2 Supertag Evaluation Before evaluating the parse trees as a whole, we evaluate the categories assigned to words. In the supertagging literature, POS tagging and supertagging are distinguished – POS tags are the traditional Penn treebank tags (e.g. NN, VBZ and DT) and supertags are CCG categories. However, because the Petrov parser trained on CCGbank has no notion of Penn treebank POS tags, we can only evaluate the accuracy of the supertags. The results are shown in figures 3 and 4 where the “Accuracy” column shows accuracy of the supertags against the CCGbank categories and the “No feats” column shows accuracy when features are ignored. Despite the lack of POS tags in the Petrov parser, we can see that it performs slightly better than the Clark and Curran parser. The difference in accuracy is only statistically significant between Clark and Curran’s Normal Form model ignoring features and the Petrov parser trained on CCGbank without features (p-value = 0.013). 3.3 Constituent Evaluation In this section we evaluate the parsers using the traditional PARSEVAL measures which measure recall, precision and F-score on constituents in 6http://www.cis.upenn.edu/ dbikel/software.html both labeled and unlabeled versions. In addition, we report a variant of the labeled PARSEVAL measures where we ignore the features on the categories. For reasons of brevity, we report the PARSEVAL measures for all sentences in sections 00 and 23, rather than for sentences of length is less than 40 or less than 100. The results are essentially identical for those two sets of sentences. Figure 5 gives the PARSEVALmeasures on section 00 for Clark and Curran’s two best models and the Petrov parser trained on the original CCGbank and the version without features after various numbers of training iterations. Figure 7 gives the accuracies on section 23. In the case of Clark and Curran’s hybrid model, the poor accuracy relative to the Petrov parsers can be attributed to the fact that this model chooses derivations based on the associated dependencies at the expense of constituent accuracy (see section 3.4). In the case of Clark and Curran’s normal form model, the large difference between labeled and unlabeled accuracy is primarily due to the mislabeling of a small number of features (specifically, NP[nb] and NP[num]). The labeled accuracies without features gives the results when features are disregarded. Due to the similarity of the accuracies and the difference in the coverage between I-5 of the Petrov parser on CCGbank and I-6 of the Petrov parser on CCGbank without features, we reevaluate their results on only those sentences for which they both return derivations in figures 6 and 8. These results show that the features in CCGbank actually inhibit accuracy (to a statistically significant degree in the case of unlabeled accuracy on section 00) when used as training data for the Petrov parser. Figure 9 gives a comparison between the Petrov parser trained on the Penn treebank and on CCGbank. These numbers should not be directly compared, but the similarity of the unlabeled measures indicates that the difference between the structure of the Penn treebank and CCGbank is not large.7 3.4 Dependency Evaluation The constituent-based PARSEVAL measures are simple to calculate from the output of the Petrov parser but the relationship of the PARSEVAL 7Because punctuation in CCG can have grammatical function, we include it in our accuracy calculations resulting in lower scores for the Petrov parser trained on the Penn treebank than those reported in Petrov and Klein (2007). 340 Labeled % Labeled no feats % Unlabeled % Parser R P F R P F R P F Cover C&C Normal Form 71.14 70.76 70.95 80.66 80.24 80.45 86.16 85.71 85.94 98.95 C&C Hybrid 50.08 49.47 49.77 58.13 57.43 57.78 61.27 60.53 60.90 98.95 Petrov I-0 74.19 74.27 74.23 74.66 74.74 74.70 78.65 78.73 78.69 99.95 Petrov I-4 85.86 85.78 85.82 86.36 86.29 86.32 89.96 89.88 89.92 99.90 Petrov I-5 86.30 86.16 86.23 86.84 86.70 86.77 90.28 90.13 90.21 99.90 Petrov I-6 85.95 85.68 85.81 86.51 86.23 86.37 90.22 89.93 90.08 99.22 Petrov no feats I-0 72.16 72.59 72.37 76.52 76.97 76.74 99.95 Petrov no feats I-5 86.67 86.57 86.62 90.30 90.20 90.25 99.90 Petrov no feats I-6 87.45 87.37 87.41 90.99 90.91 90.95 99.84 Figure 5: Constituent accuracy on all sentences from section 00. Labeled % Labeled no feats % Unlabeled % Parser R P F R P F R P F Petrov I-5 86.56 86.46 86.51 87.10 87.01 87.05 90.43 90.33 90.38 Petrov no feats I-6 87.45 87.37 87.41 90.99 90.91 90.95 p-value 0.089 0.090 0.088 0.006 0.008 0.007 Figure 6: Constituent accuracy on the sentences in section 00 that receive a derivation from both parsers. Labeled % Labeled no feats % Unlabeled % Parser R P F R P F R P F Cover C&C Normal Form 71.15 70.79 70.97 80.73 80.32 80.53 86.31 85.88 86.10 99.58 Petrov I-5 86.94 86.80 86.87 87.47 87.32 87.39 90.75 90.59 90.67 99.83 Petrov no feats I-6 87.49 87.49 87.49 90.81 90.82 90.81 99.96 Figure 7: Constituent accuracy on all sentences from section 23. Labeled % Labeled no feats % Unlabeled % Parser R P F R P F R P F Petrov I-5 86.94 86.80 86.87 87.47 87.32 87.39 90.75 90.59 90.67 Petrov no feats I-6 87.48 87.49 87.49 90.81 90.82 90.81 p-value 0.463 0.215 0.327 0.364 0.122 0.222 Figure 8: Constituent accuracy on the sentences in section 23 that receive a derivation from both parsers. Labeled % Unlabeled % Parser R P F R P F Cover Petrov on PTB I-6 89.65 89.97 89.81 90.80 91.13 90.96 100.00 Petrov on CCGbank I-5 86.94 86.80 86.87 90.75 90.59 90.67 99.83 Petrov on CCGbank no feats I-6 87.49 87.49 87.49 90.81 90.82 90.81 99.96 Figure 9: Constituent accuracy for the Petrov parser on the corpora on all sentences from Section 23. Mr. Vinken is chairman of Elsevier N.V. , the Dutch publishing group . N/N N S[dcl]\NP/NP N NP\NP/NP N/N N , NP[nb]/N N/N N/N N . Figure 10: The argument-functor relations for the CCG derivation in figure 1. 341 Mr. Vinken is chairman of Elsevier N.V. , the Dutch publishing group . N/N N S[dcl]\NP/NP N NP\NP/NP N/N N , NP[nb]/N N/N N/N N . Figure 11: The set of dependencies obtained by reorienting the argument-functor edges in figure 10. Labeled % Unlabeled % Parser R P F R P F Cover C&C Normal Form 84.39 85.28 84.83 90.93 91.89 91.41 98.95 C&C Hybrid 84.53 86.20 85.36 90.84 92.63 91.73 98.95 Petrov I-0 79.87 78.81 79.34 87.68 86.53 87.10 96.45 Petrov I-4 84.76 85.27 85.02 91.69 92.25 91.97 96.81 Petrov I-5 85.30 85.87 85.58 92.00 92.61 92.31 96.65 Petrov I-6 84.86 85.46 85.16 91.79 92.44 92.11 96.65 Figure 12: Dependency accuracy on CCGbank dependencies on all sentences from section 00. Labeled % Unlabeled % Parser R P F R P F C&C Hybrid 84.71 86.35 85.52 90.96 92.72 91.83 Petrov I-5 85.50 86.08 85.79 92.12 92.75 92.44 p-value 0.005 0.189 0.187 < 0.001 0.437 0.001 Figure 13: Dependency accuracy on the section 00 sentences that receive an analysis from both parsers. Labeled % Unlabeled % Parser R P F R P F C&C Hybrid 85.11 86.46 85.78 91.15 92.60 91.87 Petrov I-5 85.73 86.29 86.01 92.04 92.64 92.34 p-value 0.013 0.278 0.197 < 0.001 0.404 0.005 Figure 14: Dependency accuracy on the section 23 sentences that receive an analysis from both parsers. Training Time Parsing Time Training RAM Parser in CPU minutes in CPU minutes in gigabytes Clark and Curran Normal Form Model 1152 2 28 Clark and Curran Hybrid Model 2672 4 37 Petrov on PTB I-0 1 5 2 Petrov on PTB I-5 180 20 8 Petrov on PTB I-6 660 21 16 Petrov on CCGbank I-0 1 5 2 Petrov on CCGbank I-4 103 70 8 Petrov on CCGbank I-5 410 600 14 Petrov on CCGbank I-6 2760 2880 24 Petrov on CCGbank no feats I-0 1 5 2 Petrov on CCGbank no feats I-5 360 240 7 Petrov on CCGbank no feats I-6 1980 390 13 Figure 15: Time and space usage when training on sections 02-21 and parsing on section 00. 342 scores to the quality of a parse is not entirely clear. For this reason, the word to word dependencies of categorial grammar parsers are often evaluated. This evaluation is aided by the fact that in addition to the CCG derivation for each sentence, CCGbank also includes a set of dependencies. Furthermore, extracting dependencies from a CCG derivation is well-established (Clark et al., 2002). A CCG derivation can be converted into dependencies by, first, determining which arguments go with which functors as specified by the CCG derivation. This can be represented as in figure 10. Although this is not difficult, some care must be taken with respect to punctuation and the conjunction rules. Next, we reorient some of the edges according to information in the lexical categories. A language for specifying these instructions using variables and indices is given in Clark et al. (2002). This process is shown in figures 1, 10 and 11 with the directions of the dependencies reversed from Clark et al. (2002). We used the CCG derivation to dependency converter generate included in the C&C tools package to convert the output of the Petrov parser to dependencies. Other than a CCG derivation, their system requires only the lexicon of edge reorientation instructions and methods for converting the unrestricted rules of CCGbank into the argument-functor relations. Important for the purpose of comparison, this system does not depend on their parser. An unlabeled dependency is correct if the ordered pair of words is correct. A labeled dependency is correct if the ordered pair of words is correct, the head word has the correct category and the position of the category that is the source of that edge is correct. Figure 12 shows accuracies from the Petrov parser trained on CCGbank along with accuracies for the Clark and Curran parser. We only show accuracies for the Petrov parser trained on the original version of CCGbank because the dependency converter cannot currently generate dependencies for featureless derivations. The relatively poor coverage of the Petrov parser is due to the failure of the dependency converter to output dependencies from valid CCG derivations. However, the coverage of the dependency converter is actually lower when run on the gold standard derivations indicating that this coverage problem is not indicative of inaccuracies in the Petrov parser. Due to the difference in coverage, we again evaluate the top two parsers on only those sentences that they both generate dependencies for and report those results in figures 13 and 14. The Petrov parser has better results by a statistically significant margin for both labeled and unlabeled recall and unlabeled F-score. 3.5 Time and Space Evaluation As a final evaluation, we compare the resources that are required to both train and parse with the Petrov parser on the Penn Treebank, the Petrov parser on the original version of CCGbank, the Petrov parser on CCGbank without features and the Clark and Curran parser using the two models. All training and parsing was done on a 64-bit machine with 8 dual core 2.8 Ghz Opteron 8220 CPUs and 64GB of RAM. Our training times are much larger than those reported in Clark and Curran (2007) because we report the cumulative time spent on all CPUs rather than the maximum time spent on a CPU. Figure 15 shows the results. As can be seen, the Clark and Curran parser has similar training times, although significantly greater RAM requirements than the Petrov parsers. In contrast, the Clark and Curran parser is significantly faster than the Petrov parsers, which we hypothesize to be attributed to the degree to which Clark and Curran have optimized their code, their use of C++ as opposed to Java and their use of a supertagger to prune the lexicon. 4 Conclusion We have provided a number of theoretical results proving that CCGbank contains no non-contextfree structure and that the Clark and Curran parser is actually a context-free parser. Based on these results, we trained the Petrov parser on CCGbank and achieved state of the art results in terms of supertagging accuracy, PARSEVAL measures and dependency accuracy. This demonstrates the following. First, the ability to extract semantic representations from CCG derivations is not dependent on the language class of a CCG. Second, using a dedicated supertagger, as opposed to simply using a general purpose tagger, is not necessary to accurately parse with CCG. Acknowledgments We would like to thank Stephen Clark, James Curran, Jackie C. K. Cheung and our three anonymous reviewers for their insightful comments. 343 References J. Baldridge. 2002. Lexically Specified Derivational Control in Combinatory Categorial Grammar. Ph.D. thesis, University of Edinburgh. J. Bos, S. Clark, M. Steedman, J. R Curran, and J. Hockenmaier. 2004. Wide-coverage semantic representations from a CCG parser. In Proceedings of COLING, volume 4, page 1240–1246. S. Clark and J. R. Curran. 2007. Wide-Coverage efficient statistical parsing with CCG and Log-Linear models. Computational Linguistics, 33(4):493–552. S. Clark, J. Hockenmaier, and M. Steedman. 2002. Building deep dependency structures with a widecoverage CCG parser. In Proceedings of the 40th Meeting of the ACL, page 327–334. J. Hockenmaier and M. Steedman. 2007. CCGbank: a corpus of CCG derivations and dependency structures extracted from the penn treebank. Computational Linguistics, 33(3):355–396. F. Hoyt and J. Baldridge. 2008. A logical basis for the d combinator and normal form in CCG. In Proceedings of ACL-08: HLT, page 326–334, Columbus, Ohio. Association for Computational Linguistics. J. Lambek. 1958. The mathematics of sentence structure. American Mathematical Monthly, 65(3):154–170. S. Petrov and D. Klein. 2007. Improved inference for unlexicalized parsing. In Proceedings of NAACL HLT 2007, page 404–411. M. Steedman. 2000. The syntactic process. MIT Press. K. Vijay-Shanker and D. Weir. 1994. The equivalence of four extensions of context-free grammars. Mathematical Systems Theory, 27(6):511–546. W. Zielonka. 1981. Axiomatizability of AjdukiewiczLambek calculus by means of cancellation schemes. Zeitschrift fur Mathematische Logik und Grundlagen der Mathematik, 27:215–224. 344
|
2010
|
35
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 345–355, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Faster Parsing by Supertagger Adaptation Jonathan K. Kummerfelda Jessika Roesnerb Tim Dawborna James Haggertya James R. Currana∗ Stephen Clarkc∗ School of Information Technologiesa Department of Computer Scienceb Computer Laboratoryc University of Sydney University of Texas at Austin University of Cambridge NSW 2006, Australia Austin, TX, USA Cambridge CB3 0FD, UK [email protected]∗ [email protected]∗ Abstract We propose a novel self-training method for a parser which uses a lexicalised grammar and supertagger, focusing on increasing the speed of the parser rather than its accuracy. The idea is to train the supertagger on large amounts of parser output, so that the supertagger can learn to supply the supertags that the parser will eventually choose as part of the highestscoring derivation. Since the supertagger supplies fewer supertags overall, the parsing speed is increased. We demonstrate the effectiveness of the method using a CCG supertagger and parser, obtaining significant speed increases on newspaper text with no loss in accuracy. We also show that the method can be used to adapt the CCG parser to new domains, obtaining accuracy and speed improvements for Wikipedia and biomedical text. 1 Introduction In many NLP tasks and applications, e.g. distributional similarity (Curran, 2004) and question answering (Dumais et al., 2002), large volumes of text and detailed syntactic information are both critical for high performance. To avoid a tradeoff between these two, we need to increase parsing speed, but without losing accuracy. Parsing with lexicalised grammar formalisms, such as Lexicalised Tree Adjoining Grammar and Combinatory Categorial Grammar (CCG; Steedman, 2000), can be made more efficient using a supertagger. Bangalore and Joshi (1999) call supertagging almost parsing because of the significant reduction in ambiguity which occurs once the supertags have been assigned. In this paper, we focus on the CCG parser and supertagger described in Clark and Curran (2007). Since the CCG lexical category set used by the supertagger is much larger than the Penn Treebank POS tag set, the accuracy of supertagging is much lower than POS tagging; hence the CCG supertagger assigns multiple supertags1 to a word, when the local context does not provide enough information to decide on the correct supertag. The supertagger feeds lexical categories to the parser, and the two interact, sometimes using multiple passes over a sentence. If a spanning analysis cannot be found by the parser, the number of lexical categories supplied by the supertagger is increased. The supertagger-parser interaction influences speed in two ways: first, the larger the lexical ambiguity, the more derivations the parser must consider; second, each further pass is as costly as parsing a whole extra sentence. Our goal is to increase parsing speed without loss of accuracy. The technique we use is a form of self-training, in which the output of the parser is used to train the supertagger component. The existing literature on self-training reports mixed results. Clark et al. (2003) were unable to improve the accuracy of POS tagging using self-training. In contrast, McClosky et al. (2006a) report improved accuracy through self-training for a twostage parser and re-ranker. Here our goal is not to improve accuracy, only to maintain it, which we achieve through an adaptive supertagger. The adaptive supertagger produces lexical categories that the parser would have used in the final derivation when using the baseline model. However, it does so with much lower ambiguity levels, and potentially during an earlier pass, which means sentences are parsed faster. By increasing the ambiguity level of the adaptive models to match the baseline system, we can also slightly increase supertagging accuracy, which can lead to higher parsing accuracy. 1We use supertag and lexical category interchangeably. 345 Using the parser to generate training data also has the advantage that it is not a domain specific process. Previous work has shown that parsers typically perform poorly outside of their training domain (Gildea, 2001). Using a newspapertrained parser, we constructed new training sets for Wikipedia and biomedical text. These were used to create new supertagging models adapted to the different domains. The self-training method of adapting the supertagger to suit the parser increased parsing speed by more than 50% across all three domains, without loss of accuracy. Using an adapted supertagger with ambiguity levels tuned to match the baseline system, we were also able to increase F-score on labelled grammatical relations by 0.75%. 2 Background Many statistical parsers use two stages: a tagging stage that labels each word with its grammatical role, and a parsing stage that uses the tags to form a parse tree. Lexicalised grammars typically contain a much smaller set of rules than phrase-structure grammars, relying on tags (supertags) that contain a more detailed description of each word’s role in the sentence. This leads to much larger tag sets, and shifts a large proportion of the search for an optimal derivation to the tagging component of the parser. Figure 1 gives two sentences and their CCG derivations, showing how some of the syntactic ambiguity is transferred to the supertagging component in a lexicalised grammar. Note that the lexical category assigned to with is different in each case, reflecting the fact that the prepositional phrase attaches differently. Either we need a tagging model that can resolve this ambiguity, or both lexical categories must be supplied to the parser which can then attempt to resolve the ambiguity by eventually selecting between them. 2.1 Supertagging Supertaggers typically use standard linear-time tagging algorithms, and only consider words in the local context when assigning a supertag. The C&C supertagger is similar to the Ratnaparkhi (1996) tagger, using features based on words and POS tags in a five-word window surrounding the target word, and defining a local probability distribution over supertags for each word in the sentence, given the previous two supertags. The Viterbi algorithm I ate pizza with cutlery NP (S\NP)/NP NP ((S\NP)\(S\NP))/NP NP > > S\NP (S\NP)\(S\NP) < S\NP < S I ate pizza with anchovies NP (S\NP)/NP NP (NP\NP)/NP NP > NP\NP < NP > S\NP < S Figure 1: Two CCG derivations with PP ambiguity. can be used to find the most probable supertag sequence. Alternatively the Forward-Backward algorithm can be used to efficiently sum over all sequences, giving a probability distribution over supertags for each word which is conditional only on the input sentence. Supertaggers can be made accurate enough for wide coverage parsing using multi-tagging (Chen et al., 1999), in which more than one supertag can be assigned to a word; however, as more supertags are supplied by the supertagger, parsing efficiency decreases (Chen et al., 2002), demonstrating the influence of lexical ambiguity on parsing complexity (Sarkar et al., 2000). Clark and Curran (2004) applied supertagging to CCG, using a flexible multi-tagging approach. The supertagger assigns to a word all lexical categories whose probabilities are within some factor, β, of the most probable category for that word. When the supertagger is integrated with the C&C parser, several progressively lower β values are considered. If a sentence is not parsed on one pass then the parser attempts to parse the sentence again with a lower β value, using a larger set of categories from the supertagger. Since most sentences are parsed at the first level (in which the average number of supertags assigned to each word is only slightly greater than one), this provides some of the speed benefit of single tagging, but without loss of coverage (Clark and Curran, 2004). Supertagging has since been effectively applied to other formalisms, such as HPSG (Blunsom and Baldwin, 2006; Zhang et al., 2009), and as an information source for tasks such as Statistical Machine Translation (Hassan et al., 2007). The use of parser output for supertagger training has been explored for LTAG by Sarkar (2007). However, the focus of that work was on improving parser and supertagger accuracy rather than speed. 346 Previously , watch imports were denied such duty-free treatment S/S , N /N N (S[dcl]\NP)/(S[pss]\NP) (S[pss]\NP)/NP NP/NP N/N N N N (S[dcl]\NP)/NP S[pss]\NP (N /N )/(N /N ) S[adj]\NP (S[dcl]\NP)/(S[adj]\NP) (S[pss]\NP)/NP N /N (S[pt]\NP)/NP (S[dcl]\NP)/NP Figure 2: An example sentence and the sets of categories assigned by the supertagger. The first category in each column is correct and the categories used by the parser are marked in bold. The correct category for watch is included here, for expository purposes, but in fact was not provided by the supertagger. 2.2 Semi-supervised training Previous exploration of semi-supervised training in NLP has focused on improving accuracy, often for the case where only small amounts of manually labelled training data are available. One approach is co-training, in which two models with independent views of the data iteratively inform each other by labelling extra training data. Sarkar (2001) applied co-training to LTAG parsing, in which the supertagger and parser provide the two views. Steedman et al. (2003) extended the method to a variety of parser pairs. Another method is to use a re-ranker (Collins and Koo, 2002) on the output of a system to generate new training data. Like co-training, this takes advantage of a different view of the data, but the two views are not independent as the re-ranker is limited to the set of options produced by the system. This method has been used effectively to improve parsing performance on newspaper text (McClosky et al., 2006a), as well as adapting a Penn Treebank parser to a new domain (McClosky et al., 2006b). As well as using independent views of data to generate extra training data, multiple views can be used to provide constraints at test time. Hollingshead and Roark (2007) improved the accuracy of a parsing pipeline by using the output of later stages to constrain earlier stages. The only work we are aware of that uses selftraining to improve the efficiency of parsers is van Noord (2009), who adopts a similar idea to the one in this paper for improving the efficiency of a Dutch parser based on a manually constructed HPSG grammar. 3 Adaptive Supertagging The purpose of the supertagger is to cut down the search space for the parser by reducing the set of categories that must be considered for each word. A perfect supertagger would assign the correct category to every word. CCG supertaggers are about 92% accurate when assigning a single lexical category to each word (Clark and Curran, 2004). This is not accurate enough for wide coverage parsing and so a multi-tagging approach is used instead. In the final derivation, the parser uses one category from each set, and it is important to note that having the correct category in the set does not guarantee that the parser will use it. Figure 2 gives an example sentence and the sets of lexical categories supplied by the supertagger, for a particular value of β.2 The usual target of the supertagging task is to produce the top row of categories in Figure 2, the correct categories. We propose a new task that instead aims for the categories the parser will use, which are marked in bold for this case. The purpose of this new task is to improve speed. The reason speed will be improved is that we can construct models that will constrain the set of possible derivations more than the baseline model. We can construct these models because we can obtain much more of our target output, parserannotated sentences, than we could for the goldstandard supertagging task. The new target data will contain tagging errors, and so supertagging accuracy measured against the correct categories may decrease. If we obtained perfect accuracy on our new task then we would be removing all of the categories not chosen by the parser. However, parsing accuracy will not decrease since the parser will still receive the categories it would have used, and will therefore be able to form the same highest-scoring derivation (and hence will choose it). To test this idea we parsed millions of sentences 2Two of the categories for such have been left out for reasons of space, and the correct category for watch has been included for expository reasons. The fact that the supertagger does not supply this category is the reason that the parser does not analyse the sentence correctly. 347 in three domains, producing new data annotated with the categories that the parser used with the baseline model. We constructed new supertagging models that are adapted to suit the parser by training on the combination of these sets and the standard training corpora. We applied standard evaluation metrics for speed and accuracy, and explored the source of the changes in parsing performance. 4 Data In this work, we consider three domains: newswire, Wikipedia text and biomedical text. 4.1 Training and accuracy evaluation We have used Sections 02-21 of CCGbank (Hockenmaier and Steedman, 2007), the CCG version of the Penn Treebank (Marcus et al., 1993), as training data for the newspaper domain. Sections 00 and 23 were used for development and test evaluation. A further 113,346,430 tokens (4,566,241 sentences) of raw data from the Wall Street Journal section of the North American News Corpus (Graff, 1995) were parsed to produce the training data for adaptation. This text was tokenised using the C&C tools tokeniser and parsed using our baseline models. For the smaller training sets, sentences from 1988 were used as they would be most similar in style to the evaluation corpus. In all experiments the sentences from 1989 were excluded to ensure no overlap occurred with CCGbank. As Wikipedia text we have used 794,024,397 tokens (51,673,069 sentences) from Wikipedia articles. This text was processed in the same way as the NANC data to produce parser-annotated training data. For supertagger evaluation, one thousand sentences were manually annotated with CCG lexical categories and POS tags. For parser evaluation, three hundred of these sentences were manually annotated with DepBank grammatical relations (King et al., 2003) in the style of Briscoe and Carroll (2006). Both sets of annotations were produced by manually correcting the output of the baseline system. The annotation was performed by Stephen Clark and Laura Rimell. For the biomedical domain we have used several different resources. As gold standard data for supertagger evaluation we have used supertagged GENIA data (Kim et al., 2003), annotated by Rimell and Clark (2008). For parsing evaluation, grammatical relations from the BioInfer corpus were used (Pyysalo et al., 2007), with the Source Sentence Length Corpus % Range Average Variance 0-4 3.26 0.64 1.2 5-20 14.04 17.41 39.2 News 21-40 28.76 29.27 49.4 41-250 49.73 86.73 10.2 All 24.83 152.15 100.0 0-4 2.81 0.60 22.4 5-20 11.64 21.56 48.9 Wiki 21-40 28.02 28.48 24.3 41-250 49.69 77.70 4.5 All 15.33 154.57 100.0 0-4 2.98 0.75 0.9 5-20 14.54 15.14 41.3 Bio 21-40 28.49 29.34 48.0 41-250 49.17 68.34 9.8 All 24.53 139.35 100.0 Table 1: Statistics for sentences in the supertagger training data. Sentences containing more than 250 tokens were not included in our data sets. same post-processing process as Rimell and Clark (2009) to convert the C&C parser output to Stanford format grammatical relations (de Marneffe et al., 2006). For adaptive training we have used 1,900,618,859 tokens (76,739,723 sentences) from the MEDLINE abstracts tokenised by McIntosh and Curran (2008). These sentences were POS-tagged and parsed twice, once as for the newswire and Wikipedia data, and then again, using the bio-specific models developed by Rimell and Clark (2009). Statistics for the sentences in the training sets are given in Table 1. 4.2 Speed evaluation data For speed evaluation we held out three sets of sentences from each domain-specific corpus. Specifically, we used 30,000, 4,000 and 2,000 unique sentences of length 5-20, 21-40 and 41-250 tokens respectively. Speeds on these length controlled sets were combined to calculate an overall parsing speed for the text in each domain. Note that more than 20% of the Wikipedia sentences were less than five words in length and the overall distribution is skewed towards shorter sentences compared to the other corpora. 5 Evaluation We used the hybrid parsing model described in Clark and Curran (2007), and the Viterbi decoder to find the highest-scoring derivation. The multipass supertagger-parser interaction was also used. The test data was excluded from training data for the supertagger for all of the newswire and Wikipedia models. For the biomedical models ten348 fold cross validation was used. The accuracy of supertagging is measured by multi-tagging at the first β level and considering a word correct if the correct tag is amongst any of the assigned tags. For the biomedical parser evaluation we have used the parsing model and grammatical relation conversion script from Rimell and Clark (2009). Our timing measurements are calculated in two ways. Overall times were measured using the C&C parser’s timers. Individual sentence measurements were made using the Intel timing registers, since standard methods are not accurate enough for the short time it takes to parse a single sentence. To check whether changes were statistically significant we applied the test described by Chinchor (1995). This measures the probability that two sets of responses are drawn from the same distribution, where a score below 0.05 is considered significant. Models were trained on an Intel Core2Duo 3GHz with 4GB of RAM. The evaluation was performed on a dual quad-core Intel Xeon 2.27GHz with 16GB of RAM. 5.1 Tagging ambiguity optimisation The number of lexical categories assigned to a word by the CCG supertagger depends on the probabilities calculated for each category and the β level being used. Each lexical category with a probability within a factor of β of the most probable category is included. This means that the choice of β level determines the tagging ambiguity, and so has great influence on parsing speed, accuracy and coverage. Also, the tagging ambiguity produced by a β level will vary between models. A more confident model will have a more peaked distribution of category probabilities for a word, and therefore need a smaller β value to assign the same number of categories. Additionally, the C&C parser uses multiple β levels. The first pass over a sentence is at a high β level, resulting in a low tagging ambiguity. If the categories assigned are too restrictive to enable a spanning analysis, the system makes another pass with a lower β level, resulting in a higher tagging ambiguity. A maximum of five passes are made, with the β levels varying from 0.075 to 0.001. We have taken two approaches to choosing β levels. When the aim of an experiment is to improve speed, we use the system’s default β levels. While this choice means a more confident model will assign fewer tags, this simply reflects the fact that the model is more confident. It should produce similar accuracy results, but with lower ambiguity, which will lead to higher speed. For accuracy optimisation experiments we tune the β levels to produce the same average tagging ambiguity as the baseline model on Section 00 of CCGbank. Accuracy depends heavily on the number of categories supplied, so the new models are at an accuracy disadvantage if they propose fewer categories. By matching the ambiguity of the default model, we can increase accuracy at the cost of some of the speed improvements the new models obtain. 6 Results We have performed four primary sets of experiments to explore the ability of an adaptive supertagger to improve parsing speed or accuracy. In the first two experiments, we explore performance on the newswire domain, which is the source of training data for the parsing model and the baseline supertagging model. In the second set of experiments, we train on a mixture of gold standard newswire data and parser-annotated data from the target domain. In both cases we perform two experiments. The first aimed to improve speed, keeping the β levels the same. This should lead to an increase in speed as the extra training data means the models are more confident and so have lower ambiguity than the baseline model for a given β value. The second experiment aimed to improve accuracy, tuning the β levels as described in the previous section. 6.1 Newswire speed improvement In our first experiment, we trained supertagger models using Generalised Iterative Scaling (GIS) (Darroch and Ratcliff, 1972), the limited memory BFGS method (BFGS) (Nocedal and Wright, 1999), the averaged perceptron (Collins, 2002), and the margin infused relaxed algorithm (MIRA) (Crammer and Singer, 2003). Note that these are all alternative methods for estimating the local log-linear probability distributions used by the Ratnaparkhi-style tagger. We do not use global tagging models as in Lafferty et al. (2001) or Collins (2002). The training data consisted of Sections 02–21 of CCGbank and progressively larger quantities of parser-annotated NANC data – from zero to four million extra sentences. The results of these tests are presented in Table 2. 349 Ambiguity (%) Tagging Accuracy (%) F-score Speed (sents / sec) Data 0k 40k 400k 4m 0k 40k 400k 4m 0k 40k 400k 4m 0k 40k 400k 4m Baseline 1.27 96.34 85.46 39.6 BFGS 1.27 1.23 1.19 1.18 96.33 96.18 95.95 95.93 85.45 85.51 85.57 85.68 39.8 49.6 71.8 60.0 GIS 1.28 1.24 1.21 1.20 96.44 96.27 96.09 96.11 85.44 85.46 85.58 85.62 37.4 44.1 51.3 54.1 MIRA 1.30 1.24 1.17 1.13 96.44 96.14 95.56 95.18 85.44 85.40 85.38 85.42 34.1 44.8 60.2 73.3 Table 2: Speed improvements on newswire, using various amounts of parser-annotated NANC data. Sentences Av. Time Change (ms) Total Time Change (s) Sentence length 5-20 21-40 41-250 5-20 21-40 41-250 5-20 21-40 41-250 Lower tag amb. 1166 333 281 -7.54 -71.42 -183.23 -1.1 -29 -26 Earlier pass Same tag amb. 248 38 8 -2.94 -27.08 -108.28 -0.095 -1.3 -0.44 Higher tag amb. 530 33 14 -5.84 -32.25 -44.10 -0.40 -1.3 -0.31 Lower tag amb. 19288 3120 1533 -1.13 -5.18 -38.05 -2.8 -20 -30 Same pass Same tag amb. 7285 259 35 -0.29 0.94 24.57 -0.28 0.30 0.44 Higher tag amb. 1133 101 24 -0.25 2.70 8.09 -0.037 0.34 0.099 Lower tag amb. 334 114 104 0.90 7.60 -46.34 0.039 1.1 -2.5 Later pass Same tag amb. 14 1 0 1.06 4.26 n/a 0.0019 0.0053 0.0 Higher tag amb. 2 1 1 -0.13 26.43 308.03 -3.4e-05 0.033 0.16 Table 3: Breakdown of the source of changes in speed. The test sentences are divided into nine sets based on the change in parsing behaviour between the baseline model and a model trained using MIRA, Sections 02-21 of CCGbank and 4,000,000 NANC sentences. Using the default β levels we found that the perceptron-trained models lost accuracy, disqualifying them from this test. The BFGS, GIS and MIRA models produced mixed results, but no statistically significant decrease in accuracy, and as the amount of parser-annotated data was increased, parsing speed increased by up to 85%. To determine the source of the speed improvement we considered the times recorded by the timing registers. In Table 3, we have aggregated these measurements based on the change in the pass at which the sentence is parsed, and how the tagging ambiguity changes on that pass. For sentences parsed on two different passes the ambiguity comparison is at the earlier pass. The “Total Time Change” section of the table is the change in parsing time for sentences of that type when parsing ten thousand sentences from the corpus. This takes into consideration the actual distribution of sentence lengths in the corpus. Several effects can be observed in these results. 72% of sentences are parsed on the same pass, but with lower tag ambiguity (5th row in Table 3). This provides 44% of the speed improvement. Three to six times as many sentences are parsed on an earlier pass than are parsed on a later pass. This means the sentences parsed later have very little effect on the overall speed. At the same time, the average gain for sentences parsed earlier is almost always larger than the average cost for sentences parsed later. These effects combine to produce a particularly large improvement for the sentences parsed at an earlier pass. In fact, despite making up only 7% of sentences in the set, those parsed earlier with lower ambiguity provide 50% of the speed improvement. It is also interesting to note the changes for sentences parsed on the same pass, with the same ambiguity. We may expect these sentences to be parsed in approximately the same amount of time, and this is the case for the short set, but not for the two larger sets, where we see an increase in parsing time. This suggests that the categories being supplied are more productive, leading to a larger set of possible derivations. 6.2 Newswire accuracy optimised Any decrease in tagging ambiguity will generally lead to a decrease in accuracy. The parser uses a more sophisticated algorithm with global knowledge of the sentence and so we would expect it to be better at choosing categories than the supertagger. Unlike the supertagger it will exclude categories that cannot be used in a derivation. In the previous section, we saw that training the supertagger on parser output allowed us to develop models that produced the same categories, despite lower tagging ambiguity. Since they were trained on the categories the parser was able to use in derivations, these models should also now be providing categories that are more likely to be useful. This leads us to our second experiment, opti350 Tagging Accuracy (%) F-score Speed (sents / sec) NANC sents 0k 40k 400k 4m 0k 40k 400k 4m 0k 40k 400k 4m Baseline 96.34 85.46 39.6 BFGS 96.33 96.42 96.42 96.66 85.45 85.55 85.64 85.98 39.5 43.7 43.9 42.7 GIS 96.34 96.43 96.53 96.62 85.36 85.47 85.84 85.87 39.1 41.4 41.7 42.6 Perceptron 95.82 95.99 96.30 85.28 85.39 85.64 45.9 48.0 45.2 MIRA 96.23 96.29 96.46 96.63 85.47 85.45 85.55 85.84 37.7 41.4 41.4 42.9 Table 4: Accuracy optimisation on newswire, using various amounts of parser-annotated NANC data. Train Corpus Ambiguity Tag. Acc. F-score Speed (sents / sec) News Wiki Bio News Wiki Bio News Wiki Bio News Wiki Bio Baseline 1.267 1.317 1.281 96.34 94.52 90.70 85.46 80.8 75.0 39.6 50.9 35.1 News 1.126 1.151 1.130 95.18 93.56 90.07 85.42 81.2 75.2 73.3 83.9 60.3 Wiki 1.147 1.154 1.129 95.06 93.52 90.03 84.70 81.4 75.5 62.4 73.9 58.7 Bio 1.134 1.146 1.114 94.66 93.15 89.88 84.23 80.7 75.9 66.2 90.4 59.3 Table 5: Cross-corpus speed improvement, models trained with MIRA and 4,000,000 sentences. The highlighted values are the top speed for each evaluation set and results that are statistically indistinguishable from it. mising accuracy on newswire. We used the same models as in the previous experiment, but tuned the β levels as described in Section 5.1. Comparing Tables 2 and 4 we can see the influence of β level choice, and therefore tagging ambiguity. When the default β values were used ambiguity dropped consistently as more parserannotated data was used, and category accuracy dropped in the same way. Tuning the β levels to match ambiguity produces the opposite trend. Interestingly, while the decrease in supertag accuracy in the previous experiment did not translate into a decrease in F-score, the increase in tag accuracy here does translate into an increase in F-score. This indicates that the supertagger is adapting to suit the parser. In the previous experiment, the supertagger was still providing the categories the parser would have used with the baseline supertagging model, but it provided fewer other categories. Since the parser is not a perfect supertagger these other categories may in fact have been incorrect, and so supertagger accuracy goes down, without changing parsing results. Here we have allowed the supertagger to assign extra categories, which will only increase its accuracy. The increase in F-score has two sources. First, our supertagger is more accurate, and so the parser is more likely to receive category sets that can be combined into the correct derivation. Also, the supertagger has been trained on categories that the parser is able to use in derivations, which means they are more productive. As Table 6 shows, this change translates into an improvement of up to 0.75% in F-score on Section Model Tag. Acc. F-score Speed (%) (%) (sents/sec) Baseline 96.51 85.20 39.6 GIS, 4,000k NANC 96.83 85.95 42.6 BFGS, 4,000k NANC 96.91 85.90 42.7 MIRA, 4,000k NANC 96.84 85.79 42.9 Table 6: Evaluation of top models on Section 23 of CCGbank. All changes in F-score are statistically significant. 23 of CCGbank. All of the new models in the table make a statistically significant improvement over the baseline. It is also interesting to note that the results in Tables 2, 4 and 6, are similar for all of the training algorithms. However, the training times differ considerably. For all four algorithms the training time is proportional to the amount of data, but the GIS and BFGS models trained on only CCGbank took 4,500 and 4,200 seconds to train, while the equivalent perceptron and MIRA models took 90 and 95 seconds to train. 6.3 Annotation method comparison To determine whether these improvements were dependent on the annotations being produced by the parser we performed a set of tests with supertagger, rather than parser, annotated data. Three extra training sets were created by annotating newswire sentences with supertags using the baseline supertagging model. One set used the one-best tagger, and two were produced using the most probable tag for each word out of the set supplied by the multi-tagger, with variations in the β value and dictionary cutoff for the two sets. 351 Train Corpus Ambiguity Tag. Acc. F-score Speed (sents / sec) Wiki Bio News Wiki Bio News Wiki Bio News Wiki Bio Baseline 1.317 1.281 96.34 94.52 90.70 85.46 80.8 75.0 39.6 50.9 35.1 News 1.331 1.322 96.53 94.86 91.32 85.84 80.1 75.2 41.8 32.6 31.4 Wiki 1.293 1.251 96.28 94.79 91.08 85.02 81.7 75.8 40.4 37.2 37.2 Bio 1.287 1.195 96.15 94.28 91.03 84.95 80.6 76.1 39.2 52.9 26.2 Table 7: Cross-corpus accuracy optimisation, models trained with GIS and 400,000 sentences. Annotation method Tag. Acc. F-score Baseline 96.34 85.46 Parser 96.46 85.55 One-best super 95.94 85.24 Multi-tagger a 95.91 84.98 Multi-tagger b 96.00 84.99 Table 8: Comparison of annotation methods for extra data. The multi-taggers used β values 0.075 and 0.001, and dictionary cutoffs 20 and 150, for taggers a and b respectively. Corpus Speed (sents / sec) Sent length 5-20 21-40 41-250 News 242 44.8 8.24 Wiki 224 42.0 6.10 Bio 268 41.5 6.48 Table 9: Cross-corpus speed for the baseline model on data sets balanced on sentence length. As Table 8 shows, in all cases the use of supertagger-annotated data led to poorer performance than the baseline system, while the use of parser-annotated data led to an improvement in Fscore. The parser has access to a range of information that the supertagger does not, producing a different view of the data that the supertagger can productively learn from. 6.4 Cross-domain speed improvement When applying parsers out of domain they are typically slower and less accurate (Gildea, 2001). In this experiment, we attempt to increase speed on out-of-domain data. Note that for some of the results presented here it may appear that the C&C parser does not lose speed when out of domain, since the Wikipedia and biomedical corpora contain shorter sentences on average than the news corpus. However, by testing on balanced sets it is clear that speed does decrease, particularly for longer sentences, as shown in Table 9. For our domain adaptation development experiments, we considered a collection of different models; here we only present results for the best set of models. For speed improvement these were MIRA models trained on 4,000,000 parserannotated sentences from the target domain. As Table 5 shows, this training method produces models adapted to the new domain. In particular, note that models trained on Wikipedia or the biomedical data produce lower F-scores3 than the baseline on newswire. Meanwhile, on the target domain they are adapted to, these models achieve a higher F-score and parse sentences at least 45% faster than the baseline. The changes in tagging ambiguity and accuracy also show that adaptation has occurred. In all cases, the new models have lower tagging ambiguity, and lower supertag accuracy. However, on the corpus of the extra data, the performance of the adapted models is comparable to the baseline model, which means the parser is probably still be receiving the same categories that it used from the sets provided by the baseline system. 6.5 Cross-domain accuracy optimised The ambiguity tuning method used to improve accuracy on the newspaper domain can also be applied to the models trained on other domains. In Table 7, we have tested models trained using GIS and 400,000 sentences of parsed target-domain text, with β levels tuned to match ambiguity with the baseline. As for the newspaper domain, we observe increased supertag accuracy and F-score. Also, in almost every case the new models perform worse than the baseline on domains other than the one they were trained on. In some cases the models in Table 7 are less accurate than those in Table 5. This is because as well as optimising the β levels we have changed training methods. All of the training methods were tried, but only the method with the best results in newswire is included here, which for F-score when trained on 400,000 sentences was GIS. The accuracy presented so far for the biomedi3Note that the F-scores for Wikipedia and biomedical text are reported to only three significant figures as only 300 and 500 sentences respectively were available for parser evaluation. 352 Train Corpus F-score Rimell and Clark (2009) 81.5 Baseline 80.7 CCGbank + Genia 81.5 + Newswire 81.9 + Wikipedia 82.2 + Biomedical 81.7 + R&C annotated Bio 82.3 Table 10: Performance comparison for models using extra gold standard biomedical data. Models were trained with GIS and 4,000,000 extra sentences, and are tested using a POS-tagger trained on biomedical data. cal model is considerably lower than that reported by Rimell and Clark (2009). This is because no gold standard biomedical training data was used in our experiments. Table 10 shows the results of adding Rimell and Clark’s gold standard biomedical supertag data and using their biomedical POStagger. The table also shows how accuracy can be further improved by adding our parser-annotated data from the biomedical domain as well as the additional gold standard data. 7 Conclusion This work has demonstrated that an adapted supertagger can improve parsing speed and accuracy. The purpose of the supertagger is to reduce the search space for the parser. By training the supertagger on parser output, we allow the parser to reach the derivation it would have found, sooner. This approach also enables domain adaptation, improving speed and accuracy outside the original domain of the parser. The perceptron-based algorithms used in this work are also able to function online, modifying the model weights after each sentence is parsed. This could be used to construct a system that continuously adapts to the domain it is parsing. By training on parser-annotated NANC data we constructed models that were adapted to the newspaper-trained parser. The fastest model parsed sentences 1.85 times as fast and was as accurate as the baseline system. Adaptive training is also an effective method of improving performance on other domains. Models trained on parser-annotated Wikipedia text and MEDLINE text had improved performance on these target domains, in terms of both speed and accuracy. Optimising for speed or accuracy can be achieved by modifying the β levels used by the supertagger, which controls the lexical category ambiguity at each level used by the parser. The result is an accurate and efficient widecoverage CCG parser that can be easily adapted for NLP applications in new domains without manually annotating data. Acknowledgements We thank the reviewers for their helpful feedback. This work was supported by Australian Research Council Discovery grants DP0665973 and DP1097291, the Capital Markets Cooperative Research Centre, and a University of Sydney Merit Scholarship. Part of the work was completed at the Johns Hopkins University Summer Workshop and (partially) supported by National Science Foundation Grant Number IIS-0833652. References Srinivas Bangalore and Aravind K. Joshi. 1999. Supertagging: An approach to almost parsing. Computational Linguistics, 25(2):237–265. Phil Blunsom and Timothy Baldwin. 2006. Multilingual deep lexical acquisition for HPSGs via supertagging. In Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing, pages 164–171, Sydney, Australia. Ted Briscoe and John Carroll. 2006. Evaluating the accuracy of an unlexicalized statistical parser on the PARC DepBank. In Proceedings of the Poster Session of the 21st International Conference on Computational Linguistics, Sydney, Australia. John Chen, Srinivas Bangalore, and Vijay K. Shanker. 1999. New models for improving supertag disambiguation. In Proceedings of the Ninth Conference of the European Chapter of the Association for Computational Linguistics, pages 188–195, Bergen, Norway. John Chen, Srinivas Bangalore, Michael Collins, and Owen Rambow. 2002. Reranking an n-gram supertagger. In Proceedings of the 6th International Workshop on Tree Adjoining Grammars and Related Frameworks, pages 259–268, Venice, Italy. Nancy Chinchor. 1995. Statistical significance of MUC-6 results. In Proceedings of the Sixth Message Understanding Conference, pages 39–43, Columbia, MD, USA. Stephen Clark and James R. Curran. 2004. The importance of supertagging for wide-coverage CCG parsing. In Proceedings of the 20th International Conference on Computational Linguistics, pages 282– 288, Geneva, Switzerland. 353 Stephen Clark and James R. Curran. 2007. Widecoverage efficient statistical parsing with CCG and log-linear models. Computational Linguistics, 33(4):493–552. Stephen Clark, James R. Curran, and Miles Osborne. 2003. Bootstrapping POS-taggers using unlabelled data. In Proceedings of the seventh Conference on Natural Language Learning, pages 49–55, Edmonton, Canada. Michael Collins and Terry Koo. 2002. Discriminative reranking for natural language parsing. Computational Linguistics, 31(1):25–69. Michael Collins. 2002. Discriminative training methods for Hidden Markov Models: Theory and experiments with perceptron algorithms. In Proceedings of the 2002 Conference on Empirical Methods in Natural Language Processing, pages 1–8, Philadelphia, PA, USA. Koby Crammer and Yoram Singer. 2003. Ultraconservative online algorithms for multiclass problems. Journal of Machine Learning Research, 3:951–991. James R. Curran. 2004. From Distributional to Semantic Similarity. Ph.D. thesis, University of Edinburgh. John N. Darroch and David Ratcliff. 1972. Generalized iterative scaling for log-linear models. The Annals of Mathematical Statistics, 43(5):1470–1480. Marie-Catherine de Marneffe, Bill MacCartney, and Christopher D. Manning. 2006. Generating typed dependency parses from phrase structure parses. In Proceedings of the 5th International Conference on Language Resources and Evaluation, pages 449–54, Genoa, Italy. Susan Dumais, Michele Banko, Eric Brill, Jimmy Lin, and Andrew Ng. 2002. Web question answering: Is more always better? In Proceedings of the 25th International ACMSIGIR Conference on Research and Development, Tampere, Finland. Daniel Gildea. 2001. Corpus variation and parser performance. In Proceedings of the 2001 Conference on Empirical Methods in Natural Language Processing, Pittsburgh, PA, USA. David Graff. 1995. North American News Text Corpus. LDC95T21. Linguistic Data Consortium. Philadelphia, PA, USA. Hany Hassan, Khalil Sima’an, and Andy Way. 2007. Supertagged phrase-based statistical machine translation. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 288–295, Prague, Czech Republic. Julia Hockenmaier and Mark Steedman. 2007. CCGbank: A corpus of CCG derivations and dependency structures extracted from the Penn Treebank. Computational Linguistics, 33(3):355–396. Kristy Hollingshead and Brian Roark. 2007. Pipeline iteration. In Proceedings of the 45th Meeting of the Association for Computational Linguistics, pages 952–959, Prague, Czech Republic. Jin-Dong Kim, Tomoko Ohta, Yuka Teteisi, and Jun’ichi Tsujii. 2003. GENIA corpus - a semantically annotated corpus for bio-textmining. Bioinformatics, 19(1):180–182. Tracy H. King, Richard Crouch, Stefan Riezler, Mary Dalrymple, and Ronald M. Kaplan. 2003. The PARC 700 Dependency Bank. In Proceedings of the 4th International Workshop on Linguistically Interpreted Corpora, Budapest, Hungary. John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, pages 282–289, San Francisco, CA, USA. Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19(2):313–330. David McClosky, Eugene Charniak, and Mark Johnson. 2006a. Effective self-training for parsing. In Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics, Brooklyn, NY, USA. David McClosky, Eugene Charniak, and Mark Johnson. 2006b. Reranking and self-training for parser adaptation. In Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the Association for Computational Linguistics, pages 337–344, Sydney, Australia. Tara McIntosh and James R. Curran. 2008. Weighted mutual exclusion bootstrapping for domain independent lexicon and template acquisition. In Proceedings of the Australasian Language Technology Workshop, Hobart, Australia. Jorge Nocedal and Stephen J. Wright. 1999. Numerical Optimization. Springer. Sampo Pyysalo, Filip Ginter, Veronika Laippala, Katri Haverinen, Juho Heimonen, and Tapio Salakoski. 2007. On the unification of syntactic annotations under the Stanford dependency scheme: a case study on bioinfer and GENIA. In Proceedings of the ACL workshop on biological, translational, and clinical language processing, pages 25–32, Prague, Czech Republic. Adwait Ratnaparkhi. 1996. A maximum entropy partof-speech tagger. In Proceedings of the 1996 Conference on Empirical Methods in Natural Language Processing, pages 133–142, Philadelphia, PA, USA. 354 Laura Rimell and Stephen Clark. 2008. Adapting a lexicalized-grammar parser to contrasting domains. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 475–484, Honolulu, HI, USA. Laura Rimell and Stephen Clark. 2009. Porting a lexicalized-grammar parser to the biomedical domain. Journal of Biomedical Informatics, 42(5):852–865. Anoop Sarkar, Fel Xia, and Aravind K. Joshi. 2000. Some experiments on indicators of parsing complexity for lexicalized grammars. In Proceedings of the COLING Workshop on Efficiency in Large-scale Parsing Systems, pages 37–42, Luxembourg. Anoop Sarkar. 2001. Applying co-training methods to statistical parsing. In Proceedings of the Second Meeting of the North American Chapter of the Association for Computational Linguistics, pages 1–8, Pittsburgh, PA, USA. Anoop Sarkar. 2007. Combining supertagging and lexicalized tree-adjoining grammar parsing. In Srinivas Bangalore and Aravind Joshi, editors, Complexity of Lexical Descriptions and its Relevance to Natural Language Processing: A Supertagging Approach. MIT Press, Boston, MA, USA. Mark Steedman, Miles Osborne, Anoop Sarkar, Stephen Clark, Rebecca Hwa, Julia Hockenmaier, Paul Ruhlen, Stephen Baker, and Jeremiah Crim. 2003. Bootstrapping statistical parsers from small datasets. In Proceedings of the 10th Conference of the European Chapter of the Association for Computational Linguistics, pages 331–338, Budapest, Hungary. Geertjan van Noord. 2009. Learning efficient parsing. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics, pages 817–825. Association for Computational Linguistics. Yao-zhong Zhang, Takuya Matsuzaki, and Jun’ichi Tsujii. 2009. HPSG supertagging: A sequence labeling view. In Proceedings of the 11th International Conference on Parsing Technologies, pages 210–213, Paris, France. 355
|
2010
|
36
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 356–365, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Using Smaller Constituents Rather Than Sentences in Active Learning for Japanese Dependency Parsing Manabu Sassano Yahoo Japan Corporation Midtown Tower, 9-7-1 Akasaka, Minato-ku, Tokyo 107-6211, Japan [email protected] Sadao Kurohashi Graduate School of Informatics, Kyoto University Yoshida-honmachi, Sakyo-ku, Kyoto 606-8501, Japan [email protected] Abstract We investigate active learning methods for Japanese dependency parsing. We propose active learning methods of using partial dependency relations in a given sentence for parsing and evaluate their effectiveness empirically. Furthermore, we utilize syntactic constraints of Japanese to obtain more labeled examples from precious labeled ones that annotators give. Experimental results show that our proposed methods improve considerably the learning curve of Japanese dependency parsing. In order to achieve an accuracy of over 88.3%, one of our methods requires only 34.4% of labeled examples as compared to passive learning. 1 Introduction Reducing annotation cost is very important because supervised learning approaches, which have been successful in natural language processing, require typically a large number of labeled examples. Preparing many labeled examples is time consuming and labor intensive. One of most promising approaches to this issue is active learning. Recently much attention has been paid to it in the field of natural language processing. Various tasks have been targeted in the research on active learning. They include word sense disambiguation, e.g., (Zhu and Hovy, 2007), POS tagging (Ringger et al., 2007), named entity recognition (Laws and Sch¨utze, 2008), word segmentation, e.g., (Sassano, 2002), and parsing, e.g., (Tang et al., 2002; Hwa, 2004). It is the main purpose of this study to propose methods of improving active learning for parsing by using a smaller constituent than a sentence as a unit that is selected at each iteration of active learning. Typically in active learning for parsing a sentence has been considered to be a basic unit for selection. Small constituents such as chunks have not been used in sample selection for parsing. We use Japanese dependency parsing as a target task in this study since a simple and efficient algorithm of parsing is proposed and, to our knowledge, active learning for Japanese dependency parsing has never been studied. The remainder of the paper is organized as follows. Section 2 describes the basic framework of active learning which is employed in this research. Section 3 describes the syntactic characteristics of Japanese and the parsing algorithm that we use. Section 4 briefly reviews previous work on active learning for parsing and discusses several research challenges. In Section 5 we describe our proposed methods and others of active learning for Japanese dependency parsing. Section 6 describes experimental evaluation and discussion. Finally, in Section 7 we conclude this paper and point out some future directions. 2 Active Learning 2.1 Pool-based Active Learning Our base framework of active learning is based on the algorithm of (Lewis and Gale, 1994), which is called pool-based active learning. Following their sequential sampling algorithm, we show in Figure 1 the basic flow of pool-based active learning. Various methods for selecting informative examples can be combined with this framework. 2.2 Selection Algorithm for Large Margin Classifiers One of the most accurate approaches to classification tasks is an approach with large margin classifiers. Suppose that we are given data points {xi} such that the associated label yi will be either −1 or 1, and we have a hyperplane of some large margin classifier defined by {x : f(x) = 0} where the 356 1. Build an initial classifier from an initial labeled training set. 2. While resources for labeling examples are available (a) Apply the current classifier to each unlabeled example (b) Find the m examples which are most informative for the classifier (c) Have annotators label the m examples (d) Train a new classifier on all labeled examples Figure 1: Flow of the pool-based active learning Lisa-ga kare-ni ano pen-wo age-ta. Lisa-subj to him that pen-acc give-past. ID 0 1 2 3 4 Head 4 4 3 4 Figure 2: Sample sentence. An English translation is “Lisa gave that pen to him.” classification function is G(x) = sign{f(x)}. In pool-based active learning with large margin classifiers, selection of examples can be done as follows: 1. Compute f(xi) over all unlabeled examples xi in the pool. 2. Sort xi with |f(xi)| in ascending order. 3. Select top m examples. This type of selection methods with SVMs is discussed in (Tong and Koller, 2000; Schohn and Cohn, 2000). They obtain excellent results on text classification. These selection methods are simple but very effective. 3 Japanese Parsing 3.1 Syntactic Units A basic syntactic unit used in Japanese parsing is a bunsetsu, the concept of which was initially introduced by Hashimoto (1934). We assume that in Japanese we have a sequence of bunsetsus before parsing a sentence. A bunsetsu contains one or more content words and zero or more function words. A sample sentence in Japanese is shown in Figure 2. This sentence consists of five bunsetsus: Lisa-ga, kare-ni, ano, pen-wo, and age-ta where ga, ni, and wo are postpositions and ta is a verb ending for past tense. 3.2 Constraints of Japanese Dependency Analysis Japanese is a head final language and in written Japanese we usually hypothesize the following: • Each bunsetsu has only one head except the rightmost one. • Dependency links between bunsetsus go from left to right. • Dependencies do not cross one another. We can see that these constraints are satisfied in the sample sentence in Figure 2. In this paper we also assume that the above constraints hold true when we discuss algorithms of Japanese parsing and active learning for it. 3.3 Algorithm of Japanese Dependency Parsing We use Sassano’s algorithm (Sassano, 2004) for Japanese dependency parsing. The reason for this is that it is very accurate and efficient1. Furthermore, it is easy to implement. His algorithm is one of the simplest form of shift-reduce parsers and runs in linear-time.2 Since Japanese is a head final language and its dependencies are projective as described in Section 3.2, that simplification can be made. The basic flow of Sassano’s algorithm is shown in Figure 3, which is slightly simplified from the original by Sassano (2004). When we use this algorithm with a machine learning-based classifier, function Dep() in Figure 3 uses the classifier to decide whether two bunsetsus have a dependency relation. In order to prepare training examples for the trainable classifier used with his algorithm, we first have to convert a treebank to suitable labeled instances by using the algorithm in Figure 4. Note 1Iwatate et al. (2008) compare their proposed algorithm with various ones that include Sassano’s, cascaded chunking (Kudo and Matsumoto, 2002), and one in (McDonald et al., 2005). Kudo and Matsumoto (2002) compare cascaded chunking with the CYK method (Kudo and Matsumoto, 2000). After considering these results, we have concluded so far that Sassano’s is a reasonable choice for our purpose. 2Roughly speaking, Sassano’s is considered to be a simplified version, which is modified for head final languages, of Nivre’s (Nivre, 2003). Classifiers with Nivre’s are required to handle multiclass prediction, while binary classifiers can work with Sassano’s for Japanese. 357 Input: wi: bunsetsus in a given sentence. N: the number of bunsetsus. Output: hj: the head IDs of bunsetsus wj. Functions: Push(i, s): pushes i on the stack s. Pop(s): pops a value off the stack s. Dep(j, i, w): returns true when wj should modify wi. Otherwise returns false. procedure Analyze(w, N, h) var s: a stack for IDs of modifier bunsetsus begin {−1 indicates no modifier candidate} Push(−1, s); Push(0, s); for i ←1 to N −1 do begin j ←Pop(s); while (j ̸= −1 and ((i = N −1) or Dep(j, i, w)) ) do begin hj ←i; j ←Pop(s) end Push(j, s); Push(i, s) end end Figure 3: Algorithm of Japanese dependency parsing that the algorithm in Figure 4 does not generate every pair of bunsetsus.3 4 Active Learning for Parsing Most of the methods of active learning for parsing in previous work use selection of sentences that seem to contribute to the improvement of accuracy (Tang et al., 2002; Hwa, 2004; Baldridge and Osborne, 2004). Although Hwa suggests that sample selection for parsing would be improved by selecting finer grained constituents rather than sentences (Hwa, 2004), such methods have not been investigated so far. Typical methods of selecting sentences are 3We show a sample set of generated examples for training the classifier of the parser in Figure 3. By using the algorithm in Figure 4, we can obtain labeled examples from the sample sentences in Figure 2: {0, 1, “O”}, {1, 2, “O”}, {2, 3, “D”}, and {1, 3, “O”}. Please see Section 5.2 for the notation used here. For example, an actual labeled instance generated from {2, 3, “D”} will be like ”label=D, features={modifiercontent-word=ano, ..., head-content-word=pen, ...}.” Input: hi: the head IDs of bunsetsus wi. Function: Dep(j, i, w, h): returns true if hj = i. Otherwise returns false. Also prints a feature vector with a label according to hj. procedure Generate(w, N, h) begin Push(−1, s); Push(0, s); for i ←1 to N −1 do begin j ←Pop(s); while (j ̸= −1 and ((i = N −1) or Dep(j, i, w, h)) ) do begin j ←Pop(s) end Push(j, s); Push(i, s) end end Figure 4: Algorithm of generating training examples based on some entropy-based measure of a given sentence (e.g., (Tang et al., 2002)). We cannot use this kind of measures when we want to select other smaller constituents than sentences. Other bigger problem is an algorithm of parsing itself. If we sample smaller units rather than sentences, we have partially annotated sentences and have to use a parsing algorithm that can be trained from incompletely annotated sentences. Therefore, it is difficult to use some of probabilistic models for parsing. 4 5 Active Learning for Japanese Dependency Parsing In this section we describe sample selection methods which we investigated. 5.1 Sentence-wise Sample Selection Passive Selection (Passive) This method is to select sequentially sentences that appear in the training corpus. Since it gets harder for the readers to reproduce the same experimental setting, we 4We did not employ query-by-committee (QBC) (Seung et al., 1992), which is another important general framework of active learning, since the selection strategy with large margin classifiers (Section 2.2) is much simpler and seems more practical for active learning in Japanese dependency parsing with smaller constituents. 358 avoid to use random sampling in this paper. Minimum Margin Selection (Min) This method is to select sentences that contain bunsetsu pairs which have smaller margin values of outputs of the classifier used in parsing. The procedure of selection of MIN are summarized as follows. Assume that we have sentences si in the pool of unlabeled sentences. 1. Parse si in the pool with the current model. 2. Sort si with min |f(xk)| where xk are bunsetsu pairs in the sentence si. Note that xk are not all possible bunsetsu pairs in si and they are limited to bunsetsu pairs checked in the process of parsing si. 3. Select top m sentences. Averaged Margin Selection (Avg) This method is to select sentences that have smaller values of averaged margin values of outputs of the classifier in a give sentences over the number of decisions which are carried out in parsing. The difference between AVG and MIN is that for AVG we use ∑|f(xk)|/l where l is the number of calling Dep() in Figure 3 for the sentence si instead of min |f(xk)| for MIN. 5.2 Chunk-wise Sample Selection In chunk-wise sample selection, we select bunsetsu pairs rather than sentences. Bunsetsu pairs are selected from different sentences in a pool. This means that structures of sentences in the pool are partially annotated. Note that we do not use every bunsetsu pair in a sentence. When we use Sassano’s algorithm, we have to generate training examples for the classifier by using the algorithm in Figure 4. In other words, we should not sample bunsetsu pairs independently from a given sentence. Therefore, we select bunsetsu pairs that have smaller margin values of outputs given by the classifier during the parsing process. All the sentences in the pool are processed by the current parser. We cannot simply split the sentences in the pool into labeled and unlabeled ones because we do not select every bunsetsu pair in a given sentence. Naive Selection (Naive) This method is to select bunsetsu pairs that have smaller margin values of outputs of the classifier. Then it is assumed that annotators would label either “D” for the two bunsetsu having a dependency relation or “O”, which represents the two does not. Modified Simple Selection (ModSimple) Although NAIVE seems to work well, it did not (discussed later). MODSIMPLE is to select bunsetsu pairs that have smaller margin values of outputs of the classifier, which is the same as in NAIVE. The difference between MODSIMPLE and NAIVE is the way annotators label examples. Assume that we have an annotator and the learner selects some bunsetsu pair of the j-th bunsetsu and the i-th bunsetsu such that j < i. The annotator is then asked what the head of the j-th bunsetsu is. We define here the head bunsetsu is the k-th one. We differently generate labeled examples from the information annotators give according to the relation among bunsetsus j, i, and k. Below we use the notation {s, t, “D”} to denote that the s-th bunsetsu modifies the t-th one. The use of “O” instead of “D” indicates that the s-th does not modify the t-th. That is generating {s, t, “D”} means outputting an example with the label “D”. Case 1 if j < i < k, then generate {j, i, “O”} and {j, k, “D”}. Case 2 if j < i = k, then generate {j, k, “D”}. Case 3 if j < k < i, then generate {j, k, “D”}. Note that we do not generate {j, i, “O”} in this case because in Sassano’s algorithm we do not need such labeled examples if j depends on k such that k < i. Syntactically Extended Selection (Syn) This selection method is one based on MODSIMPLE and extended to generate more labeled examples for the classifier. You may notice that more labeled examples for the classifier can be generated from a single label which the annotator gives. Syntactic constraints of the Japanese language allow us to extend labeled examples. For example, suppose that we have four bunsetsus A, B, C, and D in this order. If A depends on C, i.e., the head of A is C, then it is automatically derived that B also should depend on C because the Japanese language has the no-crossing constraint for dependencies (Section 3.2). By utilizing this property we can obtain more labeled examples from a single labeled one annotators give. In the example above, we obtain {A, B, “O”} and {B, C, “D”} from {A, C, “D”}. 359 Although we can employ various extensions to MODSIMPLE, we use a rather simple extension in this research. Case 1 if (j < i < k), then generate • {j, i, “O”}, • {k −1, k, “D”} if k −1 > j, • and {j, k, “D”}. Case 2 if (j < i = k), then generate • {k −1, k, “D”} if k −1 > j, • and {j, k, “D”}. Case 3 if (j < k < i), then generate • {k −1, k, “D”} if k −1 > j, • and {j, k, “D”}. In SYN as well as MODSIMPLE, we generate examples with ”O” only for bunsetsu pairs that occur to the left of the correct head (i.e., case 1). 6 Experimental Evaluation and Discussion 6.1 Corpus In our experiments we used the Kyoto University Corpus Version 2 (Kurohashi and Nagao, 1998). Initial seed sentences and a pool of unlabeled sentences for training are taken from the articles on January 1st through 8th (7,958 sentences) and the test data is a set of sentences in the articles on January 9th (1,246 sentences). The articles on January 10th were used for development. The split of these articles for training/test/development is the same as in (Uchimoto et al., 1999). 6.2 Averaged Perceptron We used the averaged perceptron (AP) (Freund and Schapire, 1999) with polynomial kernels. We set the degree of the kernels to 3 since cubic kernels with SVM have proved effective for Japanese dependency parsing (Kudo and Matsumoto, 2000; Kudo and Matsumoto, 2002). We found the best value of the epoch T of the averaged perceptron by using the development set. We fixed T = 12 through all experiments for simplicity. 6.3 Features There are features that have been commonly used for Japanese dependency parsing among related papers, e.g., (Kudo and Matsumoto, 2002; Sassano, 2004; Iwatate et al., 2008). We also used the same features here. They are divided into three groups: modifier bunsetsu features, head bunsetsu features, and gap features. A summary of the features is described in Table 1. 6.4 Implementation We implemented a parser and a tool for the averaged perceptron in C++ and used them for experiments. We wrote the main program of active learning and some additional scripts in Perl and sh. 6.5 Settings of Active Learning For initial seed sentences, first 500 sentences are taken from the articles on January 1st. In experiments about sentence wise selection, 500 sentences are selected at each iteration of active learning and labeled5 and added into the training data. In experiments about chunk wise selection 4000 pairs of bunsetsus, which are roughly equal to the averaged number of bunsetsus in 500 sentences, are selected at each iteration of active learning. 6.6 Dependency Accuracy We use dependency accuracy as a performance measure of a parser. The dependency accuracy is the percentage of correct dependencies. This measure is commonly used for the Kyoto University Corpus. 6.7 Results and Discussion Learning Curves First we compare methods for sentence wise selection. Figure 5 shows that MIN is the best among them, while AVG is not good and similar to PASSIVE. It is observed that active learning with large margin classifiers also works well for Sassano’s algorithm of Japanese dependency parsing. Next we compare chunk-wise selection with sentence-wise one. The comparison is shown in Figure 6. Note that we must carefully consider how to count labeled examples. In sentence wise selection we obviously count the number of sentences. However, it is impossible to count such number when we label bunsetsus pairs. Therefore, we use the number of bunsetsus that have an annotated head. Although we know this may not be a completely fair comparison, we believe our choice in this experiment is reasonable 5In our experiments human annotators do not give labels. Instead, labels are given virtually from correct ones that the Kyoto University Corpus has. 360 Bunsetsu features for modifiers rightmost content word, rightmost function word, punctuation, and heads parentheses, location (BOS or EOS) Gap features distance (1, 2–5, or 6 ≤), particles, parentheses, punctuation Table 1: Features for deciding a dependency relation between two bunsetsus. Morphological features for each word (morpheme) are major part-of-speech (POS), minor POS, conjugation type, conjugation form, and surface form. for assessing the effect of reduction by chunk-wise selection. In Figure 6 NAIVE has a better learning curve compared to MIN at the early stage of learning. However, the curve of NAIVE declines at the later stage and gets worse than PASSIVE and MIN. Why does this phenomenon occur? It is because each bunsetsu pair is not independent and pairs in the same sentence are related to each other. They satisfy the constraints discussed in Section 3.2. Furthermore, the algorithm we use, i.e., Sassano’s, assumes these constraints and has the specific order for processing bunsetsu pairs as we see in Figure 3. Let us consider the meaning of {j, i, “O”} if the head of the j-th bunsetsu is the k-th one such that j < k < i. In the context of the algorithm in Figure 3, {j, i, “O”} actually means that the j-th bunsetsu modifies th l-th one such that i < l. That is “O” does not simply mean that two bunsetsus does not have a dependency relation. Therefore, we should not generate {j, i, “O”} in the case of j < k < i. Such labeled instances are not needed and the algorithm in Figure 4 does not generate them even if a fully annotated sentence is given. Based on the analysis above, we modified NAIVE and defined MODSIMPLE, where unnecessary labeled examples are not generated. Now let us compare NAIVE with MODSIMPLE (Figure 7). MODSIMPLE is almost always better than PASSIVE and does not cause a significant deterioration of accuracy unlike NAIVE.6 Comparison of MODSIMPLE and SYN is shown in Figure 8. Both exhibit a similar curve. Figure 9 shows the same comparison in terms of required queries to human annotators. It shows that SYN is better than MODSIMPLE especially at the earlier stage of active learning. Reduction of Annotations Next we examined the number of labeled bunsetsus to be required in 6We have to carefully see the curves of NAIVE and MODSIMPLE. In Figure 7 at the early stage NAIVE is slightly better than MODSIMPLE, while in Figure 9 NAIVE does not outperform MODSIMPLE. This is due to the difference of the way of accessing annotation efforts. 0.855 0.86 0.865 0.87 0.875 0.88 0.885 0.89 0 1000 2000 3000 4000 5000 6000 7000 8000 Accuracy Number of Labeled Sentences Passive Min Average Figure 5: Learning curves of methods for sentence wise selection 0.855 0.86 0.865 0.87 0.875 0.88 0.885 0.89 0 10000 20000 30000 40000 50000 Accuracy Number of bunsetsus which have a head Passive Min Naive Figure 6: Learning curves of MIN (sentence-wise) and NAIVE (chunk-wise). 361 0.855 0.86 0.865 0.87 0.875 0.88 0.885 0.89 0 10000 20000 30000 40000 50000 Accuracy Number of bunsetsus which have a head Passive ModSimple Naive Figure 7: Learning curves of NAIVE, MODSIMPLE and PASSIVE in terms of the number of bunsetsus that have a head. 0.855 0.86 0.865 0.87 0.875 0.88 0.885 0.89 0 10000 20000 30000 40000 50000 Accuracy Number of bunsetsus which have a head Passive ModSimple Syntax Figure 8: Learning curves of MODSIMPLE and SYN in terms of the number of bunsetsus which have a head. 0.855 0.86 0.865 0.87 0.875 0.88 0.885 0.89 0 10000 20000 30000 40000 50000 60000 Accuracy Number of queris to human annotators ModSimple Syntax Naive Figure 9: Comparison of MODSIMPLE and SYN in terms of the number of queries to human annotators 0 5000 10000 15000 20000 25000 30000 35000 40000 Passive Min Avg Naive Mod Simple Syn # of bunsetsus that have a head Selection strategy Figure 10: Number of labeled bunsetsus to be required to achieve an accuracy of over 88.3%. 0 5000 10000 15000 20000 25000 0 1000 2000 3000 4000 5000 6000 7000 8000 Number of Support Vectors Number of Labeled Sentences Passive Min Figure 11: Changes of number of support vectors in sentence-wise active learning 0 5000 10000 15000 20000 25000 0 10000 20000 30000 40000 50000 60000 Number of Support Vectors Number of Queries ModSimple Figure 12: Changes of number of support vectors in chunk-wise active learning (MODSIMPLE) 362 order to achieve a certain level of accuracy. Figure 10 shows that the number of labeled bunsetsus to achieve an accuracy of over 88.3% depending on the active learning methods discussed in this research. PASSIVE needs 37766 labeled bunsetsus which have a head to achieve an accuracy of 88.48%, while SYN needs 13021 labeled bunsetsus to achieve an accuracy of 88.56%. SYN requires only 34.4% of the labeled bunsetsu pairs that PASSIVE requires. Stopping Criteria It is known that increment rate of the number of support vectors in SVM indicates saturation of accuracy improvement during iterations of active learning (Schohn and Cohn, 2000). It is interesting to examine whether the observation for SVM is also useful for support vectors7 of the averaged perceptron. We plotted changes of the number of support vectors in the cases of both PASSIVE and MIN in Figure 11 and changes of the number of support vectors in the case of MODSIMPLE in Figure 12. We observed that the increment rate of support vectors mildly gets smaller. However, it is not so clear as in the case of text classification in (Schohn and Cohn, 2000). Issues on Accessing the Total Cost of Annotation In this paper, we assume that each annotation cost for dependency relations is constant. It is however not true in an actual annotation work.8 In addition, we have to note that it may be easier to annotate a whole sentence than some bunsetsu pairs in a sentence9. In a real annotation task, it will be better to show a whole sentence to annotators even when annotating some part of the sentence. Nevertheless, it is noteworthy that our research shows the minimum number of annotations in preparing training examples for Japanese dependency parsing. The methods we have proposed must be helpful when checking repeatedly annotations that are important and might be wrong or difficult to label while building an annotated cor7Following (Freund and Schapire, 1999), we use the term “support vectors” for AP as well as SVM. “Support vectors” of AP means vectors which are selected in the training phase and contribute to the prediction. 8Thus it is very important to construct models for estimating the actual annotation cost as Haertel et al. (2008) do. 9Hwa (2004) discusses similar aspects of researches on active learning. pus. They also will be useful for domain adaptation of a dependency parser.10 Applicability to Other Languages and Other Parsing Algorithms We discuss here whether or not the proposed methods and the experiments are useful for other languages and other parsing algorithms. First we take languages similar to Japanese in terms of syntax, i.e., Korean and Mongolian. These two languages are basically headfinal languages and have similar constraints in Section 3.2. Although no one has reported application of (Sassano, 2004) to the languages so far, we believe that similar parsing algorithms will be applicable to them and the discussion in this study would be useful. On the other hand, the algorithm of (Sassano, 2004) cannot be applied to head-initial languages such as English. If target languages are assumed to be projective, the algorithm of (Nivre, 2003) can be used. It is highly likely that we will invent the effective use of finer-grained constituents, e.g., head-modifier pairs, rather than sentences in active learning for Nivre’s algorithm with large margin classifiers since Sassano’s seems to be a simplified version of Nivre’s and they have several properties in common. However, syntactic constraints in European languages like English may be less helpful than those in Japanese because their dependency links do not have a single direction. Even though the use of syntactic constraints is limited, smaller constituents will still be useful for other parsing algorithms that use some deterministic methods with machine learning-based classifiers. There are many algorithms that have such a framework, which include (Yamada and Matsumoto, 2003) for English and (Kudo and Matsumoto, 2002; Iwatate et al., 2008) for Japanese. Therefore, effective use of smaller constituents in active learning would not be limited to the specific algorithm. 7 Conclusion We have investigated that active learning methods for Japanese dependency parsing. It is observed that active learning of parsing with the averaged perceptron, which is one of the large margin classifiers, works also well for Japanese dependency analysis. 10Ohtake (2006) examines heuristic methods of selecting sentences. 363 In addition, as far as we know, we are the first to propose the active learning methods of using partial dependency relations in a given sentence for parsing and we have evaluated the effectiveness of our methods. Furthermore, we have tried to obtain more labeled examples from precious labeled ones that annotators give by utilizing syntactic constraints of the Japanese language. It is noteworthy that linguistic constraints have been shown useful for reducing annotations in active learning for NLP. Experimental results show that our proposed methods have improved considerably the learning curve of Japanese dependency parsing. We are currently building a new annotated corpus with an annotation tool. We have a plan to incorporate our proposed methods to the annotation tool. We will use it to accelerate building of the large annotated corpus to improved our Japanese parser. It would be interesting to explore the use of partially labeled constituents in a sentence in another language, e.g., English, for active learning. Acknowledgements We would like to thank the anonymous reviewers and Tomohide Shibata for their valuable comments. References Jason Baldridge and Miles Osborne. 2004. Active learning and the total cost of annotation. In Proc. of EMNLP 2004, pages 9–16. Yoav Freund and Robert E. Schapire. 1999. Large margin classification using the perceptron algorithm. Machine Learning, 37(3):277–296. Robbie Haertel, Eric Ringger, Kevin Seppi, James Carroll, and Peter McClanahan. 2008. Assessing the costs of sampling methods in active learning for annotation. In Proc. of ACL-08: HLT, short papers (Companion Volume), pages 65–68. Shinkichi Hashimoto. 1934. Essentials of Japanese Grammar (Kokugoho Yousetsu) (in Japanese). Rebecca Hwa. 2004. Sample selection for statistical parsing. Computational Linguistics, 30(3):253–276. Masakazu Iwatate, Masayuki Asahara, and Yuji Matsumoto. 2008. Japanese dependency parsing using a tournament model. In Proc. of COLING 2008, pages 361–368. Taku Kudo and Yuji Matsumoto. 2000. Japanese dependency structure analysis based on support vector machines. In Proc. of EMNLP/VLC 2000, pages 18– 25. Taku Kudo and Yuji Matsumoto. 2002. Japanese dependency analysis using cascaded chunking. In Proc. of CoNLL-2002, pages 63–69. Sadao Kurohashi and Makoto Nagao. 1998. Building a Japanese parsed corpus while improving the parsing system. In Proc. of LREC-1998, pages 719–724. Florian Laws and Hinrich Sch¨utze. 2008. Stopping criteria for active learning of named entity recognition. In Proc. of COLING 2008, pages 465–472. David D. Lewis and William A. Gale. 1994. A sequential algorithm for training text classifiers. In Proc. of the Seventeenth Annual International ACMSIGIR Conference on Research and Development in Information Retrieval, pages 3–12. Ryan McDonald, Koby Crammer, and Fernando Pereira. 2005. Online large-margin training of dependency parsers. In Proc. of ACL-2005, pages 523–530. Joakim Nivre. 2003. An efficient algorithm for projective dependency parsing. In Proc. of IWPT-03, pages 149–160. Kiyonori Ohtake. 2006. Analysis of selective strategies to build a dependency-analyzed corpus. In Proc. of COLING/ACL 2006 Main Conf. Poster Sessions, pages 635–642. Eric Ringger, Peter McClanahan, Robbie Haertel, George Busby, Marc Carmen, James Carroll, Kevin Seppi, and Deryle Lonsdale. 2007. Active learning for part-of-speech tagging: Accelerating corpus annotation. In Proc. of the Linguistic Annotation Workshop, pages 101–108. Manabu Sassano. 2002. An empirical study of active learning with support vector machines for Japanese word segmentation. In Proc. of ACL-2002, pages 505–512. Manabu Sassano. 2004. Linear-time dependency analysis for Japanese. In Proc. of COLING 2004, pages 8–14. Greg Schohn and David Cohn. 2000. Less is more: Active learning with support vector machines. In Proc. of ICML-2000, pages 839–846. H. S. Seung, M. Opper, and H. Sompolinsky. 1992. Query by committee. In Proc. of COLT ’92, pages 287–294. Min Tang, Xaoqiang Luo, and Salim Roukos. 2002. Active learning for statistical natural language parsing. In Proc. of ACL-2002, pages 120–127. 364 Simon Tong and Daphne Koller. 2000. Support vector machine active learning with applications to text classification. In Proc. of ICML-2000, pages 999– 1006. Kiyotaka Uchimoto, Satoshi Sekine, and Hitoshi Isahara. 1999. Japanese dependency structure analysis based on maximum entropy models. In Proc. of EACL-99, pages 196–203. Hiroyasu Yamada and Yuji Matsumoto. 2003. Statistical dependency analysis with support vector machines. In Proc. of IWPT 2003, pages 195–206. Jingbo Zhu and Eduard Hovy. 2007. Active learning for word sense disambiguation with methods for addressing the class imbalance problem. In Proc. of EMNLP-CoNLL 2007, pages 783–790. 365
|
2010
|
37
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 366–374, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Conditional Random Fields for Word Hyphenation Nikolaos Trogkanis Computer Science and Engineering University of California, San Diego La Jolla, California 92093-0404 [email protected] Charles Elkan Computer Science and Engineering University of California, San Diego La Jolla, California 92093-0404 [email protected] Abstract Finding allowable places in words to insert hyphens is an important practical problem. The algorithm that is used most often nowadays has remained essentially unchanged for 25 years. This method is the TEX hyphenation algorithm of Knuth and Liang. We present here a hyphenation method that is clearly more accurate. The new method is an application of conditional random fields. We create new training sets for English and Dutch from the CELEX European lexical resource, and achieve error rates for English of less than 0.1% for correctly allowed hyphens, and less than 0.01% for Dutch. Experiments show that both the Knuth/Liang method and a leading current commercial alternative have error rates several times higher for both languages. 1 Introduction The task that we investigate is learning to split words into parts that are conventionally agreed to be individual written units. In many languages, it is acceptable to separate these units with hyphens, but it is not acceptable to split words arbitrarily. Another way of stating the task is that we want to learn to predict for each letter in a word whether or not it is permissible for the letter to be followed by a hyphen. This means that we tag each letter with either 1, for hyphen allowed following this letter, or 0, for hyphen not allowed after this letter. The hyphenation task is also called orthographic syllabification (Bartlett et al., 2008). It is an important issue in real-world text processing, as described further in Section 2 below. It is also useful as a preprocessing step to improve letter-tophoneme conversion, and more generally for textto-speech conversion. In the well-known NETtalk system, for example, syllable boundaries are an input to the neural network in addition to letter identities (Sejnowski and Rosenberg, 1988). Of course, orthographic syllabification is not a fundamental scientific problem in linguistics. Nevertheless, it is a difficult engineering task that is worth studying for both practical and intellectual reasons. The goal in performing hyphenation is to predict a sequence of 0/1 values as a function of a sequence of input characters. This sequential prediction task is significantly different from a standard (non-sequential) supervised learning task. There are at least three important differences that make sequence prediction difficult. First, the set of all possible sequences of labels is an exponentially large set of possible outputs. Second, different inputs have different lengths, so it is not obvious how to represent every input by a vector of the same fixed length, as is almost universal in supervised learning. Third and most important, too much information is lost if we learn a traditional classifier that makes a prediction for each letter separately. Even if the traditional classifier is a function of the whole input sequence, this remains true. In order to achieve high accuracy, correlations between neighboring predicted labels must be taken into account. Learning to predict a sequence of output labels, given a sequence of input data items, is an instance of a structured learning problem. In general, structured learning means learning to predict outputs that have internal structure. This structure can be modeled; to achieve high predictive accuracy, when there are dependencies between parts of an output, it must be modeled. Research on structured learning has been highly successful, with sequence classification as its most important and successful subfield, and with conditional random fields (CRFs) as the most influential approach to learning sequence classifiers. In the present paper, 366 we show that CRFs can achieve extremely good performance on the hyphenation task. 2 History of automated hyphenation The earliest software for automatic hyphenation was implemented for RCA 301 computers, and used by the Palm Beach Post-Tribune and Los Angeles Times newspapers in 1962. These were two different systems. The Florida system had a dictionary of 30,000 words; words not in the dictionary were hyphenated after the third, fifth, or seventh letter, because the authors observed that this was correct for many words. The California system (Friedlander, 1968) used a collection of rules based on the rules stated in a version of Webster’s dictionary. The earliest hyphenation software for a language other than English may have been a rule-based program for Finnish first used in 1964 (Jarvi, 2009). The first formal description of an algorithm for hyphenation was in a patent application submitted in 1964 (Damerau, 1964). Other early publications include (Ocker, 1971; Huyser, 1976). The hyphenation algorithm that is by far the most widely used now is due to Liang (Liang, 1983). Although this method is well-known now as the one used in TEX and its derivatives, the first version of TEX used a different, simpler method. Liang’s method was used also in troff and groff, which were the main original competitors of TEX, and is part of many contemporary software products, supposedly including Microsoft Word. Any major improvement over Liang’s method is therefore of considerable practical and commercial importance. Over the years, various machine learning methods have been applied to the hyphenation task. However, none have achieved high accuracy. One paper that presents three different learning methods is (van den Bosch et al., 1995). The lowest per-letter test error rate reported is about 2%. Neural networks have been used, but also without great success. For example, the authors of (Kristensen and Langmyhr, 2001) found that the TEX method is a better choice for hyphenating Norwegian. The highest accuracy achieved until now for the hyphenation task is by (Bartlett et al., 2008), who use a large-margin structured learning approach. Our work is similar, but was done fully independently. The accuracy we achieve is slightly higher: word-level accuracy of 96.33% compared to their 95.65% for English. Moreover, (Bartlett et al., 2008) do not address the issue that false positive hyphens are worse mistakes than false negative hyphens, which we address below. Also, they report that training on 14,000 examples requires about an hour, compared to 6.2 minutes for our method on 65,828 words. Perhaps more important for largescale publishing applications, our system is about six times faster at syllabifying new text. The speed comparison is fair because the computer we use is slightly slower than the one they used. Methods inspired by nonstatistical natural language processing research have also been proposed for the hyphenation task, in particular (Bouma, 2003; Tsalidis et al., 2004; Woestenburg, 2006; Haralambous, 2006). However, the methods for Dutch presented in (Bouma, 2003) were found to have worse performance than TEX. Moreover, our experimental results below show that the commercial software of (Woestenburg, 2006) allows hyphens incorrectly almost three times more often than TEX. In general, a dictionary based approach has zero errors for words in the dictionary, but fails to work for words not included in it. A rule-based approach requires an expert to define manually the rules and exceptions for each language, which is laborious work. Furthermore, for languages such as English where hyphenation does not systematically follow general rules, such an approach does not have good results. A pattern-learning approach, like that of TEX, infers patterns from a training list of hyphenated words, and then uses these patterns to hyphenate text. Although useful patterns are learned automatically, both the TEX learning algorithm and the learned patterns must be hand-tuned to perform well (Liang, 1983). Liang’s method is implemented in a program named PATGEN, which takes as input a training set of hyphenated words, and outputs a collection of interacting hyphenation patterns. The standard pattern collections are named hyphen.tex for American English, ukhyphen.tex for British English, and nehyph96.tex for Dutch. The precise details of how different versions of TEX and LATEX use these pattern collections to do hyphenation in practice are unclear. At a minimum, current variants of TEX improve hyphenation accuracy by disallowing hyphens in the first and last two or three letters of every word, regardless of what the PATGEN patterns recommend. 367 Despite the success of Liang’s method, incorrect hyphenations remain an issue with TEX and its current variants and competitors. For instance, incorrect hyphenations are common in the Wall Street Journal, which has the highest circulation of any newspaper in the U.S. An example is the hyphenation of the word “sudden” in this extract: It is the case that most hyphenation mistakes in the Wall Street Journal and other media are for proper nouns such as “Netflix” that do not appear in standard dictionaries, or in compound words such as “sudden-acceleration” above. 3 Conditional random fields A linear-chain conditional random field (Lafferty et al., 2001) is a way to use a log-linear model for the sequence prediction task. We use the bar notation for sequences, so ¯x means a sequence of variable length. Specifically, let ¯x be a sequence of n letters and let ¯y be a corresponding sequence of n tags. Define the log-linear model p(¯y|¯x; w) = 1 Z(¯x, w) exp X j wjFj(¯x, ¯y). The index j ranges over a large set of featurefunctions. Each such function Fj is a sum along the output sequence for i = 1 to i = n: Fj(¯x, ¯y) = n X i=1 fj(yi−1, yi, ¯x, i) where each function fj is a 0/1 indicator function that picks out specific values for neighboring tags yi−1 and yi and a particular substring of ¯x. The denominator Z(¯x, w) is a normalizing constant: Z(¯x, w) = X ¯y exp X j wjFj(¯x, ¯y) where the outer sum is over all possible labelings ¯y of the input sequence ¯x. Training a CRF means finding a weight vector w that gives the best possible predictions ¯y∗= arg max ¯y p(¯y|¯x; w) for each training example ¯x. The software we use as an implementation of conditional random fields is named CRF++ (Kudo, 2007). This implementation offers fast training since it uses L-BFGS (Nocedal and Wright, 1999), a state-of-the-art quasi-Newton method for large optimization problems. We adopt the default parameter settings of CRF++, so no development set or tuning set is needed in our work. We define indicator functions fj that depend on substrings of the input word, and on whether or not a hyphen is legal after the current and/or the previous letter. The substrings are of length 2 to 5, covering up to 4 letters to the left and right of the current letter. From all possible indicator functions we use only those that involve a substring that occurs at least once in the training data. As an example, consider the word hy-phen-ate. For this word ¯x = hyphenate and ¯y = 010001000. Suppose i = 3 so p is the current letter. Then exactly two functions fj that depend on substrings of length 2 have value 1: I(yi−1 = 1 and yi = 0 and x2x3 = yp) = 1, I(yi−1 = 1 and yi = 0 and x3x4 = ph) = 1. All other similar functions have value 0: I(yi−1 = 1 and yi = 1 and x2x3 = yp) = 0, I(yi−1 = 1 and yi = 0 and x2x3 = yq) = 0, and so on. There are similar indicator functions for substrings up to length 5. In total, 2,916,942 different indicator functions involve a substring that appears at least once in the English dataset. One finding of our work is that it is preferable to use a large number of low-level features, that is patterns of specific letters, rather than a smaller number of higher-level features such as consonant-vowel patterns. This finding is consistent with an emerging general lesson about many natural language processing tasks: the best performance is achieved with models that are discriminative, that are trained on as large a dataset as possible, and that have a very large number of parameters but are regularized (Halevy et al., 2009). When evaluating the performance of a hyphenation algorithm, one should not just count how many words are hyphenated in exactly the same way as in a reference dictionary. One should also measure separately how many legal hyphens are actually predicted, versus how many predicted hyphens are in fact not legal. Errors of the second type are false positives. For any hyphenation 368 method, a false positive hyphen is a more serious mistake than a false negative hyphen, i.e. a hyphen allowed by the lexicon that the method fails to identify. The standard Viterbi algorithm for making predictions from a trained CRF is not tuned to minimize false positives. To address this difficulty, we use the forward-backward algorithm (Sha and Pereira, 2003; Culotta and McCallum, 2004) to estimate separately for each position the probability of a hyphen at that position. Then, we only allow a hyphen if this probability is over a high threshold such as 0.9. Each hyphenation corresponds to one path through a graph that defines all 2k−1 hyphenations that are possible for a word of length k. The overall probability of a hyphen at any given location is the sum of the weights of all paths that do have a hyphen at this position, divided by the sum of the weights of all paths. The forward-backward algorithm uses the sum operator to compute the weight of a set of paths, instead of the max operator to compute the weight of a single highestweight path. In order to compute the weight of all paths that contain a hyphen at a specific location, weight 0 is assigned to all paths that do not have a hyphen at this location. 4 Dataset creation We start with the lexicon for English published by the Dutch Centre for Lexical Information at http://www.mpi.nl/world/celex. We download all English word forms with legal hyphenation points indicated by hyphens. These include plurals of nouns, conjugated forms of verbs, and compound words such as “off-line”. We separate the components of compound words and phrases, leading to 204,466 words, of which 68,744 are unique. In order to eliminate abbreviations and proper names which may not be English, we remove all words that are not fully lower-case. In particular, we exclude words that contain capital letters, apostrophes, and/or periods. This leaves 66,001 words. Among these words, 86 have two different hyphenations, and one has three hyphenations. For most of the 86 words with alternative hyphenations, these alternatives exist because different meanings of the words have different pronunciations, and the different pronunciations have different boundaries between syllables. This fact implies that no algorithm that operates on words in isolation can be a complete solution for the hyphenation task.1 We exclude the few words that have two or more different hyphenations from the dataset. Finally, we obtain 65,828 spellings. These have 550,290 letters and 111,228 hyphens, so the average is 8.36 letters and 1.69 hyphens per word. Informal inspection suggests that the 65,828 spellings contain no mistakes. However, about 1000 words follow British as opposed to American spelling. The Dutch dataset of 293,681 words is created following the same procedure as for the English dataset, except that all entries from CELEX that are compound words containing dashes are discarded instead of being split into parts, since many of these are not in fact Dutch words.2 5 Experimental design We use ten-fold cross validation for the experiments. In order to measure accuracy, we compute the confusion matrix for each method, and from this we compute error rates. We report both word-level and letter-level error rates. The wordlevel error rate is the fraction of words on which a method makes at least one mistake. The letterlevel error rate is the fraction of letters for which the method predicts incorrectly whether or not a hyphen is legal after this letter. Table 1 explains the terminology that we use in presenting our results. Precision, recall, and F1 can be computed easily from the reported confusion matrices. As an implementation of Liang’s method we use TEX Hyphenator in Java software available at http://texhyphj.sourceforge.net. We evaluate this algorithm on our entire English and Dutch datasets using the appropriate language pattern files, and not allowing a hyphen to be placed between the first lefthyphenmin and last righthyphenmin letters of each word. For 1The single word with more than two alternative hyphenations is “invalid” whose three hyphenations are in-va-lid in-val-id and in-valid. Interestingly, the Merriam–Webster online dictionary also gives three hyphenations for this word, but not the same ones: in-va-lid in-val-id invalid. The American Heritage dictionary agrees with Merriam-Webster. The disagreement illustrates that there is a certain irreducible ambiguity or subjectivity concerning the correctness of hyphenations. 2Our English and Dutch datasets are available for other researchers and practitioners to use at http://www.cs. ucsd.edu/users/elkan/hyphenation. Previously a similar but smaller CELEX-based English dataset was created by (van den Bosch et al., 1995), but that dataset is not available online currently. 369 Abbr Name Description TP true positives #hyphens predicted correctly FP false positives #hyphens predicted incorrectly TN true negatives #hyphens correctly not predicted FN false negatives #hyphens failed to be predicted owe overall word-level errors #words with at least one FP or FN swe serious word-level errors #words with at least one FP ower overall word-level error rate owe / (total #words) swer serious word-level error rate swe / (total #words) oler overall letter-level error rate (FP+FN) / (TP+TN+FP+FN) sler serious letter-level error rate FP / (TP+TN+FP+FN) Table 1: Alternative measures of accuracy. TP, TN, FP, and FN are computed by summing over the test sets of each fold of cross-validation. English the default values are 2 and 3 respectively. For Dutch the default values are both 2. The hyphenation patterns used by TeXHyphenator, which are those currently used by essentially all variants of TEX, may not be optimal for our new English and Dutch datasets. Therefore, we also do experiments with the PATGEN tool (Liang and Breitenlohner, 2008). These are learning experiments so we also use ten-fold cross validation in the same way as with CRF++. Specifically, we create a pattern file from 90% of the dataset using PATGEN, and then hyphenate the remaining 10% of the dataset using Liang’s algorithm and the learned pattern file. The PATGEN tool has many user-settable parameters. As is the case with many machine learning methods, no strong guidance is available for choosing values for these parameters. For English we use the parameters reported in (Liang, 1983). For Dutch we use the parameters reported in (Tutelaers, 1999). Preliminary informal experiments found that these parameters work better than alternatives. We also disallow hyphens in the first two letters of every word, and the last three letters for English, or last two for Dutch. We also evaluate the TALO commercial software (Woestenburg, 2006). We know of one other commercial hyphenation application, which is named Dashes.3 Unfortunately we do not have access to it for evaluation. We also cannot do a precise comparison with the method of (Bartlett et al., 2008). We do know that their training set was also derived from CELEX, and their maximum reported accuracy is slightly lower. Specifically, for English our word-level accuracy (“ower”) is 96.33% while their best (“WA”) is 95.65%. 3http://www.circlenoetics.com/dashes. aspx 6 Experimental results In Table 2 and Table 3 we report the performance of the different methods on the English and Dutch datasets respectively. Figure 1 shows how the error rate is affected by increasing the CRF probability threshold for each language. Figure 1 shows confidence intervals for the error rates. These are computed as follows. For a single Bernoulli trial the mean is p and the variance is p(1 −p). If N such trials are taken, then the observed success rate f = S/N is a random variable with mean p and variance p(1 −p)/N. For large N, the distribution of the random variable f approaches the normal distribution. Hence we can derive a confidence interval for p using the formula Pr[−z ≤ f −p p p(1 −p)/N ≤z] = c where for a 95% confidence interval, i.e. for c = 0.95, we set z = 1.96. All differences between rows in Table 2 are significant, with one exception: the serious error rates for PATGEN and TALO are not statistically significantly different. A similar conclusion applies to Table 3. For the English language, the CRF using the Viterbi path has overall error rate of 0.84%, compared to 6.81% for the TEX algorithm using American English patterns, which is eight times worse. However, the serious error rate for the CRF is less good: 0.41% compared to 0.24%. This weakness is remedied by predicting that a hyphen is allowable only if it has high probability. Figure 1 shows that the CRF can use a probability threshold up to 0.99, and still have lower overall error rate than the TEX algorithm. Fixing the probability threshold at 0.99, the CRF serious error rate is 0.04% (224 false positives) compared to 0.24% (1343 false positives) for the TEX algorithm. 370 1 2 3 4 5 6 7 8 % oler English PATGEN TeX TALO CRF 0.90 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 Probability threshold 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 % sler 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Dutch 0.90 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 Probability threshold 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 Figure 1: Total letter-level error rate and serious letter-level error rate for different values of threshold for the CRF. The left subfigures are for the English dataset, while the right ones are for the Dutch dataset. The TALO and PATGEN lines are almost identical in the bottom left subfigure. Method TP FP TN FN owe swe % ower % swer % oler % sler Place no hyphen 0 0 439062 111228 57541 0 87.41 0.00 20.21 0.00 TEX (hyphen.tex) 75093 1343 437719 36135 30337 1311 46.09 1.99 6.81 0.24 TEX (ukhyphen.tex) 70307 13872 425190 40921 31337 11794 47.60 17.92 9.96 2.52 TALO 104266 3970 435092 6962 7213 3766 10.96 5.72 1.99 0.72 PATGEN 74397 3934 435128 36831 32348 3803 49.14 5.78 7.41 0.71 CRF 108859 2253 436809 2369 2413 2080 3.67 3.16 0.84 0.41 CRF (threshold = 0.99) 83021 224 438838 28207 22992 221 34.93 0.34 5.17 0.04 Table 2: Performance on the English dataset. Method TP FP TN FN owe swe % ower % swer % oler % sler Place no hyphen 0 0 2438913 742965 287484 0 97.89 0.00 23.35 0.00 TEX (nehyph96.tex) 722789 5580 2433333 20176 20730 5476 7.06 1.86 0.81 0.18 TALO 727145 3638 2435275 15820 16346 3596 5.57 1.22 0.61 0.11 PATGEN 730720 9660 2429253 12245 20318 9609 6.92 3.27 0.69 0.30 CRF 741796 1230 2437683 1169 1443 1207 0.49 0.41 0.08 0.04 CRF (threshold = 0.99) 719710 149 2438764 23255 22067 146 7.51 0.05 0.74 0.00 Table 3: Performance on the Dutch dataset. Method TP FP TN FN owe swe % ower % swer % oler % sler PATGEN 70357 6763 432299 40871 35013 6389 53.19 9.71 8.66 1.23 CRF 104487 6518 432544 6741 6527 5842 9.92 8.87 2.41 1.18 CRF (threshold = 0.99) 75651 654 438408 35577 27620 625 41.96 0.95 6.58 0.12 Table 4: Performance on the English dataset (10-fold cross validation dividing by stem). Method TP FP TN FN owe swe % ower % swer % oler % sler PATGEN 727306 13204 2425709 15659 25363 13030 8.64 4.44 0.91 0.41 CRF 740331 2670 2436243 2634 3066 2630 1.04 0.90 0.17 0.08 CRF (threshold = 0.99) 716596 383 2438530 26369 24934 373 8.49 0.13 0.84 0.01 Table 5: Performance on the Dutch dataset (10-fold cross validation dividing by stem). Method TP FP TN FN owe swe % ower % swer % oler % sler TEX 2711 43 21433 1420 1325 43 33.13 1.08 5.71 0.17 PATGEN 2590 113 21363 1541 1466 113 36.65 2.83 6.46 0.44 CRF 4129 2 21474 2 2 2 0.05 0.05 0.02 0.01 CRF (threshold = 0.9) 4065 0 21476 66 63 0 1.58 0.00 0.26 0.00 Table 6: Performance on the 4000 most frequent English words. 371 For the English language, TALO yields overall error rate 1.99% with serious error rate 0.72%, so the standard CRF using the Viterbi path is better on both measures. The dominance of the CRF method can be increased further by using a probability threshold. Figure 1 shows that the CRF can use a probability threshold up to 0.94, and still have lower overall error rate than TALO. Using this threshold, the CRF serious error rate is 0.12% (657 false positives) compared to 0.72% (3970 false positives) for TALO. For the Dutch language, the standard CRF using the Viterbi path has overall error rate 0.08%, compared to 0.81% for the TEX algorithm. The serious error rate for the CRF is 0.04% while for TEX it is 0.18%. Figure 1 shows that any probability threshold for the CRF of 0.99 or below yields lower error rates than the TEX algorithm. Using the threshold 0.99, the CRF has serious error rate only 0.005%. For the Dutch language, the TALO method has overall error rate 0.61%. The serious error rate for TALO is 0.11%. The CRF dominance can again be increased via a high probability threshold. Figure 1 shows that this threshold can range up to 0.98, and still give lower overall error rate than TALO. Using the 0.98 threshold, the CRF has serious error rate 0.006% (206 false positives); in comparison the serious error rate of TALO is 0.11% (3638 false positives). For both languages, PATGEN has higher serious letter-level and word-level error rates than TEX using the existing pattern files. This is expected since the pattern collections included in TEX distributions have been tuned over the years to minimize objectionable errors. The difference is especially pronounced for American English, for which the standard pattern collection has been manually improved over more than two decades by many people (Beeton, 2002). Initially, Liang optimized this pattern collection extensively by upweighting the most common words and by iteratively adding exception words found by testing the algorithm against a large dictionary from an unknown publisher (Liang, 1983). One can tune PATGEN to yield either better overall error rate, or better serious error rate, but not both simultaneously, compared to the TEX algorithm using the existing pattern files for both languages. For the English dataset, if we use Liang’s parameters for PATGEN as reported in (Sojka and Sevecek, 1995), we obtain overall error rate of 6.05% and serious error rate of 0.85%. It is possible that the specific patterns used in TEX implementations today have been tuned by hand to be better than anything the PATGEN software is capable of. 7 Additional experiments This section presents empirical results following two experimental designs that are less standard, but that may be more appropriate for the hyphenation task. First, the experimental design used above has an issue shared by many CELEX-based tagging or transduction evaluations: words are randomly divided into training and test sets without being grouped by stem. This means that a method can get credit for hyphenating “accents” correctly, when “accent” appears in the training data. Therefore, we do further experiments where the folds for evaluation are divided by stem, and not by word; that is, all versions of a base form of a word appear in the same fold. Stemming uses the English and Dutch versions of the Porter stemmer (Porter, 1980).4 The 65,828 English words in our dictionary produce 27,100 unique stems, while the 293,681 Dutch words produce 169,693 unique stems. The results of these experiments are shown in Tables 4 and 5. The main evaluation in the previous section is based on a list of unique words, which means that in the results each word is equally weighted. Because cross validation is applied, errors are always measured on testing subsets that are disjoint from the corresponding training subsets. Hence, the accuracy achieved can be interpreted as the performance expected when hyphenating unknown words, i.e. rare future words. However, in real documents common words appear repeatedly. Therefore, the second lessstandard experimental design for which we report results restricts attention to the most common English words. Specifically, we consider the top 4000 words that make up about three quarters of all word appearances in the American National Corpus, which consists of 18,300,430 words from written texts of all genres.5 From the 4,471 most 4Available at http://snowball.tartarus.org/. A preferable alternative might be to use the information about the lemmas of words available directly in CELEX. 5Available at americannationalcorpus.org/ SecondRelease/data/ANC-written-count.txt 372 frequent words in this list, if we omit the words not in our dataset of 89,019 hyphenated English words from CELEX, we get 4,000 words. The words that are omitted are proper names, contractions, incomplete words containing apostrophes, and abbreviations such as DNA. These 4,000 most frequent words make up 74.93% of the whole corpus. We evaluate the following methods on the 4000 words: Liang’s method using the American patterns file hyphen.tex, Liang’s method using the patterns derived from PATGEN when trained on the whole English dataset, our CRF trained on the whole English dataset, and the same CRF with a probability threshold of 0.9. Results are shown in Table 6. In summary, TEX and PATGEN make serious errors on 43 and 113 of the 4000 words, respectively. With a threshold of 0.9, the CRF approach makes zero serious errors on these words. 8 Timings Table 7 shows the speed of the alternative methods for the English dataset. The column “Features/Patterns” in the table reports the number of feature-functions used for the CRF, or the number of patterns used for the TEX algorithm. Overall, the CRF approach is about ten times slower than the TEX algorithm, but its performance is still acceptable on a standard personal computer. All experiments use a machine having a Pentium 4 CPU at 3.20GHz and 2GB memory. Moreover, informal experiments show that CRF training would be about eight times faster if we used CRFSGD rather than CRF++ (Bottou, 2008). From a theoretical perspective, both methods have almost-constant time complexity per word if they are implemented using appropriate data structures. In TEX, hyphenation patterns are stored in a data structure that is a variant of a trie. The CRF software uses other data structures and optimizations that allow a word to be hyphenated in time that is almost independent of the number of feature-functions used. 9 Conclusions Finding allowable places in words to insert hyphens is a real-world problem that is still not fully solved in practice. The main contribution of this paper is a hyphenation method that is clearly more accurate than the currently used Knuth/Liang method. The new method is an apFeatures/ Training Testing Speed Method Patterns time (s) time (s) (ms/word) CRF 2916942 372.67 25.386 0.386 TEX (us) 4447 2.749 0.042 PATGEN 4488 33.402 2.889 0.044 TALO 8.400 0.128 Table 7: Timings for the English dataset (training and testing on the whole dataset that consists of 65,828 words). plication of CRFs, which are a major advance of recent years in machine learning. We hope that the method proposed here is adopted in practice, since the number of serious errors that it makes is about a sixfold improvement over what is currently in use. A second contribution of this paper is to provide training sets for hyphenation in English and Dutch, so other researchers can, we hope, soon invent even more accurate methods. A third contribution of our work is a demonstration that current CRF methods can be used straightforwardly for an important application and outperform state-of-the-art commercial and open-source software; we hope that this demonstration accelerates the widespread use of CRFs. References Susan Bartlett, Grzegorz Kondrak, and Colin Cherry. 2008. Automatic syllabification with structured SVMs for letter-to-phoneme conversion. Proceedings of ACL-08: HLT, pages 568–576. Barbara Beeton. 2002. Hyphenation exception log. TUGboat, 23(3). L´eon Bottou. 2008. Stochastic gradient CRF software CRFSGD. Available at http://leon.bottou. org/projects/sgd. Gosse Bouma. 2003. Finite state methods for hyphenation. Natural Language Engineering, 9(1):5–20, March. Aron Culotta and Andrew McCallum. 2004. Confidence Estimation for Information Extraction. In Susan Dumais, Daniel Marcu, and Salim Roukos, editors, HLT-NAACL 2004: Short Papers, pages 109– 112, Boston, Massachusetts, USA, May. Association for Computational Linguistics. Fred J. Damerau. 1964. Automatic Hyphenation Scheme. U.S. patent 3537076 filed June 17, 1964, issued October 1970. Gordon D. Friedlander. 1968. Automation comes to the printing and publishing industry. IEEE Spectrum, 5:48–62, April. 373 Alon Halevy, Peter Norvig, and Fernando Pereira. 2009. The Unreasonable Effectiveness of Data. IEEE Intelligent Systems, 24(2):8–12. Yannis Haralambous. 2006. New hyphenation techniques in Ω2. TUGboat, 27:98–103. Steven L. Huyser. 1976. AUTO-MA-TIC WORD DIVI-SION. SIGDOC Asterisk Journal of Computer Documentation, 3(5):9–10. Timo Jarvi. 2009. Computerized Typesetting and Other New Applications in a Publishing House. In History of Nordic Computing 2, pages 230–237. Springer. Terje Kristensen and Dag Langmyhr. 2001. Two regimes of computer hyphenation–a comparison. In Proceedings of the International Joint Conference on Neural Networks (IJCNN), volume 2, pages 1532–1535. Taku Kudo, 2007. CRF++: Yet Another CRF Toolkit. Version 0.5 available at http://crfpp. sourceforge.net/. John Lafferty, Andrew McCallum, and Fernando Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the 18th International Conference on Machine Learning (ICML), pages 282–289. Franklin M. Liang and Peter Breitenlohner, 2008. PATtern GENeration Program for the TEX82 Hyphenator. Electronic documentation of PATGEN program version 2.3 from web2c distribution on CTAN, retrieved 2008. Franklin M. Liang. 1983. Word Hy-phen-a-tion by Com-put-er. Ph.D. thesis, Stanford University. Jorge Nocedal and Stephen J. Wright. 1999. Limited memory BFGS. In Numerical Optimization, pages 222–247. Springer. Wolfgang A. Ocker. 1971. A program to hyphenate English words. IEEE Transactions on Engineering, Writing and Speech, 14(2):53–59, June. Martin Porter. 1980. An algorithm for suffix stripping. Program, 14(3):130–137. Terrence J. Sejnowski and Charles R. Rosenberg, 1988. NETtalk: A parallel network that learns to read aloud, pages 661–672. MIT Press, Cambridge, MA, USA. Fei Sha and Fernando Pereira. 2003. Shallow parsing with conditional random fields. Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology-Volume 1, pages 134– 141. Petr Sojka and Pavel Sevecek. 1995. Hyphenation in TEX–Quo Vadis? TUGboat, 16(3):280–289. Christos Tsalidis, Giorgos Orphanos, Anna Iordanidou, and Aristides Vagelatos. 2004. Proofing Tools Technology at Neurosoft S.A. ArXiv Computer Science e-prints, (cs/0408059), August. P.T.H. Tutelaers, 1999. Afbreken in TEX, hoe werkt dat nou? Available at ftp://ftp.tue.nl/pub/ tex/afbreken/. Antal van den Bosch, Ton Weijters, Jaap Van Den Herik, and Walter Daelemans. 1995. The profit of learning exceptions. In Proceedings of the 5th Belgian-Dutch Conference on Machine Learning (BENELEARN), pages 118–126. Jaap C. Woestenburg, 2006. *TALO’s Language Technology, November. Available at http://www.talo.nl/talo/download/ documents/Language_Book.pdf. 374
|
2010
|
38
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 375–383, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Enhanced word decomposition by calibrating the decision threshold of probabilistic models and using a model ensemble Sebastian Spiegler Intelligent Systems Laboratory, University of Bristol, U.K. [email protected] Peter A. Flach Intelligent Systems Laboratory, University of Bristol, U.K. [email protected] Abstract This paper demonstrates that the use of ensemble methods and carefully calibrating the decision threshold can significantly improve the performance of machine learning methods for morphological word decomposition. We employ two algorithms which come from a family of generative probabilistic models. The models consider segment boundaries as hidden variables and include probabilities for letter transitions within segments. The advantage of this model family is that it can learn from small datasets and easily generalises to larger datasets. The first algorithm PROMODES, which participated in the Morpho Challenge 2009 (an international competition for unsupervised morphological analysis) employs a lower order model whereas the second algorithm PROMODES-H is a novel development of the first using a higher order model. We present the mathematical description for both algorithms, conduct experiments on the morphologically rich language Zulu and compare characteristics of both algorithms based on the experimental results. 1 Introduction Words are often considered as the smallest unit of a language when examining the grammatical structure or the meaning of sentences, referred to as syntax and semantics, however, words themselves possess an internal structure denominated by the term word morphology. It is worthwhile studying this internal structure since a language description using its morphological formation is more compact and complete than listing all possible words. This study is called morphological analysis. According to Goldsmith (2009) four tasks are assigned to morphological analysis: word decomposition into morphemes, building morpheme dictionaries, defining morphosyntactical rules which state how morphemes can be combined to valid words and defining morphophonological rules that specify phonological changes morphemes undergo when they are combined to words. Results of morphological analysis are applied in speech synthesis (Sproat, 1996) and recognition (Hirsimaki et al., 2006), machine translation (Amtrup, 2003) and information retrieval (Kettunen, 2009). 1.1 Background In the past years, there has been a lot of interest and activity in the development of algorithms for morphological analysis. All these approaches have in common that they build a morphological model which is then applied to analyse words. Models are constructed using rule-based methods (Mooney and Califf, 1996; Muggleton and Bain, 1999), connectionist methods (Rumelhart and McClelland, 1986; Gasser, 1994) or statistical or probabilistic methods (Harris, 1955; Hafer and Weiss, 1974). Another way of classifying approaches is based on the learning aspect during the construction of the morphological model. If the data for training the model has the same structure as the desired output of the morphological analysis, in other words, if a morphological model is learnt from labelled data, the algorithm is classified under supervised learning. An example for a supervised algorithm is given by Oflazer et al. (2001). If the input data has no information towards the desired output of the analysis, the algorithm uses unsupervised learning. Unsupervised algorithms for morphological analysis are Linguistica (Goldsmith, 2001), Morfessor (Creutz, 2006) and Paramor (Monson, 2008). Minimally or semi-supervised algorithms are provided with partial information during the learning process. This 375 has been done, for instance, by Shalonova et al. (2009) who provided stems in addition to a word list in order to find multiple pre- and suffixes. A comparison of different levels of supervision for morphology learning on Zulu has been carried out by Spiegler et al. (2008). Our two algorithms, PROMODES and PROMODES-H, perform word decomposition and are based on probabilistic methods by incorporating a probabilistic generative model.1 Their parameters can be estimated from either labelled data, using maximum likelihood estimates, or from unlabelled data by expectation maximization2 which makes them either supervised or unsupervised algorithms. The purpose of this paper is an analysis of the underlying probabilistic models and the types of errors committed by each one. Furthermore, it is investigated how the decision threshold can be calibrated and a model ensemble is tested. The remainder is structured as follows. In Section 2 we introduce the probabilistic generative process and show in Sections 2.1 and 2.2 how we incorporate this process in PROMODES and PROMODES-H. We start our experiments with examining the learning behaviour of the algorithms in 3.1. Subsequently, we perform a position-wise comparison of predictions in 3.2, show how we find a better decision threshold for placing morpheme boundaries in 3.3 and combine both algorithms using a model ensemble to leverage individual strengths in 3.4. In 3.5 we examine how the single algorithms contribute to the result of the ensemble. In Section 4 we will compare our approaches to related work and in Section 5 we will draw our conclusions. 2 Probabilistic generative model Intuitively, we could say that our models describe the process of word generation from the left to the right by alternately using two dice, the first for deciding whether to place a morpheme boundary in the current word position and the second to get a corresponding letter transition. We are trying to reverse this process in order to find the underlying sequence of tosses which determine the morpheme boundaries. We are applying the notion of a prob1PROMODES stands for PRObabilistic MOdel for different DEgrees of Supervision. The H of PROMODES-H refers to Higher order. 2In (Spiegler et al., 2009; Spiegler et al., 2010a) we have presented an unsupervised version of PROMODES. abilistic generative process consisting of words as observed variables X and their hidden segmentation as latent variables Y. If a generative model is fully parameterised it can be reversed to find the underlying word decomposition by forming the conditional probability distribution Pr(Y|X). Let us first define the model-independent components. A given word w j ∈W with 1 ≤j ≤|W| consists of n letters and has m = n−1 positions for inserting boundaries. A word’s segmentation is depicted as a boundary vector bj = (bj1,...,bjm) consisting of boundary values bji ∈{0,1} with 1 ≤i ≤m which disclose whether or not a boundary is placed in position i. A letter lj,i-1 precedes the position i in w j and a letter lji follows it. Both letters l j,i-1 and lji are part of an alphabet. Furthermore, we introduce a letter transition tji which goes from l j,i-1 to lji. 2.1 PROMODES PROMODES is based on a zero-order model for boundaries bji and on a first-order model for letter transitions tji. It describes a word’s segmentation by its morpheme boundaries and resulting letter transitions within morphemes. A boundary vector bj is found by evaluating each position i with argmax bji Pr(bji|tji) = (1) argmax bji Pr(bji)Pr(tji|bji) . The first component of the equation above is the probability distribution over non-/boundaries Pr(bji). We assume that a boundary in i is inserted independently from other boundaries (zeroorder) and the graphemic representation of the word, however, is conditioned on the length of the word m j which means that the probability distribution is in fact Pr(b ji|m j). We guarantee ∑1 r=0 Pr(bji=r|m j) = 1. To simplify the notation in later explanations, we will refer to Pr(bji|m j) as Pr(bji). The second component is the letter transition probability distribution Pr(tji|bji). We suppose a first-order Markov chain consisting of transitions tji from letter lj,i-1 ∈AB to letter l ji ∈A where A is a regular letter alphabet and AB=A∪{B} includes B as an abstract morpheme start symbol which can occur in lj,i-1. For instance, the suffix ‘s’ of the verb form gets, marking 3rd person singular, would be modelled as B →s whereas a morpheme internal transition could be g →e. We 376 guarantee ∑lji∈A Pr(tji|bji)=1 with tji being a transition from a certain l j,i−1 ∈AB to lji. The advantage of the model is that instead of evaluating an exponential number of possible segmentations (2m), the best segmentation b∗ j=(b∗ j1,...,b∗ jm) is found with 2m position-wise evaluations using b∗ ji = argmax b ji Pr(bji|t ji) (2) = 1, if Pr(bji=1)Pr(tji|bji=1) > Pr(bji=0)Pr(tji|bji=0) 0, otherwise . The simplifying assumptions made, however, reduce the expressive power of the model by not allowing any dependencies on preceding boundaries or letters. This can lead to over-segmentation and therefore influences the performance of PROMODES. For this reason, we have extended the model which led to PROMODES-H, a higher-order probabilistic model. 2.2 PROMODES-H In contrast to the original PROMODES model, we also consider the boundary value bj,i-1 and modify our transition assumptions for PROMODESH in such a way that the new algorithm applies a first-order boundary model and a second-order transition model. A transition t ji is now defined as a transition from an abstract symbol in lj,i-1 ∈ {N ,B} to a letter in lji ∈A. The abstract symbol is N or B depending on whether bji is 0 or 1. This holds equivalently for letter transitions tj,i-1. The suffix of our previous example gets would be modelled N →t →B →s. Our boundary vector bj is then constructed from argmax b ji Pr(bji|tji,tj,i-1,bj,i-1) = (3) argmax b ji Pr(bji|bj,i-1)Pr(tji|bji,tj,i-1,bj,i-1) . The first component, the probability distribution over non-/boundaries Pr(bji|b j,i-1), satisfies ∑1 r=0 Pr(bji=r|bj,i-1)=1 with bj,i-1,b ji ∈{0,1}. As for PROMODES, Pr(bji|bj,i-1) is shorthand for Pr(bji|bj,i-1,m j). The second component, the letter transition probability distribution Pr(tji|bji,bj,i-1), fulfils ∑lji∈A Pr(tji|bji,tj,i-1,bj,i-1)=1 with tji being a transition from a certain lj,i−1 ∈AB to lji. Once again, we find the word’s best segmentation b∗ j in 2m evaluations with b∗ ji = argmax bji Pr(bji|tji,tj,i-1,bj,i-1) = (4) 1, if Pr(bji=1|bj,i-1)Pr(t ji|bji=1,t j,i-1,bj,i-1) > Pr(bji=0|bj,i-1)Pr(tji|bji=0,tj,i-1,bj,i-1) 0, otherwise . We will show in the experimental results that increasing the memory of the algorithm by looking at bj,i−1 leads to a better performance. 3 Experiments and Results In the Morpho Challenge 2009, PROMODES achieved competitive results on Finnish, Turkish, English and German – and scored highest on nonvowelized and vowelized Arabic compared to 9 other algorithms (Kurimo et al., 2009). For the experiments described below, we chose the South African language Zulu since our research work mainly aims at creating morphological resources for under-resourced indigenous languages. Zulu is an agglutinative language with a complex morphology where multiple prefixes and suffixes contribute to a word’s meaning. Nevertheless, it seems that segment boundaries are more likely in certain word positions. The PROMODES family harnesses this characteristic in combination with describing morphemes by letter transitions. From the Ukwabelana corpus (Spiegler et al., 2010b) we sampled 2500 Zulu words with a single segmentation each. 3.1 Learning with increasing experience In our first experiment we applied 10-fold crossvalidation on datasets ranging from 500 to 2500 words with the goal of measuring how the learning improves with increasing experience in terms of training set size. We want to remind the reader that our two algorithms are aimed at small datasets. We randomly split each dataset into 10 subsets where each subset was a test set and the corresponding 9 remaining sets were merged to a training set. We kept the labels of the training set to determine model parameters through maximum likelihood estimates and applied each model to the test set from which we had removed the answer keys. We compared results on the test set against the ground truth by counting true positive (TP), false positive (FP), true negative (TN) and 377 false negative (FN) morpheme boundary predictions. Counts were summarised using precision3, recall4 and f-measure5, as shown in Table 1. Data Precision Recall F-measure 500 0.7127±0.0418 0.3500±0.0272 0.4687±0.0284 1000 0.7435±0.0556 0.3350±0.0197 0.4614±0.0250 1500 0.7460±0.0529 0.3160±0.0150 0.4435±0.0206 2000 0.7504±0.0235 0.3068±0.0141 0.4354±0.0168 2500 0.7557±0.0356 0.3045±0.0138 0.4337±0.0163 (a) PROMODES Data Precision Recall F-measure 500 0.6983±0.0511 0.4938±0.0404 0.5776±0.0395 1000 0.6865±0.0298 0.5177±0.0177 0.5901±0.0205 1500 0.6952±0.0308 0.5376±0.0197 0.6058±0.0173 2000 0.7008±0.0140 0.5316±0.0146 0.6044±0.0110 2500 0.6941±0.0184 0.5396±0.0218 0.6068±0.0151 (b) PROMODES-H Table 1: 10-fold cross-validation on Zulu. For PROMODES we can see in Table 1a that the precision increases slightly from 0.7127 to 0.7557 whereas the recall decreases from 0.3500 to 0.3045 going from dataset size 500 to 2500. This suggests that to some extent fewer morpheme boundaries are discovered but the ones which are found are more likely to be correct. We believe that this effect is caused by the limited memory of the model which uses order zero for the occurrence of a boundary and order one for letter transitions. It seems that the model gets quickly saturated in terms of incorporating new information and therefore precision and recall do not drastically change for increasing dataset sizes. In Table 1b we show results for PROMODES-H. Across the datasets precision stays comparatively constant around a mean of 0.6949 whereas the recall increases from 0.4938 to 0.5396. Compared to PROMODES we observe an increase in recall between 0.1438 and 0.2351 at a cost of a decrease in precision between 0.0144 and 0.0616. Since both algorithms show different behaviour with increasing experience and PROMODES-H yields a higher f-measure across all datasets, we will investigate in the next experiments how these differences manifest themselves at the boundary level. 3precision = TP TP+FP. 4recall = TP TP+FN . 5 f-measure = 2·precision·recall precision+recall . TNPH
=
0.8726
TNP
=
0.9472
TPPH=
0.5394
TPP
=
0.3045
FPPH=
0.1274
FPP
=
0.0528
FNPH
=
0.4606
FNP
=
0.6955
0.3109
0.7889
0.2111
0.6891
+
0.0819
(net)
+
0.0486
(net)
0.5698
0.8828
0.4302
0.1172
Figure 1: Contingency table for PROMODES [grey with subscript P] and PROMODES-H [black with subscript PH] results including gross and net changes of PROMODES-H. 3.2 Position-wise comparison of algorithmic predictions In the second experiment, we investigated which aspects of PROMODES-H in comparison to PROMODES led to the above described differences in performance. For this reason we broke down the summary measures of precision and recall into their original components: true/false positive (TP/FP) and negative (TN/FN) counts presented in the 2 × 2 contingency table of Figure 1. For general evidence, we averaged across all experiments using relative frequencies. Note that the relative frequencies of positives (TP + FN) and negatives (TN + FP) each sum to one. The goal was to find out how predictions in each word position changed when applying PROMODES-H instead of PROMODES. This would show where the algorithms agree and where they disagree. PROMODES classifies nonboundaries in 0.9472 of the times correctly as TN and in 0.0528 of the times falsely as boundaries (FP). The algorithm correctly labels 0.3045 of the positions as boundaries (TP) and 0.6955 falsely as non-boundaries (FN). We can see that PROMODES follows a rather conservative approach. When applying PROMODES-H, the majority of the FP’s are turned into non-boundaries, however, a slightly higher number of previously correctly labelled non-boundaries are turned into false boundaries. The net change is a 0.0486 increase in FP’s which is the reason for the decrease in precision. On the other side, more false non378 boundaries (FN) are turned into boundaries than in the opposite direction with a net increase of 0.0819 of correct boundaries which led to the increased recall. Since the deduction of precision is less than the increase of recall, a better over-all performance of PROMODES-H is achieved. In summary, PROMODES predicts more accurately non-boundaries whereas PROMODES-H is better at finding morpheme boundaries. So far we have based our decision for placing a boundary in a certain word position on Equation 2 and 4 assuming that P(bji=1|...) > P(bji=0|...)6 gives the best result. However, if the underlying distribution for boundaries given the evidence is skewed, it might be possible to improve results by introducing a certain decision threshold for inserting morpheme boundaries. We will put this idea to the test in the following section. 3.3 Calibration of the decision threshold For the third experiment we slightly changed our experimental setup. Instead of dividing datasets during 10-fold cross-validation into training and test subsets with the ratio of 9:1 we randomly split the data into training, validation and test sets with the ratio of 8:1:1. We then run our experiments and measured contingency table counts. Rather than placing a boundary if P(bji=1|...) > P(bji=0|...) which corresponds to P(bji=1|...) > 0.50 we introduced a decision threshold P(bji=1|...) > h with 0 ≤h ≤1. This is based on the assumption that the underlying distribution P(bji|...) might be skewed and an optimal decision can be achieved at a different threshold. The optimal threshold was sought on the validation set and evaluated on the test set. An overview over the validation and test results is given in Table 2. We want to point out that the threshold which yields the best f-measure result on the validation set returns almost the same result on the separate test set for both algorithms which suggests the existence of a general optimal threshold. Since this experiment provided us with a set of data points where the recall varied monotonically with the threshold and the precision changed accordingly, we reverted to precision-recall curves (PR curves) from machine learning. Following Davis and Goadrich (2006) the algorithmic perfor6Based on Equation 2 and 4 we use the notation P(bji|...) if we do not want to specify the algorithm. mance can be analysed more informatively using these kinds of curves. The PR curve is plotted with recall on the x-axis and precision on the y-axis for increasing thresholds h. The PR curves for PROMODES and PROMODES-H are shown in Figure 2 on the validation set from which we learnt our optimal thresholds h∗. Points were connected for readability only – points on the PR curve cannot be interpolated linearly. In addition to the PR curves, we plotted isometrics for corresponding f-measure values which are defined as precision= f-measure·recall 2recall−f-measure and are hyperboles. For increasing f-measure values the isometrics are moving further to the top-right corner of the plot. For a threshold of h = 0.50 (marked by ‘3’) PROMODES-H has a better performance than PROMODES. Nevertheless, across the entire PR curve none of the algorithms dominates. One curve would dominate another if all data points of the dominated curve were beneath or equal to the dominating one. PROMODES has its optimal threshold at h∗= 0.36 and PROMODES-H at h∗= 0.37 where PROMODES has a slightly higher f-measure than PROMODES-H. The points of optimal f-measure performance are marked with ‘△’ on the PR curve. Prec. Recall F-meas. PROMODES validation (h=0.50) 0.7522 0.3087 0.4378 PROMODES test (h=0.50) 0.7540 0.3084 0.4378 PROMODES validation (h∗=0.36) 0.5857 0.7824 0.6699 PROMODES test (h∗=0.36) 0.5869 0.7803 0.6699 PROMODES-H validation (h=0.50) 0.6983 0.5333 0.6047 PROMODES-H test (h=0.50) 0.6960 0.5319 0.6030 PROMODES-H validation (h∗=0.37) 0.5848 0.7491 0.6568 PROMODES-H test (h∗=0.37) 0.5857 0.7491 0.6574 Table 2: PROMODES and PROMODES-H on validation and test set. Summarizing, we have shown that both algorithms commit different errors at the word position level whereas PROMODES is better in predicting non-boundaries and PROMODES-H gives better results for morpheme boundaries at the default threshold of h = 0.50. In this section, we demonstrated that across different decision thresholds h for P(bji=1|...) > h none of algorithms dominates the other one, and at the optimal threshold PROMODES achieves a slightly higher performance than PROMODES-H. The question which arises is whether we can combine PROMODES and PROMODES-H in an ensemble that leverages individual strengths of both. 379 0.4 0.5 0.6 0.7 0.8 0.9 1 0.4 0.5 0.6 0.7 0.8 0.9 1 Recall Precision Promodes Promodes−H Promodes−E F−measure isometrics Default result Optimal result (h*) Figure 2: Precision-recall curves for algorithms on validation set. 3.4 A model ensemble to leverage individual strengths A model ensemble is a set of individually trained classifiers whose predictions are combined when classifying new instances (Opitz and Maclin, 1999). The idea is that by combining PROMODES and PROMODES-H, we would be able to avoid certain errors each model commits by consulting the other model as well. We introduce PROMODES-E as the ensemble of PROMODES and PROMODESH. PROMODES-E accesses the individual probabilities Pr(bji=1|...) and simply averages them: Pr(bji=1|tji)+Pr(bji=1|tji,bj,i-1,tj,i-1) 2 > h . As before, we used the default threshold h = 0.50 and found the calibrated threshold h∗= 0.38, marked with ‘3’ and ‘△’ in Figure 2 and shown in Table 3. The calibrated threshold improves the f-measure over both PROMODES and PROMODES-H. Prec. Recall F-meas. PROMODES-E validation (h=0.50) 0.8445 0.4328 0.5723 PROMODES-E test (h=0.50) 0.8438 0.4352 0.5742 PROMODES-E validation (h∗=0.38) 0.6354 0.7625 0.6931 PROMODES-E test (h∗=0.38) 0.6350 0.7620 0.6927 Table 3: PROMODES-E on validation and test set. The optimal solution applying h∗= 0.38 is more balanced between precision and recall and boosted the original result by 0.1185 on the test set. Compared to its components PROMODES and PROMODES-H the f-measure increased by 0.0228 and 0.0353 on the test set. In short, we have shown that by combining PROMODES and PROMODES-H and finding the optimal threshold, the ensemble PROMODES-E gives better results than the individual models themselves and therefore manages to leverage the individual strengths of both to a certain extend. However, can we pinpoint the exact contribution of each individual algorithm to the improved result? We try to find an answer to this question in the analysis of the subsequent section. 3.5 Analysis of calibrated algorithms and their model ensemble For the entire dataset of 2500 words, we have examined boundary predictions dependent on the relative word position. In Figure 3 and 4 we have plotted the absolute counts of correct boundaries (TP) and non-boundaries (TN) which PROMODES predicted but not PROMODES-H, and vice versa, as continuous lines. We furthermore provided the number of individual predictions which were ultimately adopted by PROMODES-E in the ensemble as dashed lines. In Figure 3a we can see for the default threshold that PROMODES performs better in predicting non-boundaries in the middle and the end of the word in comparison to PROMODES-H. Figure 3b 380 shows the statistics for correctly predicted boundaries. Here, PROMODES-H outperforms PROMODES in predicting correct boundaries across the entire word length. After the calibration, shown in Figure 4a, PROMODES-H improves the correct prediction of non-boundaries at the beginning of the word whereas PROMODES performs better at the end. For the boundary prediction in Figure 4b the signal disappears after calibration. Concluding, it appears that our test language Zulu has certain features which are modelled best with either a lower or higher-order model. Therefore, the ensemble leveraged strengths of both algorithms which led to a better overall performance with a calibrated threshold. 4 Related work We have presented two probabilistic generative models for word decomposition, PROMODES and PROMODES-H. Another generative model for morphological analysis has been described by Snover and Brent (2001) and Snover et al. (2002), however, they were interested in finding paradigms as sets of mutual exclusive operations on a word form whereas we are describing a generative process using morpheme boundaries and resulting letter transitions. Moreover, our probabilistic models seem to resemble Hidden Markov Models (HMMs) by having certain states and transitions. The main difference is that we have dependencies between states as well as between emissions whereas in HMMs emissions only depend on the underlying state. Combining different morphological analysers has been performed, for example, by Atwell and Roberts (2006) and Spiegler et al. (2009). Their approaches, though, used majority vote to decide whether a morpheme boundary is inserted in a certain word position or not. The algorithms themselves were treated as black-boxes. Monson et al. (2009) described an indirect approach to probabilistically combine ParaMor (Monson, 2008) and Morfessor (Creutz, 2006). They used a natural language tagger which was trained on the output of ParaMor and Morfessor. The goal was to mimic each algorithm since ParaMor is rule-based and there is no access to Morfessor’s internally used probabilities. The tagger would then return a probability for starting a new morpheme in a certain position based on the original algorithm. These probabilities in combination with a threshold, learnt on a different dataset, were used to merge word analyses. In contrast, our ensemble algorithm PROMODES-E directly accesses the probabilistic framework of each algorithm and combines them based on an optimal threshold learnt on a validation set. 5 Conclusions We have presented a method to learn a calibrated decision threshold from a validation set and demonstrated that ensemble methods in connection with calibrated decision thresholds can give better results than the individual models themselves. We introduced two algorithms for word decomposition which are based on generative probabilistic models. The models consider segment boundaries as hidden variables and include probabilities for letter transitions within segments. PROMODES contains a lower order model whereas PROMODES-H is a novel development of PROMODES with a higher order model. For both algorithms, we defined the mathematical model and performed experiments on language data of the morphologically complex language Zulu. We compared the performance on increasing training set sizes and analysed for each word position whether their boundary prediction agreed or disagreed. We found out that PROMODES was better in predicting non-boundaries and PROMODESH gave better results for morpheme boundaries at a default decision threshold. At an optimal decision threshold, however, both yielded a similar f-measure result. We then performed a further analysis based on relative word positions and found out that the calibrated PROMODES-H predicted non-boundaries better for initial word positions whereas the calibrated PROMODES for midand final word positions. For boundaries, the calibrated algorithms had a similar behaviour. Subsequently, we showed that a model ensemble of both algorithms in conjunction with finding an optimal threshold exceeded the performance of the single algorithms at their individually optimal threshold. Acknowledgements We would like to thank Narayanan Edakunni and Bruno Gol´enia for discussions concerning this paper as well as the anonymous reviewers for their comments. The research described was sponsored by EPSRC grant EP/E010857/1 Learning the morphology of complex synthetic languages. 381 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 100 200 300 400 500 600 700 800 Relative word position Absolute true negatives (TN) Performance on non−boundaries, default threshold Promodes (unique TN) Promodes−H (unique TN) Promodes and Promodes−E (unique TN) Promodes−H and Promodes−E (unique TN) (a) True negatives, default 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 100 200 300 400 500 600 700 800 Relative word position Absolute true positives (TP) Performance on boundaries, default threshold Promodes (unique TP) Promodes−H (unique TP) Promodes and Promodes−E (unique TP) Promodes−H and Promodes−E (unique TP) (b) True positives, default Figure 3: Analysis of results using default threshold. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 100 200 300 400 500 600 700 800 Relative word position Absolute true negatives (TN) Performance on non−boundaries, calibrated threshold Promodes (unique TN) Promodes−H (unique TN) Promodes and Promodes−E (unique TN) Promodes−H and Promodes−E (unique TN) (a) True negatives, calibrated 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 100 200 300 400 500 600 700 800 Relative word position Absolute true positives (TP) Performance on boundaries, calibrated threshold Promodes (unique TP) Promodes−H (unique TP) Promodes and Promodes−E (unique TP) Promodes−H and Promodes−E (unique TP) (b) True positives, calibrated Figure 4: Analysis of results using calibrated threshold. 382 References J. W. Amtrup. 2003. Morphology in machine translation systems: Efficient integration of finite state transducers and feature structure descriptions. Machine Translation, 18(3):217–238. E. Atwell and A. Roberts. 2006. Combinatory hybrid elementary analysis of text (CHEAT). Proceedings of the PASCAL Challenges Workshop on Unsupervised Segmentation of Words into Morphemes, Venice, Italy. M. Creutz. 2006. Induction of the Morphology of Natural Language: Unsupervised Morpheme Segmentation with Application to Automatic Speech Recognition. Ph.D. thesis, Helsinki University of Technology, Espoo, Finland. J. Davis and M. Goadrich. 2006. The relationship between precision-recall and ROC curves. International Conference on Machine Learning, Pittsburgh, PA, 233–240. M. Gasser. 1994. Modularity in a connectionist model of morphology acquisition. Proceedings of the 15th conference on Computational linguistics, 1:214–220. J. Goldsmith. 2001. Unsupervised learning of the morphology of a natural language. Computational Linguistics, 27:153–198. J. Goldsmith. 2009. The Handbook of Computational Linguistics, chapter Segmentation and morphology. Blackwell. M. A. Hafer and S. F. Weiss. 1974. Word segmentation by letter successor varieties. Information Storage and Retrieval, 10:371–385. Z. S. Harris. 1955. From phoneme to morpheme. Language, 31(2):190–222. T. Hirsimaki, M. Creutz, V. Siivola, M. Kurimo, S. Virpioja, and J. Pylkkonen. 2006. Unlimited vocabulary speech recognition with morph language models applied to Finnish. Computer Speech And Language, 20(4):515–541. K. Kettunen. 2009. Reductive and generative approaches to management of morphological variation of keywords in monolingual information retrieval: An overview. Journal of Documentation, 65:267 – 290. M. Kurimo, S. Virpioja, and V. T. Turunen. 2009. Overview and results of Morpho Challenge 2009. Working notes for the CLEF 2009 Workshop, Corfu, Greece. C. Monson, K. Hollingshead, and B. Roark. 2009. Probabilistic ParaMor. Working notes for the CLEF 2009 Workshop, Corfu, Greece. C. Monson. 2008. ParaMor: From Paradigm Structure To Natural Language Morphology Induction. Ph.D. thesis, Language Technologies Institute, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA. R. J. Mooney and M. E. Califf. 1996. Learning the past tense of English verbs using inductive logic programming. Symbolic, Connectionist, and Statistical Approaches to Learning for Natural Language Processing, 370–384. S. Muggleton and M. Bain. 1999. Analogical prediction. Inductive Logic Programming: 9th International Workshop, ILP-99, Bled, Slovenia, 234. K. Oflazer, S. Nirenburg, and M. McShane. 2001. Bootstrapping morphological analyzers by combining human elicitation and machine learning. Computational. Linguistics, 27(1):59–85. D. Opitz and R. Maclin. 1999. Popular ensemble methods: An empirical study. Journal of Artificial Intelligence Research, 11:169–198. D. E. Rumelhart and J. L. McClelland. 1986. On learning the past tenses of English verbs. MIT Press, Cambridge, MA, USA. K. Shalonova, B. Gol´enia, and P. A. Flach. 2009. Towards learning morphology for under-resourced fusional and agglutinating languages. IEEE Transactions on Audio, Speech, and Language Processing, 17(5):956965. M. G. Snover and M. R. Brent. 2001. A Bayesian model for morpheme and paradigm identification. Proceedings of the 39th Annual Meeting on Association for Computational Linguistics, 490 – 498. M. G. Snover, G. E. Jarosz, and M. R. Brent. 2002. Unsupervised learning of morphology using a novel directed search algorithm: Taking the first step. Proceedings of the ACL-02 workshop on Morphological and phonological learning, 6:11–20. S. Spiegler, B. Gol´enia, K. Shalonova, P. A. Flach, and R. Tucker. 2008. Learning the morphology of Zulu with different degrees of supervision. IEEE Workshop on Spoken Language Technology. S. Spiegler, B. Gol´enia, and P. A. Flach. 2009. Promodes: A probabilistic generative model for word decomposition. Working Notes for the CLEF 2009 Workshop, Corfu, Greece. S. Spiegler, B. Gol´enia, and P. A. Flach. 2010a. Unsupervised word decomposition with the Promodes algorithm. In Multilingual Information Access Evaluation Vol. I, CLEF 2009, Corfu, Greece, Lecture Notes in Computer Science, Springer. S. Spiegler, A. v. d. Spuy, and P. A. Flach. 2010b. Ukwabelana - An open-source morphological Zulu corpus. in review. R. Sproat. 1996. Multilingual text analysis for text-tospeech synthesis. Nat. Lang. Eng., 2(4):369–380. 383
|
2010
|
39
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 30–39, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Computing weakest readings Alexander Koller Cluster of Excellence Saarland University [email protected] Stefan Thater Dept. of Computational Linguistics Saarland University [email protected] Abstract We present an efficient algorithm for computing the weakest readings of semantically ambiguous sentences. A corpus-based evaluation with a large-scale grammar shows that our algorithm reduces over 80% of sentences to one or two readings, in negligible runtime, and thus makes it possible to work with semantic representations derived by deep large-scale grammars. 1 Introduction Over the past few years, there has been considerable progress in the ability of manually created large-scale grammars, such as the English Resource Grammar (ERG, Copestake and Flickinger (2000)) or the ParGram grammars (Butt et al., 2002), to parse wide-coverage text and assign it deep semantic representations. While applications should benefit from these very precise semantic representations, their usefulness is limited by the presence of semantic ambiguity: On the Rondane Treebank (Oepen et al., 2002), the ERG computes an average of several million semantic representations for each sentence, even when the syntactic analysis is fixed. The problem of appropriately selecting one of them to work with would ideally be solved by statistical methods (Higgins and Sadock, 2003) or knowledge-based inferences. However, no such approach has been worked out in sufficient detail to support the disambiguation of treebank sentences. As an alternative, Bos (2008) proposes to compute the weakest reading of each sentence and then use it instead of the “true” reading of the sentence. This is based on the observation that the readings of a semantically ambiguous sentence are partially ordered with respect to logical entailment, and the weakest readings – the minimal (least informative) readings with respect to this order – only express “safe” information that is common to all other readings as well. However, when a sentence has millions of readings, finding the weakest reading is a hard problem. It is of course completely infeasible to compute all readings and compare all pairs for entailment; but even the best known algorithm in the literature (Gabsdil and Striegnitz, 1999) is only an optimization of this basic strategy, and would take months to compute the weakest readings for the sentences in the Rondane Treebank. In this paper, we propose a new, efficient approach to the problem of computing weakest readings. We follow an underspecification approach to managing ambiguity: Rather than deriving all semantic representations from the syntactic analysis, we work with a single, compact underspecified semantic representation, from which the semantic representations can then be extracted by need. We then approximate entailment with a rewrite system that rewrites readings into logically weaker readings; the weakest readings are exactly those readings that cannot be rewritten into some other reading any more (the relative normal forms). We present an algorithm that computes the relative normal forms, and evaluate it on the underspecified descriptions that the ERG derives on a 624-sentence subcorpus of the Rondane Treebank. While the mean number of scope readings in the subcorpus is in the millions, our system computes on average 4.5 weakest readings for each sentence, in less than twenty milliseconds; over 80% of all sentences are reduced to at most two weakest readings. In other words, we make it feasible for the first time to build an application that uses the individual (weakest) semantic representations computed by the ERG, both in terms of the remaining ambiguity and in terms of performance. Our technique is not limited to the ERG, but should be applicable to other underspecification-based grammars as well. Technically, we use underspecified descriptions that are regular tree grammars derived from dominance graphs (Althaus et al., 2003; Koller et al., 30 2008). We compute the weakest readings by intersecting these grammars with other grammars representing the rewrite rules. This approach can be used much more generally than just for the computation of weakest readings; we illustrate this by showing how a more general version of the redundancy elimination algorithm by Koller et al. (2008) can be seen as a special case of our construction. Thus our system can serve as a general framework for removing unintended readings from an underspecified representation. The paper is structured as follows. Section 2 starts by reviewing related work. We recall dominance graphs, regular tree grammars, and the basic ideas of underspecification in Section 3, before we show how to compute weakest readings (Section 4) and logical equivalences (Section 5). In Section 6, we define a weakening rewrite system for the ERG and evaluate it on the Rondane Treebank. Section 7 concludes and points to future work. 2 Related work The idea of deriving a single approximative semantic representation for ambiguous sentences goes back to Hobbs (1983); however, Hobbs only works his algorithm out for a restricted class of quantifiers, and his representations can be weaker than our weakest readings. Rules that weaken one reading into another were popular in the 1990s underspecification literature (Reyle, 1995; Monz and de Rijke, 2001; van Deemter, 1996) because they simplify logical reasoning with underspecified representations. From a linguistic perspective, Kempson and Cormack (1981) even go so far as to claim that the weakest reading should be taken as the “basic” reading of a sentence, and the other readings only seen as pragmatically licensed special cases. The work presented here is related to other approaches that reduce the set of readings of an underspecified semantic representation (USR). Koller and Niehren (2000) showed how to strengthen a dominance constraint using information about anaphoric accessibility; later, Koller et al. (2008) presented and evaluated an algorithm for redundancy elimination, which removes readings from an USR based on logical equivalence. Our system generalizes the latter approach and applies it to a new inference problem (weakest readings) which they could not solve. This paper builds closely upon Koller and Thater (2010), which lays the formal groundwork for the ∀x sampley seex,y ∃y repr-ofx,z ∃z compz 2 4 3 5 6 7 8 ¬ 1 Figure 1: A dominance graph describing the five readings of the sentence “it is not the case that every representative of a company saw a sample.” work presented here. Here we go beyond that paper by applying a concrete implementation of our RTG construction for weakest readings to a real-world grammar, evaluating the system on practical inputs, and combining weakest readings with redundancy elimination. 3 Underspecification This section briefly reviews two formalisms for specifying sets of trees: dominance graphs and regular tree grammars. Both of these formalisms can be used to model scope ambiguities compactly by regarding the semantic representations of a sentence as trees. Some example trees are shown in Fig. 2. These trees can be read as simplified formulas of predicate logic, or as formulas involving generalized quantifiers (Barwise and Cooper, 1981). Formally, we assume a ranked signature Σ of tree constructors { f,g,a,...}, each of which is equipped with an arity ar(f) ≥0. We take a (finite constructor) tree t as a finite tree in which each node is labelled with a symbol of Σ, and the number of children of the node is exactly the arity of this symbol. For instance, the signature of the trees in Fig. 1 is {∀x|2,∃y|2,compz|0,...}. Finite constructor trees can be seen as ground terms over Σ that respect the arities. We write T(Σ) for the finite constructor trees over Σ. 3.1 Dominance graphs A (labelled) dominance graph D (Althaus et al., 2003) is a directed graph that consists of a collection of trees called fragments, plus dominance edges relating nodes in different fragments. We distinguish the roots WD of the fragments from their holes, which are the unlabelled leaves. We write LD : WD →Σ for the labeling function of D. The basic idea behind using dominance graphs to model scope underspecification is to specify 31 (a) (b) ∃y ∀x repr-ofx,z compz sampley seex,y ¬ repr-ofx,z compz seex,y sampley ∃z ¬ ∃y ∀x ∃z [+] [-] [-] [-] [-] [-] [-] [-] [+] [+] [-] [-] [-] [+] [+] [+] (c) compz repr-ofx,z seex,y sampley ¬ ∃y ∀x ∃z [+] [-] [-] [-] [-] [-] [-] [+] (e) sampley seex,y repr-ofx,z compz ¬ ∃y ∀x ∃z [+] [-] [+] [-] [-] [-] [+] [+] (d) compz repr-ofx,z seex,y sampley ¬ ∃y ∀x ∃z [+] [-] [-] [-] [-] [-] [-] [+] Figure 2: The five configurations of the dominance graph in Fig. 1. the “semantic material” common to all readings as fragments, plus dominance relations between these fragments. An example dominance graph D is shown in Fig. 1. It represents the five readings of the sentence “it is not the case that every representative of a company saw a sample.” Each reading is encoded as a (labeled) configuration of the dominance graph, which can be obtained by “plugging” the tree fragments into each other, in a way that respects the dominance edges: The source node of each dominance edge must dominate (be an ancestor of) the target node in each configuration. The trees in Fig. 2 are the five labeled configurations of the example graph. 3.2 Regular tree grammars Regular tree grammars (RTGs) are a general grammar formalism for describing languages of trees (Comon et al., 2007). An RTG is a 4-tuple G = (S,N,Σ,P), where N and Σ are nonterminal and terminal alphabets, S ∈N is the start symbol, and P is a finite set of production rules. Unlike in context-free string grammars (which look superficially the same), the terminal symbols are tree constructors from Σ. The production rules are of the form A →t, where A is a nonterminal and t is a tree from T(Σ∪N); nonterminals count as having arity zero, i.e. they must label leaves. A derivation starts with a tree containing a single node labeled with S. Then in each step of the derivation, some leaf u which is labelled with a nonterminal A is expanded with a rule A →t; this results in a new tree in which u has been replaced by t, and the derivation proceeds with this new tree. The language L(G) generated by the grammar is the set of all trees in T(Σ) that can be derived in this way. Fig. 3 shows an RTG as an example. This grammar uses sets of root names from D as nonterminal symbols, and generates exactly the five configurations of the graph in Fig. 1. The languages that can be accepted by regular tree grammars are called regular tree languages {1,2,3,4,5,6,7,8} →¬({2,3,4,5,6,7,8}) {2,3,4,5,6,7,8} →∀x({4,5,6},{3,7,8}) {2,3,4,5,6,7,8} →∃y({7},{2,4,5,6,8}) {2,3,4,5,6,7,8} →∃z({5},{2,3,6,7,8}) {2,4,5,6,8} →∀x({4,5,6},{8}) | ∃z({5},{2,6,8}) {2,3,6,7,8} →∀x({6},{3,7,8}) | ∃y({7},{2,6,8}) {2,6,8} →∀x({6},{8}) {3,7,8} →∃y({7},{8}) {4,5,6} →∃z({5},{6}) {5} →compz {7} →sampley {6} →repr-ofx,z {8} →seex,y Figure 3: A regular tree grammar that generates the five trees in Fig. 2. (RTLs), and regular tree grammars are equivalent to finite tree automata, which are defined essentially like the well-known finite string automata, except that they assign states to the nodes in a tree rather than the positions in a string. Regular tree languages enjoy many of the closure properties of regular string languages. In particular, we will later exploit that RTLs are closed under intersection and complement. 3.3 Dominance graphs as RTGs An important class of dominance graphs are hypernormally connected (hnc) dominance graphs (Koller et al., 2003). The precise definition of hnc graphs is not important here, but note that virtually all underspecified descriptions that are produced by current grammars are hypernormally connected (Flickinger et al., 2005), and we will restrict ourselves to hnc graphs for the rest of the paper. Every hypernormally connected dominance graph D can be automatically translated into an equivalent RTG GD that generates exactly the same configurations (Koller et al., 2008); the RTG in Fig. 3 is an example. The nonterminals of GD are 32 always hnc subgraphs of D. In the worst case, GD can be exponentially bigger than D, but in practice it turns out that the grammar size remains manageable: even the RTG for the most ambiguous sentence in the Rondane Treebank, which has about 4.5 × 1012 scope readings, has only about 75 000 rules and can be computed in a few seconds. 4 Computing weakest readings Now we are ready to talk about computing the weakest readings of a hypernormally connected dominance graph. We will first explain how we approximate logical weakening with rewrite systems. We will then discuss how weakest readings can be computed efficiently as the relative normal forms of these rewrite systems. 4.1 Weakening rewrite systems The different readings of a sentence with a scope ambiguity are not a random collection of formulas; they are partially ordered with respect to logical entailment, and are structurally related in a way that allows us to model this entailment relation with simpler technical means. To illustrate this, consider the five configurations in Fig. 2. The formula represented by (d) logically entails (c); we say that (c) is a weaker reading than (d) because it is satisfied by more models. Similar entailment relations hold between (d) and (e), (e) and (b), and so on (see also Fig. 5). We can define the weakest readings of the dominance graph as the minimal elements of the entailment order; in the example, these are (b) and (c). Weakest readings capture “safe” information in that whichever reading of the sentence the speaker had in mind, any model of this reading also satisfies at least one weakest reading; in the absence of convincing disambiguation methods, they can therefore serve as a practical approximation of the intended meaning of the sentence. A naive algorithm for computing weakest readings would explicitly compute the entailment order, by running a theorem prover on each pair of configurations, and then pick out the minimal elements. But this algorithm is quadratic in the number of configurations, and therefore impractically slow for real-life sentences. Here we develop a fast algorithm for this problem. The fundamental insight we exploit is that entailment among the configurations of a dominance graph can be approximated with rewriting rules (Baader and Nipkow, 1999). Consider the relation between (d) and (c). We can explain that (d) entails (c) by observing that (c) can be built from (d) by exchanging the positions of the adjacent quantifiers ∀x and ∃y; more precisely, by applying the following rewrite rule: [−] ∀x(Q,∃y(P,R)) →∃y(P,∀x(Q,R)) (1) The body of the rule specifies that an occurrence of ∀x which is the direct parent of an occurrence of ∃y may change positions with it; the subformulas P, Q, and R must be copied appropriately. The annotation [−] specifies that we must only apply the rule to subformulas in negative logical polarity: If the quantifiers in (d) were not in the scope of a negation, then applying the rule would actually make the formula stronger. We say that the rule (1) is logically sound because applying it to a subformula with the correct polarity of some configuration t always makes the result t′ logically weaker than t. We formalize these rewrite systems as follows. We assume a finite annotation alphabet Ann with a special starting annotation a0 ∈Ann; in the example, we had Ann = {+,−} and a0 = +. We also assume an annotator function ann : Ann×Σ×N → Ann. The function ann can be used to traverse a tree top-down and compute the annotation of each node from the annotation of its parent: Its first argument is the annotation and its second argument the node label of the parent, and the third argument is the position of the child among the parent’s children. In our example, the annotator ann models logical polarity by mapping, for instance, ann(+,∃z,1) = ann(+,∃z,2) = ann(+,∃y,2) = +, ann(−,∃z,1) = ann(−,∃z,2) = ann(+,∀x,1) = −, etc. We have labelled each node of the configurations in Fig. 1 with the annotations that are computed in this way. Now we can define an annotated rewrite system R to be a finite set of pairs (a,r) where a is an annotation and r is an ordinary rewrite rule. The rule (1) above is an example of an annotated rewrite rule with a = −. A rewrite rule (a,r) can be applied at the node u of a tree t if ann assigns the annotation a to u and r is applicable at u as usual. The rule then rewrites t as described above. In other words, annotated rewrite systems are rewrite systems where rule applications are restricted to subtrees with specific annotations. We write t →R t′ if some rule of R can be applied at a node of t, and the result of rewriting is t′. The rewrite system R is called linear 33 if every variable that occurs on the left-hand side of a rule occurs on its right-hand side exactly once. 4.2 Relative normal forms The rewrite steps of a sound weakening rewrite system are related to the entailment order: Because every rewrite step transforms a reading into a weaker reading, an actual weakest readings must be such that there is no other configuration into which it can be rewritten. The converse is not always true, i.e. there can be non-rewritable configurations that are not weakest readings, but we will see in Section 6 that this approximation is good enough for practical use. So one way to solve the problem of computing weakest readings is to find readings that cannot be rewritten further. One class of configurations that “cannot be rewritten” with a rewrite system R is the set of normal forms of R, i.e. those configurations to which no rule in R can be applied. In our example, (b) and (c) are indeed normal forms with respect to a rewrite system that consists only of the rule (1). However, this is not exactly what we need here. Consider a rewrite system that also contains the following annotated rewrite rule, which is also sound for logical entailment: [+] ¬(∃z(P,Q)) →∃z(P,¬(Q)), (2) This rule would allow us to rewrite the configuration (c) into the tree ∃z(compz,¬(∃y(sampley,∀x(repr−ofx,z,seex,y)))). But this is no longer a configuration of the graph. If we were to equate weakest readings with normal forms, we would erroneously classify (c) as not being a weakest reading. The correct concept for characterizing weakest readings in terms of rewriting is that of a relative normal form. We define a configuration t of a dominance graph D to be a R-relative normal form of (the configurations of) D iff there is no other configuration t′ of D such that t →R t′. These are the configurations that can’t be weakened further without obtaining a tree that is no longer a configuration of D. In other words, if R approximates entailment, then the R-relative normal forms approximate the weakest readings. 4.3 Computing relative normal forms We now show how the relative normal forms of a dominance graph can be computed efficiently. For lack of space, we only sketch the construction and omit all proofs. Details can be found in Koller and Thater (2010). The key idea of the construction is to represent the relation →R in terms of a context tree transducer M, and characterize the relative normal forms of a tree language L in terms of the pre-image of L under M. Like ordinary regular tree transducers (Comon et al., 2007), context tree transducers read an input tree, assigning states to the nodes, while emitting an output tree. But while ordinary transducers read the input tree symbol by symbol, a context tree transducer can read multiple symbols at once. In this way, they are equivalent to the extended left-hand side transducers of Graehl et al. (2008). We will now define context tree transducers. Let Σ be a ranked signature, and let Xm be a set of m variables. We write Con(m)(Σ) for the contexts with m holes, i.e. those trees in T(Σ∪Xm) in which each element of Xm occurs exactly once, and always as a leaf. If C ∈Con(m)(Σ), then C[t1,...,tm] = C[t1/x1,...,tm/xm], where x1,...,xm are the variables from left to right. A (top-down) context tree transducer from Σ to ∆ is a 5-tuple M = (Q,Σ,∆,q0,δ). Σ and ∆are ranked signatures, Q is a finite set of states, and q0 ∈Q is the start state. δ is a finite set of transition rules of the form q(C[x1,...,xn]) →D[q1(xi1),...,qm(xim)], where C ∈Con(n)(Σ) and D ∈Con(m)(∆). If t ∈T(Σ∪∆∪Q), then we say that M derives t′ in one step from t, t →M t′, if t is of the form C′[q(C[t1,...,tn])] for some C′ ∈Con(1)(Σ), t′ is of the form C′[D[q1(ti1),...,qm(tim)]], and there is a rule q(C[x1,...,xn]) →D[q1(xi1),...,qm(xim)] in δ. The derivation relation →∗ M is the reflexive, transitive closure of →M. The translation relation τM of M is τM = {(t,t′) |t ∈T(Σ) andt′ ∈T(∆) and q0(t) →∗t′}. For each linear annotated rewrite system R, we can now build a context tree transducer MR such that t →R t′ iff (t,t′) ∈τMR. The idea is that MR traverses t from the root to the leaves, keeping track of the current annotation in its state. MR can nondeterministically choose to either copy the current symbol to the output tree unchanged, or to apply a rewrite rule from R. The rules are built in such a way that in each run, exactly one rewrite rule must be applied. We achieve this as follows. MR takes as its states the set { ¯q}∪{qa | a ∈Ann} and as its start state the state qa0. If MR reads a node u in state qa, this means that the annotator assigns annotation a to u and MR will rewrite a subtree at or 34 below u. If MR reads u in state ¯q, this means that MR will copy the subtree below u unchanged because the rewriting has taken place elsewhere. Thus MR has three types of rewrite rules. First, for any f ∈Σ, we have a rule ¯q(f(x1,...,xn)) → f( ¯q(x1),..., ¯q(xn)). Second, for any f and 1 ≤i ≤n, we have a rule qa(f(x1,...,xn)) → f( ¯q(x1),...,qann(a,f,i)(xi),..., ¯q(xn)), which nondeterministically chooses under which child the rewriting should take place, and assigns it the correct annotation. Finally, we have a rule qa(C[x1,...,xn]) →C′[ ¯q(xi1),..., ¯q(xin)] for every rewrite rule C[x1,...,xn] →C′[xi1,...,xin] with annotation a in R. Now let’s put the different parts together. We know that for each hnc dominance graph D, there is a regular tree grammar GD such that L(GD) is the set of configurations of D. Furthermore, the preimage τ−1 M (L) = {t | exists t′ ∈L with (t,t′) ∈τM} of a regular tree language L is also regular (Koller and Thater, 2010) if M is linear, and regular tree languages are closed under intersection and complement (Comon et al., 2007). So we can compute another RTG G′ such that L(G′) = L(GD)∩τ−1 MR(L(GD)). L(G′) consists of the members of L(GD) which cannot be rewritten by MR into members of L(GD); that is, L(G′) is exactly the set of R-relative normal forms of D. In general, the complement construction requires exponential time in the size of MR and GD. However, it can be shown that if the rules in R have at most depth two and GD is deterministic, then the entire above construction can be computed in time O(|GD|·|R|) (Koller and Thater, 2010). In other words, we have shown how to compute the weakest readings of a hypernormally connected dominance graph D, as approximated by a weakening rewrite system R, in time linear in the size of GD and linear in the size of R. This is a dramatic improvement over the best previous algorithm, which was quadratic in |conf(D)|. 4.4 An example Consider an annotated rewrite system that contains rule (1) plus the following rewrite rule: [−] ∃z(P,∀x(Q,R)) →∀x(∃z(P,Q),R) (3) This rewrite system translates into a top-down context tree transducer MR with the following transition rules, omitting most rules of the first two {1,2,3,4,5,6,7,8}F →¬({2,3,4,5,6,7,8}F) {2,3,4,5,6,7,8}F →∃y({7}{ ¯q},{2,4,5,6,8}F) | ∃z({5}{ ¯q},{2,3,6,7,8}F) {2,3,6,7,8}F →∃y({7}{ ¯q},∀x({6}{ ¯q},{8}{ ¯q})) {2,4,5,6,8}F →∀x({4,5,6}{ ¯q},{8}{ ¯q}) {4,5,6}{ ¯q} →∃z({5}{ ¯q},{6}{ ¯q}) {5}{ ¯q} →compz {6}{ ¯q} →repr-ofx,z {7}{ ¯q} →sampley {8}{ ¯q} →seex,y Figure 4: RTG for the weakest readings of Fig. 1. types for lack of space. q−(∀x(x1,∃y(x2,x3))) →∃y( ¯q(x2),∀x( ¯q(x1), ¯q(x3))) q−(∃y(x1,∀x(x2,x3))) →∀x(∃y( ¯q(x1), ¯q(x2)), ¯q(x3)) ¯q(¬(x1)) →¬( ¯q(x1)) q+(¬(x1)) →¬(q−(x1)) ¯q(∀x(x1,x2)) →∀x( ¯q(x1), ¯q(x2)) q+(∀x(x1,x2)) →∀x( ¯q(x1),q+(x2)) q+(∀x(x1,x2)) →∀x(q−(x1), ¯q(x2)) ... The grammar G′ for the relative normal forms is shown in Fig. 4 (omitting rules that involve unproductive nonterminals). We obtain it by starting with the example grammar GD in Fig. 3; then computing a deterministic RTG GR for τ−1 MR(L(GD)); and then intersecting the complement of GR with GD. The nonterminals of G′ are subgraphs of D, marked either with a set of states of MR or the symbol F, indicating that GR had no production rule for a given left-hand side. The start symbol of G′ is marked with F because G′ should only generate trees that GR cannot generate. As expected, G′ generates precisely two trees, namely (b) and (c). 5 Redundancy elimination, revisited The construction we just carried out – characterize the configurations we find interesting as the relative normal forms of an annotated rewrite system R, translate it into a transducer MR, and intersect conf(D) with the complement of the pre-image under MR – is more generally useful than just for the computation of weakest readings. We illustrate this on the problem of redundancy elimination (Vestre, 1991; Chaves, 2003; Koller et al., 2008) by showing how a variant of the algorithm of Koller et al. (2008) falls out of our technique as a special case. Redundancy elimination is the problem of computing, from a dominance graph D, another dominance graph D′ such that conf(D′) ⊆conf(D) and 35 every formula in conf(D) is logically equivalent to some formula in conf(D′). We can approximate logical equivalence using a finite system of equations such as ∃y(P,∃z(Q,R)) = ∃z(Q,∃y(P,R)), (4) indicating that ∃y and ∃z can be permuted without changing the models of the formula. Following the approach of Section 4, we can solve the redundancy elimination problem by transforming the equation system into a rewrite system R such that t →R t′ implies that t and t′ are equivalent. To this end, we assume an arbitrary linear order < on Σ, and orient all equations into rewrite rules that respect this order. If we assume ∃y < ∃z, the example rule (4) translates into the annotated rewrite rules [a] ∃z(P,∃y(Q,R)) →∃y(Q,∃z(P,R)) (5) for all annotations a ∈Ann; logical equivalence is not sensitive to the annotation. Finally, we can compute the relative normal forms of conf(D) under this rewrite system as above. The result will be an RTG G′ describing a subset of conf(D). Every tree t in conf(D) that is not in L(G′) is equivalent to some tree t′ in L(G′), because if t could not be rewritten into such a t′, then t would be in relative normal form. That is, the algorithm solves the redundancy elimination problem. Furthermore, if the oriented rewrite system is confluent (Baader and Nipkow, 1999), no two trees in L(G′) will be equivalent to each other, i.e. we achieve complete reduction in the sense of Koller et al. (2008). This solution shares much with that of Koller et al. (2008), in that we perform redundancy elimination by intersecting tree grammars. However, the construction we present here is much more general: The algorithmic foundation for redundancy elimination is now exactly the same as that for weakest readings, we only have to use an equivalencepreserving rewrite system instead of a weakening one. This new formal clarity also simplifies the specification of certain equations, as we will see in Section 6. In addition, we can now combine the weakening rules (1), (3), and (5) into a single rewrite system, and then construct a tree grammar for the relative normal forms of the combined system. This algorithm performs redundancy elimination and computes weakest readings at the same time, and in our example retains only a single configuration, namely (5) (e) ¬∀x(∃z,∃y) (a) ¬∃y∃z∀x (3) (1) (1) (b) ¬∃y∀x∃z (c) ¬∃z∃y∀x (d) ¬∃z∀x∃y (3) Figure 5: Structure of the configuration set of Fig. 1 in terms of rewriting. (b); the configuration (c) is rejected because it can be rewritten to (a) with (5). The graph in Fig. 5 illustrates how the equivalence and weakening rules conspire to exclude all other configurations. 6 Evaluation In this section, we evaluate the effectiveness and efficiency of our weakest readings algorithm on a treebank. We compute RTGs for all sentences in the treebank and measure how many weakest readings remain after the intersection, and how much time this computation takes. Resources. For our experiment, we use the Rondane treebank (version of January 2006), a “Redwoods style” (Oepen et al., 2002) treebank containing underspecified representations (USRs) in the MRS formalism (Copestake et al., 2005) for sentences from the tourism domain. Our implementation of the relative normal forms algorithm is based on Utool (Koller and Thater, 2005), which (among other things) can translate a large class of MRS descriptions into hypernormally connected dominance graphs and further into RTGs as in Section 3. The implementation exploits certain properties of RTGs computed from dominance graphs to maximize efficiency. We will make this implementation publically available as part of the next Utool release. We use Utool to automatically translate the 999 MRS descriptions for which this is possible into RTGs. To simplify the specification of the rewrite systems, we restrict ourselves to the subcorpus in which all scope-taking operators (labels with arity > 0) occur at least ten times. This subset contains 624 dominance graphs. We refer to this subset as “RON10.” Signature and annotations. For each dominance graph D that we obtain by converting an MRS description, we take GD as a grammar over the signature Σ = { fu | u ∈WD, f = LD(u)}. That is, we distinguish possible different occurrences of the same symbol in D by marking each occur36 rence with the name of the node. This makes GD a deterministic grammar. We then specify an annotator over Σ that assigns polarities for the weakening rewrite system. We distinguish three polarities: + for positive occurrences, −for negative occurrences (as in predicate logic), and ⊥for contexts in which a weakening rule neither weakens or strengthens the entire formula. The starting annotation is +. Finally, we need to decide upon each scopetaking operator’s effects on these annotations. To this end, we build upon Barwise and Cooper’s (1981) classification of the monotonicity properties of determiners. A determiner is upward (downward) monotonic if making the denotation of the determiner’s argument bigger (smaller) makes the sentence logically weaker. For instance, every is downward monotonic in its first argument and upward monotonic in its second argument, i.e. every girl kissed a boy entails every blond girl kissed someone. Thus ann(everyu,a,1) = −a and ann(everyu,a,2) = a (where u is a node name as above). There are also determiners with nonmonotonic argument positions, which assign the annotation ⊥to this argument. Negation reverses positive and negative polarity, and all other nonquantifiers simply pass on their annotation to the arguments. Weakest readings. We use the following weakening rewrite system for our experiment, where i ∈{1,2}: 1. [+] (E/i,D/1), (D/2,D/1) 2. [+] (E/i,P/1), (D/2,P/1) 3. [+] (E/i,A/2), (D/1,A/2) 4. [+] (A/2,N/1) 5. [+] (N/1,E/i), (N/1,D/2) 6. [+] (E/i,M/1), (D/1,M/1) Here the symbols E, D, etc. stand for classes of labels in Σ, and a rule schema [a] (C/i,C′/k) is to be read as shorthand for a set of rewrite rules which rearrange a tree where the i-th child of a symbol from C is a symbol from C′ into a tree where the symbol from C becomes the k-th child of the symbol from C′. For example, because we have allu ∈A and notv ∈N, Schema 4 licenses the following annotated rewrite rule: [+] allu(P,notv(Q)) →notv(allu(P,Q)). We write E and D for existential and definite determiners. P stands for proper names and pronouns, A stands for universal determiners like all and each, N for the negation not, and M for modal operators like can or would. M also includes intensional verbs like have to and want. Notice that while the reverse rules are applicable in negative polarities, no rules are applicable in polarity ⊥. Rule schema 1 states, for instance, that the specific (wide-scope) reading of the indefinite in the president of a company is logically stronger than the reading in which a company is within the restriction of the definite determiner. The schema is intuitively plausible, and it can also be proved to be logically sound if we make the standard assumption that the definite determiner the means “exactly one” (Montague, 1974). A similar argument applies to rule schema 2. Rule schema 3 encodes the classical entailment (1). Schema 4 is similar to the rule (2). Notice that it is not, strictly speaking, logically sound; however, because strong determiners like all or every carry a presupposition that their restrictions have a non-empty denotation (Lasersohn, 1993), the schema becomes sound for all instances that can be expressed in natural language. Similar arguments apply to rule schemas 5 and 6, which are potentially unsound for subtle reasons involving the logical interpretation of intensional expressions. However, these cases of unsoundness did not occur in our test corpus. Redundancy elimination. In addition, we assume the following equation system for redundancy elimination for i, j ∈{1,2} and k ∈N (again written in an analogous shorthand as above): 7. E/i = E/ j 8. D/1 = E/i, E/i = D/1 9. D/1 = D/1 10. Σ/k = P/2 These rule schemata state that permuting existential determiners with each other is an equivalence transformation, and so is permuting definite determiners with existential and definite determiners if one determiner is the second argument (in the scope) of a definite. Schema 10 states that proper names and pronouns, which the ERG analyzes as scope-bearing operators, can permute with any other label. We orient these equalities into rewrite rules by ordering symbols in P before symbols that are not 37 All KRT08 RE RE+WR #conf = 1 8.5% 23.4% 34.9% 66.7% #conf ≤2 20.5% 40.9% 57.9% 80.6% avg(#conf) 3.2M 7603.1 119.0 4.5 med(#conf) 25 4 2 1 runtime 8.1s 9.4s 8.7s 9.1s Figure 6: Analysis of the numbers of configurations in RON10. in P, and otherwise ordering a symbol fu before a symbol gv if u < v by comparison of the (arbitrary) node names. Results. We used these rewrite systems to compute, for each USR in RON10, the number of all configurations, the number of configurations that remain after redundancy elimination, and the number of weakest readings (i.e., the relative normal forms of the combined equivalence and weakening rewrite systems). The results are summarized in Fig. 6. By computing weakest readings (WR), we reduce the ambiguity of over 80% of all sentences to one or two readings; this is a clear improvement even over the results of the redundancy elimination (RE). Computing weakest readings reduces the mean number of readings from several million to 4.5, and improves over the RE results by a factor of 30. Notice that the RE algorithm from Section 5 is itself an improvement over Koller et al.’s (2008) system (“KRT08” in the table), which could not process the rule schema 10. Finally, computing the weakest readings takes only a tiny amount of extra runtime compared to the RE elimination or even the computation of the RTGs (reported as the runtime for “All”).1 This remains true on the entire Rondane corpus (although the reduction factor is lower because we have no rules for the rare scope-bearers): RE+WR computation takes 32 seconds, compared to 30 seconds for RE. In other words, our algorithm brings the semantic ambiguity in the Rondane Treebank down to practically useful levels at a mean runtime investment of a few milliseconds per sentence. It is interesting to note how the different rule schemas contribute to this reduction. While the instances of Schemata 1 and 2 are applicable in 340 sentences, the other schemas 3–6 together are only 1Runtimes were measured on an Intel Core 2 Duo CPU at 2.8 GHz, under MacOS X 10.5.6 and Apple Java 1.5.0_16, after allowing the JVM to just-in-time compile the bytecode. applicable in 44 sentences. Nevertheless, where these rules do apply, they have a noticeable effect: Without them, the mean number of configurations in RON10 after RE+WR increases to 12.5. 7 Conclusion In this paper, we have shown how to compute the weakest readings of a dominance graph, characterized by an annotated rewrite system. Evaluating our algorithm on a subcorpus of the Rondane Treebank, we reduced the mean number of configurations of a sentence from several million to 4.5, in negligible runtime. Our algorithm can be applied to other problems in which an underspecified representation is to be disambiguated, as long as the remaining readings can be characterized as the relative normal forms of a linear annotated rewrite system. We illustrated this for the case of redundancy elimination. The algorithm presented here makes it possible, for the first time, to derive a single meaningful semantic representation from the syntactic analysis of a deep grammar on a large scale. In the future, it will be interesting to explore how these semantic representations can be used in applications. For instance, it seems straightforward to adapt MacCartney and Manning’s (2008) “natural logic”-based Textual Entailment system, because our annotator already computes the polarities needed for their monotonicity inferences. We could then perform such inferences on (cleaner) semantic representations, rather than strings (as they do). On the other hand, it may be possible to reduce the set of readings even further. We retain more readings than necessary in many treebank sentences because the combined weakening and equivalence rewrite system is not confluent, and therefore may not recognize a logical relation between two configurations. The rewrite system could be made more powerful by running the Knuth-Bendix completion algorithm (Knuth and Bendix, 1970). Exploring the practical tradeoff between the further reduction in the number of remaining configurations and the increase in complexity of the rewrite system and the RTG would be worthwhile. Acknowledgments. We are indebted to Joachim Niehren, who pointed out a crucial simplification in the algorithm to us. We also thank our reviewers for their constructive comments. 38 References E. Althaus, D. Duchier, A. Koller, K. Mehlhorn, J. Niehren, and S. Thiel. 2003. An efficient graph algorithm for dominance constraints. Journal of Algorithms, 48:194–219. F. Baader and T. Nipkow. 1999. Term rewriting and all that. Cambridge University Press. J. Barwise and R. Cooper. 1981. Generalized quantifiers and natural language. Linguistics and Philosophy, 4:159–219. J. Bos. 2008. Let’s not argue about semantics. In Proceedings of the 6th international conference on Language Resources and Evaluation (LREC 2008). M. Butt, H. Dyvik, T. Holloway King, H. Masuichi, and C. Rohrer. 2002. The parallel grammar project. In Proceedings of COLING-2002 Workshop on Grammar Engineering and Evaluation. R. P. Chaves. 2003. Non-redundant scope disambiguation in underspecified semantics. In Proceedings of the 8th ESSLLI Student Session. H. Comon, M. Dauchet, R. Gilleron, C. Löding, F. Jacquemard, D. Lugiez, S. Tison, and M. Tommasi. 2007. Tree automata techniques and applications. Available on: http://www.grappa. univ-lille3.fr/tata. A. Copestake and D. Flickinger. 2000. An opensource grammar development environment and broad-coverage english grammar using HPSG. In Proceedings of the 2nd International Conference on Language Resources and Evaluation (LREC). A. Copestake, D. Flickinger, C. Pollard, and I. Sag. 2005. Minimal recursion semantics: An introduction. Journal of Language and Computation. D. Flickinger, A. Koller, and S. Thater. 2005. A new well-formedness criterion for semantics debugging. In Proceedings of the 12th International Conference on HPSG, Lisbon. M. Gabsdil and K. Striegnitz. 1999. Classifying scope ambiguities. In Proceedings of the First Intl. Workshop on Inference in Computational Semantics. J. Graehl, K. Knight, and J. May. 2008. Training tree transducers. Computational Linguistics, 34(3):391– 427. D. Higgins and J. Sadock. 2003. A machine learning approach to modeling scope preferences. Computational Linguistics, 29(1). J. Hobbs. 1983. An improper treatment of quantification in ordinary English. In Proceedings of the 21st Annual Meeting of the Association for Computational Linguistics (ACL’83). R. Kempson and A. Cormack. 1981. Ambiguity and quantification. Linguistics and Philosophy, 4:259– 309. D. Knuth and P. Bendix. 1970. Simple word problems in universal algebras. In J. Leech, editor, Computational Problems in Abstract Algebra, pages 263–297. Pergamon Press, Oxford. A. Koller and J. Niehren. 2000. On underspecified processing of dynamic semantics. In Proceedings of the 18th International Conference on Computational Linguistics (COLING-2000). A. Koller and S. Thater. 2005. Efficient solving and exploration of scope ambiguities. In ACL-05 Demonstration Notes, Ann Arbor. A. Koller and S. Thater. 2010. Computing relative normal forms in regular tree languages. In Proceedings of the 21st International Conference on Rewriting Techniques and Applications (RTA). A. Koller, J. Niehren, and S. Thater. 2003. Bridging the gap between underspecification formalisms: Hole semantics as dominance constraints. In Proceedings of the 10th EACL. A. Koller, M. Regneri, and S. Thater. 2008. Regular tree grammars as a formalism for scope underspecification. In Proceedings of ACL-08: HLT. P. Lasersohn. 1993. Existence presuppositions and background knowledge. Journal of Semantics, 10:113–122. B. MacCartney and C. Manning. 2008. Modeling semantic containment and exclusion in natural language inference. In Proceedings of the 22nd International Conference on Computational Linguistics (COLING). R. Montague. 1974. The proper treatment of quantification in ordinary English. In R. Thomason, editor, Formal Philosophy. Selected Papers of Richard Montague. Yale University Press, New Haven. C. Monz and M. de Rijke. 2001. Deductions with meaning. In Michael Moortgat, editor, Logical Aspects of Computational Linguistics, Third International Conference (LACL’98), volume 2014 of LNAI. Springer-Verlag, Berlin/Heidelberg. S. Oepen, K. Toutanova, S. Shieber, C. Manning, D. Flickinger, and T. Brants. 2002. The LinGO Redwoods treebank: Motivation and preliminary applications. In Proceedings of the 19th International Conference on Computational Linguistics (COLING). Uwe Reyle. 1995. On reasoning with ambiguities. In Proceedings of the 7th Conference of the European Chapter of the Association for Computational Linguistics (EACL’95). K. van Deemter. 1996. Towards a logic of ambiguous expressions. In Semantic Ambiguity and Underspecification. CSLI Publications, Stanford. E. Vestre. 1991. An algorithm for generating nonredundant quantifier scopings. In Proc. of EACL, Berlin. 39
|
2010
|
4
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 384–394, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Word representations: A simple and general method for semi-supervised learning Joseph Turian D´epartement d’Informatique et Recherche Op´erationnelle (DIRO) Universit´e de Montr´eal Montr´eal, Qu´ebec, Canada, H3T 1J4 [email protected] Lev Ratinov Department of Computer Science University of Illinois at Urbana-Champaign Urbana, IL 61801 [email protected] Yoshua Bengio D´epartement d’Informatique et Recherche Op´erationnelle (DIRO) Universit´e de Montr´eal Montr´eal, Qu´ebec, Canada, H3T 1J4 [email protected] Abstract If we take an existing supervised NLP system, a simple and general way to improve accuracy is to use unsupervised word representations as extra word features. We evaluate Brown clusters, Collobert and Weston (2008) embeddings, and HLBL (Mnih & Hinton, 2009) embeddings of words on both NER and chunking. We use near state-of-the-art supervised baselines, and find that each of the three word representations improves the accuracy of these baselines. We find further improvements by combining different word representations. You can download our word features, for off-the-shelf use in existing NLP systems, as well as our code, here: http://metaoptimize. com/projects/wordreprs/ 1 Introduction By using unlabelled data to reduce data sparsity in the labeled training data, semi-supervised approaches improve generalization accuracy. Semi-supervised models such as Ando and Zhang (2005), Suzuki and Isozaki (2008), and Suzuki et al. (2009) achieve state-of-the-art accuracy. However, these approaches dictate a particular choice of model and training regime. It can be tricky and time-consuming to adapt an existing supervised NLP system to use these semi-supervised techniques. It is preferable to use a simple and general method to adapt existing supervised NLP systems to be semi-supervised. One approach that is becoming popular is to use unsupervised methods to induce word features—or to download word features that have already been induced—plug these word features into an existing system, and observe a significant increase in accuracy. But which word features are good for what tasks? Should we prefer certain word features? Can we combine them? A word representation is a mathematical object associated with each word, often a vector. Each dimension’s value corresponds to a feature and might even have a semantic or grammatical interpretation, so we call it a word feature. Conventionally, supervised lexicalized NLP approaches take a word and convert it to a symbolic ID, which is then transformed into a feature vector using a one-hot representation: The feature vector has the same length as the size of the vocabulary, and only one dimension is on. However, the one-hot representation of a word suffers from data sparsity: Namely, for words that are rare in the labeled training data, their corresponding model parameters will be poorly estimated. Moreover, at test time, the model cannot handle words that do not appear in the labeled training data. These limitations of one-hot word representations have prompted researchers to investigate unsupervised methods for inducing word representations over large unlabeled corpora. Word features can be hand-designed, but our goal is to learn them. One common approach to inducing unsupervised word representation is to use clustering, perhaps hierarchical. This technique was used by a variety of researchers (Miller et al., 2004; Liang, 2005; Koo et al., 2008; Ratinov & Roth, 2009; Huang & Yates, 2009). This leads to a one-hot representation over a smaller vocabulary size. Neural language models (Bengio et al., 2001; Schwenk & Gauvain, 2002; Mnih & Hinton, 2007; Collobert & Weston, 2008), on the other hand, induce dense real-valued low-dimensional 384 word embeddings using unsupervised approaches. (See Bengio (2008) for a more complete list of references on neural language models.) Unsupervised word representations have been used in previous NLP work, and have demonstrated improvements in generalization accuracy on a variety of tasks. But different word representations have never been systematically compared in a controlled way. In this work, we compare different techniques for inducing word representations, evaluating them on the tasks of named entity recognition (NER) and chunking. We retract former negative results published in Turian et al. (2009) about Collobert and Weston (2008) embeddings, given training improvements that we describe in Section 7.1. 2 Distributional representations Distributional word representations are based upon a cooccurrence matrix F of size W×C, where W is the vocabulary size, each row Fw is the initial representation of word w, and each column Fc is some context. Sahlgren (2006) and Turney and Pantel (2010) describe a handful of possible design decisions in contructing F, including choice of context types (left window? right window? size of window?) and type of frequency count (raw? binary? tf-idf?). Fw has dimensionality W, which can be too large to use Fw as features for word w in a supervised model. One can map F to matrix f of size W × d, where d ≪C, using some function g, where f = g(F). fw represents word w as a vector with d dimensions. The choice of g is another design decision, although perhaps not as important as the statistics used to initially construct F. The self-organizing semantic map (Ritter & Kohonen, 1989) is a distributional technique that maps words to two dimensions, such that syntactically and semantically related words are nearby (Honkela et al., 1995; Honkela, 1997). LSA (Dumais et al., 1988; Landauer et al., 1998), LSI, and LDA (Blei et al., 2003) induce distributional representations over F in which each column is a document context. In most of the other approaches discussed, the columns represent word contexts. In LSA, g computes the SVD of F. Hyperspace Analogue to Language (HAL) is another early distributional approach (Lund et al., 1995; Lund & Burgess, 1996) to inducing word representations. They compute F over a corpus of 160 million word tokens with a vocabulary size W of 70K word types. There are 2·W types of context (columns): The first or second W are counted if the word c occurs within a window of 10 to the left or right of the word w, respectively. f is chosen by taking the 200 columns (out of 140K in F) with the highest variances. ICA is another technique to transform F into f. (V¨ayrynen & Honkela, 2004; V¨ayrynen & Honkela, 2005; V¨ayrynen et al., 2007). ICA is expensive, and the largest vocabulary size used in these works was only 10K. As far as we know, ICA methods have not been used when the size of the vocab W is 100K or more. Explicitly storing cooccurrence matrix F can be memory-intensive, and transforming F to f can be time-consuming. It is preferable that F never be computed explicitly, and that f be constructed incrementally. ˇReh˚uˇrek and Sojka (2010) describe an incremental approach to inducing LSA and LDA topic models over 270 millions word tokens with a vocabulary of 315K word types. This is similar in magnitude to our experiments. Another incremental approach to constructing f is using a random projection: Linear mapping g is multiplying F by a random matrix chosen a priori. This random indexing method is motivated by the Johnson-Lindenstrauss lemma, which states that for certain choices of random matrix, if d is sufficiently large, then the original distances between words in F will be preserved in f (Sahlgren, 2005). Kaski (1998) uses this technique to produce 100-dimensional representations of documents. Sahlgren (2001) was the first author to use random indexing using narrow context. Sahlgren (2006) does a battery of experiments exploring different design decisions involved in constructing F, prior to using random indexing. However, like all the works cited above, Sahlgren (2006) only uses distributional representation to improve existing systems for one-shot classification tasks, such as IR, WSD, semantic knowledge tests, and text categorization. It is not well-understood what settings are appropriate to induce distributional word representations for structured prediction tasks (like parsing and MT) and sequence labeling tasks (like chunking and NER). Previous research has achieved repeated successes on these tasks using clustering representations (Section 3) and distributed representations (Section 4), so we focus on these representations in our work. 3 Clustering-based word representations Another type of word representation is to induce a clustering over words. Clustering methods and 385 distributional methods can overlap. For example, Pereira et al. (1993) begin with a cooccurrence matrix and transform this matrix into a clustering. 3.1 Brown clustering The Brown algorithm is a hierarchical clustering algorithm which clusters words to maximize the mutual information of bigrams (Brown et al., 1992). So it is a class-based bigram language model. It runs in time O(V·K2), where V is the size of the vocabulary and K is the number of clusters. The hierarchical nature of the clustering means that we can choose the word class at several levels in the hierarchy, which can compensate for poor clusters of a small number of words. One downside of Brown clustering is that it is based solely on bigram statistics, and does not consider word usage in a wider context. Brown clusters have been used successfully in a variety of NLP applications: NER (Miller et al., 2004; Liang, 2005; Ratinov & Roth, 2009), PCFG parsing (Candito & Crabb´e, 2009), dependency parsing (Koo et al., 2008; Suzuki et al., 2009), and semantic dependency parsing (Zhao et al., 2009). Martin et al. (1998) presents algorithms for inducing hierarchical clusterings based upon word bigram and trigram statistics. Ushioda (1996) presents an extension to the Brown clustering algorithm, and learn hierarchical clusterings of words as well as phrases, which they apply to POS tagging. 3.2 Other work on cluster-based word representations Lin and Wu (2009) present a K-means-like non-hierarchical clustering algorithm for phrases, which uses MapReduce. HMMs can be used to induce a soft clustering, specifically a multinomial distribution over possible clusters (hidden states). Li and McCallum (2005) use an HMM-LDA model to improve POS tagging and Chinese Word Segmentation. Huang and Yates (2009) induce a fully-connected HMM, which emits a multinomial distribution over possible vocabulary words. They perform hard clustering using the Viterbi algorithm. (Alternately, they could keep the soft clustering, with the representation for a particular word token being the posterior probability distribution over the states.) However, the CRF chunker in Huang and Yates (2009), which uses their HMM word clusters as extra features, achieves F1 lower than a baseline CRF chunker (Sha & Pereira, 2003). Goldberg et al. (2009) use an HMM to assign POS tags to words, which in turns improves the accuracy of the PCFG-based Hebrew parser. Deschacht and Moens (2009) use a latent-variable language model to improve semantic role labeling. 4 Distributed representations Another approach to word representation is to learn a distributed representation. (Not to be confused with distributional representations.) A distributed representation is dense, lowdimensional, and real-valued. Distributed word representations are called word embeddings. Each dimension of the embedding represents a latent feature of the word, hopefully capturing useful syntactic and semantic properties. A distributed representation is compact, in the sense that it can represent an exponential number of clusters in the number of dimensions. Word embeddings are typically induced using neural language models, which use neural networks as the underlying predictive model (Bengio, 2008). Historically, training and testing of neural language models has been slow, scaling as the size of the vocabulary for each model computation (Bengio et al., 2001; Bengio et al., 2003). However, many approaches have been proposed in recent years to eliminate that linear dependency on vocabulary size (Morin & Bengio, 2005; Collobert & Weston, 2008; Mnih & Hinton, 2009) and allow scaling to very large training corpora. 4.1 Collobert and Weston (2008) embeddings Collobert and Weston (2008) presented a neural language model that could be trained over billions of words, because the gradient of the loss was computed stochastically over a small sample of possible outputs, in a spirit similar to Bengio and S´en´ecal (2003). This neural model of Collobert and Weston (2008) was refined and presented in greater depth in Bengio et al. (2009). The model is discriminative and nonprobabilistic. For each training update, we read an n-gram x = (w1, . . . , wn) from the corpus. The model concatenates the learned embeddings of the n words, giving e(w1) ⊕. . . ⊕e(wn), where e is the lookup table and ⊕is concatenation. We also create a corrupted or noise n-gram ˜x = (w1, . . . , wn−q, ˜wn), where ˜wn , wn is chosen uniformly from the vocabulary.1 For convenience, 1In Collobert and Weston (2008), the middle word in the 386 we write e(x) to mean e(w1) ⊕. . . ⊕e(wn). We predict a score s(x) for x by passing e(x) through a single hidden layer neural network. The training criterion is that n-grams that are present in the training corpus like x must have a score at least some margin higher than corrupted n-grams like ˜x. Specifically: L(x) = max(0, 1−s(x)+ s(˜x)). We minimize this loss stochastically over the n-grams in the corpus, doing gradient descent simultaneously over the neural network parameters and the embedding lookup table. We implemented the approach of Collobert and Weston (2008), with the following differences: • We did not achieve as low log-ranks on the English Wikipedia as the authors reported in Bengio et al. (2009), despite initially attempting to have identical experimental conditions. • We corrupt the last word of each n-gram. • We had a separate learning rate for the embeddings and for the neural network weights. We found that the embeddings should have a learning rate generally 1000–32000 times higher than the neural network weights. Otherwise, the unsupervised training criterion drops slowly. • Although their sampling technique makes training fast, testing is still expensive when the size of the vocabulary is large. Instead of cross-validating using the log-rank over the validation data as they do, we instead used the moving average of the training loss on training examples before the weight update. 4.2 HLBL embeddings The log-bilinear model (Mnih & Hinton, 2007) is a probabilistic and linear neural model. Given an n-gram, the model concatenates the embeddings of the n −1 first words, and learns a linear model to predict the embedding of the last word. The similarity between the predicted embedding and the current actual embedding is transformed into a probability by exponentiating and then normalizing. Mnih and Hinton (2009) speed up model evaluation during training and testing by using a hierarchy to exponentially filter down the number of computations that are performed. This hierarchical evaluation technique was first proposed by Morin and Bengio (2005). The model, combined with this optimization, is called the hierarchical log-bilinear (HLBL) model. n-gram is corrupted. In Bengio et al. (2009), the last word in the n-gram is corrupted. 5 Supervised evaluation tasks We evaluate the hypothesis that one can take an existing, near state-of-the-art, supervised NLP system, and improve its accuracy by including word representations as word features. This technique for turning a supervised approach into a semi-supervised one is general and task-agnostic. However, we wish to find out if certain word representations are preferable for certain tasks. Lin and Wu (2009) finds that the representations that are good for NER are poor for search query classification, and vice-versa. We apply clustering and distributed representations to NER and chunking, which allows us to compare our semi-supervised models to those of Ando and Zhang (2005) and Suzuki and Isozaki (2008). 5.1 Chunking Chunking is a syntactic sequence labeling task. We follow the conditions in the CoNLL-2000 shared task (Sang & Buchholz, 2000). The linear CRF chunker of Sha and Pereira (2003) is a standard near-state-of-the-art baseline chunker. In fact, many off-the-shelf CRF implementations now replicate Sha and Pereira (2003), including their choice of feature set: • CRF++ by Taku Kudo (http://crfpp. sourceforge.net/) • crfsgd by L´eon Bottou (http://leon. bottou.org/projects/sgd) • CRFsuite by by Naoaki Okazaki (http:// www.chokkan.org/software/crfsuite/) We use CRFsuite because it makes it simple to modify the feature generation code, so one can easily add new features. We use SGD optimization, and enable negative state features and negative transition features. (“feature.possible transitions=1, feature.possible states=1”) Table 1 shows the features in the baseline chunker. As you can see, the Brown and embedding features are unigram features, and do not participate in conjunctions like the word features and tag features do. Koo et al. (2008) sees further accuracy improvements on dependency parsing when using word representations in compound features. The data comes from the Penn Treebank, and is newswire from the Wall Street Journal in 1989. Of the 8936 training sentences, we used 1000 randomly sampled sentences (23615 words) for development. We trained models on the 7936 387 • Word features: wi for i in {−2, −1, 0, +1, +2}, wi ∧wi+1 for i in {−1, 0}. • Tag features: wi for i in {−2, −1, 0, +1, +2}, ti ∧ti+1 for i in {−2, −1, 0, +1}. ti ∧ti+1 ∧ti+2 for i in {−2, −1, 0}. • Embedding features [if applicable]: ei[d] for i in {−2, −1, 0, +1, +2}, where d ranges over the dimensions of the embedding ei. • Brown features [if applicable]: substr(bi, 0, p) for i in {−2, −1, 0, +1, +2}, where substr takes the p-length prefix of the Brown cluster bi. Table 1: Features templates used in the CRF chunker. training partition sentences, and evaluated their F1 on the development set. After choosing hyperparameters to maximize the dev F1, we would retrain the model using these hyperparameters on the full 8936 sentence training set, and evaluate on test. One hyperparameter was l2-regularization sigma, which for most models was optimal at 2 or 3.2. The word embeddings also required a scaling hyperparameter, as described in Section 7.2. 5.2 Named entity recognition NER is typically treated as a sequence prediction problem. Following Ratinov and Roth (2009), we use the regularized averaged perceptron model. Ratinov and Roth (2009) describe different sequence encoding like BILOU and BIO, and show that the BILOU encoding outperforms BIO, and the greedy inference performs competitively to Viterbi while being significantly faster. Accordingly, we use greedy inference and BILOU text chunk representation. We use the publicly available implementation from Ratinov and Roth (2009) (see the end of this paper for the URL). In our baseline experiments, we remove gazetteers and non-local features (Krishnan & Manning, 2006). However, we also run experiments that include these features, to understand if the information they provide mostly overlaps with that of the word representations. After each epoch over the training set, we measured the accuracy of the model on the development set. Training was stopped after the accuracy on the development set did not improve for 10 epochs, generally about 50–80 epochs total. The epoch that performed best on the development set was chosen as the final model. We use the following baseline set of features from Zhang and Johnson (2003): • Previous two predictions yi−1 and yi−2 • Current word xi • xi word type information: all-capitalized, is-capitalized, all-digits, alphanumeric, etc. • Prefixes and suffixes of xi, if the word contains hyphens, then the tokens between the hyphens • Tokens in the window c = (xi−2, xi−1, xi, xi+1, xi+2) • Capitalization pattern in the window c • Conjunction of c and yi−1. Word representation features, if present, are used the same way as in Table 1. When using the lexical features, we normalize dates and numbers. For example, 1980 becomes *DDDD* and 212-325-4751 becomes *DDD**DDD*-*DDDD*. This allows a degree of abstraction to years, phone numbers, etc. This delexicalization is performed separately from using the word representation. That is, if we have induced an embedding for 12/3/2008 , we will use the embedding of 12/3/2008 , and *DD*/*D*/*DDDD* in the baseline features listed above. Unlike in our chunking experiments, after we chose the best model on the development set, we used that model on the test set too. (In chunking, after finding the best hyperparameters on the development set, we would combine the dev and training set and training a model over this combined set, and then evaluate on test.) The standard evaluation benchmark for NER is the CoNLL03 shared task dataset drawn from the Reuters newswire. The training set contains 204K words (14K sentences, 946 documents), the test set contains 46K words (3.5K sentences, 231 documents), and the development set contains 51K words (3.3K sentences, 216 documents). We also evaluated on an out-of-domain (OOD) dataset, the MUC7 formal run (59K words). MUC7 has a different annotation standard than the CoNLL03 data. It has several NE types that don’t appear in CoNLL03: money, dates, and numeric quantities. CoNLL03 has MISC, which is not present in MUC7. To evaluate on MUC7, we perform the following postprocessing steps prior to evaluation: 1. In the gold-standard MUC7 data, discard (label as ‘O’) all NEs with type NUMBER/MONEY/DATE. 2. In the predicted model output on MUC7 data, discard (label as ‘O’) all NEs with type MISC. 388 These postprocessing steps will adversely affect all NER models across-the-board, nonetheless allowing us to compare different models in a controlled manner. 6 Unlabled Data Unlabeled data is used for inducing the word representations. We used the RCV1 corpus, which contains one year of Reuters English newswire, from August 1996 to August 1997, about 63 millions words in 3.3 million sentences. We left case intact in the corpus. By comparison, Collobert and Weston (2008) downcases words and delexicalizes numbers. We use a preprocessing technique proposed by Liang, (2005, p. 51), which was later used by Koo et al. (2008): Remove all sentences that are less than 90% lowercase a–z. We assume that whitespace is not counted, although this is not specified in Liang’s thesis. We call this preprocessing step cleaning. In Turian et al. (2009), we found that all word representations performed better on the supervised task when they were induced on the clean unlabeled data, both embeddings and Brown clusters. This is the case even though the cleaning process was very aggressive, and discarded more than half of the sentences. According to the evidence and arguments presented in Bengio et al. (2009), the non-convex optimization process for Collobert and Weston (2008) embeddings might be adversely affected by noise and the statistical sparsity issues regarding rare words, especially at the beginning of training. For this reason, we hypothesize that learning representations over the most frequent words first and gradually increasing the vocabulary—a curriculum training strategy (Elman, 1993; Bengio et al., 2009; Spitkovsky et al., 2010)—would provide better results than cleaning. After cleaning, there are 37 million words (58% of the original) in 1.3 million sentences (41% of the original). The cleaned RCV1 corpus has 269K word types. This is the vocabulary size, i.e. how many word representations were induced. Note that cleaning is applied only to the unlabeled data, not to the labeled data used in the supervised tasks. RCV1 is a superset of the CoNLL03 corpus. For this reason, NER results that use RCV1 word representations are a form of transductive learning. 7 Experiments and Results 7.1 Details of inducing word representations The Brown clusters took roughly 3 days to induce, when we induced 1000 clusters, the baseline in prior work (Koo et al., 2008; Ratinov & Roth, 2009). We also induced 100, 320, and 3200 Brown clusters, for comparison. (Because Brown clustering scales quadratically in the number of clusters, inducing 10000 clusters would have been prohibitive.) Because Brown clusters are hierarchical, we can use cluster supersets as features. We used clusters at path depth 4, 6, 10, and 20 (Ratinov & Roth, 2009). These are the prefixes used in Table 1. The Collobert and Weston (2008) (C&W) embeddings were induced over the course of a few weeks, and trained for about 50 epochs. One of the difficulties in inducing these embeddings is that there is no stopping criterion defined, and that the quality of the embeddings can keep improving as training continues. Collobert (p.c.) simply leaves one computer training his embeddings indefinitely. We induced embeddings with 25, 50, 100, or 200 dimensions over 5-gram windows. In comparison to Turian et al. (2009), we use improved C&W embeddings in this work: • They were trained for 50 epochs, not just 20 epochs. • We initialized all embedding dimensions uniformly in the range [-0.01, +0.01], not [-1,+1]. For rare words, which are typically updated only 143 times per epoch2, and given that our embedding learning rate was typically 1e-6 or 1e-7, this means that rare word embeddings will be concentrated around zero, instead of spread out randomly. The HLBL embeddings were trained for 100 epochs (7 days).3 Unlike our Collobert and Weston (2008) embeddings, we did not extensively tune the learning rates for HLBL. We used a learning rate of 1e-3 for both model parameters and embedding parameters. We induced embeddings with 100 dimensions over 5-gram windows, and embeddings with 50 dimensions over 5-gram windows. Embeddings were induced over one pass 2A rare word will appear 5 (window size) times per epoch as a positive example, and 37M (training examples per epoch) / 269K (vocabulary size) = 138 times per epoch as a corruption example. 3The HLBL model updates require fewer matrix multiplies than Collobert and Weston (2008) model updates. Additionally, HLBL models were trained on a GPGPU, which is faster than conventional CPU arithmetic. 389 approach using a random tree, not two passes with an updated tree and embeddings re-estimation. 7.2 Scaling of Word Embeddings Like many NLP systems, the baseline system contains only binary features. The word embeddings, however, are real numbers that are not necessarily in a bounded range. If the range of the word embeddings is too large, they will exert more influence than the binary features. We generally found that embeddings had zero mean. We can scale the embeddings by a hyperparameter, to control their standard deviation. Assume that the embeddings are represented by a matrix E: E ←σ · E/stddev(E) (1) σ is a scaling constant that sets the new standard deviation after scaling the embeddings. (a) 93.6 93.8 94 94.2 94.4 94.6 94.8 0.001 0.01 0.1 1 Validation F1 Scaling factor σ C&W, 50-dim HLBL, 50-dim C&W, 200-dim C&W, 100-dim HLBL, 100-dim C&W, 25-dim baseline (b) 89 89.5 90 90.5 91 91.5 92 92.5 0.001 0.01 0.1 1 Validation F1 Scaling factor σ C&W, 200-dim C&W, 100-dim C&W, 25-dim C&W, 50-dim HLBL, 100-dim HLBL, 50-dim baseline Figure 1: Effect as we vary the scaling factor σ (Equation 1) on the validation set F1. We experiment with Collobert and Weston (2008) and HLBL embeddings of various dimensionality. (a) Chunking results. (b) NER results. Figure 1 shows the effect of scaling factor σ on both supervised tasks. We were surprised to find that on both tasks, across Collobert and Weston (2008) and HLBL embeddings of various dimensionality, that all curves had similar shapes and optima. This is one contributions of our work. In Turian et al. (2009), we were not able to prescribe a default value for scaling the embeddings. However, these curves demonstrate that a reasonable choice of scale factor is such that the embeddings have a standard deviation of 0.1. 7.3 Capacity of Word Representations (a) 94.1 94.2 94.3 94.4 94.5 94.6 94.7 100 320 1000 3200 25 50 100 200 Validation F1 # of Brown clusters # of embedding dimensions C&W HLBL Brown baseline (b) 90 90.5 91 91.5 92 92.5 100 320 1000 3200 25 50 100 200 Validation F1 # of Brown clusters # of embedding dimensions C&W Brown HLBL baseline Figure 2: Effect as we vary the capacity of the word representations on the validation set F1. (a) Chunking results. (b) NER results. There are capacity controls for the word representations: number of Brown clusters, and number of dimensions of the word embeddings. Figure 2 shows the effect on the validation F1 as we vary the capacity of the word representations. In general, it appears that more Brown clusters are better. We would like to induce 10000 Brown clusters, however this would take several months. In Turian et al. (2009), we hypothesized on the basis of solely the HLBL NER curve that higher-dimensional word embeddings would give higher accuracy. Figure 2 shows that this hypothesis is not true. For NER, the C&W curve is almost flat, and we were suprised to find the even 25-dimensional C&W word embeddings work so well. For chunking, 50-dimensional embeddings had the highest validation F1 for both C&W and HLBL. These curves indicates that the optimal capacity of the word embeddings is task-specific. 390 System Dev Test Baseline 94.16 93.79 HLBL, 50-dim 94.63 94.00 C&W, 50-dim 94.66 94.10 Brown, 3200 clusters 94.67 94.11 Brown+HLBL, 37M 94.62 94.13 C&W+HLBL, 37M 94.68 94.25 Brown+C&W+HLBL, 37M 94.72 94.15 Brown+C&W, 37M 94.76 94.35 Ando and Zhang (2005), 15M 94.39 Suzuki and Isozaki (2008), 15M 94.67 Suzuki and Isozaki (2008), 1B 95.15 Table 2: Final chunking F1 results. In the last section, we show how many unlabeled words were used. System Dev Test MUC7 Baseline 90.03 84.39 67.48 Baseline+Nonlocal 91.91 86.52 71.80 HLBL 100-dim 92.00 88.13 75.25 Gazetteers 92.09 87.36 77.76 C&W 50-dim 92.27 87.93 75.74 Brown, 1000 clusters 92.32 88.52 78.84 C&W 200-dim 92.46 87.96 75.51 C&W+HLBL 92.52 88.56 78.64 Brown+HLBL 92.56 88.93 77.85 Brown+C&W 92.79 89.31 80.13 HLBL+Gaz 92.91 89.35 79.29 C&W+Gaz 92.98 88.88 81.44 Brown+Gaz 93.25 89.41 82.71 Lin and Wu (2009), 3.4B 88.44 Ando and Zhang (2005), 27M 93.15 89.31 Suzuki and Isozaki (2008), 37M 93.66 89.36 Suzuki and Isozaki (2008), 1B 94.48 89.92 All (Brown+C&W+HLBL+Gaz), 37M 93.17 90.04 82.50 All+Nonlocal, 37M 93.95 90.36 84.15 Lin and Wu (2009), 700B 90.90 Table 3: Final NER F1 results, showing the cumulative effect of adding word representations, non-local features, and gazetteers to the baseline. To speed up training, in combined experiments (C&W plus another word representation), we used the 50-dimensional C&W embeddings, not the 200-dimensional ones. In the last section, we show how many unlabeled words were used. 7.4 Final results Table 2 shows the final chunking results and Table 3 shows the final NER F1 results. We compare to the state-of-the-art methods of Ando and Zhang (2005), Suzuki and Isozaki (2008), and—for NER—Lin and Wu (2009). Tables 2 and 3 show that accuracy can be increased further by combining the features from different types of word representations. But, if only one word representation is to be used, Brown clusters have the highest accuracy. Given the improvements to the C&W embeddings since Turian et al. (2009), C&W embeddings outperform the HLBL embeddings. On chunking, there is only a minute difference between Brown clusters and the embeddings. Com(a) 0 50 100 150 200 250 0 1 10 100 1K 10K 100K 1M # of per-token errors (test set) Frequency of word in unlabeled data C&W, 50-dim Brown, 3200 clusters (b) 0 50 100 150 200 250 0 1 10 100 1K 10K 100K 1M # of per-token errors (test set) Frequency of word in unlabeled data C&W, 50-dim Brown, 1000 clusters Figure 3: For word tokens that have different frequency in the unlabeled data, what is the total number of per-token errors incurred on the test set? (a) Chunking results. (b) NER results. bining representations leads to small increases in the test F1. In comparison to chunking, combining different word representations on NER seems gives larger improvements on the test F1. On NER, Brown clusters are superior to the word embeddings. Since much of the NER F1 is derived from decisions made over rare words, we suspected that Brown clustering has a superior representation for rare words. Brown makes a single hard clustering decision, whereas the embedding for a rare word is close to its initial value since it hasn’t received many training updates (see Footnote 2). Figure 3 shows the total number of per-token errors incurred on the test set, depending upon the frequency of the word token in the unlabeled data. For NER, Figure 3 (b) shows that most errors occur on rare words, and that Brown clusters do indeed incur fewer errors for rare words. This supports our hypothesis that, for rare words, Brown clustering produces better representations than word embeddings that haven’t received sufficient training updates. For chunking, Brown clusters and C&W embeddings incur almost identical numbers of errors, and errors are concentrated around the more common 391 words. We hypothesize that non-rare words have good representations, regardless of the choice of word representation technique. For tasks like chunking in which a syntactic decision relies upon looking at several token simultaneously, compound features that use the word representations might increase accuracy more (Koo et al., 2008). Using word representations in NER brought larger gains on the out-of-domain data than on the in-domain data. We were surprised by this result, because the OOD data was not even used during the unsupervised word representation induction, as was the in-domain data. We are curious to investigate this phenomenon further. Ando and Zhang (2005) present a semisupervised learning algorithm called alternating structure optimization (ASO). They find a lowdimensional projection of the input features that gives good linear classifiers over auxiliary tasks. These auxiliary tasks are sometimes specific to the supervised task, and sometimes general language modeling tasks like “predict the missing word”. Suzuki and Isozaki (2008) present a semisupervised extension of CRFs. (In Suzuki et al. (2009), they extend their semi-supervised approach to more general conditional models.) One of the advantages of the semi-supervised learning approach that we use is that it is simpler and more general than that of Ando and Zhang (2005) and Suzuki and Isozaki (2008). Their methods dictate a particular choice of model and training regime and could not, for instance, be used with an NLP system based upon an SVM classifier. Lin and Wu (2009) present a K-means-like non-hierarchical clustering algorithm for phrases, which uses MapReduce. Since they can scale to millions of phrases, and they train over 800B unlabeled words, they achieve state-of-the-art accuracy on NER using their phrase clusters. This suggests that extending word representations to phrase representations is worth further investigation. 8 Conclusions Word features can be learned in advance in an unsupervised, task-inspecific, and model-agnostic manner. These word features, once learned, are easily disseminated with other researchers, and easily integrated into existing supervised NLP systems. The disadvantage, however, is that accuracy might not be as high as a semi-supervised method that includes task-specific information and that jointly learns the supervised and unsupervised tasks (Ando & Zhang, 2005; Suzuki & Isozaki, 2008; Suzuki et al., 2009). Unsupervised word representations have been used in previous NLP work, and have demonstrated improvements in generalization accuracy on a variety of tasks. Ours is the first work to systematically compare different word representations in a controlled way. We found that Brown clusters and word embeddings both can improve the accuracy of a near-state-of-the-art supervised NLP system. We also found that combining different word representations can improve accuracy further. Error analysis indicates that Brown clustering induces better representations for rare words than C&W embeddings that have not received many training updates. Another contribution of our work is a default method for setting the scaling parameter for word embeddings. With this contribution, word embeddings can now be used off-the-shelf as word features, with no tuning. Future work should explore methods for inducing phrase representations, as well as techniques for increasing in accuracy by using word representations in compound features. Replicating our experiments You can visit http://metaoptimize.com/ projects/wordreprs/ to find: The word representations we induced, which you can download and use in your experiments; The code for inducing the word representations, which you can use to induce word representations on your own data; The NER and chunking system, with code for replicating our experiments. Acknowledgments Thank you to Magnus Sahlgren, Bob Carpenter, Percy Liang, Alexander Yates, and the anonymous reviewers for useful discussion. Thank you to Andriy Mnih for inducing his embeddings on RCV1 for us. Joseph Turian and Yoshua Bengio acknowledge the following agencies for research funding and computing support: NSERC, RQCHP, CIFAR. Lev Ratinov was supported by the Air Force Research Laboratory (AFRL) under prime contract no. FA8750-09-C-0181. Any opinions, findings, and conclusion or recommendations expressed in this material are those of the author and do not necessarily reflect the view of the Air Force Research Laboratory (AFRL). 392 References Ando, R., & Zhang, T. (2005). A highperformance semi-supervised learning method for text chunking. ACL. Bengio, Y. (2008). Neural net language models. Scholarpedia, 3, 3881. Bengio, Y., Ducharme, R., & Vincent, P. (2001). A neural probabilistic language model. NIPS. Bengio, Y., Ducharme, R., Vincent, P., & Jauvin, C. (2003). A neural probabilistic language model. Journal of Machine Learning Research, 3, 1137–1155. Bengio, Y., Louradour, J., Collobert, R., & Weston, J. (2009). Curriculum learning. ICML. Bengio, Y., & S´en´ecal, J.-S. (2003). Quick training of probabilistic neural nets by importance sampling. AISTATS. Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent dirichlet allocation. Journal of Machine Learning Research, 3, 993–1022. Brown, P. F., deSouza, P. V., Mercer, R. L., Pietra, V. J. D., & Lai, J. C. (1992). Class-based n-gram models of natural language. Computational Linguistics, 18, 467–479. Candito, M., & Crabb´e, B. (2009). Improving generative statistical parsing with semi-supervised word clustering. IWPT (pp. 138–141). Collobert, R., & Weston, J. (2008). A unified architecture for natural language processing: Deep neural networks with multitask learning. ICML. Deschacht, K., & Moens, M.-F. (2009). Semisupervised semantic role labeling using the Latent Words Language Model. EMNLP (pp. 21–29). Dumais, S. T., Furnas, G. W., Landauer, T. K., Deerwester, S., & Harshman, R. (1988). Using latent semantic analysis to improve access to textual information. SIGCHI Conference on Human Factors in Computing Systems (pp. 281–285). ACM. Elman, J. L. (1993). Learning and development in neural networks: The importance of starting small. Cognition, 48, 781–799. Goldberg, Y., Tsarfaty, R., Adler, M., & Elhadad, M. (2009). Enhancing unlexicalized parsing performance using a wide coverage lexicon, fuzzy tag-set mapping, and EM-HMM-based lexical probabilities. EACL. Honkela, T. (1997). Self-organizing maps of words for natural language processing applications. Proceedings of the International ICSC Symposium on Soft Computing. Honkela, T., Pulkki, V., & Kohonen, T. (1995). Contextual relations of words in grimm tales, analyzed by self-organizing map. ICANN. Huang, F., & Yates, A. (2009). Distributional representations for handling sparsity in supervised sequence labeling. ACL. Kaski, S. (1998). Dimensionality reduction by random mapping: Fast similarity computation for clustering. IJCNN (pp. 413–418). Koo, T., Carreras, X., & Collins, M. (2008). Simple semi-supervised dependency parsing. ACL (pp. 595–603). Krishnan, V., & Manning, C. D. (2006). An effective two-stage model for exploiting nonlocal dependencies in named entity recognition. COLING-ACL. Landauer, T. K., Foltz, P. W., & Laham, D. (1998). An introduction to latent semantic analysis. Discourse Processes, 259–284. Li, W., & McCallum, A. (2005). Semi-supervised sequence modeling with syntactic topic models. AAAI. Liang, P. (2005). Semi-supervised learning for natural language. Master’s thesis, Massachusetts Institute of Technology. Lin, D., & Wu, X. (2009). Phrase clustering for discriminative learning. ACL-IJCNLP (pp. 1030–1038). Lund, K., & Burgess, C. (1996). Producing highdimensional semantic spaces from lexical co-occurrence. Behavior Research Methods, Instrumentation, and Computers, 28, 203–208. Lund, K., Burgess, C., & Atchley, R. A. (1995). Semantic and associative priming in highdimensional semantic space. Cognitive Science Proceedings, LEA (pp. 660–665). Martin, S., Liermann, J., & Ney, H. (1998). Algorithms for bigram and trigram word clustering. Speech Communication, 24, 19–37. Miller, S., Guinness, J., & Zamanian, A. (2004). Name tagging with word clusters and discriminative training. HLT-NAACL (pp. 337–342). 393 Mnih, A., & Hinton, G. E. (2007). Three new graphical models for statistical language modelling. ICML. Mnih, A., & Hinton, G. E. (2009). A scalable hierarchical distributed language model. NIPS (pp. 1081–1088). Morin, F., & Bengio, Y. (2005). Hierarchical probabilistic neural network language model. AISTATS. Pereira, F., Tishby, N., & Lee, L. (1993). Distributional clustering of english words. ACL (pp. 183–190). Ratinov, L., & Roth, D. (2009). Design challenges and misconceptions in named entity recognition. CoNLL. Ritter, H., & Kohonen, T. (1989). Self-organizing semantic maps. Biological Cybernetics, 241–254. Sahlgren, M. (2001). Vector-based semantic analysis: Representing word meanings based on random labels. Proceedings of the Semantic Knowledge Acquisition and Categorisation Workshop, ESSLLI. Sahlgren, M. (2005). An introduction to random indexing. Methods and Applications of Semantic Indexing Workshop at the 7th International Conference on Terminology and Knowledge Engineering (TKE). Sahlgren, M. (2006). The word-space model: Using distributional analysis to represent syntagmatic and paradigmatic relations between words in high-dimensional vector spaces. Doctoral dissertation, Stockholm University. Sang, E. T., & Buchholz, S. (2000). Introduction to the CoNLL-2000 shared task: Chunking. CoNLL. Schwenk, H., & Gauvain, J.-L. (2002). Connectionist language modeling for large vocabulary continuous speech recognition. International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 765–768). Orlando, Florida. Sha, F., & Pereira, F. C. N. (2003). Shallow parsing with conditional random fields. HLT-NAACL. Spitkovsky, V., Alshawi, H., & Jurafsky, D. (2010). From baby steps to leapfrog: How “less is more” in unsupervised dependency parsing. NAACL-HLT. Suzuki, J., & Isozaki, H. (2008). Semi-supervised sequential labeling and segmentation using giga-word scale unlabeled data. ACL-08: HLT (pp. 665–673). Suzuki, J., Isozaki, H., Carreras, X., & Collins, M. (2009). An empirical study of semi-supervised structured conditional models for dependency parsing. EMNLP. Turian, J., Ratinov, L., Bengio, Y., & Roth, D. (2009). A preliminary evaluation of word representations for named-entity recognition. NIPS Workshop on Grammar Induction, Representation of Language and Language Learning. Turney, P. D., & Pantel, P. (2010). From frequency to meaning: Vector space models of semantics. Journal of Artificial Intelligence Research. Ushioda, A. (1996). Hierarchical clustering of words. COLING (pp. 1159–1162). V¨ayrynen, J., & Honkela, T. (2005). Comparison of independent component analysis and singular value decomposition in word context analysis. AKRR’05, International and Interdisciplinary Conference on Adaptive Knowledge Representation and Reasoning. V¨ayrynen, J. J., & Honkela, T. (2004). Word category maps based on emergent features created by ICA. Proceedings of the STeP’2004 Cognition + Cybernetics Symposium (pp. 173–185). Finnish Artificial Intelligence Society. V¨ayrynen, J. J., Honkela, T., & Lindqvist, L. (2007). Towards explicit semantic features using independent component analysis. Proceedings of the Workshop Semantic Content Acquisition and Representation (SCAR). Stockholm, Sweden: Swedish Institute of Computer Science. ˇReh˚uˇrek, R., & Sojka, P. (2010). Software framework for topic modelling with large corpora. LREC. Zhang, T., & Johnson, D. (2003). A robust risk minimization based named entity recognition system. CoNLL. Zhao, H., Chen, W., Kit, C., & Zhou, G. (2009). Multilingual dependency learning: a huge feature engineering method to semantic dependency parsing. CoNLL (pp. 55–60). 394
|
2010
|
40
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 395–403, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Identifying Text Polarity Using Random Walks Ahmed Hassan University of Michigan Ann Arbor Ann Arbor, Michigan, USA [email protected] Dragomir Radev University of Michigan Ann Arbor Ann Arbor, Michigan, USA [email protected] Abstract Automatically identifying the polarity of words is a very important task in Natural Language Processing. It has applications in text classification, text filtering, analysis of product review, analysis of responses to surveys, and mining online discussions. We propose a method for identifying the polarity of words. We apply a Markov random walk model to a large word relatedness graph, producing a polarity estimate for any given word. A key advantage of the model is its ability to accurately and quickly assign a polarity sign and magnitude to any word. The method could be used both in a semi-supervised setting where a training set of labeled words is used, and in an unsupervised setting where a handful of seeds is used to define the two polarity classes. The method is experimentally tested using a manually labeled set of positive and negative words. It outperforms the state of the art methods in the semi-supervised setting. The results in the unsupervised setting is comparable to the best reported values. However, the proposed method is faster and does not need a large corpus. 1 Introduction Identifying emotions and attitudes from unstructured text is a very important task in Natural Language Processing. This problem has a variety of possible applications. For example, there has been a great body of work for mining product reputation on the Web (Morinaga et al., 2002; Turney, 2002). Knowing the reputation of a product is very important for marketing and customer relation management (Morinaga et al., 2002). Manually handling reviews to identify reputation is a very costly, and time consuming process given the overwhelming amount of reviews on the Web. A list of words with positive/negative polarity is a very valuable resource for such an application. Another interesting application is mining online discussions. A threaded discussion is an electronic discussion in which software tools are used to help individuals post messages and respond to other messages. Threaded discussions include e-mails, e-mail lists, bulletin boards, newsgroups, or Internet forums. Threaded discussions act as a very important tool for communication and collaboration in the Web. An enormous number of discussion groups exists on the Web. Millions of users post content to these groups covering pretty much every possible topic. Tracking participant attitude towards different topics and towards other participants is a very interesting task. For example,Tong (2001) presented the concept of sentiment timelines. His system classifies discussion posts about movies as either positive or negative. This is used to produce a plot of the number of positive and negative sentiment messages over time. All those applications could benefit much from an automatic way of identifying semantic orientation of words. In this paper, we study the problem of automatically identifying semantic orientation of any word by analyzing its relations to other words. Automatically classifying words as either positive or negative enables us to automatically identify the polarity of larger pieces of text. This could be a very useful building block for mining surveys, product reviews and online discussions. We apply a Markov random walk model to a large semantic word graph, producing a polarity estimate for any given word. Previous work on identifying the semantic orientation of words has addressed the problem as both a semi-supervised (Takamura et al., 2005) and an unsupervised (Turney and Littman, 2003) learning problem. In the semisupervised setting, a training set of labeled words 395 is used to train the model. In the unsupervised setting, only a handful of seeds is used to define the two polarity classes. The proposed method could be used both in a semi-supervised and in an unsupervised setting. Empirical experiments on a labeled set of words show that the proposed method outperforms the state of the art methods in the semi-supervised setting. The results in the unsupervised setting are comparable to the best reported values. The proposed method has the advantages that it is faster and it does not need a large training corpus. The rest of the paper is structured as follows. In Section 2, we discuss related work. Section 3 presents our method for identifying word polarity. Section 4 describes our experimental setup. We conclude in Section 5. 2 Related Work Hatzivassiloglou and McKeown (1997) proposed a method for identifying word polarity of adjectives. They extract all conjunctions of adjectives from a given corpus and then they classify each conjunctive expression as either the same orientation such as “simple and well-received” or different orientation such as “simplistic but wellreceived”. The result is a graph that they cluster into two subsets of adjectives. They classify the cluster with the higher average frequency as positive. They created and labeled their own dataset for experiments. Their approach will probably works only with adjectives because there is nothing wrong with conjunctions of nouns or verbs with opposite polarities (e.g., “war and peace”, “rise and fall”, ..etc). Turney and Littman (2003) identify word polarity by looking at its statistical association with a set of positive/negative seed words. They use two statistical measures for estimating association: Pointwise Mutual Information (PMI) and Latent Semantic Analysis (LSA). To get co-occurrence statistics, they submit several queries to a search engine. Each query consists of the given word and one of the seed words. They use the search engine near operator to look for instances where the given word is physically close to the seed word in the returned document. They present their method as an unsupervised method where a very small amount of seed words are used to define semantic orientation rather than train the model. One of the limitations of their method is that it requires a large corpus of text to achieve good performance. They use several corpora, the size of the best performing dataset is roughly one hundred billion words (Turney and Littman, 2003). Takamura et al. (2005) proposed using spin models for extracting semantic orientation of words. They construct a network of words using gloss definitions, thesaurus, and co-occurrence statistics. They regard each word as an electron. Each electron has a spin and each spin has a direction taking one of two values: up or down. Two neighboring spins tend to have the same orientation from an energetic point of view. Their hypothesis is that as neighboring electrons tend to have the same spin direction, neighboring words tend to have similar polarity. They pose the problem as an optimization problem and use the mean field method to find the best solution. The analogy with electrons leads them to assume that each word should be either positive or negative. This assumption is not accurate because most of the words in the language do not have any semantic orientation. They report that their method could get misled by noise in the gloss definition and their computations sometimes get trapped in a local optimum because of its greedy optimization flavor. Kamps et al. (2004) construct a network based on WordNet synonyms and then use the shortest paths between any given word and the words ’good’ and ’bad’ to determine word polarity. They report that using shortest paths could be very noisy. For example. ’good’ and ’bad’ themselves are closely related in WordNet with a 5long sequence “good, sound, heavy, big, bad”. A given word w may be more connected to one set of words (e.g., positive words), yet have a shorter path connecting it to one word in the other set. Restricting seed words to only two words affects their accuracy. Adding more seed words could help but it will make their method extremely costly from the computation point of view. They evaluate their method only using adjectives. Hu and Liu (2004) use WordNet synonyms and antonyms to predict the polarity of words. For any word, whose polarity is unknown, they search WordNet and a list of seed labeled words to predict its polarity. They check if any of the synonyms of the given word has known polarity. If so, they label it with the label of its synonym. Otherwise, they check if any of the antonyms of the given word has known polarity. If so, they label it 396 with the opposite label of the antonym. They continue in a bootstrapping manner till they label all possible word. This method is quite similar to the shortest-path method proposed in (Kamps et al., 2004). There are some other methods that try to build lexicons of polarized words. Esuli and Sebastiani (2005; 2006) use a textual representation of words by collating all the glosses of the word as found in some dictionary. Then, a binary text classifier is trained using the textual representation and applied to new words. Kim and Hovy (2004) start with two lists of positive and negative seed words. WordNet is used to expand these lists. Synonyms of positive words and antonyms of negative words are considered positive, while synonyms of negative words and antonyms of positive words are considered negative. A similar method is presented in (Andreevskaia and Bergler, 2006) where WordNet synonyms, antonyms, and glosses are used to iteratively expand a list of seeds. The sentiment classes are treated as fuzzy categories where some words are very central to one category, while others may be interpreted differently. Kanayama and Nasukawa (2006) use syntactic features and context coherency, the tendency for same polarities to appear successively , to acquire polar atoms. Other related work is concerned with subjectivity analysis. Subjectivity analysis is the task of identifying text that present opinions as opposed to objective text that present factual information (Wiebe, 2000). Text could be either words, phrases, sentences, or any other chunks. There are two main categories of work on subjectivity analysis. In the first category, subjective words and phrases are identified without considering their context (Wiebe, 2000; Hatzivassiloglou and Wiebe, 2000; Banea et al., 2008). In the second category, the context of subjective text is used (Riloff and Wiebe, 2003; Yu and Hatzivassiloglou, 2003; Nasukawa and Yi, 2003; Popescu and Etzioni, 2005) Wiebe et al. (2001) lists a lot of applications of subjectivity analysis such as classifying emails and mining reviews. Subjectivity analysis is related to the proposed method because identifying the polarity of text is the natural next step that should follow identifying subjective text. 3 Word Polarity We use a Markov random walk model to identify polarity of words. Assume that we have a network of words, some of which are labeled as either positive or negative. In this network, two words are connecting if they are related. Different sources of information could be used to decide whether two words are related or not. For example, the synonyms of any word are semantically related to it. The intuition behind that connecting semantically related words is that those words tend to have similar polarity. Now imagine a random surfer walking along the network starting from an unlabeled word w. The random walk continues until the surfer hits a labeled word. If the word w is positive then the probability that the random walk hits a positive word is higher and if w is negative then the probability that the random walk hits a negative word is higher. Similarly, if the word w is positive then the average time it takes a random walk starting at w to hit a positive node is less than the average time it takes a random walk starting at w to hit a negative node. In the rest of this section, we will describe how we can construct a word relatedness graph in Section 3.1. The random walk model is described in Section 3.2. Hitting time is defined in Section‘3.3. Finally, an algorithm for computing a sign and magnitude for the polarity of any given word is described in Section 3.4. 3.1 Network Construction We construct a network where two nodes are linked if they are semantically related. Several sources of information could be used as indicators of the relatedness of words. One such important source is WordNet (Miller, 1995). WordNet is a large lexical database of English. Nouns, verbs, adjectives and adverbs are grouped into sets of cognitive synonyms (synsets), each expressing a distinct concept (Miller, 1995). Synsets are interlinked by means of conceptual-semantic and lexical relations. The simplest approach is to connect words that occur in the same WordNet synset. We can collect all words in WordNet, and add links between any two words that occurr in the same synset. The resulting graph is a graph G(W, E) where W is a set of word / part-of-speech pairs for all the words in WordNet. E is the set of edges connecting each pair of synonymous words. Nodes represent word/pos pairs rather than words because the part of speech tags are helpful in disambiguating the different senses for a given word. For example, 397 the word “fine” has two different meanings when used as an adjective and as a noun. Several other methods could be used to link words. For example, we can use other WordNet relations: hypernyms, similar to,...etc. Another source of links between words is co-occurrence statistics from corpus. Following the method presented in (Hatzivassiloglou and McKeown, 1997), we can connect words if they appear in a conjunctive form in the corpus. This method is only applicable to adjectives. If two adjectives are connected by “and” in conjunctive form, it is highly likely that they have the same semantic orientation. In all our experiments, we restricted the network to only WordNet relations. We study the effect of using co-occurrence statistics to connect words later at the end of our experiments. If more than one relation exists between any two words, the strength of the corresponding edge is adjusted accordingly. 3.2 Random Walk Model Imagine a random surfer walking along the word relatedness graph G. Starting from a word with unknown polarity i , it moves to a node j with probability Pij after the first step. The walk continues until the surfer hits a word with a known polarity. Seed words with known polarity act as an absorbing boundary for the random walk. If we repeat the number of random walks N times, the percentage of time at which the walk ends at a positive/negative word could be used as an indicator of its positive/negative polarity. The average time a random walk starting at w takes to hit the set of positive/negative nodes is also an indicator of its polarity. This view is closely related to the partially labeled classification with random walks approach in (Szummer and Jaakkola, 2002) and the semi-supervised learning using harmonic functions approach in (Zhu et al., 2003). Let W be the set of words in our lexicon. We construct a graph whose nodes V are all words in W The edges E correspond to relatedness between words We define transition probabilities Pt+1|t(j|i) from i to j by normalizing the weights of the edges out of node i, so: Pt+1|t(j|i) = Wij/ X k Wik (1) where k represents all nodes in the neighborhood of i. Pt2|t1(j|i) denotes the transition probability from node i at step t1 to node j at time step t2. We note that the weights Wij are symmetric and the transition probabilities Pt+1|t(j|i) are not necessarily symmetric because of the node out degree normalization. 3.3 First-Passage Time The mean first-passage (hitting) time h(i|k) is defined as the average number of steps a random walker, starting in state i ̸= k, will take to enter state k for the first time (Norris, 1997). Let G = (V, E) be a graph with a set of vertices V , and a set of edges E. Consider a subset of vertices S ⊂V , Consider a random walk on G starting at node i ̸∈S. Let Nt denote the position of the random surfer at time t. Let h(i|S) be the the average number of steps a random walker, starting in state i ̸∈S, will take to enter a state k ∈S for the first time. Let T S be the first-passage for any vertex in S. P(TS = t|N0 = i) = X j∈V pij × P(TS = t −1|N0 = j) (2) h(i|S) is the expectation of TS. Hence: h(i|S) = E(TS|N0 = i) = ∞ X t=1 t × P(TS = t|N0 = i) = ∞ X t=1 t X j∈V pijP(TS = t −1|N0 = j) = X j∈V ∞ X t=1 (t −1)pijP(TS = t −1|N0 = j) + X j∈V ∞ X t=1 pijP(TS = t −1|N0 = j) = X j∈V pij ∞ X t=1 tP(TS = t|N0 = j) + 1 = X j∈V pij × h(j|S) + 1 (3) Hence the first-passage (hitting) time can be formally defined as: h(i|S) = ( 0 i ∈S P j∈V pij × h(j|S) + 1 otherwise (4) 3.4 Word Polarity Calculation Based on the description of the random walk model and the first-passage (hitting) time above, 398 we now propose our word polarity identification algorithm. We begin by constructing a word relatedness graph and defining a random walk on that graph as described above. Let S+ and S−be two sets of vertices representing seed words that are already labeled as either positive or negative respectively. For any given word w, we compute the hitting time h(w|S+), and h(w|S−) for the two sets iteratively as described earlier. if h(w|S+) is greater than h(w|S−), the word is classified as negative, otherwise it is classified as positive. The ratio between the two hitting times could be used as an indication of how positive/negative the given word is. This is useful in case we need to provide a confidence measure for the prediction. This could be used to allow the model to abstain from classifying words with when the confidence level is low. Computing hitting time as described earlier may be time consuming especially if the graph is large. To overcome this problem, we propose a Monte Carlo based algorithm for estimating it. The algorithm is shown in Algorithm 1. Algorithm 1 Word Polarity using Random Walks Require: A word relatedness graph G 1: Given a word w in V 2: Define a random walk on the graph. the transition probability between any two nodes i, and j is defined as: Pt+1|t(j|i) = Wij/ P k Wik 3: Start k independent random walks from w with a maximum number of steps m 4: Stop when a positive word is reached 5: Let h∗(w|S+) be the estimated value for h(w|S+) 6: Repeat for negative words computing h∗(w|S−) 7: if h∗(w|S+) ≤h∗(w|S−) then 8: Classify w as positive 9: else 10: Classify w as negative 11: end if 4 Experiments We performed experiments on the General Inquirer lexicon (Stone et al., 1966). We used it as a gold standard data set for positive/negative words. The dataset contains 4206 words, 1915 of which are positive and 2291 are negative. Some of the ambiguous words were removed like (Turney, 2002; Takamura et al., 2005). We use WordNet (Miller, 1995) as a source of synonyms and hypernyms for the word relatedness graph. We used 10-fold cross validation for all tests. We evaluate our results in terms of accuracy. Statistical significance was tested using a 2-tailed paired t-test. All reported results are statistically significant at the 0.05 level. We perform experiments varying the parameters and the network. We also look at the performance of the proposed method for different parts of speech, and for different confidence levels We compare our method to the Semantic Orientation from PMI (SO-PMI) method described in (Turney, 2002), the Spin model (Spin) described in (Takamura et al., 2005), the shortest path (short-path) described in (Kamps et al., 2004), and the bootstrapping (bootstrap) method described in (Hu and Liu, 2004). 4.1 Comparisons with other methods This method could be used in a semi-supervised setting where a set of labeled words are used and the system learns from these labeled nodes and from other unlabeled nodes. Under this setting, we compare our method to the spin model described in (Takamura et al., 2005). Table 2 compares the performance using 10-fold cross validation. The table shows that the proposed method outperforms the spin model. The spin model approach uses word glosses, WordNet synonym, hypernym, and antonym relations, in addition to co-occurrence statistics extracted from corpus. The proposed method achieves better performance by only using WordNet synonym, hypernym and similar to relations. Adding co-occurrence statistics slightly improved performance, while using glosses did not help at all. We also compare our method to the SO-PMI method presented in (Turney, 2002). They describe this setting as unsupervised (Turney, 2002) because they only use 14 seeds as paradigm words that define the semantic orientation rather than train the model. After (Turney, 2002), we use our method to predict semantic orientation of words in the General Inquirer lexicon (Stone et al., 1966) using only 14 seed words. The network we used contains only WordNet relations. No glosses or co-occurrence statistics are used. The results comparing the SO-PMI method with different dataset sizes, the spin model, and the proposed method using only 14 seeds is shown in Table 2. We no399 Table 1: Accuracy for adjectives only for the spin model, the bootstrap method, and the random walk model. spin-model bootstrap short-path rand-walks 83.6 72.8 68.8 88.8 tice that the random walk method outperforms SOPMI when SO-PMI uses datasets of sizes 1 × 107 and 2 × 109 words. The performance of SO-PMI and the random walk methods are comparable when SO-PMI uses a very large dataset (1 × 1011 words). The performance of the spin model approach is also comparable to the other 2 methods. The advantages of the random walk method over SO-PMI is that it is faster and it does not need a very large corpus like the one used by SOPMI. Another advantage is that the random walk method can be used along with the labeled data from the General Inquirer lexicon (Stone et al., 1966) to get much better performance. This is costly for the SO-PMI method because that will require the submission of almost 4000 queries to a commercial search engine. We also compare our method to the bootstrapping method described in (Hu and Liu, 2004), and the shortest path method described in (Kamps et al., 2004). We build a network using only WordNet synonyms and hypernyms. We restrict the test set to the set of adjectives in the General Inquirer lexicon (Stone et al., 1966) because this method is mainly interested in classifying adjectives. The performance of the spin model method, the bootstrapping method, the shortest path method, and the random walk method for only adjectives is shown in Table 1. We notice from the table that the random walk method outperforms both the spin model, the bootstrapping method, and the shortest path method for adjectives. The reported accuracy for the shortest path method only considers the words it could assign a non-zero orientation value. If we consider all words, the accuracy will drop to around 61%. 4.1.1 Varying Parameters As we mentioned in Section 3.4, we use a parameter m to put an upper bound on the length of random walks. In this section, we explore the impact Table 2: Accuracy for SO-PMI with different dataset sizes, the spin model, and the random walks model for 10-fold cross validation and 14 seeds. CV 14 seeds SO-PMI (1 × 107) 61.3 SO-PMI (2 × 109) 76.1 SO-PMI (1 × 1011) 82.8 Spin Model 91.5 81.9 Random Walks 93.1 82.1 of this parameter on our method’s performance. Figure 1 shows the accuracy of the random walk method as a function of the maximum number of steps m. m varies from 5 to 50. We use a network built from WordNet synonyms and hypernyms only. The number of samples k was set to 1000. We perform 10-fold cross validation using the General Inquirer lexicon. We notice that the maximum number of steps m has very little impact on performance until it rises above 30. When it does, the performance drops by no more than 1%, and then it does not change anymore as m increases. An interesting observation is that the proposed method performs quite well with a very small number of steps (around 10). We looked at the dataset to understand why increasing the number of steps beyond 30 negatively affects performance. We found out that when the number of steps is very large, compared to the diameter of the graph, the random walk that starts at ambiguous words, that are hard to classify, have the chance of moving till it hits a node in the opposite class. That does not happen when the limit on the number of steps is smaller because those walks are then terminated without hitting any labeled nodes and hence ignored. Next, we study the effect of the random of samples k on our method’s performance. As explained in Section 3.4, k is the number of samples used by the Monte Carlo algorithm to find an estimate for the hitting time. Figure 2 shows the accuracy of the random walks method as a function of the number of samples k. We use the same settings as in the previous experiment. the only difference is that we fix m at 15 and vary k from 10 to 20000 (note the logarithmic scale). We notice that the performance is badly affected, when the value of k is very small (less than 100). We also notice that 400 after 1000, varying k has very little, if any, effect on performance. This shows that the Monte Carlo algorithm for computing the random walks hitting time performs quite well with values of the number of samples as small as 1000. The preceding experiments suggest that the parameter have very little impact on performance. This suggests that the approach is fairly robust (i.e., it is quite insensitive to different parameter settings). Figure 1: The effect of varying the maximum number of steps (m) on accuracy. Figure 2: The effect of varying the number of samples (k) on accuracy. 4.1.2 Other Experiments We now measure the performance of the proposed method when the system is allowed to abstain from classifying the words for which it have low confidence. We regard the ratio between the hitting time to positive words and hitting time to negative words as a confidence measure and evaluate the top words with the highest confidence level at different values of threshold. Figure 4 shows the accuracy for 10-fold cross validation and for using only 14 seeds at different thresholds. We notice that the accuracy improves by abstaining from classifying the difficult words. The figure shows that the top 60% words are classified with an accuracy greater than 99% for 10-fold cross validation and 92% with 14 seed words. This may be compared to the work descibed in (Takamura et al., 2005) where they achieve the 92% level when they only consider the top 1000 words (28%). Figure 3 shows a learning curve displaying how the performance of the proposed method is affected with varying the labeled set size (i.e., the number of seeds). We notice that the accuracy exceeds 90% when the training set size rises above 20%. The accuracy steadily increases as the labeled data increases. We also looked at the classification accuracy for different parts of speech in Figure 5. we notice that, in the case of 10-fold cross validation, the performance is consistent across parts of speech. However, when we only use 14 seeds all of which are adjectives, similar to (Turney and Littman, 2003), we notice that the performance on adjectives is much better than other parts of speech. When we use 14 seeds but replace some of the adjectives with verbs and nouns like (love, harm, friend, enemy), the performance for nouns and verbs improves considerably at the cost of losing a little bit of the performance on adjectives. We had a closer look at the results to find out what are the reasons behind incorrect predictions. We found two main reasons. First, some words are ambiguous and has more than one sense, possible with different orientations. Disambiguating the sense of words given their context before trying to predict their polarity should solve this problem. The second reason is that some words have very few connection in thesaurus. A possible solution to this might be identifying those words and adding more links to them from glosses of co-occurrence statistics in corpus. Figure 3: The effect of varying the number of seeds on accuracy. 401 Figure 4: Accuracy for words with high confidence measure. Figure 5: Accuracy for different parts of speech. 5 Conclusions Predicting the semantic orientation of words is a very interesting task in Natural Language Processing and it has a wide variety of applications. We proposed a method for automatically predicting the semantic orientation of words using random walks and hitting time. The proposed method is based on the observation that a random walk starting at a given word is more likely to hit another word with the same semantic orientation before hitting a word with a different semantic orientation. The proposed method can be used in a semi-supervised setting where a training set of labeled words is used, and in an unsupervised setting where only a handful of seeds is used to define the two polarity classes. We predict semantic orientation with high accuracy. The proposed method is fast, simple to implement, and does not need any corpus. Acknowledgments This research was funded by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), through the U.S. Army Research Lab. All statements of fact, opinion or conclusions contained herein are those of the authors and should not be construed as representing the official views or policies of IARPA, the ODNI or the U.S. Government. References Alina Andreevskaia and Sabine Bergler. 2006. Mining wordnet for fuzzy sentiment: Sentiment tag extraction from wordnet glosses. In Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics (EACL 2006). Carmen Banea, Rada Mihalcea, and Janyce Wiebe. 2008. A bootstrapping method for building subjectivity lexicons for languages with scarce resources. In Proceedings of the Sixth International Language Resources and Evaluation (LREC’08). Andrea Esuli and Fabrizio Sebastiani. 2005. Determining the semantic orientation of terms through gloss classification. In Proceedings of the 14th Conference on Information and Knowledge Management (CIKM 2005), pages 617–624. Andrea Esuli and Fabrizio Sebastiani. 2006. Sentiwordnet: A publicly available lexical resource for opinion mining. In Proceedings of the 5th Conference on Language Resources and Evaluation (LREC 2006), pages 417–422. Vasileios Hatzivassiloglou and Kathleen R. McKeown. 1997. Predicting the semantic orientation of adjectives. In Proceedings of the eighth conference on European chapter of the Association for Computational Linguistics, pages 174–181. Vasileios Hatzivassiloglou and Janyce Wiebe. 2000. Effects of adjective orientation and gradability on sentence subjectivity. In COLING, pages 299–305. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In KDD ’04: Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 168–177. Jaap Kamps, Maarten Marx, Robert J. Mokken, and Maarten De Rijke. 2004. Using wordnet to measure semantic orientations of adjectives. In National Institute for, pages 1115–1118. Hiroshi Kanayama and Tetsuya Nasukawa. 2006. Fully automatic lexicon expansion for domainoriented sentiment analysis. In Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 355– 363. Soo-Min Kim and Eduard Hovy. 2004. Determining the sentiment of opinions. In Proceedings of the 20th international conference on Computational Linguistics (COLING 2004), pages 1367–1373. 402 George A. Miller. 1995. Wordnet: a lexical database for english. Commun. ACM, 38(11):39–41. Satoshi Morinaga, Kenji Yamanishi, Kenji Tateishi, and Toshikazu Fukushima. 2002. Mining product reputations on the web. In KDD ’02: Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 341–349. Tetsuya Nasukawa and Jeonghee Yi. 2003. Sentiment analysis: capturing favorability using natural language processing. In K-CAP ’03: Proceedings of the 2nd international conference on Knowledge capture, pages 70–77. J. Norris. 1997. Markov chains. Cambridge University Press. Ana-Maria Popescu and Oren Etzioni. 2005. Extracting product features and opinions from reviews. In HLT ’05: Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, pages 339–346. Ellen Riloff and Janyce Wiebe. 2003. Learning extraction patterns for subjective expressions. In Proceedings of the 2003 conference on Empirical methods in natural language processing, pages 105–112. Philip Stone, Dexter Dunphy, Marchall Smith, and Daniel Ogilvie. 1966. The general inquirer: A computer approach to content analysis. The MIT Press. Martin Szummer and Tommi Jaakkola. 2002. Partially labeled classification with markov random walks. In Advances in Neural Information Processing Systems, pages 945–952. Hiroya Takamura, Takashi Inui, and Manabu Okumura. 2005. Extracting semantic orientations of words using spin model. In ACL ’05: Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, pages 133–140. Richard M. Tong. 2001. An operational system for detecting and tracking opinions in on-line discussion. Workshop note, SIGIR 2001 Workshop on Operational Text Classification. Peter Turney and Michael Littman. 2003. Measuring praise and criticism: Inference of semantic orientation from association. ACM Transactions on Information Systems, 21:315–346. Peter D. Turney. 2002. Thumbs up or thumbs down?: semantic orientation applied to unsupervised classification of reviews. In ACL ’02: Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, pages 417–424. Janyce Wiebe, Rebecca Bruce, Matthew Bell, Melanie Martin, and Theresa Wilson. 2001. A corpus study of evaluative and speculative language. In Proceedings of the Second SIGdial Workshop on Discourse and Dialogue, pages 1–10. Janyce Wiebe. 2000. Learning subjective adjectives from corpora. In Proceedings of the Seventeenth National Conference on Artificial Intelligence and Twelfth Conference on Innovative Applications of Artificial Intelligence, pages 735–740. Hong Yu and Vasileios Hatzivassiloglou. 2003. Towards answering opinion questions: separating facts from opinions and identifying the polarity of opinion sentences. In Proceedings of the 2003 conference on Empirical methods in natural language processing, pages 129–136. Xiaojin Zhu, Zoubin Ghahramani, and John Lafferty. 2003. Semi-supervised learning using gaussian fields and harmonic functions. In In ICML, pages 912–919. 403
|
2010
|
41
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 404–413, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Sentiment Learning on Product Reviews via Sentiment Ontology Tree Wei Wei Department of Computer and Information Science Norwegian University of Science and Technology [email protected] Jon Atle Gulla Department of Computer and Information Science Norwegian University of Science and Technology [email protected] Abstract Existing works on sentiment analysis on product reviews suffer from the following limitations: (1) The knowledge of hierarchical relationships of products attributes is not fully utilized. (2) Reviews or sentences mentioning several attributes associated with complicated sentiments are not dealt with very well. In this paper, we propose a novel HL-SOT approach to labeling a product’s attributes and their associated sentiments in product reviews by a Hierarchical Learning (HL) process with a defined Sentiment Ontology Tree (SOT). The empirical analysis against a humanlabeled data set demonstrates promising and reasonable performance of the proposed HL-SOT approach. While this paper is mainly on sentiment analysis on reviews of one product, our proposed HLSOT approach is easily generalized to labeling a mix of reviews of more than one products. 1 Introduction As the internet reaches almost every corner of this world, more and more people write reviews and share opinions on the World Wide Web. The usergenerated opinion-rich reviews will not only help other users make better judgements but they are also useful resources for manufacturers of products to keep track and manage customer opinions. However, as the number of product reviews grows, it becomes difficult for a user to manually learn the panorama of an interesting topic from existing online information. Faced with this problem, research works, e.g., (Hu and Liu, 2004; Liu et al., 2005; Lu et al., 2009), of sentiment analysis on product reviews were proposed and have become a popular research topic at the crossroads of information retrieval and computational linguistics. Carrying out sentiment analysis on product reviews is not a trivial task. Although there have already been a lot of publications investigating on similar issues, among which the representatives are (Turney, 2002; Dave et al., 2003; Hu and Liu, 2004; Liu et al., 2005; Popescu and Etzioni, 2005; Zhuang et al., 2006; Lu and Zhai, 2008; Titov and McDonald, 2008; Zhou and Chaovalit, 2008; Lu et al., 2009), there is still room for improvement on tackling this problem. When we look into the details of each example of product reviews, we find that there are some intrinsic properties that existing previous works have not addressed in much detail. First of all, product reviews constitute domainspecific knowledge. The product’s attributes mentioned in reviews might have some relationships between each other. For example, for a digital camera, comments on image quality are usually mentioned. However, a sentence like “40D handles noise very well up to ISO 800”, also refers to image quality of the camera 40D. Here we say “noise” is a sub-attribute factor of “image quality”. We argue that the hierarchical relationship between a product’s attributes can be useful knowledge if it can be formulated and utilized in product reviews analysis. Secondly, Vocabularies used in product reviews tend to be highly overlapping. Especially, for same attribute, usually same words or synonyms are involved to refer to them and to describe sentiment on them. We believe that labeling existing product reviews with attributes and corresponding sentiment forms an effective training resource to perform sentiment analysis. Thirdly, sentiments expressed in a review or even in a sentence might be opposite on different attributes and not every attributes mentioned are with sentiments. For example, it is common to find a fragment of a review as follows: Example 1: “...I am very impressed with this camera except for its a bit heavy weight especially with 404 camera + camera design and usability image quality lens camera design and usability + weight interface design and usability image quality + noise resolution image quality lens + lens weight + weight interface + menu button interface menu + menu button + button noise + noise resolution + resolution Figure 1: an example of part of a SOT for digital camera extra lenses attached. It has many buttons and two main dials. The first dial is thumb dial, located near shutter button. The second one is the big round dial located at the back of the camera...” In this example, the first sentence gives positive comment on the camera as well as a complaint on its heavy weight. Even if the words “lenses” appears in the review, it is not fair to say the customer expresses any sentiment on lens. The second sentence and the rest introduce the camera’s buttons and dials. It’s also not feasible to try to get any sentiment from these contents. We argue that when performing sentiment analysis on reviews, such as in the Example 1, more attention is needed to distinguish between attributes that are mentioned with and without sentiment. In this paper, we study the problem of sentiment analysis on product reviews through a novel method, called the HL-SOT approach, namely Hierarchical Learning (HL) with Sentiment Ontology Tree (SOT). By sentiment analysis on product reviews we aim to fulfill two tasks, i.e., labeling a target text1 with: 1) the product’s attributes (attributes identification task), and 2) their corresponding sentiments mentioned therein (sentiment annotation task). The result of this kind of labeling process is quite useful because it makes it possible for a user to search reviews on particular attributes of a product. For example, when considering to buy a digital camera, a prospective user who cares more about image quality probably wants to find comments on the camera’s image quality in other users’ reviews. SOT is a tree-like ontology structure that formulates the relationships between a product’s attributes. For example, Fig. 1 is a SOT for a digital camera2. The root node of the SOT is 1Each product review to be analyzed is called target text in the following of this paper. 2Due to the space limitation, not all attributes of a digital camera are enumerated in this SOT; m+/m- means posia camera itself. Each of the non-leaf nodes (white nodes) of the SOT represents an attribute of a camera3. All leaf nodes (gray nodes) of the SOT represent sentiment (positive/negative) nodes respectively associated with their parent nodes. A formal definition on SOT is presented in Section 3.1. With the proposed concept of SOT, we manage to formulate the two tasks of the sentiment analysis to be a hierarchical classification problem. We further propose a specific hierarchical learning algorithm, called HL-SOT algorithm, which is developed based on generalizing an online-learning algorithm H-RLS (Cesa-Bianchi et al., 2006). The HL-SOT algorithm has the same property as the H-RLS algorithm that allows multiple-path labeling (input target text can be labeled with nodes belonging to more than one path in the SOT) and partial-path labeling (the input target text can be labeled with nodes belonging to a path that does not end on a leaf). This property makes the approach well suited for the situation where complicated sentiments on different attributes are expressed in one target text. Unlike the H-RLS algorithm , the HL-SOT algorithm enables each classifier to separately learn its own specific threshold. The proposed HL-SOT approach is empirically analyzed against a human-labeled data set. The experimental results demonstrate promising and reasonable performance of our approach. This paper makes the following contributions: • To the best of our knowledge, with the proposed concept of SOT, the proposed HL-SOT approach is the first work to formulate the tasks of sentiment analysis to be a hierarchical classification problem. • A specific hierarchical learning algorithm is tive/negative sentiment associated with an attribute m. 3A product itself can be treated as an overall attribute of the product. 405 further proposed to achieve tasks of sentiment analysis in one hierarchical classification process. • The proposed HL-SOT approach can be generalized to make it possible to perform sentiment analysis on target texts that are a mix of reviews of different products, whereas existing works mainly focus on analyzing reviews of only one type of product. The remainder of the paper is organized as follows. In Section 2, we provide an overview of related work on sentiment analysis. Section 3 presents our work on sentiment analysis with HLSOT approach. The empirical analysis and the results are presented in Section 4, followed by the conclusions, discussions, and future work in Section 5. 2 Related Work The task of sentiment analysis on product reviews was originally performed to extract overall sentiment from the target texts. However, in (Turney, 2002), as the difficulty shown in the experiments, the whole sentiment of a document is not necessarily the sum of its parts. Then there came up with research works shifting focus from overall document sentiment to sentiment analysis based on product attributes (Hu and Liu, 2004; Popescu and Etzioni, 2005; Ding and Liu, 2007; Liu et al., 2005). Document overall sentiment analysis is to summarize the overall sentiment in the document. Research works related to document overall sentiment analysis mainly rely on two finer levels sentiment annotation: word-level sentiment annotation and phrase-level sentiment annotation. The wordlevel sentiment annotation is to utilize the polarity annotation of words in each sentence and summarize the overall sentiment of each sentimentbearing word to infer the overall sentiment within the text (Hatzivassiloglou and Wiebe, 2000; Andreevskaia and Bergler, 2006; Esuli and Sebastiani, 2005; Esuli and Sebastiani, 2006; Hatzivassiloglou and McKeown, 1997; Kamps et al., 2004; Devitt and Ahmad, 2007; Yu and Hatzivassiloglou, 2003). The phrase-level sentiment annotation focuses sentiment annotation on phrases not words with concerning that atomic units of expression is not individual words but rather appraisal groups (Whitelaw et al., 2005). In (Wilson et al., 2005), the concepts of prior polarity and contextual polarity were proposed. This paper presented a system that is able to automatically identify the contextual polarity for a large subset of sentiment expressions. In (Turney, 2002), an unsupervised learning algorithm was proposed to classify reviews as recommended or not recommended by averaging sentiment annotation of phrases in reviews that contain adjectives or adverbs. However, the performances of these works are not good enough for sentiment analysis on product reviews, where sentiment on each attribute of a product could be so complicated that it is unable to be expressed by overall document sentiment. Attributes-based sentiment analysis is to analyze sentiment based on each attribute of a product. In (Hu and Liu, 2004), mining product features was proposed together with sentiment polarity annotation for each opinion sentence. In that work, sentiment analysis was performed on product attributes level. In (Liu et al., 2005), a system with framework for analyzing and comparing consumer opinions of competing products was proposed. The system made users be able to clearly see the strengths and weaknesses of each product in the minds of consumers in terms of various product features. In (Popescu and Etzioni, 2005), Popescu and Etzioni not only analyzed polarity of opinions regarding product features but also ranked opinions based on their strength. In (Liu et al., 2007), Liu et al. proposed Sentiment-PLSA that analyzed blog entries and viewed them as a document generated by a number of hidden sentiment factors. These sentiment factors may also be factors based on product attributes. In (Lu and Zhai, 2008), Lu et al. proposed a semi-supervised topic models to solve the problem of opinion integration based on the topic of a product’s attributes. The work in (Titov and McDonald, 2008) presented a multi-grain topic model for extracting the ratable attributes from product reviews. In (Lu et al., 2009), the problem of rated attributes summary was studied with a goal of generating ratings for major aspects so that a user could gain different perspectives towards a target entity. All these research works concentrated on attribute-based sentiment analysis. However, the main difference with our work is that they did not sufficiently utilize the hierarchical relationships among a product attributes. Although a method of ontologysupported polarity mining, which also involved 406 ontology to tackle the sentiment analysis problem, was proposed in (Zhou and Chaovalit, 2008), that work studied polarity mining by machine learning techniques that still suffered from a problem of ignoring dependencies among attributes within an ontology’s hierarchy. In the contrast, our work solves the sentiment analysis problem as a hierarchical classification problem that fully utilizes the hierarchy of the SOT during training and classification process. 3 The HL-SOT Approach In this section, we first propose a formal definition on SOT. Then we formulate the HL-SOT approach. In this novel approach, tasks of sentiment analysis are to be achieved in a hierarchical classification process. 3.1 Sentiment Ontology Tree As we discussed in Section 1, the hierarchial relationships among a product’s attributes might help improve the performance of attribute-based sentiment analysis. We propose to use a tree-like ontology structure SOT, i.e., Sentiment Ontology Tree, to formulate relationships among a product’s attributes. Here,we give a formal definition on what a SOT is. Definition 1 [SOT] SOT is an abbreviation for Sentiment Ontology Tree that is a tree-like ontology structure T(v, v+, v−, T). v is the root node of T which represents an attribute of a given product. v+ is a positive sentiment leaf node associated with the attribute v. v−is a negative sentiment leaf node associated with the attribute v. T is a set of subtrees. Each element of T is also a SOT T ′(v′, v′+, v′−, T′) which represents a subattribute of its parent attribute node. By the Definition 1, we define a root of a SOT to represent an attribute of a product. The SOT’s two leaf child nodes are sentiment (positive/negative) nodes associated with the root attribute. The SOT recursively contains a set of sub-SOTs where each root of a sub-SOT is a non-leaf child node of the root of the SOT and represent a sub-attribute belonging to its parent attribute. This definition successfully describes the hierarchical relationships among all the attributes of a product. For example, in Fig. 1 the root node of the SOT for a digital camera is its general overview attribute. Comments on a digital camera’s general overview attribute appearing in a review might be like “this camera is great”. The “camera” SOT has two sentiment leaf child nodes as well as three non-leaf child nodes which are respectively root nodes of sub-SOTs for sub-attributes “design and usability”, “image quality”, and “lens”. These sub-attributes SOTs recursively repeat until each node in the SOT does not have any more non-leaf child node, which means the corresponding attributes do not have any sub-attributes, e.g., the attribute node “button” in Fig. 1. 3.2 Sentiment Analysis with SOT In this subsection, we present the HL-SOT approach. With the defined SOT, the problem of sentiment analysis is able to be formulated to be a hierarchial classification problem. Then a specific hierarchical learning algorithm is further proposed to solve the formulated problem. 3.2.1 Problem Formulation In the proposed HL-SOT approach, each target text is to be indexed by a unit-norm vector x ∈ X, X = Rd. Let Y = {1, ..., N} denote the finite set of nodes in SOT. Let y = {y1, ..., yN} ∈ {0, 1}N be a label vector to a target text x, where ∀i ∈Y : yi = { 1, if x is labeled by the classifier of node i, 0, if x is not labeled by the classifier of node i. A label vector y ∈{0, 1}N is said to respect SOT if and only if y satisfies ∀i ∈Y , ∀j ∈ A(i) : if yi = 1 then yj = 1, where A(i) represents a set ancestor nodes of i, i.e.,A(i) = {x|ancestor(i, x)}. Let Y denote a set of label vectors that respect SOT. Then the tasks of sentiment analysis can be formulated to be the goal of a hierarchical classification that is to learn a function f : X →Y, that is able to label each target text x ∈X with classifier of each node and generating with x a label vector y ∈Y that respects SOT. The requirement of a generated label vector y ∈Y ensures that a target text is to be labeled with a node only if its parent attribute node is labeled with the target text. For example, in Fig. 1 a review is to be labeled with “image quality +” requires that the review should be successively labeled as related to “camera” and “image quality”. This is reasonable and consistent with intuition, because if a review cannot be identified to be related to a camera, it is not safe to infer that the review is commenting a camera’s image quality with positive sentiment. 407 3.2.2 HL-SOT Algorithm The algorithm H-RLS studied in (Cesa-Bianchi et al., 2006) solved a similar hierarchical classification problem as we formulated above. However, the H-RLS algorithm was designed as an onlinelearning algorithm which is not suitable to be applied directly in our problem setting. Moreover, the algorithm H-RLS defined the same value as the threshold of each node classifier. We argue that if the threshold values could be learned separately for each classifiers, the performance of classification process would be improved. Therefore we propose a specific hierarchical learning algorithm, named HL-SOT algorithm, that is able to train each node classifier in a batch-learning setting and allows separately learning for the threshold of each node classifier. Defining the f function Let w1, ..., wN be weight vectors that define linear-threshold classifiers of each node in SOT. Let W = (w1, ..., wN)⊤ be an N × d matrix called weight matrix. Here we generalize the work in (Cesa-Bianchi et al., 2006) and define the hierarchical classification function f as: ˆy = f(x) = g(W · x), where x ∈X, ˆy ∈Y. Let z = W · x. Then the function ˆy = g(z) on an N-dimensional vector z defines: ∀i = 1, ..., N : ˆyi = B(zi ≥θi), if i is a root node in SOT or yj = 1 for j = P(i), 0, else where P(i) is the parent node of i in SOT and B(S) is a boolean function which is 1 if and only if the statement S is true. Then the hierarchical classification function f is parameterized by the weight matrix W = (w1, ..., wN)⊤and threshold vector θ = (θ1, ..., θN)⊤. The hierarchical learning algorithm HL-SOT is proposed for learning the parameters of W and θ. Parameters Learning for f function Let D denote the training data set: D = {(r, l)|r ∈X, l ∈ Y}. In the HL-SOT learning process, the weight matrix W is firstly initialized to be a 0 matrix, where each row vector wi is a 0 vector. The threshold vector is initialized to be a 0 vector. Each instance in the training set D goes into the training process. When a new instance rt is observed, each row vector wi,t of the weight matrix Wt is updated by a regularized least squares estimator given by: wi,t = (I + Si,Q(i,t−1)S⊤ i,Q(i,t−1) + rtr⊤ t )−1 ×Si,Q(i,t−1)(li,i1, li,i2, ..., li,iQ(i,t−1))⊤ (1) where I is a d × d identity matrix, Q(i, t −1) denotes the number of times the parent of node i observes a positive label before observing the instance rt, Si,Q(i,t−1) = [ri1, ..., riQ(i,t−1)] is a d × Q(i, t−1) matrix whose columns are the instances ri1, ..., riQ(i,t−1), and (li,i1, li,i2, ..., li,iQ(i,t−1))⊤is a Q(i, t−1)-dimensional vector of the corresponding labels observed by node i. The Formula 1 restricts that the weight vector wi,t of the classifier i is only updated on the examples that are positive for its parent node. Then the label vector ˆyrt is computed for the instance rt, before the real label vector lrt is observed. Then the current threshold vector θt is updated by: θt+1 = θt + ϵ(ˆyrt −lrt), (2) where ϵ is a small positive real number that denotes a corrective step for correcting the current threshold vector θt. To illustrate the idea behind the Formula 2, let y′ t = ˆyrt −lrt. Let y′ i,t denote an element of the vector y′ t. The Formula 2 correct the current threshold θi,t for the classifier i in the following way: • If y′ i,t = 0, it means the classifier i made a proper classification for the current instance rt. Then the current threshold θi does not need to be adjusted. • If y′ i,t = 1, it means the classifier i made an improper classification by mistakenly identifying the attribute i of the training instance rt that should have not been identified. This indicates the value of θi is not big enough to serve as a threshold so that the attribute i in this case can be filtered out by the classifier i. Therefore, the current threshold θi will be adjusted to be larger by ϵ. • If y′ i,t = −1, it means the classifier i made an improper classification by failing to identify the attribute i of the training instance rt that should have been identified. This indicates the value of θi is not small enough to serve as a threshold so that the attribute i in this case 408 Algorithm 1 Hierarchical Learning Algorithm HL-SOT INITIALIZATION: 1: Each vector wi,1, i = 1, ..., N of weight matrix W1 is set to be 0 vector 2: Threshold vector θ1 is set to be 0 vector BEGIN 3: for t = 1, ..., |D| do 4: Observe instance rt ∈X 5: for i = 1, ...N do 6: Update each row wi,t of weight matrix Wt by Formula 1 7: end for 8: Compute ˆyrt = f(rt) = g(Wt · rt) 9: Observe label vector lrt ∈Y of the instance rt 10: Update threshold vector θt by Formula 2 11: end for END can be recognized by the classifier i. Therefore, the current threshold θi will be adjusted to be smaller by ϵ. The hierarchial learning algorithm HL-SOT is presented as in Algorithm 1. The HL-SOT algorithm enables each classifier to have its own specific threshold value and allows this threshold value can be separately learned and corrected through the training process. It is not only a batchlearning setting of the H-RLS algorithm but also a generalization to the latter. If we set the algorithm HL-SOT’s parameter ϵ to be 0, the HL-SOT becomes the H-RLS algorithm in a batch-learning setting. 4 Empirical Analysis In this section, we conduct systematic experiments to perform empirical analysis on our proposed HLSOT approach against a human-labeled data set. In order to encode each text in the data set by a d-dimensional vector x ∈Rd, we first remove all the stop words and then select the top d frequency terms appearing in the data set to construct the index term space. Our experiments are intended to address the following questions:(1) whether utilizing the hierarchical relationships among labels help to improve the accuracy of the classification? (2) whether the introduction of separately learning threshold for each classifier help to improve the accuracy of the classification? (3) how does the corrective step ϵ impact the performance of the proposed approach?(4)how does the dimensionality d of index terms space impact the proposed approach’s computing efficiency and accuracy? 4.1 Data Set Preparation The data set contains 1446 snippets of customer reviews on digital cameras that are collected from a customer review website4. We manually construct a SOT for the product of digital cameras. The constructed SOT (e.g., Fig. 1) contains 105 nodes that include 35 non-leaf nodes representing attributes of the digital camera and 70 leaf nodes representing associated sentiments with attribute nodes. Then we label all the snippets with corresponding labels of nodes in the constructed SOT complying with the rule that a target text is to be labeled with a node only if its parent attribute node is labeled with the target text. We randomly divide the labeled data set into five folds so that each fold at least contains one example snippets labeled by each node in the SOT. For each experiment setting, we run 5 experiments to perform cross-fold evaluation by randomly picking three folds as the training set and the other two folds as the testing set. All the testing results are averages over 5 running of experiments. 4.2 Evaluation Metrics Since the proposed HL-SOT approach is a hierarchical classification process, we use three classic loss functions for measuring classification performance. They are the One-error Loss (O-Loss) function, the Symmetric Loss (S-Loss) function, and the Hierarchical Loss (H-Loss) function: • One-error loss (O-Loss) function is defined as: LO(ˆy, l) = B(∃i : ˆyi ̸= li), where ˆy is the prediction label vector and l is the true label vector; B is the boolean function as defined in Section 3.2.2. • Symmetric loss (S-Loss) function is defined as: LS(ˆy, l) = N ∑ i=1 B(ˆyi ̸= li), • Hierarchical loss (H-Loss) function is defined as: LH(ˆy, l) = N ∑ i=1 B(ˆyi ̸= li ∧∀j ∈A(i), ˆyj = lj), 4http://www.consumerreview.com/ 409 Table 1: Performance Comparisons (A Smaller Loss Value Means a Better Performance) Metrics Dimensinality=110 Dimensinality=220 H-RLS HL-flat HL-SOT H-RLS HL-flat HL-SOT O-Loss 0.9812 0.8772 0.8443 0.9783 0.8591 0.8428 S-Loss 8.5516 2.8921 2.3190 7.8623 2.8449 2.2812 H-Loss 3.2479 1.1383 1.0366 3.1029 1.1298 1.0247 0 0.02 0.04 0.06 0.08 0.1 0.838 0.84 0.842 0.844 0.846 0.848 0.85 0.852 Corrective Step O−Loss d=110 d=220 (a) O-Loss 0 0.02 0.04 0.06 0.08 0.1 2.15 2.2 2.25 2.3 2.35 2.4 Corrective Step S−Loss d=110 d=220 (b) S-Loss 0 0.02 0.04 0.06 0.08 0.1 1.02 1.025 1.03 1.035 1.04 1.045 1.05 Corrective Step H−Loss d=110 d=220 (c) H-Loss Figure 2: Impact of Corrective Step ϵ where A denotes a set of nodes that are ancestors of node i in SOT. Unlike the O-Loss function and the S-Loss function, the H-Loss function captures the intuition that loss should only be charged on a node whenever a classification mistake is made on a node of SOT but no more should be charged for any additional mistake occurring in the subtree of that node. It measures the discrepancy between the prediction labels and the true labels with consideration on the SOT structure defined over the labels. In our experiments, the recorded loss function values for each experiment running are computed by averaging the loss function values of each testing snippets in the testing set. 4.3 Performance Comparison In order to answer the questions (1), (2) in the beginning of this section, we compare our HLSOT approach with the following two baseline approaches: • HL-flat: The HL-flat approach involves an algorithm that is a “flat” version of HL-SOT algorithm by ignoring the hierarchical relationships among labels when each classifier is trained. In the training process of HL-flat, the algorithm reflexes the restriction in the HL-SOT algorithm that requires the weight vector wi,t of the classifier i is only updated on the examples that are positive for its parent node. • H-RLS: The H-RLS approach is implemented by applying the H-RLS algorithm studied in (Cesa-Bianchi et al., 2006). Unlike our proposed HL-SOT algorithm that enables the threshold values to be learned separately for each classifiers in the training process, the H-RLS algorithm only uses an identical threshold values for each classifiers in the classification process. Experiments are conducted on the performance comparison between the proposed HL-SOT approach with HL-flat approach and the H-RLS approach. The dimensionality d of the index term space is set to be 110 and 220. The corrective step ϵ is set to be 0.005. The experimental results are summarized in Table 1. From Table 1, we can observe that the HL-SOT approach generally beats the H-RLS approach and HL-flat approach on OLoss, S-Loss, and H-Loss respectively. The HRLS performs worse than the HL-flat and the HLSOT, which indicates that the introduction of separately learning threshold for each classifier did improve the accuracy of the classification. The HLSOT approach performs better than the HL-flat, which demonstrates the effectiveness of utilizing the hierarchical relationships among labels. 4.4 Impact of Corrective Step ϵ The parameter ϵ in the proposed HL-SOT approach controls the corrective step of the classifiers’ thresholds when any mistake is observed in the training process. If the corrective step ϵ is set too large, it might cause the algorithm to be too 410 50 100 150 200 250 300 0.84 0.841 0.842 0.843 0.844 0.845 0.846 Dimensionality of Index Term Space O−Loss (a) O-Loss 50 100 150 200 250 300 2.26 2.27 2.28 2.29 2.3 2.31 2.32 2.33 2.34 2.35 Dimensionality of Index Term Space S−Loss (b) S-Loss 50 100 150 200 250 300 1.01 1.015 1.02 1.025 1.03 1.035 1.04 1.045 Dimensionality of Index Term Space H−Loss (c) H-Loss Figure 3: Impact of Dimensionality d of Index Term Space (ϵ = 0.005) sensitive to each observed mistake. On the contrary, if the corrective step is set too small, it might cause the algorithm not sensitive enough to the observed mistakes. Hence, the corrective step ϵ is a factor that might impact the performance of the proposed approach. Fig. 2 demonstrates the impact of ϵ on O-Loss, S-Loss, and H-Loss. The dimensionality of index term space d is set to be 110 and 220. The value of ϵ is set to vary from 0.001 to 0.1 with each step of 0.001. Fig. 2 shows that the parameter ϵ impacts the classification performance significantly. As the value of ϵ increase, the O-Loss, S-Loss, and H-Loss generally increase (performance decrease). In Fig. 2c it is obviously detected that the H-Loss decreases a little (performance increase) at first before it increases (performance decrease) with further increase of the value of ϵ. This indicates that a finer-grained value of ϵ will not necessarily result in a better performance on the H-loss. However, a fine-grained corrective step generally makes a better performance than a coarse-grained corrective step. 4.5 Impact of Dimensionality d of Index Term Space In the proposed HL-SOT approach, the dimensionality d of the index term space controls the number of terms to be indexed. If d is set too small, important useful terms will be missed that will limit the performance of the approach. However, if d is set too large, the computing efficiency will be decreased. Fig. 3 shows the impacts of the parameter d respectively on O-Loss, S-Loss, and H-Loss, where d varies from 50 to 300 with each step of 10 and the ϵ is set to be 0.005. From Fig. 3, we observe that as the d increases the O-Loss, S-Loss, and H-Loss generally decrease (performance increase). This means that when more terms are indexed better performance can be achieved by the HL-SOT approach. However, 50 100 150 200 250 300 0 2 4 6 8 10 12 x 10 6 Dimensionality of Index Term Space Time Consuming (ms) Figure 4: Time Consuming Impacted by d considering the computing efficiency impacted by d, Fig. 4 shows that the computational complexity of our approach is non-linear increased with d’s growing, which indicates that indexing more terms will improve the accuracy of our proposed approach although this is paid by decreasing the computing efficiency. 5 Conclusions, Discussions and Future Work In this paper, we propose a novel and effective approach to sentiment analysis on product reviews. In our proposed HL-SOT approach, we define SOT to formulate the knowledge of hierarchical relationships among a product’s attributes and tackle the problem of sentiment analysis in a hierarchical classification process with the proposed algorithm. The empirical analysis on a humanlabeled data set demonstrates the promising results of our proposed approach. The performance comparison shows that the proposed HL-SOT approach outperforms two baselines: the HL-flat and the H-RLS approach. This confirms two intuitive motivations based on which our approach is proposed: 1) separately learning threshold values for 411 each classifier improve the classification accuracy; 2) knowledge of hierarchical relationships of labels improve the approach’s performance. The experiments on analyzing the impact of parameter ϵ indicate that a fine-grained corrective step generally makes a better performance than a coarsegrained corrective step. The experiments on analyzing the impact of the dimensionality d show that indexing more terms will improve the accuracy of our proposed approach while the computing efficiency will be greatly decreased. The focus of this paper is on analyzing review texts of one product. However, the framework of our proposed approach can be generalized to deal with a mix of review texts of more than one products. In this generalization for sentiment analysis on multiple products reviews, a “big” SOT is constructed and the SOT for each product reviews is a sub-tree of the “big” SOT. The sentiment analysis on multiple products reviews can be performed the same way the HL-SOT approach is applied on single product reviews and can be tackled in a hierarchical classification process with the “big” SOT. This paper is motivated by the fact that the relationships among a product’s attributes could be a useful knowledge for mining product review texts. The SOT is defined to formulate this knowledge in the proposed approach. However, what attributes to be included in a product’s SOT and how to structure these attributes in the SOT is an effort of human beings. The sizes and structures of SOTs constructed by different individuals may vary. How the classification performance will be affected by variances of the generated SOTs is worthy of study. In addition, an automatic method to learn a product’s attributes and the structure of SOT from existing product review texts will greatly benefit the efficiency of the proposed approach. We plan to investigate on these issues in our future work. Acknowledgments The authors would like to thank the anonymous reviewers for many helpful comments on the manuscript. This work is funded by the Research Council of Norway under the VERDIKT research programme (Project No.: 183337). References Alina Andreevskaia and Sabine Bergler. 2006. Mining wordnet for a fuzzy sentiment: Sentiment tag extraction from wordnet glosses. In Proceedings of 11th Conference of the European Chapter of the Association for Computational Linguistics (EACL’06), Trento, Italy. Nicol`o Cesa-Bianchi, Claudio Gentile, and Luca Zaniboni. 2006. Incremental algorithms for hierarchical classification. Journal of Machine Learning Research (JMLR), 7:31–54. Kushal Dave, Steve Lawrence, and David M. Pennock. 2003. Mining the peanut gallery: opinion extraction and semantic classification of product reviews. In Proceedings of 12nd International World Wide Web Conference (WWW’03), Budapest, Hungary. Ann Devitt and Khurshid Ahmad. 2007. Sentiment polarity identification in financial news: A cohesionbased approach. In Proceedings of 45th Annual Meeting of the Association for Computational Linguistics (ACL’07), Prague, Czech Republic. Xiaowen Ding and Bing Liu. 2007. The utility of linguistic rules in opinion mining. In Proceedings of 30th Annual International ACM Special Interest Group on Information Retrieval Conference (SIGIR’07), Amsterdam, The Netherlands. Andrea Esuli and Fabrizio Sebastiani. 2005. Determining the semantic orientation of terms through gloss classification. In Proceedings of 14th ACM Conference on Information and Knowledge Management (CIKM’05), Bremen, Germany. Andrea Esuli and Fabrizio Sebastiani. 2006. Sentiwordnet: A publicly available lexical resource for opinion mining. In Proceedings of 5th International Conference on Language Resources and Evaluation (LREC’06), Genoa, Italy. Vasileios Hatzivassiloglou and Kathleen R. McKeown. 1997. Predicting the semantic orientation of adjectives. In Proceedings of 35th Annual Meeting of the Association for Computational Linguistics (ACL’97), Madrid, Spain. Vasileios Hatzivassiloglou and Janyce M. Wiebe. 2000. Effects of adjective orientation and gradability on sentence subjectivity. In Proceedings of 18th International Conference on Computational Linguistics (COLING’00), Saarbr¨uken, Germany. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In Proceedings of 10th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD’04), Seattle, USA. Jaap Kamps, Maarten Marx, R. ort. Mokken, and Maarten de Rijke. 2004. Using WordNet to measure semantic orientation of adjectives. In Proceedings of 4th International Conference on Language Resources and Evaluation (LREC’04), Lisbon, Portugal. 412 Bing Liu, Minqing Hu, and Junsheng Cheng. 2005. Opinion observer: analyzing and comparing opinions on the web. In Proceedings of 14th International World Wide Web Conference (WWW’05), Chiba, Japan. Yang Liu, Xiangji Huang, Aijun An, and Xiaohui Yu. 2007. ARSA: a sentiment-aware model for predicting sales performance using blogs. In Proceedings of the 30th Annual International ACM Special Interest Group on Information Retrieval Conference (SIGIR’07), Amsterdam, The Netherlands. Yue Lu and Chengxiang Zhai. 2008. Opinion integration through semi-supervised topic modeling. In Proceedings of 17th International World Wide Web Conference (WWW’08), Beijing, China. Yue Lu, ChengXiang Zhai, and Neel Sundaresan. 2009. Rated aspect summarization of short comments. In Proceedings of 18th International World Wide Web Conference (WWW’09), Madrid, Spain. Ana-Maria Popescu and Oren Etzioni. 2005. Extracting product features and opinions from reviews. In Proceedings of Human Language Technology Conference and Empirical Methods in Natural Language Processing Conference (HLT/EMNLP’05), Vancouver, Canada. Ivan Titov and Ryan T. McDonald. 2008. Modeling online reviews with multi-grain topic models. In Proceedings of 17th International World Wide Web Conference (WWW’08), Beijing, China. Peter D. Turney. 2002. Thumbs up or thumbs down? semantic orientation applied to unsupervised classification of reviews. In Proceedings of 40th Annual Meeting of the Association for Computational Linguistics (ACL’02), Philadelphia, USA. Casey Whitelaw, Navendu Garg, and Shlomo Argamon. 2005. Using appraisal taxonomies for sentiment analysis. In Proceedings of 14th ACM Conference on Information and Knowledge Management (CIKM’05), Bremen, Germany. Theresa Wilson, Janyce Wiebe, and Paul Hoffmann. 2005. Recognizing contextual polarity in phraselevel sentiment analysis. In Proceedings of Human Language Technology Conference and Empirical Methods in Natural Language Processing Conference (HLT/EMNLP’05), Vancouver, Canada. Hong Yu and Vasileios Hatzivassiloglou. 2003. Towards answering opinion questions: Separating facts from opinions and identifying the polarity of opinion sentences. In Proceedings of 8th Conference on Empirical Methods in Natural Language Processing (EMNLP’03), Sapporo, Japan. Lina Zhou and Pimwadee Chaovalit. 2008. Ontologysupported polarity mining. Journal of the American Society for Information Science and Technology (JASIST), 59(1):98–110. Li Zhuang, Feng Jing, and Xiao-Yan Zhu. 2006. Movie review mining and summarization. In Proceedings of the 15th ACM International Conference on Information and knowledge management (CIKM’06), Arlington, USA. 413
|
2010
|
42
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 414–423, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Employing Personal/Impersonal Views in Supervised and Semi-supervised Sentiment Classification Shoushan Li†‡ Chu-Ren Huang† Guodong Zhou‡ Sophia Yat Mei Lee† †Department of Chinese and Bilingual Studies The Hong Kong Polytechnic University {shoushan.li,churenhuang, sophiaym}@gmail.com ‡ Natural Language Processing Lab School of Computer Science and Technology Soochow University, China [email protected] Abstract In this paper, we adopt two views, personal and impersonal views, and systematically employ them in both supervised and semi-supervised sentiment classification. Here, personal views consist of those sentences which directly express speaker’s feeling and preference towards a target object while impersonal views focus on statements towards a target object for evaluation. To obtain them, an unsupervised mining approach is proposed. On this basis, an ensemble method and a co-training algorithm are explored to employ the two views in supervised and semi-supervised sentiment classification respectively. Experimental results across eight domains demonstrate the effectiveness of our proposed approach. 1 Introduction As a special task of text classification, sentiment classification aims to classify a text according to the expressed sentimental polarities of opinions such as ‘thumb up’ or ‘thumb down’ on the movies (Pang et al., 2002). This task has recently received considerable interests in the Natural Language Processing (NLP) community due to its wide applications. In general, the objective of sentiment classification can be represented as a kind of binary relation R, defined as an ordered triple (X, Y, G), where X is an object set including different kinds of people (e.g. writers, reviewers, or users), Y is another object set including the target objects (e.g. products, events, or even some people), and G is a subset of the Cartesian product X Y × . The concerned relation in sentiment classification is X ’s evaluation on Y, such as ‘thumb up’, ‘thumb down’, ‘favorable’, and ‘unfavorable’. Such relation is usually expressed in text by stating the information involving either a person (one element in X ) or a target object itself (one element in Y ). The first type of statement called personal view, e.g. ‘I am so happy with this book’, contains X ’s “subjective” feeling and preference towards a target object, which directly expresses sentimental evaluation. This kind of information is normally domain-independent and serves as highly relevant clues to sentiment classification. The latter type of statement called impersonal view, e.g. ‘it is too small’, contains Y ’s “objective” (i.e. or at least criteria-based) evaluation of the target object. This kind of information tends to contain much domain-specific classification knowledge. Although such information is sometimes not as explicit as personal views in classifying the sentiment of a text, speaker’s sentiment is usually implied by the evaluation result. It is well-known that sentiment classification is very domain-specific (Blitzer et al., 2007), so it is critical to eliminate its dependence on a large-scale labeled data for its wide applications. Since the unlabeled data is ample and easy to collect, a successful semi-supervised sentiment classification system would significantly minimize the involvement of labor and time. Therefore, given the two different views mentioned above, one promising application is to adopt them in co-training algorithms, which has been proven to be an effective semi-supervised learning strategy of incorporating unlabeled data to further improve the classification performance (Zhu, 2005). In addition, we would show that personal/impersonal views are linguistically marked and mining them in text can be easily performed without special annotation. 414 In this paper, we systematically employ personal/impersonal views in supervised and semi-supervised sentiment classification. First, an unsupervised bootstrapping method is adopted to automatically separate one document into personal and impersonal views. Then, both views are employed in supervised sentiment classification via an ensemble of individual classifiers generated by each view. Finally, a co-training algorithm is proposed to incorporate unlabeled data for semi-supervised sentiment classification. The remainder of this paper is organized as follows. Section 2 introduces the related work of sentiment classification. Section 3 presents our unsupervised approach for mining personal and impersonal views. Section 4 and Section 5 propose our supervised and semi-supervised methods on sentiment classification respectively. Experimental results are presented and analyzed in Section 6. Section 7 discusses on the differences between personal/impersonal and subjective/objective. Finally, Section 8 draws our conclusions and outlines the future work. 2 Related Work Recently, a variety of studies have been reported on sentiment classification at different levels: word level (Esuli and Sebastiani, 2005), phrase level (Wilson et al., 2009), sentence level (Kim and Hovy, 2004; Liu et al., 2005), and document level (Turney, 2002; Pang et al., 2002). This paper focuses on the document-level sentiment classification. Generally, document-level sentiment classification methods can be categorized into three types: unsupervised, supervised, and semi-supervised. Unsupervised methods involve deriving a sentiment classifier without any labeled documents. Most of previous work use a set of labeled sentiment words called seed words to perform unsupervised classification. Turney (2002) determines the sentiment orientation of a document by calculating point-wise mutual information between the words in the document and the seed words of ‘excellent’ and ‘poor’. Kennedy and Inkpen (2006) use a term-counting method with a set of seed words to determine the sentiment. Zagibalov and Carroll (2008) first propose a seed word selection approach and then apply the same term-counting method for Chinese sentiment classifications. These unsupervised approaches are believed to be domain-independent for sentiment classification. Supervised methods consider sentiment classification as a standard classification problem in which labeled data in a domain are used to train a domain-specific classifier. Pang et al. (2002) are the first to apply supervised machine learning methods to sentiment classification. Subsequently, many other studies make efforts to improve the performance of machine learning-based classifiers by various means, such as using subjectivity summarization (Pang and Lee, 2004), seeking new superior textual features (Riloff et al., 2006), and employing document subcomponent information (McDonald et al., 2007). As far as the challenge of domain-dependency is concerned, Blitzer et al. (2007) present a domain adaptation approach for sentiment classification. Semi-supervised methods combine unlabeled data with labeled training data (often small-scaled) to improve the models. Compared to the supervised and unsupervised methods, semi-supervised methods for sentiment classification are relatively new and have much less related studies. Dasgupta and Ng (2009) integrate various methods in semi-supervised sentiment classification including spectral clustering, active learning, transductive learning, and ensemble learning. They achieve a very impressive improvement across five domains. Wan (2009) applies a co-training method to semi-supervised learning with labeled English corpus and unlabeled Chinese corpus for Chinese sentiment classification. 3 Unsupervised Mining of Personal and Impersonal Views As mentioned in Section 1, the objective of sentiment classification is to classify a specific binary relation: X ’s evaluation on Y, where X is an object set including different kinds of persons and Y is another object set including the target objects to be evaluated. First of all, we focus on an analysis on sentences in product reviews regarding the two views: personal and impersonal views. The personal view consists of personal sentences (i.e. X ’s sentences) exemplified below: I. Personal preference: E1: I love this breadmaker! E2: I disliked it from the beginning. II. Personal emotion description: E3: Very disappointed! E4: I am happy with the product. III. Personal actions: 415 E5: Do not waste your money. E6: I have recommended this machine to all my friends. The impersonal view consists of impersonal sentences (i.e.Y ’s sentences) exemplified below: I. Impersonal feature description: E7: They are too thin to start with. E8: This product is extremely quiet. II. Impersonal evaluation: E9: It's great. E10: The product is a waste of time and money. III. Impersonal actions: E11: This product not even worth a penny. E12: It broke down again and again. We find that the subject of a sentence presents important cues for personal/impersonal views, even though a formal and computable definition of this contrast cannot be found. Here, subject refers to one of the two main constituents in the traditional English grammar (the other constituent being the predicate) (Crystal, 2003)1. For example, the subjects in the above examples of E1, E7 and E11 are ‘I’, ‘they’, and ‘this product’ respectively. For automatic mining the two views, personal/impersonal sentences can be defined according to their subjects: Personal sentence: the sentence whose subject is (or represents) a person. Impersonal sentence: the sentence whose subject is not (does not represent) a person. In this study, we mainly focus on product review classification where the target object in the set Y is not a person. The definitions need to be adjusted when the evaluation target itself is a person, e.g. the political sentiment classification by Durant and Smith (2007). Our unsupervised mining approach for mining personal and impersonal sentences consists of two main steps. First, we extract an initial set of personal and impersonal sentences with some heuristic rules: If the first word of one sentence is (or implies) a personal pronoun including ‘I’, ‘we’, and ‘do’, then the sentence is extracted as a personal sentence; If the first word of one sentence is an impersonal pronoun including 'it', 'they', 'this', and 'these', then the sentence is extracted as an impersonal sentence. Second, we apply the classifier which is trained with the initial set of personal and impersonal sentences to classify the remaining sentences. This step aims to classify the sentences without pronouns 1 The subject has the grammatical function in a sentence of relating its constituent (a noun phrase) by means of the verb to any other elements present in the sentence, i.e. objects, complements, and adverbials. (e.g. E3). Figure 1 shows the unsupervised mining algorithm. Input: The training data D Output: All personal and impersonal sentences, i.e. sentence sets personal S and impersonal S . Procedure: (1). Segment all documents in D to sentences S using punctuations (such as periods and interrogation marks) (2). Apply the heuristic rules to classify the sentences S with proper pronouns into, 1 p S and 1iS (3). Train a binary classifier p i f − with 1 p S and 1iS (4). Use p i f − to classify the remaining sentences into 2 p S and 2 iS (5). 1 2 personal p p S S S = ∪ , 1 2 impersonal i i S S S = ∪ Figure 1: The algorithm for unsupervised mining personal and impersonal sentences from a training data 4 Employing Personal/Impersonal Views in Supervised Sentiment Classification After unsupervised mining of personal and impersonal sentences, the training data is divided into two views: the personal view, which contains personal sentences, and the impersonal view, which contains impersonal sentences. Obviously, these two views can be used to train two different classifiers, 1f and 2f , for sentiment classification respectively. Since our mining approach is unsupervised, there inevitably exist some noises. In addition, the sentences of different views may share the same information for sentiment classification. For example, consider the following two sentences: ‘It is a waste of money.’ and ‘Do not waste your money.’ Apparently, the first one belongs to the impersonal view while the second one belongs to personal view, according to our heuristic rules. However, these two sentences share the same word, ‘waste’, which conveys strong negative sentiment information. This suggests that training a single-view classifier 3f with all sentences should help. Therefore, three base classifiers, 1f , 2f , and 3f , are eventually derived from the personal view, the impersonal 416 view and the single view, respectively. Each base classifier provides not only the class label outputs but also some kinds of confidence measurements, e.g. posterior probabilities of the testing sample belonging to each class. Formally, each base classifier ( 1,2,3) lf l = assigns a test sample (denoted as lx ) a posterior probability vector ( ) l P x : 1 2 ( ) ( | ), ( | ) t l l l P x p c x p c x = < > where 1 ( | ) l p c x denotes the probability that the -th l base classifier considers the sample belonging to 1c . In the ensemble learning literature, various methods have been presented for combining base classifiers. The combining methods are categorized into two groups (Duin, 2002): fixed rules such as voting rule, product rule, and sum rule (Kittler et al., 1998), and trained rules such as weighted sum rule (Fumera and Roli, 2005) and meta-learning approaches (Vilalta and Drissi, 2002). In this study, we choose a fixed rule and a trained rule to combine the three base classifiers 1f , 2f , and 3f . The chosen fixed rule is product rule which combine base classifiers by multiplying the posterior possibilities and using the multiplied possibility for decision, i.e. 3 1 argmax ( | ) j i l i l assign y c where j p c x = → = ∏ The chosen trained rule is stacking (Vilalta and Drissi, 2002; Džeroski and Ženko, 2004) where a meta-classifier is trained with the output of the base classifiers as the input. Formally, let 'x denote a feature vector of a sample from the development data. The output of the -th l base classifier lf on this sample is the probability distribution over the category set 1 2 { , } c c , i.e. 1 2 ( ' ) ( | ' ), ( | ' ) l l l l P x p c x p c x =< > Then, a meta-classifier is trained using the development data with the meta-level feature vector 2 3 meta x R × ∈ 1 2 3 ( ' ), ( ' ), ( ' ) meta l l l x P x P x P x = = = =< > In our experiments, we perform stacking with 4-fold cross validation to generate meta-training data where each fold is used as the development data and the other three folds are used to train the base classifiers in the training phase. 5 Employing Personal/Impersonal Views in Semi-Supervised Sentiment Classification Semi-supervised learning is a strategy which combines unlabeled data with labeled training data to improve the models. Given the two-view classifiers 1f and 2f along with the single-view classifier 3f , we perform a co-training algorithm for semi-supervised sentiment classification. The co-training algorithm is a specific semi-supervised learning approach which starts with a set of labeled data and increases the amount of labeled data using the unlabeled data by bootstrapping (Blum and Mitchell, 1998). Figure 2 shows the co-training algorithm in our semi-supervised sentiment classification. Input: The labeled data L containing personal sentence set L personal S − and impersonal sentence set L impersonal S − The unlabeled data U containing personal sentence set U personal S − and impersonal sentence set U impersonal S − Output: New labeled data L Procedure: Loop for N iterations untilU φ = (1). Learn the first classifier 1f with L personal S − (2). Use 1f to label samples from U with U personal S − (3). Choose 1n positive and 1n negative most confidently predicted samples 1A (4). Learn the second classifier 2f with L impersonal S − (5). Use 2f to label samples from U with U impersonal S − (6). Choose 2n positive and 2n negative most confidently predicted samples 2 A (7). Learn the third classifier 3f with L (8). Use 3f to label samples from U (9). Choose 3n positive and 3n negative most confidently predicted samples 3A (10). Add samples 1 2 3 A A A ∪ ∪ with the corresponding labels into L (11). Update L personal S − and L impersonal S − Figure 2: Our co-training algorithm for semi-supervised sentiment classification 417 After obtaining the new labeled data, we can either adopt one classifier (i.e. 3f ) or a combined classifier (i.e. 1 2 3 f f f + + ) in further training and testing. In our experimentation, we explore both of them with the former referred to as co-training and single classifier and the latter referred to as co-training and combined classifier. 6 Experimental Studies We have systematically explored our method on product reviews from eight domains: book, DVD, electronic appliances, kitchen appliances, health, network, pet and software. 6.1 Experimental Setting The product reviews on the first four domains (book, DVD, electronic, and kitchen appliances) come from the multi-domain sentiment classification corpus, collected from http://www.amazon.com/ by Blitzer et al. (2007)2. Besides, we also collect the product views from http://www.amazon.com/ on other four domains (health, network, pet and software)3. Each of the eight domains contains 1000 positive and 1000 negative reviews. Figure 3 gives the distribution of personal and impersonal sentences in the training data (75% labeled data of all data). It shows that there are more impersonal sentences than personal ones in each domain, in particular in the DVD domain, where the number of impersonal sentences is at least twice as many as that of personal sentences. This unusual phenomenon is mainly attributed to the fact that many objective descriptions, e.g. the movie plot introductions, are expressed in the DVD domain which makes the extracted personal and impersonal sentences rather unbalanced. We apply both support vector machine (SVM) and Maximum Entropy (ME) algorithms with the help of the SVM-light4 and Mallet5 tools. All parameters are set to their default values. We find that ME performs slightly better than SVM on the average. Furthermore, ME offers posterior probability information which is required for 2 http://www.seas.upenn.edu/~mdredze/datasets/sentiment/ 3 Note that the second version of multi-domain sentiment classification corpus does contain data from many other domains. However, we find that the reviews in the other domains contain many duplicated samples. Therefore, we re-collect the reviews from http://www.amazon.com/ and filter those duplicated ones. The new collection is here: http://llt.cbs.polyu.edu.hk/~lss/ACL2010_Data_SSLi.zip 4 http://svmlight.joachims.org/ 5 http://mallet.cs.umass.edu/ combination methods. Thus we apply the ME classification algorithm for further combination and co-training. In particular, we only employ Boolean features, representing the presence or absence of a word in a document. Finally, we perform t-test to evaluate the significance of the performance difference between two systems with different methods (Yang and Liu, 1999). Sentence Number in the Training Data 16134 8477 8337 8843 13097 29290 14852 14414 12691 11941 13818 14265 16441 14753 15573 27714 0 10000 20000 30000 40000 Book DVD Electronic Kitchen Health Network Pet Software Number Number of personal sentences Number of impersonal sentences Figure 3: Distribution of personal and impersonal sentences in the training data of each domain 6.2 Experimental Results on Supervised Sentiment Classification 4-fold cross validation is performed for supervised sentiment classification. For comparison, we generate two random views by randomly splitting the whole feature space into two parts. Each part is seen as a view and used to train a classifier. The combination (two random view classifiers along with the single-view classifier 3f ) results are shown in the last column of Table 1. The comparison between random two views and our proposed two views will clarify whether the performance gain comes truly from our proposed two-view mining, or simply from using the classifier combination strategy. Table 1 shows the performances of different classifiers, where the single-view classifier 3f which uses all sentences for training and testing, is considered as our baseline. Note that the baseline performances of the first four domains are worse than the ones reported in Blitzer et al. (2007). But their experiment is performed with only one split on the data with 80% as the training data and 20% as the testing data, which means the size of their training data is larger than ours. Also, we find that our performances are similar to the ones (described as fully supervised results) reported in Dasgupta and Ng (2009) where the same data in the four domains are used and 10-fold cross validation is performed. 418 Domain Personal View Classifier 1f Impersonal View Classifier 2f Single View Classifier (baseline) 3f Combination (Stacking) 1 2 3 f f f + + Combination (Product rule) 1 2 3 f f f + + Combination with two random views (Product rule) Book 0.7004 0.7474 0.7654 0.7919 0.7949 0.7546 DVD 0.6931 0.7663 0.7884 0.8079 0.8165 0.8054 Electronic 0.7414 0.7844 0.8074 0.8304 0.8364 0.8210 Kitchen 0.7430 0.8030 0.8290 0.8555 0.8565 0.8152 Health 0.7000 0.7370 0.7559 0.7780 0.7815 0.7548 Network 0.7655 0.7710 0.8265 0.8360 0.8435 0.8312 Pet 0.6940 0.7145 0.7390 0.7565 0.7665 0.7423 Software 0.7035 0.7205 0.7470 0.7730 0.7715 0.7615 AVERAGE 0.7176 0.7555 0.7823 0.8037 0.8084 0.7858 Table 1: Performance of supervised sentiment classification From Table 1, we can see that impersonal view classifier 1f consistently performs better than personal view classifier 2f . Similar to the sentence distributions, the difference in the classification performances between these two views in the DVD domain is the largest (0.6931 vs. 0.7663). Both the combination methods (stacking and product rule) significantly outperform the baseline in each domain (p-value<0.01) with a decent average performance improvement of 2.61%. Although the performance difference between the product rule and stacking is not significant, the product rule is obviously a better choice as it involves much easier implementation. Therefore, in the semi-supervised learning process, we only use the product rule to combine the individual classifiers. Finally, it shows that random generation of two views with the combination method of the product rule only slightly outperforms the baseline on the average (0.7858 vs. 0.7823) but performs much worse than our unsupervised mining of personal and impersonal views. 6.3 Experimental Results on Semi-supervised Sentiment Classification We systematically evaluate and compare our two-view learning method with various semi-supervised ones as follows: Self-training, which uses the unlabeled data in a bootstrapping way like co-training yet limits the number of classifiers and the number of views to one. Only the baseline classifier 3f is used to select most confident unlabeled samples in each iteration. Transductive SVM, which seeks the largest separation between labeled and unlabeled data through regularization (Joachims, 1999). We implement it with the help of the SVM-light tool. Co-training with random two-view generation (briefly called co-training with random views), where two views are generated by randomly splitting the whole feature space into two parts. In semi-supervised sentiment classification, the data are randomly partitioned into labeled training data, unlabeled data, and testing data with the proportion of 10%, 70% and 20% respectively. Figure 4 reports the classification accuracies in all iterations, where baseline indicates the supervised classifier 3f trained on the 10% data; both co-training and single classifier and co-training and combined classifier refer to co-training using our proposed personal and impersonal views. But the former merely applies the baseline classifier 3f trained the new labeled data to test on the testing data while the latter applies the combined classifier 1 2 3 f f f + + . In each iteration, two top-confident samples in each category are chosen, i.e. 1 2 3 2 n n n = = = . For clarity, results of other methods (e.g. self-training, transductive SVM) are not shown in Figure 4 but will be reported in Figure 5 later. Figure 4 shows that co-training and combined classifier always outperforms co-training and single classifier. This again justifies the effectiveness of our two-view learning on supervised sentiment classification. 419 25 50 75 100 125 0.62 0.64 0.66 0.68 0.7 0.72 0.74 0.76 Domain: Book Iteration Number Accuracy 25 50 75 100 125 0.58 0.6 0.62 0.64 0.66 0.68 0.7 Domain: DVD Iteration Number Accuracy 25 50 75 100 125 0.7 0.72 0.74 0.76 0.78 0.8 Domain: Electronic Iteration Number Accuracy 25 50 75 100 125 0.72 0.74 0.76 0.78 0.8 0.82 Domain: Kitchen Iteration Number Accuracy 25 50 75 100 125 0.54 0.56 0.58 0.6 0.62 0.64 0.66 Domain: Health Iteration Number Accuracy 25 50 75 100 125 0.72 0.74 0.76 0.78 0.8 0.82 0.84 0.86 Domain: Network Iteration Number Accuracy Baseline Co-traning and single classifier Co-traning and combined classifier 25 50 75 100 125 0.58 0.6 0.62 0.64 0.66 0.68 Domain: Pet Iteration Number Accuracy 25 50 75 100 125 0.62 0.64 0.66 0.68 0.7 0.72 Domain: Software Iteration Number Accuracy Figure 4: Classification performance vs. iteration numbers (using 10% labeled data as training data) One open question is whether the unlabeled data improve the performance. Let us set aside the influence of the combination strategy and focus on the effectiveness of semi-supervised learning by comparing the baseline and co-training and single classifier. Figure 4 shows different results on different domains. Semi-supervised learning fails on the DVD domain while on the three domains of book, electronic, and software, semi-supervised learning benefits slightly (p-value>0.05). In contrast, semi-supervised learning benefits much on the other four domains (health, kitchen, network, and pet) from using unlabeled data and the performance improvements are statistically significant (p-value<0.01). Overall speaking, we think that the unlabeled data are very helpful as they lead to about 4% accuracy improvement on the average except for the DVD domain. Along with the supervised combination strategy, our approach can significantly improve the performance more than 7% on the average compared to the baseline. Figure 5 shows the classification results of different methods with different sizes of the labeled data: 5%, 10%, and 15% of all data, where the testing data are kept the same (20% of all data). Specifically, the results of other methods including self-training, transductive SVM, and random views are presented when 10% labeled data are used in training. It shows that self-training performs much worse than our approach and fails to improve the performance of five of the eight domains. Transductive SVM performs even worse and can only improve the performance of the “software” domain. Although co-training with random views outperforms the baseline on four of the eight domains, it performs worse than co-training and single classifier. This suggests that the impressive improvements are mainly due to our unsupervised two-view mining rather than the combination strategy. 420 Using 10% labeled data as training data 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 Book DVD Electronic Kitchen Health Network Pet Software Accuracy Baseline Transductive SVM Self-training Co-training with random views Co-training and single classifier Co-training and combined classifier Using 5% labeled data as training data 0.69 0.747 0.584 0.525 0.67 0.653 0.626 0.55 0.564 0.683 0.495 0.615 0.8675 0.7855 0.7 0.601 0.45 0.55 0.65 0.75 0.85 Book DVD Electronic Kitchen Health Network Pet Software Accuracy Using 15% labeled data as training data 0.763 0.6925 0.765 0.5925 0.679 0.564 0.677 0.7375 0.6625 0.735 0.655 0.615 0.8625 0.8325 0.782 0.716 0.45 0.55 0.65 0.75 0.85 Book DVD Electronic Kitchen Health Network Pet Software Accuracy Figure 5: Performance of semi-supervised sentiment classification when 5%, 10%, and 15% labeled data are used Figure 5 also shows that our approach is rather robust and achieves excellent performances in different training data sizes, although our approach fails on two domains, i.e. book and DVD, when only 5% of the labeled data are used. This failure may be due to that some of the samples in these two domains are too ambiguous and hard to classify. Manual checking shows that quite a lot of samples on these two domains are even too difficult for professionals to give a high-confident label. Another possible reason is that there exist too many objective descriptions in these two domains, thus introducing too much noisy information for semi-supervised learning. The effectiveness of different sizes of chosen samples in each iteration is also evaluated like 1 2 3 6 n n n = = = and 1 2 3 3, 6 n n n = = = (This assignment is considered because the personal view classifier performs worse than the other two classifiers). Our experimental results are still unsuccessful in the DVD domain and do not show much difference on other domains. We also test the co-training approach without the single-view classifier 3f . Experimental results show that the inclusion of the single-view classifier 3f slightly helps the co-training approach. The detailed discussion of the results is omitted due to space limit. 6.4 Why our approach is effective? One main reason for the effectiveness of our approach on supervised learning is the way how personal and impersonal views are dealt with. As personal and impersonal views have different ways of expressing opinions, splitting them into two separations can filter some classification noises. For example, in the sentence of “I have seen amazing dancing, and good dancing. This was TERRIBLE dancing!”. The first sentence is classified as a personal sentence and the second one is an impersonal sentence. Although the words ‘amazing’ and ‘good’ convey strong positive sentiment information, the whole text is negative. If we get the bag-of-words from the whole text, the classification result will be wrong. Rather, splitting the text into two parts based on different views allows correct classification as the personal view rarely contains impersonal words such as ‘amazing’ and ‘good’. The classification result will thus be influenced by the impersonal view. In addition, a document may contain both personal and impersonal sentences, and each of them, to a certain extent, , provides classification evidence. In fact, we randomly select 50 documents in the domain of kitchen appliances and find that 80% of the documents take both personal and impersonal sentences in which both of them express explicit opinions. That is to say, the two views provide different, complementary information for classification. This qualifies the success requirement of co-training algorithm to some extend. This might be the reason for the effectiveness of our approach on semi-supervised learning. 421 7 Discussion on Personal/Impersonal vs. Subjective/Objective As mentioned in Section 1, personal view contains X ’s “subjective” feeling, and impersonal view containsY ’s “objective” (i.e. or at least criteria-based) evaluation of the target object. However, our technically-defined concepts of personal/impersonal are definitely different from subjective/objective: Personal view can certainly contain many objective expressions, e.g. ‘I bought this electric kettle’ and impersonal view can contain many subjective expressions, e.g. ‘It is disappointing’. Our technically-defined personal/impersonal views are two different ways to describe opinions. Personal sentences are often used to express opinions in a direct way and their target object should be one of X. Impersonal ones are often used to express opinions in an indirect way and their target object should be one of Y. The ideal definition of personal (or impersonal) view given in Section 1 is believed to be a subset of our technical definition of personal (or impersonal) view. Thus impersonal view may contain both Y ’s objective evaluation (more likely to be domain independent) and subjective Y’s description. In addition, simply splitting text into subjective/objective views is not particularly helpful. Since a piece of objective text provides rather limited implicit classification information, the classification abilities of the two views are very unbalanced. This makes the co-training process unfeasible. Therefore, we believe that our technically-defined personal/impersonal views are more suitable for two-view learning compared to subjective/objective views. 8 Conclusion and Future Work In this paper, we propose a robust and effective two-view model for sentiment classification based on personal/impersonal views. Here, the personal view consists of subjective sentences whose subject is a person, whereas the impersonal view consists of objective sentences whose subject is not a person. Such views are lexically cued and can be obtained without pre-labeled data and thus we explore an unsupervised learning approach to mine them. Combination methods and a co-training algorithm are proposed to deal with supervised and semi-supervised sentiment classification respectively. Evaluation on product reviews from eight domains shows that our approach significantly improves the performance across all eight domains on supervised sentiment classification and greatly outperforms the baseline with more than 7% accuracy improvement on the average across seven of eight domains (except the DVD domain) on semi-supervised sentiment classification. In the future work, we will integrate the subjectivity summarization strategy (Pang and Lee, 2004) to help discard noisy objective sentences. Moreover, we need to consider the cases when both X and Y appear in a sentence. For example, the sentence “I think they're poor” should be an impersonal view but wrongly classified as a personal one according to our technical rules. We believe that these will help improve our approach and hopefully are applicable to the DVD domain. Another interesting and practical idea is to integrate active learning (Settles, 2009), another popular but principally different kind of semi-supervised learning approach, with our two-view learning approach to build high-performance systems with the least labeled data. Acknowledgments The research work described in this paper has been partially supported by Start-up Grant for Newly Appointed Professors, No. 1-BBZM in the Hong Kong Polytechnic University and two NSFC grants, No. 60873150 and No. 90920004. We also thank the three anonymous reviewers for their invaluable comments. References Blitzer J., M. Dredze, and F. Pereira. 2007. Biographies, Bollywood, Boom-boxes and Blenders: Domain Adaptation for Sentiment Classification. In Proceedings of ACL-07. Blum A. and T. Mitchell. 1998. Combining labeled and unlabeled data with co-training. In Proceedings of COLT-98. Crystal D. 2003. The Cambridge Encyclopedia of the English Language. Cambridge University Press. Dasgupta S. and V. Ng. 2009. Mine the Easy and Classify the Hard: Experiments with Automatic Sentiment Classification. In Proceedings of ACL-IJCNLP-09. Duin R. 2002. The Combining Classifier: To Train Or Not To Train? In Proceedings of 16th International Conference on Pattern Recognition (ICPR-02). Durant K. and M. Smith. 2007. Predicting the Political Sentiment of Web Log Posts using 422 Supervised Machine Learning Techniques Coupled with Feature Selection. In Processing of Advances in Web Mining and Web Usage Analysis. Džeroski S. and B. Ženko. 2004. Is Combining Classifiers with Stacking Better than Selecting the Best One? Machine Learning, vol.54(3), pp.255-273, 2004. Esuli A. and F. Sebastiani. 2005. Determining the Semantic Orientation of Terms through Gloss Classification. In Proceedings of CIKM-05. Fumera G. and F. Roli. 2005. A Theoretical and Experimental Analysis of Linear Combiners for Multiple Classifier Systems. IEEE Trans. PAMI, vol.27, pp.942–956, 2005 Joachims, T. 1999. Transductive Inference for Text Classification using Support Vector Machines. ICML1999. Kennedy A. and D. Inkpen. 2006. Sentiment Classification of Movie Reviews using Contextual Valence Shifters. Computational Intelligence, vol.22(2), pp.110-125, 2006. Kim S. and E. Hovy. 2004. Determining the Sentiment of Opinions. In Proceedings of COLING-04. Kittler J., M. Hatef, R. Duin, and J. Matas. 1998. On Combining Classifiers. IEEE Trans. PAMI, vol.20, pp.226-239, 1998 Liu B., M. Hu, and J. Cheng. 2005. Opinion Observer: Analyzing and Comparing Opinions on the Web. In Proceedings of WWW-05. McDonald R., K. Hannan, T. Neylon, M. Wells, and J. Reynar. 2007. Structured Models for Fine-to-coarse Sentiment Analysis. In Proceedings of ACL-07. Pang B. and L. Lee. 2004. A Sentimental Education: Sentiment Analysis using Subjectivity Summarization based on Minimum Cuts. In Proceedings of ACL-04. Pang B., L. Lee, and S. Vaithyanathan. 2002. Thumbs up? Sentiment Classification using Machine Learning Techniques. In Proceedings of EMNLP-02. Riloff E., S. Patwardhan, and J. Wiebe. 2006. Feature Subsumption for Opinion Analysis. In Proceedings of EMNLP-06. Settles B. 2009. Active Learning Literature Survey. Technical Report 1648, Department of Computer Sciences, University of Wisconsin at Madison, Wisconsin. Turney P. 2002. Thumbs Up or Thumbs Down? Semantic Orientation Applied to Unsupervised Classification of Reviews. In Proceedings of ACL-02. Vilalta R. and Y. Drissi. 2002. A Perspective View and Survey of Meta-learning. Artificial Intelligence Review, 18(2): 77–95. Wan X. 2009. Co-Training for Cross-Lingual Sentiment Classification. In Proceedings of ACL-IJCNLP-09. Wilson T., J. Wiebe, and P. Hoffmann. 2009. Recognizing Contextual Polarity: An Exploration of Features for Phrase-Level Sentiment Analysis. Computational Linguistics, vol.35(3), pp.399-433, 2009. Yang Y. and X. Liu. 1999. A Re-Examination of Text Categorization methods. In Proceedings of SIGIR-99. Zagibalov T. and J. Carroll. 2008. Automatic Seed Word Selection for Unsupervised Sentiment Classification of Chinese Test. In Proceedings of COLING-08. Zhu X. 2005. Semi-supervised Learning Literature Survey. Technical Report Computer Sciences 1530, University of Wisconsin – Madison. 423
|
2010
|
43
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 424–434, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics A Latent Dirichlet Allocation method for Selectional Preferences Alan Ritter, Mausam and Oren Etzioni Department of Computer Science and Engineering Box 352350, University of Washington, Seattle, WA 98195, USA {aritter,mausam,etzioni}@cs.washington.edu Abstract The computation of selectional preferences, the admissible argument values for a relation, is a well-known NLP task with broad applicability. We present LDA-SP, which utilizes LinkLDA (Erosheva et al., 2004) to model selectional preferences. By simultaneously inferring latent topics and topic distributions over relations, LDA-SP combines the benefits of previous approaches: like traditional classbased approaches, it produces humaninterpretable classes describing each relation’s preferences, but it is competitive with non-class-based methods in predictive power. We compare LDA-SP to several state-ofthe-art methods achieving an 85% increase in recall at 0.9 precision over mutual information (Erk, 2007). We also evaluate LDA-SP’s effectiveness at filtering improper applications of inference rules, where we show substantial improvement over Pantel et al.’s system (Pantel et al., 2007). 1 Introduction Selectional Preferences encode the set of admissible argument values for a relation. For example, locations are likely to appear in the second argument of the relation X is headquartered in Y and companies or organizations in the first. A large, high-quality database of preferences has the potential to improve the performance of a wide range of NLP tasks including semantic role labeling (Gildea and Jurafsky, 2002), pronoun resolution (Bergsma et al., 2008), textual inference (Pantel et al., 2007), word-sense disambiguation (Resnik, 1997), and many more. Therefore, much attention has been focused on automatically computing them based on a corpus of relation instances. Resnik (1996) presented the earliest work in this area, describing an information-theoretic approach that inferred selectional preferences based on the WordNet hypernym hierarchy. Recent work (Erk, 2007; Bergsma et al., 2008) has moved away from generalization to known classes, instead utilizing distributional similarity between nouns to generalize beyond observed relation-argument pairs. This avoids problems like WordNet’s poor coverage of proper nouns and is shown to improve performance. These methods, however, no longer produce the generalized class for an argument. In this paper we describe a novel approach to computing selectional preferences by making use of unsupervised topic models. Our approach is able to combine benefits of both kinds of methods: it retains the generalization and humaninterpretability of class-based approaches and is also competitive with the direct methods on predictive tasks. Unsupervised topic models, such as latent Dirichlet allocation (LDA) (Blei et al., 2003) and its variants are characterized by a set of hidden topics, which represent the underlying semantic structure of a document collection. For our problem these topics offer an intuitive interpretation – they represent the (latent) set of classes that store the preferences for the different relations. Thus, topic models are a natural fit for modeling our relation data. In particular, our system, called LDA-SP, uses LinkLDA (Erosheva et al., 2004), an extension of LDA that simultaneously models two sets of distributions for each topic. These two sets represent the two arguments for the relations. Thus, LDA-SP is able to capture information about the pairs of topics that commonly co-occur. This information is very helpful in guiding inference. We run LDA-SP to compute preferences on a massive dataset of binary relations r(a1, a2) ex424 tracted from the Web by TEXTRUNNER (Banko and Etzioni, 2008). Our experiments demonstrate that LDA-SP significantly outperforms state of the art approaches obtaining an 85% increase in recall at precision 0.9 on the standard pseudodisambiguation task. Additionally, because LDA-SP is based on a formal probabilistic model, it has the advantage that it can naturally be applied in many scenarios. For example, we can obtain a better understanding of similar relations (Table 1), filter out incorrect inferences based on querying our model (Section 4.3), as well as produce a repository of class-based preferences with a little manual effort as demonstrated in Section 4.4. In all these cases we obtain high quality results, for example, massively outperforming Pantel et al.’s approach in the textual inference task.1 2 Previous Work Previous work on selectional preferences can be broken into four categories: class-based approaches (Resnik, 1996; Li and Abe, 1998; Clark and Weir, 2002; Pantel et al., 2007), similarity based approaches (Dagan et al., 1999; Erk, 2007), discriminative (Bergsma et al., 2008), and generative probabilistic models (Rooth et al., 1999). Class-based approaches, first proposed by Resnik (1996), are the most studied of the four. They make use of a pre-defined set of classes, either manually produced (e.g. WordNet), or automatically generated (Pantel, 2003). For each relation, some measure of the overlap between the classes and observed arguments is used to identify those that best describe the arguments. These techniques produce a human-interpretable output, but often suffer in quality due to an incoherent taxonomy, inability to map arguments to a class (poor lexical coverage), and word sense ambiguity. Because of these limitations researchers have investigated non-class based approaches, which attempt to directly classify a given noun-phrase as plausible/implausible for a relation. Of these, the similarity based approaches make use of a distributional similarity measure between arguments and evaluate a heuristic scoring function: Srel(arg)= X arg′∈Seen(rel) sim(arg, arg′) · wtrel(arg) 1Our repository of selectional preferences is available at http://www.cs.washington.edu/research/ ldasp. Erk (2007) showed the advantages of this approach over Resnik’s information-theoretic classbased method on a pseudo-disambiguation evaluation. These methods obtain better lexical coverage, but are unable to obtain any abstract representation of selectional preferences. Our solution fits into the general category of generative probabilistic models, which model each relation/argument combination as being generated by a latent class variable. These classes are automatically learned from the data. This retains the class-based flavor of the problem, without the knowledge limitations of the explicit classbased approaches. Probably the closest to our work is a model proposed by Rooth et al. (1999), in which each class corresponds to a multinomial over relations and arguments and EM is used to learn the parameters of the model. In contrast, we use a LinkLDA framework in which each relation is associated with a corresponding multinomial distribution over classes, and each argument is drawn from a class-specific distribution over words; LinkLDA captures co-occurrence of classes in the two arguments. Additionally we perform full Bayesian inference using collapsed Gibbs sampling, in which parameters are integrated out (Griffiths and Steyvers, 2004). Recently, Bergsma et. al. (2008) proposed the first discriminative approach to selectional preferences. Their insight that pseudo-negative examples could be used as training data allows the application of an SVM classifier, which makes use of many features in addition to the relation-argument co-occurrence frequencies used by other methods. They automatically generated positive and negative examples by selecting arguments having high and low mutual information with the relation. Since it is a discriminative approach it is amenable to feature engineering, but needs to be retrained and tuned for each task. On the other hand, generative models produce complete probability distributions of the data, and hence can be integrated with other systems and tasks in a more principled manner (see Sections 4.2.2 and 4.3.1). Additionally, unlike LDA-SP Bergsma et al.’s system doesn’t produce human-interpretable topics. Finally, we note that LDA-SP and Bergsma’s system are potentially complimentary – the output of LDA-SP could be used to generate higher-quality training data for Bergsma, potentially improving their results. 425 Topic models such as LDA (Blei et al., 2003) and its variants have recently begun to see use in many NLP applications such as summarization (Daum´e III and Marcu, 2006), document alignment and segmentation (Chen et al., 2009), and inferring class-attribute hierarchies (Reisinger and Pasca, 2009). Our particular model, LinkLDA, has been applied to a few NLP tasks such as simultaneously modeling the words appearing in blog posts and users who will likely respond to them (Yano et al., 2009), modeling topic-aligned articles in different languages (Mimno et al., 2009), and word sense induction (Brody and Lapata, 2009). Finally, we highlight two systems, developed independently of our own, which apply LDA-style models to similar tasks. ´O S´eaghdha (2010) proposes a series of LDA-style models for the task of computing selectional preferences. This work learns selectional preferences between the following grammatical relations: verb-object, nounnoun, and adjective-noun. It also focuses on jointly modeling the generation of both predicate and argument, and evaluation is performed on a set of human-plausibility judgments obtaining impressive results against Keller and Lapata’s (2003) Web hit-count based system. Van Durme and Gildea (2009) proposed applying LDA to general knowledge templates extracted using the KNEXT system (Schubert and Tong, 2003). In contrast, our work uses LinkLDA and focuses on modeling multiple arguments of a relation (e.g., the subject and direct object of a verb). 3 Topic Models for Selectional Prefs. We present a series of topic models for the task of computing selectional preferences. These models vary in the amount of independence they assume between a1 and a2. At one extreme is IndependentLDA, a model which assumes that both a1 and a2 are generated completely independently. On the other hand, JointLDA, the model at the other extreme (Figure 1) assumes both arguments of a specific extraction are generated based on a single hidden variable z. LinkLDA (Figure 2) lies between these two extremes, and as demonstrated in Section 4, it is the best model for our relation data. We are given a set R of binary relations and a corpus D = {r(a1, a2)} of extracted instances for these relations. 2 Our task is to compute, for each argument ai of each relation r, a set of usual argument values (noun phrases) that it takes. For example, for the relation is headquartered in the first argument set will include companies like Microsoft, Intel, General Motors and second argument will favor locations like New York, California, Seattle. 3.1 IndependentLDA We first describe the straightforward application of LDA to modeling our corpus of extracted relations. In this case two separate LDA models are used to model a1 and a2 independently. In the generative model for our data, each relation r has a corresponding multinomial over topics θr, drawn from a Dirichlet. For each extraction, a hidden topic z is first picked according to θr, and then the observed argument a is chosen according to the multinomial βz. Readers familiar with topic modeling terminology can understand our approach as follows: we treat each relation as a document whose contents consist of a bags of words corresponding to all the noun phrases observed as arguments of the relation in our corpus. Formally, LDA generates each argument in the corpus of relations as follows: for each topic t = 1 . . . T do Generate βt according to symmetric Dirichlet distribution Dir(η). end for for each relation r = 1 . . . |R| do Generate θr according to Dirichlet distribution Dir(α). for each tuple i = 1 . . . Nr do Generate zr,i from Multinomial(θr). Generate the argument ar,i from multinomial βzr,i. end for end for One weakness of IndependentLDA is that it doesn’t jointly model a1 and a2 together. Clearly this is undesirable, as information about which topics one of the arguments favors can help inform the topics chosen for the other. For example, class pairs such as (team, game), (politician, political issue) form much more plausible selectional preferences than, say, (team, political issue), (politician, game). 2We focus on binary relations, though the techniques presented in the paper are easily extensible to n-ary relations. 426 3.2 JointLDA As a more tightly coupled alternative, we first propose JointLDA, whose graphical model is depicted in Figure 1. The key difference in JointLDA (versus LDA) is that instead of one, it maintains two sets of topics (latent distributions over words) denoted by β and γ, one for classes of each argument. A topic id k represents a pair of topics, βk and γk, that co-occur in the arguments of extracted relations. Common examples include (Person, Location), (Politician, Political issue), etc. The hidden variable z = k indicates that the noun phrase for the first argument was drawn from the multinomial βk, and that the second argument was drawn from γk. The per-relation distribution θr is a multinomial over the topic ids and represents the selectional preferences, both for arg1s and arg2s of a relation r. Although JointLDA has many desirable properties, it has some drawbacks as well. Most notably, in JointLDA topics correspond to pairs of multinomials (βk, γk); this leads to a situation in which multiple redundant distributions are needed to represent the same underlying semantic class. For example consider the case where we we need to represent the following selectional preferences for our corpus of relations: (person, location), (person, organization), and (person, crime). Because JointLDA requires a separate pair of multinomials for each topic, it is forced to use 3 separate multinomials to represent the class person, rather than learning a single distribution representing person and choosing 3 different topics for a2. This results in poor generalization because the data for a single class is divided into multiple topics. In order to address this problem while maintaining the sharing of influence between a1 and a2, we next present LinkLDA, which represents a compromise between IndependentLDA and JointLDA. LinkLDA is more flexible than JointLDA, allowing different topics to be chosen for a1, and a2, however still models the generation of topics from the same distribution for a given relation. 3.3 LinkLDA Figure 2 illustrates the LinkLDA model in the plate notation, which is analogous to the model in (Erosheva et al., 2004). In particular note that each ai is drawn from a different hidden topic zi, however the zi’s are drawn from the same distribution θr for a given relation r. To facilitate learnθ a1 a2 β |R| N α η1 γ T η2 z Figure 1: JointLDA θ z1 z2 a1 a2 β |R| N α η1 γ T η2 Figure 2: LinkLDA ing related topic pairs between arguments we employ a sparse prior over the per-relation topic distributions. Because a few topics are likely to be assigned most of the probability mass for a given relation it is more likely (although not necessary) that the same topic number k will be drawn for both arguments. When comparing LinkLDA with JointLDA the better model may not seem immediately clear. On the one hand, JointLDA jointly models the generation of both arguments in an extracted tuple. This allows one argument to help disambiguate the other in the case of ambiguous relation strings. LinkLDA, however, is more flexible; rather than requiring both arguments to be generated from one of |Z| possible pairs of multinomials (βz, γz), LinkLDA allows the arguments of a given extraction to be generated from |Z|2 possible pairs. Thus, instead of imposing a hard constraint that z1 = z2 (as in JointLDA), LinkLDA simply assigns a higher probability to states in which z1 = z2, because both hidden variables are drawn from the same (sparse) distribution θr. LinkLDA can thus re-use argument classes, choosing different combinations of topics for the arguments if it fits the data better. In Section 4 we show experimentally that LinkLDA outperforms JointLDA (and IndependentLDA) by wide margins. We use LDA-SP to refer to LinkLDA in all the experiments below. 3.4 Inference For all the models we use collapsed Gibbs sampling for inference in which each of the hidden variables (e.g., zr,i,1 and zr,i,2 in LinkLDA) are sampled sequentially conditioned on a fullassignment to all others, integrating out the parameters (Griffiths and Steyvers, 2004). This produces robust parameter estimates, as it allows computation of expectations over the posterior distribution 427 as opposed to estimating maximum likelihood parameters. In addition, the integration allows the use of sparse priors, which are typically more appropriate for natural language data. In all experiments we use hyperparameters α = η1 = η2 = 0.1. We generated initial code for our samplers using the Hierarchical Bayes Compiler (Daume III, 2007). 3.5 Advantages of Topic Models There are several advantages to using topic models for our task. First, they naturally model the class-based nature of selectional preferences, but don’t take a pre-defined set of classes as input. Instead, they compute the classes automatically. This leads to better lexical coverage since the issue of matching a new argument to a known class is side-stepped. Second, the models naturally handle ambiguous arguments, as they are able to assign different topics to the same phrase in different contexts. Inference in these models is also scalable – linear in both the size of the corpus as well as the number of topics. In addition, there are several scalability enhancements such as SparseLDA (Yao et al., 2009), and an approximation of the Gibbs Sampling procedure can be efficiently parallelized (Newman et al., 2009). Finally we note that, once a topic distribution has been learned over a set of training relations, one can efficiently apply inference to unseen relations (Yao et al., 2009). 4 Experiments We perform three main experiments to assess the quality of the preferences obtained using topic models. The first is a task-independent evaluation using a pseudo-disambiguation experiment (Section 4.2), which is a standard way to evaluate the quality of selectional preferences (Rooth et al., 1999; Erk, 2007; Bergsma et al., 2008). We use this experiment to compare the various topic models as well as the best model with the known state of the art approaches to selectional preferences. Secondly, we show significant improvements to performance at an end-task of textual inference in Section 4.3. Finally, we report on the quality of a large database of Wordnet-based preferences obtained after manually associating our topics with Wordnet classes (Section 4.4). 4.1 Generalization Corpus For all experiments we make use of a corpus of r(a1, a2) tuples, which was automatically extracted by TEXTRUNNER (Banko and Etzioni, 2008) from 500 million Web pages. To create a generalization corpus from this large dataset. We first selected 3,000 relations from the middle of the tail (we used the 2,0005,000 most frequent ones)3 and collected all instances. To reduce sparsity, we discarded all tuples containing an NP that occurred fewer than 50 times in the data. This resulted in a vocabulary of about 32,000 noun phrases, and a set of about 2.4 million tuples in our generalization corpus. We inferred topic-argument and relation-topic multinomials (β, γ, and θ) on the generalization corpus by taking 5 samples at a lag of 50 after a burn in of 750 iterations. Using multiple samples introduces the risk of topic drift due to lack of identifiability, however we found this to not be a problem in practice. During development we found that the topics tend to remain stable across multiple samples after sufficient burn in, and multiple samples improved performance. Table 1 lists sample topics and high ranked words for each (for both arguments) as well as relations favoring those topics. 4.2 Task Independent Evaluation We first compare the three LDA-based approaches to each other and two state of the art similarity based systems (Erk, 2007) (using mutual information and Jaccard similarity respectively). These similarity measures were shown to outperform the generative model of Rooth et al. (1999), as well as class-based methods such as Resnik’s. In this pseudo-disambiguation experiment an observed tuple is paired with a pseudo-negative, which has both arguments randomly generated from the whole vocabulary (according to the corpus-wide distribution over arguments). The task is, for each relation-argument pair, to determine whether it is observed, or a random distractor. 4.2.1 Test Set For this experiment we gathered a primary corpus by first randomly selecting 100 high-frequency relations not in the generalization corpus. For each relation we collected all tuples containing arguments in the vocabulary. We held out 500 randomly selected tuples as the test set. For each tu3Many of the most frequent relations have very weak selectional preferences, and thus provide little signal for inferring meaningful topics. For example, the relations has and is can take just about any arguments. 428 Topic t Arg1 Relations which assign highest probability to t Arg2 18 The residue - The mixture - The reaction mixture - The solution - the mixture - the reaction mixture - the residue - The reaction the solution - The filtrate - the reaction - The product - The crude product - The pellet The organic layer - Thereto - This solution - The resulting solution - Next - The organic phase - The resulting mixture - C. ) was treated with, is treated with, was poured into, was extracted with, was purified by, was diluted with, was filtered through, is disolved in, is washed with EtOAc - CH2Cl2 - H2O - CH.sub.2Cl.sub.2 - H.sub.2O - water - MeOH - NaHCO3 Et2O - NHCl - CHCl.sub.3 - NHCl - dropwise - CH2Cl.sub.2 - Celite - Et.sub.2O Cl.sub.2 - NaOH - AcOEt - CH2C12 - the mixture - saturated NaHCO3 - SiO2 - H2O - N hydrochloric acid - NHCl - preparative HPLC - to0 C 151 the Court - The Court - the Supreme Court - The Supreme Court - this Court - Court - The US Supreme Court - the court - This Court - the US Supreme Court - The court - Supreme Court - Judge - the Court of Appeals - A federal judge will hear, ruled in, decides, upholds, struck down, overturned, sided with, affirms the case - the appeal - arguments - a case evidence - this case - the decision - the law - testimony - the State - an interview - an appeal - cases - the Court - that decision Congress - a decision - the complaint - oral arguments - a law - the statute 211 President Bush - Bush - The President Clinton - the President - President Clinton - President George W. Bush - Mr. Bush The Governor - the Governor - Romney McCain - The White House - President Schwarzenegger - Obama hailed, vetoed, promoted, will deliver, favors, denounced, defended the bill - a bill - the decision - the war - the idea - the plan - the move - the legislation legislation - the measure - the proposal - the deal - this bill - a measure - the program the law - the resolution - efforts - the agreement - gay marriage - the report - abortion 224 Google - Software - the CPU - Clicking Excel - the user - Firefox - System - The CPU - Internet Explorer - the ability - Program - users - Option - SQL Server - Code - the OS - the BIOS will display, to store, to load, processes, cannot find, invokes, to search for, to delete data - files - the data - the file - the URL information - the files - images - a URL - the information - the IP address - the user - text - the code - a file - the page - IP addresses PDF files - messages - pages - an IP address Table 1: Example argument lists from the inferred topics. For each topic number t we list the most probable values according to the multinomial distributions for each argument (βt and γt). The middle column reports a few relations whose inferred topic distributions θr assign highest probability to t. ple r(a1, a2) in the held-out set, we removed all tuples in the training set containing either of the rel-arg pairs, i.e., any tuple matching r(a1, ∗) or r(∗, a2). Next we used collapsed Gibbs sampling to infer a distribution over topics, θr, for each of the relations in the primary corpus (based solely on tuples in the training set) using the topics from the generalization corpus. For each of the 500 observed tuples in the testset we generated a pseudo-negative tuple by randomly sampling two noun phrases from the distribution of NPs in both corpora. 4.2.2 Prediction Our prediction system needs to determine whether a specific relation-argument pair is admissible according to the selectional preferences or is a random distractor (D). Following previous work, we perform this experiment independently for the two relation-argument pairs (r, a1) and (r, a2). We first compute the probability of observing a1 for first argument of relation r given that it is not a distractor, P(a1|r, ¬D), which we approximate by its probability given an estimate of the parameters inferred by our model, marginalizing over hidden topics t. The analysis for the second argument is similar. P(a1|r, ¬D) ≈PLDA(a1|r) = T X t=0 P(a1|t)P(t|r) = T X t=0 βt(a1)θr(t) A simple application of Bayes Rule gives the probability that a particular argument is not a distractor. Here the distractor-related probabilities are independent of r, i.e., P(D|r) = P(D), P(a1|D, r) = P(a1|D), etc. We estimate P(a1|D) according to their frequency in the generalization corpus. P(¬D|r, a1) = P(¬D|r)P(a1|r, ¬D) P(a1|r) ≈ P(¬D)PLDA(a1|r) P(D)P(a1|D) + P(¬D)PLDA(a1|r) 4.2.3 Results Figure 3 plots the precision-recall curve for the pseudo-disambiguation experiment comparing the three different topic models. LDA-SP, which uses LinkLDA, substantially outperforms both IndependentLDA and JointLDA. Next, in figure 4, we compare LDA-SP with mutual information and Jaccard similarities using both the generalization and primary corpus for 429 0.0 0.2 0.4 0.6 0.8 1.0 0.4 0.6 0.8 1.0 recall precision LDA−SP IndependentLDA JointLDA Figure 3: Comparison of LDA-based approaches on the pseudo-disambiguation task. LDA-SP (LinkLDA) substantially outperforms the other models. 0.0 0.2 0.4 0.6 0.8 1.0 0.4 0.6 0.8 1.0 recall precision LDA−SP Jaccard Mutual Information Figure 4: Comparison to similarity-based selectional preference systems. LDA-SP obtains 85% higher recall at precision 0.9. computation of similarities. We find LDA-SP significantly outperforms these methods. Its edge is most noticed at high precisions; it obtains 85% more recall at 0.9 precision compared to mutual information. Overall LDA-SP obtains an 15% increase in the area under precision-recall curve over mutual information. All three systems’ AUCs are shown in Table 2; LDA-SP’s improvements over both Jaccard and mutual information are highly significant with a significance level less than 0.01 using a paired t-test. In addition to a superior performance in selectional preference evaluation LDA-SP also produces a set of coherent topics, which can be useful in their own right. For instance, one could use them for tasks such as set-expansion (Carlson et al., 2010) or automatic thesaurus induction (EtLDA-SP MI-Sim Jaccard-Sim AUC 0.833 0.727 0.711 Table 2: Area under the precision recall curve. LDA-SP’s AUC is significantly higher than both similarity-based methods according to a paired ttest with a significance level below 0.01. zioni et al., 2005; Kozareva et al., 2008). 4.3 End Task Evaluation We now evaluate LDA-SP’s ability to improve performance at an end-task. We choose the task of improving textual entailment by learning selectional preferences for inference rules and filtering inferences that do not respect these. This application of selectional preferences was introduced by Pantel et. al. (2007). For now we stick to inference rules of the form r1(a1, a2) ⇒r2(a1, a2), though our ideas are more generally applicable to more complex rules. As an example, the rule (X defeats Y) ⇒(X plays Y) holds when X and Y are both sports teams, however fails to produce a reasonable inference if X and Y are Britain and Nazi Germany respectively. 4.3.1 Filtering Inferences In order for an inference to be plausible, both relations must have similar selectional preferences, and further, the arguments must obey the selectional preferences of both the antecedent r1 and the consequent r2.4 Pantel et al. (2007) made use of these intuitions by producing a set of classbased selectional preferences for each relation, then filtering out any inferences where the arguments were incompatible with the intersection of these preferences. In contrast, we take a probabilistic approach, evaluating the quality of a specific inference by measuring the probability that the arguments in both the antecedent and the consequent were drawn from the same hidden topic in our model. Note that this probability captures both the requirement that the antecedent and consequent have similar selectional preferences, and that the arguments from a particular instance of the rule’s application match their overlap. We use zri,j to denote the topic that generates the jth argument of relation ri. The probability that the two arguments a1, a2 were drawn from the same hidden topic factorizes as follows due to the conditional independences in our model:5 P(zr1,1 = zr2,1, zr1,2 = zr2,2|a1, a2) = P(zr1,1 = zr2,1|a1)P(zr1,2 = zr2,2|a2) 4Similarity-based and discriminative methods are not applicable to this task as they offer no straightforward way to compare the similarity between selectional preferences of two relations. 5Note that all probabilities are conditioned on an estimate of the parameters θ, β, γ from our model, which are omitted for compactness. 430 To compute each of these factors we simply marginalize over the hidden topics: P(zr1,j = zr2,j|aj) = T X t=1 P(zr1,j = t|aj)P(zr2,j = t|aj) where P(z = t|a) can be computed using Bayes rule. For example, P(zr1,1 = t|a1) = P(a1|zr1,1 = t)P(zr1,1 = t) P(a1) = βt(a1)θr1(t) P(a1) 4.3.2 Experimental Conditions In order to evaluate LDA-SP’s ability to filter inferences based on selectional preferences we need a set of inference rules between the relations in our corpus. We therefore mapped the DIRT Inference rules (Lin and Pantel, 2001), (which consist of pairs of dependency paths) to TEXTRUNNER relations as follows. We first gathered all instances in the generalization corpus, and for each r(a1, a2) created a corresponding simple sentence by concatenating the arguments with the relation string between them. Each such simple sentence was parsed using Minipar (Lin, 1998). From the parses we extracted all dependency paths between nouns that contain only words present in the TEXTRUNNER relation string. These dependency paths were then matched against each pair in the DIRT database, and all pairs of associated relations were collected producing about 26,000 inference rules. Following Pantel et al. (2007) we randomly sampled 100 inference rules. We then automatically filtered out any rules which contained a negation, or for which the antecedent and consequent contained a pair of antonyms found in WordNet (this left us with 85 rules). For each rule we collected 10 random instances of the antecedent, and generated the consequent. We randomly sampled 300 of these inferences to hand-label. 4.3.3 Results In figure 5 we compare the precision and recall of LDA-SP against the top two performing systems described by Pantel et al. (ISP.IIM-∨and ISP.JIM, both using the CBC clusters (Pantel, 2003)). We find that LDA-SP achieves both higher precision and recall than ISP.IIM-∨. It is also able to achieve the high-precision point of ISP.JIM and can trade precision to get a much larger recall. 0.0 0.2 0.4 0.6 0.8 1.0 0.4 0.6 0.8 1.0 recall precision X O X O LDA−SP ISP.JIM ISP.IIM−OR Figure 5: Precision and recall on the inference filtering task. Top 10 Inference Rules Ranked by LDA-SP antecedent consequent KL-div will begin at will start at 0.014999 shall review shall determine 0.129434 may increase may reduce 0.214841 walk from walk to 0.219471 consume absorb 0.240730 shall keep shall maintain 0.264299 shall pay to will notify 0.290555 may apply for may obtain 0.313916 copy download 0.316502 should pay must pay 0.371544 Bottom 10 Inference Rules Ranked by LDA-SP antecedent consequent KL-div lose to shall take 10.011848 should play could do 10.028904 could play get in 10.048857 will start at move to 10.060994 shall keep will spend 10.105493 should play get in 10.131299 shall pay to leave for 10.131364 shall keep return to 10.149797 shall keep could do 10.178032 shall maintain have spent 10.221618 Table 3: Top 10 and Bottom 10 ranked inference rules ranked by LDA-SPafter automatically filtering out negations and antonyms (using WordNet). In addition we demonstrate LDA-SP’s ability to rank inference rules by measuring the Kullback Leibler Divergence6 between the topicdistributions of the antecedent and consequent, θr1 and θr2 respectively. Table 3 shows the top 10 and bottom 10 rules out of the 26,000 ranked by KL Divergence after automatically filtering antonyms (using WordNet) and negations. For slight variations in rules (e.g., symmetric pairs) we mention only one example to show more variety. 6KL-Divergence is an information-theoretic measure of the similarity between two probability distributions, and defined as follows: KL(P||Q) = P x P(x) log P (x) Q(x). 431 4.4 A Repository of Class-Based Preferences Finally we explore LDA-SP’s ability to produce a repository of human interpretable class-based selectional preferences. As an example, for the relation was born in, we would like to infer that the plausible arguments include (person, location) and (person, date). Since we already have a set of topics, our task reduces to mapping the inferred topics to an equivalent class in a taxonomy (e.g., WordNet). We experimented with automatic methods such as Resnik’s, but found them to have all the same problems as directly applying these approaches to the SP task.7 Guided by the fact that we have a relatively small number of topics (600 total, 300 for each argument) we simply chose to label them manually. By labeling this small number of topics we can infer class-based preferences for an arbitrary number of relations. In particular, we applied a semi-automatic scheme to map topics to WordNet. We first applied Resnik’s approach to automatically shortlist a few candidate WordNet classes for each topic. We then manually picked the best class from the shortlist that best represented the 20 top arguments for a topic (similar to Table 1). We marked all incoherent topics with a special symbol ∅. This process took one of the authors about 4 hours to complete. To evaluate how well our topic-class associations carry over to unseen relations we used the same random sample of 100 relations from the pseudo-disambiguation experiment.8 For each argument of each relation we picked the top two topics according to frequency in the 5 Gibbs samples. We then discarded any topics which were labeled with ∅; this resulted in a set of 236 predictions. A few examples are displayed in table 4. We evaluated these classes and found the accuracy to be around 0.88. We contrast this with Pantel’s repository,9 the only other released database of selectional preferences to our knowledge. We evaluated the same 100 relations from his website and tagged the top 2 classes for each argument and evaluated the accuracy to be roughly 0.55. 7Perhaps recent work on automatic coherence ranking (Newman et al., 2010) and labeling (Mei et al., 2007) could produce better results. 8Recall that these 100 were not part of the original 3,000 in the generalization corpus, and are, therefore, representative of new “unseen” relations. 9http://demo.patrickpantel.com/ Content/LexSem/paraphrase.htm arg1 class relation arg2 class politician#1 was running for leader#1 people#1 will love show#3 organization#1 has responded to accusation#2 administrative unit#1 has appointed administrator#3 Table 4: Class-based Selectional Preferences. We emphasize that tagging a pair of class-based preferences is a highly subjective task, so these results should be treated as preliminary. Still, these early results are promising. We wish to undertake a larger scale study soon. 5 Conclusions and Future Work We have presented an application of topic modeling to the problem of automatically computing selectional preferences. Our method, LDA-SP, learns a distribution over topics for each relation while simultaneously grouping related words into these topics. This approach is capable of producing human interpretable classes, however, avoids the drawbacks of traditional class-based approaches (poor lexical coverage and ambiguity). LDA-SP achieves state-of-the-art performance on predictive tasks such as pseudo-disambiguation, and filtering incorrect inferences. Because LDA-SP generates a complete probabilistic model for our relation data, its results are easily applicable to many other tasks such as identifying similar relations, ranking inference rules, etc. In the future, we wish to apply our model to automatically discover new inference rules and paraphrases. Finally, our repository of selectional preferences for 10,000 relations is available at http://www.cs.washington.edu/ research/ldasp. Acknowledgments We would like to thank Tim Baldwin, Colin Cherry, Jesse Davis, Elena Erosheva, Stephen Soderland, Dan Weld, in addition to the anonymous reviewers for helpful comments on a previous draft. This research was supported in part by NSF grant IIS-0803481, ONR grant N00014-081-0431, DARPA contract FA8750-09-C-0179, a National Defense Science and Engineering Graduate (NDSEG) Fellowship 32 CFR 168a, and carried out at the University of Washington’s Turing Center. 432 References Michele Banko and Oren Etzioni. 2008. The tradeoffs between open and traditional relation extraction. In ACL-08: HLT. Shane Bergsma, Dekang Lin, and Randy Goebel. 2008. Discriminative learning of selectional preference from unlabeled text. In EMNLP. David M. Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent dirichlet allocation. J. Mach. Learn. Res. Samuel Brody and Mirella Lapata. 2009. Bayesian word sense induction. In EACL, pages 103–111, Morristown, NJ, USA. Association for Computational Linguistics. Andrew Carlson, Justin Betteridge, Richard C. Wang, Estevam R. Hruschka Jr., and Tom M. Mitchell. 2010. Coupled semi-supervised learning for information extraction. In WSDM 2010. Harr Chen, S. R. K. Branavan, Regina Barzilay, and David R. Karger. 2009. Global models of document structure using latent permutations. In NAACL. Stephen Clark and David Weir. 2002. Class-based probability estimation using a semantic hierarchy. Comput. Linguist. Ido Dagan, Lillian Lee, and Fernando C. N. Pereira. 1999. Similarity-based models of word cooccurrence probabilities. In Machine Learning. Hal Daum´e III and Daniel Marcu. 2006. Bayesian query-focused summarization. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics. Hal Daume III. 2007. hbc: Hierarchical bayes compiler. http://hal3.name/hbc. Katrin Erk. 2007. A simple, similarity-based model for selectional preferences. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics. Elena Erosheva, Stephen Fienberg, and John Lafferty. 2004. Mixed-membership models of scientific publications. Proceedings of the National Academy of Sciences of the United States of America. Oren Etzioni, Michael Cafarella, Doug Downey, Ana maria Popescu, Tal Shaked, Stephen Soderl, Daniel S. Weld, and Alex Yates. 2005. Unsupervised named-entity extraction from the web: An experimental study. Artificial Intelligence. Daniel Gildea and Daniel Jurafsky. 2002. Automatic labeling of semantic roles. Comput. Linguist. T. L. Griffiths and M. Steyvers. 2004. Finding scientific topics. Proc Natl Acad Sci U S A. Frank Keller and Mirella Lapata. 2003. Using the web to obtain frequencies for unseen bigrams. Comput. Linguist. Zornitsa Kozareva, Ellen Riloff, and Eduard Hovy. 2008. Semantic class learning from the web with hyponym pattern linkage graphs. In ACL-08: HLT. Hang Li and Naoki Abe. 1998. Generalizing case frames using a thesaurus and the mdl principle. Comput. Linguist. Dekang Lin and Patrick Pantel. 2001. Dirt-discovery of inference rules from text. In KDD. Dekang Lin. 1998. Dependency-based evaluation of minipar. In Proc. Workshop on the Evaluation of Parsing Systems. Qiaozhu Mei, Xuehua Shen, and ChengXiang Zhai. 2007. Automatic labeling of multinomial topic models. In KDD. David Mimno, Hanna M. Wallach, Jason Naradowsky, David A. Smith, and Andrew McCallum. 2009. Polylingual topic models. In EMNLP. David Newman, Arthur Asuncion, Padhraic Smyth, and Max Welling. 2009. Distributed algorithms for topic models. JMLR. David Newman, Jey Han Lau, Karl Grieser, and Timothy Baldwin. 2010. Automatic evaluation of topic coherence. In NAACL-HLT. Diarmuid ´O S´eaghdha. 2010. Latent variable models of selectional preference. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics. Patrick Pantel, Rahul Bhagat, Bonaventura Coppola, Timothy Chklovski, and Eduard H. Hovy. 2007. Isp: Learning inferential selectional preferences. In HLT-NAACL. Patrick Andre Pantel. 2003. Clustering by committee. Ph.D. thesis, University of Alberta, Edmonton, Alta., Canada. Joseph Reisinger and Marius Pasca. 2009. Latent variable models of concept-attribute attachment. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP. P. Resnik. 1996. Selectional constraints: an information-theoretic model and its computational realization. Cognition. Philip Resnik. 1997. Selectional preference and sense disambiguation. In Proc. of the ACL SIGLEX Workshop on Tagging Text with Lexical Semantics: Why, What, and How? 433 Mats Rooth, Stefan Riezler, Detlef Prescher, Glenn Carroll, and Franz Beil. 1999. Inducing a semantically annotated lexicon via em-based clustering. In Proceedings of the 37th annual meeting of the Association for Computational Linguistics on Computational Linguistics. Lenhart Schubert and Matthew Tong. 2003. Extracting and evaluating general world knowledge from the brown corpus. In In Proc. of the HLT-NAACL Workshop on Text Meaning, pages 7–13. Benjamin Van Durme and Daniel Gildea. 2009. Topic models for corpus-centric knowledge generalization. In Technical Report TR-946, Department of Computer Science, University of Rochester, Rochester. Tae Yano, William W. Cohen, and Noah A. Smith. 2009. Predicting response to political blog posts with topic models. In NAACL. L. Yao, D. Mimno, and A. Mccallum. 2009. Efficient methods for topic model inference on streaming document collections. In KDD. 434
|
2010
|
44
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 435–444, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Latent variable models of selectional preference Diarmuid ´O S´eaghdha University of Cambridge Computer Laboratory United Kingdom [email protected] Abstract This paper describes the application of so-called topic models to selectional preference induction. Three models related to Latent Dirichlet Allocation, a proven method for modelling document-word cooccurrences, are presented and evaluated on datasets of human plausibility judgements. Compared to previously proposed techniques, these models perform very competitively, especially for infrequent predicate-argument combinations where they exceed the quality of Web-scale predictions while using relatively little data. 1 Introduction Language researchers have long been aware that many words place semantic restrictions on the words with which they can co-occur in a syntactic relationship. Violations of these restrictions make the sense of a sentence odd or implausible: (1) Colourless green ideas sleep furiously. (2) The deer shot the hunter. Recognising whether or not a selectional restriction is satisfied can be an important trigger for metaphorical interpretations (Wilks, 1978) and also plays a role in the time course of human sentence processing (Rayner et al., 2004). A more relaxed notion of selectional preference captures the idea that certain classes of entities are more likely than others to fill a given argument slot of a predicate. In Natural Language Processing, knowledge about probable, less probable and wholly infelicitous predicateargument pairs is of value for numerous applications, for example semantic role labelling (Gildea and Jurafsky, 2002; Zapirain et al., 2009). The notion of selectional preference is not restricted to surface-level predicates such as verbs and modifiers, but also extends to semantic frames (Erk, 2007) and inference rules (Pantel et al., 2007). The fundamental problem that selectional preference models must address is data sparsity: in many cases insufficient corpus data is available to reliably measure the plausibility of a predicate-argument pair by counting its observed frequency. A rarely seen pair may be fundamentally implausible (a carrot laughed) or plausible but rarely expressed (a manservant laughed).1 In general, it is beneficial to smooth plausibility estimates by integrating knowledge about the frequency of other, similar predicate-argument pairs. The task thus share some of the nature of language modelling; however, it is a task less amenable to approaches that require very large training corpora and one where the semantic quality of a model is of greater importance. This paper takes up tools (“topic models”) that have been proven successful in modelling document-word co-occurrences and adapts them to the task of selectional preference learning. Advantages of these models include a well-defined generative model that handles sparse data well, the ability to jointly induce semantic classes and predicate-specific distributions over those classes, and the enhanced statistical strength achieved by sharing knowledge across predicates. Section 2 surveys prior work on selectional preference modelling and on semantic applications of topic models. Section 3 describes the models used in our experiments. Section 4 provides details of the experimental design. Section 5 presents results for our models on the task of predicting human plausibility judgements for predicate-argument combinations; we show that performance is generally competi1At time of writing, Google estimates 855 hits for “a|the carrot|carrots laugh|laughs|laughed” and 0 hits for “a|the manservant|manservants|menservants laugh|laughs|laughed”; many of the carrot hits are false positives but a significant number are true subject-verb observations. 435 tive with or superior to a number of other models, including models using Web-scale resources, especially for low-frequency examples. In Section 6 we wrap up by summarising the paper’s conclusions and sketching directions for future research. 2 Related work 2.1 Selectional preference learning The representation (and latterly, learning) of selectional preferences for verbs and other predicates has long been considered a fundamental problem in computational semantics (Resnik, 1993). Many approaches to the problem use lexical taxonomies such as WordNet to identify the semantic classes that typically fill a particular argument slot for a predicate (Resnik, 1993; Clark and Weir, 2002; Schulte im Walde et al., 2008). In this paper, however, we focus on methods that do not assume the availability of a comprehensive taxonomy but rather induce semantic classes automatically from a corpus of text. Such methods are more generally applicable, for example in domains or languages where handbuilt semantic lexicons have insufficient coverage or are non-existent. Rooth et al. (1999) introduced a model of selectional preference induction that casts the problem in a probabilistic latent-variable framework. In Rooth et al.’s model each observed predicateargument pair is probabilistically generated from a latent variable, which is itself generated from an underlying distribution on variables. The use of latent variables, which correspond to coherent clusters of predicate-argument interactions, allow probabilities to be assigned to predicate-argument pairs which have not previously been observed by the model. The discovery of these predicate-argument clusters and the estimation of distributions on latent and observed variables are performed simultaneously via an Expectation Maximisation procedure. The work presented in this paper is inspired by Rooth et al.’s latent variable approach, most directly in the model described in Section 3.3. Erk (2007) and Pad´o et al. (2007) describe a corpusdriven smoothing model which is not probabilistic in nature but relies on similarity estimates from a “semantic space” model that identifies semantic similarity with closeness in a vector space of cooccurrences. Bergsma et al. (2008) suggest learning selectional preferences in a discriminative way, by training a collection of SVM classifiers to recognise likely and unlikely arguments for predicates of interest. Keller and Lapata (2003) suggest a simple alternative to smoothing-based approaches. They demonstrate that noisy counts from a Web search engine can yield estimates of plausibility for predicate-argument pairs that are superior to models learned from a smaller parsed corpus. The assumption inherent in this approach is that given sufficient text, all plausible predicate-argument pairs will be observed with frequency roughly correlated with their degree of plausibility. While the model is undeniably straightforward and powerful, it has a number of drawbacks: it presupposes an extremely large corpus, the like of which will only be available for a small number of domains and languages, and it is only suitable for relations that are identifiable by searching raw text for specific lexical patterns. 2.2 Topic modelling The task of inducing coherent semantic clusters is common to many research areas. In the field of document modelling, a class of methods known as “topic models” have become a de facto standard for identifying semantic structure in documents. These include the Latent Dirichlet Allocation (LDA) model of Blei et al. (2003) and the Hierarchical Dirichlet Process model of Teh et al. (2006). Formally seen, these are hierarchical Bayesian models which induce a set of latent variables or topics that are shared across documents. The combination of a well-defined probabilistic model and Gibbs sampling procedure for estimation guarantee (eventual) convergence and the avoidance of degenerate solutions. As a result of intensive research in recent years, the behaviour of topic models is well-understood and computationally efficient implementations have been developed. The tools provided by this research are used in this paper as the building blocks of our selectional preference models. Hierarchical Bayesian modelling has recently gained notable popularity in many core areas of natural language processing, from morphological segmentation (Goldwater et al., 2009) to opinion modelling (Lin et al., 2006). Yet so far there have been relatively few applications to traditional lexical semantic tasks. Boyd-Graber et al. (2007) integrate a model of random walks on the WordNet graph into an LDA topic model to build an unsupervised word sense disambiguation system. Brody 436 and Lapata (2009) adapt the basic LDA model for application to unsupervised word sense induction; in this context, the topics learned by the model are assumed to correspond to distinct senses of a particular lemma. Zhang et al. (2009) are also concerned with inducing multiple senses for a particular term; here the goal is to identify distinct entity types in the output of a pattern-based entity set discovery system. Reisinger and Pas¸ca (2009) use LDA-like models to map automatically acquired attribute sets onto the WordNet hierarchy. Griffiths et al. (2007) demonstrate that topic models learned from document-word co-occurrences are good predictors of semantic association judgements by humans. Simultaneously to this work, Ritter et al. (2010) have also investigated the use of topic models for selectional preference learning. Their goal is slightly different to ours in that they wish to model the probability of a binary predicate taking two specified arguments, i.e., P(n1, n2|v), whereas we model the joint and conditional probabilities of a predicate taking a single specified argument. The model architecture they propose, LinkLDA, falls somewhere between our LDA and DUAL-LDA models. Hence LinkLDA could be adapted to estimate P(n, v|r) as DUAL-LDA does, but a preliminary investigation indicates that it does not perform well in this context. The most likely explanation is that LinkLDA generates its two arguments independently, which may be suitable for distinct argument positions of a given predicate but is unsuitable when one of those “arguments” is in fact the predicate. The models developed in this paper, though intended for semantic modelling, also bear some similarity to the internals of generative syntax models such as the “infinite tree” (Finkel et al., 2007). In some ways, our models are less ambitious than comparable syntactic models as they focus on specific fragments of grammatical structure rather than learning a more general representation of sentence syntax. It would be interesting to evaluate whether this restricted focus improves the quality of the learned model or whether general syntax models can also capture fine-grained knowledge about combinatorial semantics. 3 Three selectional preference models 3.1 Notation In the model descriptions below we assume a predicate vocabulary of V types, an argument vocabulary of N types and a relation vocabulary of R types. Each predicate type is associated with a singe relation; for example the predicate type eat:V:dobj (the direct object of the verb eat) is treated as distinct from eat:V:subj (the subject of the verb eat). The training corpus consists of W observations of argument-predicate pairs. Each model has at least one vocabulary of Z arbitrarily labelled latent variables. fzn is the number of observations where the latent variable z has been associated with the argument type n, fzv is the number of observations where z has been associated with the predicate type v and fzr is the number of observations where z has been associated with the relation r. Finally, fz· is the total number of observations associated with z and f·v is the total number of observations containing the predicate v. 3.2 Latent Dirichlet Allocation As noted above, LDA was originally introduced to model sets of documents in terms of topics, or clusters of terms, that they share in varying proportions. For example, a research paper on bioinformatics may use some vocabulary that is shared with general computer science papers and some vocabulary that is shared with biomedical papers. The analogical move from modelling document-term cooccurrences to modelling predicate-argument cooccurrences is intuitive: we assume that each predicate is associated with a distribution over semantic classes (“topics”) and that these classes are shared across predicates. The high-level “generative story” for the LDA selectional preference model is as follows: (1) For each predicate v, draw a multinomial distribution Θv over argument classes from a Dirichlet distribution with parameters α. (2) For each argument class z, draw a multinomial distribution Φz over argument types from a Dirichlet with parameters β. (3) To generate an argument for v, draw an argument class z from Θv and then draw an argument type n from Φz The resulting model can be written as: P(n|v, r) = X z P(n|z)P(z|v, r) (1) ∝ X z fzn + β fz· + Nβ fzv + αz f·v + P z′ αz′ (2) 437 Due to multinomial-Dirichlet conjugacy, the distributions Θv and Φz can be integrated out and do not appear explicitly in the above formula. The first term in (2) can be seen as a smoothed estimate of the probability that class z produces the argument n; the second is a smoothed estimate of the probability that predicate v takes an argument belonging to class z. One important point is that the smoothing effects of the Dirichlet priors on Θv and Φz are greatest for predicates and arguments that are rarely seen, reflecting an intuitive lack of certainty. We assume an asymmetric Dirichlet prior on Θv (the α parameters can differ for each class) and a symmetric prior on Φz (all β parameters are equal); this follows the recommendations of Wallach et al. (2009) for LDA. This model estimates predicate-argument probabilities conditional on a given predicate v; it cannot by itself provide joint probabilities P(n, v|r), which are needed for our plausibility evaluation. Given a dataset of predicate-argument combinations and values for the hyperparameters α and β, the probability model is determined by the class assignment counts fzn and fzv. Following Griffiths and Steyvers (2004), we estimate the model by Gibbs sampling. This involves resampling the topic assignment for each observation in turn using probabilities estimated from all other observations. One efficiency bottleneck in the basic sampler described by Griffiths and Steyvers is that the entire set of topics must be iterated over for each observation. Yao et al. (2009) propose a reformulation that removes this bottleneck by separating the probability mass p(z|n, v) into a number of buckets, some of which only require iterating over the topics currently assigned to instances of type n, typically far fewer than the total number of topics. It is possible to apply similar reformulations to the models presented in Sections 3.3 and 3.4 below; depending on the model and parameterisation this can reduce the running time dramatically. Unlike some topic models such as HDP (Teh et al., 2006), LDA is parametric: the number of topics Z must be set by the user in advance. However, Wallach et al. (2009) demonstrate that LDA is relatively insensitive to larger-than-necessary choices of Z when the Dirichlet parameters α are optimised as part of model estimation. In our implementation we use the optimisation routines provided as part of the Mallet library, which use an iterative procedure to compute a maximum likelihood estimate of these hyperparameters.2 3.3 A Rooth et al.-inspired model In Rooth et al.’s (1999) selectional preference model, a latent variable is responsible for generating both the predicate and argument types of an observation. The basic LDA model can be extended to capture this kind of predicate-argument interaction; the generative story for the resulting ROOTH-LDA model is as follows: (1) For each relation r, draw a multinomial distribution Θr over interaction classes from a Dirichlet distribution with parameters α. (2) For each class z, draw a multinomial Φz over argument types from a Dirichlet distribution with parameters β and a multinomial Ψz over predicate types from a Dirichlet distribution with parameters γ. (3) To generate an observation for r, draw a class z from Θr, then draw an argument type n from Φz and a predicate type v from Ψz. The resulting model can be written as: P(n, v|r) = X z P(n|z)P(v|z)P(z|r) (3) ∝ X z fzn + β fz· + Nβ fzv + γ fz· + V γ fzr + αz f·r + P z′ αz′ (4) As suggested by the similarity between (4) and (2), the ROOTH-LDA model can be estimated by an LDA-like Gibbs sampling procedure. Unlike LDA, ROOTH-LDA does model the joint probability P(n, v|r) of a predicate and argument co-occurring. Further differences are that information about predicate-argument co-occurrence is only shared within a given interaction class rather than across the whole dataset and that the distribution Φz is not specific to the predicate v but rather to the relation r. This could potentially lead to a loss of model quality, but in practice the ability to induce “tighter” clusters seems to counteract any deterioration this causes. 3.4 A “dual-topic” model In our third model, we attempt to combine the advantages of LDA and ROOTH-LDA by clustering arguments and predicates according to separate 2http://mallet.cs.umass.edu/ 438 class vocabularies. Each observation is generated by two latent variables rather than one, which potentially allows the model to learn more flexible interactions between arguments and predicates.: (1) For each relation r, draw a multinomial distribution Ξr over predicate classes from a Dirichlet with parameters κ. (2) For each predicate class c, draw a multinomial Ψc over predicate types and a multinomial Θc over argument classes from Dirichlets with parameters γ and α respectively. (3) For each argument class z, draw a multinomial distribution Φz over argument types from a Dirichlet with parameters β. (4) To generate an observation for r, draw a predicate class c from Ξr, a predicate type from Ψc, an argument class z from Θc and an argument type from Φz. The resulting model can be written as: P(n, v|r) = X c X z P(n|z)P(z|c)P(v|c)P(c|r) (5) ∝ X c X z fzn + β fz· + Nβ fzc + αz f·c + P z′ αz′ × fcv + γ fc· + V γ fcr + κc f·r + P c′ κc′ (6) To estimate this model, we first resample the class assignments for all arguments in the data and then resample class assignments for all predicates. Other approaches are possible – resampling argument and then predicate class assignments for each observation in turn, or sampling argument and predicate assignments together by blocked sampling – though from our experiments it does not seem that the choice of scheme makes a significant difference. 4 Experimental setup In the document modelling literature, probabilistic topic models are often evaluated on the likelihood they assign to unseen documents; however, it has been shown that higher log likelihood scores do not necessarily correlate with more semantically coherent induced topics (Chang et al., 2009). One popular method for evaluating selectional preference models is by testing the correlation between their predictions and human judgements of plausibility on a dataset of predicate-argument pairs. This can be viewed as a more semantically relevant measurement of model quality than likelihood-based methods, and also permits comparison with nonprobabilistic models. In Section 5, we use two plausibility datasets to evaluate our models and compare to other previously published results. We trained our models on the 90-million word written component of the British National Corpus (Burnard, 1995), parsed with the RASP toolkit (Briscoe et al., 2006). Predicates occurring with just one argument type were removed, as were all tokens containing non-alphabetic characters; no other filtering was done. The resulting datasets consisted of 3,587,172 verb-object observations with 7,954 predicate types and 80,107 argument types, 3,732,470 noun-noun observations with 68,303 predicate types and 105,425 argument types, and 3,843,346 adjective-noun observations with 29,975 predicate types and 62,595 argument types. During development we used the verb-noun plausibility dataset from Pad´o et al. (2007) to direct the design of the system. Unless stated otherwise, all results are based on runs of 1,000 iterations with 100 classes, with a 200-iteration burnin period after which hyperparameters were reestimated every 50 iterations.3 The probabilities estimated by the models (P(n|v, r) for LDA and P(n, v|r) for ROOTH- and DUAL-LDA) were sampled every 50 iterations post-burnin and averaged over three runs to smooth out variance. To compare plausibility scores for different predicates, we require the joint probability P(n, v|r); as LDA does not provide this, we approximate PLDA(n, v|r) = PBNC(v|r)PLDA(n|v, r), where PBNC(v|r) is proportional to the frequency with which predicate v is observed as an instance of relation r in the BNC. For comparison, we reimplemented the methods of Rooth et al. (1999) and Pad´o et al. (2007). As mentioned above, Rooth et al. use a latent-variable model similar to (4) but without priors, trained via EM. Our implementation (henceforth ROOTHEM) chooses the number of classes from the range (20, 25, . . . , 50) through 5-fold cross-validation on a held-out log-likelihood measure. Settings outside this range did not give good results. Again, we run for 1,000 iterations and average predictions over 3These settings were based on the MALLET defaults; we have not yet investigated whether modifying the simulation length or burnin period is beneficial. 439 LDA 0 Nouns: agreement, contract, permission, treaty, deal, ... 1 Nouns information, datum, detail, evidence, material, ... 2 Nouns skill, knowledge, country, technique, understanding, ... ROOTH-LDA 0 Nouns force, team, army, group, troops, .. . 0 Verbs join, arm, lead, beat, send, ... 1 Nouns door, eye, mouth, window, gate, ... 1 Verbs open, close, shut, lock, slam, ... DUAL-LDA 0N Nouns house, building, site, home, station, . .. 1N Nouns stone, foot, bit, breath, line, ... 0V Verbs involve, join, lead, represent, concern, ... 1V Verbs see, break, have, turn, round, ... ROOTH-EM 0 Nouns system, method, technique, skill, model, . .. 0 Verbs use, develop, apply, design, introduce, ... 1 Nouns eye, door, page, face, chapter,... 1 Verbs see, open, close, watch, keep,... Table 1: Most probable words for sample semantic classes induced from verb-object observations three runs. Pad´o et al. (2007), a refinement of Erk (2007), is a non-probabilistic method that smooths predicate-argument counts with counts for other observed arguments of the same predicate, weighted by the similarity between arguments. Following their description, we use a 2,000-dimensional space of syntactic co-occurrence features appropriate to the relation being predicted, weight features with the G2 transformation and compute similarity with the cosine measure. 5 Results 5.1 Induced semantic classes Table 1 shows sample semantic classes induced by models trained on the corpus of BNC verb-object co-occurrences. LDA clusters nouns only, while ROOTH-LDA and ROOTH-EM learn classes that generate both nouns and verbs and DUAL-LDA clusters nouns and verbs separately. The LDA clusters are generally sensible: class 0 is exemplified by agreement and contract and class 1 by information and datum. There are some unintuitive blips, for example country appears between knowledge and understanding in class 2. The ROOTH-LDA classes also feel right: class 0 deals with nouns such as force, team and army which one might join, arm or lead and class 1 corresponds to “things that can be opened or closed” such as a door, an eye or a mouth (though the model also makes the questionable prediction that all these items can plausibly be locked or slammed). The DUAL-LDA classes are notably less coherent, especially when it comes to clustering verbs: DUAL-LDA’s class 0V, like ROOTH-LDA’s class 0, has verbs that take groups as objects but its class 1V mixes sensible conflations (turn, round) with very common verbs such as see and have and the unrelated break. The general impression given by inspection of the DUAL-LDA model is that it has problems with mixing and does not manage to learn a good model; we have tried a number of solutions (e.g., blocked sampling of argument and predicate classes), without overcoming this brittleness. Unsurprisingly, ROOTH-EM’s classes have a similar feel to ROOTH-LDA; our general impression is that some of ROOTH-EM’s classes look even more coherent than the LDAbased models, presumably because it does not use priors to smooth its per-class distributions. 5.2 Comparison with Keller and Lapata (2003) Keller and Lapata (2003) collected a dataset of human plausibility judgements for three classes of grammatical relation: verb-object, noun-noun modification and adjective-noun modification. The items in this dataset were not chosen to balance plausibility and implausibility (as in prior psycholinguistic experiments) but according to their corpus frequency, leading to a more realistic task. 30 predicates were selected for each relation; each predicate was matched with three arguments from different co-occurrence bands in the BNC, e.g., naughty-girl (high frequency), naughty-dog (medium) and naughty-lunch (low). Each predicate was also matched with three random arguments 440 Verb-object Noun-noun Adjective-noun Seen Unseen Seen Unseen Seen Unseen r ρ r ρ r ρ r ρ r ρ r ρ AltaVista (KL) .641 – .551 – .700 – .578 – .650 – .480 – Google (KL) .624 – .520 – .692 – .595 – .641 – .473 – BNC (RASP) .620 .614 .196 .222 .544 .604 .114 .125 .543 .622 .135 .102 ROOTH-EM .455 .487 .479 .520 .503 .491 .586 .625 .514 .463 .395 .355 Pad´o et al. .484 .490 .398 .430 .431 .503 .558 .533 .479 .570 .120 .138 LDA .504 .541 .558 .603 .615 .641 .636 .666 .594 .558 .468 .459 ROOTH-LDA .520 .548 .564 .605 .607 .622 .691 .722 .575 .599 .501 .469 DUAL-LDA .453 .494 .446 .516 .496 .494 .553 .573 .460 .400 .334 .278 Table 2: Results (Pearson r and Spearman ρ correlations) on Keller and Lapata’s (2003) plausibility data with which it does not co-occur in the BNC (e.g., naughty-regime, naughty-rival, naughty-protocol). In this way two datasets (Seen and Unseen) of 90 items each were assembled for each predicate. Table 2 presents results for a variety of predictive models – the Web frequencies reported by Keller and Lapata (2003) for two search engines, frequencies from the RASP-parsed BNC,4 the reimplemented methods of Rooth et al. (1999) and Pad´o et al. (2007), and the LDA, ROOTH-LDA and DUALLDA topic models. Following Keller and Lapata, we report Pearson correlation coefficients between log-transformed predicted frequencies and the goldstandard plausibility scores (which are already logtransformed). We also report Spearman rank correlations except where we do not have the original predictions (the Web count models), for completeness and because the predictions of preference models are may not be log-normally distributed as corpus counts are. Zero values (found only in the BNC frequency predictions) were smoothed by 0.1 to facilitate the log transformation; it seems natural to take a zero prediction as a non-specific prediction of very low plausibility rather than a “missing value” as is done in other work (e.g., Pad´o et al., 2007). Despite their structural differences, LDA and ROOTH-LDA perform similarly - indeed, their predictions are highly correlated. ROOTH-LDA scores best overall, outperforming Pad´o et al.’s (2007) method and ROOTH-EM on every dataset and evaluation measure, and outperforming Keller and Lapata’s (2003) Web predictions on every Un4The correlations presented here for BNC counts are notably better than those reported by Keller and Lapata (2003), presumably reflecting our use of full parsing rather than shallow parsing. seen dataset. LDA also performs consistently well, surpassing ROOTH-EM and Pad´o et al. on all but one occasion. For frequent predicate-argument pairs (Seen datasets), Web counts are clearly better; however, the BNC counts are unambiguously superior to LDA and ROOTH-LDA (whose predictions are based entirely on the generative model even for observed items) for the Seen verb-object data only. As might be suspected from the mixing problems observed with DUAL-LDA, this model does not perform as well as LDA and ROOTH-LDA, though it does hold its own against the other selectional preference methods. To identify significant differences between models, we use the statistical test for correlated correlation coefficients proposed by Meng et al. (1992), which is appropriate for correlations that share the same gold standard.5 For the seen data there are few significant differences: ROOTH-LDA and LDA are significantly better (p < 0.01) than Pad´o et al.’s model for Pearson’s r on seen noun-noun data, and ROOTH-LDA is also significantly better (p < 0.01) using Spearman’s ρ. For the unseen datasets, the BNC frequency predictions are unsurprisingly significantly worse at the p < 0.01 level than all smoothing models. LDA and ROOTHLDA are significantly better (p < 0.01) than Pad´o et al. on every unseen dataset; ROOTH-EM is significantly better (p < 0.01) than Pad´o et al. on Unseen adjectives for both correlations. Meng et al.’s test does not find significant differences between ROOTH-EM and the LDA models despite the latter’s clear advantages (a number of conditions do come close). This is because their predictions are highly correlated, which is perhaps 5We cannot compare our data to Keller and Lapata’s Web counts as we do not possess their per-item scores. 441 50 100 150 200 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 No. of classes ρ (a) Verb-object 50 100 150 200 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 No. of classes ρ (b) Noun-noun 50 100 150 200 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 No. of classes ρ (c) Adjective-noun Figure 1: Effect of number of argument classes on Spearman rank correlation with LDA: the solid and dotted lines show the Seen and Unseen datasets respectively; bars show locations of individual samples unsurprising given that they are structurally similar models trained on the same data. We hypothesise that the main reason for the superior numerical performance of the LDA models over EM is the principled smoothing provided by the use of Dirichlet priors, which has a small but discriminative effect on model predictions. Collating the significance scores, we find that ROOTH-LDA achieves the most positive outcomes, followed by LDA and then by ROOTH-EM. DUAL-LDA is found significantly better than Pad´o et al.’s model on unseen adjectivenoun combinations, and significantly worse than the same model on seen adjective-noun data. Latent variable models that use EM for inference can be very sensitive to the number of latent variables chosen. For example, the performance of ROOTH-EM worsens quickly if the number of clusters is overestimated; for the Keller and Lapata datasets, settings above 50 classes lead to clear overfitting and a precipitous drop in Pearson correlation scores. On the other hand, Wallach et al. (2009) demonstrate that LDA is relatively insensitive to the choice of topic vocabulary size Z when the α and β hyperparameters are optimised appropriately during estimation. Figure 1 plots the effect of Z on Spearman correlation for the LDA model. In general, Wallach et al.’s finding for document modelling transfers to selectional preference models; within the range Z = 50–200 performance remains at a roughly similar level. In fact, we do not find that performance becomes significantly less robust when hyperparameter reestimation is deactiviated; correlation scores simply drop by a small amount (1–2 points), irrespective of the Z chosen. ROOTH-LDA (not graphed) seems slightly more sensitive to Z; this may be because the α parameters in this model operate on the relation level rather than the document level and thus fewer “observations” of class distributions are available when reestimating them. 5.3 Comparison with Bergsma et al. (2008) As mentioned in Section 2.1, Bergsma et al. (2008) propose a discriminative approach to preference learning. As part of their evaluation, they compare their approach to a number of others, including that of Erk (2007), on a plausibility dataset collected by Holmes et al. (1989). This dataset consists of 16 verbs, each paired with one plausible object (e.g., write-letter) and one implausible object (write-market). Bergsma et al.’s model, trained on the 3GB AQUAINT corpus, is the only model reported to achieve perfect accuracy on distinguishing plausible from implausible arguments. It would be interesting to do a full comparison that controls for size and type of corpus data; in the meantime, we can report that the LDA and ROOTH-LDA models trained on verb-object observations in the BNC (about 4 times smaller than AQUAINT) also achieve a perfect score on the Holmes et al. data.6 6 Conclusions and future work This paper has demonstrated how Bayesian techniques originally developed for modelling the topical structure of documents can be adapted to learn probabilistic models of selectional preference. These models are especially effective for estimating plausibility of low-frequency items, thus distinguishing rarity from clear implausibility. The models presented here derive their predictions by modelling predicate-argument plausibility through the intermediary of latent variables. As observed in Section 5.2 this may be a suboptimal 6Bergsma et al. report that all plausible pairs were seen in their corpus; three were unseen in ours, as well as 12 of the implausible pairs. 442 strategy for frequent combinations, where corpus counts are probably reliable and plausibility judgements may be affected by lexical collocation effects. One principled method for folding corpus counts into LDA-like models would be to use hierarchical priors, as in the n-gram topic model of Wallach (2006). Another potential direction for system improvement would be an integration of our generative model with Bergsma et al.’s (2008) discriminative model – this could be done in a number of ways, including using the induced classes of a topic model as features for a discriminative classifier or using the discriminative classifier to produce additional high-quality training data from noisy unparsed text. Comparison to plausibility judgements gives an intrinsic measure of model quality. As mentioned in the Introduction, selectional preferences have many uses in NLP applications, and it will be interesting to evaluate the utility of Bayesian preference models in contexts such as semantic role labelling or human sentence processing modelling. The probabilistic nature of topic models, coupled with an appropriate probabilistic task model, may facilitate the integration of class induction and task learning in a tight and principled way. We also anticipate that latent variable models will prove effective for learning selectional preferences of semantic predicates (e.g., FrameNet roles) where direct estimation from a large corpus is not a viable option. Acknowledgements This work was supported by EPSRC grant EP/G051070/1. I am grateful to Frank Keller and Mirella Lapata for sharing their plausibility data, and to Andreas Vlachos and the anonymous ACL and CoNLL reviewers for their helpful comments. References Shane Bergsma, Dekang Lin, and Randy Goebel. 2008. Discriminative learning of selectional preferences from unlabeled text. In Proceedings of EMNLP-08, Honolulu, HI. David M. Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent Dirichlet allocation. Journal of Machine Learning Research, 3:993–1022. Jordan Boyd-Graber, David Blei, and Xiaojin Zhu. 2007. A topic model for word sense disambiguation. In Proceedings of EMNLP-CoNLL-07, Prague, Czech Republic. Ted Briscoe, John Carroll, and Rebecca Watson. 2006. The second release of the RASP system. In Proceedings of the ACL-06 Interactive Presentation Sessions, Sydney, Australia. Samuel Brody and Mirella Lapata. 2009. Bayesian word sense induction. In Proceedings of EACL-09, Athens, Greece. Lou Burnard, 1995. Users’ Guide for the British National Corpus. British National Corpus Consortium, Oxford University Computing Service, Oxford, UK. Jonathan Chang, Jordan Boyd-Graber, Sean Gerrish, Chong Wang, and David M. Blei. 2009. Reading tea leaves: How humans interpret topic models. In Proceedings of NIPS-09, Vancouver, BC. Stephen Clark and David Weir. 2002. Class-based probability estimation using a semantic hierarchy. Computational Linguistics, 28(2):187–206. Katrin Erk. 2007. A simple, similarity-based model for selectional preferences. In Proceedings of ACL07, Prague, Czech Republic. Jenny Rose Finkel, Trond Grenager, and Christopher D. Manning. 2007. The infinite tree. In Proceedings of ACL-07, Prague, Czech Republic. Daniel Gildea and Daniel Jurafsky. 2002. Automatic labeling of semantic roles. Computational Linguistics, 28(3):245–288. Sharon Goldwater, Thomas L. Griffiths, and Mark Johnson. 2009. A Bayesian framework for word segmentation: Exploring the effects of context. Cognition, 112(1):21–54. Thomas L. Griffiths and Mark Steyvers. 2004. Finding scientific topics. Proceedings of the National Academy of Sciences, 101(suppl. 1):5228–5235. Thomas L. Griffiths, Mark Steyvers, and Joshua B. Tenenbaum. 2007. Topics in semantic representation. Psychological Review, 114(2):211–244. Virginia M. Holmes, Laurie Stowe, and Linda Cupples. 1989. Lexical expectations in parsing complementverb sentences. Journal of Memory and Language, 28(6):668–689. Frank Keller and Mirella Lapata. 2003. Using the Web to obtain frequencies for unseen bigrams. Computational Linguistics, 29(3):459–484. Wei-Hao Lin, Theresa Wilson, Janyce Wiebe, and Alexander Hauptmann. 2006. Which side are you on? Identifying perspectives at the document and sentence levels. In Proceedings of CoNLL-06, New York, NY. Xiao-Li Meng, Robert Rosenthal, and Donald B. Rubin. 1992. Comparing correlated correlation coefficients. Psychological Bulletin, 111(1):172–175. 443 Sebastian Pad´o, Ulrike Pad´o, and Katrin Erk. 2007. Flexible, corpus-based modelling of human plausibility judgements. In Proceedings of EMNLPCoNLL-07, Prague, Czech Republic. Patrick Pantel, Rahul Bhagat, Bonaventura Coppola, Timothy Chklovski, and Eduard Hovy. 2007. ISP: Learning inferential selectional preferences. In Proceedings of NAACL-HLT-07, Rochester, NY. Keith Rayner, Tessa Warren, Barbara J. Juhasz, and Simon P. Liversedge. 2004. The effect of plausibility on eye movements in reading. Journal of Experimental Psychology: Learning Memory and Cognition, 30(6):1290–1301. Joseph Reisinger and Marius Pas¸ca. 2009. Latent variable models of concept-attribute attachment. In Proceedings of ACL-IJCNLP-09, Singapore. Philip S. Resnik. 1993. Selection and Information: A Class-Based Approach to Lexical Relationships. Ph.D. thesis, University of Pennsylvania. Alan Ritter, Mausam, and Oren Etzioni. 2010. A Latent Dirichlet Allocation method for selectional preferences. In Proceedings of ACL-10, Uppsala, Sweden. Mats Rooth, Stefan Riezler, Detlef Prescher, Glenn Carroll, and Franz Beil. 1999. Inducing a semantically annotated lexicon via EM-based clustering. In Proceedings of ACL-99, College Park, MD. Sabine Schulte im Walde, Christian Hying, Christian Scheible, and Helmut Schmid. 2008. Combining EM training and the MDL principle for an automatic verb classification incorporating selectional preferences. In Proceedings of ACL-08:HLT, Columbus, OH. Yee W. Teh, Michael I. Jordan, Matthew J. Beal, and David M. Blei. 2006. Hierarchical Dirichlet processes. Journal of the American Statistical Association, 101(476):1566–1581. Hanna Wallach, David Mimno, and Andrew McCallum. 2009. Rethinking LDA: Why priors matter. In Proceedings of NIPS-09, Vancouver, BC. Hanna Wallach. 2006. Topic modeling: Beyond bagof-words. In Proceedings of ICML-06, Pittsburgh, PA. Yorick Wilks. 1978. Making preferences more active. Artificial Intelligence, 11:197–225. Limin Yao, David Mimno, and Andrew McCallum. 2009. Efficient methods for topic model inference on streaming document collections. In Proceedings of KDD-09, Paris, France. Be˜nat Zapirain, Eneko Agirre, and Llu´ıs M`arquez. 2009. Generalizing over lexical features: Selectional preferences for semantic role classification. In Proceedings of ACL-IJCNLP-09, Singapore. Huibin Zhang, Mingjie Zhu, Shuming Shi, and Ji-Rong Wen. 2009. Employing topic models for patternbased semantic class discovery. In Proceedings of ACL-IJCNLP-09, Singapore. 444
|
2010
|
45
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 445–453, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Improving the Use of Pseudo-Words for Evaluating Selectional Preferences Nathanael Chambers and Dan Jurafsky Department of Computer Science Stanford University {natec,jurafsky}@stanford.edu Abstract This paper improves the use of pseudowords as an evaluation framework for selectional preferences. While pseudowords originally evaluated word sense disambiguation, they are now commonly used to evaluate selectional preferences. A selectional preference model ranks a set of possible arguments for a verb by their semantic fit to the verb. Pseudo-words serve as a proxy evaluation for these decisions. The evaluation takes an argument of a verb like drive (e.g. car), pairs it with an alternative word (e.g. car/rock), and asks a model to identify the original. This paper studies two main aspects of pseudoword creation that affect performance results. (1) Pseudo-word evaluations often evaluate only a subset of the words. We show that selectional preferences should instead be evaluated on the data in its entirety. (2) Different approaches to selecting partner words can produce overly optimistic evaluations. We offer suggestions to address these factors and present a simple baseline that outperforms the state-ofthe-art by 13% absolute on a newspaper domain. 1 Introduction For many natural language processing (NLP) tasks, particularly those involving meaning, creating labeled test data is difficult or expensive. One way to mitigate this problem is with pseudowords, a method for automatically creating test corpora without human labeling, originally proposed for word sense disambiguation (Gale et al., 1992; Schutze, 1992). While pseudo-words are now less often used for word sense disambigation, they are a common way to evaluate selectional preferences, models that measure the strength of association between a predicate and its argument filler, e.g., that the noun lunch is a likely object of eat. Selectional preferences are useful for NLP tasks such as parsing and semantic role labeling (Zapirain et al., 2009). Since evaluating them in isolation is difficult without labeled data, pseudoword evaluations can be an attractive evaluation framework. Pseudo-word evaluations are currently used to evaluate a variety of language modeling tasks (Erk, 2007; Bergsma et al., 2008). However, evaluation design varies across research groups. This paper studies the evaluation itself, showing how choices can lead to overly optimistic results if the evaluation is not designed carefully. We show in this paper that current methods of applying pseudo-words to selectional preferences vary greatly, and suggest improvements. A pseudo-word is the concatenation of two words (e.g. house/car). One word is the original in a document, and the second is the confounder. Consider the following example of applying pseudo-words to the selectional restrictions of the verb focus: Original: This story focuses on the campaign. Test: This story/part focuses on the campaign/meeting. In the original sentence, focus has two arguments: a subject story and an object campaign. In the test sentence, each argument of the verb is replaced by pseudo-words. A model is evaluated by its success at determining which of the two arguments is the original word. Two problems exist in the current use of 445 pseudo-words to evaluate selectional preferences. First, selectional preferences historically focus on subsets of data such as unseen words or words in certain frequency ranges. While work on unseen data is important, evaluating on the entire dataset provides an accurate picture of a model’s overall performance. Most other NLP tasks today evaluate all test examples in a corpus. We will show that seen arguments actually dominate newspaper articles, and thus propose creating test sets that include all verb-argument examples to avoid artificial evaluations. Second, pseudo-word evaluations vary in how they choose confounders. Previous work has attempted to maintain a similar corpus frequency to the original, but it is not clear how best to do this, nor how it affects the task’s difficulty. We argue in favor of using nearest-neighbor frequencies and show how using random confounders produces overly optimistic results. Finally, we present a surprisingly simple baseline that outperforms the state-of-the-art and is far less memory and computationally intensive. It outperforms current similarity-based approaches by over 13% when the test set includes all of the data. We conclude with a suggested backoff model based on this baseline. 2 History of Pseudo-Word Disambiguation Pseudo-words were introduced simultaneously by two papers studying statistical approaches to word sense disambiguation (WSD). Sch¨utze (1992) simply called the words, ‘artificial ambiguous words’, but Gale et al. (1992) proposed the succinct name, pseudo-word. Both papers cited the sparsity and difficulty of creating large labeled datasets as the motivation behind pseudo-words. Gale et al. selected unambiguous words from the corpus and paired them with random words from different thesaurus categories. Sch¨utze paired his words with confounders that were ‘comparable in frequency’ and ‘distinct semantically’. Gale et al.’s pseudo-word term continues today, as does Sch¨utze’s frequency approach to selecting the confounder. Pereira et al. (1993) soon followed with a selectional preference proposal that focused on a language model’s effectiveness on unseen data. The work studied clustering approaches to assist in similarity decisions, predicting which of two verbs was the correct predicate for a given noun object. One verb v was the original from the source document, and the other v′ was randomly generated. This was the first use of such verb-noun pairs, as well as the first to test only on unseen pairs. Several papers followed with differing methods of choosing a test pair (v, n) and its confounder v′. Dagan et al. (1999) tested all unseen (v, n) occurrences of the most frequent 1000 verbs in his corpus. They then sorted verbs by corpus frequency and chose the neighboring verb v′ of v as the confounder to ensure the closest frequency match possible. Rooth et al. (1999) tested 3000 random (v, n) pairs, but required the verbs and nouns to appear between 30 and 3000 times in training. They also chose confounders randomly so that the new pair was unseen. Keller and Lapata (2003) specifically addressed the impact of unseen data by using the web to first ‘see’ the data. They evaluated unseen pseudowords by attempting to first observe them in a larger corpus (the Web). One modeling difference was to disambiguate the nouns as selectional preferences instead of the verbs. Given a test pair (v, n) and its confounder (v, n′), they used web searches such as “v Det n” to make the decision. Results beat or matched current results at the time. We present a similarly motivated, but new webbased approach later. Very recent work with pseudo-words (Erk, 2007; Bergsma et al., 2008) further blurs the lines between what is included in training and test data, using frequency-based and semantic-based reasons for deciding what is included. We discuss this further in section 5. As can be seen, there are two main factors when devising a pseudo-word evaluation for selectional preferences: (1) choosing (v, n) pairs from the test set, and (2) choosing the confounding n′ (or v′). The confounder has not been looked at in detail and as best we can tell, these factors have varied significantly. Many times the choices are well motivated based on the paper’s goals, but in other cases the motivation is unclear. 3 How Frequent is Unseen Data? Most NLP tasks evaluate their entire datasets, but as described above, most selectional preference evaluations have focused only on unseen data. This section investigates the extent of unseen examples in a typical training/testing environment 446 of newspaper articles. The results show that even with a small training size, seen examples dominate the data. We argue that, absent a system’s need for specialized performance on unseen data, a representative test set should include the dataset in its entirety. 3.1 Unseen Data Experiment We use the New York Times (NYT) and Associated Press (APW) sections of the Gigaword Corpus (Graff, 2002), as well as the British National Corpus (BNC) (Burnard, 1995) for our analysis. Parsing and SRL evaluations often focus on newspaper articles and Gigaword is large enough to facilitate analysis over varying amounts of training data. We parsed the data with the Stanford Parser1 into dependency graphs. Let (vd, n) be a verb v with grammatical dependency d ∈ {subject, object, prep} filled by noun n. Pairs (vd, n) are chosen by extracting every such dependency in the graphs, setting the head predicate as v and the head word of the dependent d as n. All prepositions are condensed into prep. We randomly selected documents from the year 2001 in the NYT portion of the corpus as development and test sets. Training data for APW and NYT include all years 1994-2006 (minus NYT development and test documents). We also identified and removed duplicate documents2. The BNC in its entirety is also used for training as a single data point. We then record every seen (vd, n) pair during training that is seen two or more times3 and then count the number of unseen pairs in the NYT development set (1455 tests). Figure 1 plots the percentage of unseen arguments against training size when trained on either NYT or APW (the APW portion is smaller in total size, and the smaller BNC is provided for comparison). The first point on each line (the highest points) contains approximately the same number of words as the BNC (100 million). Initially, about one third of the arguments are unseen, but that percentage quickly falls close to 10% as additional training is included. This suggests that an evaluation focusing only on unseen data is not representative, potentially missing up to 90% of the data. 1http://nlp.stanford.edu/software/lex-parser.shtml 2Any two documents whose first two paragraphs in the corpus files are identical. 3Our results are thus conservative, as including all single occurrences would achieve even smaller unseen percentages. 0 2 4 6 8 10 12 0 5 10 15 20 25 30 35 40 45 Number of Tokens in Training (hundred millions) Percent Unseen Unseen Arguments in NYT Dev BNC AP NYT Google Figure 1: Percentage of NYT development set that is unseen when trained on varying amounts of data. The two lines represent training with NYT or APW data. The APW set is smaller in size from the NYT. The dotted line uses Google n-grams as training. The x-axis represents tokens × 108. 0 2 4 6 8 10 12 0 5 10 15 20 25 30 35 40 Number of Tokens in Training (hundred millions) Percent Unseen Unseen Arguments by Type Preps Subjects Objects Figure 2: Percentage of subject/object/preposition arguments in the NYT development set that is unseen when trained on varying amounts of NYT data. The x-axis represents tokens × 108. 447 The third line across the bottom of the figure is the number of unseen pairs using Google n-gram data as proxy argument counts. Creating argument counts from n-gram counts is described in detail below in section 5.2. We include these Web counts to illustrate how an openly available source of counts affects unseen arguments. Finally, figure 2 compares which dependency types are seen the least in training. Prepositions have the largest unseen percentage, but not surprisingly, also make up less of the training examples overall. In order to analyze why pairs are unseen, we analyzed the distribution of rare words across unseen and seen examples. To define rare nouns, we order head words by their individual corpus frequencies. A noun is rare if it occurs in the lowest 10% of the list. We similarly define rare verbs over their ordered frequencies (we count verb lemmas, and do not include the syntactic relations). Corpus counts covered 2 years of the AP section, and we used the development set of the NYT section to extract the seen and unseen pairs. Figure 3 shows the percentage of rare nouns and verbs that occur in unseen and seen pairs. 24.6% of the verbs in unseen pairs are rare, compared to only 4.5% in seen pairs. The distribution of rare nouns is less contrastive: 13.3% vs 8.9%. This suggests that many unseen pairs are unseen mainly because they contain low-frequency verbs, rather than because of containing low-frequency argument heads. Given the large amount of seen data, we believe evaluations should include all data examples to best represent the corpus. We describe our full evaluation results and include a comparison of different training sizes below. 4 How to Select a Confounder Given a test set S of pairs (vd, n) ∈S, we now address how best to select a confounder n′. Work in WSD has shown that confounder choice can make the pseudo-disambiguation task significantly easier. Gaustad (2001) showed that human-generated pseudo-words are more difficult to classify than random choices. Nakov and Hearst (2003) further illustrated how random confounders are easier to identify than those selected from semantically ambiguous, yet related concepts. Our approach evaluates selectional preferences, not WSD, but our results complement these findings. We identified three methods of confounder selection based on varying levels of corpus freverbs nouns Unseen Tests Seen Tests Distribution of Rare Verbs and Nouns in Tests Percent Rare Words 0 5 10 15 20 25 30 Figure 3: Comparison between seen and unseen tests (verb,relation,noun). 24.6% of unseen tests have rare verbs, compared to just 4.5% in seen tests. The rare nouns are more evenly distributed across the tests. quency: (1) choose a random noun, (2) choose a random noun from a frequency bucket similar to the original noun’s frequency, and (3) select the nearest neighbor, the noun with frequency closest to the original. These methods evaluate the range of choices used in previous work. Our experiments compare the three. 5 Models 5.1 A New Baseline The analysis of unseen slots suggests a baseline that is surprisingly obvious, yet to our knowledge, has not yet been evaluated. Part of the reason is that early work in pseudo-word disambiguation explicitly tested only unseen pairs4. Our evaluation will include seen data, and since our analysis suggests that up to 90% is seen, a strong baseline should address this seen portion. 4Recent work does include some seen data. Bergsma et al. (2008) test pairs that fall below a mutual information threshold (might include some seen pairs), and Erk (2007) selects a subset of roles in FrameNet (Baker et al., 1998) to test and uses all labeled instances within this subset (unclear what portion of subset of data is seen). Neither evaluates all of the seen data, however. 448 We propose a conditional probability baseline: P(n|vd) = ( C(vd,n) C(vd,∗) if C(vd, n) > 0 0 otherwise where C(vd, n) is the number of times the head word n was seen as an argument to the predicate v, and C(vd, ∗) is the number of times vd was seen with any argument. Given a test (vd, n) and its confounder (vd, n′), choose n if P(n|vd) > P(n′|vd), and n′ otherwise. If P(n|vd) = P(n′|vd), randomly choose one. Lapata et al. (1999) showed that corpus frequency and conditional probability correlate with human decisions of adjective-noun plausibility, and Dagan et al. (1999) appear to propose a very similar baseline for verb-noun selectional preferences, but the paper evaluates unseen data, and so the conditional probability model is not studied. We later analyze this baseline against a more complicated smoothing approach. 5.2 A Web Baseline If conditional probability is a reasonable baseline, better performance may just require more data. Keller and Lapata (2003) proposed using the web for this task, querying for specific phrases like ‘Verb Det N’ to find syntactic objects. Such a web corpus would be attractive, but we’d like to find subjects and prepositional objects as well as objects, and also ideally we don’t want to limit ourselves to patterns. Since parsing the web is unrealistic, a reasonable compromise is to make rough counts when pairs of words occur in close proximity to each other. Using the Google n-gram corpus, we recorded all verb-noun co-occurrences, defined by appearing in any order in the same n-gram, up to and including 5-grams. For instance, the test pair (throwsubject, ball) is considered seen if there exists an n-gram such that throw and ball are both included. We count all such occurrences for all verb-noun pairs. We also avoided over-counting co-occurrences in lower order n-grams that appear again in 4 or 5-grams. This crude method of counting has obvious drawbacks. Subjects are not distinguished from objects and nouns may not be actual arguments of the verb. However, it is a simple baseline to implement with these freely available counts. Thus, we use conditional probability as defined in the previous section, but define the count C(vd, n) as the number of times v and n (ignoring d) appear in the same n-gram. 5.3 Smoothing Model We implemented the current state-of-the-art smoothing model of Erk (2007). The model is based on the idea that the arguments of a particular verb slot tend to be similar to each other. Given two potential arguments for a verb, the correct one should correlate higher with the arguments observed with the verb during training. Formally, given a verb v and a grammatical dependency d, the score for a noun n is defined: Svd(n) = X w∈Seen(vd) sim(n, w) ∗C(vd, w) (1) where sim(n, w) is a noun-noun similarity score, Seen(vd) is the set of seen head words filling the slot vd during training, and C(vd, n) is the number of times the noun n was seen filling the slot vd The similarity score sim(n, w) can thus be one of many vector-based similarity metrics5. We evaluate both Jaccard and Cosine similarity scores in this paper, but the difference between the two is small. 6 Experiments Our training data is the NYT section of the Gigaword Corpus, parsed into dependency graphs. We extract all (vd, n) pairs from the graph, as described in section 3. We randomly chose 9 documents from the year 2001 for a development set, and 41 documents for testing. The test set consisted of 6767 (vd, n) pairs. All verbs and nouns are stemmed, and the development and test documents were isolated from training. 6.1 Varying Training Size We repeated the experiments with three different training sizes to analyze the effect data size has on performance: • Train x1: Year 2001 of the NYT portion of the Gigaword Corpus. After removing duplicate documents, it contains approximately 110 million tokens, comparable to the 100 million tokens in the BNC corpus. 5A similar type of smoothing was proposed in earlier work by Dagan et al. (1999). A noun is represented by a vector of verb slots and the number of times it is observed filling each slot. 449 • Train x2: Years 2001 and 2002 of the NYT portion of the Gigaword Corpus, containing approximately 225 million tokens. • Train x10: The entire NYT portion of Gigaword (approximately 1.2 billion tokens). It is an order of magnitude larger than Train x1. 6.2 Varying the Confounder We generated three different confounder sets based on word corpus frequency from the 41 test documents. Frequency was determined by counting all tokens with noun POS tags. As motivated in section 4, we use the following approaches: • Random: choose a random confounder from the set of nouns that fall within some broad corpus frequency range. We set our range to eliminate (approximately) the top 100 most frequent nouns, but otherwise arbitrarily set the lower range as previous work seems to do. The final range was [30, 400000]. • Buckets: all nouns are bucketed based on their corpus frequencies6. Given a test pair (vd, n), choose the bucket in which n belongs and randomly select a confounder n′ from that bucket. • Neighbor: sort all seen nouns by frequency and choose the confounder n′ that is the nearest neighbor of n with greater frequency. 6.3 Model Implementation None of the models can make a decision if they identically score both potential arguments (most often true when both arguments were not seen with the verb in training). As a result, we extend all models to randomly guess (50% performance) on pairs they cannot answer. The conditional probability is reported as Baseline. For the web baseline (reported as Google), we stemmed all words in the Google n-grams and counted every verb v and noun n that appear in Gigaword. Given two nouns, the noun with the higher co-occurrence count with the verb is chosen. As with the other models, if the two nouns have the same counts, it randomly guesses. The smoothing model is named Erk in the results with both Jaccard and Cosine as the similarity metric. Due to the large vector representations of the nouns, it is computationally wise to 6We used frequency buckets of 4, 10, 25, 200, 1000, >1000. Adding more buckets moves the evaluation closer to Neighbor, less is closer to Random. trim their vectors, but also important to do so for best performance. A noun’s representative vector consists of verb slots and the number of times the noun was seen in each slot. We removed any verb slot not seen more than x times, where x varied based on all three factors: the dataset, confounder choice, and similarity metric. We optimized x on the development data with a linear search, and used that cutoff on each test. Finally, we trimmed any vectors over 2000 in size to reduce the computational complexity. Removing this strict cutoff appears to have little effect on the results. Finally, we report backoff scores for Google and Erk. These consist of always choosing the Baseline if it returns an answer (not a guessed unseen answer), and then backing off to the Google/Erk result for Baseline unknowns. These are labeled Backoff Google and Backoff Erk. 7 Results Results are given for the two dimensions: confounder choice and training size. Statistical significance tests were calculated using the approximate randomization test (Yeh, 2000) with 1000 iterations. Figure 4 shows the performance change over the different confounder methods. Train x2 was used for training. Each model follows the same progression: it performs extremely well on the random test set, worse on buckets, and the lowest on the nearest neighbor. The conditional probability Baseline falls from 91.5 to 79.5, a 12% absolute drop from completely random to neighboring frequency. The Erk smoothing model falls 27% from 93.9 to 68.1. The Google model generally performs the worst on all sets, but its 74.3% performance with random confounders is significantly better than a 50-50 random choice. This is notable since the Google model only requires n-gram counts to implement. The Backoff Erk model is the best, using the Baseline for the majority of decisions and backing off to the Erk smoothing model when the Baseline cannot answer. Figure 5 (shown on the next page) varies the training size. We show results for both Bucket Frequencies and Neighbor Frequencies. The only difference between columns is the amount of training data. As expected, the Baseline improves as the training size is increased. The Erk model, somewhat surprisingly, shows no continual gain with more training data. The Jaccard and Cosine simi450 Varying the Confounder Frequency Random Buckets Neighbor Baseline 91.5 89.1 79.5 Erk-Jaccard 93.9* 82.7* 68.1* Erk-Cosine 91.2 81.8* 65.3* Google 74.3* 70.4* 59.4* Backoff Erk 96.6* 91.8* 80.8* Backoff Goog 92.7† 89.7 79.8 Figure 4: Trained on two years of NYT data (Train x2). Accuracy of the models on the same NYT test documents, but with three different ways of choosing the confounders. * indicates statistical significance with the column’s Baseline at the p < 0.01 level, † at p < 0.05. Random is overly optimistic, reporting performance far above more conservative (selective) confounder choices. Baseline Details Train Train x2 Train x10 Precision 96.1 95.5* 95.0† Accuracy 78.2 82.0* 88.1* Accuracy +50% 87.5 89.1* 91.7* Figure 6: Results from the buckets confounder test set. Baseline precision, accuracy (the same as recall), and accuracy when you randomly guess the tests that Baseline does not answer. All numbers are statistically significant * with p-value < 0.01 from the number to their left. larity scores perform similarly in their model. The Baseline achieves the highest accuracies (91.7% and 81.2%) with Train x10, outperforming the best Erk model by 5.2% and 13.1% absolute on buckets and nearest neighbor respectively. The backoff models improve the baseline by just under 1%. The Google n-gram backoff model is almost as good as backing off to the Erk smoothing model. Finally, figure 6 shows the Baseline’s precision and overall accuracy. Accuracy is the same as recall when the model does not guess between pseudo words that have the same conditional probabilities. Accuracy +50% (the full Baseline in all other figures) shows the gain from randomly choosing one of the two words when uncertain. Precision is extremely high. 8 Discussion Confounder Choice: Performance is strongly influenced by the method used when choosing confounders. This is consistent with findings for WSD that corpus frequency choices alter the task (Gaustad, 2001; Nakov and Hearst, 2003). Our results show the gradation of performance as one moves across the spectrum from completely random to closest in frequency. The Erk model dropped 27%, Google 15%, and our baseline 12%. The overly optimistic performance on random data suggests using the nearest neighbor approach for experiments. Nearest neighbor avoids evaluating on ‘easy’ datasets, and our baseline (at 79.5%) still provides room for improvement. But perhaps just as important, the nearest neighbor approach facilitates the most reproducibile results in experiments since there is little ambiguity in how the confounder is selected. Realistic Confounders: Despite its overoptimism, the random approach to confounder selection may be the correct approach in some circumstances. For some tasks that need selectional preferences, random confounders may be more realistic. It’s possible, for example, that the options in a PP-attachment task might be distributed more like the random rather than nearest neighbor models. In any case, this is difficult to decide without a specific application in mind. Absent such specific motiviation, a nearest neighbor approach is the most conservative, and has the advantage of creating a reproducible experiment, whereas random choice can vary across design. Training Size: Training data improves the conditional probability baseline, but does not help the smoothing model. Figure 5 shows a lack of improvement across training sizes for both jaccard and cosine implementations of the Erk model. The Train x1 size is approximately the same size used in Erk (2007), although on a different corpus. We optimized argument cutoffs for each training size, but the model still appears to suffer from additional noise that the conditional probability baseline does not. This may suggest that observing a test argument with a verb in training is more reliable than a smoothing model that compares all training arguments against that test example. High Precision Baseline: Our conditional probability baseline is very precise. It outperforms the smoothed similarity based Erk model and gives high results across tests. The only combination when Erk is better is when the training data includes just one year (one twelfth of the NYT section) and the confounder is chosen com451 Varying the Training Size Bucket Frequency Neighbor Frequency Train x1 Train x2 Train x10 Train x1 Train x2 Train x10 Baseline 87.5 89.1 91.7 78.4 79.5 81.2 Erk-Jaccard 86.5* 82.7* 83.1* 66.8* 68.1* 65.5* Erk-Cosine 82.1* 81.8* 81.1* 66.1* 65.3* 65.7* Google 70.4* 59.4* Backoff Erk 92.6* 91.8* 92.6* 79.4* 80.8* 81.7* Backoff Google 88.6 89.7 91.9† 78.7 79.8 81.2 Figure 5: Accuracy of varying NYT training sizes. The left and right tables represent two confounder choices: choose the confounder with frequency buckets, and choose by nearest frequency neighbor. Trainx1 starts with year 2001 of NYT data, Trainx2 doubles the size, and Trainx10 is 10 times larger. * indicates statistical significance with the column’s Baseline at the p < 0.01 level, † at p < 0.05. pletely randomly. These results appear consistent with Erk (2007) because that work used the BNC corpus (the same size as one year of our data) and Erk chose confounders randomly within a broad frequency range. Our reported results include every (vd, n) in the data, not a subset of particular semantic roles. Our reported 93.9% for ErkJaccard is also significantly higher than their reported 81.4%, but this could be due to the random choices we made for confounders, or most likely corpus differences between Gigaword and the subset of FrameNet they evaluated. Ultimately we have found that complex models for selectional preferences may not be necessary, depending on the task. The higher computational needs of smoothing approaches are best for backing off when unseen data is encountered. Conditional probability is the best choice for seen examples. Further, analysis of the data shows that as more training data is made available, the seen examples make up a much larger portion of the test data. Conditional probability is thus a very strong starting point if selectional preferences are an internal piece to a larger application, such as semantic role labeling or parsing. Perhaps most important, these results illustrate the disparity in performance that can come about when designing a pseudo-word disambiguation evaluation. It is crucially important to be clear during evaluations about how the confounder was generated. We suggest the approach of sorting nouns by frequency and using a neighbor as the confounder. This will also help avoid evaluations that produce overly optimistic results. 9 Conclusion Current performance on various natural language tasks is being judged and published based on pseudo-word evaluations. It is thus important to have a clear understanding of the evaluation’s characteristics. We have shown that the evaluation is strongly affected by confounder choice, suggesting a nearest frequency neighbor approach to provide the most reproducible performance and avoid overly optimistic results. We have shown that evaluating entire documents instead of subsets of the data produces vastly different results. We presented a conditional probability baseline that is both novel to the pseudo-word disambiguation task and strongly outperforms state-of-the-art models on entire documents. We hope this provides a new reference point to the pseudo-word disambiguation task, and enables selectional preference models whose performance on the task similarly transfers to larger NLP applications. Acknowledgments This work was supported by the National Science Foundation IIS-0811974, and the Air Force Research Laboratory (AFRL) under prime contract no. FA8750-09-C-0181. Any opinions, ndings, and conclusion or recommendations expressed in this material are those of the authors and do not necessarily reect the view of the AFRL. Thanks to Sebastian Pad´o, the Stanford NLP Group, and the anonymous reviewers for very helpful suggestions. 452 References Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998. The Berkeley FrameNet project. In Christian Boitet and Pete Whitelock, editors, ACL-98, pages 86–90, San Francisco, California. Morgan Kaufmann Publishers. Shane Bergsma, Dekang Lin, and Randy Goebel. 2008. Discriminative learning of selectional preference from unlabeled text. In Empirical Methods in Natural Language Processing, pages 59–68, Honolulu, Hawaii. Lou Burnard. 1995. User Reference Guide for the British National Corpus. Oxford University Press, Oxford. Ido Dagan, Lillian Lee, and Fernando C. N. Pereira. 1999. Similarity-based models of word cooccurrence probabilities. Machine Learning, 34(1):43– 69. Katrin Erk. 2007. A simple, similarity-based model for selectional preferences. In 45th Annual Meeting of the Association for Computational Linguistics, Prague, Czech Republic. William A. Gale, Kenneth W. Church, and David Yarowsky. 1992. Work on statistical methods for word sense disambiguation. In AAAI Fall Symposium on Probabilistic Approaches to Natural Language, pages 54–60. Tanja Gaustad. 2001. Statistical corpus-based word sense disambiguation: Pseudowords vs. real ambiguous words. In 39th Annual Meeting of the Association for Computational Linguistics - Student Research Workshop. David Graff. 2002. English Gigaword. Linguistic Data Consortium. Frank Keller and Mirella Lapata. 2003. Using the web to obtain frequencies for unseen bigrams. Computational Linguistics, 29(3):459–484. Maria Lapata, Scott McDonald, and Frank Keller. 1999. Determinants of adjective-noun plausibility. In European Chapter of the Association for Computational Linguistics (EACL). Preslav I. Nakov and Marti A. Hearst. 2003. Categorybased pseudowords. In Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology, pages 67–69, Edmonton, Canada. Fernando Pereira, Naftali Tishby, and Lillian Lee. 1993. Distributional clustering of english words. In 31st Annual Meeting of the Association for Computational Linguistics, pages 183–190, Columbus, Ohio. Mats Rooth, Stefan Riezler, Detlef Prescher, Glenn Carroll, and Franz Beil. 1999. Inducing a semantically annotated lexicon via em-based clustering. In 37th Annual Meeting of the Association for Computational Linguistics, pages 104–111. Hinrich Schutze. 1992. Context space. In AAAI Fall Symposium on Probabilistic Approaches to Natural Language, pages 113–120. Alexander Yeh. 2000. More accurate tests for the statistical significance of result differences. In International Conference on Computational Linguistics (COLING). Beat Zapirain, Eneko Agirre, and Llus Mrquez. 2009. Generalizing over lexical features: Selectional preferences for semantic role classification. In Joint Conference of the 47th Annual Meeting of the Association for Computational Linguistics and the 4th International Joint Conference on Natural Language Processing, Singapore. 453
|
2010
|
46
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 454–464, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Syntax-to-Morphology Mapping in Factored Phrase-Based Statistical Machine Translation from English to Turkish Reyyan Yeniterzi Language Technologies Institute Carnegie Mellon University Pittsburgh, PA, 15213, USA [email protected] Kemal Oflazer Computer Science Carnegie Mellon University-Qatar PO Box 24866, Doha, Qatar [email protected] Abstract We present a novel scheme to apply factored phrase-based SMT to a language pair with very disparate morphological structures. Our approach relies on syntactic analysis on the source side (English) and then encodes a wide variety of local and non-local syntactic structures as complex structural tags which appear as additional factors in the training data. On the target side (Turkish), we only perform morphological analysis and disambiguation but treat the complete complex morphological tag as a factor, instead of separating morphemes. We incrementally explore capturing various syntactic substructures as complex tags on the English side, and evaluate how our translations improve in BLEU scores. Our maximal set of source and target side transformations, coupled with some additional techniques, provide an 39% relative improvement from a baseline 17.08 to 23.78 BLEU, all averaged over 10 training and test sets. Now that the syntactic analysis on the English side is available, we also experiment with more long distance constituent reordering to bring the English constituent order close to Turkish, but find that these transformations do not provide any additional consistent tangible gains when averaged over the 10 sets. 1 Introduction Statistical machine translation into a morphologically complex language such as Turkish, Finnish or Arabic, involves the generation of target words with the proper morphology, in addition to properly ordering the target words. Earlier work on translation from English to Turkish (Oflazer and Durgar-El-Kahlout, 2007; Oflazer, 2008; DurgarEl-Kahlout and Oflazer, 2010) has used an approach which relied on identifying the contextually correct parts-of-speech, roots and any morphemes on the English side, and the complete sequence of roots and overt derivational and inflectional morphemes for each word on the Turkish side. Once these were identified as separate tokens, they were then used as “words” in a standard phrase-based framework (Koehn et al., 2003). They have reported that, given the typical complexity of Turkish words, there was a substantial percentage of words whose morphological structure was incorrect: either the morphemes were not applicable for the part-of-speech category of the root word selected, or the morphemes were in the wrong order. The main reason given for these problems was that the same statistical translation, reordering and language modeling mechanisms were being employed to both determine the morphological structure of the words and, at the same time, get the global order of the words correct. Even though a significant improvement of a standard word-based baseline was achieved, further analysis hinted at a direction where morphology and syntax on the Turkish side had to be dealt with using separate mechanisms. Motivated by the observation that many local and some nonlocal syntactic structures in English essentially map to morphologically complex words in Turkish, we present a radically different approach which does not segment Turkish words into morphemes, but uses a representation equivalent to the full word form. On the English side, we rely on a full syntactic analysis using a dependency parser. This analysis then lets us abstract and encode many local and some nonlocal syntactic structures as complex tags (dynamically, as opposed to the static complex tags as proposed by Birch et al. (2007) and Hassan et al. (2007)). Thus 454 we can bring the representation of English syntax closer to the Turkish morphosyntax. Such an approach enables the following: (i) Driven by the pattern of morphological structures of full word forms on the Turkish side represented as root words and complex tags, we can identify and reorganize phrases on the English side, to “align” English syntax to Turkish morphology wherever possible. (ii) Continuous and discontinuous variants of certain (syntactic) phrases can be conflated during the SMT phrase extraction process. (iii) The length of the English sentences can be dramatically reduced, as most function words encoding syntax are now abstracted into complex tags. (iv) The representation of both the source and the target sides of the parallel corpus can now be mostly normalized. This facilitates the use of factored phrase-based translation that was not previously applicable due to the morphological complexity on the target side and mismatch between source and target morphologies. We find that with the full set of syntax-tomorphology transformations and some additional techniques we can get about 39% relative improvement in BLEU scores over a word-based baseline and about 28% improvement of a factored baseline, all experiments being done over 10 training and test sets. We also find that further constituent reordering taking advantage of the syntactic analysis of the source side, does not provide tangible improvements when averaged over the 10 data sets. This paper is organized as follows: Section 2 presents the basic idea behind syntaxto-morphology alignment. Section 3 describes our experimental set-up and presents results from a sequence of incremental syntax-to-morphology transformations, and additional techniques. Section 4 summarizes our constituent reordering experiments and their results. Section 5 presents a review of related work and situates our approach. We assume that the reader is familiar with the basics of phrase-based statistical machine translation (Koehn et al., 2003) and factored statistical machine translation (Koehn and Hoang, 2007). 2 Syntax-to-Morphology Mapping In this section, we describe how we map between certain source language syntactic structures and target words with complex morphological structures. At the top of Figure 1, we see a pair of (syntactic) phrases, where we have (positionally) aligned the words that should be translated to each other. We can note that the function words on and Figure 1: Transformation of an English prepositional phrase their are not really aligned to any of the Turkish words as they really correspond to two of the morphemes of the last Turkish word. When we tag and syntactically analyze the English side into dependency relations, and morphologically analyze and disambiguate the Turkish phrase, we get the representation in the middle of Figure 1, where we have co-indexed components that should map to each other, and some of the syntactic relations that the function words are involved in are marked with dependency links.1 The basic idea in our approach is to take various function words on the English side, whose syntactic relationships are identified by the parser, and then package them as complex tags on the related content words. So, in this example, if we move the first two function words from the English side and attach as syntactic tags to the word they are in dependency relation with, we get the aligned representation at the bottom of Figure 1.2,3 Here we can note that all root words and tags that correspond to each other are nicely structured and are in the same relative order. In fact, we can treat each token as being composed of two factors: the roots and the accompanying tags. The tags on the Turkish side encode morphosyntactic information encoded in the morphology of the words, while the 1The meanings of various tags are as follows: Dependency Labels: PMOD - Preposition Modifier; POS - Possessive. Part-of-Speech Tags for the English words: +IN Preposition; +PRP$ - Possessive Pronoun; +JJ - Adjective; +NN - Noun; +NNS - Plural Noun. Morphological Feature Tags in the Turkish Sentence: +A3pl - 3rd person plural; +P3sg - 3rd person singular possessive; +Loc - Locative case. Note that we mark an English plural noun as +NN NNS to indicate that the root is a noun and there is a plural morpheme on it. Note also that economic is also related to relations but we are not interested in such content words and their relations. 2We use to prefix such syntactic tags on the English side. 3The order is important in that we would like to attach the same sequence of function words in the same order so that the resulting tags on the English side are the same. 455 (complex) tags on the English side encode local (and sometimes, non-local) syntactic information. Furthermore, we can see that before the transformations, the English side has 4 words, while afterwards it has only 2 words. We find (and elaborate later) that this reduction in the English side of the training corpus, in general, is about 30%, and is correlated with improved BLEU scores. We believe the removal of many function words and their folding into complex tags (which do not get involved in GIZA++ alignment – we only align the root words) seems to improve alignment as there are less number of “words” to worry about during that process.4 Another interesting side effect of this representation is the following. As the complex syntactic tags on the English side are based on syntactic relations and not necessarily positional proximity, the tag for relations in a phrase like in their cultural, historical and economic relations would be exactly the same as above. Thus phrase extraction algorithms can conflate all constructs like in their . . . economic relations as one phrase, regardless of the intervening modifiers, assuming that parser does its job properly. Not all cases can be captured as cleanly as the example above, but most transformations capture local and nonlocal syntax involving many function words and then encode syntax with complex tags resembling full morphological tags on the Turkish side. These transformations, however, are not meant to perform sentence level constituent reordering on the English side. We explore these later. We developed set of about 20 linguisticallymotivated syntax-to-morphology transformations which had variants parameterized depending on what, for instance, the preposition or the adverbial was, and how they map to morphological structure on the Turkish side. For instance, one general rule handles cases like while . . . verb and if ...verb etc., mapping these to appropriate complex tags. It is also possible that multiple transformations can apply to generate a single English complex tag: a portion of the tag can come from a verb complex transformation, and another from an adverbial phrase transformation involving a marked such as while. Our transformations handle the following cases: • Prepositions attach to the head-word of their 4Fraser (2009) uses the first four letters of German words after morphological stripping and compound decomposition to help with alignment in German to English and reverse translation. complement noun phrase as a component in its complex tag. • Possessive pronouns attach to the head-word they specify. • The possessive markers following a noun (separated by the tokenizer) attached to the noun. • Auxiliary verbs and negation markers attach to the lexical verb that they form a verb complex with. • Modals attach to the lexical verb they modify. • Forms of be used as predicates with adjectival or nominal dependents attach to the dependent. • Forms of be or have used to form passive voice with past participle verbs, and forms of be used with -ing verbs to form present continuous verbs, attach to the verb. • Various adverbial clauses formed with if, while, when, etc., are reorganized so that these markers attach to the head verb of the clause. As stated earlier, these rules are linguistically motivated and are based on the morphological structure of the target language words. Hence for different target languages these rules will be different. The rules recognize various local and nonlocal syntactic structures in the source side parse tree that correspond to complex morphological of target words and then remove source function words folding them into complex tags. For instance, the transformations in Figure 1 are handled by scripts that process Malt Parser’s dependency structure output and that essentially implement the following sequence of rules expressed as pseudo code: 1) if (<Y>+PRP$ POS <Z>+NN<TAG>) then { APPEND <Y>+PRP$ TO <Z>+NN<TAG> REMOVE <Y>+PRP$ } 2) if (<X>+IN PMOD <Z>+NN<TAG>) then { APPEND <X>+IN TO <Z>+NN<TAG> REMOVE <X>+IN } Here <X>, <Y> and <Z> can be considered as Prolog like-variables that bind to patterns (mostly root words), and the conditions check for specified dependency relations (e.g., PMOD) between the left and the right sides. When the condition is satisfied, then the part matching the function word is removed and its syntactic information is appended to form the complex tag on the noun (<TAG> would either match null string or any previously appended function word markers.)5 5We outline two additional rules later when we see a more complex example in Figure 2. 456 There are several other rules that handle more mundane cases of date and time constructions (for which, the part of the date construct which the parser attaches a preposition, is usually different than the part on the Turkish side that gets inflected with case markers, and these have to be reconciled by overriding the parser output.) The next section presents an example of a sentence with multiple transformations applied, after discussing the preprocessing steps. 3 Experimental Setup and Results 3.1 Data Preparation We worked on an English-Turkish parallel corpus which consists of approximately 50K sentences with an average of 23 words in English sentences and 18 words in Turkish sentences. This is the same parallel data that has been used in earlier SMT work on Turkish (Durgar-El-Kahlout and Oflazer, 2010). Let’s assume we have the following pair of parallel sentences: E: if a request is made orally the authority must make a record of it T: istek s¨ozl¨u olarak yapılmıs¸sa yetkili makam bunu kaydetmelidir On the English side of the data, we use the Stanford Log-Linear Tagger (Toutanova et al., 2003), to tag the text with Penn Treebank Tagset. On the Turkish side, we perform a full morphological analysis, (Oflazer, 1994), and morphological disambiguation (Yuret and T¨ure, 2006) to select the contextually salient interpretation of words. We then remove any morphological features that are not explicitly marked by an overt morpheme.6 So for both sides we get, E: if+IN a+DT request+NN is+VBZ made+VBN orally+RB the+DT authority+NN must+MD make+VB a+DT record+NN of+IN it+PRP T: istek+Noun s¨ozl¨u+Adj olarak+Verb+ByDoingSo yap+Verb+Pass+Narr+Cond yetkili+Adj makam+Noun bu+Pron+Acc kaydet+Verb+Neces+Cop Finally we parse the English sentences using MaltParser (Nivre et al., 2007), which gives us labeled dependency parses. On the output of the parser, we make one more transformation. We replace each word with its root, and possibly add an additional tag for any inflectional information conveyed by overt morphemes or exceptional forms. This is done by running the TreeTagger (Schmid, 1994) on the English side which provides the roots in addition to the tags, and then carrying over this information to the parser output. For example, is is tagged as be+VB VBZ, made is tagged as make+VB VBN, and a word like books is tagged 6For example, the morphological analyzer outputs +A3sg to mark a singular noun, if there is no explicit plural morpheme. Such markers are removed. as book+NN NNS (and not as books+NNS). On the Turkish side, each marker with a preceding + is a morphological feature. The first marker is the part-of-speech tag of the root and the remainder are the overt inflectional and derivational markers of the word. For example, the analysis kitap+Noun+P2pl+A3pl+Gen for a word like kitap-lar-ınız-ın7 (of your books) represents the root kitap (book), a Noun, with third person plural agreement A3pl, second person plural possessive agreement, P2pl and genitive case Gen. The sentence representations in the middle part of Figure 2 show these sentences with some of the dependency relations (relevant to our transformations) extracted by the parser, explicitly marked as labeled links. The representation at the bottom of this figure (except for the co-indexation markers) corresponds to the final transformed form of the parallel training and test data. The co-indexation is meant to show which root words on one side map to which on the other side. Ultimately we would want the alignment process to uncover the root word alignments indicated here. We can also note that the initial form of the English sentence has 14 words and the final form after transformations, has 7 words (with complex tags).8 3.2 Experiments We evaluated the impact of the transformations in factored phrase-based SMT with an EnglishTurkish data set which consists of 52712 parallel sentences. In order to have more confidence in the impact of our transformations, we randomly generated 10 training, test and tune set combinations. For each combination, the latter two were 1000 sentences each and the remaining 50712 sentences were used as training sets.9,10 We performed our experiments with the Moses toolkit (Koehn et al., 2007). In order to encourage long distance reordering in the decoder, we used a distortion limit of -1 and a distortion weight of 7- shows surface morpheme boundaries. 8We could give two more examples of rules to process the if-clause in the example in Figure 2. These rules would be applied sequentially: The first rule recognizes the passive construction mediated by be+VB<AGR> forming a verb complex (VC) with <Y>+VB_VBN and appends the former to the complex tag on the latter and then deletes the former token. The second rule then recognizes <X>+IN relating to <Y>+VB<TAGS>with VMOD and appends the former to the complex tag on the latter and then deletes the former token. 9The tune set was not used in this work but reserved for future work so that meaningful comparisons could be made. 10It is possible that the 10 test sets are not mutually exclusive. 457 Figure 2: An English-Turkish sentence pair with multiple transformations applied 0.1.11 We did not use MERT to further optimize our model.12 For evaluation, we used the BLEU metric (Papineni et al., 2001). Each experiment was repeated over the 10 data sets. Wherever meaningful, we report the average BLEU scores over 10 data sets along with the maximum and minimum values and the standard deviation. 11These allow and do not penalize unlimited distortions. 12The experience with MERT for this language pair has not been very positive. Earlier work on Turkish indicates that starting with default Moses parameters and applying MERT to the resulting model does not even come close to the performance of the model with those two specific parameters set as such (distortion limit -1 and distortion weight 0.1), most likely because the default parameters do not encourage the range of distortions that are needed to deal with the constituent order differences. Earlier work on Turkish also shows that even when the weight-d parameter is initialized with this specific value, the space explored for distortion weight and other parameters do not produce any improvements on the test set, even though MERT claims there are improvements on the tune set. The other practical reasons for not using MERT were the following: at the time we performed this work, the discussion thread at http://www.mail-archive. com/[email protected]/msg01012.html indicated that MERT was not tested on multiple factors. The discussion thread at http://www.mail-archive. com/[email protected]/msg00262.html claimed that MERT does not help very much with factored models. With these observations, we opted not to experiment with MERT with the multiple factor approach we employed, given that it would be risky and time consuming to run MERT needed for 10 different models and then not necessarily see any (consistent) improvements. MERT however is orthogonal to the improvements we achieve here and can always be applied on top of the best model we get. 3.2.1 The Baseline Systems As a baseline system, we built a standard phrasebased system, using the surface forms of the words without any transformations, and with a 3-gram LM in the decoder. We also built a second baseline system with a factored model. Instead of using just the surface form of the word, we included the root, part-of-speech and morphological tag information into the corpus as additional factors alongside the surface form.13 Thus, a token is represented with three factors as Surface|Root|Tags where Tags are complex tags on the English side, and morphological tags on the Turkish side.14 Moses lets word alignment to align over any of the factors. We aligned our training sets using only the root factor to conflate statistics from different forms of the same root. The rest of the factors are then automatically assumed to be aligned, based on the root alignment. Furthermore, in factored models, we can employ different language models for different factors. For the initial set of experiments we used 3-gram LMs for all the factors. For factored decoding, we employed a model whereby we let the decoder translate a surface form directly, but if/when that fails, the decoder can back-off with a generation model that builds a target word from independent translations of the root and tags. 13In Moses, factors are separated by a ‘|’ symbol. 14Concatenating Root and Tags gives the Surface form, in that the surface is unique given this concatenation. 458 The results of our baseline models are given in top two rows of Table 1. As expected, the wordbased baseline performs worse than the factored baseline. We believe that the use of multiple language models (some much less sparse than the surface LM) in the factored baseline is the main reason for the improvement. 3.2.2 Applying Syntax-to-Morphology Mapping Transformations To gauge the effects of transformations separately, we first performed them in batches on the English side. These batches were (i) transformations involving nouns and adjectives (Noun+Adj), (ii) transformations involving verbs (Verb), (iii) transformations involving adverbs (Adv), and (iv) transformations involving verbs and adverbs (Verb+Adv). We also performed one set of transformations on the Turkish side. In general, English prepositions translate as case markers on Turkish nouns. However, there are quite a number of lexical postpositions in Turkish which also correspond to English prepositions. To normalize these with the handling of case-markers, we treated these postpositions as if they were case-markers and attached them to the immediately preceding noun, and then aligned the resulting training data (PostP).15 The results of these experiments are presented in Table 1. We can observe that the combined syntax-to-morphology transformations on the source side provide a substantial improvement by themselves and a simple target side transformation on top of those provides a further boost to 21.96 BLEU which represents a 28.57% relative improvement over the word-based baseline and a 18.00% relative improvement over the factored baseline. Experiment Ave. STD Max. Min. Baseline 17.08 0.60 17.99 15.97 Factored Baseline 18.61 0.76 19.41 16.80 Noun+Adj 21.33 0.62 22.27 20.05 Verb 19.41 0.62 20.19 17.99 Adv 18.62 0.58 19.24 17.30 Verb+Adv 19.42 0.59 20.17 18.13 Noun+Adj 21.67 0.72 22.66 20.38 +Verb+Adv Noun+Adj+Verb 21.96 0.72 22.91 20.67 +Adv+PostP Table 1: BLEU scores for a variety of transformation combinations We can see that every transformation improves 15Note than in this case, the translations would be generated in the same format, but we then split such postpositions from the words they are attached to, during decoding, and then evaluate the BLEU score. the baseline system and the highest performance is attained when all transformations are performed. However when we take a closer look at the individual transformations performed on English side, we observe that not all of them have the same effect. While Noun+Adj transformations give us an increase of 2.73 BLEU points, Verbs improve the result by only 0.8 points and improvement with Adverbs is even lower. To understand why we get such a difference, we investigated the correlation of the decrease in the number of tokens on both sides of the parallel data, with the change in BLEU scores. The graph in Figure 3 plots the BLEU scores and the number of tokens in the two sides of the training data as the data is modified with transformations. We can see that as the number of tokens in English decrease, the BLEU score increases. In order to measure the relationship between these two variables statistically, we performed a correlation analysis and found that there is a strong negative correlation of -0.99 between the BLEU score and the number of English tokens. We can also note that the largest reduction in the number of tokens comes with the application of the Noun+Adj transformations, which correlates with the largest increase in BLEU score. It is also interesting to look at the n-gram precision components of the BLEU scores (again averaged). In Table 2, we list these for words (actual BLEU), roots (BLEU-R) to see how effective we are in getting the root words right, and morphological tags, (BLEU-M), to see how effective we are in getting just the morphosyntax right. It 1-gr. 2-gr. 3-gr. 4-gr. BLEU 21.96 55.73 27.86 16.61 10.68 BLEU-R 27.63 68.60 35.49 21.08 13.47 BLEU-M 27.93 67.41 37.27 21.40 13.41 Table 2: Details of Word, Root and Morphology BLEU Scores seems we are getting almost 69% of the root words and 68% of the morphological tags correct, but not necessarily getting the combination equally as good, since only about 56% of the full word forms are correct. One possible way to address is to use longer distance constraints on the morphological tag factors, to see if we can select them better. 3.2.3 Experiments with higher-order language models Factored phrase-based SMT allows the use of multiple language models for the target side, for different factors during decoding. Since the number of possible distinct morphological tags (the morphological tag vocabulary size) in our training data 459 Figure 3: BLEU scores vs number of tokens in the training sets (about 3700) is small compared to distinct number of surface forms (about 52K) and distinct roots (about 15K including numbers), it makes sense to investigate the contribution of higher order n-gram language models for the morphological tag factor on the target side, to see if we can address the observation in the previous section. Using the data transformed with Noun+Adj+Verb+Adv+PostP transformations which previously gave us the best results overall, we experimented with using higher order models (4-grams to 9-grams) during decoding, for the morphological tag factor models, keeping the surface and root models at 3-gram. We observed that for all the 10 data sets, the improvements were consistent for up to 8-gram. The BLEU with the 8-gram for only the morphological tag factor averaged over the 10 data sets was 22.61 (max: 23.66, min: 21.37, std: 0.72) compared to the 21.96 in Table 1. Using a 4gram root LM, considerably less sparse than word forms but more sparse that tags, we get a BLEU score of 22.80 (max: 24.07, min: 21.57, std: 0.85). The details of the various BLEU scores are shown in the two halves of Table 3. It seems that larger n-gram LMs contribute to the larger n-gram precisions contributing to the BLEU but not to the unigram precision. 3-gram root LM 1-gr. 2-gr. 3-gr. 4-gr. BLEU 22.61 55.85 28.21 17.16 11.36 BLEU-R 28.21 68.67 35.80 21.55 14.07 BLEU-M 28.68 67.50 37.59 22.02 14.22 4-gram root LM 1-gr. 2-gr. 3-gr. 4-gr. BLEU 22.80 55.85 28.39 17.34 11.54 BLEU-R 28.48 68.68 35.97 21.79 14.35 BLEU-M 28.82 67.49 37.63 22.17 14.40 Table 3: Details of Word, Root and Morphology BLEU Scores, with 8-gram tag LM and 3/4-gram root LMs 3.2.4 Augmenting the Training Data In order to alleviate the lack of large scale parallel corpora for the English–Turkish language pair, we experimented with augmenting the training data with reliable phrase pairs obtained from a previous alignment. Phrase table entries for the surface factors produced by Moses after it does an alignment on the roots, contain the English (e) and Turkish (t) parts of a pair of aligned phrases, and the probabilities, p(e|t), the conditional probability that the English phrase is e given that the Turkish phrase is t, and p(t|e), the conditional probability that the Turkish phrase is t given the English phrase is e. Among these phrase table entries, those with p(e|t) ≈p(t|e) and p(t|e) + p(e|t) larger than some threshold, can be considered as reliable mutual translations, in that they mostly translate to each other and not much to others. We extracted 460 from the phrase table those phrases with 0.9 ≤ p(e|t)/p(t|e) ≤1.1 and p(t|e) + p(e|t) ≥1.5 and added them to the training data to further bias the alignment process. The resulting BLEU score was 23.78 averaged over 10 data sets (max: 24.52, min: 22.25, std: 0.71).16 4 Experiments with Constituent Reordering The transformations in the previous section do not perform any constituent level reordering, but rather eliminate certain English function words as tokens in the text and fold them into complex syntactic tags. That is, no transformations reorder the English SVO order to Turkish SOV,17 for instance, or move postnominal prepositional phrase modifiers in English, to prenominal phrasal modifiers in Turkish. Now that we have the parses of the English side, we have also investigated a more comprehensive set of reordering transformations which perform the following constituent reorderings to bring English constituent order more in line with the Turkish constitent order at the top and embedded phrase levels: • Object reordering (ObjR), in which the objects and their dependents are moved in front of the verb. • Adverbial phrase reordering (AdvR), which involve moving post-verbal adverbial phrases in front of the verb. • Passive sentence agent reordering (PassAgR), in which any post-verbal agents marked by by, are moved in front of the verb. • Subordinate clause reordering (SubCR) which involve moving postnominal relative clauses or prepositional phrase modifers in front of any modifiers of the head noun. Similarly any prepositional phrases attached to verbs are moved to in front of the verb. We performed these reorderings on top of the data obtained with the Noun+Adj+Verb+Adv+PostP transformations earlier in Section 3.2.2 and used the same decoder parameters. Table 4 shows the performance obtained after various combination of reordering operations over the 10 data sets. Although there were some improvements for certain cases, none 16These experiments were done on top of the model in 3.2.3 with a 3-gram word and root LMs and 8-gram tag LM. 17Although Turkish is a free-constituent order language, SOV is the dominant order in text. of reordering gave consistent improvements for all the data sets. A cursory examinations of the alignments produced after these reordering transformations indicated that the resulting root alignments were not necessarily that close to being monotonic as we would have expected. Experiment Ave. STD Max. Min. Baseline 21.96 0.72 22.91 20.67 ObjR 21.94 0.71 23.12 20.56 ObjR+AdvR 21.73 0.50 22.44 20.69 ObjR+PassAgR 21.88 0.73 23.03 20.51 ObjR+SubCR 21.88 0.61 22.77 20.92 Table 4: BLEU scores of after reordering transformations 5 Related Work Statistical Machine Translation into a morphologically rich language is a challenging problem in that, on the target side, the decoder needs to generate both the right sequence of constituents and the right sequence of morphemes for each word. Furthermore, since for such languages one can generate tens of hundreds of inflected variants, standard word-based alignment approaches suffer from sparseness issues. Koehn (2005) applied standard phrase-based SMT to Finnish using the Europarl corpus and reported that translation to Finnish had the worst BLEU scores. Using morphology in statistical machine translation has been addressed by many researchers for translation from or into morphologically rich(er) languages. Niessen and Ney (2004) used morphological decomposition to get better alignments. Yang and Kirchhoff (2006) have used phrasebased backoff models to translate unknown words by morphologically decomposing the unknown source words. Lee (2004) and Zolmann et al. (2006) have exploited morphology in ArabicEnglish SMT. Popovic and Ney (2004) investigated improving translation quality from inflected languages by using stems, suffixes and part-ofspeech tags. Goldwater and McClosky (2005) use morphological analysis on the Czech side to get improvements in Czech-to-English statistical machine translation. Minkov et al. (2007) have used morphological postprocessing on the target side, to improve translation quality. Avramidis and Koehn (2008) have annotated English with additional morphological information extracted from a syntactic tree, and have used this in translation to Greek and Czech. Recently, Bisazza and Federico (2009) have applied morphological segmentation in Turkish-to-English statistical machine translation and found that it provides nontrivial BLEU 461 score improvements. In the context of translation from English to Turkish, Durgar-El Kahlout and Oflazer (2010) have explored different representational units of the lexical morphemes and found that selectively splitting morphemes on the target side provided nontrivial improvement in the BLEU score. Their approach was based on splitting the target Turkish side, into constituent morphemes while our approach in this paper is the polar opposite: we do not segment morphemes on the Turkish side but rather join function words on the English side to the related content words. Our approach is somewhat similar to recent approaches that use complex syntactically-motivated complex tags. Birch et al. (2007) have integrated more syntax in a factored translation approach by using CCG supertags as a separate factor and have reported a 0.46 BLEU point improvement in Dutch-toEnglish translations. Although they used supertags, these were obtained not via syntactic analysis but by supertagging, while we determine, on the fly, the appropriate syntactic tags based on syntactic structure. A similar approach based on supertagging was proposed by Hassan et al. (2007). They used both CCG supertags and LTAG supertags in Arabic-to-English phrase-based translation and have reported about 6% relative improvement in BLEU scores. In the context of reordering, one recent work (Xu et al., 2009), was able to get an improvement of 0.6 BLEU points by using source syntactic analysis and a constituent reordering scheme like ours for English-to-Turkish translation, but without using any morphology. 6 Conclusions We have presented a novel way to incorporate source syntactic structure in English-to-Turkish phrase-based machine translation by parsing the source sentences and then encoding many local and nonlocal source syntactic structures as additional complex tag factors. Our goal was to obtain representations of source syntactic structures that parallel target morphological structures, and enable us to extend factored translation, in applicability, to languages with very disparate morphological structures. In our experiments over a limited amount training data, but repeated with 10 different training and test sets, we found that syntax-to-morphology mapping transformations on the source side sentences, along with a very small set of transformations on the target side, coupled with some additional techniques provided about 39% relative improvement in BLEU scores over a word-based baseline and about 28% improvement of a factored baseline. We also experimented with numerous additional syntactic reordering transformation on the source to further bring the constituent order in line with the target order but found that these did not provide any tangible improvements when averaged over the 10 different data sets. It is possible that the techniques presented in this paper may be less effective if the available data is much larger, but we have reasons to believe that they will still be effective then also. The reduction in size of the source language side of the training corpus seems to be definitely effective and there no reason why such a reduction (if not more) will not be observed in larger data. Also, the preprocessing of English prepositional phrases and many adverbial phrases usually involve rather long distance relations in the source side syntactic structure18 and when such structures are coded as complex tags on the nominal or verbal heads, such long distance syntax is effectively “localized” and thus can be better captured with the limited window size used for phrase extraction. One limitation of the approach presented here is that it is not directly applicable in the reverse direction. The data encoding and set-up can directly be employed to generate English “translation” expressed as a sequence of root and complex tag combinations, but then some of the complex tags could encode various syntactic constructs. To finalize the translation after the decoding step, the function words/tags in the complex tag would then have to be unattached and their proper positions in the sentence would have to be located. The problem is essentially one of generating multiple candidate sentences with the unattached function words ambiguously positioned (say in a lattice) and then use a second language model to rerank these sentences to select the target sentence. This is an avenue of research that we intend to look at in the very near future. Acknowledgements We thank Joakim Nivre for providing us with the parser. This publication was made possible by the generous support of the Qatar Foundation through Carnegie Mellon University’s Seed Research program. The statements made herein are solely the responsibility of the authors. 18For instance, consider the example in Figure 2 involving if with some additional modifiers added to the intervening noun phrase. 462 References Eleftherios Avramidis and Philipp Koehn. 2008. Enriching morphologically poor languages for statistical machine translation. In Proceedings of ACL08/HLT, pages 763–770, Columbus, Ohio, June. Alexandra Birch, Miles Osborne, and Philipp Koehn. 2007. CCG supertags in factored translation models. In Proceedings of SMT Workshop at the 45th ACL. Arianna Bisazza and Marcello Federico. 2009. Morphological pre-processing for Turkish to English statistical machine translation. In Proceedings of the International Workshop on Spoken Language Translation, Tokyo, Japan, December. ˙Ilknur Durgar-El-Kahlout and Kemal Oflazer. 2010. Exploiting morphology and local word reordering in English to Turkish phrase-based statistical machine translation. IEEE Transactions on Audio, Speech, and Language Processing. To Appear. Alexander Fraser. 2009. Experiments in morphosyntactic processing for translating to and from German. In Proceedings of the Fourth Workshop on Statistical Machine Translation, pages 115–119, Athens, Greece, March. Association for Computational Linguistics. Sharon Goldwater and David McClosky. 2005. Improving statistical MT through morphological analysis. In Proceedings of HLT/EMNLP-2005, pages 676–683, Vancouver, British Columbia, Canada, October. Hany Hassan, Khalil Sima’an, and Andy Way. 2007. Supertagged phrase-based statistical machine translation. In Proceedings of the 45th ACL, pages 288– 295, Prague, Czech Republic, June. Association for Computational Linguistics. Philipp Koehn and Hieu Hoang. 2007. Factored translation models. In Proceedings of EMNLP. Philipp Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of HLT/NAACL-2003. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th ACL–demonstration session, pages 177–180. Philipp Koehn. 2005. Europarl: A parallel corpus for statistical machine translation. In MT Summit X. Young-Suk Lee. 2004. Morphological analysis for statistical machine translation. In Proceedings of HLT/NAACL-2004 – Companion Volume, pages 57– 60. Einat Minkov, Kristina Toutanova, and Hisami Suzuki. 2007. Generating complex morphology for machine translation. In Proceedings of the 45th ACL, pages 128–135, Prague, Czech Republic, June. Association for Computational Linguistics. Sonja Niessen and Hermann Ney. 2004. Statistical machine translation with scarce resources using morpho-syntatic information. Computational Linguistics, 30(2):181–204. Joakim Nivre, Hall Johan, Nilsson Jens, Chanev Atanas, G¨uls¸en Eryi˘git, Sandra K¨ubler, Marinov Stetoslav, and Erwin Marsi. 2007. Maltparser: A language-independent system for data-driven dependency parsing. Natural Language Engineering Journal, 13(2):99–135. Kemal Oflazer and ˙Ilknur Durgar-El-Kahlout. 2007. Exploring different representational units in English-to-Turkish statistical machine translation. In Proceedings of Statistical Machine Translation Workshop at the 45th Annual Meeting of the Association for Computational Linguistics, pages 25–32. Kemal Oflazer. 1994. Two-level description of Turkish morphology. Literary and Linguistic Computing, 9(2):137–148. Kemal Oflazer. 2008. Statistical machine translation into a morphologically complex language. In Proceedings of the Conference on Intelligent Text Processing and Computational Linguistics (CICLing), pages 376–387. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2001. BLEU: A method for automatic evaluation of machine translation. In Proceedings of the 40th ACL, pages 311–318. Maja Popovic and Hermann Ney. 2004. Towards the use of word stems and suffixes for statistical machine translation. In Proceedings of the 4th LREC, pages 1585–1588, May. Helmut Schmid. 1994. Probabilistic part-of-speech tagging using decision trees. In Proceedings of International Conference on New Methods in Language Processing. Kristina Toutanova, Dan Klein, Christopher D. Manning, and Yoram Singer. 2003. Feature-rich part-ofspeech tagging with a cyclic dependency network. In Proceedings of HLT/NAACL-2003, pages 252– 259. Peng Xu, Jaeho Kang, Michael Ringgaard, and Franz Och. 2009. Using a dependency parser to improve SMT for subject-object-verb languages. In Proceedings HLT/NAACL-2009, pages 245–253, June. Mei Yang and Katrin Kirchhoff. 2006. Phrase-based backoff models for machine translation of highly inflected languages. In Proceedings of EACL-2006, pages 41–48. 463 Deniz Yuret and Ferhan T¨ure. 2006. Learning morphological disambiguation rules for Turkish. In Proceedings of HLT/NAACL-2006, pages 328–334, New York City, USA, June. Andreas Zollmann, Ashish Venugopal, and Stephan Vogel. 2006. Bridging the inflection morphology gap for Arabic statistical machine translation. In Proceedings of HLT/NAACL-2006 – Companion Volume, pages 201–204, New York City, USA, June. 464
|
2010
|
47
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 465–474, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Hindi-to-Urdu Machine Translation Through Transliteration Nadir Durrani Hassan Sajjad Alexander Fraser Helmut Schmid Institute for Natural Language Processing University of Stuttgart {durrani,sajjad,fraser,schmid}@ims.uni-stuttgart.de Abstract We present a novel approach to integrate transliteration into Hindi-to-Urdu statistical machine translation. We propose two probabilistic models, based on conditional and joint probability formulations, that are novel solutions to the problem. Our models consider both transliteration and translation when translating a particular Hindi word given the context whereas in previous work transliteration is only used for translating OOV (out-of-vocabulary) words. We use transliteration as a tool for disambiguation of Hindi homonyms which can be both translated or transliterated or transliterated differently based on different contexts. We obtain final BLEU scores of 19.35 (conditional probability model) and 19.00 (joint probability model) as compared to 14.30 for a baseline phrase-based system and 16.25 for a system which transliterates OOV words in the baseline system. This indicates that transliteration is useful for more than only translating OOV words for language pairs like Hindi-Urdu. 1 Introduction Hindi is an official language of India and is written in Devanagari script. Urdu is the national language of Pakistan, and also one of the state languages in India, and is written in Perso-Arabic script. Hindi inherits its vocabulary from Sanskrit while Urdu descends from several languages including Arabic, Farsi (Persian), Turkish and Sanskrit. Hindi and Urdu share grammatical structure and a large proportion of vocabulary that they both inherited from Sanskrit. Most of the verbs and closed-class words (pronouns, auxiliaries, casemarkers, etc) are the same. Because both languages have lived together for centuries, some Urdu words which originally came from Arabic and Farsi have also mixed into Hindi and are now part of the Hindi vocabulary. The spoken form of the two languages is very similar. The extent of overlap between Hindi and Urdu vocabulary depends upon the domain of the text. Text coming from the literary domain like novels or history tend to have more Sanskrit (for Hindi) and Persian/Arabic (for Urdu) vocabulary. However, news wire that contains text related to media, sports and politics, etc., is more likely to have common vocabulary. In an initial study on a small news corpus of 5000 words, randomly selected from BBC1 News, we found that approximately 62% of the Hindi types are also part of Urdu vocabulary and thus can be transliterated while only 38% have to be translated. This provides a strong motivation to implement an end-to-end translation system which strongly relies on high quality transliteration from Hindi to Urdu. Hindi and Urdu have similar sound systems but transliteration from Hindi to Urdu is still very hard because some phonemes in Hindi have several orthographic equivalents in Urdu. For example the “z” sound2 can only be written as whenever it occurs in a Hindi word but can be written as , , and in an Urdu word. Transliteration becomes non-trivial in cases where the multiple orthographic equivalents for a Hindi word are all valid Urdu words. Context is required to resolve ambiguity in such cases. Our transliterator (described in sections 3.1.2 and 4.1.3) gives an accuracy of 81.6% and a 25-best accuracy of 92.3%. Transliteration has been previously used only as a back-off measure to translate NEs (Name Entities) and OOV words in a pre- or post-processing step. The problem we are solving is more difficult than techniques aimed at handling OOV words, 1http://www.bbc.co.uk/hindi/index.shtml 2All sounds are represented using SAMPA notation. 465 Hindi Urdu SAMPA Gloss / Am Mango/Ordinary / d ZAli Fake/Net / Ser Lion/Verse Table 1: Hindi Words That Can Be Transliterated Differently in Different Contexts Hindi Urdu SAMPA Gloss / simA Border/Seema / Amb@r Sky/Ambar / vId Ze Victory/Vijay Table 2: Hindi Words That Can Be Translated or Transliterated in Different Contexts which focus primarily on name transliteration, because we need different transliterations in different contexts; in their case context is irrelevant. For example: consider the problem of transliterating the English word “read” to a phoneme representation in the context “I will read” versus the context “I have read”. An example of this for Hindi to Urdu transliteration: the two Urdu words (face/condition) and (chapter of the Koran) are both written as (sur@t d) in Hindi. The two are pronounced identically in Urdu but written differently. In such cases we hope to choose the correct transliteration by using context. Some other examples are shown in Table 1. Sometimes there is also an ambiguity of whether to translate or transliterate a particular word. The Hindi word , for example, will be translated to (peace, s@kun) when it is a common noun but transliterated to (Shanti, SAnt di) when it is a proper name. We try to model whether to translate or transliterate in a given situation. Some other examples are shown in Table 2. The remainder of this paper is organized as follows. Section 2 provides a review of previous work. Section 3 introduces two probabilistic models for integrating translations and transliterations into a translation model which are based on conditional and joint probability distributions. Section 4 discusses the training data, parameter optimization and the initial set of experiments that compare our two models with a baseline Hindi-Urdu phrasebased system and with two transliteration-aided phrase-based systems in terms of BLEU scores (Papineni et al., 2001). Section 5 performs an error analysis showing interesting weaknesses in the initial formulations. We remedy the problems by adding some heuristics and modifications to our models which show improvements in the results as discussed in section 6. Section 7 gives two examples illustrating how our model decides whether to translate or transliterate and how it is able to choose among different valid transliterations given the context. Section 8 concludes the paper. 2 Previous Work There has been a significant amount of work on transliteration. We can break down previous work into three groups. The first group is generic transliteration work, which is evaluated outside of the context of translation. This work uses either grapheme or phoneme based models to transliterate words lists (Knight and Graehl, 1998; Li et al., 2004; Ekbal et al., 2006; Malik et al., 2008). The work by Malik et al. addresses Hindi to Urdu transliteration using hand-crafted rules and a phonemic representation; it ignores translation context. A second group deals with out-of-vocabulary words for SMT systems built on large parallel corpora, and therefore focuses on name transliteration, which is largely independent of context. AlOnaizan and Knight (2002) transliterate Arabic NEs into English and score them against their respective translations using a modified IBM Model 1. The options are further re-ranked based on different measures such as web counts and using coreference to resolve ambiguity. These re-ranking methodologies can not be performed in SMT at the decoding time. An efficient way to compute and re-rank the transliterations of NEs and integrate them on the fly might be possible. However, this is not practical in our case as our model considers transliterations of all input words and not just NEs. A log-linear block transliteration model is applied to OOV NEs in Arabic to English SMT by Zhao et al. (2007). This work is also transliterating only NEs and not doing any disambiguation. The best method proposed by Kashani et al. (2007) integrates translations provided by external sources such as transliteration or rule-base translation of numbers and dates, for an arbitrary number of entries within the input text. Our work is different from Kashani et al. (2007) in that our model compares transliterations with translations 466 on the fly whereas transliterations in Kashani et al. do not compete with internal phrase tables. They only compete amongst themselves during a second pass of decoding. Hermjakob et al. (2008) use a tagger to identify good candidates for transliteration (which are mostly NEs) in input text and add transliterations to the SMT phrase table dynamically such that they can directly compete with translations during decoding. This is closer to our approach except that we use transliteration as an alternative to translation for all Hindi words. Our focus is disambiguation of Hindi homonyms whereas they are concentrating only on transliterating NE’s. Moreover, they are working with a large bitext so they can rely on their translation model and only need to transliterate NEs and OOVs. Our translation model is based on data which is both sparse and noisy. Therefore we pit transliterations against translations for every input word. Sinha (2009) presents a rule-based MT system that uses Hindi as a pivot to translate from English to Urdu. This work also uses transliteration only for the translation of unknown words. Their work can not be used for direct translation from Hindi to Urdu (independently of English) “due to various ambiguous mappings that have to be resolved”. The third group uses transliteration models inside of a cross-lingual IR system (AbdulJaleel and Larkey, 2003; Virga and Khudanpur, 2003; Pirkola et al., 2003). Picking a single best transliteration or translation in context is not important in an IR system. Instead, all the options are used by giving them weights and context is typically not taken into account. 3 Our Approach Both of our models combine a character-based transliteration model with a word-based translation model. Our models look for the most probable Urdu token sequence un 1 for a given Hindi token sequence hn 1. We assume that each Hindi token is mapped to exactly one Urdu token and that there is no reordering. The assumption of no reordering is reasonable given the fact that Hindi and Urdu have identical grammar structure and the same word order. An Urdu token might consist of more than one Urdu word3. The following sections give a math3This occurs frequently in case markers with nouns, derivational affixes and compounds etc. These are written as single words in Hindi as opposed to Urdu where they are ematical formulation of our two models, Model-1 and Model-2. 3.1 Model-1 : Conditional Probability Model Applying a noisy channel model to compute the most probable translation ˆun 1, we get: arg max un 1 p(un 1|hn 1) = arg max un 1 p(un 1)p(hn 1|un 1) (1) 3.1.1 Language Model The language model (LM) p(un 1) is implemented as an n-gram model using the SRILM-Toolkit (Stolcke, 2002) with Kneser-Ney smoothing. The parameters of the language model are learned from a monolingual Urdu corpus. The language model is defined as: p(un 1) = n Y i=1 pLM(ui|ui−1 i−k) (2) where k is a parameter indicating the amount of context used (e.g., k = 4 means 5-gram model). ui can be a single or a multi-word token. A multi-word token consists of two or more Urdu words. For a multi-word ui we do multiple language model look-ups, one for each uix in ui = ui1, . . . , uim and take their product to obtain the value pLM(ui|ui−1 i−k). Language Model for Unknown Words: Our model generates transliterations that can be known or unknown to the language model and the translation model. We refer to the words known to the language model and to the translation model as LM-known and TM-known words respectively and to words that are unknown as LM-unknown and TM-unknown respectively. We assign a special value ψ to the LM-unknown words. If one or more uix in a multi-word ui are LM-unknown we assign a language model score pLM(ui|ui−1 i−k) = ψ for the entire ui, meaning that we consider partially known transliterations to be as bad as fully unknown transliterations. The parameter ψ controls the trade-off between LMknown and LM-unknown transliterations. It does not influence translation options because they are always LM-known in our case. This is because our monolingual corpus also contains the Urdu part of translation corpus. The optimization of ψ is described in section 4.2.1. written as two words. For example (beautiful ; xubsur@t d) and (your’s ; ApkA) are written as and respectively in Urdu. 467 3.1.2 Translation Model The translation model (TM) p(hn 1|un 1) is approximated with a context-independent model: p(hn 1|un 1) = n Y i=1 p(hi|ui) (3) where hi and ui are Hindi and Urdu tokens respectively. Our model estimates the conditional probability p(hi|ui) by interpolating a wordbased model and a character-based (transliteration) model. p(hi|ui) = λpw(hi|ui) + (1 −λ)pc(hi|ui) (4) The parameters of the word-based translation model pw(h|u) are estimated from the word alignments of a small parallel corpus. We only retain 1-1/1-N (1 Hindi word, 1 or more Urdu words) alignments and throw away N-1 and M-N alignments for our models. This is further discussed in section 4.1.1. The character-based transliteration model pc(h|u) is computed in terms of pc(h, u), a joint character model, which is also used for ChineseEnglish back-transliteration (Li et al., 2004) and Bengali-English name transliteration (Ekbal et al., 2006). The character-based transliteration probability is defined as follows: pc(h, u) = X an 1 ∈align(h,u) p(an 1) = X an 1 ∈align(h,u) n Y i=1 p(ai|ai−1 i−k) (5) where ai is a pair consisting of the i-th Hindi character hi and the sequence of 0 or more Urdu characters that it is aligned with. A sample alignment is shown in Table 3(b) in section 4.1.3. Our best results are obtained with a 5-gram model. The parameters p(ai|ai−1 i−k) are estimated from a small transliteration corpus which we automatically extracted from the translation corpus. The extraction details are also discussed in section 4.1.3. Because our overall model is a conditional probability model, joint-probabilities are marginalized using character-based prior probabilities: pc(h|u) = pc(h, u) pc(u) (6) The prior probability pc(u) of the character sequence u = cm 1 is defined with a character-based language model: pc(u) = m Y i=1 p(ci|ci−1 i−k) (7) The parameters p(ci|ci−1 i−k) are estimated from the Urdu part of the character-aligned transliteration corpus. Replacing (6) in (4) we get: p(hi|ui) = λpw(hi|ui) + (1 −λ)pc(hi, ui) pc(ui) (8) Having all the components of our model defined we insert (8) and (2) in (1) to obtain the final equation: ˆun 1 = arg max un 1 n Y i=1 pLM(ui|ui−1 i−k)[λpw(hi|ui) + (1 −λ)pc(hi, ui) pc(ui) ] (9) The optimization of the interpolating factor λ is discussed in section 4.2.1. 3.2 Model-2 : Joint Probability Model This section briefly defines a variant of our model where we interpolate joint probabilities instead of conditional probabilities. Again, the translation model p(hn 1|un 1) is approximated with a contextindependent model: p(hn 1|un 1) = n Y i=1 p(hi|ui) = n Y i=1 p(hi, ui) p(ui) (10) The joint probability p(hi, ui) of a Hindi and an Urdu word is estimated by interpolating a wordbased model and a character-based model. p(hi, ui) = λpw(hi, ui)+(1−λ)pc(hi, ui) (11) and the prior probability p(ui) is estimated as: p(ui) = λpw(ui) + (1 −λ)pc(ui) (12) The parameters of the translation model pw(hi, ui) and the word-based prior probabilities pw(ui) are estimated from the 1-1/1-N word-aligned corpus (the one that we also used to estimate translation probabilities pw(hi|ui) previously). The character-based transliteration probability pc(hi, ui) and the character-based prior probability pc(ui) are defined by (5) and (7) respectively in 468 the previous section. Putting (11) and (12) in (10) we get p(hn 1|un 1) = n Y i=1 λpw(hi, ui) + (1 −λ)pc(hi, ui) λpw(ui) + (1 −λ)pc(ui) (13) The idea is to interpolate joint probabilities and divide them by the interpolated marginals. The final equation for Model-2 is given as: ˆun 1 = arg max un 1 n Y i=1 pLM(ui|ui−1 i−k)× λpw(hi, ui) + (1 −λ)pc(hi, ui) λpw(ui) + (1 −λ)pc(ui) (14) 3.3 Search The decoder performs a stack-based search using a beam-search algorithm similar to the one used in Pharoah (Koehn, 2004a). It searches for an Urdu string that maximizes the product of translation probability and the language model probability (equation 1) by translating one Hindi word at a time. It is implemented as a two-level process. At the lower level, it computes n-best transliterations for each Hindi word hi according to pc(h, u). The joint probabilities given by pc(h, u) are marginalized for each Urdu transliteration to give pc(h|u). At the higher level, transliteration probabilities are interpolated with pw(h|u) and then multiplied with language model probabilities to give the probability of a hypothesis. We use 20-best translations and 25-best transliterations for pw(h|u) and pc(h|u) respectively and a 5-gram language model. To keep the search space manageable and time complexity polynomial we apply pruning and recombination. Since our model uses monotonic decoding we only need to recombine hypotheses that have the same context (last n-1 words). Next we do histogram-based pruning, maintaining the 100best hypotheses for each stack. 4 Evaluation 4.1 Training This section discusses the training of the different model components. 4.1.1 Translation Corpus We used the freely available EMILLE Corpus as our bilingual resource which contains roughly 13,000 Urdu and 12,300 Hindi sentences. From these we were able to sentence-align 7000 sentence pairs using the sentence alignment algorithm given by Moore (2002). The word alignments for this task were extracted by using GIZA++ (Och and Ney, 2003) in both directions. We extracted a total of 107323 alignment pairs (5743 N-1 alignments, 8404 MN alignments and 93176 1-1/1-N alignments). Of these alignments M-N and N-1 alignment pairs were ignored. We manually inspected a sample of 1000 instances of M-N/N-1 alignments and found that more than 70% of these were (totally or partially) wrong. Of the 30% correct alignments, roughly one-third constitute N-1 alignments. Most of these are cases where the Urdu part of the alignment actually consists of two (or three) words but was written without space because of lack of standard writing convention in Urdu. For example (can go ; d ZA s@kt de) is alternatively written as (can go ; d ZAs@kt de) i.e. without space. We learned that these N-1 translations could be safely dropped because we can generate a separate Urdu word for each Hindi word. For valid M-N alignments we observed that these could be broken into 1-1/1-N alignments in most of the cases. We also observed that we usually have coverage of the resulting 1-1 and 1-N alignments in our translation corpus. Looking at the noise in the incorrect alignments we decided to drop N-1 and M-N cases. We do not model deletions and insertions so we ignored null alignments. Also 1-N alignments with gaps were ignored. Only the alignments with contiguous words were kept. 4.1.2 Monolingual Corpus Our monolingual Urdu corpus consists of roughly 114K sentences. This comprises 108K sentences from the data made available by the University of Leipzig4 + 5600 sentences from the training data of each fold during cross validation. 4.1.3 Transliteration Corpus The training corpus for transliteration is extracted from the 1-1/1-N word-alignments of the EMILLE corpus discussed in section 4.1.1. We use an edit distance algorithm to align this training corpus at the character level and we eliminate translation pairs with high edit distance which are unlikely to be transliterations. 4http://corpora.informatik.uni-leipzig.de/ 469 We used our knowledge of the Hindi and Urdu scripts to define the initial character mapping. The mapping was further extended by looking into available Hindi-Urdu transliteration systems[5,6] and other resources (Gupta, 2004; Malik et al., 2008; Jawaid and Ahmed, 2009). Each pair in the character map is assigned a cost. A Hindi character that always map to only one Urdu character is assigned a cost of 0 whereas the Hindi characters that map to different Urdu characters are assigned a cost of 0.2. The edit distance metric allows insert, delete and replace operations. The handcrafted pairs define the cost of replace operations. We set a cost of 0.6 for deletions and insertions. These costs were optimized on held out data. The details of optimization are not mentioned due to limited space. Using this metric we filter out the word pairs with high edit-distance to extract our transliteration corpus. We were able to extract roughly 2100 unique pairs along with their alignments. The resulting alignments are modified by merging unaligned ∅→1 (no character on source side, 1 character on target side) or ∅→N alignments with the preceding alignment pair. If there is no preceding alignment pair then it is merged with the following pair. Table 3 gives an example showing initial alignment (a) and the final alignment (b) after applying the merge operation. Our model retains 1 →∅and N →∅alignments as deletion operations. a) Hindi ∅ b c ∅ e f Urdu A XY C D ∅ F b) Hindi b c e f Urdu AXY CD ∅ F Table 3: Alignment (a) Before (b) After Merge The parameters pc(h, u) and pc(u) are trained on the aligned corpus using the SRILM toolkit. We use Add-1 smoothing for unigrams and Kneser-Ney smoothing for higher n-grams. 4.1.4 Diacritic Removal and Normalization In Urdu, short vowels are represented with diacritics but these are rarely written in practice. In order to keep the data consistent, all diacritics are removed. This loss of information is not harmful when transliterating/translating from Hindi to Urdu because undiacritized text is equally read5CRULP: http://www.crulp.org/software/langproc.htm 6Malerkotla.org: http://translate.malerkotla.co.in able to native speakers as its diacritized counter part. However leaving occasional diacritics in the corpus can worsen the problem of data sparsity by creating spurious ambiguity7. There are a few Urdu characters that have multiple equivalent Unicodes. All such forms are normalized to have only one representation8. 4.2 Experimental Setup We perform a 5-fold cross validation taking 4/5 of the data as training and 1/5 as test data. Each fold comprises roughly 1400 test sentences and 5600 training sentences. 4.2.1 Parameter Optimization Our model contains two parameters λ (the interpolating factor between translation and transliteration modules) and ψ (the factor that controls the trade-off between LM-known and LM-unknown transliterations). The interpolating factor λ is initialized, inspired by Written-Bell smoothing, with a value of N N+B 9. We chose a very low value 1e−40 for the factor ψ initially, favoring LMknown transliterations very strongly. Both of these parameters are optimized as described below. Because our training data is very sparse we do not use held-out data for parameter optimization. Instead we optimize these parameters by performing a 2-fold optimization for each of the 5 folds. Each fold is divided into two halves. The parameters λ and ψ are optimized on the first half and the other half is used for testing, then optimization is done on the second half and the first half is used for testing. The optimal value for parameter λ occurs between 0.7-0.84 and for the parameter ψ between 1e−5 and 1e−10. 4.2.2 Results Baseline Pb0: We ran Moses (Koehn et al., 2007) using Koehn’s training scripts10, doing a 5-fold cross validation with no reordering11. For the other parameters we use the default values i.e. 5-gram language model and maximum phraselength= 6. Again, the language model is imple7It should be noted though that diacritics play a very important role when transliterating in the reverse direction because these are virtually always written in Hindi as dependent vowels. 8www.crulp.org/software/langproc/urdunormalization.htm 9N is the number of aligned word pairs (tokens) and B is the number of different aligned word pairs (types). 10http://statmt.org/wmt08/baseline.html 11Results are worse with reordering enabled. 470 M Pb0 Pb1 Pb2 M1 M2 BLEU 14.3 16.25 16.13 18.6 17.05 Table 4: Comparing Model-1 and Model-2 with Phrase-based Systems mented as an n-gram model using the SRILMToolkit with Kneser-Ney smoothing. Each fold comprises roughly 1400 test sentences, 5000 in training and 600 in dev12. We also used two methods to incorporate transliterations in the phrasebased system: Post-process Pb1: All the OOV words in the phrase-based output are replaced with their topcandidate transliteration as given by our transliteration system. Pre-process Pb2: Instead of adding transliterations as a post process we do a second pass by adding the unknown words with their topcandidate transliteration to the training corpus and rerun Koehn’s training script with the new training corpus. Table 4 shows results (taking arithmetic average over 5 folds) from Model-1 and Model2 in comparison with three baselines discussed above. Both our systems (Model-1 and Model-2) beat the baseline phrase-based system with a BLEU point difference of 4.30 and 2.75 respectively. The transliteration aided phrase-based systems Pb1 and Pb2 are closer to our Model-2 results but are way below Model-1 results. The difference of 2.35 BLEU points between M1 and Pb1 indicates that transliteration is useful for more than only translating OOV words for language pairs like Hindi-Urdu. Our models choose between translations and transliterations based on context unlike the phrase-based systems Pb1 and Pb2 which use transliteration only as a tool to translate OOV words. 5 Error Analysis Based on preliminary experiments we found three major flaws in our initial formulations. This section discusses each one of them and provides some heuristics and modifications that we employ to try to correct deficiencies we found in the two models described in section 3.1 and 3.2. 12After having the MERT parameters, we add the 600 dev sentences back into the training corpus, retrain GIZA, and then estimate a new phrase table on all 5600 sentences. We then use the MERT parameters obtained before together with the newer (larger) phrase-table set. 5.1 Heuristic-1 A lot of errors occur because our translation model is built on very sparse and noisy data. The motivation for this heuristic is to counter wrong alignments at least in the case of verbs and functional words (which are often transliterations). This heuristic favors translations that also appear in the n-best transliteration list over only-translation and only-transliteration options. We modify the translation model for both the conditional and the joint model by adding another factor which strongly weighs translation+transliteration options by taking the square-root of the product of the translation and transliteration probabilities. Thus modifying equations (8) and (11) in Model-1 and Model-2 we obtain equations (15) and (16) respectively: p(hi|ui) = λ1pw(hi|ui) + λ2 pc(hi, ui) pc(ui) + λ3 s pw(hi|ui)pc(hi, ui) pc(ui) (15) p(hi, ui) = λ1pw(hi, ui) + λ2pc(hi, ui) + λ3 p pw(hi, ui)pc(hi, ui) (16) For the optimization of lambda parameters we hold the value of the translation coefficient λ113 and the transliteration coefficient λ2 constant (using the optimized values as discussed in section 4.2.1) and optimize λ3 again using 2-fold optimization on all the folds as described above14. 5.2 Heuristic-2 When an unknown Hindi word occurs for which all transliteration options are LM-unknown then the best transliteration should be selected. The problem in our original models is that a fixed LM probability ψ is used for LM-unknown transliterations. Hence our model selects the transliteration that has the best pc(hi,ui) pc(ui) score i.e. we maximize pc(hi|ui) instead of pc(ui|hi) (or equivalently pc(hi, ui)). The reason is an inconsistency in our models. The language model probability of unknown words is uniform (and equal to ψ) whereas the translation model uses the nonuniform prior probability pc(ui) for these words. There is another reason why we can not use the 13The translation coefficient λ1 is same as λ used in previous models and the transliteration coefficient λ2 = 1 −λ 14After optimization we normalize the lambdas to make their sum equal to 1. 471 value ψ in this case. Our transliterator model also produces space inserted words. The value of ψ is very small because of which transliterations that are actually LM-unknown, but are mistakenly broken into constituents that are LM-known, will always be preferred over their counter parts. An example of this is (America) for which two possible transliterations as given by our model are (AmerIkA, without space) and (AmerI kA, with space). The latter version is LM-known as its constituents are LM-known. Our models always favor the latter version. Space insertion is an important feature of our transliteration model. We want our transliterator to tackle compound words, derivational affixes, case-markers with nouns that are written as one word in Hindi but as two or more words in Urdu. Examples were already shown in section 3’s footnote. We eliminate the inconsistency by using pc(ui) as the 0-gram back-off probability distribution in the language model. For an LM-unknown transliterations we now get in Model-1: p(ui|ui−1 i−k)[λpw(hi|ui) + (1 −λ)pc(hi, ui) pc(ui) ] = p(ui|ui−1 i−k)[(1 −λ)pc(hi, ui) pc(ui) ] = k Y j=0 α(ui−1 i−j)pc(ui)[(1 −λ)pc(hi, ui) pc(ui) ] = k Y j=0 α(ui−1 i−j)[(1 −λ)pc(hi, ui)] where Qk j=0 α(ui−1 i−j) is just the constant that SRILM returns for unknown words. The last line of the calculation shows that we simply drop pc(ui) if ui is LM-unknown and use the constant Qk j=0 α(ui−1 i−j) instead of ψ. A similar calculation for Model-2 gives Qk j=0 α(ui−1 i−j)pc(hi, ui). 5.3 Heuristic-3 This heuristic discusses a flaw in Model-2. For transliteration options that are TM-unknown, the pw(h, u) and pw(u) factors becomes zero and the translation model probability as given by equation (13) becomes: (1 −λ)pc(hi, ui) (1 −λ)pc(ui) = pc(hi, ui) pc(ui) In such cases the λ factor cancels out and no weighting of word translation vs. transliteration H1 H2 H12 M1 18.86 18.97 19.35 M2 17.56 17.85 18.34 Table 5: Applying Heuristics 1 and 2 and their Combinations to Model-1 and Model-2 H3 H13 H23 H123 M2 18.52 18.93 18.55 19.00 Table 6: Applying Heuristic 3 and its Combinations with other Heuristics to Model-2 occurs anymore. As a result of this, transliterations are sometimes incorrectly favored over their translation alternatives. In order to remedy this problem we assign a minimal probability β to the word-based prior pw(ui) in case of TM-unknown transliterations, which prevents it from ever being zero. Because of this addition the translation model probability for LM-unknown words becomes: (1 −λ)pc(hi, ui) λβ + (1 −λ)pc(ui) where β = 1 Urdu Types in TM 6 Final Results This section shows the improvement in BLEU score by applying heuristics and combinations of heuristics in both the models. Tables 5 and 6 show the improvements achieved by using the different heuristics and modifications discussed in section 5. We refer to the results as MxHy where x denotes the model number, 1 for the conditional probability model and 2 for the joint probability model and y denotes a heuristic or a combination of heuristics applied to that model15. Both heuristics (H1 and H2) show improvements over their base models M1 and M2. Heuristic-1 shows notable improvement for both models in parts of test data which has high number of common vocabulary words. Using heuristic 2 we were able to properly score LM-unknown transliterations against each other. Using these heuristics together we obtain a gain of 0.75 over M-1 and a gain of 1.29 over M-2. Heuristic-3 remedies the flaw in M2 by assigning a special value to the word-based prior pw(ui) for TM-unknown words which prevents the cancelation of interpolating parameter λ. M2 combined with heuristic 3 (M2H3) results in a 1.47 15For example M1H1 refers to the results when heuristic1 is applied to model-1 whereas M2H12 refers to the results when heuristics 1 and 2 are together applied to model 2. 472 BLEU point improvement and combined with all the heuristics (M2H123) gives an overall gain of 1.95 BLEU points and is close to our best results (M1H12). We also performed significance test by concatenating all the fold results. Both our best systems M1H12 and M2H123 are statistically significant (p < 0.05)16 over all the baselines discussed in section 4.2.2. One important issue that has not been investigated yet is that BLEU has not yet been shown to have good performance in morphologically rich target languages like Urdu, but there is no metric known to work better. We observed that sometimes on data where the translators preferred to translate rather than doing transliteration our system is penalized by BLEU even though our output string is a valid translation. For other parts of the data where the translators have heavily used transliteration, the system may receive a higher BLEU score. We feel that this is an interesting area of research for automatic metric developers, and that a large scale task of translation to Urdu which would involve a human evaluation campaign would be very interesting. 7 Sample Output This section gives two examples showing how our model (M1H2) performs disambiguation. Given below are some test sentences that have Hindi homonyms (underlined in the examples) along with Urdu output given by our system. In the first example (given in Figure 1) Hindi word can be transliterated to ( Lion) or (Verse) depending upon the context. Our model correctly identifies which transliteration to choose given the context. In the second example (shown in Figure 2) Hindi word can be translated to (peace, s@kun) when it is a common noun but transliterated to (Shanti, SAnt di) when it is a proper name. Our model successfully decides whether to translate or transliterate given the context. 8 Conclusion We have presented a novel way to integrate transliterations into machine translation. In closely related language pairs such as Hindi-Urdu with a significant amount of vocabulary overlap, 16We used Kevin Gimpel’s tester (http://www.ark.cs.cmu.edu/MT/) which uses bootstrap resampling (Koehn, 2004b), with 1000 samples. Ser d Z@ngl kA rAd ZA he “Lion is the king of jungle” AIqbAl kA Aek xub sur@t d Ser he “There is a beautiful verse from Iqbal” Figure 1: Different Transliterations in Different Contexts p hIr b hi vh s@kun se n@he˜rh s@kt dA “Even then he can’t live peacefully” Aom SAnt di Aom frhA xAn ki d dusri fIl@m he “Om Shanti Om is Farah Khan’s second film” Figure 2: Translation or Transliteration transliteration can be very effective in machine translation for more than just translating OOV words. We have addressed two problems. First, transliteration helps overcome the problem of data sparsity and noisy alignments. We are able to generate word translations that are unseen in the translation corpus but known to the language model. Additionally, we can generate novel transliterations (that are LM-Unknown). Second, generating multiple transliterations for homograph Hindi words and using language model context helps us solve the problem of disambiguation. We found that the joint probability model performs almost as well as the conditional probability model but that it was more complex to make it work well. Acknowledgments The first two authors were funded by the Higher Education Commission (HEC) of Pakistan. The third author was funded by Deutsche Forschungsgemeinschaft grants SFB 732 and MorphoSynt. The fourth author was funded by Deutsche Forschungsgemeinschaft grant SFB 732. 473 References Nasreen AbdulJaleel and Leah S. Larkey. 2003. Statistical transliteration for English-Arabic cross language information retrieval. In CIKM 03: Proceedings of the twelfth international conference on Information and knowledge management, pages 139– 146. Yaser Al-Onaizan and Kevin Knight. 2002. Translating named entities using monolingual and bilingual resources. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 400–408. Asif Ekbal, Sudip Kumar Naskar, and Sivaji Bandyopadhyay. 2006. A modified joint source-channel model for transliteration. In Proceedings of the COLING/ACL poster sessions, pages 191–198, Sydney, Australia. Association for Computational Linguistics. Swati Gupta. 2004. Aligning Hindi and Urdu bilingual corpora for robust projection. Masters project dissertation, Department of Computer Science, University of Sheffield. Ulf Hermjakob, Kevin Knight, and Hal Daum´e III. 2008. Name translation in statistical machine translation - learning when to transliterate. In Proceedings of ACL-08: HLT, pages 389–397, Columbus, Ohio. Association for Computational Linguistics. Bushra Jawaid and Tafseer Ahmed. 2009. Hindi to Urdu conversion: beyond simple transliteration. In Conference on Language and Technology 2009, Lahore, Pakistan. Mehdi M. Kashani, Eric Joanis, Roland Kuhn, George Foster, and Fred Popowich. 2007. Integration of an Arabic transliteration module into a statistical machine translation system. In Proceedings of the Second Workshop on Statistical Machine Translation, pages 17–24, Prague, Czech Republic. Association for Computational Linguistics. Kevin Knight and Jonathan Graehl. 1998. Machine transliteration. Computational Linguistics, 24(4):599–612. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics, Demonstration Program, Prague, Czech Republic. Philipp Koehn. 2004a. Pharaoh: A beam search decoder for phrase-based statistical machine translation models. In AMTA, pages 115–124. Philipp Koehn. 2004b. Statistical significance tests for machine translation evaluation. In Dekang Lin and Dekai Wu, editors, Proceedings of EMNLP 2004, pages 388–395, Barcelona, Spain, July. Association for Computational Linguistics. Haizhou Li, Zhang Min, and Su Jian. 2004. A joint source-channel model for machine transliteration. In ACL ’04: Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics, pages 159–166, Barcelona, Spain. Association for Computational Linguistics. M G Abbas Malik, Christian Boitet, and Pushpak Bhattacharyya. 2008. Hindi Urdu machine transliteration using finite-state transducers. In Proceedings of the 22nd International Conference on Computational Linguistics, Manchester, UK. Robert C. Moore. 2002. Fast and accurate sentence alignment of bilingual corpora. In Conference of the Association for Machine Translation in the Americas (AMTA). Franz J. Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1):19–51. Kishore A. Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2001. BLEU: a method for automatic evaluation of machine translation. Technical Report RC22176 (W0109-022), IBM Research Division, Thomas J. Watson Research Center, Yorktown Heights, NY. Ari Pirkola, Jarmo Toivonen, Heikki Keskustalo, Kari Visala, and Kalervo J¨arvelin. 2003. Fuzzy translation of cross-lingual spelling variants. In SIGIR ’03: Proceedings of the 26th annual international ACM SIGIR conference on Research and development in informaion retrieval, pages 345–352, New York, NY, USA. ACM. R. Mahesh K. Sinha. 2009. Developing English-Urdu machine translation via Hindi. In Third Workshop on Computational Approaches to Arabic Scriptbased Languages (CAASL3), MT Summit XII, Ottawa, Canada. Andreas Stolcke. 2002. SRILM - an extensible language modeling toolkit. In Intl. Conf. Spoken Language Processing, Denver, Colorado. Paola Virga and Sanjeev Khudanpur. 2003. Transliteration of proper names in cross-lingual information retrieval. In Proceedings of the ACL 2003 workshop on Multilingual and mixed-language named entity recognition, pages 57–64, Morristown, NJ, USA. Association for Computational Linguistics. Bing Zhao, Nguyen Bach, Ian Lane, and Stephan Vogel. 2007. A log-linear block transliteration model based on bi-stream HMMs. In Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 364–371, Rochester, New York. Association for Computational Linguistics. 474
|
2010
|
48
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 475–484, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Training Phrase Translation Models with Leaving-One-Out Joern Wuebker and Arne Mauser and Hermann Ney Human Language Technology and Pattern Recognition Group RWTH Aachen University, Germany <surname>@cs.rwth-aachen.de Abstract Several attempts have been made to learn phrase translation probabilities for phrasebased statistical machine translation that go beyond pure counting of phrases in word-aligned training data. Most approaches report problems with overfitting. We describe a novel leavingone-out approach to prevent over-fitting that allows us to train phrase models that show improved translation performance on the WMT08 Europarl German-English task. In contrast to most previous work where phrase models were trained separately from other models used in translation, we include all components such as single word lexica and reordering models in training. Using this consistent training of phrase models we are able to achieve improvements of up to 1.4 points in BLEU. As a side effect, the phrase table size is reduced by more than 80%. 1 Introduction A phrase-based SMT system takes a source sentence and produces a translation by segmenting the sentence into phrases and translating those phrases separately (Koehn et al., 2003). The phrase translation table, which contains the bilingual phrase pairs and the corresponding translation probabilities, is one of the main components of an SMT system. The most common method for obtaining the phrase table is heuristic extraction from automatically word-aligned bilingual training data (Och et al., 1999). In this method, all phrases of the sentence pair that match constraints given by the alignment are extracted. This includes overlapping phrases. At extraction time it does not matter, whether the phrases are extracted from a highly probable phrase alignment or from an unlikely one. Phrase model probabilities are typically defined as relative frequencies of phrases extracted from word-aligned parallel training data. The joint counts C( ˜f, ˜e) of the source phrase ˜f and the target phrase ˜e in the entire training data are normalized by the marginal counts of source and target phrase to obtain a conditional probability pH( ˜f|˜e) = C( ˜f, ˜e) C(˜e) . (1) The translation process is implemented as a weighted log-linear combination of several models hm(eI 1, sK 1 , fJ 1 ) including the logarithm of the phrase probability in source-to-target as well as in target-to-source direction. The phrase model is combined with a language model, word lexicon models, word and phrase penalty, and many others. (Och and Ney, 2004) The best translation ˆeˆI 1 as defined by the models then can be written as ˆe ˆI 1 = argmax I,eI 1 ( M X m=1 λmhm(eI 1, sK 1 , fJ 1 ) ) (2) In this work, we propose to directly train our phrase models by applying a forced alignment procedure where we use the decoder to find a phrase alignment between source and target sentences of the training data and then updating phrase translation probabilities based on this alignment. In contrast to heuristic extraction, the proposed method provides a way of consistently training and using phrase models in translation. We use a modified version of a phrase-based decoder to perform the forced alignment. This way we ensure that all models used in training are identical to the ones used at decoding time. An illustration of the basic 475 Figure 1: Illustration of phrase training with forced alignment. idea can be seen in Figure 1. In the literature this method by itself has been shown to be problematic because it suffers from over-fitting (DeNero et al., 2006), (Liang et al., 2006). Since our initial phrases are extracted from the same training data, that we want to align, very long phrases can be found for segmentation. As these long phrases tend to occur in only a few training sentences, the EM algorithm generally overestimates their probability and neglects shorter phrases, which better generalize to unseen data and thus are more useful for translation. In order to counteract these effects, our training procedure applies leaving-one-out on the sentence level. Our results show, that this leads to a better translation quality. Ideally, we would produce all possible segmentations and alignments during training. However, this has been shown to be infeasible for real-world data (DeNero and Klein, 2008). As training uses a modified version of the translation decoder, it is straightforward to apply pruning as in regular decoding. Additionally, we consider three ways of approximating the full search space: 1. the single-best Viterbi alignment, 2. the n-best alignments, 3. all alignments remaining in the search space after pruning. The performance of the different approaches is measured and compared on the German-English Europarl task from the ACL 2008 Workshop on Statistical Machine Translation (WMT08). Our results show that the proposed phrase model training improves translation quality on the test set by 0.9 BLEU points over our baseline. We find that by interpolation with the heuristically extracted phrases translation performance can reach up to 1.4 BLEU improvement over the baseline on the test set. After reviewing the related work in the following section, we give a detailed description of phrasal alignment and leaving-one-out in Section 3. Section 4 explains the estimation of phrase models. The empirical evaluation of the different approaches is done in Section 5. 2 Related Work It has been pointed out in literature, that training phrase models poses some difficulties. For a generative model, (DeNero et al., 2006) gave a detailed analysis of the challenges and arising problems. They introduce a model similar to the one we propose in Section 4.2 and train it with the EM algorithm. Their results show that it can not reach a performance competitive to extracting a phrase table from word alignment by heuristics (Och et al., 1999). Several reasons are revealed in (DeNero et al., 2006). When given a bilingual sentence pair, we can usually assume there are a number of equally correct phrase segmentations and corresponding alignments. For example, it may be possible to transform one valid segmentation into another by splitting some of its phrases into sub-phrases or by shifting phrase boundaries. This is different from word-based translation models, where a typical assumption is that each target word corresponds to only one source word. As a result of this ambiguity, different segmentations are recruited for different examples during training. That in turn leads to over-fitting which shows in overly determinized estimates of the phrase translation probabilities. In addition, (DeNero et al., 2006) found that the trained phrase table shows a highly peaked distribution in opposition to the more flat distribution resulting from heuristic extraction, leaving the decoder only few translation options at decoding time. Our work differs from (DeNero et al., 2006) in a number of ways, addressing those problems. 476 To limit the effects of over-fitting, we apply the leaving-one-out and cross-validation methods in training. In addition, we do not restrict the training to phrases consistent with the word alignment, as was done in (DeNero et al., 2006). This allows us to recover from flawed word alignments. In (Liang et al., 2006) a discriminative translation system is described. For training of the parameters for the discriminative features they propose a strategy they call bold updating. It is similar to our forced alignment training procedure described in Section 3. For the hierarchical phrase-based approach, (Blunsom et al., 2008) present a discriminative rule model and show the difference between using only the viterbi alignment in training and using the full sum over all possible derivations. Forced alignment can also be utilized to train a phrase segmentation model, as is shown in (Shen et al., 2008). They report small but consistent improvements by incorporating this segmentation model, which works as an additional prior probability on the monolingual target phrase. In (Ferrer and Juan, 2009), phrase models are trained by a semi-hidden Markov model. They train a conditional “inverse” phrase model of the target phrase given the source phrase. Additionally to the phrases, they model the segmentation sequence that is used to produce a phrase alignment between the source and the target sentence. They used a phrase length limit of 4 words with longer phrases not resulting in further improvements. To counteract over-fitting, they interpolate the phrase model with IBM Model 1 probabilities that are computed on the phrase level. We also include these word lexica, as they are standard components of the phrase-based system. It is shown in (Ferrer and Juan, 2009), that Viterbi training produces almost the same results as full Baum-Welch training. They report improvements over a phrase-based model that uses an inverse phrase model and a language model. Experiments are carried out on a custom subset of the English-Spanish Europarl corpus. Our approach is similar to the one presented in (Ferrer and Juan, 2009) in that we compare Viterbi and a training method based on the ForwardBackward algorithm. But instead of focusing on the statistical model and relaxing the translation task by using monotone translation only, we use a full and competitive translation system as starting point with reordering and all models included. In (Marcu and Wong, 2002), a joint probability phrase model is presented. The learned phrases are restricted to the most frequent n-grams up to length 6 and all unigrams. Monolingual phrases have to occur at least 5 times to be considered in training. Smoothing is applied to the learned models so that probabilities for rare phrases are non-zero. In training, they use a greedy algorithm to produce the Viterbi phrase alignment and then apply a hill-climbing technique that modifies the Viterbi alignment by merge, move, split, and swap operations to find an alignment with a better probability in each iteration. The model shows improvements in translation quality over the singleword-based IBM Model 4 (Brown et al., 1993) on a subset of the Canadian Hansards corpus. The joint model by (Marcu and Wong, 2002) is refined by (Birch et al., 2006) who use high-confidence word alignments to constrain the search space in training. They observe that due to several constraints and pruning steps, the trained phrase table is much smaller than the heuristically extracted one, while preserving translation quality. The work by (DeNero et al., 2008) describes a method to train the joint model described in (Marcu and Wong, 2002) with a Gibbs sampler. They show that by applying a prior distribution over the phrase translation probabilities they can prevent over-fitting. The prior is composed of IBM1 lexical probabilities and a geometric distribution over phrase lengths which penalizes long phrases. The two approaches differ in that we apply the leaving-one-out procedure to avoid overfitting, as opposed to explicitly defining a prior distribution. 3 Alignment The training process is divided into three parts. First we obtain all models needed for a normal translations system. We perform minimum error rate training with the downhill simplex algorithm (Nelder and Mead, 1965) on the development data to obtain a set of scaling factors that achieve a good BLEU score. We then use these models and scaling factors to do a forced alignment, where we compute a phrase alignment for the training data. From this alignment we then estimate new phrase models, while keeping all other models un477 changed. In this section we describe our forced alignment procedure that is the basic training procedure for the models proposed here. 3.1 Forced Alignment The idea of forced alignment is to perform a phrase segmentation and alignment of each sentence pair of the training data using the full translation system as in decoding. What we call segmentation and alignment here corresponds to the “concepts” used by (Marcu and Wong, 2002). We apply our normal phrase-based decoder on the source side of the training data and constrain the translations to the corresponding target sentences from the training data. Given a source sentence fJ 1 and target sentence eI 1, we search for the best phrase segmentation and alignment that covers both sentences. A segmentation of a sentence into K phrase is defined by k →sk := (ik, bk, jk), for k = 1, . . . , K where for each segment ik is last position of kth target phrase, and (bk, jk) are the start and end positions of the source phrase aligned to the kth target phrase. Consequently, we can modify Equation 2 to define the best segmentation of a sentence pair as: ˆs ˆ K 1 = argmax K,sK 1 ( M X m=1 λmhm(eI 1, sK 1 , fJ 1 ) ) (3) The identical models as in search are used: conditional phrase probabilities p( ˜fk|˜ek) and p(˜ek| ˜fk), within-phrase lexical probabilities, distance-based reordering model as well as word and phrase penalty. A language model is not used in this case, as the system is constrained to the given target sentence and thus the language model score has no effect on the alignment. In addition to the phrase matching on the source sentence, we also discard all phrase translation candidates, that do not match any sequence in the given target sentence. Sentences for which the decoder can not find an alignment are discarded for the phrase model training. In our experiments, this is the case for roughly 5% of the training sentences. 3.2 Leaving-one-out As was mentioned in Section 2, previous approaches found over-fitting to be a problem in phrase model training. In this section, we describe a leaving-one-out method that can improve the phrase alignment in situations, where the probability of rare phrases and alignments might be overestimated. The training data that consists of N parallel sentence pairs fn and en for n = 1, . . . , N is used for both the initialization of the translation model p( ˜f|˜e) and the phrase model training. While this way we can make full use of the available data and avoid unknown words during training, it has the drawback that it can lead to overfitting. All phrases extracted from a specific sentence pair fn, en can be used for the alignment of this sentence pair. This includes longer phrases, which only match in very few sentences in the data. Therefore those long phrases are trained to fit only a few sentence pairs, strongly overestimating their translation probabilities and failing to generalize. In the extreme case, whole sentences will be learned as phrasal translations. The average length of the used phrases is an indicator of this kind of over-fitting, as the number of matching training sentences decreases with increasing phrase length. We can see an example in Figure 2. Without leaving-one-out the sentence is segmented into a few long phrases, which are unlikely to occur in data to be translated. Phrase boundaries seem to be unintuitive and based on some hidden structures. With leaving-one-out the phrases are shorter and therefore better suited for generalization to unseen data. Previous attempts have dealt with the overfitting problem by limiting the maximum phrase length (DeNero et al., 2006; Marcu and Wong, 2002) and by smoothing the phrase probabilities by lexical models on the phrase level (Ferrer and Juan, 2009). However, (DeNero et al., 2006) experienced similar over-fitting with short phrases due to the fact that the same word sequence can be segmented in different ways, leading to specific segmentations being learned for specific training sentence pairs. Our results confirm these findings. To deal with this problem, instead of simple phrase length restriction, we propose to apply the leavingone-out method, which is also used for language modeling techniques (Kneser and Ney, 1995). When using leaving-one-out, we modify the phrase translation probabilities for each sentence pair. For a training example fn, en, we have to remove all phrases Cn( ˜f, ˜e) that were extracted from this sentence pair from the phrase counts that 478 Figure 2: Segmentation example from forced alignment. Top: without leaving-one-out. Bottom: with leaving-one-out. we used to construct our phrase translation table. The same holds for the marginal counts Cn(˜e) and Cn( ˜f). Starting from Equation 1, the leaving-oneout phrase probability for training sentence pair n is pl1o,n( ˜f|˜e) = C( ˜f, ˜e) −Cn( ˜f, ˜e) C(˜e) −Cn(˜e) (4) To be able to perform the re-computation in an efficient way, we store the source and target phrase marginal counts for each phrase in the phrase table. A phrase extraction is performed for each training sentence pair separately using the same word alignment as for the initialization. It is then straightforward to compute the phrase counts after leaving-one-out using the phrase probabilities and marginal counts stored in the phrase table. While this works well for more frequent observations, singleton phrases are assigned a probability of zero. We refer to singleton phrases as phrase pairs that occur only in one sentence. For these sentences, the decoder needs the singleton phrase pairs to produce an alignment. Therefore we retain those phrases by assigning them a positive probability close to zero. We evaluated with two different strategies for this, which we call standard and length-based leaving-one-out. Standard leavingone-out assigns a fixed probability α to singleton phrase pairs. This way the decoder will prefer using more frequent phrases for the alignment, but is able to resort to singletons if necessary. However, we found that with this method longer singleton phrases are preferred over shorter ones, because fewer of them are needed to produce the target sentence. In order to better generalize to unseen data, we would like to give the preference to shorter phrases. This is done by length-based leavingone-out, where singleton phrases are assigned the probability β(| ˜f|+|˜e|) with the source and target Table 1: Avg. source phrase lengths in forced alignment without leaving-one-out and with standard and length-based leaving-one-out. avg. phrase length without l1o 2.5 standard l1o 1.9 length-based l1o 1.6 phrase lengths | ˜f| and |˜e| and fixed β < 1. In our experiments we set α = e−20 and β = e−5. Table 1 shows the decrease in average source phrase length by application of leaving-one-out. 3.3 Cross-validation For the first iteration of the phrase training, leaving-one-out can be implemented efficiently as described in Section 3.2. For higher iterations, phrase counts obtained in the previous iterations would have to be stored on disk separately for each sentence and accessed during the forced alignment process. To simplify this procedure, we propose a cross-validation strategy on larger batches of data. Instead of recomputing the phrase counts for each sentence individually, this is done for a whole batch of sentences at a time. In our experiments, we set this batch-size to 10000 sentences. 3.4 Parallelization To cope with the runtime and memory requirements of phrase model training that was pointed out by previous work (Marcu and Wong, 2002; Birch et al., 2006), we parallelized the forced alignment by splitting the training corpus into blocks of 10k sentence pairs. From the initial phrase table, each of these blocks only loads the phrases that are required for alignment. The align479 ment and the counting of phrases are done separately for each block and then accumulated to build the updated phrase model. 4 Phrase Model Training The produced phrase alignment can be given as a single best alignment, as the n-best alignments or as an alignment graph representing all alignments considered by the decoder. We have developed two different models for phrase translation probabilities which make use of the force-aligned training data. Additionally we consider smoothing by different kinds of interpolation of the generative model with the state-of-the-art heuristics. 4.1 Viterbi The simplest of our generative phrase models estimates phrase translation probabilities by their relative frequencies in the Viterbi alignment of the data, similar to the heuristic model but with counts from the phrase-aligned data produced in training rather than computed on the basis of a word alignment. The translation probability of a phrase pair ( ˜f, ˜e) is estimated as pFA( ˜f|˜e) = CFA( ˜f, ˜e) X ˜f′ CFA( ˜f′, ˜e) (5) where CFA( ˜f, ˜e) is the count of the phrase pair ( ˜f, ˜e) in the phrase-aligned training data. This can be applied to either the Viterbi phrase alignment or an n-best list. For the simplest model, each hypothesis in the n-best list is weighted equally. We will refer to this model as the count model as we simply count the number of occurrences of a phrase pair. We also experimented with weighting the counts with the estimated likelihood of the corresponding entry in the the n-best list. The sum of the likelihoods of all entries in an n-best list is normalized to 1. We will refer to this model as the weighted count model. 4.2 Forward-backward Ideally, the training procedure would consider all possible alignment and segmentation hypotheses. When alternatives are weighted by their posterior probability. As discussed earlier, the run-time requirements for computing all possible alignments is prohibitive for large data tasks. However, we can approximate the space of all possible hypotheses by the search space that was used for the alignment. While this might not cover all phrase translation probabilities, it allows the search space and translation times to be feasible and still contains the most probable alignments. This search space can be represented as a graph of partial hypotheses (Ueffing et al., 2002) on which we can compute expectations using the Forward-Backward algorithm. We will refer to this alignment as the full alignment. In contrast to the method described in Section 4.1, phrases are weighted by their posterior probability in the word graph. As suggested in work on minimum Bayes-risk decoding for SMT (Tromble et al., 2008; Ehling et al., 2007), we use a global factor to scale the posterior probabilities. 4.3 Phrase Table Interpolation As (DeNero et al., 2006) have reported improvements in translation quality by interpolation of phrase tables produced by the generative and the heuristic model, we adopt this method and also report results using log-linear interpolation of the estimated model with the original model. The log-linear interpolations pint( ˜f|˜e) of the phrase translation probabilities are estimated as pint( ˜f|˜e) = pH( ˜f|˜e) 1−ω · pgen( ˜f|˜e) (ω) (6) where ω is the interpolation weight, pH the heuristically estimated phrase model and pgen the count model. The interpolation weight ω is adjusted on the development corpus. When interpolating phrase tables containing different sets of phrase pairs, we retain the intersection of the two. As a generalization of the fixed interpolation of the two phrase tables we also experimented with adding the two trained phrase probabilities as additional features to the log-linear framework. This way we allow different interpolation weights for the two translation directions and can optimize them automatically along with the other feature weights. We will refer to this method as featurewise combination. Again, we retain the intersection of the two phrase tables. With good loglinear feature weights, feature-wise combination should perform at least as well as fixed interpolation. However, the results presented in Table 5 480 Table 2: Statistics for the Europarl GermanEnglish data German English TRAIN Sentences 1 311 815 Run. Words 34 398 651 36 090 085 Vocabulary 336 347 118 112 Singletons 168 686 47 507 DEV Sentences 2 000 Run. Words 55 118 58 761 Vocabulary 9 211 6 549 OOVs 284 77 TEST Sentences 2 000 Run. Words 56 635 60 188 Vocabulary 9 254 6 497 OOVs 266 89 show a slightly lower performance. This illustrates that a higher number of features results in a less reliable optimization of the log-linear parameters. 5 Experimental Evaluation 5.1 Experimental Setup We conducted our experiments on the GermanEnglish data published for the ACL 2008 Workshop on Statistical Machine Translation (WMT08). Statistics for the Europarl data are given in Table 2. We are given the three data sets TRAIN, DEV and TEST. For the heuristic phrase model, we first use GIZA++ (Och and Ney, 2003) to compute the word alignment on TRAIN. Next we obtain a phrase table by extraction of phrases from the word alignment. The scaling factors of the translation models have been optimized for BLEU on the DEV data. The phrase table obtained by heuristic extraction is also used to initialize the training. The forced alignment is run on the training data TRAIN from which we obtain the phrase alignments. Those are used to build a phrase table according to the proposed generative phrase models. Afterward, the scaling factors are trained on DEV for the new phrase table. By feeding back the new phrase table into forced alignment we can reiterate the training procedure. When training is finished the resulting phrase model is evaluated on DEV Table 3: Comparison of different training setups for the count model on DEV . leaving-one-out max phr.len. BLEU TER baseline 6 25.7 61.1 none 2 25.2 61.3 3 25.7 61.3 4 25.5 61.4 5 25.5 61.4 6 25.4 61.7 standard 6 26.4 60.9 length-based 6 26.5 60.6 and TEST. Additionally, we can apply smoothing by interpolation of the new phrase table with the original one estimated heuristically, retrain the scaling factors and evaluate afterwards. The baseline system is a standard phrase-based SMT system with eight features: phrase translation and word lexicon probabilities in both translation directions, phrase penalty, word penalty, language model score and a simple distance-based reordering model. The features are combined in a log-linear way. To investigate the generative models, we replace the two phrase translation probabilities and keep the other features identical to the baseline. For the feature-wise combination the two generative phrase probabilities are added to the features, resulting in a total of 10 features. We used a 4-gram language model with modified Kneser-Ney discounting for all experiments. The metrics used for evaluation are the case-sensitive BLEU (Papineni et al., 2002) score and the translation edit rate (TER) (Snover et al., 2006) with one reference translation. 5.2 Results In this section, we investigate the different aspects of the models and methods presented before. We will focus on the proposed leaving-oneout technique and show that it helps in finding good phrasal alignments on the training data that lead to improved translation models. Our final results show an improvement of 1.4 BLEU over the heuristically extracted phrase model on the test data set. In Section 3.2 we have discussed several methods which aim to overcome the over-fitting prob481 Figure 3: Performance on DEV in BLEU of the count model plotted against size n of n-best list on a logarithmic scale. lems described in (DeNero et al., 2006). Table 3 shows translation scores of the count model on the development data after the first training iteration for both leaving-one-out strategies we have introduced and for training without leaving-one-out with different restrictions on phrase length. We can see that by restricting the source phrase length to a maximum of 3 words, the trained model is close to the performance of the heuristic phrase model. With the application of leaving-one-out, the trained model is superior to the baseline, the length-based strategy performing slightly better than standard leaving-one-out. For these experiments the count model was estimated with a 100best list. The count model we describe in Section 4.1 estimates phrase translation probabilities using counts from the n-best phrase alignments. For smaller n the resulting phrase table contains fewer phrases and is more deterministic. For higher values of n more competing alignments are taken into account, resulting in a bigger phrase table and a smoother distribution. We can see in Figure 3 that translation performance improves by moving from the Viterbi alignment to n-best alignments. The variations in performance with sizes between n = 10 and n = 10000 are less than 0.2 BLEU. The maximum is reached for n = 100, which we used in all subsequent experiments. An additional benefit of the count model is the smaller phrase table size compared to the heuristic phrase extraction. This is consistent with the findings of (Birch et al., 2006). Table 4 shows the phrase table sizes for different n. With n = 100 we retain only 17% of the original phrases. Even for the full model, we Table 4: Phrase table size of the count model for different n-best list sizes, the full model and for heuristic phrase extraction. N # phrases % of full table 1 4.9M 5.3 10 8.4M 9.1 100 15.9M 17.2 1000 27.1M 29.2 10000 40.1M 43.2 full 59.6M 64.2 heuristic 92.7M 100.0 do not retain all phrase table entries. Due to pruning in the forced alignment step, not all translation options are considered. As a result experiments can be done more rapidly and with less resources than with the heuristically extracted phrase table. Also, our experiments show that the increased performance of the count model is partly derived from the smaller phrase table size. In Table 5 we can see that the performance of the heuristic phrase model can be increased by 0.6 BLEU on TEST by filtering the phrase table to contain the same phrases as the count model and reoptimizing the log-linear model weights. The experiments on the number of different alignments taken into account were done with standard leaving-one-out. The final results are given in Table 5. We can see that the count model outperforms the baseline by 0.8 BLEU on DEV and 0.9 BLEU on TEST after the first training iteration. The performance of the filtered baseline phrase table shows that part of that improvement derives from the smaller phrase table size. Application of crossvalidation (cv) in the first iteration yields a performance close to training with leaving-one-out (l1o), which indicates that cross-validation can be safely applied to higher training iterations as an alternative to leaving-one-out. The weighted count model clearly under-performs the simpler count model. A second iteration of the training algorithm shows nearly no changes in BLEU score, but a small improvement in TER. Here, we used the phrase table trained with leaving-one-out in the first iteration and applied cross-validation in the second iteration. Log-linear interpolation of the count model with the heuristic yields a further increase, showing an improvement of 1.3 BLEU on DEV and 1.4 BLEU on TEST over the baseline. The interpo482 Table 5: Final results for the heuristic phrase table filtered to contain the same phrases as the count model (baseline filt.), the count model trained with leaving-one-out (l1o) and cross-validation (cv), the weighted count model and the full model. Further, scores for fixed log-linear interpolation of the count model trained with leaving-one-out with the heuristic as well as a feature-wise combination are shown. The results of the second training iteration are given in the bottom row. DEV TEST BLEU TER BLEU TER baseline 25.7 61.1 26.3 60.9 baseline filt. 26.0 61.6 26.9 61.2 count (l1o) 26.5 60.6 27.2 60.5 count (cv) 26.4 60.7 27.0 60.7 weight. count 25.9 61.4 26.4 61.3 full 26.3 60.0 27.0 60.2 fixed interpol. 27.0 59.4 27.7 59.2 feat. comb. 26.8 60.1 27.6 59.9 count, iter. 2 26.4 60.3 27.2 60.0 lation weight is adjusted on the development set and was set to ω = 0.6. Integrating both models into the log-linear framework (feat. comb.) yields a BLEU score slightly lower than with fixed interpolation on both DEV and TEST. This might be attributed to deficiencies in the tuning procedure. The full model, where we extract all phrases from the search graph, weighted with their posterior probability, performs comparable to the count model with a slightly worse BLEU and a slightly better TER. 6 Conclusion We have shown that training phrase models can improve translation performance on a state-ofthe-art phrase-based translation model. This is achieved by training phrase translation probabilities in a way that they are consistent with their use in translation. A crucial aspect here is the use of leaving-one-out to avoid over-fitting. We have shown that the technique is superior to limiting phrase lengths and smoothing with lexical probabilities alone. While models trained from Viterbi alignments already lead to good results, we have demonstrated that considering the 100-best alignments allows to better model the ambiguities in phrase segmentation. The proposed techniques are shown to be superior to previous approaches that only used lexical probabilities to smooth phrase tables or imposed limits on the phrase lengths. On the WMT08 Europarl task we show improvements of 0.9 BLEU points with the trained phrase table and 1.4 BLEU points when interpolating the newly trained model with the original, heuristically extracted phrase table. In TER, improvements are 0.4 and 1.7 points. In addition to the improved performance, the trained models are smaller leading to faster and smaller translation systems. Acknowledgments This work was partly realized as part of the Quaero Programme, funded by OSEO, French State agency for innovation, and also partly based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001-06-C-0023. Any opinions, ndings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reect the views of the DARPA. References Alexandra Birch, Chris Callison-Burch, Miles Osborne, and Philipp Koehn. 2006. Constraining the phrase-based, joint probability statistical translation model. In smt2006, pages 154–157, Jun. Phil Blunsom, Trevor Cohn, and Miles Osborne. 2008. A discriminative latent variable model for statistical machine translation. In Proceedings of ACL-08: HLT, pages 200–208, Columbus, Ohio, June. Association for Computational Linguistics. P. F. Brown, V. J. Della Pietra, S. A. Della Pietra, and R. L. Mercer. 1993. The mathematics of statistical machine translation: parameter estimation. Computational Linguistics, 19(2):263–312, June. John DeNero and Dan Klein. 2008. The complexity of phrase alignment problems. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics on Human Language Technologies: Short Papers, pages 25–28, Morristown, NJ, USA. Association for Computational Linguistics. John DeNero, Dan Gillick, James Zhang, and Dan Klein. 2006. Why Generative Phrase Models Underperform Surface Heuristics. In Proceedings of the 483 Workshop on Statistical Machine Translation, pages 31–38, New York City, June. John DeNero, Alexandre Buchard-Cˆot´e, and Dan Klein. 2008. Sampling Alignment Structure under a Bayesian Translation Model. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 314–323, Honolulu, October. Nicola Ehling, Richard Zens, and Hermann Ney. 2007. Minimum bayes risk decoding for bleu. In ACL ’07: Proceedings of the 45th Annual Meeting of the ACL on Interactive Poster and Demonstration Sessions, pages 101–104, Morristown, NJ, USA. Association for Computational Linguistics. Jes´us-Andr´es Ferrer and Alfons Juan. 2009. A phrasebased hidden semi-markov approach to machine translation. In Procedings of European Association for Machine Translation (EAMT), Barcelona, Spain, May. European Association for Machine Translation. Reinhard Kneser and Hermann Ney. 1995. Improved Backing-Off for M-gram Language Modelling. In IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP), pages 181–184, Detroit, MI, May. Philipp Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology - Volume 1, pages 48–54, Morristown, NJ, USA. Association for Computational Linguistics. Percy Liang, Alexandre Buchard-Cˆot´e, Dan Klein, and Ben Taskar. 2006. An End-to-End Discriminative Approach to Machine Translation. In Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the Association for Computational Linguistics, pages 761– 768, Sydney, Australia. Daniel Marcu and William Wong. 2002. A phrasebased, joint probability model for statistical machine translation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP-2002), July. J.A. Nelder and R. Mead. 1965. A Simplex Method for Function Minimization. The Computer Journal), 7:308–313. Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1):19–51, March. Franz Josef Och and Hermann Ney. 2004. The alignment template approach to statistical machine translation. Computational Linguistics, 30(4):417–449, December. F.J. Och, C. Tillmann, and H. Ney. 1999. Improved alignment models for statistical machine translation. In Proc. of the Joint SIGDAT Conf. on Empirical Methods in Natural Language Processing and Very Large Corpora (EMNLP99), pages 20–28, University of Maryland, College Park, MD, USA, June. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, pages 311–318, Morristown, NJ, USA. Association for Computational Linguistics. Wade Shen, Brian Delaney, Tim Anderson, and Ray Slyh. 2008. The MIT-LL/AFRL IWSLT-2008 MT System. In Proceedings of IWSLT 2008, pages 69– 76, Hawaii, U.S.A., October. Matthew Snover, Bonnie Dorr, Rich Schwartz, Linnea Micciulla, and John Makhoul. 2006. A study of translation edit rate with targeted human annotation. In Proc. of AMTA, pages 223–231, Aug. Roy Tromble, Shankar Kumar, Franz Och, and Wolfgang Macherey. 2008. Lattice Minimum BayesRisk decoding for statistical machine translation. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 620–629, Honolulu, Hawaii, October. Association for Computational Linguistics. N. Ueffing, F.J. Och, and H. Ney. 2002. Generation of word graphs in statistical machine translation. In Proc. of the Conference on Empirical Methods for Natural Language Processing, pages 156–163, Philadelphia, PA, USA, July. 484
|
2010
|
49
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 40–49, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Identifying Generic Noun Phrases Nils Reiter and Anette Frank Department of Computational Linguistics Heidelberg University, Germany {reiter,frank}@cl.uni-heidelberg.de Abstract This paper presents a supervised approach for identifying generic noun phrases in context. Generic statements express rulelike knowledge about kinds or events. Therefore, their identification is important for the automatic construction of knowledge bases. In particular, the distinction between generic and non-generic statements is crucial for the correct encoding of generic and instance-level information. Generic expressions have been studied extensively in formal semantics. Building on this work, we explore a corpus-based learning approach for identifying generic NPs, using selections of linguistically motivated features. Our results perform well above the baseline and existing prior work. 1 Introduction Generic expressions come in two basic forms: generic noun phrases and generic sentences. Both express rule-like knowledge, but in different ways. A generic noun phrase is a noun phrase that does not refer to a specific (set of) individual(s), but rather to a kind or class of individuals. Thus, the NP The lion in (1.a)1 is understood as a reference to the class “lion” instead of a specific individual. Generic NPs are not restricted to occur with kind-related predicates as in (1.a). As seen in (1.b), they may equally well be combined with predicates that denote specific actions. In contrast to (1.a), the property defined by the verb phrase in (1.b) may hold of individual lions. (1) a. The lion was the most widespread mammal. b. Lions eat up to 30 kg in one sitting. 1All examples are taken from Wikipedia unless stated otherwise. Generic sentences are characterising sentences that quantify over situations or events, expressing rule-like knowledge about habitual actions or situations (2.a). This is in contrast with sentences that refer to specific events and individuals, as in (2.b). (2) a. After 1971 [Paul Erd˝os] also took amphetamines. b. Paul Erd˝os was born [...] on March 26, 1913. The genericity of an expression may arise from the generic (kind-referring, class-denoting) interpretation of the NP or the characterising interpretation of the sentence predicate. Both sources may concur in a single sentence, as illustrated in Table 1, where we have cross-classified the examples above according to the genericity of the NP and the sentence. This classification is extremely difficult, because (i) the criteria for generic interpretation are far from being clear-cut and (ii) both sources of genericity may freely interact. S[gen+] S[gen-] NP[gen+] (1.b) (1.a) NP[gen-] (2.a) (2.b) Table 1: Generic NPs and generic sentences The above classification of generic expressions is well established in traditional formal semantics (cf. Krifka et al. (1995))2. As we argue in this paper, these distinctions are relevant for semantic processing in computational linguistics, especially for information extraction and ontology learning and population tasks. With appropriate semantic analysis of generic statements, we can not only formally capture and exploit generic knowledge, 2The literature draws some finer distinctions including aspects like specificity, which we will ignore in this work. 40 but also distinguish between information pertaining to individuals vs. classes. We will argue that the automatic identification of generic expressions should be cast as a machine learning problem instead of a rule-based approach, as there is (i) no transparent marking of genericity in English (as in most other European languages) and (ii) the phenomenon is highly context dependent. In this paper, we build on insights from formal semantics to establish a corpus-based machine learning approach for the automatic classification of generic expressions. In principle our approach is applicable to the detection of both generic NPs and generic sentences, and in fact it would be highly desirable and possibly advantageous to cover both types of genericity simultaneously. Our current work is confined to generic NPs, as there are no corpora available at present that contain annotations for genericity at the sentence level. The paper is organised as follows. Section 2 introduces generic expressions and motivates their relevance for knowledge acquisition and semantic processing tasks in computational linguistics. Section 3 reviews prior and related work. In section 4 we motivate the choice of feature sets for the automatic identification of generic NPs in context. Sections 5 and 6 present our experiments and results obtained for this task on the ACE-2 data set. Section 7 concludes. 2 Generic Expressions & their Relevance for Computational Linguistics 2.1 Interpretation of generic expressions Generic NPs There are two contrasting views on how to formally interpret generic NPs. According to the first one, a generic NP involves a special form of quantification. Quine (1960), for example, proposes a universally quantified reading for generic NPs. This view is confronted with the most important problem of all quantificationbased approaches, namely that the exact determination of the quantifier restriction (QR) is highly dependent on the context, as illustrated in (3)3. (3) a. Lions are mammals. QR: all lions b. Mammals give birth to live young. QR: less than half of all mammals 3Some of these examples are taken from Carlson (1977). c. Rats are bothersome to people. QR: few rats4 In view of this difficulty, several approaches restrict the quantification to only “relevant” (Declerck, 1991) or “normal” (Dahl, 1975) individuals. According to the second view, generic noun phrases denote kinds. Following Carlson (1977), a kind can be considered as an individual that has properties on its own. On this view, the generic NP cannot be analysed as a quantifier over individuals pertaining to the kind. For some predicates, this is clearly marked. (1.a), for instance, attributes a property to the kind lion that cannot be attributed to individual lions. Generic sentences are usually analysed using a special dyadic operator, as first proposed by Heim (1982). The dyadic operator relates two semantic constituents, the restrictor and the matrix: Q[x1, ..., xi]([x1, ..., xi] | {z } Restrictor ; ∃y1, ..., yi[x1, .., xi, y1, ..., yi] | {z } Matrix ) By choosing GEN as a generic dyadic operator, it is possible to represent the two readings (a) and (b) of the characterising sentence (4) by variation in the specification of restrictor and matrix (Krifka et al., 1995). (4) Typhoons arise in this part of the pacific. (a) Typhoons in general have a common origin in this part of the pacific. (b) There arise typhoons in this part of the pacific. (a’) GEN[x; y](Typhoon(x);this-part-of-thepacific(y)∧arise-in(x, y)) (b’) GEN[x; y](this-part-of-thepacific(x);Typhoon(y)∧arise-in(y, x)) In order to cope with characterising sentences as in (2.a), we must allow the generic operator to quantify over situations or events, in this case, “normal” situations which were such that Erd˝os took amphetamines. 2.2 Relevance for computational linguistics Knowledge acquisition The automatic acquisition of formal knowledge for computational applications is a major endeavour in current research 4Most rats are not even noticed by people. 41 and could lead to big improvements of semanticsbased processing. Bos (2009), e.g., describes systems using automated deduction for language understanding tasks using formal knowledge. There are manually built formal ontologies such as SUMO (Niles and Pease, 2001) or Cyc (Lenat, 1995) and linguistic ontologies like WordNet (Fellbaum, 1998) that capture linguistic and world knowledge to a certain extent. However, these resources either lack coverage or depth. Automatically constructed ontologies or taxonomies, on the other hand, are still of poor quality (Cimiano, 2006; Ponzetto and Strube, 2007). Attempts to automatically induce knowledge bases from text or encyclopaedic sources are currently not concerned with the distinction between generic and non-generic expressions, concentrating mainly on factual knowledge. However, rulelike knowledge can be found in textual sources in the form of generic expressions5. In view of the properties of generic expressions discussed above, this lack of attention bears two types of risks. The first concerns the distinction between classes and instances, regarding the attribution of properties. The second concerns modelling exceptions in both representation and inferencing. The distinction between classes and instances is a serious challenge even for the simplest methods in automatic ontology construction, e.g., Hearst (1992) patterns. The so-called IS-A patterns do not only identify subclasses, but also instances. Shakespeare, e.g., would be recognised as a hyponym of author in the same way as temple is recognised as a hyponym of civic building. Such a missing distinction between classes and instances is problematic. First, there are predicates that can only attribute properties to a kind (1.a). Second, even for properties that in principle can be attributed to individuals of the class, this is highly dependent on the selection of the quantifier’s restriction in context (3). In both cases, it holds that properties attributed to a class are not necessarily 5In the field of cognitive science, research on the acquisition of generic knowledge in humans has shown that adult speakers tend to use generic expressions very often when talking to children (Pappas and Gelman, 1998). We are not aware of any detailed assessment of the proportion of generic noun phrases in educational text genres or encyclopaedic resources like Wikipedia. Concerning generic sentences, Mathew and Katz (2009) report that 19.9% of the sentences in their annotated portion of the Penn Treebank are habitual (generic) and 80.1% episodic (non-generic). inherited by any or all instances pertaining to the class. Zirn et al. (2008) are the first to present fully automatic, heuristic methods to distinguish between classes and instances in the Wikipedia taxonomy derived by Ponzetto and Strube (2007). They report an accuracy of 81.6% and 84.5% for different classification schemes. However, apart from a plural feature, all heuristics are tailored to specific properties of the Wikipedia resource. Modelling exceptions is a cumbersome but necessary problem to be handled in ontology building, be it manually or by automatic means, and whether or not the genericity of knowledge is formalised explicitly. In artificial intelligence research, this area has been tackled for many years. Default reasoning (Reiter, 1980) is confronted with severe efficiency problems and therefore has not extended beyond experimental systems. However, the emerging paradigm of Answer Set Programming (ASP, Lifschitz (2008)) seems to be able to model exceptions efficiently. In ASP a given problem is cast as a logic program, and an answer set solver calculates all possible answer sets, where an answer set corresponds to a solution of the problem. Efficient answer set solvers have been proposed (Gelfond, 2007). Although ASP may provide us with very efficient reasoning systems, it is still necessary to distinguish and mark default rules explicitly (Lifschitz, 2002). Hence, the recognition of generic expressions is an important precondition for the adequate representation and processing of generic knowledge. 3 Prior Work Suh (2006) applied a rule-based approach to automatically identify generic noun phrases. Suh used patterns based on part of speech tags that identify bare plural noun phrases, reporting a precision of 28.9% for generic entities, measured against an annotated corpus, the ACE 2005 (Ferro et al., 2005). Neither recall nor f-measure are reported. To our knowledge, this is the single prior work on the task of identifying generic NPs. Next to the ACE corpus (described in more detail below), Herbelot and Copestake (2008) offer a study on annotating genericity in a corpus. Two annotators annotated 48 noun phrases from the British National Corpus for their genericity (and specificity) properties, obtaining a kappa value of 0.744. Herbelot and Copestake (2008) leave su42 pervised learning for the identification of generic expressions as future work. Recent work by Mathew and Katz (2009) presents automatic classification of generic and non-generic sentences, yet restricted to habitual interpretations of generic sentences. They use a manually annotated part of the Penn TreeBank as training and evaluation set6. Using a selection of syntactic and semantic features operating mainly on the sentence level, they achieved precision between 81.2% and 84.3% and recall between 60.6% and 62.7% for the identification of habitual generic sentences. 4 Characterising Generic Expressions for Automatic Classification 4.1 Properties of generic expressions Generic NPs come in various syntactic forms. These include definite and indefinite singular count nouns, bare plural count and singular and plural mass nouns as in (5.a-f). (5.f) shows a construction that makes the kind reading unambiguous. As Carlson (1977) observed, the generic reading of “well-established” kinds seems to be more prominent (g vs. h). (5) a. The lion was the most widespread mammal. b. A lioness is weaker [...] than a male. c. Lions died out in northern Eurasia. d. Metals are good conductors. e. Metal is also used for heat sinks. f. The zoo has one kind of tiger. g. The Coke bottle has a narrow neck. h. The green bottle has a narrow neck. Apart from being all NPs, there is no obvious syntactic property that is shared by all examples. Similarly, generic sentences come in a range of syntactic forms (6). (6) a. John walks to work. b. John walked to work (when he lived in California). c. John will walk to work (when he moves to California). 6The corpus has not been released. Although generic NPs and generic sentences can be combined freely (cf. Section 1; Table 1), both phenomena highly interact and quite often appear in the same sentence (Krifka et al., 1995). Also, genericity is highly dependent on contextual factors. Present tense, e.g., may be indicative for genericity, but with appropriate temporal modification, generic sentences may occur in past or future tense (6). Presence of a copular construction as in (5.a,b,d) may indicate a generic NP reading, but again we find generic NPs with event verbs, as in (5.e) or (1.b). Lexical semantic factors, such as the semantic type of the clause predicate (5.c,e), or “well-established” kinds (5.g) may favour a generic reading, but such lexical factors are difficult to capture in a rule-based setting. In our view, these observations call for a corpusbased machine learning approach that is able to capture a variety of factors indicating genericity in combination and in context. 4.2 Feature set and feature classes In Table 2 we give basic information about the individual features we investigate for identifying generic NPs. In the following, we will structure this feature space along two dimensions, distinguishing NP- and sentence-level factors as well as syntactic and semantic (including lexical semantic) factors. Table 3 displays the grouping into corresponding feature classes. NP-level features are extracted from the local NP without consideration of the sentence context. Sentence-level features are extracted from the clause (in which the NP appears), as well as sentential and non-sentential adjuncts of the clause. We also included the (dependency) relations between the target NP and its governing clause. Syntactic features are extracted from a parse tree or shallow surface-level features. The feature set includes NP-local and global features. Semantic features include semantic features abstracted from syntax, such as tense and aspect or type of modification, but also lexical semantic features such as word sense classes, sense granularity or verbal predicates. Our aim is to determine indicators for genericity from combinations of these feature classes. 43 Feature Description Number sg, pl Person 1, 2, 3 Countability ambig, no noun, count, uncount Noun Type common, proper, pronoun Determiner Type def, indef, demon Granularity The number of edges in the WordNet hypernymy graph between the synset of the entity and a top node Part of Speech POS-tag (Penn TreeBank tagset; Marcus et al. (1993)) of the head of the phrase Bare Plural false, true Sense[0-3] WordNet sense. Sense[0] represents the sense of the head of the entity, Sense[1] its direct hypernym sense and so forth. Sense[Top] The top sense in the hypernym hierarchy (often referred to as “super sense”) Dependency Relation [0-4] Dependency Relations. Relation[0] represents the relation between entity and its governor, Relation[1] the relation between the governor and its governor and so forth. Embedded Predicate.Pred Lemma of the head of the directly governing predicate of the entity C.Tense past, pres, fut C.Progressive false, true C.Perfective false, true C.Mood indicative, imperative, subjunctive C.Passive false, true C.Temporal Modifier? false, true C.Number of Modifiers numeric C.Part of Speech POS-tag (Penn TreeBank tagset; Marcus et al. (1993)) of the head of the phrase C.Pred Lemma of the head of the clause C.Adjunct.Time true, false C.Adjunct.VType main, copular C.Adjunct.Adverbial Type vpadv, sadv C.Adjunct.Degree positive, comparative, superlative C.Adjunct.Pred Lemma of the head of the adjunct of the clause XLE.Quality How complete is the parse by the XLE parser? fragmented, complete, no parse Table 2: The features used in our system. C stands for the clause in which the noun phrase appears, “Embedding Predicate” its direct predicate. In most cases, we just give the value range, if necessary, we give descriptions. All features may have a NULL value. Syntactic Semantic NP-level Number, Person, Part of Speech, Determiner Type, Bare Plural Countability, Granularity, Sense[0-3, Top] S-level Clause.{Part of Speech, Passive, Number of Modifiers}, Dependency Relation[0-4], Clause.Adjunct.{Verbal Type, Adverbial Type}, XLE.Quality Clause.{Tense, Progressive, Perfective, Mood, Pred, Has temporal Modifier}, Clause.Adjunct.{Time, Pred}, Embedded Predicate.Pred Table 3: Feature classes Name Descriptions and Features Set 1 Five best single features: Bare Plural, Person, Sense [0], Clause.Pred, Embedding Predicate.Pred Set 2 Five best feature tuples: a. Number, Part of Speech b. Countability, Part of Speech c. Sense [0], Part of Speech d. Number, Countability e. Noun Type, Part of Speech Set 3 Five best feature triples: a. Number, Clause.Tense, Part of Speech b. Number, Clause.Tense, Noun Type c. Number, Clause.Part of Speech, Part of Speech d. Number, Part of Speech, Noun Type e. Number, Clause.Part of Speech, Noun Type Set 4 Features, that appear most often among the single, tuple and triple tests: Number, Noun Type, Part of Speech, Clause.Tense, Clause.Part of Speech, Clause.Pred, Embedding Predicate.Pred, Person, Sense [0], Sense [1], Sense[2] Set 5 Features performing best in the ablation test: Number, Person, Clause.Part of Speech, Clause.Pred, Embedding Predicate.Pred, Clause.Tense, Determiner Type, Part of Speech, Bare Plural, Dependency Relation [2], Sense [0] Table 4: Derived feature sets 44 5 Experiments 5.1 Dataset As data set we are using the ACE-2 (Mitchell et al., 2003) corpus, a collection of newspaper texts annotated with entities marked for their genericity. In this version of the corpus, the classification of entities is a binary one. Annotation guidelines The ACE-2 annotation guidelines describe generic NPs as referring to an arbitrary member of the set in question, rather than to a particular individual. Thus, a property attributed to a generic NP is in principle applicable to arbitrary members of the set (although not to all of them). The guidelines list several tests that are either local syntactic tests involving determiners or tests that cannot be operationalised as they involve world knowledge and context information. The guidelines give a number of criteria to identify generic NPs referring to specific properties. These are (i) types of entities (lions in 3.a), (ii) suggested attributes of entities (mammals in 3.a), (iii) hypothetical entities (7) and (iv) generalisations across sets of entities (5.d). (7) If a person steps over the line, they must be punished. The general description of generic NPs as denoting arbitrary members of sets obviously does not capture kind-referring readings. However, the properties characterised (i) can be understood to admit kinds. Also, some illustrations in the guidelines explicitly characterise kind-referring NPs as generic. Thus, while at first sight the guidelines do not fully correspond to the characterisation of generics we find in the formal semantics literature, we argue that both characterisations have similar extensions, i.e., include largely overlapping sets of noun phrases. In fact, all of the examples for generic noun phrases presented in this paper would also be classified as generic according to the ACE-2 guidelines. We also find annotated examples of generic NPs that are not discussed in the formal semantics literature (8.a), but that are well captured by the ACE-2 guidelines. However, there are also cases that are questionable (8.b). (8) a. “It’s probably not the perfect world, but you kind of have to deal with what you have to work with,” he said. b. Even more remarkable is the Internet, where information of all kinds is available about the government and the economy. This shows that the annotation of generics is difficult, but also highlights the potential benefit of a corpus-driven approach that allows us to gather a wider range of realisations. This in turn can contribute to novel insights and discussion. Data analysis A first investigation of the corpus shows that generic NPs are much less common than non-generic ones, at least in the newspaper genre at hand. Of the 40,106 annotated entities, only 5,303 (13.2%) are marked as generic. In order to control for bias effects in our classifier, we will experiment with two different training sets, a balanced and an unbalanced one. 5.2 Preprocessing The texts have been (pre-)processed to add several layers of linguistic annotation (Table 5). We use MorphAdorner for sentence splitting and TreeTagger with the standard parameter files for part of speech tagging and lemmatisation. As we do not have a word sense disambiguation system available that outperforms the most frequent sense baseline, we simply used the most frequent sense (MFS). The countability information is taken from Celex. Parsing was done using the English LFG grammar (cf. Butt et al. (2002)) in the XLE parsing platform and the Stanford Parser. Task Tool Sentence splitting MorphAdorner 7 POS, lemmatisation TreeTagger (Schmid, 1994) WSD MFS (according to WordNet 3.0) Countability Celex (Baayen et al., 1996) Parsing XLE (Crouch et al., 2010) Stanford (Klein and Manning, 2003) Table 5: Preprocessing pipeline As the LFG-grammar produced full parses only for the sentences of 56% of the entities (partial parses: 37% of the entities), we chose to integrate the Stanford parser as a fallback. If we are unable to extract feature values from the f-structure produced by the XLE parser, we extract them from the Stanford Parser, if possible. Experimentation showed using the two parsers in tandem yields best results, compared to individual use. 7http://morphadorner.northwestern.edu 45 Feature Set Generic Non generic Overall P R F P R F P R F Baseline Majority 0 0 0 86.8 100 92.9 75.3 86.8 80.6 Baseline Person 60.5 10.2 17.5 87.9 99.0 93.1 84.3 87.2 85.7 Baseline Suh 28.9 Feature Classes Unbalanced NP 31.7 56.6 40.7 92.5 81.4 86.6 84.5 78.2 81.2 S 32.2 50.7 39.4 91.8 83.7 87.6 83.9 79.4 81.6 NP/Syntactic 39.2 58.4 46.9 93.2 86.2 89.5 86.0 82.5 84.2 S/Syntactic 31.9 22.1 26.1 88.7 92.8 90.7 81.2 83.5 82.3 NP/Semantic 28.2 53.5 36.9 91.8 79.2 85 83.4 75.8 79.4 S/Semantic 32.1 36.6 34.2 90.1 88.2 89.2 82.5 81.4 81.9 Syntactic 40.1 66.6 50.1 94.3 84.8 89.3 87.2 82.4 84.7 Semantic 34.5 56.0 42.7 92.6 83.8 88.0 84.9 80.1 82.4 All 37.0 72.1 49.0 81.3 87.6 87.4 80.1 80.1 83.6 Balanced NP 30.1 71.0 42.2 94.4 74.8 83.5 85.9 74.3 79.7 S 26.9 73.1 39.3 94.4 69.8 80.3 85.5 70.2 77.1 NP/Syntactic 35.4 76.3 48.4 95.6 78.8 86.4 87.7 78.5 82.8 S/Syntactic 23.1 77.1 35.6 94.6 61.0 74.2 85.1 63.1 72.5 NP/Semantic 24.7 60.0 35.0 92.2 72.1 80.9 83.3 70.5 76.4 S/Semantic 26.4 66.3 37.7 93.3 71.8 81.2 84.5 71.1 77.2 Syntactic 30.8 85.3 45.3 96.9 70.8 81.9 88.2 72.8 79.7 Semantic 30.1 67.5 41.6 93.9 76.1 84.1 85.5 75.0 79.9 All 33.7 81.0 47.6 96.3 75.8 84.8 88.0 76.5 81.8 Feature Selection Unbalanced Set 1 49.5 37.4 42.6 90.8 94.2 92.5 85.3 86.7 86.0 Set 2a 37.3 42.7 39.8 91.1 89.1 90.1 84.0 82.9 83.5 Set 3a 42.6 54.1 47.7 92.7 88.9 90.8 86.1 84.3 85.2 Set 4 42.7 69.6 52.9 94.9 85.8 90.1 88.0 83.6 85.7 Set 5 45.7 64.8 53.6 94.3 88.3 91.2 87.9 85.2 86.5 Balanced Set 1 29.7 71.1 41.9 94.4 74.4 83.2 85.9 73.9 79.5 Set 2a 36.5 70.5 48.1 94.8 81.3 87.5 87.1 79.8 83.3 Set 3a 36.2 70.8 47.9 94.8 81.0 87.4 87.1 79.7 83.2 Set 4 35.9 83.1 50.1 96.8 77.4 86.0 88.7 78.2 83.1 Set 5 37.0 81.9 51.0 96.6 78.7 86.8 88.8 79.2 83.7 Table 6: Results of the classification, using different feature and training sets 5.3 Experimental setup Given the unclear dependencies of features, we chose to use a Bayesian network. A Bayesian network represents the dependencies of random variables in a directed acyclic graph, where each node represents a random variable and each edge a dependency between variables. In fact, a number of feature selection tests uncovered feature dependencies (see below). We used the Weka (Witten and Frank, 2002) implementation BayesNet in all our experiments. To control for bias effects, we created balanced data sets by oversampling the number of generic entities and simultaneously undersampling nongeneric entities. This results in a dataset of 20,053 entities with approx. 10,000 entities for each class. All experiments are performed on balanced and unbalanced data sets using 10-fold crossvalidation, where balancing has been performed for each training fold separately (if any). Feature classes We performed evaluation runs for different combinations of feature sets: NP- vs. S-level features (with further distinction between syntactic and semantic NP-/S-level features), as well as overall syntactic vs. semantic features. This was done in order to determine the effect of different types of linguistic factors for the detection of genericity (cf. Table 3). 46 Feature selection We experimented with two methods for feature selection. Table 4 shows the resulting feature sets. In ablation testing, a single feature in turn is temporarily omitted from the feature set. The feature whose omission causes the biggest drop in fmeasure is set aside as a strong feature. This process is repeated until we are left with an empty feature set. From the ranked list of features f1 to fn we evaluate increasingly extended feature sets f1..fi for i = 2..n. We select the feature set that yields the best balanced performance, at 45.7% precision and 53.6% f-measure. The features are given as Set 5 in Table 4. As ablation testing does not uncover feature dependencies, we also experimented with single, tuple and triple feature combinations to determine features that perform well in combination. We ran evaluations using features in isolation and each possible pair and triple of features. We select the resulting five best features, tuples and triples of features. The respective feature sets are given as Set 1 to Set 3 in Table 4. The features that appear most often in Set 1 to Set 3 are grouped in Set 4. Baseline Our results are evaluated against three baselines. Since the class distribution is unequal, a majority baseline consists in classifying each entity as non-generic. As a second baseline we chose the performance of the feature Person, as this feature gave the best performance in precision among those that are similarly easy to extract. Finally, we compare our results to (Suh, 2006). 6 Results and Discussion The results of classification are summarised in Table 6. The columns Generic and Non-generic give the results for the respective class. Overall shows the weighted average of the classes. Comparison to baselines Given the bias for non-generic NPs in the unbalanced data, the majority baseline achieves high performance overall (F: 80.6). Of course, it does not detect any generic NPs. The Person-based baseline also suffers from very low recall (R: 10.2%), but achieves the highest precision (P: 60.5 %). (Suh, 2006) reported only precision of the generic class, so we can only compare against this value (28.9 %). Most of the features and feature sets yield precision values above the results of Suh. Feature classes, unbalanced data For the identification of generic NPs, syntactic features achieve the highest precision and recall (P: 40.1%, R: 66.6 %). Using syntactic features on the NPor sentence-level only, however, leads to a drop in precision as well as recall. The recall achieved by syntactic features can be improved at the cost of precision by adding semantic features (R: 66.6 → 72.1, P: 40.1 →37). Semantic features in separation perform lower than the syntactic ones, in terms of recall and precision. Even though our results achieve a lower precision than the Person baseline, in terms of fmeasure, we achieve a result of over 50%, which is almost three times the baseline. Feature classes, balanced data Balancing the training data leads to a moderate drop in performance. All feature classes perform lower than on the unbalanced data set, yielding an increase in recall and a drop in precision. The overall performance differences between the balanced and unbalanced data for the best achieved values for the generic class are -4.7 (P), +13.2 (R) and -1.7 (F). This indicates that (i) the features prove to perform rather effectively, and (ii) the distributional bias in the data can be exploited in practical experiments, as long as the data distribution remains constant. We observe that generally, the recall for the generic class improves for the balanced data. This is most noticeable for the S-level features with an increase of 55 (syntactic) and 29.7 (semantic). This could indicate that S-level features are useful for detecting genericity, but are too sparse in the non-oversampled data to become prominent. This holds especially for the lexical semantic features. As a general conclusion, syntactic features prove most important in both setups. We also observe that the margin between syntactic and semantic features reduces in the balanced dataset, and that both NP- and S-level features contribute to classification performance, with NP-features generally outperforming the S-level features. This confirms our hypothesis that all feature classes contribute important information. Feature selection While the above figures were obtained for the entire feature space, we now discuss the effects of feature selection both on performance and the distribution over feature classes. The results for each feature set are given in Table 6. In general, we find a behaviour similar to 47 Syntactic Semantic NP Number, Person, Part of Speech, Determiner Type, Bare Plural Sense[0] S Clause.Part of Speech, Dependency Relation[2] Clause.{Tense, Pred} Table 7: Best performing features by feature class the homogeneous classes, in that balanced training data increases recall at the cost of precision. With respect to overall f-measure, the best single features are strong on the unbalanced data. They even yield a relatively high precision for the generic NPs (49.5%), the highest value among the selected feature sets. This, however, comes at the price of one of the lowest recalls. The best performing feature in terms of f-measure on both balanced and unbalanced data is Set 5 with Set 4 as a close follow-up. Set 5 achieves an f-score of 53.6 (unbalanced) and 51.0 (balanced). The highest recall is achieved using Set 4 (69.6% on the unbalanced and 83.1% on the balanced dataset). The results for Set 5 represent an improvement of 3.5 respectively 2.6 (unbalanced and balanced) over the best achieved results on homogeneous feature classes. In fact, Table 7 shows that these features, selected by ablation testing, distribute over all homogeneous classes. We trained a decision tree to gain insights into the dependencies among these features. Figure 1 shows an excerpt of the obtained tree. The classifier learned to classify singular proper names as non-generic, while the genericity of singular nouns depends on their predicate. At this point, the classifier can correctly classify some of the NPs in (5) as kind-referring (given the training data contains predicates like “widespread”, “die out”, ...). 7 Conclusions and Future Work This paper addresses a linguistic phenomenon that has been thoroughly studied in the formal semantics literature but only recently is starting to be addressed as a task in computational linguistics. We presented a data-driven machine learning approach for identifying generic NPs in context that in turn can be used to improve tasks such as knowledge acquisition and organisation. The classification of generic NPs has proven difficult even for humans. Therefore, a machine learning approach seemed promising, both for the identification of relevant features as for capturing contexFigure 1: A decision tree trained on feature Set 5 tual factors. We explored a range of features using homogeneous and mixed classes gained by alternative methods of feature selection. In terms of f-measure on the generic class, all feature sets performed above the baseline(s). In the overall classification, the selected sets perform above the majority and close to or above the Person baseline. The final feature set that we established characterises generic NPs as a phenomenon that exhibits both syntactic and semantic as well as sentenceand NP-level properties. Although our results are satisfying, in future work we will extend the range of features for further improvements. In particular, we will address lexical semantic features, as they tend to be effected by sparsity. As a next step, we will apply our approach to the classification of generic sentences. Treating both cases simultaneously could reveal insights into dependencies between them. The classification of generic expressions is only a first step towards a full treatment of the challenges involved in their semantic processing. As discussed, this requires a contextually appropriate selection of the quantifier restriction8, as well as determining inheritance of properties from classes to individuals and the formalisation of defaults. References R. Harald Baayen, Richard Piepenbrock, and Leon Gulikers. 1996. CELEX2. Linguistic Data Consortium, Philadelphia. Johan Bos. 2009. Applying automated deduction to natural language understanding. Journal of Applied 8Consider example (1.a), which is contextually restricted to a certain time and space. 48 Logic, 7(1):100 – 112. Special Issue: Empirically Successful Computerized Reasoning. Miriam Butt, Helge Dyvik, Tracy Holloway King, Hiroshi Marsuichi, and Christian Rohrer. 2002. The Parallel Grammar Project. In Proceedings of Grammar Engineering and Evaluation Workshop. Gregory Norman Carlson. 1977. Reference to Kinds in English. Ph.D. thesis, University of Massachusetts. Philipp Cimiano. 2006. Ontology Learning and Populating from Text. Springer. Dick Crouch, Mary Dalrymple, Ron Kaplan, Tracy King, John Maxwell, and Paula Newman, 2010. XLE Documentation. www2.parc.com/isl/groups/nltt/xle/doc/xle toc.html. ¨Osten Dahl. 1975. On Generics. In Edward Keenan, editor, Formal Semantics of Natural Language, pages 99–111. Cambridge University Press, Cambridge. Renaat Declerck. 1991. The Origins of Genericity. Linguistics, 29:79–102. Christiane Fellbaum. 1998. WordNet: An Electronic Lexical Database. MIT Press. Lisa Ferro, Laurie Gerber, Janet Hitzeman, Elizabeth Lima, and Beth Sundheim. 2005. ACE English Training Data. Linguistic Data Consortium, Philadelphia. Michael Gelfond. 2007. Answer sets. In Handbook of Knowledge Representation. Elsevier Science. Marti A. Hearst. 1992. Automatic acquisition of hyponyms from large text corpora. In Proceedings of the 14th International Conference on Computational Linguistics, pages 539–545. Irene Heim. 1982. The Semantics of Definite and Indefinite Noun Phrases. Ph.D. thesis, University of Massachusetts, Amherst. Aurelie Herbelot and Ann Copestake. 2008. Annotating genericity: How do humans decide? (a case study in ontology extraction). In Sam Featherston and Susanne Winkler, editors, The Fruits of Empirical Linguistics, volume 1. de Gruyter. Dan Klein and Christopher Manning. 2003. Accurate unlexicalized parsing. In Proceedings of the 41st Meeting of the Association for Computational Linguistics, pages 423–430. Manfred Krifka, Francis Jeffry Pelletier, Gregory N. Carlson, Alice ter Meulen, Gennaro Chierchia, and Godehard Link. 1995. Genericity: An Introduction. In Gregory Norman Carlson and Francis Jeffry Pelletier, editors, The Generic Book. University of Chicago Press, Chicago. Douglas B. Lenat. 1995. Cyc: a large-scale investment in knowledge infrastructure. Commun. ACM, 38(11):33–38. Vladimir Lifschitz. 2002. Answer set programming and plan generation. Artificial Intelligence, 138(12):39 – 54. Vladimir Lifschitz. 2008. What is Answer Set Programming? In Proceedings of AAAI. Mitchell P. Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. 1993. Building a large annotated corpus of English: the Penn treebank. Computational Linguistics, 19(2):313–330. Thomas Mathew and Graham Katz. 2009. Supervised Categorization of Habitual and Episodic Sentences. In Sixth Midwest Computational Linguistics Colloquium. Bloomington, Indiana: Indiana University. Alexis Mitchell, Stephanie Strassel, Mark Przybocki, JK Davis, George Doddington, Ralph Grishman, Adam Meyers, Ada Brunstein, Lisa Ferro, and Beth Sundheim. 2003. ACE-2 Version 1.0. Linguistic Data Consortium, Philadelphia. Ian Niles and Adam Pease. 2001. Towards a Standard Upper Ontology. In Proceedings of the 2nd International Conference on Formal Ontology in Information Systems. Athina Pappas and Susan A. Gelman. 1998. Generic noun phrases in mother–child conversations. Journal of Child Language, 25(1):19–33. Simone Paolo Ponzetto and Michael Strube. 2007. Deriving a large scale taxonomy from wikipedia. In Proceedings of the 22nd Conference on the Advancement of Artificial Intelligence, pages 1440– 1445, Vancouver, B.C., Canada, July. Willard Van Orman Quine. 1960. Word and Object. MIT Press, Cambridge, Massachusetts. Raymond Reiter. 1980. A logic for default reasoning. Artificial Intelligence, 13:81–132. Helmut Schmid. 1994. Probabilistic part-of-speech tagging using decision trees. Proceedings of the conference on New Methods in Language Processing, 12. Sangweon Suh. 2006. Extracting Generic Statements for the Semantic Web. Master’s thesis, University of Edinburgh. Ian H. Witten and Eibe Frank. 2002. Data mining: practical machine learning tools and techniques with Java implementations. ACM SIGMOD Record, 31(1):76–77. C¨acilia Zirn, Vivi Nastase, and Michael Strube. 2008. Distinguishing between instances and classes in the Wikipedia taxonomy. In Proceedings of the 5th European Semantic Web Conference. 49
|
2010
|
5
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 485–494, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Efficient Staggered Decoding for Sequence Labeling Nobuhiro Kaji Yasuhiro Fujiwara Naoki Yoshinaga Masaru Kitsuregawa Institute of Industrial Science, The University of Tokyo, 4-6-1, Komaba, Meguro-ku, Tokyo, 153-8505 Japan {kaji,fujiwara,ynaga,kisture}@tkl.iis.u-tokyo.ac.jp Abstract The Viterbi algorithm is the conventional decoding algorithm most widely adopted for sequence labeling. Viterbi decoding is, however, prohibitively slow when the label set is large, because its time complexity is quadratic in the number of labels. This paper proposes an exact decoding algorithm that overcomes this problem. A novel property of our algorithm is that it efficiently reduces the labels to be decoded, while still allowing us to check the optimality of the solution. Experiments on three tasks (POS tagging, joint POS tagging and chunking, and supertagging) show that the new algorithm is several orders of magnitude faster than the basic Viterbi and a state-of-the-art algorithm, CARPEDIEM (Esposito and Radicioni, 2009). 1 Introduction In the past decade, sequence labeling algorithms such as HMMs, CRFs, and Collins’ perceptrons have been extensively studied in the field of NLP (Rabiner, 1989; Lafferty et al., 2001; Collins, 2002). Now they are indispensable in a wide range of NLP tasks including chunking, POS tagging, NER and so on (Sha and Pereira, 2003; Tsuruoka and Tsujii, 2005; Lin and Wu, 2009). One important task in sequence labeling is how to find the most probable label sequence from among all possible ones. This task, referred to as decoding, is usually carried out using the Viterbi algorithm (Viterbi, 1967). The Viterbi algorithm has O(NL2) time complexity,1 where N is the input size and L is the number of labels. Although the Viterbi algorithm is generally efficient, 1The first-order Markov assumption is made throughout this paper, although our algorithm is applicable to higherorder Markov models as well. it becomes prohibitively slow when dealing with a large number of labels, since its computational cost is quadratic in L (Dietterich et al., 2008). Unfortunately, several sequence-labeling problems in NLP involve a large number of labels. For example, there are more than 40 and 2000 labels in POS tagging and supertagging, respectively (Brants, 2000; Matsuzaki et al., 2007). These tasks incur much higher computational costs than simpler tasks like NP chunking. What is worse, the number of labels grows drastically if we jointly perform multiple tasks. As we shall see later, we need over 300 labels to reduce joint POS tagging and chunking into the single sequence labeling problem. Although joint learning has attracted much attention in recent years, how to perform decoding efficiently still remains an open problem. In this paper, we present a new decoding algorithm that overcomes this problem. The proposed algorithm has three distinguishing properties: (1) It is much more efficient than the Viterbi algorithm when dealing with a large number of labels. (2) It is an exact algorithm, that is, the optimality of the solution is always guaranteed unlike approximate algorithms. (3) It is automatic, requiring no taskdependent hyperparameters that have to be manually adjusted. Experiments evaluate our algorithm on three tasks: POS tagging, joint POS tagging and chunking, and supertagging2. The results demonstrate that our algorithm is up to several orders of magnitude faster than the basic Viterbi algorithm and a state-of-the-art algorithm (Esposito and Radicioni, 2009); it makes exact decoding practical even in labeling problems with a large label set. 2 Preliminaries We first provide a brief overview of sequence labeling and introduce related work. 2Our implementation is available at http://www.tkl.iis.utokyo.ac.jp/˜kaji/staggered 485 2.1 Models Sequence labeling is the problem of predicting label sequence y = {yn}N n=1 for given token sequence x = {xn}N n=1. This is typically done by defining a score function f(x, y) and locating the best label sequence: ymax = argmax y f(x, y). The form of f(x, y) is dependent on the learning model used. Here, we introduce two models widely used in the literature. Generative models HMM is the most famous generative model for labeling token sequences (Rabiner, 1989). In HMMs, the score function f(x, y) is the joint probability distribution over (x, y). If we assume a one-to-one correspondence between the hidden states and the labels, the score function can be written as: f(x, y) = log p(x, y) = log p(x|y) + log p(y) = N n=1 log p(xn|yn)+ N n=1 log p(yn|yn−1). The parameters log p(xn|yn) and log p(yn|yn−1) are usually estimated using maximum likelihood or the EM algorithm. Since parameter estimation lies outside the scope of this paper, a detailed description is omitted. Discriminative models Recent years have seen the emergence of discriminative training methods for sequence labeling (Lafferty et al., 2001; Tasker et al., 2003; Collins, 2002; Tsochantaridis et al., 2005). Among them, we focus on the perceptron algorithm (Collins, 2002). Although we do not discuss the other discriminative models, our algorithm is equivalently applicable to them. The major difference between those models lies in parameter estimation; the decoding process is virtually the same. In the perceptron, the score function f(x, y) is given as f(x, y) = w · φ(x, y) where w is the weight vector, and φ(x, y) is the feature vector representation of the pair (x, y). By making the first-order Markov assumption, we have f(x, y) = w · φ(x, y) = N n=1 K k=1 wkφk(x, yn−1, yn), where K = |φ(x, y)| is the number of features, φk is the k-th feature function, and wk is the weight corresponding to it. Parameter w can be estimated in the same way as in the conventional perceptron algorithm. See (Collins, 2002) for details. 2.2 Viterbi decoding Given the score function f(x, y), we have to locate the best label sequence. This is usually performed by applying the Viterbi algorithm. Let ω(yn) be the best score of the partial label sequence ending with yn. The idea of the Viterbi algorithm is to use dynamic programming to compute ω(yn). In HMMs, ω(yn) can be can be defined as max yn−1{ω(yn−1) + log p(yn|yn−1)} + log p(xn|yn). Using this recursive definition, we can evaluate ω(yn) for all yn. This results in the identification of the best label sequence. Although the Viterbi algorithm is commonly adopted in past studies, it is not always efficient. The computational cost of the Viterbi algorithm is O(NL2), where N is the input length and L is the number of labels; it is efficient enough if L is small. However, if there are many labels, the Viterbi algorithm becomes prohibitively slow because of its quadratic dependence on L. 2.3 Related work To the best of our knowledge, the Viterbi algorithm is the only algorithm widely adopted in the NLP field that offers exact decoding. In other communities, several exact algorithms have already been proposed for handling large label sets. While they are successful to some extent, they demand strong assumptions that are unusual in NLP. Moreover, none were challenged with standard NLP tasks. Felzenszwalb et al. (2003) presented a fast inference algorithm for HMMs based on the assumption that the hidden states can be embedded in a grid space, and the transition probability corresponds to the distance on that space. This type of probability distribution is not common in NLP tasks. Lifshits et al. (2007) proposed a compression-based approach to speed up HMM decoding. It assumes that the input sequence is highly repetitive. Amongst others, CARPEDIEM (Esposito and Radicioni, 2009) is the algorithm closest to our work. It accelerates decoding by assuming that the adjacent labels are not strongly correlated. This assumption is appropriate for 486 some NLP tasks. For example, as suggested in (Liang et al., 2008), adjacent labels do not provide strong information in POS tagging. However, the applicability of this idea to other NLP tasks is still unclear. Approximate algorithms, such as beam search or island-driven search, have been proposed for speeding up decoding. Tsuruoka and Tsujii (2005) proposed easiest-first deterministic decoding. Siddiqi and Moore (2005) presented the parameter tying approach for fast inference in HMMs. A similar idea was applied to CRFs as well (Cohn, 2006; Jeong et al., 2009). In general, approximate algorithms have the advantage of speed over exact algorithms. However, both types of algorithms are still widely adopted by practitioners, since exact algorithms have merits other than speed. First, the optimality of the solution is always guaranteed. It is hard for most of the approximate algorithms to even bound the error rate. Second, approximate algorithms usually require hyperparameters, which control the tradeoff between accuracy and efficiency (e.g., beam width), and these have to be manually adjusted. On the other hand, most of the exact algorithms, including ours, do not require such a manual effort. Despite these advantages, exact algorithms are rarely used when dealing with a large number of labels. This is because exact algorithms become considerably slower than approximate algorithms in such situations. The paper presents an exact algorithm that avoids this problem; it provides the research community with another option for handling a lot of labels. 3 Algorithm This section presents the new decoding algorithm. The key is to reduce the number of labels examined. Our algorithm locates the best label sequence by iteratively solving labeling problems with a reduced label set. This results in significant time savings in practice, because each iteration becomes much more efficient than solving the original labeling problem. More importantly, our algorithm always obtains the exact solution. This is because the algorithm allows us to check the optimality of the solution achieved by using only the reduced label set. In the following discussions, we restrict our focus to HMMs for presentation clarity. Extension to A A A A B B B B C D C D C D C D D E D E D E D E E F E F E F E F G G G G H H H H (a) A B C D A A B A B C D D D (b) Figure 1: (a) An example of a lattice, where the letters {A, B, C, D, E, F, G, H} represent labels associated with nodes. (b) The degenerate lattice. the perceptron algorithm is presented in Section 4. 3.1 Degenerate lattice We begin by introducing the degenerate lattice, which plays a central role in our algorithm. Consider the lattice in Figure 1(a). Following convention, we regard each path on the lattice as a label sequence. Note that the label set is {A, B, C, D, E, F, G, H}. By aggregating several nodes in the same column of the lattice, we can transform the original lattice into a simpler form, which we call the degenerate lattice (Figure 1(b)). Let us examine the intuition behind the degenerate lattice. Aggregating nodes can be viewed as grouping several labels into a new one. Here, a label is referred to as an active label if it is not aggregated (e.g., A, B, C, and D in the first column of Figure 1(b)), and otherwise as an inactive label (i.e., dotted nodes). The new label, which is made by grouping the inactive labels, is referred to as a degenerate label (i.e., large nodes covering the dotted ones). Two degenerate labels can be seen as equivalent if their corresponding inactive label sets are the same (e.g., degenerate labels in the first and the last column). In this approach, each path of the degenerate lattice can also be interpreted as a label sequence. In this case, however, the label to be assigned is either an active label or a degenerate label. We then define the parameters associated with degenerate label z. For reasons that will become clear later, they are set to the maxima among the parameters of the inactive labels: log p(x|z) = max y′∈I(z) log p(x|y′), (1) log p(z|y) = max y′∈I(z) log p(y′|y), (2) log p(y|z) = max y′∈I(z) log p(y|y′), (3) log p(z|z′) = max y′∈I(z),y′′∈I(z′) log p(y′|y′′), (4) 487 A A A A B B B B C D C D C D C D D E D E D E D E E F E F E F E F G G G G H H H H (a) A B C D A A B A B C D (b) Figure 2: (a) The path y = {A, E, G, C} of the original lattice. (b) The path z of the degenerate lattice that corresponds to y. where y is an active label, z and z′ are degenerate labels, and I(z) denotes one-to-one mapping from z to its corresponding inactive label set. The degenerate lattice has an important property which is the key to our algorithm: Lemma 1. If the best path of the degenerate lattice does not include any degenerate label, it is equivalent to the best path of the original lattice. Proof. Let zmax be the best path of the degenerate lattice. Our goal is to prove that if zmax does not include any degenerate label, then ∀y ∈Y, log p(x, y) ≤log p(x, zmax) (5) where Y is the set of all paths on the original lattice. We prove this by partitioning Y into two disjoint sets: Y0 and Y1, where Y0 is the subset of Y appearing in the degenerate lattice. Notice that zmax ∈Y0. Since zmax is the best path of the degenerate lattice, we have ∀y ∈Y0, log p(x, y) ≤log p(x, zmax). The equation holds when y = zmax. We next examine the label sequence y such that y ∈Y1. For each path y ∈Y1, there exists a unique path z on the degenerate lattice that corresponds to y (Figure 2). Therefore, we have ∀y ∈Y1, ∃z ∈Z, log p(x, y) ≤log p(x, z) < log p(x, zmax) where Z is the set of all paths of the degenerate lattice. The inequality log p(x, y) ≤log p(x, z) can be proved by using Equations (1)-(4). Using these results, we can complete (5). A A A A (a) A A B A B A B B (b) A A B C D A B A B C D B C D C D (c) Figure 3: (a) The best path of the initial degenerate lattice, which is denoted by the line, is located. (b) The active labels are expanded and the best path is searched again. (c) The best path without degenerate labels is obtained. 3.2 Staggered decoding Now we can describe our algorithm, which we call staggered decoding. The algorithm successively constructs degenerate lattices and checks whether the best path includes degenerate labels. In building each degenerate lattice, labels with high probability p(y), estimated from training data, are preferentially selected as the active label; the expectation is that such labels are likely to belong to the best path. The algorithm is detailed as follows: Initialization step The algorithm starts by building a degenerate lattice in which there is only one active label in each column. We select label y with the highest p(y) as the active label. Search step The best path of the degenerate lattice is located (Figure 3(a)). This is done by using the Viterbi algorithm (and pruning technique, as we describe in Section 3.3). If the best path does not include any degenerate label, we can terminate the algorithm since it is identical with the best path of the original lattice according to Lemma 1. Otherwise, we proceed to the next step. Expansion step We double the number of the active labels in the degenerate lattice. The new active labels are selected from the current inactive label set in descending order of p(y). If the inactive label set becomes empty, we simply reconstructed the original lattice. After expanding the active labels, we go back to the previous step (Figure 3(b)). This procedure is repeated until the termination condition in the search step is satisfied, i.e., the best path has no degenerate label (Figure 3(c)). Compared to the Viterbi algorithm, staggered decoding requires two additional computations for 488 training. First, we have to estimate p(y) so as to select active labels in the initialization and expansion step. Second, we have to compute the parameters regarding degenerate labels according to Equations (1)-(4). Both impose trivial computation costs. 3.3 Pruning To achieve speed-up, it is crucial that staggered decoding efficiently performs the search step. For this purpose, we can basically use the Viterbi algorithm. In earlier iterations, the Viterbi algorithm is indeed efficient because the label set to be handled is much smaller than the original one. In later iterations, however, our algorithm drastically increases the number of labels, making Viterbi decoding quite expensive. To handle this problem, we propose a method of pruning the lattice nodes. This technique is motivated by the observation that the degenerate lattice shares many active labels with the previous iteration. In the remainder of Section3.3, we explain the technique by taking the following steps: • Section 3.3.1 examines a lower bound l such that l ≤maxy log p(x, y). • Section 3.3.2 examines the maximum score MAX(yn) in case token xn takes label yn: MAX(yn) = max y′n=yn log p(x, y′). • Section 3.3.3 presents our pruning procedure. The idea is that if MAX(yn) < l, then the node corresponding to yn can be removed from consideration. 3.3.1 Lower bound Lower bound l can be trivially calculated in the search step. This can be done by retaining the best path among those consisting of only active labels. The score of that path is obviously the lower bound. Since the search step is repeated until the termination criteria is met, we can update the lower bound at every search step. As the iteration proceeds, the degenerate lattice becomes closer to the original one, so the lower bound becomes tighter. 3.3.2 Maximum score The maximum score MAX(yn) can be computed from the original lattice. Let ω(yn) be the best score of the partial label sequence ending with yn. Presuming that we traverse the lattice from left to right, ω(yn) can be defined as max yn−1{ω(yn−1) + log p(yn|yn−1)} + log p(xn|yn). If we traverse the lattice from right to left, an analogous score ¯ω(yn) can be defined as log p(xn|yn) + max yn+1{¯ω(yn+1) + log p(yn|yn+1)}. Using these two scores, we have MAX(yn) = ω(yn) + ¯ω(yn) −log p(xn|yn). Notice that updating ω(yn) or ¯ω(yn) is equivalent to the forward or backward Viterbi algorithm, respectively. Although it is expensive to compute ω(yn) and ¯ω(yn), we can efficiently estimate their upper bounds. Let λ(yn) and ¯λ(yn) be scores analogous to ω(yn) and ¯ω(yn) that are computed using the degenerate lattice. We have ω(yn) ≤λ(yn) and ¯ω(yn) ≤¯λ(yn), by following similar discussions as raised in the proof of Lemma 1. Therefore, we can still check whether MAX(yn) is smaller than l by using λ(yn) and ¯λ(yn): MAX(yn) = ω(yn) + ¯ω(yn) −log p(xn|yn) ≤λ(yn) + ¯λ(yn) −log p(xn|yn) < l. For the sake of simplicity, we assume that yn is an active label. Although we do not discuss the other cases, our pruning technique is also applicable to them. We just point out that, if yn is an inactive label, then there exists a degenerate label zn in the n-th column such that yn ∈I(zn), and we can use λ(zn) and ¯λ(zn) instead of λ(yn) and ¯λ(yn). We compute λ(yn) and ¯λ(yn) by using the forward and backward Viterbi algorithm, respectively. In the search step immediately following initialization, we perform the forward Viterbi algorithm to find the best path, that is, λ(yn) is updated for all yn. In the next search step, the backward Viterbi algorithm is carried out, and ¯λ(yn) is updated. In the succeeding search steps, these updates are alternated. As the algorithm progresses, λ(yn) and ¯λ(yn) become closer to ω(yn) and ¯ω(yn). 3.3.3 Pruning procedure We make use of the bounds in pruning the lattice nodes. To do this, we keep the values of l, λ(yn) 489 and ¯λ(yn). They are set as l = −∞and λ(yn) = ¯λ(yn) = ∞in the initialization step, and are updated in the search step. The lower bound l is updated at the end of the search step, while λ(yn) and ¯λ(yn) can be updated during the running of the Viterbi algorithm. When λ(yn) or ¯λ(yn) is changed, we check whether MAX(yn) < l holds and the node is pruned if the condition is met. 3.4 Analysis We provide here a theoretical analysis of staggered decoding. In the following proofs, L, V , and N represent the number of original labels, the number of distinct tokens, and the length of input token sequence, respectively. To simplify the discussion, we assume that log2 L is an integer (e.g., L = 64). We first introduce three lemmas: Lemma 2. Staggered decoding requires at most (log2 L + 1) iterations to terminate. Proof. We have 2m−1 active labels in the m-th search step (m = 1, 2 . . . ), which means we have L active labels and no degenerate labels in the (log2 L + 1)-th search step. Therefore, the algorithm always terminates within (log2 L + 1) iterations. Lemma 3. The number of degenerate labels is log2 L. Proof. Since we create one new degenerate label in all but the last expansion step, we have log2 L degenerate labels. Lemma 4. The Viterbi algorithm requires O(L2+ LV ) memory space and has O(NL2) time complexity. Proof. Since we need O(L2) and O(LV ) space to keep the transition and emission probability matrices, we need O(L2 + LV ) space to perform the Viterbi algorithm. The time complexity of the Viterbi algorithm is O(NL2) since there are NL nodes in the lattice and it takes O(L) time to evaluate the score of each node. The above statements allow us to establish our main results: Theorem 1. Staggered decoding requires O(L2 + LV ) memory space. Proof. Since we have L original labels and log2 L degenerate labels, staggered decoding requires O((L+log2 L)2+(L+log2 L)V ) = O(L2+LV ) A A A A (a) A A B A B A B (b) A A B C D A B A B C D (c) Figure 4: Staggered decoding with column-wise expansion: (a) The best path of the initial degenerate lattice, which does not pass through the degenerate label in the first column. (b) Columnwise expansion is performed and the best path is searched again. Notice that the active label in the first column is not expanded. (c) The final result. memory space to perform Viterbi decoding in the search step. Theorem 2. Staggered decoding has O(N) best case time complexity and O(NL2) worst case time complexity. Proof. To perform the m-th search step, staggered decoding requires the order of O(N4m−1) time because we have 2m−1 active labels. Therefore, it has O(M m=1 N4m−1) time complexity if it terminates after the M-th search step. In the best case, M = 1, the time complexity is O(N). In the worst case, M = log2 L + 1, the time complexity is the order of O(NL2) because log2 L+1 m=1 N4m−1 < 4 3NL2. Theorem 1 shows that staggered decoding asymptotically requires the same order of memory space as the Viterbi algorithm. Theorem 2 reveals that staggered decoding has the same order of time complexity as the Viterbi algorithm even in the worst case. 3.5 Heuristic techniques We present two heuristic techniques for further speeding up our algorithm. First, we can initialize the value of lower bound l by selecting a path from the original lattice in some way, and then computing the score of that path. In our experiments, we use the path located by the left-to-right deterministic decoding (i.e., beam search with a beam width of 1). Although this method requires an additional cost to locate the path, it is very effective in practice. If l is initialized in this manner, the best case time complexity of our algorithm becomes O(NL). 490 The second technique is for the expansion step. Instead of the expansion technique described in Section 3.2, we can expand the active labels in a heuristic manner to keep the number of active labels small: Column-wise expansion step We double the number of the active labels in the column only if the best path of the degenerate lattice passes through the degenerate label of that column (Figure 4). A drawback of this strategy is that the algorithm requires N(log2 L+1) iterations in the worst case. As the result, we can no longer derive a reasonable upper bound for the time complexity. Nevertheless, column-wise expansion is highly effective in practice as we will demonstrate in the experiment. Note that Theorem 1 still holds true even if we use column-wise expansion. 4 Extension to the Perceptron The discussion we have made so far can be applied to perceptrons. This can be clarified by comparing the score functions f(x, y). In HMMs, the score function can be written as N n=1 log(xn|yn) + log(yn|yn−1) . In perceptrons, on the other hand, it is given as N n=1 k w1 kφ1 k(x, yn) + k w2 kφ2 k(x, yn−1, yn) where we explicitly distinguish the unigram feature function φ1 k and bigram feature function φ2 k. Comparing the form of the two functions, we can see that our discussion on HMMs can be extended to perceptrons by substituting k w1 kφ1 k(x, yn) and k w2 kφ2 k(x, yn−1, yn) for log p(xn|yn) and log p(yn|yn−1). However, implementing the perceptron algorithm is not straightforward. The problem is that it is difficult, if not impossible, to compute k w1 kφ1 k(x, y) and k w2 kφ2 k(x, y, y′) offline because they are dependent on the entire token sequence x, unlike log p(x|y) and log p(y|y′). Consequently, we cannot evaluate the maxima analogous to Equations (1)-(4) offline either. For unigram features, we compute the maximum, maxy k w1 kφ1 k(x, y), as a preprocess in the initialization step (cf. Equation (1)). This preprocess requires O(NL) time, which is negligible compared with the cost required by the Viterbi algorithm. Unfortunately, we cannot use the same technique for computing maxy,y′ k w2 kφ2 k(x, y, y′) because a similar computation would take O(NL2) time (cf. Equation (4)). For bigram features, we compute its upper bound offline. For example, the following bound was proposed by Esposito and Radicioni (2009): max y,y′ k w2 kφ2 k(x, y, y′) ≤max y,y′ k w2 kδ(0 < w2 k) where δ(·) is the delta function and the summations are taken over all feature functions associated with both y and y′. Intuitively, the upper bound corresponds to an ideal case in which all features with positive weight are activated.3 It can be computed without any task-specific knowledge. In practice, however, we can compute better bounds based on task-specific knowledge. The simplest case is that the bigram features are independent of the token sequence x. In such a situation, we can trivially compute the exact maxima offline, as we did in the case of HMMs. Fortunately, such a feature set is quite common in NLP problems and we could use this technique in our experiments. Even if bigram features are dependent on x, it is still possible to compute better bounds if several features are mutually exclusive, as discussed in (Esposito and Radicioni, 2009). Finally, it is worth noting that we can use staggered decoding in training perceptrons as well, although such application lies outside the scope of this paper. The algorithm does not support training acceleration for other discriminative models. 5 Experiments and Discussion 5.1 Setting The proposed algorithm was evaluated with three tasks: POS tagging, joint POS tagging and chunking (called joint tagging for short), and supertagging. To reduce joint tagging into a single sequence labeling problem, we produced the labels by concatenating the POS tag and the chunk tag (BIO format), e.g., NN/B-NP. In the two tasks other than supertagging, the input token is the word. In supertagging, the token is the pair of the word and its oracle POS tag. 3We assume binary feature functions. 491 Table 1: Decoding speed (sent./sec). POS tagging Joint tagging Supertagging VITERBI 4000 77 1.1 CARPEDIEM 8600 51 0.26 SD 8800 850 121 SD+C-EXP. 14,000 1600 300 The data sets we used for the three experiments are the Penn TreeBank (PTB) corpus, CoNLL 2000 corpus, and an HPSG treebank built from the PTB corpus (Matsuzaki et al., 2007). We used sections 02-21 of PTB for training, and section 23 for testing. The number of labels in the three tasks is 45, 319 and 2602, respectively. We used the perceptron algorithm for training. The models were averaged over 10 iterations (Collins, 2002). For features, we basically followed previous studies (Tsuruoka and Tsujii, 2005; Sha and Pereira, 2003; Ninomiya et al., 2006). In POS tagging, we used unigrams of the current and its neighboring words, word bigrams, prefixes and suffixes of the current word, capitalization, and tag bigrams. In joint tagging, we also used the same features. In supertagging, we used POS unigrams and bigrams in addition to the same features other than capitalization. As the evaluation measure, we used the average decoding speed (sentences/sec) to two significant digits over five trials. To strictly measure the time spent for decoding, we ignored the preprocessing time, that is, the time for loading the model file and converting the features (i.e., strings) into integers. We note that the accuracy was comparable to the state-of-the-art in the three tasks: 97.08, 93.21, and 91.20% respectively. 5.2 Results and discussions Table 1 presents the performance of our algorithm. SD represents the proposed algorithm without column-wise expansion, while SD+C-EXP. uses column-wise expansion. For comparison, we present the results of two baseline algorithms as well: VITERBI and CARPEDIEM (Esposito and Radicioni, 2009). In almost all settings, we see that both of our algorithms outperformed the other two. We also find that SD+C-EXP. performed consistently better than SD. This indicates the effectiveness of column-wise expansion. Following VITERBI, CARPEDIEM is the most relevant algorithm, for sequence labeling in NLP, as discussed in Section 2.3. However, our results Table 2: The average number of iterations. POS tagging Joint tagging Supertagging SD 6.02 8.15 10.0 SD+C-EXP. 6.12 8.62 10.6 Table 3: Training time. POS tagging Joint tagging Supertagging VITERBI 100 sec. 20 min. 100 hour SD+C-EXP. 37 sec. 1.5 min. 5.3 hour demonstrated that CARPEDIEM worked poorly in two of the three tasks. We consider this is because the transition information is crucial for the two tasks, and the assumption behind CARPEDIEM is violated. In contrast, the proposed algorithms performed reasonably well for all three tasks, demonstrating the wide applicability of our algorithm. Table 2 presents the average iteration numbers of SD and SD+C-EXP. We can observe that the two algorithms required almost the same number of iterations on average, although the iteration number is not tightly bounded if we use column-wise expansion. This indicates that SD+C-EXP. virtually avoided performing extra iterations, while heuristically restricting active label expansion. Table 3 compares the training time spent by VITERBI and SD+C-EXP. Although speeding up perceptron training is a by-product, it is interesting to see that our algorithm is in fact effective at reducing the training time as well. The result also indicates that the speed-up is more significant at test time. This is probably because the model is not predictive enough at the beginning of training, and the pruning is not that effective. 5.3 Comparison with approximate algorithm Table 4 compares two exact algorithms (VITERBI and SD+E-XP.) with beam search, which is the approximate algorithm widely adopted for sequence labeling in NLP. For this experiment, the beam width, B, was exhaustively calibrated: we tried B = {1, 2, 4, 8, ...} until the beam search achieved comparable accuracy to the exact algorithms, i.e., the difference fell below 0.1 in our case. We see that there is a substantial difference in the performance between VITERBI and BEAM. On the other hand, SD+C-EXP. reached speeds very close to those of BEAM. In fact, they achieved comparable performance in our experiment. These results demonstrate that we could successfully bridge the gap in the performance be492 Table 4: Comparison with beam search (sent./sec). POS tagging Joint tagging Supertagging VITERBI 4000 77 1.1 SD+C-EXP. 14,000 1600 300 BEAM 18,000 2400 180 tween exact and approximate algorithms, while retaining the advantages of exact algorithms. 6 Relation to coarse-to-fine approach Before concluding remarks, we briefly examine the relationship between staggered decoding and coarse-to-fine PCFG parsing (2006). In coarse-tofine parsing, the candidate parse trees are pruned by using the parse forest produced by a coarsegrained PCFG. Since the degenerate label can be interpreted as a coarse-level label, one may consider that staggered decoding is an instance of coarse-to-fine approach. While there is some resemblance, there are at least two essential differences. First, coarse-to-fine approach is a heuristic pruning, that is, it is not an exact algorithm. Second, our algorithm does not always perform decoding at the fine-grained level. It is designed to be able to stop decoding at the coarse-level. 7 Conclusions The sequence labeling algorithm is indispensable to modern statistical NLP. However, the Viterbi algorithm, which is the standard decoding algorithm in NLP, is not efficient when we have to deal with a large number of labels. In this paper we presented staggered decoding, which provides a principled way of resolving this problem. We consider that it is a real alternative to the Viterbi algorithm in various NLP tasks. An interesting future direction is to extend the proposed technique to handle more complex structures than the Markov chains, including semiMarkov models and factorial HMMs (Sarawagi and Cohen, 2004; Sutton et al., 2004). We hope this work opens a new perspective on decoding algorithms for a wide range of NLP problems, not just sequence labeling. Acknowledgement We wish to thank the anonymous reviewers for their helpful comments, especially on the computational complexity of our algorithm. We also thank Yusuke Miyao for providing us with the HPSG Treebank data. References Thorsten Brants. 2000. TnT - a statistical part-ofspeech tagger. In Proceedings of ANLP, pages 224– 231. Eugene Charniak, Mark Johnson, Micha Elsner, Joseph Austerweil, David Ellis, Isaac Haxton, Catherine Hill, R. Shrivaths, Jeremy Moore, Michael Pozar, and Theresa Vu. 2006. Multi-level coarse-to-fine PCFG parsing. In Proceedings of NAACL, pages 168–175. Trevor Cohn. 2006. Efficient inference in large conditional random fields. In Proceedings of ECML, pages 606–613. Michael Collins. 2002. Discriminative training methods for hidden Markov models: Theory and experiments with perceptron algorithms. In Proceedings of EMNLP, pages 1–8. Thomas G. Dietterich, Pedro Domingos, Lise Getoor, Stephen Muggleton, and Prasad Tadepalli. 2008. Structured machine learning: the next ten years. Machine Learning, 73(1):3–23. Roberto Esposito and Daniele P. Radicioni. 2009. CARPEDIEM: Optimizing the Viterbi algorithm and applications to supervised sequential learning. Jorunal of Machine Learning Research, 10:1851– 1880. Pedro F. Felzenszwalb, Daniel P. Huttenlocher, and Jon M. Kleinberg. 2003. Fast algorithms for largestate-space HMMs with applications to Web usage analysis. In Proceedings of NIPS, pages 409–416. Minwoo Jeong, Chin-Yew Lin, and Gary Geunbae Lee. 2009. Efficient inference of CRFs for large-scale natural language data. In Proceedings of ACLIJCNLP Short Papers, pages 281–284. John Lafferty, Andrew McCallum, and Fernand Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of ICML, pages 282– 289. Percy Liang, Hal Daum´e III, and Dan Klein. 2008. Structure compilation: Trading structure for features. In Proceedings of ICML, pages 592–599. Yury Lifshits, Shay Mozes, Oren Weimann, and Michal Ziv-Ukelson. 2007. Speeding up HMM decoding and training by exploiting sequence repetitions. Computational Pattern Matching, pages 4–15. Dekang Lin and Xiaoyun Wu. 2009. Phrae clustering for discriminative training. In Proceedings of ACLIJCNLP, pages 1030–1038. 493 Takuya Matsuzaki, Yusuke Miyao, and Jun’ichi Tsujii. 2007. Efficient HPSG parsing with supertagging and CFG-filtering. In Proceedings of IJCAI, pages 1671–1676. Takashi Ninomiya, Takuya Matsuzaki, Yoshimasa Tsuruoka, Yusuke Miyao, and Jun’ichi Tsujii. 2006. Extremely lexicalized models for accurate and fast HPSG parsing. In Proceedings of EMNLP, pages 155–163. Lawrence R. Rabiner. 1989. A tutorial on hidden Markov models and selected applications in speech recognition. In Proceedings of The IEEE, pages 257–286. Sunita Sarawagi and Willian W. Cohen. 2004. SemiMarkov conditional random fields for information extraction. In Proceedings of NIPS, pages 1185– 1192. Fei Sha and Fernando Pereira. 2003. Shallow parsing with conditional random fields. In Proceedings of HLT-NAACL, pages 134–141. Sajid M. Siddiqi and Andrew W. Moore. 2005. Fast inference and learning in large-state-space HMMs. In Proceedings of ICML, pages 800–807. Charles Sutton, Khashayar Rohanimanesh, and Andrew McCallum. 2004. Dynamic conditional random fields: Factorized probabilistic models for labeling and segmenting sequence data. In Proceedings of ICML. Ben Tasker, Carlos Guestrin, and Daphe Koller. 2003. Max-margin Markov networks. In Proceedings of NIPS, pages 25–32. Ioannis Tsochantaridis, Thorsten Joachims, Thomas Hofmann, and Yasemin Altun. 2005. Large margin methods for structured and interdependent output variables. Journal of Machine Learning Research, 6:1453–1484. Yoshimasa Tsuruoka and Jun’ichi Tsujii. 2005. Bidirectional inference with the easiest-first strategy for tagging sequence data. In Proceedings of HLT/EMNLP, pages 467–474. Andrew J. Viterbi. 1967. Error bounds for convolutional codes and an asymeptotically optimum decoding algorithm. IEEE Transactios on Information Theory, 13(2):260–267. 494
|
2010
|
50
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 495–503, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Minimized models and grammar-informed initialization for supertagging with highly ambiguous lexicons Sujith Ravi1 Jason Baldridge2 Kevin Knight1 1University of Southern California Information Sciences Institute Marina del Rey, California 90292 {sravi,knight}@isi.edu 2Department of Linguistics The University of Texas at Austin Austin, Texas 78712 [email protected] Abstract We combine two complementary ideas for learning supertaggers from highly ambiguous lexicons: grammar-informed tag transitions and models minimized via integer programming. Each strategy on its own greatly improves performance over basic expectation-maximization training with a bitag Hidden Markov Model, which we show on the CCGbank and CCG-TUT corpora. The strategies provide further error reductions when combined. We describe a new two-stage integer programming strategy that efficiently deals with the high degree of ambiguity on these datasets while obtaining the full effect of model minimization. 1 Introduction Creating accurate part-of-speech (POS) taggers using a tag dictionary and unlabeled data is an interesting task with practical applications. It has been explored at length in the literature since Merialdo (1994), though the task setting as usually defined in such experiments is somewhat artificial since the tag dictionaries are derived from tagged corpora. Nonetheless, the methods proposed apply to realistic scenarios in which one has an electronic part-of-speech tag dictionary or a hand-crafted grammar with limited coverage. Most work has focused on POS-tagging for English using the Penn Treebank (Marcus et al., 1993), such as (Banko and Moore, 2004; Goldwater and Griffiths, 2007; Toutanova and Johnson, 2008; Goldberg et al., 2008; Ravi and Knight, 2009). This generally involves working with the standard set of 45 POS-tags employed in the Penn Treebank. The most ambiguous word has 7 different POS tags associated with it. Most methods have employed some variant of Expectation Maximization (EM) to learn parameters for a bigram or trigram Hidden Markov Model (HMM). Ravi and Knight (2009) achieved the best results thus far (92.3% word token accuracy) via a Minimum Description Length approach using an integer program (IP) that finds a minimal bigram grammar that obeys the tag dictionary constraints and covers the observed data. A more challenging task is learning supertaggers for lexicalized grammar formalisms such as Combinatory Categorial Grammar (CCG) (Steedman, 2000). For example, CCGbank (Hockenmaier and Steedman, 2007) contains 1241 distinct supertags (lexical categories) and the most ambiguous word has 126 supertags. This provides a much more challenging starting point for the semi-supervised methods typically applied to the task. Yet, this is an important task since creating grammars and resources for CCG parsers for new domains and languages is highly labor- and knowledge-intensive. Baldridge (2008) uses grammar-informed initialization for HMM tag transitions based on the universal combinatory rules of the CCG formalism to obtain 56.1% accuracy on ambiguous word tokens, a large improvement over the 33.0% accuracy obtained with uniform initialization for tag transitions. The strategies employed in Ravi and Knight (2009) and Baldridge (2008) are complementary. The former reduces the model size globally given a data set, while the latter biases bitag transitions toward those which are more likely based on a universal grammar without reference to any data. In this paper, we show how these strategies may be combined straightforwardly to produce improvements on the task of learning supertaggers from lexicons that have not been filtered in any way.1 We demonstrate their cross-lingual effectiveness on CCGbank (English) and the Italian CCG-TUT 1See Banko and Moore (2004) for a description of how many early POS-tagging papers in fact used a number of heuristic cutoffs that greatly simplify the problem. 495 corpus (Bos et al., 2009). We find a consistent improved performance by using each of the methods compared to basic EM, and further improvements by using them in combination. Applying the approach of Ravi and Knight (2009) naively to CCG supertagging is intractable due to the high level of ambiguity. We deal with this by defining a new two-stage integer programming formulation that identifies minimal grammars efficiently and effectively. 2 Data CCGbank. CCGbank was created by semiautomatically converting the Penn Treebank to CCG derivations (Hockenmaier and Steedman, 2007). We use the standard splits of the data used in semi-supervised tagging experiments (e.g. Banko and Moore (2004)): sections 0-18 for training, 19-21 for development, and 22-24 for test. CCG-TUT. CCG-TUT was created by semiautomatically converting dependencies in the Italian Turin University Treebank to CCG derivations (Bos et al., 2009). It is much smaller than CCGbank, with only 1837 sentences. It is split into three sections: newspaper texts (NPAPER), civil code texts (CIVIL), and European law texts from the JRC-Acquis Multilingual Parallel Corpus (JRC). For test sets, we use the first 400 sentences of NPAPER, the first 400 of CIVIL, and all of JRC. This leaves 409 and 498 sentences from NPAPER and CIVIL, respectively, for training (to acquire a lexicon and run EM). For evaluation, we use two different settings of train/test splits: TEST 1 Evaluate on the NPAPER section of test using a lexicon extracted only from NPAPER section of train. TEST 2 Evaluate on the entire test using lexicons extracted from (a) NPAPER + CIVIL, (b) NPAPER, and (c) CIVIL. Table 1 shows statistics for supertag ambiguity in CCGbank and CCG-TUT. As a comparison, the POS word token ambiguity in CCGbank is 2.2: the corresponding value of 18.71 for supertags is indicative of the (challenging) fact that supertag ambiguity is greatest for the most frequent words. 3 Grammar informed initialization for supertagging Part-of-speech tags are atomic labels that in and of themselves encode no internal structure. In conData Distinct Max Type ambig Tok ambig CCGbank 1241 126 1.69 18.71 CCG-TUT NPAPER+CIVIL 849 64 1.48 11.76 NPAPER 644 48 1.42 12.17 CIVIL 486 39 1.52 11.33 Table 1: Statistics for the training data used to extract lexicons for CCGbank and CCG-TUT. Distinct: # of distinct lexical categories; Max: # of categories for the most ambiguous word; Type ambig: per word type category ambiguity; Tok ambig: per word token category ambiguity. trast, supertags are detailed, structured labels; a universal set of grammatical rules defines how categories may combine with one another to project syntactic structure.2 Because of this, properties of the CCG formalism itself can be used to constrain learning—prior to considering any particular language, grammar or data set. Baldridge (2008) uses this observation to create grammar-informed tag transitions for a bitag HMM supertagger based on two main properties. First, categories differ in their complexity and less complex categories tend to be used more frequently. For example, two categories for buy in CCGbank are (S[dcl]\NP)/NP and ((((S[b]\NP)/PP)/PP)/(S[adj]\NP))/NP; the former occurs 33 times, the latter once. Second, categories indicate the form of categories found adjacent to them; for example, the category for sentential complement verbs ((S\NP)/S) expects an NP to its left and an S to its right. Categories combine via rules such as application and composition (see Steedman (2000) for details). Given a lexicon containing the categories for each word, these allow derivations like: Ed might see a cat NP (S\NP)/(S\NP) (S\NP)/NP NP/N N >B > (S\NP)/NP NP > S\NP > S Other derivations are possible. In fact, every pair of adjacent words above may be combined directly. For example, see and a may combine through forward composition to produce the category (S\NP)/N, and Ed’s category may type-raise to S/(S\NP) and compose with might’s category. Baldridge uses these properties to define tag 2Note that supertags can be lexical categories of CCG (Steedman, 2000), elementary trees of Tree-adjoining Grammar (Joshi, 1988), or types in a feature hierarchy as in Headdriven Phrase Structure Grammar (Pollard and Sag, 1994). 496 transition distributions that have higher likelihood for simpler categories that are able to combine. For example, for the distribution p(ti|ti−1=NP), (S\NP)\NP is more likely than ((S\NP)/(N/N))\NP because both categories may combine with a preceding NP but the former is simpler. In turn, the latter is more likely than NP: it is more complex but can combine with the preceding NP. Finally, NP is more likely than (S/NP)/NP since neither can combine, but NP is simpler. By starting EM with these tag transition distributions and an unfiltered lexicon (word-tosupertag dictionary), Baldridge obtains a tagging accuracy of 56.1% on ambiguous words—a large improvement over the accuracy of 33.0% obtained by starting with uniform transition distributions. We refer to a model learned from basic EM (uniformly initialized) as EM, and to a model with grammar-informed initialization as EMGI. 4 Minimized models for supertagging The idea of searching for minimized models is related to classic Minimum Description Length (MDL) (Barron et al., 1998), which seeks to select a small model that captures the most regularity in the observed data. This modeling strategy has been shown to produce good results for many natural language tasks (Goldsmith, 2001; Creutz and Lagus, 2002; Ravi and Knight, 2009). For tagging, the idea has been implemented using Bayesian models with priors that indirectly induce sparsity in the learned models (Goldwater and Griffiths, 2007); however, Ravi and Knight (2009) show a better approach is to directly minimize the model using an integer programming (IP) formulation. Here, we build on this idea for supertagging. There are many challenges involved in using IP minimization for supertagging. The 1241 distinct supertags in the tagset result in 1.5 million tag bigram entries in the model and the dictionary contains almost 3.5 million word/tag pairs that are relevant to the test data. The set of 45 POS tags for the same data yields 2025 tag bigrams and 8910 dictionary entries. We also wish to scale our methods to larger data settings than the 24k word tokens in the test data used in the POS tagging task. Our objective is to find the smallest supertag grammar (of tag bigram types) that explains the entire text while obeying the lexicon’s constraints. However, the original IP method of Ravi and Knight (2009) is intractable for supertagging, so we propose a new two-stage method that scales to the larger tagsets and data involved. 4.1 IP method for supertagging Our goal for supertagging is to build a minimized model with the following objective: IPoriginal: Find the smallest supertag grammar (i.e., tag bigrams) that can explain the entire text (the test word token sequence). Using the full grammar and lexicon to perform model minimization results in a very large, difficult to solve integer program involving billions of variables and constraints. This renders the minimization objective IPoriginal intractable. One way of combating this is to use a reduced grammar and lexicon as input to the integer program. We do this without further supervision by using the HMM model trained using basic EM: entries are pruned based on the tag sequence it predicts on the test data. This produces an observed grammar of distinct tag bigrams (Gobs) and lexicon of observed lexical assignments (Lobs). For CCGbank, Gobs and Lobs have 12,363 and 18,869 entries, respectively—far less than the millions of entries in the full grammar and lexicon. Even though EM minimizes the model somewhat, many bad entries remain in the grammar. We prune further by supplying Gobs and Lobs as input (G, L) to the IP-minimization procedure. However, even with the EM-reduced grammar and lexicon, the IP-minimization is still very hard to solve. We thus split it into two stages. The first stage (Minimization 1) finds the smallest grammar Gmin1 ⊂G that explains the set of word bigram types observed in the data rather than the word sequence itself, and the second (Minimization 2) finds the smallest augmentation of Gmin1 that explains the full word sequence. Minimization 1 (MIN1). We begin with a simpler minimization problem than the original one (IPoriginal), with the following objective: IPmin 1: Find the smallest set of tag bigrams Gmin1 ⊂G, such that there is at least one tagging assignment possible for every word bigram type observed in the data. We formulate this as an integer program, creating binary variables gvari for every tag bigram gi = tjtk in G. Binary link variables connect tag bigrams with word bigrams; these are restricted 497 : : ti tj : : Input Grammar (G) word bigrams: w1 w2 w2 w3 : : wi wj : : MIN 1 : : ti tj : : Input Grammar (G) word bigrams: w1 w2 w2 w3 : : wi wj : : word sequence: w1 w2 w3 w4 w5 t1 t2 t3 : : tk supertags tag bigrams chosen in first minimization step (Gmin1) (does not explain the word sequence) word sequence: w1 w2 w3 w4 w5 t1 t2 t3 : : tk supertags tag bigrams chosen in second minimization step (Gmin2) MIN 2 IP Minimization 1 IP Minimization 2 Figure 1: Two-stage IP method for selecting minimized models for supertagging. to the set of links that respect the lexicon L provided as input, i.e., there exists a link variable linkjklm connecting tag bigram tjtk with word bigram wlwm only if the word/tag pairs (wl, tj) and (wm, tk) are present in L. The entire integer programming formulation is shown Figure 2. The IP solver3 solves the above integer program and we extract the set of tag bigrams Gmin1 based on the activated grammar variables. For the CCGbank test data, MIN1 yields 2530 tag bigrams. However, a second stage is needed since there is no guarantee that Gmin1 can explain the test data: it contains tags for all word bigram types, but it cannot necessarily tag the full word sequence. Figure 1 illustrates this. Using only tag bigrams from MIN1 (shown in blue), there is no fully-linked tag path through the network. There are missing links between words w2 and w3 and between words w3 and w4 in the word sequence. The next stage fills in these missing links. Minimization 2 (MIN2). This stage uses the original minimization formulation for the supertagging problem IPoriginal, again using an integer programming method similar to that proposed by Ravi and Knight (2009). If applied to the observed grammar Gobs, the resulting integer program is hard to solve.4 However, by using the partial solution Gmin1 obtained in MIN1 the IP optimization speeds up considerably. We implement this by fixing the values of all binary grammar variables present in Gmin1 to 1 before optimization. This reduces the search space signifi3We use the commercial CPLEX solver. 4The solver runs for days without returning a solution. Minimize: P ∀gi∈G gvari Subject to constraints: 1. For every word bigram wlwm, there exists at least one tagging that respects the lexicon L. P ∀tj∈L(wl), tk∈L(wm) linkjklm ≥1 where L(wl) and L(wm) represent the set of tags seen in the lexicon for words wl and wm respectively. 2. The link variable assignments are constrained to respect the grammar variables chosen by the integer program. linkjklm ≤gvari where gvari is the binary variable corresponding to tag bigram tjtk in the grammar G. Figure 2: IP formulation for Minimization 1. cantly, and CPLEX finishes in just a few hours. The details of this method are described below. We instantiate binary variables gvari and lvari for every tag bigram (in G) and lexicon entry (in L). We then create a network of possible taggings for the word token sequence w1w2....wn in the corpus and assign a binary variable to each link in the network. We name these variables linkcjk, where c indicates the column of the link’s source in the network, and j and k represent the link’s source and destination (i.e., linkcjk corresponds to tag bigram tjtk in column c). Next, we formulate the integer program given in Figure 3. Figure 1 illustrates how MIN2 augments the grammar Gmin1 (links shown in blue) with addi498 Minimize: P ∀gi∈G gvari Subject to constraints: 1. Chosen link variables form a left-to-right path through the tagging network. ∀c=1..n−2∀k P j linkcjk = P j link(c+1)kj 2. Link variable assignments should respect the chosen grammar variables. for every link: linkcjk ≤gvari where gvari corresponds to tag bigram tjtk 3. Link variable assignments should respect the chosen lexicon variables. for every link: linkcjk ≤lvarwctj for every link: linkcjk ≤lvarwc+1tk where wc is the cth word in the word sequence w1...wn, and lvarwctj is the binary variable corresponding to the word/tag pair wc/tj in the lexicon L. 4. The final solution should produce at least one complete tagging path through the network. P ∀j,k link1jk ≥1 5. Provide minimized grammar from MIN1as partial solution to the integer program. ∀gi∈Gmin1 gvari = 1 Figure 3: IP formulation for Minimization 2. tional tag bigrams (shown in red) to form a complete tag path through the network. The minimized grammar set in the final solution Gmin2 contains only 2810 entries, significantly fewer than the original grammar Gobs’s 12,363 tag bigrams. We note that the two-stage minimization procedure proposed here is not guaranteed to yield the optimal solution to our original objective IPoriginal. On the simpler task of unsupervised POS tagging with a dictionary, we compared our method versus directly solving IPoriginal and found that the minimization (in terms of grammar size) achieved by our method is close to the optimal solution for the original objective and yields the same tagging accuracy far more efficiently. Fitting the minimized model. The IPminimization procedure gives us a minimal grammar, but does not fit the model to the data. In order to estimate probabilities for the HMM model for supertagging, we use the EM algorithm but with certain restrictions. We build the transition model using only entries from the minimized grammar set Gmin2, and instantiate an emission model using the word/tag pairs seen in L (provided as input to the minimization procedure). All the parameters in the HMM model are initialized with uniform probabilities, and we run EM for 40 iterations. The trained model is used to find the Viterbi tag sequence for the corpus. We refer to this model (where the EM output (Gobs, Lobs) was provided to the IP-minimization as initial input) as EM+IP. Bootstrapped minimization. The quality of the observed grammar and lexicon improves considerably at the end of a single EM+IP run. Ravi and Knight (2009) exploited this to iteratively improve their POS tag model: since the first minimization procedure is seeded with a noisy grammar and tag dictionary, iterating the IP procedure with progressively better grammars further improves the model. We do likewise, bootstrapping a new EM+IP run using as input, the observed grammar Gobs and lexicon Lobs from the last tagging output of the previous iteration. We run this until the chosen grammar set Gmin2 does not change.5 4.2 Minimization with grammar-informed initialization There are two complementary ways to use grammar-informed initialization with the IPminimization approach: (1) using EMGI output as the starting grammar/lexicon and (2) using the tag transitions directly in the IP objective function. The first takes advantage of the earlier observation that the quality of the grammar and lexicon provided as initial input to the minimization procedure can affect the quality of the final supertagging output. For the second, we modify the objective function used in the two IP-minimization steps to be: Minimize: X ∀gi∈G wi · gvari (1) where, G is the set of tag bigrams provided as input to IP, gvari is a binary variable in the integer program corresponding to tag bigram (ti−1, ti) ∈ G, and wi is negative logarithm of pgii(ti|ti−1) as given by Baldridge (2008).6 All other parts of 5In our experiments, we run three bootstrap iterations. 6Other numeric weights associated with the tag bigrams could be considered, such as 0/1 for uncombin499 the integer program including the constraints remain unchanged, and, we acquire a final tagger in the same manner as described in the previous section. In this way, we combine the minimization and GI strategies into a single objective function that finds a minimal grammar set while keeping the more likely tag bigrams in the chosen solution. EMGI+IPGI is used to refer to the method that uses GI information in both ways: EMGI output as the starting grammar/lexicon and GI weights in the IP-minimization objective. 5 Experiments We compare the four strategies described in Sections 3 and 4, summarized below: EM HMM uniformly initialized, EM training. EM+IP IP minimization using initial grammar provided by EM. EMGI HMM with grammar-informed initialization, EM training. EMGI+IPGI IP minimization using initial grammar/lexicon provided by EMGI and additional grammar-informed IP objective. For EM+IP and EMGI+IPGI, the minimization and EM training processes are iterated until the resulting grammar and lexicon remain unchanged. Forty EM iterations are used for all cases. We also include a baseline which randomly chooses a tag from those associated with each word in the lexicon, averaged over three runs. Accuracy on ambiguous word tokens. We evaluate the performance in terms of tagging accuracy with respect to gold tags for ambiguous words in held-out test sets for English and Italian. We consider results with and without punctuation.7 Recall that unlike much previous work, we do not collect the lexicon (tag dictionary) from the test set: this means the model must handle unknown words and the possibility of having missing lexical entries for covering the test set. Precision and recall of grammar and lexicon. In addition to accuracy, we measure precision and able/combinable bigrams. 7The reason for this is that the “categories” for punctuation in CCGbank are for the most part not actual categories; for example, the period “.” has the categories “.” and “S”. As such, these supertags are outside of the categorial system: their use in derivations requires phrase structure rules that are not derivable from the CCG combinatory rules. Model ambig ambig all all -punc -punc Random 17.9 16.2 27.4 21.9 EM 38.7 35.6 45.6 39.8 EM+IP 52.1 51.0 57.3 53.9 EMGI 56.3 59.4 61.0 61.7 EMGI+IPGI 59.6 62.3 63.8 64.3 Table 2: Supertagging accuracy for CCGbank sections 22-24. Accuracies are reported for four settings—(1) ambiguous word tokens in the test corpus, (2) ambiguous word tokens, ignoring punctuation, (3) all word tokens, and (4) all word tokens except punctuation. recall for each model on the observed bitag grammar and observed lexicon on the test set. We calculate them as follows, for an observed grammar or lexicon X: Precision = |{X} ∩{Observedgold}| |{X}| Recall = |{X} ∩{Observedgold}| |{Observedgold}| This provides a measure of model performance on bitag types for the grammar and lexical entry types for the lexicon, rather than tokens. 5.1 English CCGbank results Accuracy on ambiguous tokens. Table 2 gives performance on the CCGbank test sections. All models are well above the random baseline, and both of the strategies individually boost performance over basic EM by a large margin. For the models using GI, accuracy ignoring punctuation is higher than for all almost entirely due to the fact that “.” has the supertags “.” and S, and the GI gives a preference to S since it can in fact combine with other categories, unlike “.”—the effect is that nearly every sentence-final period (˜5.5k tokens) is tagged S rather than “.”. EMGI is more effective than EM+IP; however, it should be kept in mind that IP-minimization is a general technique that can be applied to any sequence prediction task, whereas grammarinformed initialization may be used only with tasks in which the interactions of adjacent labels may be derived from the labels themselves. Interestingly, the gap between the two approaches is greater when punctuation is ignored (51.0 vs. 59.4)—this is unsurprising because, as noted already, punctuation supertags are not actual cate500 EM EM+IP EMGI EMGI+IPGI Grammar Precision 7.5 32.9 52.6 68.1 Recall 26.9 13.2 34.0 19.8 Lexicon Precision 58.4 63.0 78.0 80.6 Recall 50.9 56.0 71.5 67.6 Table 3: Comparison of grammar/lexicon observed in the model tagging vs. gold tagging in terms of precision and recall measures for supertagging on CCGbank data. gories, so EMGI is unable to model their distribution. Most importantly, the complementary effects of the two approaches can be seen in the improved results for EMGI+IPGI, which obtains about 3% better accuracy than EMGI. Accuracy on all tokens. Table 2 also gives performance when taking all tokens into account. The HMM when using full supervision obtains 87.6% accuracy (Baldridge, 2008),8 so the accuracy of 63.8% achieved by EMGI+IPGI nearly halves the gap between the supervised model and the 45.6% obtained by basic EM semi-supervised model. Effect of GI information in EM and/or IPminimization stages. We can also consider the effect of GI information in either EM training or IP-minimization to see whether it can be effectively exploited in both. The latter, EM+IPGI, obtains 53.2/51.1 for all/no-punc—a small gain compared to EM+IP’s 52.1/51.0. The former, EMGI+IP, obtains 58.9/61.6—a much larger gain. Thus, the better starting point provided by EMGI has more impact than the integer program that includes GI in its objective function. However, we note that it should be possible to exploit the GI information more effectively in the integer program than we have here. Also, our best model, EMGI+IPGI, uses GI information in both stages to obtain our best accuracy of 59.6/62.3. P/R for grammars and lexicons. We can obtain a more-fine grained understanding of how the models differ by considering the precision and recall values for the grammars and lexicons of the different models, given in Table 3. The basic EM model has very low precision for the grammar, indicating it proposes many unnecessary bitags; it 8A state-of-the-art, fully-supervised maximum entropy tagger (Clark and Curran, 2007) (which also uses part-ofspeech labels) obtains 91.4% on the same train/test split. achieves better recall because of the sheer number of bitags it proposes (12,363). EM+IP prunes that set of bitags considerably, leading to better precision at the cost of recall. EMGI’s higher recall and precision indicate the tag transition distributions do capture general patterns of linkage between adjacent CCG categories, while EM ensures that the data filters out combinable, but unnecessary, bitags. With EMGI+IPGI, we again see that IP-minimization prunes even more entries, improving precision at the loss of some recall. Similar trends are seen for precision and recall on the lexicon. IP-minimization’s pruning of inappropriate taggings means more common words are not assigned highly infrequent supertags (boosting precision) while unknown words are generally assigned more sensible supertags (boosting recall). EMGI again focuses taggings on combinable contexts, boosting precision and recall similarly to EM+IP, but in greater measure. EMGI+IPGI then prunes some of the spurious entries, boosting precision at some loss of recall. Tag frequencies predicted on the test set. Table 4 compares gold tags to tags generated by all four methods for the frequent and highly ambiguous words the and in. Basic EM wanders far away from the gold assignments; it has little guidance in the very large search space available to it. IP-minimization identifies a smaller set of tags that better matches the gold tags; this emerges because other determiners and prepositions evoke similar, but not identical, supertags, and the grammar minimization pushes (but does not force) them to rely on the same supertags wherever possible. However, the proportions are incorrect; for example, the tag assigned most frequently to in is ((S\NP)\(S\NP))/NP though (NP\NP)/NP is more frequent in the test set. EMGI’s tags correct that balance and find better proportions, but also some less common categories, such as (((N/N)\(N/N))\((N/N)\(N/N)))/N, sneak in because they combine with frequent categories like N/N and N. Bringing the two strategies together with EMGI+IPGI filters out the unwanted categories while getting better overall proportions. 5.2 Italian CCG-TUT results To demonstrate that both methods and their combination are language independent, we apply them to the Italian CCG-TUT corpus. We wanted to evaluate performance out-of-the-box because 501 Lexicon Gold EM EM+IP EMGI EMGI+IPGI the →(41 distinct tags in Ltrain) (14 tags) (18 tags) (9 tags) (25 tags) (12 tags) NP[nb]/N 5742 0 4544 4176 4666 ((S\NP)\(S\NP))/N 14 5 642 122 107 (((N/N)\(N/N))\((N/N)\(N/N)))/N 0 0 0 698 0 ((S/S)/S[dcl])/(S[adj]\NP) 0 733 0 0 0 PP/N 0 1755 0 3 1 : : : : : : in →(76 distinct tags in Ltrain) (35 tags) (20 tags) (17 tags) (37 tags) (14 tags) (NP\NP)/NP 883 0 649 708 904 ((S\NP)\(S\NP))/NP 793 0 911 320 424 PP/NP 177 1 33 12 82 ((S[adj]\NP)/(S[adj]\NP))/NP 0 215 0 0 0 : : : : : : Table 4: Comparison of tag assignments from the gold tags versus model tags obtained on the test set. The table shows tag assignments (and their counts for each method) for the and in in the CCGbank test sections. The number of distinct tags assigned by each method is given in parentheses. Ltrain is the lexicon obtained from sections 0-18 of CCGbank that is used as the basis for EM training. Model TEST 1 TEST 2 (using lexicon from:) NPAPER+CIVIL NPAPER CIVIL Random 9.6 9.7 8.4 9.6 EM 26.4 26.8 27.2 29.3 EM+IP 34.8 32.4 34.8 34.6 EMGI 43.1 43.9 44.0 40.3 EMGI+IPGI 45.8 43.6 47.5 40.9 Table 5: Comparison of supertagging results for CCG-TUT. Accuracies are for ambiguous word tokens in the test corpus, ignoring punctuation. bootstrapping a supertagger for a new language is one of the main use scenarios we envision: in such a scenario, there is no development data for changing settings and parameters. Thus, we determined a train/test split beforehand and ran the methods exactly as we had for CCGbank. The results, given in Table 5, demonstrate the same trends as for English: basic EM is far more accurate than random, EM+IP adds another 8-10% absolute accuracy, and EMGI adds an additional 810% again. The combination of the methods generally improves over EMGI, except when the lexicon is extracted from NPAPER+CIVIL. Table 6 gives precision and recall for the grammars and lexicons for CCG-TUT—the values are lower than for CCGbank (in line with the lower baseline), but exhibit the same trends. 6 Conclusion We have shown how two complementary strategies—grammar-informed tag transitions and IP-minimization—for learning of supertaggers from highly ambiguous lexicons can be straightEM EM+IP EMGI EMGI+IPGI Grammar Precision 23.1 26.4 44.9 46.7 Recall 18.4 15.9 24.9 22.7 Lexicon Precision 51.2 52.0 54.8 55.1 Recall 43.6 42.8 46.0 44.9 Table 6: Comparison of grammar/lexicon observed in the model tagging vs. gold tagging in terms of precision and recall measures for supertagging on CCG-TUT. forwardly integrated. We verify the benefits of both cross-lingually, on English and Italian data. We also provide a new two-stage integer programming setup that allows model minimization to be tractable for supertagging without sacrificing the quality of the search for minimal bitag grammars. The experiments in this paper use large lexicons, but the methodology will be particularly useful in the context of bootstrapping from smaller ones. This brings further challenges; in particular, it will be necessary to identify novel entries consisting of seen word and seen category and to predict unseen, but valid, categories which are needed to explain the data. For this, it will be necessary to forgo the assumption that the provided lexicon is always obeyed. The methods we introduce here should help maintain good accuracy while opening up these degrees of freedom. Because the lexicon is the grammar in CCG, learning new wordcategory associations is grammar generalization and is of interest for grammar acquisition. 502 Finally, such lexicon refinement and generalization is directly relevant for using CCG in syntaxbased machine translation models (Hassan et al., 2009). Such models are currently limited to languages for which corpora annotated with CCG derivations are available. Clark and Curran (2006) show that CCG parsers can be learned from sentences labeled with just supertags—without full derivations—with little loss in accuracy. The improvements we show here for learning supertaggers from lexicons without labeled data may be able to help create annotated resources more efficiently, or enable CCG parsers to be learned with less human-coded knowledge. Acknowledgements The authors would like to thank Johan Bos, Joey Frazee, Taesun Moon, the members of the UTNLL reading group, and the anonymous reviewers. Ravi and Knight acknowledge the support of the NSF (grant IIS-0904684) for this work. Baldridge acknowledges the support of a grant from the Morris Memorial Trust Fund of the New York Community Trust. References J. Baldridge. 2008. Weakly supervised supertagging with grammar-informed initialization. In Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 57–64, Manchester, UK, August. M. Banko and R. C. Moore. 2004. Part of speech tagging in context. In Proceedings of the International Conference on Computational Linguistics (COLING), page 556, Morristown, NJ, USA. A. R. Barron, J. Rissanen, and B. Yu. 1998. The minimum description length principle in coding and modeling. IEEE Transactions on Information Theory, 44(6):2743–2760. J. Bos, C. Bosco, and A. Mazzei. 2009. Converting a dependency treebank to a categorial grammar treebank for Italian. In Proceedings of the Eighth International Workshop on Treebanks and Linguistic Theories (TLT8), pages 27–38, Milan, Italy. S. Clark and J. Curran. 2006. Partial training for a lexicalized-grammar parser. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 144–151, New York City, USA, June. S. Clark and J. Curran. 2007. Wide-coverage efficient statistical parsing with CCG and log-linear models. Computational Linguistics, 33(4). M. Creutz and K. Lagus. 2002. Unsupervised discovery of morphemes. In Proceedings of the ACL Workshop on Morphological and Phonological Learning, pages 21–30, Morristown, NJ, USA. Y. Goldberg, M. Adler, and M. Elhadad. 2008. EM can find pretty good HMM POS-taggers (when given a good start). In Proceedings of the ACL, pages 746– 754, Columbus, Ohio, June. J. Goldsmith. 2001. Unsupervised learning of the morphology of a natural language. Computational Linguistics, 27(2):153–198. S. Goldwater and T. L. Griffiths. 2007. A fully Bayesian approach to unsupervised part-of-speech tagging. In Proceedings of the ACL, pages 744–751, Prague, Czech Republic, June. H. Hassan, K. Sima’an, and A. Way. 2009. A syntactified direct translation model with linear-time decoding. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1182–1191, Singapore, August. J. Hockenmaier and M. Steedman. 2007. CCGbank: A corpus of CCG derivations and dependency structures extracted from the Penn Treebank. Computational Linguistics, 33(3):355–396. A. Joshi. 1988. Tree Adjoining Grammars. In David Dowty, Lauri Karttunen, and Arnold Zwicky, editors, Natural Language Parsing, pages 206–250. Cambridge University Press, Cambridge. M. P. Marcus, M. A. Marcinkiewicz, and B. Santorini. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19(2). B. Merialdo. 1994. Tagging English text with a probabilistic model. Computational Linguistics, 20(2):155–171. C. Pollard and I. Sag. 1994. Head Driven Phrase Structure Grammar. CSLI/Chicago University Press, Chicago. S. Ravi and K. Knight. 2009. Minimized models for unsupervised part-of-speech tagging. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 504–512, Suntec, Singapore, August. M. Steedman. 2000. The Syntactic Process. MIT Press, Cambridge, MA. Kristina Toutanova and Mark Johnson. 2008. A Bayesian LDA-based model for semi-supervised part-of-speech tagging. In Proceedings of the Advances in Neural Information Processing Systems (NIPS), pages 1521–1528, Cambridge, MA. MIT Press. 503
|
2010
|
51
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 504–513, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Practical very large scale CRFs Thomas Lavergne LIMSI – CNRS [email protected] Olivier Capp´e T´el´ecom ParisTech LTCI – CNRS [email protected] Franc¸ois Yvon Universit´e Paris-Sud 11 LIMSI – CNRS [email protected] Abstract Conditional Random Fields (CRFs) are a widely-used approach for supervised sequence labelling, notably due to their ability to handle large description spaces and to integrate structural dependency between labels. Even for the simple linearchain model, taking structure into account implies a number of parameters and a computational effort that grows quadratically with the cardinality of the label set. In this paper, we address the issue of training very large CRFs, containing up to hundreds output labels and several billion features. Efficiency stems here from the sparsity induced by the use of a ℓ1 penalty term. Based on our own implementation, we compare three recent proposals for implementing this regularization strategy. Our experiments demonstrate that very large CRFs can be trained efficiently and that very large models are able to improve the accuracy, while delivering compact parameter sets. 1 Introduction Conditional Random Fields (CRFs) (Lafferty et al., 2001; Sutton and McCallum, 2006) constitute a widely-used and effective approach for supervised structure learning tasks involving the mapping between complex objects such as strings and trees. An important property of CRFs is their ability to handle large and redundant feature sets and to integrate structural dependency between output labels. However, even for simple linear chain CRFs, the complexity of learning and inference This work was partly supported by ANR projects CroTaL (ANR-07-MDCO-003) and MGA (ANR-07-BLAN-031102). grows quadratically with respect to the number of output labels and so does the number of structural features, ie. features testing adjacent pairs of labels. Most empirical studies on CRFs thus either consider tasks with a restricted output space (typically in the order of few dozens of output labels), heuristically reduce the use of features, especially of features that test pairs of adjacent labels1, and/or propose heuristics to simulate contextual dependencies, via extended tests on the observations (see discussions in, eg., (Punyakanok et al., 2005; Liang et al., 2008)). Limitating the feature set or the number of output labels is however frustrating for many NLP tasks, where the type and number of potentially relevant features are very large. A number of studies have tried to alleviate this problem. Pal et al. (2006) propose to use a “sparse” version of the forward-backward algorithm during training, where sparsity is enforced through beam pruning. Related ideas are discussed by Dietterich et al. (2004); by Cohn (2006), who considers “generalized” feature functions; and by Jeong et al. (2009), who use approximations to simplify the forward-backward recursions. In this paper, we show that the sparsity that is induced by ℓ1-penalized estimation of CRFs can be used to reduce the total training time, while yielding extremely compact models. The benefits of sparsity are even greater during inference: less features need to be extracted and included in the potential functions, speeding up decoding with a lesser memory footprint. We study and compare three different ways to implement ℓ1 penalty for CRFs that have been introduced recently: orthantwise Quasi Newton (Andrew and Gao, 2007), stochastic gradient descent (Tsuruoka et al., 2009) and coordinate descent (Sokolovska et al., 2010), concluding that these methods have complemen1In CRFsuite (Okazaki, 2007), it is even impossible to jointly test a pair of labels and a test on the observation, bigrams feature are only of the form f(yt−1, yt). 504 tary strengths and weaknesses. Based on an efficient implementation of these algorithms, we were able to train very large CRFs containing more than a hundred of output labels and up to several billion features, yielding results that are as good or better than the best reported results for two NLP benchmarks, text phonetization and part-of-speech tagging. Our contribution is therefore twofold: firstly a detailed analysis of these three algorithms, discussing implementation, convergence and comparing the effect of various speed-ups. This comparison is made fair and reliable thanks to the reimplementation of these techniques in the same software package. Second, the experimental demonstration that using large output label sets is doable and that very large feature sets actually help improve prediction accuracy. In addition, we show how sparsity in structured feature sets can be used in incremental training regimes, where long-range features are progressively incorporated in the model insofar as the shorter range features have proven useful. The rest of the paper is organized as follows: we first recall the basics of CRFs in Section 2, and discuss three ways to train CRFs with a ℓ1 penalty in Section 3. We then detail several implementation issues that need to be addressed when dealing with massive feature sets in Section 4. Our experiments are reported in Section 5. The main conclusions of this study are drawn in Section 6. 2 Conditional Random Fields In this section, we recall the basics of Conditional Random Fields (CRFs) (Lafferty et al., 2001; Sutton and McCallum, 2006) and introduce the notations that will be used throughout. 2.1 Basics CRFs are based on the following model pθ(y|x) = 1 Zθ(x) exp ( K X k=1 θkFk(x, y) ) (1) where x = (x1, . . . , xT ) and y = (y1, . . . , yT ) are, respectively, the input and output sequences2, and Fk(x, y) is equal to PT t=1 fk(yt−1, yt, xt), where {fk}1≤k≤K is an arbitrary set of feature 2Our implementation also includes a special label y0, that is always observed and marks the beginning of a sequence. functions and {θk}1≤k≤K are the associated parameter values. We denote by Y and X, respectively, the sets in which yt and xt take their values. The normalization factor in (1) is defined by Zθ(x) = X y∈Y T exp ( K X k=1 θkFk(x, y) ) . (2) The most common choice of feature functions is to use binary tests. In the sequel, we distinguish between two types of feature functions: unigram features fy,x, associated with parameters µy,x, and bigram features fy′,y,x, associated with parameters λy′,y,x. These are defined as fy,x(yt−1, yt, xt) = 1(yt = y, xt = x) fy′,y,x(yt−1, yt, xt) = 1(yt−1 = y′, yt = y, xt = x) where 1(cond.) is equal to 1 when the condition is verified and to 0 otherwise. In this setting, the number of parameters K is equal to |Y |2×|X|train, where |·| denotes the cardinal and |X|train refers to the number of configurations of xt observed during training. Thus, even in moderate size applications, the number of parameters can be very large, mostly due to the introduction of sequential dependencies in the model. This also explains why it is hard to train CRFs with dependencies spanning more than two adjacent labels. Using only unigram features {fy,x}(y,x)∈Y ×X results in a model equivalent to a simple bag-of-tokens positionby-position logistic regression model. On the other hand, bigram features {fy′,y,x}(y,x)∈Y 2×X are helpful in modelling dependencies between successive labels. The motivations for using simultaneously both types of feature functions are evaluated experimentally in Section 5. 2.2 Parameter Estimation Given N independent sequences {x(i), y(i)}N i=1, where x(i) and y(i) contain T (i) symbols, conditional maximum likelihood estimation is based on the minimization, with respect to θ, of the negated conditional log-likelihood of the observations l(θ) = − N X i=1 log pθ(y(i)|x(i)) (3) = N X i=1 ( log Zθ(x(i)) − K X k=1 θkFk(x(i), y(i)) ) This term is usually complemented with an additional regularization term so as to avoid overfitting 505 (see Section 3.1 below). The gradient of l(θ) is ∂l(θ) ∂θk = N X i=1 T (i) X t=1 Epθ(y|x(i)) fk(yt−1, yt, x(i) t ) − N X i=1 T (i) X t=1 fk(y(i) t−1, y(i) t , x(i) t ) (4) where Epθ(y|x) denotes the conditional expectation given the observation sequence, i.e. Epθ(y|x) fk(yt−1, yt, x(i) t ) = X (y′,y)∈Y 2 fk(y, y′, xt) Pθ(yt−1 = y′, yt = y|x) (5) Although l(θ) is a smooth convex function, its optimum cannot be computed in closed form, and l(θ) has to be optimized numerically. The computation of its gradient implies to repeatedly compute the conditional expectation in (5) for all input sequences x(i) and all positions t. The standard approach for computing these expectations is inspired by the forward-backward algorithm for hidden Markov models: using the notations introduced above, the algorithm implies the computation of the forward ( α1(y) = exp(µy,x1 + λy0,y,x1) αt+1(y) = P y′ αt(y′) exp(µy,xt+1 + λy′,y,xt+1) and backward recursions ( βTi(y) = 1 βt(y′) = P y βt+1(y) exp(µy,xt+1 + λy′,y,xt+1), for all indices 1 ≤t ≤T and all labels y ∈Y . Then, Zθ(x) = P y αT (y) and the pairwise probabilities Pθ(yt = y′, yt+1 = y|x) are given by αt(y′) exp(µy,xt+1 + λy′,y,xt+1)βt+1(y)/Zθ(x) These recursions require a number of operations that grows quadratically with |Y |. 3 ℓ1 Regularization in CRFs 3.1 Regularization The standard approach for parameter estimation in CRFs consists in minimizing the logarithmic loss l(θ) defined by (3) with an additional ℓ2 penalty term ρ2 2 ∥θ∥2 2, where ρ2 is a regularization parameter. The objective function is then a smooth convex function to be minimized over an unconstrained parameter space. Hence, any numerical optimization strategy may be used and practical solutions include limited memory BFGS (L-BFGS) (Liu and Nocedal, 1989), which is used in the popular CRF++ (Kudo, 2005) and CRFsuite (Okazaki, 2007) packages; conjugate gradient (Nocedal and Wright, 2006) and Stochastic Gradient Descent (SGD) (Bottou, 2004; Vishwanathan et al., 2006), used in CRFsgd (Bottou, 2007). The only caveat is to avoid numerical optimizers that require the full Hessian matrix (e.g., Newton’s algorithm) due to the size of the parameter vector in usual applications of CRFs. The most significant alternative to ℓ2 regularization is to use a ℓ1 penalty term ρ1∥θ∥1: such regularizers are able to yield sparse parameter vectors in which many component have been zeroed (Tibshirani, 1996). Using a ℓ1 penalty term thus implicitly performs feature selection, where ρ1 controls the amount of regularization and the number of extracted features. In the following, we will jointly use both penalty terms, yielding the socalled elastic net penalty (Zhou and Hastie, 2005) which corresponds to the objective function l(θ) + ρ1∥θ∥1 + ρ2 2 ∥θ∥2 2 (6) The use of both penalty terms makes it possible to control the number of non zero coefficients and to avoid the numerical problems that might occur in large dimensional parameter settings (see also (Chen, 2009)). However, the introduction of a ℓ1 penalty term makes the optimization of (6) more problematic, as the objective function is no longer differentiable in 0. Various strategies have been proposed to handle this difficulty. We will only consider here exact approaches and will not discuss heuristic strategies such as grafting (Perkins et al., 2003; Riezler and Vasserman, 2004). 3.2 Quasi Newton Methods To deal with ℓ1 penalties, a simple idea is that of (Kazama and Tsujii, 2003), originally introduced for maxent models. It amounts to reparameterizing θk as θk = θ+ k −θ− k , where θ+ k and θ− k are positive. The ℓ1 penalty thus becomes ρ1(θ+ −θ−). In this formulation, the objective function recovers its smoothness and can be optimized with conventional algorithms, subject to domain constraints. Optimization is straightforward, but the number of parameters is doubled and convergence is slow 506 (Andrew and Gao, 2007): the procedure lacks a mechanism for zeroing out useless parameters. A more efficient strategy is the orthant-wise quasi-Newton (OWL-QN) algorithm introduced in (Andrew and Gao, 2007). The method is based on the observation that the ℓ1 norm is differentiable when restricted to a set of points in which each coordinate never changes its sign (an “orthant”), and that its second derivative is then zero, meaning that the ℓ1 penalty does not change the Hessian of the objective on each orthant. An OWL-QN update then simply consists in (i) computing the Newton update in a well-chosen orthant; (ii) performing the update, which might cause some component of the parameter vector to change sign; and (iii) projecting back the parameter value onto the initial orthant, thereby zeroing out those components. In (Gao et al., 2007), the authors show that OWL-QN is faster than the algorithm proposed by Kazama and Tsujii (2003) and can perform model selection even in very high-dimensional problems, with no loss of performance compared to the use of ℓ2 penalty terms. 3.3 Stochastic Gradient Descent Stochastic gradient (SGD) approaches update the parameter vector based on an crude approximation of the gradient (4), where the computation of expectations only includes a small batch of observations. SGD updates have the following form θk ←θk + η∂l(θ) ∂θk , (7) where η is the learning rate. In (Tsuruoka et al., 2009), various ways of adapting this update to ℓ1penalized likelihood functions are discussed. Two effective ideas are proposed: (i) only update parameters that correspond to active features in the current observation, (ii) keep track of the cumulated penalty zk that θk should have received, had the gradient been computed exactly, and use this value to “clip” the parameter value. This is implemented by patching the update (7) as follows ( if (θk > 0) θk ←max(0, θk −zk) else if (θk < 0) θk ←min(0, θk −zk) (8) Based on a study of three NLP benchmarks, the authors of (Tsuruoka et al., 2009) claim this approach to be much faster than the orthant-wise approach and yet to yield very comparable performance, while selecting slightly larger feature sets. 3.4 Block Coordinate Descent The coordinate descent approach of Dud´ık et al. (2004) and Friedman et al. (2008) uses the fact that optimizing a mono-dimensional quadratic function augmented with a ℓ1 penalty can be performed analytically. For arbitrary functions, this idea can be adapted by considering quadratic approximations of the objective around the current value ¯θ lk,¯θ(θk) = ∂l(¯θ) ∂θk (θk −¯θk) + 1 2 ∂2l(¯θ) ∂θ2 k (θk −¯θk)2 + ρ1|θk| + ρ2 2 θ2 k + Cst (9) The minimizer of the approximation (9) is simply θk = s n ∂2l(¯θ) ∂θ2 k ¯θk −∂l(¯θ) ∂θk , ρ1 o ∂2l(¯θ) ∂θ2 k + ρ2 (10) where s is the soft-thresholding function s(z, ρ) = z −ρ if z > ρ z + ρ if z < −ρ 0 otherwise (11) Coordinate descent is ported to CRFs in (Sokolovska et al., 2010). Making this scheme practical requires a number of adaptations, including (i) approximating the second order term in (10), (ii) performing updates in block, where a block contains the |Y | × |Y + 1| features νy′,y,x and λy,x for a fixed test x on the observation sequence and (iii) approximating the Hessian for a block by its diagonal terms. (ii) is specially critical, as repeatedly cycling over individual features to perform the update (10) is only possible with restricted sets of features. The block update schemes uses the fact that all features within a block appear in the same set of sequences, which means that most of the computations needed to perform theses updates can be shared within the block. One advantage of the resulting algorithm, termed BCD in the following, is that the update of θk only involves carrying out the forward-backward recursions for the set of sequences that contain symbols x such that at least one {fk(y′, y, x)}(y,y′)∈Y 2 is non null, which can be much smaller than the whole training set. 507 4 Implementation Issues Efficiently processing very-large feature and observation sets requires to pay attention to many implementation details. In this section, we present several optimizations devised to speed up training. 4.1 Sparse Forward-Backward Recursions For all algorithms, the computation time is dominated by the evaluations of the gradient: our implementation takes advantage of the sparsity to accelerate these computations. Assume the set of bigram features {λy′,y,xt+1}(y′,y)∈Y 2 is sparse with only r(xt+1) ≪|Y |2 non null values and define the |Y | × |Y | sparse matrix Mt(y′, y) = exp(λy′,y,xt) −1. Using M, the forward-backward recursions are αt(y) = X y′ ut−1(y′) + X y′ ut−1(y′)Mt(y′, y) βt(y′) = X y vt+1(y) + X y Mt+1(y′, y)vt+1(y) with ut−1(y) = exp(µy,xt)αt−1(y) and vt+1(y) = exp(µy,xt+1)βt+1(y). (Sokolovska et al., 2010) explains how computational savings can be obtained using the fact that the vector/matrix products in the recursions above only involve the sparse matrix Mt+1(y′, y). They can thus be computed with exactly r(xt+1) multiplications instead of |Y |2. The same idea can be used when the set {µy,xt+1}y∈Y of unigram features is sparse. Using this implementation, the complexity of the forward-backward procedure for x(i) can be made proportional to the average number of active features per position, which can be much smaller than the number of potentially active features. For BCD, forward-backward can even be made slightly faster. When computing the gradient wrt. features λy,x and µy′,y,x (for all the values of y and y′) for sequence x(i), assuming that x only occurs once in x(i) at position t, all that is needed is α′ t(y), ∀t′ ≤t and β′ t(y), ∀t′ ≥t. Zθ(x) is then recovered as P y αt(y)βt(y). Forward-backward recursions can thus be truncated: in our experiments, this divided the computational cost by 1,8 on average. Note finally that forward-backward is performed on a per-observation basis and is easily parallelized (see also (Mann et al., 2009) for more powerful ways to distribute the computation when dealing with very large datasets). In our implementation, it is distributed on all available cores, resulting in significant speed-ups for OWL-QN and L-BFGS; for BCD the gain is less acute, as parallelization only helps when updating the parameters for a block of features that are occur in many sequences; for SGD, with batches of size one, this parallelization policy is useless. 4.2 Scaling Most existing implementations of CRFs, eg. CRF++ and CRFsgd perform the forwardbackward recursions in the log-domain, which guarantees that numerical over/underflows are avoided no matter the length T (i) of the sequence. It is however very inefficient from an implementation point of view, due to the repeated calls to the exp() and log() functions. As an alternative way of avoiding numerical problems, our implementation, like crfSuite’s, resorts to “scaling”, a solution commonly used for HMMs. Scaling amounts to normalizing the values of αt and βt to one, making sure to keep track of the cumulated normalization factors so as to compute Zθ(x) and the conditional expectations Epθ(y|x). Also note that in our implementation, all the computations of exp(x) are vectorized, which provides an additional speed up of about 20%. 4.3 Optimization in Large Parameter Spaces Processing very large feature vectors, up to billions of components, is problematic in many ways. Sparsity has been used here to speed up forwardbackward, but we have made no attempt to accelerate the computation of the OWL-QN updates, which are linear in the size of the parameter vector. Of the three algorithms, BCD is the most affected by increases in the number of features, or more precisely, in the number of features blocks, where one block correspond to a specific test of the observation. In the worst case scenario, each block may require to visit all the training instances, yielding terrible computational wastes. In practice though, most blocks only require to process a small fraction of the training set, and the actual complexity depends on the average number of blocks per observations. Various strategies have been tried to further accelerate BCD, such as processing blocks that only visit one observation in parallel and updating simultaneously all the blocks that visit all the training instances, leading to a small speed-up on the POS-tagging task. 508 Working with billions of features finally requires to worry also about memory usage. In this respect, BCD is the most efficient, as it only requires to store one K-dimensional vector for the parameter itself. SGD requires two such vectors, one for the parameter and one for storing the zk (see Eq. (8)). In comparison, OWL-QN requires much more memory, due to the internals of the update routines, which require several histories of the parameter vector and of its gradient. Typically, our implementation necessitates in the order of a dozen K-dimensional vectors. Parallelization only makes things worse, as each core will also need to maintain its own copy of the gradient. 5 Experiments Our experiments use two standard NLP tasks, phonetization and part-of-speech tagging, chosen here to illustrate two very different situations, and to allow for comparison with results reported elsewhere in the literature. Unless otherwise mentioned, the experiments use the same protocol: 10 fold cross validation, where eight folds are used for training, one for development, and one for testing. Results are reported in terms of phoneme error rates or tag error rates on the test set. Comparing run-times can be a tricky matter, especially when different software packages are involved. As discussed above, the observed runtimes depend on many small implementation details. As the three algorithms share as much code as possible, we believe the comparison reported hereafter to be fair and reliable. All experiments were performed on a server with 64G of memory and two Xeon processors with 4 cores at 2.27 Ghz. For comparison, all measures of run-times include the cumulated activity of all cores and give very pessimistic estimates of the wall time, which can be up to 7 times smaller. For OWL-QN, we use 5 past values of the gradient to approximate the inverse of the Hessian matrix: increasing this value had no effect on accuracy or convergence and was detrimental to speed; for SGD, the learning rate parameter was tuned manually. Note that we have not spent much time optimizing the values of ρ1 and ρ2. Based on a pilot study on Nettalk, we found that taking ρ1 = .5 and ρ2 in the order of 10−5 to yield nearly optimal performance, and have used these values throughout. 5.1 Tasks and Settings 5.1.1 Nettalk Our first benchmark is the word phonetization task, using the Nettalk dictionary (Sejnowski and Rosenberg, 1987). This dataset contains approximately 20,000 English word forms, their pronunciation, plus some prosodic information (stress markers for vowels, syllabic parsing for consonants). Grapheme and phoneme strings are aligned at the character level, thanks to the use of a “null sound” in the latter string when it is shorter than the former; likewise, each prosodic mark is aligned with the corresponding letter. We have derived two test conditions from this database. The first one is standard and aims at predicting the pronunciation information only. In this setting, the set of observations (X) contains 26 graphemes, and the output label set contains |Y | = 51 phonemes. The second condition aims at jointly predicting phonemic and prosodic information3. The reasons for designing this new condition are twofold: firstly, it yields a large set of composite labels (|Y | = 114) and makes the problem computationally challenging. Second, it allows to quantify how much the information provided by the prosodic marks help predict the phonemic labels. Both information are quite correlated, as the stress mark and the syllable openness, for instance, greatly influence the realization of some archi-phonemes. The features used in Nettalk experiments take the form fy,w (unigram) and fy′,y,w (bigram), where w is a n-gram of letters. The n-grm feature sets (n = {1, 3, 5, 7}) includes all features testing embedded windows of k letters, for all 0 ≤k ≤n; the n-grm- setting is similar, but only includes the window of length n; in the n-grm+ setting, we add features for odd-size windows; in the ngrm++ setting, we add all sequences of letters up to size n occurring in current window. For instance, the active bigram features at position t = 2 in the sequence x=’lemma’ are as follows: the 3grm feature set contains fy,y′, fy,y′,e and fy′,y,lem; only the latter appears in the 3-grm- setting. In the 3-grm+ feature set, we also have fy′,y,le and fy′,y,em. The 3-grm++ feature set additionally includes fy′,y,l and fy′,y,m. The number of features ranges from 360 thousands (1-grm setting) to 1.6 billion (7-grm). 3Given the design of the Nettalk dictionary, this experiment required to modify the original database so as to reassign prosodic marks to phonemes, rather than to letters. 509 Features With Without Nettalk 3-grm 10.74% 14.3M 14.59% 0.3M 5-grm 8.48% 132.5M 11.54% 2.5M POS tagging base 2.91% 436.7M 3.47% 70.2M Table 1: Features jointly testing label pairs and the observation are useful (error rates and features counts.) ℓ2 ℓ1-sparse ℓ1 % zero 1-grm 84min 41min 57min 44.6% 3-grm- 65min 16min 44min 99.6% 3-grm 72min 48min 58min 19.9% Table 2: Sparse vs standard forward-backward (training times and percentages of sparsity of M) 5.1.2 Part-of-Speech Tagging Our second benchmark is a part-of-speech (POS) tagging task using the PennTreeBank corpus (Marcus et al., 1993), which provides us with a quite different condition. For this task, the number of labels is smaller (|Y | = 45) than for Nettalk, and the set of observations is much larger (|X| = 43207). This benchmark, which has been used in many studies, allows for direct comparisons with other published work. We thus use a standard experimental set-up, where sections 0-18 of the Wall Street Journal are used for training, sections 19-21 for development, and sections 22-24 for testing. Features are also standard and follow the design of (Suzuki and Isozaki, 2008) and test the current words (as written and lowercased), prefixes and suffixes up to length 4, and typographical characteristics (case, etc.) of the words. Our baseline feature set also contains tests on individual and pairs of words in a window of 5 words. 5.2 Using Large Feature Sets The first important issue is to assess the benefits of using large feature sets, notably including features testing both a bigram of labels and an observation. Table 1 compares the results obtained with and without these features for various setting (using OWL-QN to perform the optimization), suggesting that for the tasks at hand, these features are actually helping. ℓ2 ℓ1 Elastic-net 1-grm 17.81% 17.86% 17.79% 3-grm 10.62% 10.74% 10.70% 5-grm 8.50% 8.45% 8.48% Table 3: Error rates of the three regularizers on the Nettalk task. 5.3 Speed, Sparsity, Convergence The training speed depends of two main factors: the number of iterations needed to achieve convergence and the computational cost of one iteration. In this section, we analyze and compare the runtime efficiency of the three optimizers. 5.3.1 Convergence As far as convergence is concerned, the two forms of regularization (ℓ2 and ℓ1) yield the same performance (see Table 3), and the three algorithms exhibit more or less the same behavior. They quickly reach an acceptable set of active parameters, which is often several orders of magnitude smaller than the whole parameter set (see results below in Table 4 and 5). Full convergence, reflected by a stabilization of the objective function, is however not so easily achieved. We have often observed a slow, yet steady, decrease of the log-loss, accompanied with a diminution of the number of active features as the number of iterations increases. Based on this observation, we have chosen to stop all algorithms based on their performance on an independent development set, allowing a fair comparison of the overall training time; for OWL-QN, it allowed to divide the total training time by almost 2. It has finally often been found useful to fine tune the non-zero parameters by running a final handful of L-BFGS iterations using only a small ℓ2 penalty; at this stage, all the other features are removed from the model. This had a small impact BCD and SGD’s performance and allowed them to catch up with OWL-QN’s performance. 5.3.2 Sparsity and the Forward-Backward As explained in section 4.1, the forward-backward algorithm can be written so as to use the sparsity of the matrix My,y′,x. To evaluate the resulting speed-up, we ran a series of experiments using Nettalk (see Table 2). In this table, the 3-grm- setting corresponds to maximum sparsity for M, and training with the sparse algorithm is three times faster than with the non-sparse version. Throwing 510 Method Iter. # Feat. Error Time OWL-QN 1-grm 63.4 4684 17.79% 11min 7-grm 140.2 38214 8.12% 1h02min 5-grm+ 141.0 43429 7.89% 1h37min SGD 1-grm 21.4 3540 18.21% 9min 5-grm+ 28.5 34319 8.01% 45min BCD 1-grm 28.2 5017 18.27% 27min 7-grm 9.2 3692 8.21% 1h22min 5-grm+ 8.7 47675 7.91% 2h18min Table 4: Performance on Nettalk in more features has the effect of making M much more dense, mitigating the benefits of the sparse recursions. Nevertheless, even for very large feature sets, the percentage of zeros in M averages 20% to 30%, and the sparse version remains 10 to 20% faster than the non-sparse one. Note that the non-sparse version is faster with a ℓ1 penalty term than with only the ℓ2 term: this is because exp(0) is faster to evaluate than exp(x) when x ̸= 0. 5.3.3 Training Speed and Test Accuracy Table 4 displays the results achieved on the Nettalk task. The three algorithms yield very comparable accuracy results, and deliver compact models: for the 5-gram+ setting, only 50,000 out of 250 million features are selected. SGD is the fastest of the three, up to twice as fast as OWL-QN and BCD depending on the feature set. The performance it achieves are consistently slightly worst than the other optimizers, and only catch up when the parameters are fine-tuned (see above). There are not so many comparisons for Nettalk with CRFs, due to the size of the label set. Our results compare favorably with those reported in (Pal et al., 2006), where the accuracy attains 91.7% using 19075 examples for training and 934 for testing, and with those in (Jeong et al., 2009) (88.4% accuracy with 18,000 (2,000) training (test) instances). Table 5 gives the results obtained for the larger Nettalk+prosody task. Here, we only report the results obtained with SGD and BCD. For OWL-QN, the largest model we could handle was the 3-grm model, which contained 69 million features, and took 48min to train. Here again, performance steadily increase with the number of features, showing the benefits of large-scale models. We lack comparisons for this task, which seems considerably harder than the sole phonetization task, and all systems seem to plateau around 13.5% accuracy. Interestingly, simultaMethod Error Time SGD 5-grm 14.71% / 8.11% 55min 5-grm+ 13.91% / 7.51% 2h45min BCD 5-grm 14.57% / 8.06% 2h46min 7-grm 14.12% / 7.86% 3h02min 5-grm+ 13.85% / 7.47% 7h14min 5-grm++ 13.69% / 7.36% 16h03min Table 5: Performance on Nettalk+prosody. Error is given for both joint labels and phonemic labels. neously predicting the phoneme and its prosodic markers allows to improve the accuracy on the prediction of phonemes, which improves of almost a half point as compared to the best Nettalk system. For the POS tagging task, BCD appears to be unpractically slower to train than the others approaches (SGD takes about 40min to train, OWLQN about 1 hour) due the simultaneous increase in the sequence length and in the number of observations. As a result, one iteration of BCD typically requires to repeatedly process over and over the same sequences: on average, each sequence is visited 380 times when we use the baseline feature set. This technique should reserved for tasks where the number of blocks is small, or, as below, when memory usage is an issue. 5.4 Structured Feature Sets In many tasks, the ambiguity of tokens can be reduced by looking up increasingly large windows of local context. This strategy however quickly runs into a combinatorial increase of the number of features. A side note of the Nettalk experiments is that when using embedded features, the active feature set tends to reflect this hierarchical organization. This means that when a feature testing a n-gram is active, in most cases, the features for all embedded k-grams are also selected. Based on this observation, we have designed an incremental training strategy for the POS tagging task, where more specific features are progressively incorporated into the model if the corresponding less specific feature is active. This experiment used BCD, which is the most memory efficient algorithm. The first iteration only includes tests on the current word. During the second iteration, we add tests on bigram of words, on suffixes and prefixes up to length 4. After four iterations, we throw in features testing word trigrams, subject to the corresponding unigram block being active. After 6 iterations, we finally augment the 511 model with windows of length 5, subject to the corresponding trigram being active. After 10 iterations, the model contains about 4 billion features, out of which 400,000 are active. It achieves an error rate of 2.63% (resp. 2.78%) on the development (resp. test) data, which compares favorably with some of the best results for this task (for instance (Toutanova et al., 2003; Shen et al., 2007; Suzuki and Isozaki, 2008)). 6 Conclusion and Perspectives In this paper, we have discussed various ways to train extremely large CRFs with a ℓ1 penalty term and compared experimentally the results obtained, both in terms of training speed and of accuracy. The algorithms studied in this paper have complementary strength and weaknesses: OWL-QN is probably the method of choice in small or moderate size applications while BCD is most efficient when using very large feature sets combined with limited-size observation alphabets; SGD complemented with fine tuning appears to be the preferred choice in most large-scale applications. Our analysis demonstrate that training large-scale sparse models can be done efficiently and allows to improve over the performance of smaller models. The CRF package developed in the course of this study implements many algorithmic optimizations and allows to design innovative training strategies, such as the one presented in section 5.4. This package is released as open-source software and is available at http://wapiti.limsi.fr. In the future, we intend to study how sparsity can be used to speed-up training in the face of more complex dependency patterns (such as higher-order CRFs or hierarchical dependency structures (Rozenknop, 2002; Finkel et al., 2008). From a performance point of view, it might also be interesting to combine the use of large-scale feature sets with other recent improvements such as the use of semi-supervised learning techniques (Suzuki and Isozaki, 2008) or variable-length dependencies (Qian et al., 2009). References Galen Andrew and Jianfeng Gao. 2007. Scalable training of l1-regularized log-linear models. In Proceedings of the International Conference on Machine Learning, pages 33–40, Corvalis, Oregon. L´eon Bottou. 2004. Stochastic learning. In Olivier Bousquet and Ulrike von Luxburg, editors, Advanced Lectures on Machine Learning, Lecture Notes in Artificial Intelligence, LNAI 3176, pages 146–168. Springer Verlag, Berlin. L´eon Bottou. 2007. Stochastic gradient descent (sgd) implementation. http://leon.bottou.org/projects/sgd. Stanley Chen. 2009. Performance prediction for exponential language models. In Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 450–458, Boulder, Colorado, June. Trevor Cohn. 2006. Efficient inference in large conditional random fields. In Proceedings of the 17th European Conference on Machine Learning, pages 606–613, Berlin, September. Thomas G. Dietterich, Adam Ashenfelter, and Yaroslav Bulatov. 2004. Training conditional random fields via gradient tree boosting. In Proceedings of the International Conference on Machine Learning, Banff, Canada. Miroslav Dud´ık, Steven J. Phillips, and Robert E. Schapire. 2004. Performance guarantees for regularized maximum entropy density estimation. In John Shawe-Taylor and Yoram Singer, editors, Proceedings of the 17th annual Conference on Learning Theory, volume 3120 of Lecture Notes in Computer Science, pages 472–486. Springer. Jenny Rose Finkel, Alex Kleeman, and Christopher D. Manning. 2008. Efficient, feature-based, conditional random field parsing. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, pages 959–967, Columbus, Ohio. Jerome Friedman, Trevor Hastie, and Rob Tibshirani. 2008. Regularization paths for generalized linear models via coordinate descent. Technical report, Department of Statistics, Stanford University. Jianfeng Gao, Galen Andrew, Mark Johnson, and Kristina Toutanova. 2007. A comparative study of parameter estimation methods for statistical natural language processing. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 824–831, Prague, Czech republic. Minwoo Jeong, Chin-Yew Lin, and Gary Geunbae Lee. 2009. Efficient inference of crfs for large-scale natural language data. In Proceedings of the Joint Conference of the Annual Meeting of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing, pages 281–284, Suntec, Singapore. Jun’ichi Kazama and Jun’ichi Tsujii. 2003. Evaluation and extension of maximum entropy models with inequality constraints. In Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing, pages 137–144. Taku Kudo. 2005. CRF++: Yet another CRF toolkit. http://crfpp.sourceforge.net/. 512 John Lafferty, Andrew McCallum, and Fernando Pereira. 2001. Conditional random fields: probabilistic models for segmenting and labeling sequence data. In Proceedings of the International Conference on Machine Learning, pages 282–289. Morgan Kaufmann, San Francisco, CA. Percy Liang, Hal Daum´e, III, and Dan Klein. 2008. Structure compilation: trading structure for features. In Proceedings of the 25th international conference on Machine learning, pages 592–599. Dong C. Liu and Jorge Nocedal. 1989. On the limited memory BFGS method for large scale optimization. Mathematical Programming, 45:503–528. Gideon Mann, Ryan McDonald, Mehryar Mohri, Nathan Silberman, and Dan Walker. 2009. Efficient large-scale distributed training of conditional maximum entropy models. In Y. Bengio, D. Schuurmans, J. Lafferty, C. K. I. Williams, and A.Culotta, editors, Advances in Neural Information Processing Systems 22, pages 1231–1239. Mitchell P. Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. 1993. Building a large annotated corpus of English: The Penn treebank. Computational Linguistics, 19(2):313–330. Jorge Nocedal and Stephen Wright. 2006. Numerical Optimization. Springer. Naoaki Okazaki. 2007. CRFsuite: A fast implementation of conditional random fields (CRFs). http://www.chokkan.org/software/crfsuite/. Chris Pal, Charles Sutton, and Andrew McCallum. 2006. Sparse forward-backward using minimum divergence beams for fast training of conditional random fields. In Proceedings of the International Conference on Acoustics, Speech, and Signal Processing, Toulouse, France. Simon Perkins, Kevin Lacker, and James Theiler. 2003. Grafting: Fast, incremental feature selection by gradient descent in function space. Journal of Machine Learning Research, 3:1333–1356. Vasin Punyakanok, Dan Roth, Wen tau Yih, and Dav Zimak. 2005. Learning and inference over constrained output. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 1124–1129. Xian Qian, Xiaoqian Jiang, Qi Zhang, Xuanjing Huang, and Lide Wu. 2009. Sparse higher order conditional random fields for improved sequence labeling. In Proceedings of the Annual International Conference on Machine Learning, pages 849–856. Stefan Riezler and Alexander Vasserman. 2004. Incremental feature selection and l1 regularization for relaxed maximum-entropy modeling. In Dekang Lin and Dekai Wu, editors, Proceedings of the conference on Empirical Methods in Natural Language Processing, pages 174–181, Barcelona, Spain, July. Antoine Rozenknop. 2002. Mod`eles syntaxiques probabilistes non-g´en´eratifs. Ph.D. thesis, Dpt. d’informatique, ´Ecole Polytechnique F´ed´erale de Lausanne. Terrence J. Sejnowski and Charles R. Rosenberg. 1987. Parallel networks that learn to pronounce english text. Complex Systems, 1. Libin Shen, Giorgio Satta, and Aravind Joshi. 2007. Guided learning for bidirectional sequence classification. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 760–767, Prague, Czech Republic. Nataliya Sokolovska, Thomas Lavergne, Olivier Capp´e, and Franc¸ois Yvon. 2010. Efficient learning of sparse conditional random fields for supervised sequence labelling. IEEE Selected Topics in Signal Processing. Charles Sutton and Andrew McCallum. 2006. An introduction to conditional random fields for relational learning. In Lise Getoor and Ben Taskar, editors, Introduction to Statistical Relational Learning, Cambridge, MA. The MIT Press. Jun Suzuki and Hideki Isozaki. 2008. Semi-supervised sequential labeling and segmentation using gigaword scale unlabeled data. In Proceedings of the Conference of the Association for Computational Linguistics on Human Language Technology, pages 665–673, Columbus, Ohio. Robert Tibshirani. 1996. Regression shrinkage and selection via the lasso. J.R.Statist.Soc.B, 58(1):267– 288. Kristina Toutanova, Dan Klein, Christopher D. Manning, and Yoram Singer. 2003. Feature-rich part-ofspeech tagging with a cyclic dependency network. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology, pages 173–180. Yoshimasa Tsuruoka, Jun’ichi Tsujii, and Sophia Ananiadou. 2009. Stochastic gradient descent training for l1-regularized log-linear models with cumulative penalty. In Proceedings of the Joint Conference of the Annual Meeting of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing, pages 477–485, Suntec, Singapore. S. V. N. Vishwanathan, Nicol N. Schraudolph, Mark Schmidt, and Kevin Murphy. 2006. Accelerated training of conditional random fields with stochastic gradient methods. In Proceedings of the 23th International Conference on Machine Learning, pages 969–976. ACM Press, New York, NY, USA. Hui Zhou and Trevor Hastie. 2005. Regularization and variable selection via the elastic net. J. Royal. Stat. Soc. B., 67(2):301–320. 513
|
2010
|
52
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 514–524, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics On the Computational Complexity of Dominance Links in Grammatical Formalisms Sylvain Schmitz LSV, ENS Cachan & CNRS, France [email protected] Abstract Dominance links were introduced in grammars to model long distance scrambling phenomena, motivating the definition of multiset-valued linear indexed grammars (MLIGs) by Rambow (1994b), and inspiring quite a few recent formalisms. It turns out that MLIGs have since been rediscovered and reused in a variety of contexts, and that the complexity of their emptiness problem has become the key to several open questions in computer science. We survey complexity results and open issues on MLIGs and related formalisms, and provide new complexity bounds for some linguistically motivated restrictions. 1 Introduction Scrambling constructions, as found in German and other SOV languages (Becker et al., 1991; Rambow, 1994a; Lichte, 2007), cause notorious difficulties to linguistic modeling in classical grammar formalisms like HPSG or TAG. A well-known illustration of this situation is given in the following two German sentences for “that Peter has repaired the fridge today” (Lichte, 2007), dass [Peter] heute [den K¨uhlschrank] repariert hat that Peternom today the fridgeacc repaired has dass [den K¨uhlschrank] heute [Peter] repariert hat that the fridgeacc today Peternom repaired has with a flexible word order between the two complements of repariert, namely between the nominative Peter and the accusative den K¨uhlschrank. Rambow (1994b) introduced a formalism, unordered vector grammars with dominance links (UVG-dls), for modeling such phenomena. These grammars are defined by vectors of contextfree productions along with dominance links that VP NPnom VP VP NPacc VP VP V repariert Figure 1: A vector of productions for the verb repariert together with its two complements. should be enforced during derivations; for instance, Figure 1 shows how a flexible order between the complements of repariert could be expressed in an UVG-dl. Similar dominance mechanisms have been employed in various tree description formalisms (Rambow et al., 1995; Rambow et al., 2001; Candito and Kahane, 1998; Kallmeyer, 2001; Guillaume and Perrier, 2010) and TAG extensions (Becker et al., 1991; Rambow, 1994a). However, the prime motivation for this survey is another grammatical formalism defined in the same article: multiset-valued linear indexed grammars (Rambow, 1994b, MLIGs), which can be seen as a low-level variant of UVG-dls that uses multisets to emulate unfulfilled dominance links in partial derivations. It is a natural extension of Petri nets, with broader scope than just UVG-dls; indeed, it has been independently rediscovered by de Groote et al. (2004) in the context of linear logic, and by Verma and Goubault-Larrecq (2005) in that of equational theories. Moreover, the decidability of its emptiness problem has proved to be quite challenging and is still uncertain, with several open questions depending on its resolution: • provability in multiplicative exponential linear logic (de Groote et al., 2004), • emptiness and membership of abstract categorial grammars (de Groote et al., 2004; Yoshinaka and Kanazawa, 2005), • emptiness and membership of Stabler (1997)’s minimalist grammars without 514 shortest move constraint (Salvati, 2010), • satisfiability of first-order logic on data trees (Boja´nczyk et al., 2009), and of course • emptiness and membership for the various formalisms that embed UVG-dls. Unsurprisingly in the light of their importance in different fields, several authors have started investigating the complexity of decisions problems for MLIGs (Demri et al., 2009; Lazi´c, 2010). We survey the current state of affairs, with a particular emphasis on two points: 1. the applicability of complexity results to UVG-dls, which is needed if we are to conclude anything on related formalisms with dominance links, 2. the effects of two linguistically motivated restrictions on such formalisms, lexicalization and boundedness/rankedness. The latter notion is imported from Petri nets, and turns out to offer interesting new complexity trade-offs, as we prove that k-boundedness and k-rankedness are EXPTIME-complete for MLIGs, and that the emptiness and membership problems are EXPTIME-complete for k-bounded MLIGs but PTIME-complete in the k-ranked case. This also implies an EXPTIME lower bound for emptiness and membership in minimalist grammars with shortest move constraint. We first define MLIGs formally in Section 2 and review related formalisms in Section 3. We proceed with complexity results in Section 4 before concluding in Section 5. Notations In the following, Σ denotes a finite alphabet, Σ∗the set of finite sentences over Σ, and ε the empty string. The length of a string w is noted |w|, and the number of occurrence of a symbol a in w is noted |w|a. A language is formalized as a subset of Σ∗. Let Nn denote the set of vectors of positive integers of dimension n. The i-th component of a vector x in Nn is x(i), 0 denotes the null vector, 1 the vector with 1 values, and ei the vector with 1 as its i-th component and 0 everywhere else. The ordering ≤on Nn is the componentwise ordering: x ≤y iff x(i) ≤y(i) for all 0 < i ≤n. The size of a vector refers to the size of its binary encoding: |x| = Pn i=1 1 + max(0, ⌊log2 x(i)⌋). We refer the reader unfamiliar with complexity classes and notions such as hardness or LOGSPACE reductions to classical textbooks (e.g. Papadimitriou, 1994). 2 Multiset-Valued Linear Indexed Grammars Definition 1 (Rambow, 1994b). An ndimensional multiset-valued linear indexed grammar (MLIG) is a tuple G = ⟨N, Σ, P, (S, x0)⟩ where N is a finite set of nonterminal symbols, Σ a finite alphabet disjoint from N, V = (N ×Nn)⊎Σ the vocabulary, P a finite set of productions in (N × Nn) × V ∗, and (S, x0) ∈N × Nn the start symbol. Productions are more easily written as (A,x) →u0(B1,x1)u1 · · · um(Bm,xm)um+1 (⋆) with each ui in Σ∗and each (Bi, xi) in N × Nn. The derivation relation ⇒over sequences in V ∗ is defined by δ(A,y)δ′ ⇒δu0(B1,y1)u1 · · · um(Bm,ym)um+1δ′ if δ and δ′ are in V ∗, a production of form (⋆) appears in P, x ≤y, for each 1 ≤i ≤m, xi ≤yi, and y −x = Pm i=1 yi −xi. The language of a MLIG is the set of terminal strings derived from (S, x0), i.e. L(G) = {w ∈Σ∗| (S, x0) ⇒∗w} and we denote by L(MLIG) the class of MLIG languages. Example 2. To illustrate this definition, and its relevance for free word order languages, consider the 3-dimensional MLIG with productions (S, 0) →ε | (S, 1), (S, e1) →a (S, 0), (S, e2) →b (S, 0), (S, e3) →c (S, 0) and start symbol (S, 0). It generates the MIX language of all sentences with the same number of a, b, and c’s (see Figure 2 for an example derivation): Lmix = {w ∈{a, b, c}∗| |w|a = |w|b = |w|c} . The size |G| of a MLIG G is essentially the sum of the sizes of each of its productions of form (⋆): |x0| + X P m + 1 + |x| + m X i=1 |xi| + m+1 X i=0 |ui| ! . 2.1 Normal Forms A MLIG is in extended two form (ETF) if all its productions are of form terminal (A, 0) →a or (A, 0) →ε, or 515 S, (0, 0, 0) S, (1, 1, 1) b S, (1, 0, 1) S, (2, 1, 2) c S, (2, 1, 1) a S, (1, 1, 1) a S, (0, 1, 1) b S, (0, 0, 1) c S, (0, 0, 0) ε Figure 2: A derivation for bcaabc in the grammar of Example 2. nonterminal (A, x) → (B1, x1)(B2, x2) or (A, x) →(B1, x1), with a in Σ, A, B1, B2 in N, and x, x1, x2 in Nn. Using standard constructions, any MLIG can be put into ETF in linear time or logarithmic space. A MLIG is in restricted index normal form (RINF) if the productions in P are of form (A,0) →α, (A,0) →(B,ei), or (A,ei) → (B,0), with A, B in N, 0 < i ≤n, and α in (Σ∪(N ×{0}))∗. The direct translation into RINF proposed by Rambow (1994a) is exponential if we consider a binary encoding of vectors, but using techniques developed for Petri nets (Dufourd and Finkel, 1999), this blowup can be avoided: Proposition 3. For any MLIG, one can construct an equivalent MLIG in RINF in logarithmic space. 2.2 Restrictions Two restrictions on dominance links have been suggested in an attempt to reduce their complexity, sometimes in conjunction: lexicalization and k-boundedness. We provide here characterizations for them in terms of MLIGs. We can combine the two restrictions, thus defining the class of kbounded lexicalized MLIGs. Lexicalization Lexicalization in UVG-dls reflects the strong dependence between syntactic constructions (vectors of productions representing an extended domain of locality) and lexical anchors. We define here a restriction of MLIGs with similar complexity properties: Definition 4. A terminal derivation α ⇒p w with w in Σ∗is c-lexicalized for some c > 0 if p ≤ c·|w|.1 A MLIG is lexicalized if there exists c such that any terminal derivation starting from (S, x0) is c-lexicalized, and we denote by L(MLIGℓ) the set of lexicalized MLIG languages. Looking at the grammar of Example 2, any terminal derivation (S, 0) ⇒p w verifies p = 4·|w| 3 + 1, and the grammar is thus lexicalized. Boundedness As dominance links model longdistance dependencies, bounding the number of simultaneously pending links can be motivated on competence/performance grounds (Joshi et al., 2000; Kallmeyer and Parmentier, 2008), and on complexity/expressiveness grounds (Søgaard et al., 2007; Kallmeyer and Parmentier, 2008; Chiang and Scheffler, 2008). The shortest move constraint (SMC) introduced by Stabler (1997) to enforce a strong form of minimality also falls into this category of restrictions. Definition 5. A MLIG derivation α0 ⇒α1 ⇒ · · · ⇒αp is of rank k for some k ≥0 if, no vector with a sum of components larger than k can appear in any αj, i.e. for all x in Nn such that there exist 0 ≤j ≤p, δ, δ′ in V ∗and A in N with αj = δ(A, x)δ′, one has Pn i=1 x(i) ≤k. A MLIG is k-ranked (noted kr-MLIG) if any derivation starting with α0 = (S, x0) is of rank k. It is ranked if there exists k such that it is k-ranked. A 0-ranked MLIG is simply a context-free grammar (CFG), and we have more generally the following: Lemma 6. Any n-dimensional k-ranked MLIG G can be transformed into an equivalent CFG G′ in time O(|G| · (n + 1)k3). Proof. We assume G to be in ETF, at the expense of a linear time factor. Each A in N is then mapped to at most (n + 1)k nonterminals (A, y) in N′ = N × Nn with Pn i=1 y(i) ≤k. Finally, for each production (A, x) →(B1, x1)(B2, x2) of P, at most (n + 1)k3 choices are possible for productions (A, y) →(B1, y1)(B2, y2) with (A, y), (B1, y1), and (B2, y2) in N′. A definition quite similar to k-rankedness can be found in the Petri net literature: 1This restriction is slightly stronger than that of linearly restricted derivations (Rambow, 1994b), but still allows to capture UVG-dl lexicalization. 516 Definition 7. A MLIG derivation α0 ⇒α1 ⇒ · · · ⇒αp is k-bounded for some k ≥0 if, no vector with a coordinate larger than k can appear in any αj, i.e. for all x in Nn such that there exist 0 ≤j ≤p, δ, δ′ in V ∗and A in N with αj = δ(A, x)δ′, and for all 1 ≤i ≤n, one has x(i) ≤k. A MLIG is k-bounded (noted kb-MLIG) if any derivation starting with α0 = (S, x0) is kbounded. It is bounded if there exists k such that it is k-bounded. The SMC in minimalist grammars translates exactly into 1-boundedness of the corresponding MLIGs (Salvati, 2010). Clearly, any k-ranked MLIG is also k-bounded, and conversely any n-dimensional k-bounded MLIG is (kn)-ranked, thus a MLIG is ranked iff it is bounded. The counterpart to Lemma 6 is: Lemma 8. Any n-dimensional k-bounded MLIG G can be transformed into an equivalent CFG G′ in time O(|G| · (k + 1)n2). Proof. We assume G to be in ETF, at the expense of a linear time factor. Each A in N is then mapped to at most (k +1)n nonterminals (A, y) in N′ = N × {0, . . . , k}n. Finally, for each production (A, x) →(B1, x1)(B2, x2) of P, each nonterminal (A, y) of N′ with x ≤y, and each index 0 < i ≤n, there are at most k + 1 ways to split (y(i) −x(i)) ≤k into y1(i) + y2(i) and span a production (A, y) →(B1, x1 + y1)(B2, x2 + y2) of P ′. Overall, each production is mapped to at most (k + 1)n2 context-free productions. One can check that the grammar of Example 2 is not bounded (to see this, repeatedly apply production (S, 0) →(S, 1)), as expected since MIX is not a context-free language. 2.3 Language Properties Let us mention a few more results pertaining to MLIG languages: Proposition 9 (Rambow, 1994b). L(MLIG) is a substitution closed full abstract family of languages. Proposition 10 (Rambow, 1994b). L(MLIGℓ) is a subset of the context-sensitive languages. Natural languages are known for displaying some limited cross-serial dependencies, as witnessed in linguistic analyses, e.g. of SwissGerman (Shieber, 1985), Dutch (Kroch and Santorini, 1991), or Tagalog (Maclachlan and Rambow, 2002). This includes the copy language Lcopy = {ww | w ∈{a, b}∗} , which does not seem to be generated by any MLIG: Conjecture 11 (Rambow, 1994b). Lcopy is not in L(MLIG). Finally, we obtain the following result as a consequence of Lemmas 6 and 8: Corollary 12. L(kr-MLIG) = L(kb-MLIG) = L(kb-MLIGℓ) is the set of context-free languages. 3 Related Formalisms We review formalisms connected to MLIGs, starting in Section 3.1 with Petri nets and two of their extensions, which turn out to be exactly equivalent to MLIGs. We then consider various linguistic formalisms that employ dominance links (Section 3.2). 3.1 Petri Nets Definition 13 (Petri, 1962). A marked Petri net2 is a tuple N = ⟨S, T, f, m0⟩where S and T are disjoint finite sets of places and transitions, f a flow function from (S × T) ∪(T × S) to N, and m0 an initial marking in NS. A transition t ∈T can be fired in a marking m in NS if f(p, t) ≥ m(p) for all p ∈S, and reaches a new marking m′ defined by m′(p) = m(p) −f(p, t) + f(t, p) for all p ∈S, written m [t⟩m′. Another view is that place p holds m(p) tokens, f(p, t) of which are first removed when firing t, and then f(t, p) added back. Firings are extended to sequences σ in T ∗by m [ε⟩m, and m [σt⟩m′ if there exists m′′ with m [σ⟩m′′ [t⟩m′. A labeled Petri net with reachability acceptance is endowed with a labeling homomorphism ϕ : T ∗→Σ∗and a finite acceptance set F ⊆NS, defining the language (Peterson, 1981) L(N, ϕ, F) = {ϕ(σ) ∈Σ∗| ∃m ∈F, m0 [σ⟩m} . Labeled Petri nets (with acceptance set {0}) are notational variants of right linear MLIGs, defined as having production in (N×Nn)×(Σ∗∪(Σ∗·(N× Nn))). This is is case of the MLIG of Example 2, which is given in Petri net form in Figure 3, where 2Petri nets are also equivalent to vector addition system (Karp and Miller, 1969, VAS) and vector addition systems with states (Hopcroft and Pansiot, 1979, VASS). 517 S e1 e2 e3 a b c ε ε Figure 3: The labeled Petri net corresponding to the right linear MLIG of Example 2. circles depict places (representing MLIG nonterminals and indices) with black dots for initial tokens (representing the MLIG start symbol), boxes transitions (representing MLIG productions), and arcs the flow values. For instance, production (S,e3) →c (S,0) is represented by the rightmost, c-labeled transition, with f(S, t) = f(e3, t) = f(t, S) = 1 and f(e1, t) = f(e2, t) = f(t, e1) = f(t, e2) = f(t, e3) = 0. Extensions The subsumption of Petri nets is not innocuous, as it allows to derive lower bounds on the computational complexity of MLIGs. Among several extensions of Petri net with some branching capacity (see e.g. Mayr, 1999; Haddad and Poitrenaud, 2007), two are of singular importance: It turns out that MLIGs in their full generality have since been independently rediscovered under the names vector addition tree automata (de Groote et al., 2004, VATA) and branching VASS (Verma and Goubault-Larrecq, 2005, BVASS). Semilinearity Another interesting consequence of the subsumption of Petri nets by MLIGs is that the former generate some non semilinear languages, i.e. with a Parikh image which is not a semilinear subset of N|Σ| (Parikh, 1966). Hopcroft and Pansiot (1979, Lemma 2.8) exhibit an example of a VASS with a non semilinear reachability set, which we translate as a 2-dimensional right linear MLIG with productions3 (S, e2) →(S, e1), (S, 0) →(A, 0) | (B, 0), (A, e1) →(A, 2e2), (A, 0) →a (S, 0), (B, e1) →b (B, 0) | b, (B, e2) →b (B, 0) | b 3Adding terminal symbols c in each production would result in a lexicalized grammar, still with a non semilinear language. S ε S S S a S S b S S c S Figure 4: An UVG-dl for Lmix. and (S, e2) as start symbol, that generates the non semilinear language Lnsm = {anbm | 0 ≤n, 0 < m ≤2n} . Proposition 14 (Hopcroft and Pansiot, 1979). There exist non semilinear Petri nets languages. The non semilinearity of MLIGs entails that of all the grammatical formalisms mentioned next in Section 3.2; this answers in particular a conjecture by Kallmeyer (2001) about the semilinearity of VTAGs. 3.2 Dominance Links UVG-dl Rambow (1994b) introduced UVG-dls as a formal model for scrambling and tree description grammars. Definition 15 (Rambow, 1994b). An unordered vector grammars with dominance links (UVG-dl) is a tuple G = ⟨N, Σ, W, S⟩where N and Σ are disjoint finite sets of nonterminals and terminals, V = N ∪Σ is the vocabulary, W is a set of vectors of productions with dominance links, i.e. each element of W is a pair (P, D) where each P is a multiset of productions in N × V ∗and D is a relation from nonterminals in the right parts of productions in P to nonterminals in their left parts, and S in N is the start symbol. A terminal derivation of w in Σ∗in an UVG-dl is a context-free derivation of form S p1 =⇒α1 p2 =⇒ α2 · · · αp−1 pp =⇒w such that the control word p1p2 · · · pp is a permutation of a member of W ∗ and the dominance relations of W hold in the associated derivation tree. The language L(G) of an UVG-dl G is the set of sentences w with some terminal derivation. We write L(UVG-dl) for the class of UVG-dl languages. An alternative semantics of derivations in UVGdls is simply their translation into MLIGs: associate with each nonterminal in a derivation the multiset of productions it has to spawn. Figure 4 presents the two vectors of an UVG-dl for the MIX language of Example 2, with dashed arrows indicating dominance links. Observe that production 518 S →S in the second vector has to spawn eventually one occurrence of each S →aS, S →bS, and S →cS, which corresponds exactly to the MLIG of Example 2. The ease of translation from the grammar of Figure 4 into a MLIG stems from the impossibility of splitting any of its vectors (P, D) into two nonempty ones (P1, D1) and (P2, D2) while preserving the dominance relation, i.e. with P = P1⊎P2 and D = D1⊎D2. This strictness property can be enforced without loss of generality since we can always add to each vector (P, D) a production S →S with a dominance link to each production in P. This was performed on the second vector in Figure 4; remark that the grammar without this addition is an unordered vector grammar (Cremers and Mayer, 1974, UVG), and still generates Lmix. Theorem 16 (Rambow, 1994b). Every MLIG can be transformed into an equivalent UVG-dl in logarithmic space, and conversely. Proof sketch. One can check that Rambow (1994b)’s proof of L(MLIG) ⊆ L(UVG-dl) incurs at most a quadratic blowup from a MLIG in RINF, and invoke Proposition 3. More precisely, given a MLIG in RINF, productions of form (A,0) →α with A in N and α in (Σ ∪(N × {0}))∗form singleton vectors, and productions of form (A,0) →(B,ei) with A, B in N and 0 < i ≤n need to be paired with a production of form (C,ei) →(D,0) for some C and D in N in order to form a vector with a dominance link between B and C. The converse inclusion and its complexity are immediate when considering strict UVG-dls. The restrictions to k-ranked and k-bounded grammars find natural counterparts in strict UVGdls by bounding the (total) number of pending dominance links in any derivation. Lexicalization has now its usual definition: for every vector ({pi,1, . . . , pi,ki}, Di) in W, at least one of the pi,j should contain at least one terminal in its right part—we have then L(UVG-dlℓ) ⊆L(MLIGℓ). More on Dominance Links Dominance links are quite common in tree description formalisms, where they were already in use in D-theory (Marcus et al., 1983) and in quasi-tree semantics for fbTAGs (Vijay-Shanker, 1992). In particular, D-tree substitution grammars are essentially the same as UVG-dls (Rambow et al., 2001), and quite a few other tree description formalisms subsume them (Candito and Kahane, 1998; Kallmeyer, 2001; Guillaume and Perrier, 2010). Another class of grammars are vector TAGs (V-TAGs), which extend TAGs and MCTAGs using dominance links (Becker et al., 1991; Rambow, 1994a; Champollion, 2007), subsuming again UVG-dls. 4 Computational Complexity We study in this section the complexity of several decision problems on MLIGs, prominently of emptiness and membership problems, in the general (Section 4.2), k-bounded (Section 4.3), and lexicalized cases (Section 4.4). Table 1 sums up the known complexity results. Since by Theorem 16 we can translate between MLIGs and UVG-dls in logarithmic space, the complexity results on UVG-dls will be the same. 4.1 Decision Problems Let us first review some decision problems of interest. In the following, G denotes a MLIG ⟨N, Σ, P, (S, x0)⟩: boundedness given ⟨G⟩, is G bounded? As seen in Section 2.2, this is equivalent to rankedness. k-boundedness given ⟨G, k⟩, k in N, is G kbounded? As seen in Section 2.2, this is the same as (kn)-rankedness. Here we will distinguish two cases depending on whether k is encoded in unary or binary. coverability given ⟨G, F⟩, G ε-free in ETF and F a finite subset of N×Nn, does there exist α = (A1, y1) · · · (Am, ym) in (N ×Nn)∗such that (S, x0) ⇒∗α and for each 0 < j ≤m there exists (Aj, xj) in F with xj ≤yj? reachability given ⟨G, F⟩, G ε-free in ETF and F a finite subset of N × Nn, does there exist α = (A1, y1) · · · (Am, ym) in F ∗such that (S, x0) ⇒∗α? non emptiness given ⟨G⟩, is L(G) non empty? (uniform) membership given ⟨G, w⟩, w in Σ∗, does w belong to L(G)? Boundedness and k-boundedness are needed in order to prove that a grammar is bounded, and to apply the smaller complexities of Section 4.3. Coverability is often considered for Petri nets, and allows to derive lower bounds on reachability. Emptiness is the most basic static 519 analysis one might want to perform on a grammar, and is needed for parsing as intersection approaches (Lang, 1994), while membership reduces to parsing. Note that we only consider uniform membership, since grammars for natural languages are typically considerably larger than input sentences, and their influence can hardly be neglected. There are several obvious reductions between reachability, emptiness, and membership. Let →log denote LOGSPACE reductions between decision problems; we have: Proposition 17. coverability →log reachability (1) ↔log non emptiness (2) ↔log membership (3) Proof sketch. For (1), construct a reachability instance ⟨G′, {(E, 0)}⟩from a coverability instance ⟨G, F⟩by adding to G a fresh nonterminal E and the productions {(A, x) →(E, 0) | (A, x) ∈F} ∪{(E, ei) →(E, 0) | 0 < i ≤n} . For (2), from a reachability instance ⟨G, F⟩, remove all terminal productions from G and add instead the productions {(A, x) →ε | (A, x) ∈F}; the new grammar G′ has a non empty language iff the reachability instance was positive. Conversely, from a non emptiness instance ⟨G⟩, put the grammar in ETF and define F to match all terminal productions, i.e. F = {(A, x) | (A, x) →a ∈P, a ∈ Σ∪{ε}}, and then remove all terminal productions in order to obtain a reachability instance ⟨G′, F⟩. For (3), from a non emptiness instance ⟨G⟩, replace all terminals in G by ε to obtain an empty word membership instance ⟨G′, ε⟩. Conversely, from a membership instance ⟨G, w⟩, construct the intersection grammar G′ with L(G′) = L(G)∩{w} (Bar-Hillel et al., 1961), which serves as non emptiness instance ⟨G′⟩. 4.2 General Case Verma and Goubault-Larrecq (2005) were the first to prove that coverability and boundedness were decidable for BVASS, using a covering tree construction `a la Karp and Miller (1969), thus of non primitive recursive complexity. Demri et al. (2009, Theorems 7, 17, and 18) recently proved tight complexity bounds for these problems, extending earlier results by Rackoff (1978) and Lipton (1976) for Petri nets. Theorem 18 (Demri et al., 2009). Coverability and boundedness for MLIGs are 2EXPTIMEcomplete. Regarding reachability, emptiness, and membership, decidability is still open. A 2EXPSPACE lower bound was recently found by Lazi´c (2010). If a decision procedure exists, we can expect it to be quite complex, as already in the Petri net case, the complexity of the known decision procedures (Mayr, 1981; Kosaraju, 1982) is not primitive recursive (Cardoza et al., 1976, who attribute the idea to Hack). 4.3 k-Bounded and k-Ranked Cases Since k-bounded MLIGs can be converted into CFGs (Lemma 8), emptiness and membership problems are decidable, albeit at the expense of an exponential blowup. We know from the Petri net literature that coverability and reachability problems are PSPACE-complete for k-bounded right linear MLIGs (Jones et al., 1977) by a reduction from linear bounded automaton (LBA) membership. We obtain the following for k-bounded MLIGs, using a similar reduction from membership in polynomially space bounded alternating Turing machines (Chandra et al., 1981, ATM): Theorem 19. Coverability and reachability for kbounded MLIGs are EXPTIME-complete, even for fixed k ≥1. The lower bound is obtained through an encoding of an instance of the membership problem for ATMs working in polynomial space into an instance of the coverability problem for 1-bounded MLIGs. The upper bound is a direct application of Lemma 8, coverability and reachability being reducible to the emptiness problem for a CFG of exponential size. Theorem 19 also shows the EXPTIME-hardness of emptiness and membership in minimalist grammars with SMC. Corollary 20. Let k ≥1; k-boundedness for MLIGs is EXPTIME-complete. Proof. For the lower bound, consider an instance ⟨G, F⟩of coverability for a 1-bounded MLIG G, which is EXPTIME-hard according to Theorem 19. Add to the MLIG G a fresh nonterminal E and the productions {(A, x) →(E, x) | (A, x) ∈F} ∪{(E, 0) →(E, ei) | 0 < i ≤n} , which make it non k-bounded iff the coverability instance was positive. 520 Problem Lower bound Upper bound Petri net k-Boundedness PSPACE (Jones et al., 1977) PSPACE (Jones et al., 1977) Petri net Boundedness EXPSPACE (Lipton, 1976) EXPSPACE (Rackoff, 1978) Petri net {Emptiness, Membership} EXPSPACE (Lipton, 1976) Decidable, not primitive recursive (Mayr, 1981; Kosaraju, 1982) {MLIG, MLIGℓ} k-Boundedness EXPTIME (Corollary 20) EXPTIME (Corollary 20) {MLIG, MLIGℓ} Boundedness 2EXPTIME (Demri et al., 2009) 2EXPTIME (Demri et al., 2009) {MLIG, MLIGℓ} Emptiness 2EXPSPACE (Lazi´c, 2010) Not known to be decidable MLIG Membership {kb-MLIG, kb-MLIGℓ} Emptiness EXPTIME (Theorem 19) EXPTIME (Theorem 19) kb-MLIG Membership {MLIGℓ, kb-MLIGℓ} Membership NPTIME (Koller and Rambow, 2007) NPTIME (trivial) kr-MLIG {Emptiness, Membership} PTIME (Jones and Laaser, 1976) PTIME (Lemma 6) Table 1: Summary of complexity results. For the upper bound, apply Lemma 8 with k′ = k + 1 to construct an O(|G| · 2n2 log2(k′+1))-sized CFG, reduce it in polynomial time, and check whether a nonterminal (A, x) with x(i) = k′ for some 0 < i ≤n occurs in the reduced grammar. Note that the choice of the encoding of k is irrelevant, as k = 1 is enough for the lower bound, and k only logarithmically influences the exponent for the upper bound. Corollary 20 also implies the EXPTIMEcompleteness of k-rankedness, k encoded in unary, if k can take arbitrary values. On the other hand, if k is known to be small, for instance logarithmic in the size of G, then k-rankedness becomes polynomial by Lemma 6. Observe finally that k-rankedness provides the only tractable class of MLIGs for uniform membership, using again Lemma 6 to obtain a CFG of polynomial size—actually exponential in k, but k is assumed to be fixed for this problem. An obvious lower bound is that of membership in CFGs, which is PTIME-complete (Jones and Laaser, 1976). 4.4 Lexicalized Case Unlike the high complexity lower bounds of the previous two sections, NPTIME-hardness results for uniform membership have been proved for a number of formalisms related to MLIGs, from the commutative CFG viewpoint (Huynh, 1983; Barton, 1985; Esparza, 1995), or from more specialized models (Søgaard et al., 2007; Champollion, 2007; Koller and Rambow, 2007). We focus here on this last proof, which reduces from the normal dominance graph configurability problem (Althaus et al., 2003), as it allows to derive NPTIME-hardness even in highly restricted grammars. Theorem 21 (Koller and Rambow, 2007). Uniform membership of ⟨G, w⟩for G a 1-bounded, lexicalized, UVG-dl with finite language is NPTIME-hard, even for |w| = 1. Proof sketch. Set S as start symbol and add a production S →aA to the sole vector of the grammar G constructed by Koller and Rambow (2007) from a normal dominance graph, with dominance links to all the other productions. Then G becomes strict, lexicalized, with finite language {a} or ∅, and 1-bounded, such that a belongs to L(G) iff the normal dominance graph is configurable. The fact that uniform membership is in NPTIME in the lexicalized case is clear, as we only need to guess nondeterministically a derivation of size linear in |w| and check its correctness. The weakness of lexicalized grammars is however that their emptiness problem is not any easier to solve! The effect of lexicalization is indeed to break the reduction from emptiness to membership in Proposition 17, but emptiness is as hard as ever, which means that static checks on the grammar might even be undecidable. 5 Conclusion Grammatical formalisms with dominance links, introduced in particular to model scrambling phenomena in computational linguistics, have deep connections with several open questions in an unexpected variety of fields in computer science. We hope this survey to foster cross-fertilizing exchanges; for instance, is there a relation between 521 Conjecture 11 and the decidability of reachability in MLIGs? A similar question, whether the language Lpal of even 2-letters palindromes was a Petri net language, was indeed solved using the decidability of reachability in Petri nets (Jantzen, 1979), and shown to be strongly related to the latter (Lambert, 1992). A conclusion with a more immediate linguistic value is that MLIGs and UVG-dls hardly qualify as formalisms for mildly context-sensitive languages, claimed by Joshi (1985) to be adequate for modeling natural languages, and “roughly” defined as the extensions of context-free languages that display 1. support for limited cross-serial dependencies: seems doubtful, see Conjecture 11, 2. constant growth, a requisite nowadays replaced by semilinearity: does not hold, as seen with Proposition 14, and 3. polynomial recognition algorithms: holds only for restricted classes of grammars, as seen in Section 4. Nevertheless, variants such as k-ranked V-TAGs are easily seen to fulfill all the three points above. Acknowledgements Thanks to Pierre Chambart, St´ephane Demri, and Alain Finkel for helpful discussions, and to Sylvain Salvati for pointing out the relation with minimalist grammars. References Ernst Althaus, Denys Duchier, Alexander Koller, Kurt Mehlhorn, Joachim Niehren, and Sven Thiel. 2003. An efficient graph algorithm for dominance constraints. Journal of Algorithms, 48(1):194–219. Yehoshua Bar-Hillel, Micha Perles, and Eliahu Shamir. 1961. On formal properties of simple phrase structure grammars. Zeitschrift f¨ur Phonetik, Sprachwissenschaft und Kommunikationsforschung, 14:143– 172. G. Edward Barton. 1985. The computational difficulty of ID/LP parsing. In ACL’85, pages 76–81. ACL Press. Tilman Becker, Aravind K. Joshi, and Owen Rambow. 1991. Long-distance scrambling and tree adjoining grammars. In EACL’91, pages 21–26. ACL Press. Mikołaj Boja´nczyk, Anca Muscholl, Thomas Schwentick, and Luc Segoufin. 2009. Twovariable logic on data trees and XML reasoning. Journal of the ACM, 56(3):1–48. Marie-H´el`ene Candito and Sylvain Kahane. 1998. Defining DTG derivations to get semantic graphs. In TAG+4, pages 25–28. E. Cardoza, Richard J. Lipton, and Albert R. Meyer. 1976. Exponential space complete problems for Petri nets and commutative semigroups: Preliminary report. In STOC’76, pages 50–54. ACM Press. Lucas Champollion. 2007. Lexicalized non-local MCTAG with dominance links is NP-complete. In MOL 10. Ashok K. Chandra, Dexter C. Kozen, and Larry J. Stockmeyer. 1981. Alternation. Journal of the ACM, 28(1):114–133. David Chiang and Tatjana Scheffler. 2008. Flexible composition and delayed tree-locality. In TAG+9. Armin B. Cremers and Otto Mayer. 1974. On vector languages. Journal of Computer and System Sciences, 8(2):158–166. Philippe de Groote, Bruno Guillaume, and Sylvain Salvati. 2004. Vector addition tree automata. In LICS’04, pages 64–73. IEEE Computer Society. St´ephane Demri, Marcin Jurdzi´nski, Oded Lachish, and Ranko Lazi´c. 2009. The covering and boundedness problems for branching vector addition systems. In Ravi Kannan and K. Narayan Kumar, editors, FSTTCS’09, volume 4 of Leibniz International Proceedings in Informatics, pages 181–192. Schloss Dagstuhl–Leibniz-Zentrum f¨ur Informatik. Catherine Dufourd and Alain Finkel. 1999. A polynomial λ-bisimilar normalization for reset Petri nets. Theoretical Computer Science, 222(1–2):187–194. Javier Esparza. 1995. Petri nets, commutative contextfree grammars, and basic parallel processes. In Horst Reichel, editor, FCT’95, volume 965 of Lecture Notes in Computer Science, pages 221–232. Springer. Bruno Guillaume and Guy Perrier. 2010. Interaction grammars. Research on Language and Computation. To appear. Serge Haddad and Denis Poitrenaud. 2007. Recursive Petri nets. Acta Informatica, 44(7–8):463–508. John Hopcroft and Jean-Jacques Pansiot. 1979. On the reachability problem for 5-dimensional vector addition systems. Theoretical Computer Science, 8(2):135–159. Dung T. Huynh. 1983. Commutative grammars: the complexity of uniform word problems. Information and Control, 57(1):21–39. Matthias Jantzen. 1979. On the hierarchy of Petri net languages. RAIRO Theoretical Informatics and Applications, 13(1):19–30. 522 Neil D. Jones and William T. Laaser. 1976. Complete problems for deterministic polynomial time. Theoretical Computer Science, 3(1):105–117. Neil D. Jones, Lawrence H. Landweber, and Y. Edmund Lien. 1977. Complexity of some problems in Petri nets. Theoretical Computer Science, 4(3):277– 299. Aravind K. Joshi, Tilman Becker, and Owen Rambow. 2000. Complexity of scrambling: A new twist to the competence-performance distinction. In Anne Abeill´e and Owen Rambow, editors, Tree Adjoining Grammars. Formalisms, Linguistic Analysis and Processing, chapter 6, pages 167–181. CSLI Publications. Aravind K. Joshi. 1985. Tree-adjoining grammars: How much context sensitivity is required to provide reasonable structural descriptions? In David R. Dowty, Lauri Karttunen, and Arnold M. Zwicky, editors, Natural Language Parsing: Psychological, Computational, and Theoretical Perspectives, chapter 6, pages 206–250. Cambridge University Press. Laura Kallmeyer and Yannick Parmentier. 2008. On the relation between multicomponent tree adjoining grammars with tree tuples (TT-MCTAG) and range concatenation grammars (RCG). In Carlos Mart´ınVide, Friedrich Otto, and Henning Fernau, editors, LATA’08, volume 5196 of Lecture Notes in Computer Science, pages 263–274. Springer. Laura Kallmeyer. 2001. Local tree description grammars. Grammars, 4(2):85–137. Richard M. Karp and Raymond E. Miller. 1969. Parallel program schemata. Journal of Computer and System Sciences, 3(2):147–195. Alexander Koller and Owen Rambow. 2007. Relating dominance formalisms. In FG’07. S. Rao Kosaraju. 1982. Decidability of reachability in vector addition systems. In STOC’82, pages 267– 281. ACM Press. Anthony S. Kroch and Beatrice Santorini. 1991. The derived constituent structure of the West Germanic verb-raising construction. In Robert Freidin, editor, Principles and Parameters in Comparative Grammar, chapter 10, pages 269–338. MIT Press. Jean-Luc Lambert. 1992. A structure to decide reachability in Petri nets. Theoretical Computer Science, 99(1):79–104. Bernard Lang. 1994. Recognition can be harder than parsing. Computational Intelligence, 10(4):486– 494. Ranko Lazi´c. 2010. The reachability problem for branching vector addition systems requires doublyexponential space. Manuscript. Timm Lichte. 2007. An MCTAG with tuples for coherent constructions in German. In FG’07. Richard Lipton. 1976. The reachability problem requires exponential space. Technical Report 62, Yale University. Anna Maclachlan and Owen Rambow. 2002. Crossserial dependencies in Tagalog. In TAG+6, pages 100–107. Mitchell P. Marcus, Donald Hindle, and Margaret M. Fleck. 1983. D-theory: talking about talking about trees. In ACL’83, pages 129–136. ACL Press. Ernst W. Mayr. 1981. An algorithm for the general Petri net reachability problem. In STOC’81, pages 238–246. ACM Press. Richard Mayr. 1999. Process rewrite systems. Information and Computation, 156(1–2):264–286. Christos H. Papadimitriou. 1994. Computational Complexity. Addison-Wesley. Rohit J. Parikh. 1966. On context-free languages. Journal of the ACM, 13(4):570–581. James L. Peterson. 1981. Petri Net Theory and the Modeling of Systems. Prentice Hall. Carl A. Petri. 1962. Kommunikation mit Automaten. Ph.D. thesis, University of Bonn. Charles Rackoff. 1978. The covering and boundedness problems for vector addition systems. Theoretical Computer Science, 6(2):223–231. Owen Rambow, K. Vijay-Shanker, and David Weir. 1995. D-tree grammars. In ACL’95, pages 151–158. ACL Press. Owen Rambow, David Weir, and K. Vijay-Shanker. 2001. D-tree substitution grammars. Computational Linguistics, 27(1):89–121. Owen Rambow. 1994a. Formal and Computational Aspects of Natural Language Syntax. Ph.D. thesis, University of Pennsylvania. Owen Rambow. 1994b. Multiset-valued linear index grammars: imposing dominance constraints on derivations. In ACL’94, pages 263–270. ACL Press. Sylvain Salvati. 2010. Minimalist grammars in the light of logic. Manuscript. Stuart M. Shieber. 1985. Evidence against the contextfreeness of natural language. Linguistics and Philosophy, 8(3):333–343. Anders Søgaard, Timm Lichte, and Wolfgang Maier. 2007. The complexity of linguistically motivated extensions of tree-adjoining grammar. In RANLP’07, pages 548–553. Edward P. Stabler. 1997. Derivational minimalism. In Christian Retor´e, editor, LACL’96, volume 1328 of Lecture Notes in Computer Science, pages 68–95. Springer. 523 Kumar Neeraj Verma and Jean Goubault-Larrecq. 2005. Karp-Miller trees for a branching extension of VASS. Discrete Mathematics and Theoretical Computer Science, 7(1):217–230. K. Vijay-Shanker. 1992. Using descriptions of trees in a tree adjoining grammar. Computational Linguistics, 18(4):481–517. Ryo Yoshinaka and Makoto Kanazawa. 2005. The complexity and generative capacity of lexicalized abstract categorial grammars. In Philippe Blache, Edward Stabler, Joan Busquets, and Richard Moot, editors, LACL’05, volume 3492 of Lecture Notes in Computer Science, pages 330–346. Springer. 524
|
2010
|
53
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 525–533, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Optimal rank reduction for Linear Context-Free Rewriting Systems with Fan-Out Two Benot Sagot INRIA & Universit´e Paris 7 Le Chesnay, France [email protected] Giorgio Satta Department of Information Engineering University of Padua, Italy [email protected] Abstract Linear Context-Free Rewriting Systems (LCFRSs) are a grammar formalism capable of modeling discontinuous phrases. Many parsing applications use LCFRSs where the fan-out (a measure of the discontinuity of phrases) does not exceed 2. We present an efficient algorithm for optimal reduction of the length of production right-hand side in LCFRSs with fan-out at most 2. This results in asymptotical running time improvement for known parsing algorithms for this class. 1 Introduction Linear Context-Free Rewriting Systems (LCFRSs) have been introduced by VijayShanker et al. (1987) for modeling the syntax of natural language. The formalism extends the generative capacity of context-free grammars, still remaining far below the class of context-sensitive grammars. An important feature of LCFRSs is their ability to generate discontinuous phrases. This has been recently exploited for modeling phrase structure treebanks with discontinuous constituents (Maier and Søgaard, 2008), as well as non-projective dependency treebanks (Kuhlmann and Satta, 2009). The maximum number f of tuple components that can be generated by an LCFRS G is called the fan-out of G, and the maximum number r of nonterminals in the right-hand side of a production is called the rank of G. As an example, contextfree grammars are LCFRSs with f = 1 and r given by the maximum length of a production right-hand side. Tree adjoining grammars (Joshi and Levy, 1977) can also be viewed as a special kind of LCFRS with f = 2, since each auxiliary tree generates two strings, and with r given by the maximum number of adjunction and substitution sites in an elementary tree. Beyond tree adjoining languages, LCFRSs with f = 2 can also generate languages in which pair of strings derived from different nonterminals appear in socalled crossing configurations. It has recently been observed that, in this way, LCFRSs with f = 2 can model the vast majority of data in discontinuous phrase structure treebanks and non-projective dependency treebanks (Maier and Lichte, 2009; Kuhlmann and Satta, 2009). Under a theoretical perspective, the parsing problem for LCFRSs with f = 2 is NP-complete (Satta, 1992), and in known parsing algorithms the running time is exponentially affected by the rank r of the grammar. Nonetheless, in natural language parsing applications, it is possible to achieve efficient, polynomial parsing if we succeed in reducing the rank r (number of nonterminals in the right-hand side) of individual LCFRSs’ productions (Kuhlmann and Satta, 2009). This process is called production factorization. Production factorization is very similar to the reduction of a context-free grammar production into Chomsky normal form. However, in the LCFRS case some productions might not be reducible to r = 2, and the process stops at some larger value for r, which in the worst case might as well be the rank of the source production (Rambow and Satta, 1999). Motivated by parsing efficiency, the factorization problem for LCFRSs with f = 2 has attracted the attention of many researchers in recent years. Most of the literature has been focusing on binarization algorithms, which attempt to find a reduction to r = 2 and return a failure if this is not possible. G´omez-Rodr´ıguez et al. (2009) report a general binarization algorithm for LCFRS which, in the case of f = 2, works in time O(|p|7), where |p| is the size of the input production. A more efficient binarization algorithm for the case f = 2 is presented in (G´omez-Rodr´ıguez and Satta, 2009), working in time O(|p|). 525 In this paper we are interested in general factorization algorithms, i.e., algorithms that find factorizations with the smallest possible rank (not necessarily r = 2). We present a novel technique that solves the general factorization problem in time O(|p|2) for LCFRSs with f = 2. Strong generative equivalence results between LCFRS and other finite copying parallel rewriting systems have been discussed in (Weir, 1992) and in (Rambow and Satta, 1999). Through these equivalence results, we can transfer the factorization techniques presented in this article to other finite copying parallel rewriting systems. 2 LCFRSs In this section we introduce the basic notation for LCFRS and the notion of production factorization. 2.1 Definitions Let ΣT be a finite alphabet of terminal symbols. As usual, Σ ∗ T denotes the set of all finite strings over ΣT , including the empty string ε. For integer k ≥1, (Σ ∗ T )k denotes the set of all tuples (w1, . . . , wk) of strings wi ∈Σ ∗ T . In what follows we are interested in functions mapping several tuples of strings in Σ ∗ T into tuples of strings in Σ ∗ T . Let r and f be two integers, r ≥0 and f ≥1. We say that a function g has rank r if there exist integers fi ≥1, 1 ≤i ≤r, such that g is defined on (Σ ∗ T )f1 × (Σ ∗ T )f2 × · · · × (Σ ∗ T )fr. We also say that g has fan-out f if the range of g is a subset of (Σ ∗ T )f. Let yh, xij, 1 ≤h ≤f, 1 ≤i ≤r and 1 ≤j ≤fi, be string-valued variables. A function g as above is said to be linear regular if it is defined by an equation of the form g(⟨x11, . . . , x1f1⟩, . . . , ⟨xr1, . . . , xrfr⟩) = = ⟨y1, . . . , yf⟩, (1) where ⟨y1, . . . , yf⟩represents some grouping into f sequences of all and only the variables appearing in the left-hand side of (1) (without repetitions) along with some additional terminal symbols (with possible repetitions). For a mathematical definition of LCFRS we refer the reader to (Weir, 1992, p. 137). Informally, in a LCFRS every nonterminal symbol A is associated with an integer ϕ(A) ≥1, called its fan-out, and it generates tuples in (Σ ∗ T )ϕ(A). Productions in a LCFRS have the form p : A →g(B1, B2, . . . , Bρ(p)), where ρ(p) ≥0, A and Bi, 1 ≤i ≤ρ(p), are nonterminal symbols, and g is a linear regular function having rank ρ(p) and fan-out ϕ(A), defined on (Σ ∗ T )ϕ(B1) ×· · ·×(Σ ∗ T )ϕ(Bρ(p)) and taking values in (Σ ∗ T )ϕ(A). The basic idea underlying the rewriting relation associated with LCFRS is that production p applies to any sequence of string tuples generated by the Bi’s, and provides a new string tuple in (Σ ∗ T )ϕ(A) obtained through function g. We say that ϕ(p) = ϕ(A) is the fan-out of p, and ρ(p) is the rank of p. Example 1 Let L be the language L = {anbnambmanbnambm | n, m ≥1}. A LCFRS generating L is defined by means of the nonterminals S, ϕ(S) = 1, and A, ϕ(A) = 2, and the productions in figure 1. Observe that nonterminal A generates all tuples of the form ⟨anbn, anbn⟩. 2 Recognition and parsing for a given LCFRS can be carried out in polynomial time on the length of the input string. This is usually done by exploiting standard dynamic programming techniques; see for instance (Seki et al., 1991).1 However, the polynomial degree in the running time is a monotonically strictly increasing function that depends on both the rank and the fan-out of the productions in the grammar. To optimize running time, one can then recast the source grammar in such a way that the value of the above function is kept to a minimum. One way to achieve this is by factorizing the productions of a LCFRS, as we now explain. 2.2 Factorization Consider a LCFRS production of the form p : A →g(B1, B2, . . . , Bρ(p)), where g is specified as in (1). Let also C be a subset of {B1, B2, . . . , Bρ(p)} such that |C| ̸= 0 and |C| ̸= ρ(p). We let ΣC be the alphabet of all variables xij defined as in (1), for all values of i and j such that Bi ∈C and 1 ≤j ≤fi. For each i with 1 ≤i ≤f, we rewrite each string yi in (1) in a form yi = y′ i0zi1y′ i1 · · · y′ idi−1zidiy′ idi, with di ≥0, such that the following conditions are all met: • each zij, 1 ≤j ≤di, is a string with one or more occurrences of variables, all in ΣC; • each y′ ij, 1 ≤j ≤di −1, is a non-empty string with no occurrences of symbols in ΣC; • y′ 0j and y′ 0di are (possibly empty) strings with no occurrences of symbols in ΣC. 1In (Seki et al., 1991) a syntactic variant of LCFRS is used, called multiple context-free grammars. 526 S →gS(A, A), gS(⟨x11, x12⟩, ⟨x21, x22⟩) = ⟨x11x21x12x22⟩; A →gA(A), gA(⟨x11, x12⟩) = ⟨ax11b, ax12b⟩; A →g′ A(), g′ A() = ⟨ab, ab⟩. Figure 1: A LCFRS for language L = {anbnambmanbnambm | n, m ≥1}. Let c = |C| and c = ρ(p) −|C|. Assume that C = {Bh1, . . . , Bhc}, and {B1, . . . , Bρ(p)} −C = {Bh′ 1, . . . , Bh′ c}. We introduce a fresh nonterminal C with ϕ(C) = Pf i=1 di and replace production p in our grammar by means of the two new productions p1 : C →g1(Bh1, . . . , Bhc) and p2 : A →g2(C, Bh′ 1, . . . , Bh′ c). Functions g1 and g2 are defined as: g1(⟨xh11, . . . , xh1fh1⟩, . . . , ⟨xhc1, . . . , xhcfhc⟩) = ⟨z11, · · · , z1d1, z21, · · · , zfdf ⟩; g2(⟨xh′ 11, . . . , xh′ 1fh′ 1⟩, . . . , ⟨xh′ c1, . . . , xh′ cfh′ c⟩) = ⟨y′ 10, . . . , y′ 1d1, y′ 20, . . . , y′ fdf ⟩. Note that productions p1 and p2 have rank strictly smaller than the source production p. Furthermore, if it is possible to choose set C in such a way that Pf i=0 di ≤f, then the fan-out of p1 and p2 will be no greater than the fan-out of p. We can iterate the procedure above as many times as possible, under the condition that the fanout of the productions does not increase. Example 2 Let us consider the following production with rank 4: A →gS(B, C, D, E), gA(⟨x11, x12⟩, ⟨x21, x22⟩, ⟨x31, x32⟩, ⟨x41, x42⟩) = ⟨x11x21x31x41x12x42, x22x32⟩. Applyng the above procedure twice, we obtain a factorization consisting of three productions with rank 2 (variables have been renamed to reflect our conventions): A →gA(A1, A2), gA(⟨x11, x12⟩, ⟨x21, x22⟩) = ⟨x11x21x12, x22⟩; A1 →gA1(B, E), gA1(⟨x11, x12⟩, ⟨x21, x22⟩) = ⟨x11, x21x12x22⟩; A2 →gA2(C, D), gA2(⟨x11, x12⟩, ⟨x21, x22⟩) = ⟨x11x21, x12x22⟩. 2 The factorization procedure above should be applied to all productions of a LCFRS with rank larger than two. This might result in an asymptotic improvement of the running time of existing dynamic programming algorithms for parsing based on LCFRS. The factorization technique we have discussed can also be viewed as a generalization of wellknown techniques for casting context-free grammars into binary forms. These are forms where no more than two nonterminal symbols are found in the right-hand side of productions of the grammar; see for instance (Harrison, 1978). One important difference is that, while production factorization into binary form is always possible in the contextfree case, for LCFRS there are worst case grammars in which rank reduction is not possible at all, as shown in (Rambow and Satta, 1999). 3 A graph-based representation for LCFRS productions Rather than factorizing LCFRS productions directly, in this article we work with a more abstract representation of productions based on graphs. From now on we focus on LCFRS whose nonterminals and productions all have fan-out smaller than or equal to 2. Consider then a production p : A →g(B1, B2, . . . , Bρ(p)), with ϕ(A), ϕ(Bi) ≤ 2, 1 ≤i ≤ρ(p), and with g defined as g(⟨x11, . . . , x1ϕ(B1)⟩, . . . . . . , ⟨xρ(p)1, . . . , xρ(p)ϕ(Bρ(p))⟩) = ⟨y1, . . . , yϕ(A)⟩. In what follows, if ϕ(A) = 1 then ⟨y1, . . . , yϕ(A)⟩ should be read as ⟨y1⟩and y1 · · · yϕ(A) should be read as y1. The same convention applies to all other nonterminals and tuples. We now introduce a special kind of undirected graph that is associated with a linear order defined over the set of its vertices. The p-graph associated with production p is a triple (Vp, Ep, ≺p) such that • Vp = {xij | 1 ≤i ≤ρ(p), ϕ(Bi) = 2, 1 ≤ j ≤ϕ(Bi)} is a set of vertices;2 2Here we are overloading symbols xij. It will always be clear from the context whether xij is a string-valued variable or a vertex in a p-graph. 527 • Ep = {(xi1, xi2) | xi1, xi2 ∈Vp} is a set of undirected edges; • for x, x′ ∈Vp, x ≺p x′ if x ̸= x′ and the (unique) occurrence of x in y1 · · · yϕ(A) precedes the (unique) occurrence of x′. Note that in the above definition we are ignoring all string-valued variables xij associated with nonterminals Bi with ϕ(Bi) = 1. This is because nonterminals with fan-out one can always be treated as in the context-free grammar case, as it will be explained later. Example 3 The p-graph associated with the LCFRS production in Example 2 is shown in Figure 2. Circled sets of edges indicate the factorization in that example. 2 x21 x31 x41 x11 B C D E A1 A2 x42 x12 x22 x32 Figure 2: The p-graph associated with the LCFRS production in Example 2. We close this section by introducing some additional notation related to p-graphs that will be used throughout this paper. Let E ⊆Ep be some set of edges. The cover set for E is defined as V (E) = {x | (x, x′) ∈E} (recall that our edges are unordered pairs, so (x, x′) and (x′, x) denote the same edge). Conversely, let V ⊆Vp be some set of vertices. The incident set for V is defined as E(V ) = {(x, x′) | (x, x′) ∈Ep, x ∈V }. Assume ϕ(p) = 2, and let x1, x2 ∈Vp. If x1 and x2 do not occur both in the same string y1 or y2, then we say that there is a gap between x1 and x2. If x1 ≺p x2 and there is no gap between x1 and x2, then we write [x1, x2] to denote the set {x1, x2} ∪{x | x ∈Vp, x1 ≺p x ≺p x2}. For x ∈ Vp we also let [x, x] = {x}. A set [x, x′] is called a range. Let r and r′ be two ranges. The pair (r, r′) is called a tandem if the following conditions are both satisfied: (i) r∪r′ is not a range, and (ii) there exists some edge (x, x′) ∈Ep with x ∈r and x′ ∈r′. Note that the first condition means that r and r′ are disjoint sets and, for any pair of vertices x ∈r and x′ ∈r′, either there is a gap between x and x′ or else there exists some xg ∈Vp such that x ≺p xg ≺p x′ and xg ̸∈r ∪r′. A set of edges E ⊆Ep is called a bundle with fan-out one if V (E) = [x1, x2] for some x1, x2 ∈ Vp, i.e., V (E) is a range. Set E is called a bundle with fan-out two if V (E) = [x1, x2] ∪[x3, x4] for some x1, x2, x3, x4 ∈Vp, and ([x1, x2], [x3, x4]) is a tandem. Note that if E is a bundle with fan-out two with V (E) = [x1, x2] ∪[x3, x4], then neither E([x1, x2]) nor E([x3, x4]) are bundles with fanout one, since there is at least one edge incident upon a vertex in [x1, x2] and a vertex in [x3, x4]. We also use the term bundle to denote a bundle with fan-out either one or two. Intuitively, in a p-graph associated with a LCFRS production p, a bundle E with fan-out f and with |E| > 1 identifies a set of nonterminals C in the right-hand side of p that can be factorized into a new production. The nonterminals in C are then replaced in p by a fresh nonterminal C with fan-out f, as already explained. Our factorization algorithm is based on efficient methods for the detection of bundles with fan-out one and two. 4 The algorithm In this section we provide an efficient, recursive algorithm for the decomposition of a p-graph into bundles, which corresponds to factorizing the represented LCFRS production. 4.1 Overview of the algorithm The basic idea underlying our graph-based algorithm can be described as follows. We want to compute an optimal hierarchical decomposition of an input bundle with fan-out 1 or 2. This decomposition can be represented by a tree, in which each node N corresponds to a bundle (the root node corresponds to the input bundle) and the daughters of N represent the bundles in which N is immediately decomposed. The decomposition is optimal in so far as the maximum arity of the decomposition tree is as small as possible. As already explained above, this decomposition represents a factorization of some production p of a LCFRS, resulting in optimal rank reduction. All the internal nodes in the decomposition represent fresh nonterminals that will be created during the factorization process. The construction of the decomposition tree is carried out recursively. For a given bundle with fan-out 1 or 2, we apply a procedure for decomposing this bundle in its immediate sub-bundles with fan-out 1 or 2, in an optimal way. Then, 528 we recursively apply our procedure to the obtained sub-bundles. Recursion stops when we reach bundles containing only one edge (which correspond to the nonterminals in the right-hand side of the input production). We shall prove that the result is an optimal decomposition. The procedure for computing an optimal decomposition of a bundle F into its immediate subbundles, which we describe in the first part of this section, can be sketched as follows. First, we identify and temporarily remove all maximal bundles with fan-out 1 (Section 4.3). The result is a new bundle F ′ which is a subset of the original bundle, and has the same fan-out. Next, we identify all sub-bundles with fan-out 2 in F ′ (Section 4.4). We compute the optimal decomposition of F ′, resting on the hypothesis that there are no sub-bundles with fan-out 1. Each resulting sub-bundle is later expanded with the maximal sub-bundles with fanout 1 that have been previously removed. This results in a “first level” decomposition of the original bundle F. We then recursively decompose all individual sub-bundles of F, including the bundles with fan-out 1 that have been later attached. 4.2 Backward and forward quantities For a set V ⊆Vp of vertices, we write max(V ) (resp. min(V )) the maximum (resp. minimum) vertex in V w.r.t. the ≺p total order. Let r = [x1, x2] be a range. We write r.left = x1 and r.right = x2. The set of backward edges for r is defined as Br = {(x, x′) | (x, x′) ∈ Er, x ≺p r.left, x′ ∈r}. The set of forward edges for r is defined symmetrically as Fr = {(x, x′) | (x, x′) ∈Er, x ∈r, r.right ≺p x′}. For E ∈{Br, Fr} we also define L(E) = {x | (x, x′) ∈ E, x ≺p x′} and R(E) = {x′ | (x, x′) ∈E, x ≺p x′}. Let us assume Br ̸= ∅. We write r.b.left = min(L(Br)). Intuitively, r.b.left is the leftmost vertex of the p-graph that is located at the left of range r and that is connected to some vertex in r through some edge. Similarly, we write r.b.right = max(L(Br)). If Br = ∅, then we set r.b.left = r.b.right = ⊥. Quantities r.b.left and r.b.right are called backward quantities. We also introduce local backward quantities, defined as follows. We write r.lb.left = min(R(Br)). Intuitively, r.lb.left is the leftmost vertex among all those vertices in r that are connected to some vertex to the left of r. Similarly, we write r.lb.right = max(R(Br)). If Br = ∅, then we set r.lb.left = r.lb.right = ⊥. We define forward and local forward quantities in a symmetrical way. The backward quantities r.b.left and r.b.right and the local backward quantities r.lb.left and r.lb.right for all ranges r in the p-graph can be computed efficiently as follows. We process ranges in increasing order of size, expanding each range r by one unit at a time by adding a new vertex at its right. Backward and local backward quantities for the expanded range can be expressed as a function of the same quantities for r. Therefore if we store our quantities for previously processed ranges, each new range can be annotated with the desired quantities in constant time. This algorithm runs in time O(n2), where n is the number of vertices in Vp. This is an optimal result, since O(n2) is also the size of the output. We compute in a similar way the forward quantities r.f .left and r.f .right and the local forward quantities r.lf .left and r.lf .right, this time expanding each range by one unit at its left. 4.3 Bundles with fan-out one The detection of bundles with fan-out 1 within the p-graph can be easily performed in O(n2), where n is the number of its vertices. Indeed, the incident set E(r) of a range r is a bundle with fan-out one if and only if r.b.left = r.f .left = ⊥. This immediately follows from the definitions given in Section 4.2. It is therefore possible to check all ranges the one after the other, once the backward and forward properties have been computed. These checks take constant time for each of the Θ(n2) ranges, hence the quadratic complexity. We now remove from F all bundles with fan-out 1 from the original bundle F. The result is the new bundle F ′, that has no sub-bundles with fan-out 1. 4.4 Bundles with fan-out two Efficient detection of bundles with fan-out two in F ′ is considerably more challenging. A direct generalization of the technique proposed for detecting bundles with fan-out 1 would use the following property, that is also a direct corollary of the definitions in Section 4.2: the incident set E(r ∪r′) of a tandem (r, r′) is a bundle with fan-out two if and only if all of the following conditions hold: (i) r.b.left = r′.f .left = ⊥, (ii) r.f .left ∈r′, r.f .right ∈r′, (iii) r′.b.left ∈r, r′.b.right ∈r. 529 However, checking all O(n4) tandems the one after the other would require time O(n4). Therefore, preserving the quadratic complexity of the overall algorithm requires a more complex representation. From now on, we assume that Vp = {x1, . . . , xn}, and we write [i, j] as a shorthand for the range [xi, xj]. First, we need to compute an additional data structure that will store local backward figures in a convenient way. Let us define the expansion table T as follows: for a given range r′ = [i′, j′], T(r′) is the set of all ranges r = [i, j] such that r.lb.left = i′ and r.lb.right = j′, ordered by increasing left boundary i. It turns out that the construction of such a table can be achieved in time O(n2). Moreover, it is possible to compute in O(n2) an auxiliary table T ′ that associates with r the first range r′′ in T([r.f.left, r.f.right]) such that r′′.b.right ≥r. Therefore, either (r, T ′(r)) anchors a valid bundle, or there is no bundle E such that the first component of V (E) is r. We now have all the pieces to extract bundles with fan-out 2 in time O(n2). We proceed as follows. For each range r = [i, j]: • We first retrieve r′ = [r.f.left, r.f.right] in constant time. • Then, we check in constant time whether r′.b.left lies within r. If it doesn’t, r is not the first part of a valid bundle with fan-out 2, and we move on to the next range r. • Finally, for each r′′ in the ordered set T(r′), starting with T ′(r), we check whether r′′.b.right is inside r. If it is not, we stop and move on to the next range r. If it is, we output the valid bundle (r, r′′) and move on to the next element in T(r′). Indeed, in case of a failure, the backward edge that relates a vertex in r′′ with a vertex outside r will still be included in all further elements in T(r′) since T(r′) is ordered by increasing left boundary. This step costs a constant time for each success, and a constant time for the unique failure, if any. This algorithm spends a constant time on each range plus a constant time on each bundle with fan-out 2. We shall prove in Section 5 that there are O(n2) bundles with fan-out 2. Therefore, this algorithm runs in time O(n2). Now that we have extracted all bundles, we need to extract an optimal decomposition of the input bundle F ′, i.e., a minimal size partition of all n elements (edges) in the input bundle such that each of these partition is a bundle (with fan-out 2, since bundles with fan-out 1 are excluded, except for the input bundle). By definition, a partition has minimal size if there is no other partition it is a refinment of.3 4.5 Extracting an optimal decomposition We have constructed the set of all (fan-out 2) subbundles of F ′. We now need to build one optimal decomposition of F ′ into sub-bundles. We need some more theoretical results on the properties of bundles. Lemma 1 Let E1 and E2 be two sub-bundles of F ′ (with fan-out 2) that have non-empty intersection, but that are not included the one in the other. Then E1 ∪E2 is a bundle (with fan-out 2). PROOF This lemma can be proved by considering all possible respective positions of the covers of E1 and E2, and discarding all situations that would lead to the existence of a fan-out 1 sub-bundle. ■ Theorem 1 For any bundle E, either it has at least one binary decomposition, or all its decompositions are refinements of a unique optimal one. PROOF Let us suppose that E has no binary decomposition. Its cover corresponds to the tandem (r, r′) = ([i, j], [i′, j′]). Let us consider two different decompositions of E, that correspond respectively to decompositions of the range r in two sets of sub-ranges of the form [i, k1], [k1 + 1, k2], . . . , [km, j] and [i, k′ 1], [k′ 1 + 1, k′ 2], . . . , [k′ m′, j]. For simplifying the notations, we write k0 = k′ 0 = i and km+1 = km′+1 = j. Since k0 = k′ 0, there exist an index p > 0 such that for any l < p, kl = k′ l, but kp ̸= k′ p: p is the index that identifies the first discrepancy between both decomposition. Since km+1 = km′+1, there must exist q ≤m and q′ ≤m′ such that q and q′ are strictly greater than p and that are the minimal indexes such that kq = k′ q′. By definition, all bundles of the form E[kl−1,kl] (p ≤l ≤q) have a non-empty intersection with at least one bundle of the form E[k′ l−1,k′ l] 3The term “refinement” is used in the usual way concerning partitions, i.e., a partition P1 is a refinement of another one P2 if all constituents in P1 are constituents of P2, or belongs to a subset of the partition P1 that is a partition of one element of P2. 530 (p ≤l ≤q′). The reverse is true as well. Applying Lemma 1, this shows that E([kp+1, kq]) is a bundle with fan-out 2. Therefore, by replacing all ranges involved in this union in one decomposition or the other, we get a third decomposition for which the two initial ones are strict refinements. This is a contradiction, which concludes the proof. ■ Lemma 2 Let E = V (r ∪r′) be a bundle, with r = [i, j]. We suppose it has a unique (non-binary) optimal decomposition, which decomposes [i, j] into [i, k1], [k1 + 1, k2], . . . , [km, j]. There exist no range r′′ ⊂r such that (i) Er′′ is a bundle and (ii) ∃l, 1 ≤l ≤m such that [kl, kl+1] ⊂r′′. PROOF Let us consider a range r′′ that would contradict the lemma. The union of r′′ and of the ranges in the optimal decomposition that have a non-empty intersection with r′′ is a fan-out 2 bundle that includes at least two elements of the optimal decomposition, but that is strictly included in E because the decomposition is not binary. This is a contradiction. ■ Lemma 3 Let E = V (r, r′) be a bundle, with r = [i, j]. We suppose it has a binary (optimal) decomposition (not necessarily unique). Let r′′ = [i, k] be the largest range starting in i such that k < j and such that it anchors a bundle, namely E(r′′). Then E(r′′) and E([k + 1, j]) form a binary decomposition of E. PROOF We need to prove that E([k + 1, j]) is a bundle. Each (optimal) binary decomposition of E decomposes r in 1, 2 or 3 sub-ranges. If no optimal decomposition decomposes r in at least 2 subranges, then the proof given here can be adapted by reasoning on r′ instead of r. We now suppose that at least one of them decomposes r in at least 2 sub-ranges. Therefore, it decomposes r in [i, k1] and [k1 + 1, j] or in [i, k1], [k1 + 1, k2] and [k2 + 1, j]. We select one of these optimal decomposition by taking one such that k1 is maximal. We shall now distinguish between two cases. First, let us suppose that r is decomposed into two sub-ranges [i, k1] and [k1 + 1, j] by the selected optimal decomposition. Obviously, E([i, k1]) is a “crossing” bundle, i.e., the right component of its cover is is a sub-range of r′. Since r is decomposed in two sub-ranges, it is necessarily the same for r′. Therefore, E([i, k1]) has a cover of the form [i, k1] ∪[i′, k′ 1] or [i, k1] ∪ [k′ 1 + 1, j]. Since r′′ includes [i, k1], E(r′′) has a cover of the form [i, k]∪[i′, k′] or [i, k]∪[k′ + 1, j]. This means that r′ is decomposed by E(r′′) in only 2 ranges, namely the right component of E(r′′)’s cover and another range, that we can call r′′′. Since r \ r′′ = [k + 1, j] may not anchor a bundle with fan-out 1, it must contain at least one crossing edge. All such edges necessarily fall within r′′′. Conversely, any crossing edge that falls inside r′′′ necessarily has its other end inside [k + 1, j]. Which means that E(r′′) and E(r′′′) form a binary decomposition of E. Therefore, by definition of k1, k = k1. Second, let us suppose that r is decomposed into 3 sub-ranges by the selected original decomposition (therefore, r′ is not decomposed by this decomposition). This means that this decomposition involves a bundle with a cover of the form [i, k1]∪[k2 + 1, j] and another bundle with a cover of the form [k1 + 1, k2] ∪r′ (this bundle is in fact E(r′)). If k ≥k2, then the left range of both members of the original decomposition are included in r′′, which means that E(r′′) = E, and therefore r′′ = r which is excluded. Note that k is at least as large as k1 (since [i, k1] is a valid “range starting in i such that k < j and such that it anchors a bundle”). Therefore, we have k1 ≤k < k2. Therefore, E([i, k1]) ⊂E(r′′), which means that all edges anchored inside [k2 + 1, j]) are included in E(r′′). Hence, E(r′′) can not be a crossing bundle without having a left component that is [i, j], which is excluded (it would mean E(r′′) = E). This means that E(r′′) is a bundle with a cover of the form [i, k] ∪[k′ + 1, j]. Which means that E(r′) is in fact the bundle whose cover is [k + 1, k′ + 1] ∪r′. Hence, E(r′′) and E(r′) form a binary decomposition of E. Hence, by definition of k1, k = k1. ■ As an immediate consequence of Lemmas 2 and 3, our algorithm for extracting the optimal decomposition for F ′ consists in applying the following procedure recursively, starting with F ′, and repeating it on each constructed sub-bundle E, until sub-bundles with only one edge are reached. Let E = E(r, r′) be a bundle, with r = [i, j]. One optimal decomposition of E can be obtained as follows. One selects the bundle with a left component starting in i and with the maximum length, and iterating this selection process until r is covered. The same is done with r′. We retain the optimal among both resulting decompositions (or one of them if they are both optimal). Note that this 531 decomposition is unique if and only if it has four components or more; it can not be ternary; it may be binary, and in this case it may be non-unique. This algorithm gives us a way to extract an optimal decomposition of F ′ in linear time w.r.t. the number of sub-bundles in this optimal decomposition. The only required data structure is, for each i (resp. k), the list of bundles with a cover of the form [i, j]∪[k, l] ordered by decreasing j (resp. l). This can trivially be constructed in time O(n2) from the list of all bundles we built in time O(n2) in the previous section. Since the number of bundles is bounded by O(n2) (as mentioned above and proved in Section 5), this means we can extract an optimal decomposition for F ′ in O(n2). Similar ideas apply to the simpler case of the decomposition of bundles with fan-out 1. 4.6 The main decomposition algorithm We now have to generalize our algorithm in order to handle the possible existence of fan-out 1 bundles. We achieve this by using the fan-out 2 algorithm recursively. First, we extract and remove (maximal) bundles with fan-out 1 from F, and recursively apply to each of them the complete algorithm. What remains is F ′, which is a set of bundles with no sub-bundles with fan-out 1. This means we can apply the algorithm presented above. Then, for each bundle with fan-out 1, we group it with a randomly chosen adjacent bundle with fan-out 2, which builds an expanded bundle with fan-out 2, which has a binary decomposition into the original bundle with fan-out 2 and the bundle with fan-out 1. 5 Time complexity analysis In Section 4, we claimed that there are no more than O(n2) bundles. In this section we sketch the proof of this result, which will prove the quadratic time complexity of our algorithm. Let us compute an upper bound on the number of bundles with fan-out two that can be found within the p-graph processed in Section 4.5, i.e., a p-graph with no fan-out 1 sub-bundle. Let E, E′ ⊆Ep be bundles with fan-out two. If E ⊂E′, then we say that E′ expands E. E′ is said to immediately expand E, written E →E′, if E′ expands E and there is no bundle E′′ such that E′′ expands E and E′ expands E′′. Let us represent bundles and the associated immediate expansion relation by means of a graph. Let E denote the set of all bundles (with fan-out two) in our p-graph. The e-graph associated with our LCFRS production p is the directed graph with vertices E and edges defined by the relation →. For E ∈E, we let out(E) = {E′ | E →E′} and in(E) = {E′ | E′ →E}. Lack of space prevents us from providing the proof of the following property. For any E ∈E that contains more than one edge, |out(E)| ≤2 and |in(E)| ≥2. This allows us to prove our upper bound on the size of E. Theorem 2 The e-graph associated with an LCFRS production p has at most n2 vertices, where n is the rank of p. PROOF Consider the e-graph associated with production p, with set of vertices E. For a vertex E ∈E, we define the level of E as the number |E| of edges in the corresponding bundle from the p-graph associated with p. Let d be the maximum level of a vertex in E. We thus have 1 ≤d ≤n. We now prove the following claim. For any integer k with 1 ≤k ≤d, the set of vertices in E with level k has no more than n elements. For k = 1, since there are no more than n edges in such a p-graph, the statement holds. We can now consider all vertices in E with level k > 1 (k ≤d). Let E(k−1) be the set of all vertices in E with level smaller than or equal to k −1, and let us call T (k−1) the set of all edges in the egraph that are leaving from some vertex in E(k−1). Since for each bundle E in E(k−1) we know that |out(E)| ≤2, we have |T (k−1)| ≤2|E(k−1)|. The number of vertices in E(k) with level larger than one is at least |E(k−1)| −n. Since for each E ∈E(k−1) we know that |in(E)| ≥2, we conclude that at least 2(|E(k−1)| −n) edges in T (k−1) must end up at some vertex in E(k). Let T be the set of edges in T (k−1) that impinge on some vertex in E \ E(k). Thus we have |T| ≤2|E(k−1)| − 2(|E(k−1)|−n) = 2n. Since the vertices of level k in E must have incoming edges from set T, and because each of them have at least 2 incoming edges, there cannot be more than n such vertices. This concludes the proof of our claim. Since the the level of a vertex in E is necessarily lower than n, this completes the proof. ■ The overall complexity of the complete algorithm can be computed by induction. Our induction hypothesis is that for m < n, the time complexity is in O(m2). This is obviously true for n = 1 and n = 2. Extracting the bundles 532 with fan-out 1 costs O(n2). These bundles are of length n1 . . . nm. Extracting bundles with fan-out 2 costs O((n −n1 −. . . −nm)2). Applying recursively the algorithm to bundles with fan-out 1 costs O(n2 1) + . . . + O(n2 m). Therefore, the complexity is in O(n2)+O((n −n1 −. . . −nm)2)+ Pn i=1 O(ni) = O(n2) + O(Pn i=1 ni) = O(n2). 6 Conclusion We have introduced an efficient algorithm for optimal reduction of the rank of LCFRSs with fan-out at most 2, that runs in quadratic time w.r.t. the rank of the input grammar. Given the fact that fan-out 1 bundles can be attached to any adjacent bundle in our factorization, we can show that our algorithm also optimizes time complexity for known tabular parsing algorithms for LCFRSs with fan-out 2. As for general LCFRS, it has been shown by Gildea (2010) that rank optimization and time complexity optimization are not equivalent. Furthermore, all known algorithms for rank or time complexity optimization have an exponential time complexity (G´omez-Rodr´ıguez et al., 2009). Acknowledgments Part of this work was done while the second author was a visiting scientist at Alpage (INRIA ParisRocquencourt and Universit´e Paris 7), and was financially supported by the hosting institutions. References Daniel Gildea. 2010. Optimal parsing strategies for linear context-free rewriting systems. In Human Language Technologies: The 11th Annual Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, Los Angeles, California. To appear. Carlos G´omez-Rodr´ıguez and Giorgio Satta. 2009. An optimal-time binarization algorithm for linear context-free rewriting systems with fan-out two. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 985–993, Suntec, Singapore, August. Association for Computational Linguistics. Carlos G´omez-Rodr´ıguez, Marco Kuhlmann, Giorgio Satta, and David J. Weir. 2009. Optimal reduction of rule length in linear context-free rewriting systems. In Proceedings of the North American Chapter of the Association for Computational Linguistics - Human Language Technologies Conference (NAACL’09:HLT), Boulder, Colorado. To appear. Michael A. Harrison. 1978. Introduction to Formal Language Theory. Addison-Wesley, Reading, MA. Aravind K. Joshi and Leon S. Levy. 1977. Constraints on local descriptions: Local transformations. SIAM Journal of Computing Marco Kuhlmann and Giorgio Satta. 2009. Treebank grammar techniques for non-projective dependency parsing. In Proceedings of the 12th Meeting of the European Chapter of the Association for Computational Linguistics (EACL 2009), Athens, Greece. To appear. Wolfgang Maier and Timm Lichte. 2009. Characterizing discontinuity in constituent treebanks. In Proceedings of the 14th Conference on Formal Grammar (FG 2009), Bordeaux, France. Wolfgang Maier and Anders Søgaard. 2008. Treebanks and mild context-sensitivity. In Philippe de Groote, editor, Proceedings of the 13th Conference on Formal Grammar (FG 2008), pages 61–76, Hamburg, Germany. CSLI Publications. Owen Rambow and Giorgio Satta. 1999. Independent parallelism in finite copying parallel rewriting systems. Theoretical Computer Science, 223:87–120. Giorgio Satta. 1992. Recognition of linear context-free rewriting systems. In Proceedings of the 30th Meeting of the Association for Computational Linguistics (ACL’92), Newark, Delaware. Hiroyuki Seki, Takashi Matsumura, Mamoru Fujii, and Tadao Kasami. 1991. On multiple context-free grammars. Theoretical Computer Science, 88:191– 229. K. Vijay-Shanker, David J. Weir, and Aravind K. Joshi. 1987. Characterizing structural descriptions produced by various grammatical formalisms. In Proceedings of the 25th Meeting of the Association for Computational Linguistics (ACL’87). David J. Weir. 1992. Linear context-free rewriting systems and deterministic tree-walk transducers. In Proceedings of the 30th Meeting of the Association for Computational Linguistics (ACL’92), Newark, Delaware. 533
|
2010
|
54
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 534–543, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics The Importance of Rule Restrictions in CCG Marco Kuhlmann Dept. of Linguistics and Philology Uppsala University Uppsala, Sweden Alexander Koller Cluster of Excellence Saarland University Saarbrücken, Germany Giorgio Satta Dept. of Information Engineering University of Padua Padua, Italy Abstract Combinatory Categorial Grammar (CCG) is generally construed as a fully lexicalized formalism, where all grammars use one and the same universal set of rules, and crosslinguistic variation is isolated in the lexicon. In this paper, we show that the weak generative capacity of this ‘pure’ form of CCG is strictly smaller than that of CCG with grammar-specific rules, and of other mildly context-sensitive grammar formalisms, including Tree Adjoining Grammar (TAG). Our result also carries over to a multi-modal extension of CCG. 1 Introduction Combinatory Categorial Grammar (CCG) (Steedman, 2001; Steedman and Baldridge, 2010) is an expressive grammar formalism with formal roots in combinatory logic (Curry et al., 1958) and links to the type-logical tradition of categorial grammar (Moortgat, 1997). It has been successfully used for a wide range of practical tasks, such as data-driven parsing (Hockenmaier and Steedman, 2002; Clark and Curran, 2007), wide-coverage semantic construction (Bos et al., 2004), and the modelling of syntactic priming (Reitter et al., 2006). It is well-known that CCG can generate languages that are not context-free (which is necessary to capture natural languages), but can still be parsed in polynomial time. Specifically, VijayShanker and Weir (1994) identified a version of CCG that is weakly equivalent to Tree Adjoining Grammar (TAG) (Joshi and Schabes, 1997) and other mildly context-sensitive grammar formalisms, and can generate non-context-free languages such as anbncn. The generative capacity of CCG is commonly attributed to its flexible composition rules, which allow it to model more complex word orders that context-free grammar can. The discussion of the (weak and strong) generative capacity of CCG and TAG has recently been revived (Hockenmaier and Young, 2008; Koller and Kuhlmann, 2009). In particular, Koller and Kuhlmann (2009) have shown that CCGs that are pure (i.e., they can only use generalized composition rules, and there is no way to restrict the instances of these rules that may be used) and first-order (i.e., all argument categories are atomic) can not generate anbncn. This shows that the generative capacity of at least first-order CCG crucially relies on its ability to restrict rule instantiations, and is at odds with the general conception of CCG as a fully lexicalized formalism, in which all grammars use one and the same set of universal rules. A question then is whether the result carries over to pure CCG with higher-order categories. In this paper, we answer this question to the positive: We show that the weak generative capacity of general pure CCG is still strictly smaller than that of the formalism considered by Vijay-Shanker and Weir (1994); composition rules can only achieve their full expressive potential if their use can be restricted. Our technical result is that every language L that can be generated by a pure CCG has a context-free sublanguage L0 L such that every string in L is a permutation of a string in L0, and vice versa. This means that anbncn, for instance, cannot be generated by pure CCG, as it does not have any (non-trivial) permutation-equivalent sublanguages. Conversely, we show that there are still languages that can be generated by pure CCG but not by context-free grammar. We then show that our permutation language lemma also holds for pure multi-modal CCG as defined by Baldridge and Kruijff (2003), in which the use of rules can be controlled through the lexicon entries by assigning types to slashes. Since this extension was intended to do away with the need for grammar-specific rule restrictions, it comes as quite a surprise that pure multi-modal 534 CCG in the style of Baldridge and Kruijff (2003) is still less expressive than the CCG formalism used by Vijay-Shanker and Weir (1994). This means that word order in CCG cannot be fully lexicalized with the current formal tools; some ordering constraints must be specified via language-specific combination rules and not in lexicon entries. On the other hand, as pure multi-modal CCG has been successfully applied to model the syntax of a variety of natural languages, another way to read our results is as contributions to a discussion about the exact expressiveness needed to model natural language. The remainder of this paper is structured as follows. In Section 2, we introduce the formalism of pure CCG that we consider in this paper, and illustrate the relevance of rule restrictions. We then study the generative capacity of pure CCG in Section 3; this section also presents our main result. In Section 4, we show that this result still holds for multi-modal CCG. Section 5 concludes the paper with a discussion of the relevance of our findings. 2 Combinatory Categorial Grammar We start by providing formal definitions for categories, syntactic rules, and grammars, and then discuss the relevance of rule restrictions for CCG. 2.1 Categories Given a finite set A of atomic categories, the set of categories over A is the smallest set C such that A C, and .x=y/; .xny/ 2 C whenever x; y 2 C. A category x=y represents a function that seeks a string with category y to the right (indicated by the forward slash) and returns a new string with category x; a category xny instead seeks its argument to the left (indicated by the backward slash). In the remainder of this paper, we use lowercase sansserif letters such as x; y; z as variables for categories, and the vertical bar j as a variable for slashes. In order to save some parentheses, we understand slashes as left-associative operators, and write a category such as .x=y/nz as x=ynz. The list of arguments of a category c is defined recursively as follows: If c is atomic, then it has no arguments. If c D xjy for some categories x and y, then the arguments of c are the slashed category jy, plus the arguments of x. We number the arguments of a category from outermost to innermost. The arity of a category is the number of its arguments. The target of a category c is the atomic category that remains when stripping c of its arguments. x=y y ) x forward application > y xny ) x backward application < x=y y=z ) x=z forward harmonic composition >B ynz xny ) xnz backward harmonic composition <B x=y ynz ) xnz forward crossed composition >B y=z xny ) x=z backward crossed composition <B Figure 1: The core set of rules of CCG. 2.2 Rules The syntactic rules of CCG are directed versions of combinators in the sense of combinatory logic (Curry et al., 1958). Figure 1 lists a core set of commonly assumed rules, derived from functional application and the B combinator, which models functional composition. When talking about these rules, we refer to the premise containing the argument jy as the primary premise, and to the other premise as the secondary premise of the rule. The rules in Figure 1 can be generalized into composition rules of higher degrees. These are defined as follows, where n 0 and ˇ is a variable for a sequence of n arguments. x=y yˇ ) xˇ generalized forward composition >n yˇ xny ) xˇ generalized backward composition <n We call the value n the degree of the composition rule. Note that the rules in Figure 1 are the special cases for n D 0 and n D 1. Apart from the core rules given in Figure 1, some versions of CCG also use rules derived from the S and T combinators of combinatory logic, called substitution and type-raising, the latter restricted to the lexicon. However, since our main point of reference in this paper, the CCG formalism defined by Vijay-Shanker and Weir (1994), does not use such rules, we will not consider them here, either. 2.3 Grammars and Derivations With the set of rules in place, we can define a pure combinatory categorial grammar (PCCG) as a construct G D .A; ˙; L; s/, where A is an alphabet of atomic categories, s 2 A is a distinguished atomic category called the final category, ˙ is a finite set of terminal symbols, and L is a finite relation between symbols in ˙ and categories over A, called the lexicon. The elements of the lexicon L are called lexicon entries, and we represent them using the notation ` x, where 2 ˙ and x is a category over A. A category that occurs in a lexicon entry is called a lexical category. 535 A derivation in a grammar G can be represented as a derivation tree as follows. Given a string w 2 ˙, we choose a lexicon entry for each occurrence of a symbol in w, line up the respective lexical categories from left to right, and apply admissible rules to adjacent pairs of categories. After the application of a rule, only the conclusion is available for future applications. We iterate this process until we end up with a single category. The string w is called the yield of the resulting derivation tree. A derivation tree is complete, if the last category is the final category of G. The language generated by G, denoted by L.G/, is formed by the yields of all complete derivation trees. 2.4 Degree Restrictions Work on CCG generally assumes an upper bound on the degree of composition rules that can be used in derivations. We also employ this restriction, and only consider grammars with compositions of some bounded (but arbitrary) degree n 0.1 CCG with unbounded-degree compositions is more expressive than bounded-degree CCG or TAG (Weir and Joshi, 1988). Bounded-degree grammars have a number of useful properties, one of which we mention here. The following lemma rephrases Lemma 3.1 in Vijay-Shanker and Weir (1994). Lemma 1 For every grammar G, every argument in a derivation of G is the argument of some lexical category of G. As a consequence, there is only a finite number of categories that can occur as arguments in some derivation. In the presence of a bound on the degree of composition rules, this implies the following: Lemma 2 For every grammar G, there is a finite number of categories that can occur as secondary premises in derivations of G. Proof. The arity of a secondary premise c can be written as m C n, where m is the arity of the first argument of the corresponding primary premise, and n is the degree of the rule applied. Since each argument is an argument of some lexical category of G (Lemma 1), and since n is assumed to be bounded, both m and n are bounded. Hence, there is a bound on the number of choices for c. Note that the number of categories that can occur as primary premises is generally unbounded even in a grammar with bounded degree. 1For practical grammars, n 4. 2.5 Rule Restrictions The rule set of pure CCG is universal: the difference between the grammars of different languages should be restricted to different choices of categories in the lexicon. This is what makes pure CCG a lexicalized grammar formalism (Steedman and Baldridge, 2010). However, most practical CCG grammars rely on the possibility to exclude or restrict certain rules. For example, Steedman (2001) bans the rule of forward crossed composition from his grammar of English, and stipulates that the rule of backward crossed composition may be applied only if both of its premises share the common target category s, representing sentences. Exclusions and restrictions of rules are also assumed in much of the language-theoretic work on CCG. In particular, they are essential for the formalism used in the aforementioned equivalence proof for CCG and TAG (Vijay-Shanker and Weir, 1994). To illustrate the formal relevance of rule restrictions, suppose that we wanted to write a pure CCG that generates the language L3 D f anbncn j n 1 g , which is not context-free. An attempt could be G1 D .f s; a; b; c g; f a; b; c g; L; s/ , where the lexicon L is given as follows: a ` a , b ` s=cna , b ` b=cna , b ` s=c=bna , b ` s=c=bna , c ` c . From a few sample derivations like the one given in Figure 2a, we can convince ourselves that G1 generates all strings of the form anbncn, for any n 1. However, a closer inspection reveals that it also generates other, unwanted strings—in particular, strings of the form .ab/ncn, as witnessed by the derivation given in Figure 2b. Now suppose that we would have a way to only allow those instances of generalized composition in which the secondary premise has the form b=c=bna or b=cna. Then the compositions b=c=b b=c b=c=c >1 and s=c=b b=c s=c=c >1 would be disallowed, and it is not hard to see that G1 would generate exactly anbncn. As we will show in this paper, our attempt to capture L3 with a pure CCG grammar failed not only because we could not think of one: L3 cannot be generated by any pure CCG. 536 a................... a a........... a a... a b... s=c=bna b....... b=c=bna b............... b=cna c....................... c c........................... c c............................... c <0 s=c=b >3 s=c=c=bna <0 s=c=c=b >2 s=c=c=cna <0 s=c=c=c >0 s=c=c >0 s=c >0 s (a) Derivation of the string aaabbbccc. a........... a b........... s=c=bna a... a b... b=c=bna a... a b... b=cna c........... c c................... c c....................... c <0 s=c=b <0 b=c=b <0 b=c >1 b=c=c >0 b=c >1 s=c=c >0 s=c >0 s (b) Derivation of the string abababccc. Figure 2: Two derivations of the grammar G1. 3 The Generative Capacity of Pure CCG We will now develop a formal argument showing that rule restrictions increase the weak generative capacity of CCG. We will first prove that pure CCG is still more expressive than context-free grammar. We will then spend the rest of this section working towards the result that pure CCG is strictly less expressive than CCG with rule restrictions. Our main technical result will be the following: Theorem 1 Every language that can be generated by a pure CCG has a Parikh-equivalent context-free sublanguage. Here, two languages L and L0 are called Parikhequivalent if every string in L is the permutation of a string in L0 and vice versa. 3.1 CFG ¨ PCCG Proposition 1 The class of languages generated by pure CCG properly includes the class of contextfree languages. Proof. To see the inclusion, it suffices to note that pure CCG when restricted to application rules is the same as AB-grammar, the classical categorial formalism investigated by Ajdukiewicz and BarHillel (Bar-Hillel et al., 1964). This formalism is weakly equivalent to context-free grammar. To see that the inclusion is proper, we can go back to the grammar G1 that we gave in Section 2.5. We have already discussed that the language L3 is included in L.G1/. We can also convince ourselves that all strings generated by the grammar G1 have an equal number of as, bs and cs. Consider now the regular language R D abc. From our observations, it follows that L.G1/ \ R D L3. Since context-free languages are closed under intersection with regular languages, we find that L.G1/ can be context-free only if L3 is. Since L3 is not context-free, we therefore conclude that L.G1/ is not context-free, either. Two things are worth noting. First, our result shows that the ability of CCG to generate non-context-free languages does not hinge on the availability of substitution and type-raising rules: The derivations of G1 only use generalized compositions. Neither does it require the use of functional argument categories: The grammar G1 is first-order in the sense of Koller and Kuhlmann (2009). Second, it is important to note that if the composition degree n is restricted to 0 or 1, pure CCG actually collapses to context-free expressive power. This is clear for n D 0 because of the equivalence to AB grammar. For n D 1, observe that the arity of the result of a composition is at most as high as 537 that of each premise. This means that the arity of any derived category is bounded by the maximal arity of lexical categories in the grammar, which together with Lemma 1 implies that there is only a finite set of derivable categories. The set of all valid derivations can then be simulated by a context-free grammar. In the presence of rules with n 2, the arities of derived categories can grow unboundedly. 3.2 Active and Inactive Arguments In the remainder of this section, we will develop the proof of Theorem 1, and use it to show that the generative capacity of PCCG is strictly smaller than that of CCG with rule restrictions. For the proof, we adopt a certain way to view the information flow in CCG derivations. Consider the following instance of forward harmonic composition: a=b b=c ) a=c This rule should be understood as obtaining its conclusion a=c from the primary premise a=b by the removal of the argument =b and the subsequent transfer of the argument =c from the secondary premise. With this picture in mind, we will view the two occurrences of =c in the secondary premise and in the conclusion as two occurrences of one and the same argument. Under this perspective, in a given derivation, an argument has a lifespan that starts in a lexical category and ends in one of two ways: either in the primary or in the secondary premise of a composition rule. If it ends in a primary premise, it is because it is matched against a subcategory of the corresponding secondary premise; this is the case for the argument =b in the example above. We will refer to such arguments as active. If an argument ends its life in a secondary premise, it is because it is consumed as part of a higher-order argument. This is the case for the argument =c in the secondary premise of the following rule instance: a=.b=c/ b=c=d ) a=d (Recall that we assume that slashes are left-associative.) We will refer to such arguments as inactive. Note that the status of an argument as either active or inactive is not determined by the grammar, but depends on a concrete derivation. The following lemma states an elementary property in connection with active and inactive arguments, which we will refer to as segmentation: Lemma 3 Every category that occurs in a CCG derivation has the general form a˛ˇ, where a is an atomic category, ˛ is a sequence of inactive arguments, and ˇ is a sequence of active arguments. Proof. The proof is by induction on the depth of a node in the derivation. The property holds for the root (which is labeled with the final category), and is transferred from conclusions to premises. 3.3 Transformation The fundamental reason for why the example grammar G1 from Section 2.5 overgenerates is that in the absence of rule restrictions, we have no means to control the point in a derivation at which a category combines with its arguments. Consider the examples in Figure 2: It is because we cannot ensure that the bs finish combining with the other bs before combining with the cs that the undesirable word order in Figure 2b has a derivation. To put it as a slogan: Permuting the words allows us to saturate arguments prematurely. In this section, we show that this property applies to all pure CCGs. More specifically, we show that, in a derivation of a pure CCG, almost all active arguments of a category can be saturated before that category is used as a secondary premise; at most one active argument must be transferred to the conclusion of that premise. Conversely, any derivation that still contains a category with at least two active arguments can be transformed into a new derivation that brings us closer to the special property just characterized. We formalize this transformation by means of a system of rewriting rules in the sense of Baader and Nipkow (1998). The rules are given in Figure 3. To see how they work, let us consider the first rule, R1; the other ones are symmetric. This rules states that, whenever we see a derivation in which a category of the form x=y (here marked as A) is combined with a category of the form yˇ=z (marked as B), and the result of this combination is combined with a category of the form z
(C), then the resulting category can also be obtained by ‘rotating’ the derivation to first saturate =z by combining B with C, and only then do the combination with A. When applying these rotations exhaustively, we end up with a derivation in which almost all active arguments of a category are saturated before that category is used as a secondary premise. Applying the transformation to the derivation in Figure 2a, for instance, yields the derivation in Figure 2b. We need the following result for some of the lemmas we prove below. We call a node in a deriv538 A x=y B yˇ=z xˇ=z C z
xˇ
R1 H) x=y yˇ=z z
yˇ
xˇ
B yˇ=z A xny xˇ=z C z
xˇ
R2 H) yˇ=z z
yˇ
xny xˇ
C z
A x=y B yˇnz xˇnz xˇ
R3 H) x=y z
yˇnz yˇ
xˇ
C z
B yˇnz A xny xˇnz xˇ
R4 H) z
yˇnz yˇ
xny xˇ
Figure 3: Rewriting rules used in the transformation. Here,
represents a (possibly empty) sequence of arguments, and ˇ represents a sequence of arguments in which the first (outermost) argument is active. ation critical if its corresponding category contains more than one active argument and it is the secondary premise of a rule. We say that u is a highest critical node if there is no other critical node whose distance to the root is shorter. Lemma 4 If u is a highest critical node, then we can apply one of the transformation rules to the grandparent of u. Proof. Suppose that the category at u has the form yˇ=z, where =z is an active argument, and the first argument in ˇ is active as well. (The other possible case, in which the relevant occurrence has the form yˇnz, can be treated symmetrically.) Since u is a secondary premise, it is involved in an inference of one of the following two forms: x=y yˇ=z xˇ=z yˇ=z xny xˇ=z Since u is a highest critical node, the conclusion of this inference is not a critical node itself; in particular, it is not a secondary premise. Therefore, the above inferences can be extended as follows: x=y yˇ=z xˇ=z z
xˇ
yˇ=z xny xˇ=z z
xˇ
These partial derivations match the left-hand side of the rewriting rules R1 and R2, respectively. Hence, we can apply a rewriting rule to the derivation. We now show that the transformation is welldefined, in the sense that it terminates and transforms derivations of a grammar G into new derivations of G. Lemma 5 The rewriting of a derivation tree ends after a finite number of steps. Proof. We assign natural numbers to the nodes of a derivation tree as follows. Each leaf node is assigned the number 0. For an inner node u, which corresponds to the conclusion of a composition rule, let m; n be the numbers assigned to the nodes corresponding to the primary and secondary premise, respectively. Then u is assigned the number 1 C 2m C n. Suppose now that we have associated premise A with the number x, premise B with the number y, and premise C with the number z. It is then easy to verify that the conclusion of the partial derivation on the left-hand side of each rule has the value 3 C 4x C 2y C z, while the conclusion of the right-hand side has the value 2 C 2x C 2y C z. Thus, each step decreases the value of a derivation tree under our assignment by the amount 1 C 2x. Since this value is positive for all choices of x, the rewriting ends after a finite number of steps. To convince ourselves that our transformation does not create ill-formed derivations, we need to show that none of the rewriting rules necessitates the use of composition operations whose degree is higher than the degree of the operations used in the original derivation. Lemma 6 Applying the rewriting rules from the top down does not increase the degree of the composition operations. Proof. The first composition rule used in the lefthand side of each rewriting rule has degree jˇj C 1, the second rule has degree j
j; the first rule used in the right-hand side has degree j
j, the second rule has degree jˇj C j
j. To prove the claim, it suffices to show that j
j 1. This is a consequence of the following two observations. 1. In the category xˇ
, the arguments in
occur on top of the arguments in ˇ, the first of which is active. Using the segmentation property stated in Lemma 3, we can therefore infer that
does not contain any inactive arguments. 539 2. Because we apply rules top-down, premise B is a highest critical node in the derivation (by Lemma 4). This means that the category at premise C contains at most one active argument; otherwise, premise C would be a critical node closer to the root than premise B. We conclude that, if we rewrite a derivation d of G top-down until exhaustion, then we obtain a new valid derivation d 0. We call all derivations d 0 that we can build in this way transformed. It is easy to see that a derivation is transformed if and only if it contains no critical nodes. 3.4 Properties of Transformed Derivations The special property established by our transformation has consequences for the generative capacity of pure CCG. In particular, we will now show that the set of all transformed derivations of a given grammar yields a context-free language. The crucial lemma is the following: Lemma 7 For every grammar G, there is some k 0 such that no category in a transformed derivation of G has arity greater than k. Proof. The number of inactive arguments in the primary premise of a rule does not exceed the number of inactive arguments in the conclusion. In a transformed derivation, a symmetric property holds for active arguments: Since each secondary premise contains at most one active argument, the number of active arguments in the conclusion of a rule is not greater than the number of active arguments in its primary premise. Taken together, this implies that the arity of a category that occurs in a transformed derivation is bounded by the sum of the maximal arity of a lexical category (which bounds the number of active arguments), and the maximal arity of a secondary premise (which bounds the number of inactive arguments). Both of these values are bounded in G. Lemma 8 The yields corresponding to the set of all transformed derivations of a pure CCG form a context-free language. Proof. Let G be a pure CCG. We construct a context-free grammar GT that generates the yields of the set of all transformed derivations of G. As the set of terminals of GT , we use the set of terminals of G. To form the set of nonterminals, we take all categories that can occur in a transformed derivation of G, and mark each argument as either ‘active’ (C) or ‘inactive’ ( ), in all possible ways that respect the segmentation property stated in Lemma 3. Note that, because of Lemma 7 and Lemma 1, the set of nonterminals is finite. As the start symbol, we use s, the final category of G. The set of productions of GT is constructed as follows. For each lexicon entry ` c of G, we include all productions of the form x ! , where x is some marked version of c. These productions represent all valid guesses about the activity of the arguments of c during a derivation of G. The remaining productions encode all valid instantiations of composition rules, keeping track of active and inactive arguments to prevent derivations with critical nodes. More specifically, they have the form xˇ ! x=yC yˇ or xˇ ! yˇ xnyC , where the arguments in the y-part of the secondary premise are all marked as inactive, the sequence ˇ contains at most one argument marked as active, and the annotations of the left-hand side nonterminal are copied over from the corresponding annotations on the right-hand side. The correctness of the construction of GT can be proved by induction on the length of a transformed derivation of G on the one hand, and the length of a derivation of GT on the other hand. 3.5 PCCG ¨ CCG We are now ready to prove our main result, repeated here for convenience. Theorem 1 Every language that can be generated by a pure CCG grammar has a Parikh-equivalent context-free sublanguage. Proof. Let G be a pure CCG, and let LT be the set of yields of the transformed derivations of G. Inspecting the rewriting rules, it is clear that every string of L.G/ is the permutation of a string in LT : the transformation only rearranges the yields. By Lemma 8, we also know that LT is context-free. Since every transformed derivation is a valid derivation of G, we have LT L.G/. As an immediate consequence, we find: Proposition 2 The class of languages generated by pure CCG cannot generate all languages that can be generated by CCG with rule restrictions. Proof. The CCG formalism considered by VijayShanker and Weir (1994) can generate the non-context-free language L3. However, the only Parikhequivalent sublanguage of that language is L3 itself. From Theorem 1, we therefore conclude that L3 cannot be generated by pure CCG. 540 In the light of the equivalence result established by Vijay-Shanker and Weir (1994), this means that pure CCG cannot generate all languages that can be generated by TAG. 4 Multi-Modal CCG We now extend Theorem 1 to multi-modal CCG. We will see that at least for a popular version of multi-modal CCG, the B&K-CCG formalism presented by Baldridge and Kruijff (2003), the proof can be adapted quite straightforwardly. This means that even B&K-CCG becomes less expressive when rule restrictions are disallowed. 4.1 Multi-Modal CCG The term ‘multi-modal CCG’ (MM-CCG) refers to a family of extensions to CCG which attempt to bring some of the expressive power of Categorial Type Logic (Moortgat, 1997) into CCG. Slashes in MM-CCG have slash types, and rules can be restricted to only apply to arguments that have slashes of the correct type. The idea behind this extension is that many constraints that in ordinary CCG can only be expressed in terms of rule restrictions can now be specified in the lexicon entries by giving the slashes the appropriate types. The most widely-known version of multi-modal CCG is the formalism defined by Baldridge and Kruijff (2003) and used by Steedman and Baldridge (2010); we refer to it as B&K-CCG. This formalism uses an inventory of four slash types, f ?; ; ˘; g, arranged in a simple type hierarchy: ? is the most general type, the most specific, and and ˘ are in between. Every slash in a B&K-CCG lexicon is annotated with one of these slash types. The combinatory rules in B&K-CCG, given in Figure 4, are defined to be sensitive to the slash types. In particular, slashes with the types ˘ and can only be eliminated by harmonic and crossed compositions, respectively.2 Thus, a grammar writer can constrain the application of harmonic and crossed composition rules to certain categories by assigning appropriate types to the slashes of this category in the lexicon. Application rules apply to slashes of any type. As before, we call an MM-CCG grammar pure if it only uses application and generalized compositions, and does not provide means to restrict rule applications. 2Our definitions of generalized harmonic and crossed composition are the same as the ones used by Hockenmaier and Young (2008), but see the discussion in Section 4.3. x=?y y ) x forward application y xn?y ) x backward application x=˘y y=˘zˇ ) x=˘zˇ forward harmonic composition x=y ynzˇ ) xnzˇ forward crossed composition yn˘zˇ xn˘y ) xn˘zˇ backward harmonic composition y=zˇ xny ) x=zˇ backward crossed composition Figure 4: Rules in B&K-CCG. 4.2 Rule Restrictions in B&K-CCG We will now see what happens to the proof of Theorem 1 in the context of pure B&K-CCG. There is only one point in the entire proof that could be damaged by the introduction of slash types, and that is the result that if a transformation rule from Figure 3 is applied to a correct derivation, then the result is also grammatical. For this, it must not only be the case that the degree on the composition operations is preserved (Lemma 6), but also that the transformed derivation remains consistent with the slash types. Slash types make the derivation process sensitive to word order by restricting the use of compositions to categories with the appropriate type, and the transformation rules permute the order of the words in the string. There is a chance therefore that a transformed derivation might not be grammatical in B&K-CCG. We now show that this does not actually happen, for rule R3; the other three rules are analogous. Using s1; s2; s3 as variables for the relevant slash types, rule R3 appears in B&K-CCG as follows: z
x=s1y yjs2wˇns3z xjs2wˇns3z xjs2wˇ
R3 H) x=s1y z
yjs2wˇns3z yjs2wˇ
xjs2wˇ
Because the original derivation is correct, we know that, if the slash of w is forward, then s1 and s2 are subtypes of ˘; if the slash is backward, they are subtypes of . A similar condition holds for s3 and the first slash in
; if
is empty, then s3 can be anything because the second rule is an application. After the transformation, the argument =s1y is used to compose with yjs2wˇ
. The direction of the slash in front of the w is the same as before, so the (harmonic or crossed) composition is still compatible with the slash types s1 and s2. An analogous argument shows that the correctness of combining ns3z with
carries over from the left to the right-hand side. Thus the transformation maps grammatical derivations into grammatical derivations. The rest of the proof in Section 3 continues to work literally, so we have the following result: 541 Theorem 2 Every language that can be generated by a pure B&K-CCG grammar contains a Parikhequivalent context-free sublanguage. This means that pure B&K-CCG is just as unable to generate L3 as pure CCG is. In other words, the weak generative capacity of CCG with rule restrictions, and in particular that of the formalism considered by Vijay-Shanker and Weir (1994), is strictly greater than the generative capacity of pure B&K-CCG—although we conjecture (but cannot prove) that pure B&K-CCG is still more expressive than pure non-modal CCG. 4.3 Towards More Expressive MM-CCGs To put the result of Theorem 2 into perspective, we will now briefly consider ways in which B&K-CCG might be modified in order to obtain a pure multimodal CCG that is weakly equivalent to CCG in the style of Vijay-Shanker and Weir (1994). Such a modification would have to break the proof in Section 4.2, which is harder than it may seem at first glance. For instance, simply assuming a more complex type system will not do it, because the arguments ns3z and =s1y are eliminated using the same rules in the original and the transformed derivations, so if the derivation step was valid before, it will still be valid after the transformation. Instead, we believe that it is necessary to make the composition rules sensitive to the categories inside ˇ and
instead of only the arguments ns3z and =s1y, and we can see two ways how to do this. First, one could imagine a version of multimodal CCG with unary modalities that can be used to mark certain category occurrences. In such an MM-CCG, the composition rules for a certain slash type could be made sensitive to the presence or absence of unary modalities in ˇ. Say for instance that the slash type s1 in the modalized version of R3 in Section 4.2 would require that no category in the secondary argument is marked with the unary modality ‘’, but ˇ contains a category marked with ‘’. Then the transformed derivation would be ungrammatical. A second approach concerns the precise definition of the generalized composition rules, about which there is a surprising degree of disagreement. We have followed Hockenmaier and Young (2008) in classifying instances of generalized forward composition as harmonic if the innermost slash of the secondary argument is forward and crossed if it is backward. However, generalized forward composition is sometimes only accepted as harmonic if all slashes of the secondary argument are forward (see e.g. Baldridge (2002) (40, 41), Steedman (2001) (19)). At the same time, based on the principle that CCG rules should be derived from proofs of Categorial Type Logic as Baldridge (2002) does, it can be argued that generalized composition rules of the form x=y y=znw ) x=znw, which we have considered as harmonic, should actually be classified as crossed, due to the presence of a slash of opposite directionality in front of the w. This definition would break our proof. Thus our result might motivate further research on the ‘correct’ definition of generalized composition rules, which might then strengthen the generative capacity of pure MM-CCG. 5 Conclusion In this paper, we have shown that the weak generative capacity of pure CCG and even pure B&K-CCG crucially depends on the ability to restrict the application of individual rules. This means that these formalisms cannot be fully lexicalized, in the sense that certain languages can only be described by selecting language-specific rules. Our result generalizes Koller and Kuhlmann’s (2009) result for pure first-order CCG. Our proof is not as different as it looks at first glance, as their construction of mapping a CCG derivation to a valency tree and back to a derivation provides a different transformation on derivation trees. Our transformation is also technically related to the normal form construction for CCG parsing presented by Eisner (1996). Of course, at the end of the day, the issue that is more relevant to computational linguistics than a formalism’s ability to generate artificial languages such as L3 is how useful it is for modeling natural languages. CCG, and multi-modal CCG in particular, has a very good track record for this. In this sense, our formal result can also be understood as a contribution to a discussion about the expressive power that is needed to model natural languages. Acknowledgments We have profited enormously from discussions with Jason Baldridge and Mark Steedman, and would also like to thank the anonymous reviewers for their detailed comments. 542 References Franz Baader and Tobias Nipkow. 1998. Term Rewriting and All That. Cambridge University Press. Jason Baldridge and Geert-Jan M. Kruijff. 2003. Multi-modal Combinatory Categorial Grammar. In Proceedings of the Tenth Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 211–218, Budapest, Hungary. Jason Baldridge. 2002. Lexically Specified Derivational Control in Combinatory Categorial Grammar. Ph.D. thesis, University of Edinburgh. Yehoshua Bar-Hillel, Haim Gaifman, and Eli Shamir. 1964. On categorial and phrase structure grammars. In Language and Information: Selected Essays on their Theory and Application, pages 99–115. Addison-Wesley. Johan Bos, Stephen Clark, Mark Steedman, James R. Curran, and Julia Hockenmaier. 2004. Widecoverage semantic representations from a CCG parser. In Proceedings of the 20th International Conference on Computational Linguistics (COLING), pages 176–182, Geneva, Switzerland. Stephen Clark and James Curran. 2007. Widecoverage efficient statistical parsing with CCG and log-linear models. Computational Linguistics, 33(4). Haskell B. Curry, Robert Feys, and William Craig. 1958. Combinatory Logic. Volume 1. Studies in Logic and the Foundations of Mathematics. NorthHolland. Jason Eisner. 1996. Efficient normal-form parsing for combinatory categorial grammar. In Proceedings of the 34th Annual Meeting of the Association for Computational Linguistics (ACL), pages 79–86, Santa Cruz, CA, USA. Julia Hockenmaier and Mark Steedman. 2002. Generative models for statistical parsing with Combinatory Categorial Grammar. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), pages 335–342, Philadelphia, USA. Julia Hockenmaier and Peter Young. 2008. Non-local scrambling: the equivalence of TAG and CCG revisited. In Proceedings of the 9th Internal Workshop on Tree Adjoining Grammars and Related Formalisms (TAG+9), Tübingen, Germany. Aravind K. Joshi and Yves Schabes. 1997. TreeAdjoining Grammars. In Grzegorz Rozenberg and Arto Salomaa, editors, Handbook of Formal Languages, volume 3, pages 69–123. Springer. Alexander Koller and Marco Kuhlmann. 2009. Dependency trees and the strong generative capacity of CCG. In Proceedings of the Twelfth Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 460–468, Athens, Greece. Michael Moortgat. 1997. Categorial type logics. In Handbook of Logic and Language, chapter 2, pages 93–177. Elsevier. David Reitter, Julia Hockenmaier, and Frank Keller. 2006. Priming effects in combinatory categorial grammar. In Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 308–316, Sydney, Australia. Mark Steedman and Jason Baldridge. 2010. Combinatory categorial grammar. In R. Borsley and K. Borjars, editors, Non-Transformational Syntax. Blackwell. Draft 7.0, to appear. Mark Steedman. 2001. The Syntactic Process. MIT Press. K. Vijay-Shanker and David J. Weir. 1994. The equivalence of four extensions of context-free grammars. Mathematical Systems Theory, 27(6):511–546. David J. Weir and Aravind K. Joshi. 1988. Combinatory categorial grammars: Generative power and relationship to linear context-free rewriting systems. In Proceedings of the 26th Annual Meeting of the Association for Computational Linguistics, pages 278– 285, Buffalo, NY, USA. 543
|
2010
|
55
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 544–554, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Automatic Evaluation of Linguistic Quality in Multi-Document Summarization Emily Pitler, Annie Louis, Ani Nenkova Computer and Information Science University of Pennsylvania Philadelphia, PA 19104, USA epitler,lannie,[email protected] Abstract To date, few attempts have been made to develop and validate methods for automatic evaluation of linguistic quality in text summarization. We present the first systematic assessment of several diverse classes of metrics designed to capture various aspects of well-written text. We train and test linguistic quality models on consecutive years of NIST evaluation data in order to show the generality of results. For grammaticality, the best results come from a set of syntactic features. Focus, coherence and referential clarity are best evaluated by a class of features measuring local coherence on the basis of cosine similarity between sentences, coreference information, and summarization specific features. Our best results are 90% accuracy for pairwise comparisons of competing systems over a test set of several inputs and 70% for ranking summaries of a specific input. 1 Introduction Efforts for the development of automatic text summarizers have focused almost exclusively on improving content selection capabilities of systems, ignoring the linguistic quality of the system output. Part of the reason for this imbalance is the existence of ROUGE (Lin and Hovy, 2003; Lin, 2004), the system for automatic evaluation of content selection, which allows for frequent evaluation during system development and for reporting results of experiments performed outside of the annual NIST-led evaluations, the Document Understanding Conference (DUC)1 and the Text Analysis Conference (TAC)2. Few metrics, however, have been proposed for evaluating linguistic 1http://duc.nist.gov/ 2http://www.nist.gov/tac/ quality and none have been validated on data from NIST evaluations. In their pioneering work on automatic evaluation of summary coherence, Lapata and Barzilay (2005) provide a correlation analysis between human coherence assessments and (1) semantic relatedness between adjacent sentences and (2) measures that characterize how mentions of the same entity in different syntactic positions are spread across adjacent sentences. Several of their models exhibit a statistically significant agreement with human ratings and complement each other, yielding an even higher correlation when combined. Lapata and Barzilay (2005) and Barzilay and Lapata (2008) both show the effectiveness of entity-based coherence in evaluating summaries. However, fewer than five automatic summarizers were used in these studies. Further, both sets of experiments perform evaluations of mixed sets of human-produced and machine-produced summaries, so the results may be influenced by the ease of discriminating between a human and machine written summary. Therefore, we believe it is an open question how well these features predict the quality of automatically generated summaries. In this work, we focus on linguistic quality evaluation for automatic systems only. We analyze how well different types of features can rank good and poor machine-produced summaries. Good performance on this task is the most desired property of evaluation metrics during system development. We begin in Section 2 by reviewing the various aspects of linguistic quality that are relevant for machine-produced summaries and currently used in manual evaluations. In Section 3, we introduce and motivate diverse classes of features to capture vocabulary, sentence fluency, and local coherence properties of summaries. We evaluate the predictive power of these linguistic quality metrics by training and testing models on consecutive years of NIST evaluations (data described 544 in Section 4). We test the performance of different sets of features separately and in combination with each other (Section 5). Results are presented in Section 6, showing the robustness of each class and their abilities to reproduce human rankings of systems and summaries with high accuracy. 2 Aspects of linguistic quality We focus on the five aspects of linguistic quality that were used to evaluate summaries in DUC: grammaticality, non-redundancy, referential clarity, focus, and structure/coherence.3 For each of the questions, all summaries were manually rated on a scale from 1 to 5, in which 5 is the best. The exact definitions that were provided to the human assessors are reproduced below. Grammaticality: The summary should have no datelines, system-internal formatting, capitalization errors or obviously ungrammatical sentences (e.g., fragments, missing components) that make the text difficult to read. Non-redundancy: There should be no unnecessary repetition in the summary. Unnecessary repetition might take the form of whole sentences that are repeated, or repeated facts, or the repeated use of a noun or noun phrase (e.g., “Bill Clinton”) when a pronoun (“he”) would suffice. Referential clarity: It should be easy to identify who or what the pronouns and noun phrases in the summary are referring to. If a person or other entity is mentioned, it should be clear what their role in the story is. So, a reference would be unclear if an entity is referenced but its identity or relation to the story remains unclear. Focus: The summary should have a focus; sentences should only contain information that is related to the rest of the summary. Structure and Coherence: The summary should be wellstructured and well-organized. The summary should not just be a heap of related information, but should build from sentence to sentence to a coherent body of information about a topic. These five questions get at different aspects of what makes a well-written text. We therefore predict each aspect of linguistic quality separately. 3 Indicators of linguistic quality Multiple factors influence the linguistic quality of text in general, including: word choice, the reference form of entities, and local coherence. We extract features which serve as proxies for each of the factors mentioned above (Sections 3.1 to 3.5). In addition, we investigate some models of grammaticality (Chae and Nenkova, 2009) and coherence (Graesser et al., 2004; Soricut and Marcu, 2006; Barzilay and Lapata, 2008) from prior work (Sections 3.6 to 3.9). 3http://www-nlpir.nist.gov/projects/ duc/duc2006/quality-questions.txt All of the features we investigate can be computed automatically directly from text, but some require considerable linguistic processing. Several of our features require a syntactic parse. To extract these, all summaries were parsed by the Stanford parser (Klein and Manning, 2003). 3.1 Word choice: language models Psycholinguistic studies have shown that people read frequent words and phrases more quickly (Haberlandt and Graesser, 1985; Just and Carpenter, 1987), so the words that appear in a text might influence people’s perception of its quality. Language models (LM) are a way of computing how familiar a text is to readers using the distribution of words from a large background corpus. Bigram and trigram LMs additionally capture grammaticality of sentences using properties of local transitions between words. For this reason, LMs are widely used in applications such as generation and machine translation to guide the production of sentences. Judging from the effectiveness of LMs in these applications, we expect that they will provide a strong baseline for the evaluation of at least some of the linguistic quality aspects. We built unigram, bigram, and trigram language models with Good-Turing smoothing over the New York Times (NYT) section of the English Gigaword corpus (over 900 million words). We used the SRI Language Modeling Toolkit (Stolcke, 2002) for this purpose. For each of the three ngram language models, we include the min, max, and average log probability of the sentences contained in a summary, as well as the overall log probability of the entire summary. 3.2 Reference form: Named entities This set of features examines whether named entities have informative descriptions in the summary. We focus on named entities because they appear often in summaries of news documents and are often not known to the reader beforehand. In addition, first mentions of entities in text introduce the entity into the discourse and so must be informative and properly descriptive (Prince, 1981; Fraurud, 1990; Elsner and Charniak, 2008). We run the Stanford Named Entity Recognizer (Finkel et al., 2005) and record the number of PERSONs, ORGANIZATIONs, and LOCATIONs. First mentions to people Feature exploration on our development set found that under-specified 545 references to people are much more disruptive to a summary than short references to organizations or locations. In fact, prior work in Nenkova and McKeown (2003) found that summaries that have been rewritten so that first mentions of people are informative descriptions and subsequent mentions are replaced with more concise reference forms are overwhelmingly preferred to summaries whose entity references have not been rewritten. In this class, we include features that reflect the modification properties of noun phrases (NPs) in the summary that are first mentions to people. Noun phrases can include pre-modifiers, appositives, prepositional phrases, etc. Rather than prespecifying all the different ways a person expression can be modified, we hoped to discover the best patterns automatically, by including features for the average number of each Part of Speech (POS) tag occurring before, each syntactic phrase occurring before4, each POS tag occurring after, and each syntactic phrase occurring after the head of the first mention NP for a PERSON. To measure if the lack of pre or post modification is particularly detrimental, we also include the proportion of PERSON first mention NPs with no words before and with no words after the head of the NP. Summarization specific Most summarization systems today are extractive and create summaries using complete sentences from the source documents. A subsequent mention of an entity in a source document which is extracted to be the first mention of the entity in the summary is probably not informative enough. For each type of named entity (PERSON, ORGANIZATION, LOCATION), we separately record the number of instances which appear as first mentions in the summary but correspond to non-first mentions in the source documents. 3.3 Reference form: NP syntax Some summaries might not include people and other named entities at all. To measure how entities are referred to more generally, we include features about the overall syntactic patterns found in NPs: the average number of each POS tag and each syntactic phrase occurring inside NPs. 4We define a linear order based on a preorder traversal of the tree, so syntactic phrases which dominate the head are considered occurring before the head. 3.4 Local coherence: Cohesive devices In coherent text, constituent clauses and sentences are related and depend on each other for their interpretation. Referring expressions such as pronouns link the current utterance to those where the entities were previously mentioned. In addition, discourse connectives such as “but” or “because” relate propositions or events expressed by different clauses or sentences. Both these categories are known cohesive or linking devices in humanproduced text (Halliday and Hasan, 1976). The mere presence of such items in a text would be indicative of better structure and coherence. We compute a number of shallow features that provide a cheap way of capturing the above intuitions: the number of demonstratives, pronouns, and definite descriptions as well as the number of sentence-initial discourse connectives. 3.5 Local coherence: Continuity This class of linguistic quality indicators is a combination of factors related to coreference, adjacent sentence similarity, and summary-specific context of surface cohesive devices. Summarization specific Extractive multidocument summaries often lack appropriate antecedents for pronouns and proper context for the use of discourse connectives. In fact, early work in summarization (Paice, 1980; Paice, 1990) has pointed out that the presence of cohesive devices described in the previous section might in fact be the source of problems. A manual analysis of automatic summaries (Otterbacher et al., 2002) also revealed that anaphoric references that cannot be resolved and unclear discourse relations constitute more than 30% of all revisions required to manually rewrite summaries into a more coherent form. To identify these potential problems, we adapt the features for surface cohesive devices to indicate whether referring expressions and discourse connectives appear in the summary with the same context as in the input documents. For each of the cohesive devices discussed in Section 3.4—demonstratives, pronouns, definite descriptions, and sentence-initial discourse connectives—we compare the previous sentence in the summary with the previous sentence in the input article. Two features are computed for each type of cohesive device: (1) number of times the preceding sentence in the summary is the same 546 as the preceding sentence in the input and (2) the number of times the preceding sentence in summary is different from that in the input. Since the previous sentence in the input text often contains the antecedent of pronouns in the current sentence, if the previous sentence from the input is also included in the summary, the pronoun is highly likely to have a proper antecedent. We also compute the proportion of adjacent sentences in the summary that were extracted from the same input document. Coreference Steinberger et al. (2007) compare the coreference chains in input documents and in summaries in order to locate potential problems. We instead define a set of more general features related to coreference that are not specific to summarization and are applicable for any text. Our features check the existence of proper antecedents for pronouns in the summary without reference to the text of the input documents. We use the publicly available pronoun resolution system described in Charniak and Elsner (2009) to mark possible antecedents for pronouns in the summary. We then compute as features the number of times an antecedent for a pronoun was found in the previous sentence, in the same sentence, or neither. In addition, we modified the pronoun resolution system to also output the probability of the most likely antecedent and include the average antecedent probability for the pronouns in the text. Automatic coreference systems are trained on human-produced texts and we expect their accuracies to drop when applied to automatically generated summaries. However, the predictions and confidence scores still reflect whether or not possible antecedents exist in previous sentences that match in gender/number, and so may still be useful for coherence evaluation. Cosine similarity We use cosine similarity to compute the overlap of words in adjacent sentences si and si+1 as a measure of continuity. cosθ = vsi.vsi+1 ||vsi||||vsi+1|| (1) The dimensions of the two vectors (vsi and vsi+1) are the total number of word types from both sentences si and si+1. Stop words were retained. The value of each dimension for a sentence is the number of tokens of that word type in that sentence. We compute the min, max, and average value of cosine similarity over the entire summary. While some repetition is beneficial for cohesion, too much repetition leads to redundancy in the summary. Cosine similarity is thus indicative of both continuity and redundancy. 3.6 Sentence fluency: Chae and Nenkova (2009) We test the usefulness of a suite of 38 shallow syntactic features studied by Chae and Nenkova (2009). These features are weakly but significantly correlated with the fluency of machine translated sentences. These include sentence length, number of fragments, average lengths of the different types of syntactic phrases, total length of modifiers in noun phrases, and various other syntactic features. We expect that these structural features will be better at detecting ungrammatical sentences than the local language model features. Since all of these features are calculated over individual sentences, we use the average value over all the sentences in a summary in our experiments. 3.7 Coh-Metrix: Graesser et al. (2004) The Coh-Metrix tool5 provides an implementation of 54 features known in the psycholinguistic literature to correlate with the coherence of humanwritten texts (Graesser et al., 2004). These include commonly used readability metrics based on sentence length and number of syllables in constituent words. Other measures implemented in the system are surface text properties known to contribute to text processing difficulty. Also included are measures of cohesion between adjacent sentences such as similarity under a latent semantic analysis (LSA) model (Deerwester et al., 1990), stem and content word overlap, syntactic similarity between adjacent sentences, and use of discourse connectives. Coh-Metrix has been designed with the goal of capturing properties of coherent text and has been used for grade level assessment, predicting student essay grades, and various other tasks. Given the heterogeneity of features in this class, we expect that they will provide reasonable accuracies for all the linguistic quality measures. In particular, the overlap features might serve as a measure of redundancy and local coherence. 5http://cohmetrix.memphis.edu/ 547 3.8 Word coherence: Soricut and Marcu (2006) Word co-occurrence patterns across adjacent sentences provide a way of measuring local coherence that is not linguistically informed but which can be easily computed using large amounts of unannotated text (Lapata, 2003; Soricut and Marcu, 2006). Word coherence can be considered as the analog of language models at the inter-sentence level. Specifically, we used the two features introduced by Soricut and Marcu (2006). Soricut and Marcu (2006) make an analogy to machine translation: two words are likely to be translations of each other if they often appear in parallel sentences; in texts, two words are likely to signal local coherence if they often appear in adjacent sentences. The two features we computed are forward likelihood, the likelihood of observing the words in sentence si conditioned on si−1, and backward likelihood, the likelihood of observing the words in sentence si conditioned on sentence si+1. “Parallel texts” of 5 million adjacent sentences were extracted from the NYT section of GigaWord. We used the GIZA++6 implementation of IBM Model 1 to align the words in adjacent sentences and obtain all relevant probabilities. 3.9 Entity coherence: Barzilay and Lapata (2008) Linguistic theories, and Centering theory (Grosz et al., 1995) in particular, have hypothesized that the properties of the transition of attention from entities in one sentence to those in the next, play a major role in the determination of local coherence. Barzilay and Lapata (2008), inspired by Centering, proposed a method to compute the local coherence of texts on the basis of the sequences of entity mentions appearing in them. In their Entity Grid model, a text is represented by a matrix with rows corresponding to each sentence in a text, and columns to each entity mentioned anywhere in the text. The value of a cell in the grid is the entity’s grammatical role in that sentence (Subject, Object, Neither, or Absent). An entity transition is a particular entity’s role in two adjacent sentences. The actual entity coherence features are the fraction of each type of these transitions in the entire entity grid for the text. One would expect that coherent texts would contain a certain distribution of entity transitions which 6http://www.fjoch.com/GIZA++.html would differ from those in incoherent sequences. We use the Brown Coherence Toolkit7 (Elsner et al., 2007) to construct the grids. The tool does not perform full coreference resolution. Instead, noun phrases are considered to refer to the same entity if their heads are identical. Entity coherence features are the only ones that have been previously applied with success for predicting summary coherence. They can therefore be considered to be the state-of-the-art approach for automatic evaluation of linguistic quality. 4 Summarization data For our experiments, we use data from the multi-document summarization tasks of the Document Understanding Conference (DUC) workshops (Over et al., 2007). Our training and development data comes from DUC 2006 and our test data from DUC 2007. These were the most recent years in which the summaries were evaluated according to specific linguistic quality questions. Each input consists of a set of 25 related documents on a topic and the target length of summaries is 250 words. In DUC 2006, there were 50 inputs to be summarized and 35 summarization systems which participated in the evaluation. This included 34 automatic systems submitted by participants, and a baseline system that simply extracted the leading sentences from the most recent article. In DUC 2007, there were 45 inputs and 32 different summarization systems. Apart from the leading sentences baseline, a high performance automatic summarizer from a previous year was also used as a baseline. All these automatic systems are included in our evaluation experiments. 4.1 System performance on linguistic quality Each summary was evaluated according to the five linguistic quality questions introduced in Section 2: grammaticality, non-redundancy, referential clarity, focus, and structure. For each of these questions, all summaries were manually rated on a scale from 1 to 5, in which 5 is the best. The distributions of system scores in the 2006 data are shown in Figure 1. Systems are currently the worst at structure, middling at referential clarity, and relatively better at grammaticality, focus, 7http://www.cs.brown.edu/˜melsner/ manual.html 548 Figure 1: Distribution of system scores on the five linguistic quality questions Gram Non-redun Ref Focus Struct Content .02 -.40 * .29 .28 .09 Gram .38 * .25 .24 .54 * Non-redun -.07 -.09 .27 Ref .89 * .76 * Focus .80 * Table 1: Spearman correlations between the manual ratings for systems averaged over the 50 inputs in 2006; * p < .05 and non-redundancy. Structure is the aspect of linguistic quality where there is the most room for improvement. The only system with an average structure score above 3.5 in DUC 2006 was the leading sentences baseline system. As can be expected, people are unlikely to be able to focus on a single aspect of linguistic quality exclusively while ignoring the rest. Some of the linguistic quality ratings are significantly correlated with each other, particularly referential clarity, focus, and structure (Table 1). More importantly, the systems that produce summaries with good content8 are not necessarily the systems producing the most readable summaries. Notice from the first row of Table 1 that none of the system rankings based on these measures of linguistic quality are significantly positively correlated with system rankings of content. The development of automatic linguistic quality measurements will allow researchers to optimize both content and linguistic quality. 8as measured by summary responsiveness ratings on a 1 to 5 scale, without regard to linguistic quality 5 Experimental setup We use the summaries from DUC 2006 for training and feature development and DUC 2007 served as the test set. Validating the results on consecutive years of evaluation is important, as results that hold for the data in one year might not carry over to the next, as happened for example in Conroy and Dang (2008)’s work. Following Barzilay and Lapata (2008), we report summary ranking accuracy as the fraction of correct pairwise rankings in the test set. We use a Ranking SVM (SV Mlight (Joachims, 2002)) to score summaries using our features. The Ranking SVM seeks to minimize the number of discordant pairs (pairs in which the gold standard has x1 ranked strictly higher than x2, but the learner ranks x2 strictly higher than x1). The output of the ranker is always a real valued score, so a global rank order is always obtained. The default regularization parameter was used. 5.1 Combining predictions To combine information from the different feature classes, we train a meta ranker using the predictions from each class as features. First, we use a leave-one out (jackknife) procedure to get the predictions of our features for the entire 2006 data set. To predict rankings of systems on one input, we train all the individual rankers, one for each of the classes of features introduced above, on data from the remaining inputs. We then apply these rankers to the summaries produced for the held-out input. By repeating this process for each input in turn, we obtain the predicted scores for each summary. Once this is done, we use these predicted scores as features for the meta ranker, which is trained on all 2006 data. To test on a new summary pair in 2007, we first apply each individual ranker to get its predictions, and then apply the meta ranker. In either case (meta ranker or individual feature class), all training is performed on 2006 data, and all testing is done on 2007 data which guarantees the results generalize well at least from one year of evaluation to the next. 5.2 Evaluation of rankings We examine the predictive power of our features for each of the five linguistic quality questions in two settings. In system-level evaluation, we would like to rank all participating systems according to 549 their performance on the entire test set. In inputlevel evaluation, we would like to rank all summaries produced for a single given input. For input-level evaluation, the pairs are formed from summaries of the same input. Pairs in which the gold standard ratings are tied are not included. After removing the ties, the test set consists of 13K to 16K pairs for each linguistic quality question. Note that there were 45 inputs and 32 automatic systems in DUC 2007. So, there are a total of 45· 32 2 = 22, 320 possible summary pairs. For system-level evaluation, we treat the realvalued output of the SVM ranker for each summary as the linguistic quality score. The 45 individual scores for summaries produced by a given system are averaged to obtain an overall score for the system. The gold-standard system-level quality rating is equal to the average human ratings for the system’s summaries over the 45 inputs. At the system level, there are about 500 non-tied pairs in the test set for each question. For both evaluation settings, a random baseline which ranked the summaries in a random order would have an expected pairwise accuracy of 50%. 6 Results and discussion 6.1 System-level evaluation System-level accuracies for each class of features are shown in Table 2. All classes of features perform well, with at least a 20% absolute increase in accuracy over the random baseline (50% accuracy). For each of the linguistic quality questions, the corresponding best class of features gives prediction accuracies around 90%. In other words, if these features were used to fully automatically compare systems that participated in the 2007 DUC evaluation, only one out of ten comparisons would have been incorrect. These results set a high standard for future work on automatic system-level evaluation of linguistic quality. The state-of-the-art entity coherence features perform well but are not the best for any of the five aspects of linguistic quality. As expected, sentence fluency is the best feature class for grammaticality. For all four other questions, the best feature set is Continuity, which is a combination of summarization specific features, coreference features and cosine similarity of adjacent sentences. Continuity features outperform entity coherence by 3 to 4% absolute difference on referential quality, focus, and coherence. Accuracies from the language Feature set Gram. Redun. Ref. Focus Struct. Lang. models 87.6 83.0 91.2 85.2 86.3 Named ent. 78.5 83.6 82.1 74.0 69.6 NP syntax 85.0 83.8 87.0 76.6 79.2 Coh. devices 82.1 79.5 82.7 82.3 83.7 Continuity 88.8 88.5 92.9 89.2 91.4 Sent. fluency 91.7 78.9 87.6 82.3 84.9 Coh-Metrix 87.2 86.0 88.6 83.9 86.3 Word coh. 81.7 76.0 87.8 81.7 79.0 Entity coh. 90.2 88.1 89.6 85.0 87.1 Meta ranker 92.9 87.9 91.9 87.8 90.0 Table 2: System-level prediction accuracies (%) model features are within 1% of entity coherence for these three aspects of summary quality. Coh-Metrix, which has been proposed as a comprehensive characterization of text, does not perform as well as the language model and the entity coherence classes, which contain considerably fewer features related to only one aspect of text. The classes of features specific to named entities and noun phrase syntax are the weakest predictors. It is apparent from the results that continuity, entity coherence, sentence fluency and language models are the most powerful classes of features that should be used in automation of evaluation and against which novel predictors of text quality should be compared. Combining all feature classes with the meta ranker only yields higher results for grammaticality. For the other aspects of linguistic quality, it is better to use Continuity by itself to rank systems. One certainly unexpected result is that features designed to capture one aspect of well-written text turn out to perform well for other questions as well. For instance, entity coherence and continuity features predict grammaticality with very high accuracy of around 90%, and are surpassed only by the sentence fluency features. These findings warrant further investigation because we would not expect characteristics of local transitions indicative of text structure to have anything to do with sentence grammaticality or fluency. The results are probably due to the significant correlation between structure and grammaticality (Table 1). 6.2 Input-level evaluation The results of the input-level ranking experiments are shown in Table 3. Understandably, inputlevel prediction is more difficult and the results are lower compared to the system-level predictions: even with wrong predictions for some of the summaries by two systems, the overall judgment that 550 one system is better than the other over the entire test set can still be accurate. While for system-level predictions the meta ranker was only useful for grammaticality, at the input level it outperforms every individual feature class for each of the five questions, obtaining accuracies around 70%. These input-level accuracies compare favorably with automatic evaluation metrics for other natural language processing tasks. For example, at the 2008 ACL Workshop on Statistical Machine Translation, all fifteen automatic evaluation metrics, including variants of BLEU scores, achieved between 42% and 56% pairwise accuracy with human judgments at the sentence level (CallisonBurch et al., 2008). As in system-level prediction, for referential clarity, focus, and structure, the best feature class is Continuity. Sentence fluency again is the best class for identifying grammaticality. Coh-Metrix features are now best for determining redundancy. Both Coh-Metrix and Continuity (the top two features for redundancy) include overlap measures between adjacent sentences, which serve as a good proxy for redundancy. Surprisingly, the relative performance of the feature classes at input level is not the same as for system-level prediction. For example, the language model features, which are the second best class for the system-level, do not fare as well at the input-level. Word co-occurrence which obtained good accuracies at the system level is the least useful class at the input level with accuracies just above chance in all cases. 6.3 Components of continuity The class of features capturing sentence-tosentence continuity in the summary (Section 3.5) are the most effective for predicting referential clarity, focus, and structure at the input level. We now investigate to what extent each of its components–summary-specific features, coreference, and cosine similarity between adjacent sentences–contribute to performance. Results obtained after excluding each of the components of continuity is shown in Table 4; each line in the table represents Continuity minus a feature subclass. Removing cosine overlap causes the largest drop in prediction accuracy, with results about 10% lower than those for the complete Continuity class. Summary specific feaFeature set Gram. Redun. Ref. Focus Struct. Lang. models 66.3 57.6 62.2 60.5 62.5 Named ent. 52.9 54.4 60.0 54.1 52.5 NP Syntax 59.0 50.8 59.1 54.5 55.1 Coh. devices 56.8 54.4 55.2 52.7 53.6 Continuity 61.7 62.5 69.7 65.4 70.4 Sent. fluency 69.4 52.5 64.4 61.9 62.6 Coh-Metrix 65.5 67.6 67.9 63.0 62.4 Word coh. 54.7 55.5 53.3 53.2 53.7 Entity coh. 61.3 62.0 64.3 64.2 63.6 Meta ranker 71.0 68.6 73.1 67.4 70.7 Table 3: Input-level prediction accuracies (%) tures, which compare the context of a sentence in the summary with the context in the original document where it appeared, also contribute substantially to the success of the Continuity class in predicting structure and referential clarity. Accuracies drop by about 7% when these features are excluded. However, the coreference features do not seem to contribute much towards predicting summary linguistic quality. The accuracies of the Continuity class are not affected at all when these coreference features are not included. 6.4 Impact of summarization methods In this paper, we have discussed an analysis of the outputs of current research systems. Almost all of these systems still use extractive methods. The summarization specific continuity features reward systems that include the necessary preceding context from the original document. These features have high prediction accuracies (Section 6.3) of linguistic quality, however note that the supporting context could often contain less important content. Therefore, there is a tension between strategies for optimizing linguistic quality and for optimizing content, which warrants the development of abstractive methods. As the field moves towards more abstractive summaries, we expect to see differences in both a) summary linguistic quality and b) the features predictive of linguistic aspects. As discussed in Section 4.1, systems are currently worst at structure/coherence. However, grammaticality will become more of an issue as systems use sentence compression (Knight and Marcu, 2002), reference rewriting (Nenkova and McKeown, 2003), and other techniques to produce their own sentences. The number of discourse connectives is currently significantly negatively correlated with structure/coherence (Spearman correlation of r = 551 Ref. Focus Struct. Continuity 69.7 65.4 70.4 - Sum-specific 63.9 64.2 63.5 - Coref 70.1 65.2 70.6 - Cosine 60.2 56.6 60.7 Table 4: Ablation within the Continuity class; pairwise accuracy for input-level predictions (%) -.06, p = .008 on DUC 2006 system summaries). This can be explained by the fact that they often lack proper context in an extractive summary. However, an abstractive system could plan a discourse structure and insert appropriate connectives (Saggion, 2009). In this case, we would expect the presence of discourse connectives to be a mark of a well-written summary. 6.5 Results on human-written abstracts Since abstractive summaries would have markedly different properties from extracts, it would be interesting to know how well these sets of features would work for predicting the quality of machineproduced abstracts. However, since current systems are extractive, such a data set is not available. Therefore we experiment on human-written abstracts to get an estimate of the expected performance of our features on abstractive system summaries. In both DUC 2006 and DUC 2007, ten NIST assessors wrote summaries for the various inputs. There are four human-written summaries for each input and these summaries were judged on the same five linguistic quality aspects as the machine-written summaries. We train on the human-written summaries from DUC 2006 and test on the human-written summaries from DUC 2007, using the same set-up as in Section 5. These results are shown in Table 5. We only report results on the input level, as we are interested in distinguishing between the quality of the summaries, not the NIST assessors’ writing skills. Except for grammaticality, the prediction accuracies of the best feature classes for human abstracts are better than those at input level for machine extracts. This result is promising, as it shows that similar features for evaluating linguistic quality will be valid for abstractive summaries as well. Note however that the relative performance of the feature sets changes between the machine and human results. While for the machines Continuity feature class is the best predictor of referential clarity, focus, and structure (Table 3), for humans, language models and sentence fluency are best for Feature set Gram. Redun. Ref. Focus Struct. Lang. models 52.1 60.8 76.5 71.9 78.4 Named ent. 62.5 66.7 47.1 43.9 59.1 NP Syntax 64.6 49.0 43.1 49.1 58.0 Coh. devices 54.2 68.6 66.7 49.1 64.8 Continuity 54.2 49.0 62.7 61.4 71.6 Sent. fluency 54.2 64.7 80.4 71.9 72.7 Coh-Metrix 54.2 52.9 68.6 56.1 69.3 Word coh. 62.5 58.8 62.7 70.2 60.2 Entity coh. 45.8 49.0 54.9 52.6 56.8 Meta ranker 62.5 56.9 80.4 50.9 67.0 Table 5: Input-level prediction accuracies for human-written summaries (%) these three aspects of linguistic quality. A possible explanation for this difference could be that in system-produced extracts, incoherent organization influences human perception of linguistic quality to a great extent and so local coherence features turned out very predictive. But in human summaries, sentences are clearly well-organized and here, continuity features appear less useful. Sentence level fluency seems to be more predictive of the linguistic quality of these summaries. 7 Conclusion We have presented an analysis of a wide variety of features for the linguistic quality of summaries. Continuity between adjacent sentences was consistently indicative of the quality of machine generated summaries. Sentence fluency was useful for identifying grammaticality. Language model and entity coherence features also performed well and should be considered in future endeavors for automatic linguistic quality evaluation. The high prediction accuracies for input-level evaluation and the even higher accuracies for system-level evaluation confirm that questions regarding the linguistic quality of summaries can be answered reasonably using existing computational techniques. Automatic evaluation will make testing easier during system development and enable reporting results obtained outside of the cycles of NIST evaluation. Acknowledgments This material is based upon work supported under a National Science Foundation Graduate Research Fellowship and NSF CAREER award 0953445. We would like to thank Bonnie Webber for productive discussions. 552 References R. Barzilay and M. Lapata. 2008. Modeling local coherence: An entity-based approach. Computational Linguistics, 34(1):1–34. C. Callison-Burch, C. Fordyce, P. Koehn, C. Monz, and J. Schroeder. 2008. Further meta-evaluation of machine translation. In Proceedings of the Third Workshop on Statistical Machine Translation, pages 70– 106. J. Chae and A. Nenkova. 2009. Predicting the fluency of text with shallow structural features: case studies of machine translation and human-written text. In Proceedings of EACL, pages 139–147. E. Charniak and M. Elsner. 2009. EM works for pronoun anaphora resolution. In Proceedings of EACL, pages 148–156. J.M. Conroy and H.T. Dang. 2008. Mind the gap: dangers of divorcing evaluations of summary content from linguistic quality. In Proceedings of COLING, pages 145–152. S. Deerwester, S.T. Dumais, G.W. Furnas, T.K. Landauer, and R. Harshman. 1990. Indexing by latent semantic analysis. Journal of the American Society for Information Science, 41:391–407. M. Elsner and E. Charniak. 2008. Coreferenceinspired coherence modeling. In Proceedings of ACL/HLT: Short Papers, pages 41–44. M. Elsner, J. Austerweil, and E. Charniak. 2007. A unified local and global model for discourse coherence. In Proceedings of NAACL/HLT. J.R. Finkel, T. Grenager, and C. Manning. 2005. Incorporating non-local information into information extraction systems by gibbs sampling. In Proceedings of ACL, pages 363–370. K. Fraurud. 1990. Definiteness and the processing of noun phrases in natural discourse. Journal of Semantics, 7(4):395. A.C. Graesser, D.S. McNamara, M.M. Louwerse, and Z. Cai. 2004. Coh-Metrix: Analysis of text on cohesion and language. Behavior Research Methods Instruments and Computers, 36(2):193–202. B. Grosz, A. Joshi, and S. Weinstein. 1995. Centering: a framework for modelling the local coherence of discourse. Computational Linguistics, 21(2):203– 226. K.F. Haberlandt and A.C. Graesser. 1985. Component processes in text comprehension and some of their interactions. Journal of Experimental Psychology: General, 114(3):357–374. M.A.K. Halliday and R. Hasan. 1976. Cohesion in English. Longman Group Ltd, London, U.K. T. Joachims. 2002. Optimizing search engines using clickthrough data. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 133–142. M.A. Just and P.A. Carpenter. 1987. The psychology of reading and language comprehension. Allyn and Bacon Boston, MA. D. Klein and C.D. Manning. 2003. Accurate unlexicalized parsing. In Proceedings of ACL, pages 423– 430. K. Knight and D. Marcu. 2002. Summarization beyond sentence extraction: A probabilistic approach to sentence compression. Artificial Intelligence, 139(1):91–107. M. Lapata and R. Barzilay. 2005. Automatic evaluation of text coherence: Models and representations. In International Joint Conference On Artificial Intelligence, volume 19, page 1085. M. Lapata. 2003. Probabilistic text structuring: Experiments with sentence ordering. In Proceedings of ACL, pages 545–552. C.Y. Lin and E. Hovy. 2003. Automatic evaluation of summaries using n-gram co-occurrence statistics. In Proceedings of NAACL/HLT, page 78. C.Y. Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Proceedings of the Workshop on Text Summarization Branches Out (WAS 2004), pages 25–26. A. Nenkova and K. McKeown. 2003. References to named entities: a corpus study. In Proceedings of HLT/NAACL 2003 (short paper). J. Otterbacher, D. Radev, and A. Luo. 2002. Revisions that improve cohesion in multi-document summaries: a preliminary study. In Proceedings of the Workshop on Automatic Summarization, ACL. P. Over, H. Dang, and D. Harman. 2007. Duc in context. Information Processing Management, 43(6):1506–1520. C.D. Paice. 1980. The automatic generation of literature abstracts: an approach based on the identification of self-indicating phrases. In Proceedings of the 3rd annual ACM conference on Research and development in information retrieval, pages 172–191. C.D. Paice. 1990. Constructing literature abstracts by computer: Techniques and prospects. Information Processing Management, 26(1):171–186. E.F. Prince. 1981. Toward a taxonomy of given-new information. Radical pragmatics, 223:255. H. Saggion. 2009. A Classification Algorithm for Predicting the Structure of Summaries. Proceedings of the 2009 Workshop on Language Generation and Summarisation, page 31. 553 R. Soricut and D. Marcu. 2006. Discourse generation using utility-trained coherence models. In Proceedings of ACL. J. Steinberger, M. Poesio, M.A. Kabadjov, and K. Jeek. 2007. Two uses of anaphora resolution in summarization. Information Processing Management, 43(6):1663–1680. A. Stolcke. 2002. SRILM-an extensible language modeling toolkit. In Seventh International Conference on Spoken Language Processing, volume 3. 554
|
2010
|
56
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 555–564, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Identifying Non-explicit Citing Sentences for Citation-based Summarization Vahed Qazvinian Department of EECS University of Michigan Ann Arbor, MI [email protected] Dragomir R. Radev Department of EECS and School of Information University of Michigan Ann Arbor, MI [email protected] Abstract Identifying background (context) information in scientific articles can help scholars understand major contributions in their research area more easily. In this paper, we propose a general framework based on probabilistic inference to extract such context information from scientific papers. We model the sentences in an article and their lexical similarities as a Markov Random Field tuned to detect the patterns that context data create, and employ a Belief Propagation mechanism to detect likely context sentences. We also address the problem of generating surveys of scientific papers. Our experiments show greater pyramid scores for surveys generated using such context information rather than citation sentences alone. 1 Introduction In scientific literature, scholars use citations to refer to external sources. These secondary sources are essential in comprehending the new research. Previous work has shown the importance of citations in scientific domains and indicated that citations include survey-worthy information (Siddharthan and Teufel, 2007; Elkiss et al., 2008; Qazvinian and Radev, 2008; Mohammad et al., 2009; Mei and Zhai, 2008). A citation to a paper in a scientific article may contain explicit information about the cited research. The following example is an excerpt from a CoNLL paper1 that contains information about Eisner’s work on bottom-up parsers and the notion of span in parsing: “Another use of bottom-up is due to Eisner (1996), who introduced the notion of a span.” 1Buchholz and Marsi “CoNLL-X Shared Task On Multilingual Dependency Parsing”, CoNLL 2006 However, the citation to a paper may not always include explicit information about the cited paper: “This approach is one of those described in Eisner (1996)” Although this sentence alone does not provide any information about the cited paper, it suggests that its surrounding sentences describe the proposed approach in Eisner’s paper: “... In an all pairs approach, every possible pair of two tokens in a sentence is considered and some score is assigned to the possibility of this pair having a (directed) dependency relation. Using that information as building blocks, the parser then searches for the best parse for the sentence. This approach is one of those described in Eisner (1996).” We refer to such implicit citations that contain information about a specific secondary source but do not explicitly cite it, as sentences with context information or context sentences for short. We look at the patterns that such sentences create and observe that context sentences occur withing a small neighborhood of explicit citations. We also discuss the problem of extracting context sentences for a source-reference article pair. We propose a general framework that looks at each sentence as a random variable whose value determines its state about the target paper. In summary, our proposed model is based on the probabilistic inference of these random variables using graphical models. Finally we give evidence on how such sentences can help us produce better surveys of research areas. The rest of this paper is organized as follows. Preceded by a review of prior work in Section 2, we explain the data collection and our annotation process in Section 3. Section 4 explains our methodology and is followed by experimental setup in Section 5. 555 #Refs ACL-ID Author Title Year all AAN # Sents P08-2026 McClosky & Charniak Self-Training for Biomedical Parsing 2008 12 8 102 N07-1025∗ Mihalcea Using Wikipedia for Automatic ... 2007 21 12 153 N07-3002 Wang Learning Structured Classifiers ... 2007 22 14 74 P06-1101 Snow et, al. Semantic Taxonomy Induction ... 2006 19 9 138 P06-1116 Abdalla & Teufel A Bootstrapping Approach To ... 2006 24 10 231 W06-2933 Nivre et, al. Labeled Pseudo-Projective Dependency ... 2006 27 5 84 P05-1044 Smith & Eisner Contrastive Estimation: Training Log-Linear ... 2005 30 13 262 P05-1073 Toutanova et, al. Joint Learning Improves Semantic Role Labeling 2005 14 10 185 N03-1003 Barzilay & Lee Learning To Paraphrase: An Unsupervised ... 2003 26 13 203 N03-2016∗ Kondrak et, al. Cognates Can Improve Statistical Translation ... 2003 8 5 92 Table 1: Papers chosen from AAN as source papers for the evaluation corpus, together with their publication year, number of references (in AAN) and number of sentences. Papers marked with ∗are used to calculate inter-judge agreement. 2 Prior Work Analyzing the structure of scientific articles and their relations has received a lot of attention recently. The structure of citation and collaboration networks has been studied in (Teufel et al., 2006; Newman, 2001), and summarization of scientific documents is discussed in (Teufel and Moens, 2002). In addition, there is some previous work on the importance of citation sentences. Elkiss et al, (Elkiss et al., 2008) perform a large-scale study on citations in the free PubMed Central (PMC) and show that they contain information that may not be present in abstracts. In other work, Nanba et al, (Nanba and Okumura, 1999; Nanba et al., 2004b; Nanba et al., 2004a) analyze citation sentences and automatically categorize them in order to build a tool for survey generation. The text of scientific citations has been used in previous research. Bradshaw (Bradshaw, 2002; Bradshaw, 2003) uses citations to determine the content of articles. Similarly, the text of citation sentences has been directly used to produce summaries of scientific papers in (Qazvinian and Radev, 2008; Mei and Zhai, 2008; Mohammad et al., 2009). Determining the scientific attribution of an article has also been studied before. Siddharthan and Teufel (Siddharthan and Teufel, 2007; Teufel, 2005) categorize sentences according to their role in the author’s argument into predefined classes: Own, Other, Background, Textual, Aim, Basis, Contrast. Little work has been done on automatic citation extraction from research papers. Kaplan et al, (Kaplan et al., 2009) introduces “citation-site” as a block of text in which the cited text is discussed. The mentioned work uses a machine learning method for extracting citations from research papers and evaluates the result using 4 annotated articles. In our work we use graphical models to extract context sentences. Graphical models have a number of properties and corresponding techniques and have been used before on Information Retrieval tasks. Romanello et al, (Romanello et al., 2009) use Conditional Random Fields (CRF) to extract references from unstructured text in digital libraries of classic texts. Similar work include term dependency extraction (Metzler and Croft, 2005), query expansion (Metzler and Croft, 2007), and automatic feature selection (Metzler, 2007). 3 Data The ACL Anthology Network (AAN)2 is a collection of papers from the ACL Anthology3 published in the Computational Linguistics journal and proceedings from ACL conferences and workshops and includes more than 14, 000 papers over a period of four decades (Radev et al., 2009). AAN includes the citation network of the papers in the ACL Anthology. The papers in AAN are publicly available in text format retrieved by an OCR process from the original pdf files, and are segmented into sentences. To build a corpus for our experiments we picked 10 recently published papers from various areas in NLP4, each of which had references for a total of 203 candidate paper-reference pairs. Table 1 lists these papers together with their authors, titles, publication year, number of references, number of references within AAN, and the number of sen2http://clair.si.umich.edu/clair/anthology/ 3http://www.aclweb.org/anthology-new/ 4Regardless of data selection, the methodology in this work is applicable to any of the papers in AAN. 556 L&PS&al Sentence · · · C C Jacquemin (1999) and Barzilay and McKeown (2001) identify phrase level paraphrases, while Lin and Pantel (2001) and Shinyama et al. (2002) acquire structural paraphrases encoded as templates. 1 1 These latter are the most closely related to the sentence-level paraphrases we desire, and so we focus in this section on templateinduction approaches. C 0 Lin and Pantel (2001) extract inference rules, which are related to paraphrases (for example, X wrote Y implies X is the author of Y), to improve question answering. 1 0 They assume that paths in dependency trees that take similar arguments (leaves) are close in meaning. 1 0 However, only two-argument templates are considered. 0 C Shinyama et al. (2002) also use dependency-tree information to extract templates of a limited form (in their case, determined by the underlying information extraction application). 1 1 Like us (and unlike Lin and Pantel, who employ a single large corpus), they use articles written about the same event in different newspapers as data. 1 1 Our approach shares two characteristics with the two methods just described: pattern comparison by analysis of the patterns respective arguments, and use of nonparallel corpora as a data source. 0 0 However, extraction methods are not easily extended to generation methods. 1 1 One problem is that their templates often only match small fragments of a sentence. 1 1 While this is appropriate for other applications, deciding whether to use a given template to generate a paraphrase requires information about the surrounding context provided by the entire sentence. · · · Table 2: Part of the annotation for N03-1003 with respect to two of its references “Lin and Pantel (2001)” (the first column) “Shinyama et al. (2002)” (the second column). Cs indicate explicit citations, 1s indicate implicit citations and 0s are none. tences. 3.1 Annotation Process We annotated the sentences in each paper from Table 1. Each annotation instance in our setting corresponds to a paper-reference pair, and is a vector in which each dimension corresponds to a sentence and is marked with a C if it explicitly cites the reference, and with a 1 if it implicitly talks about it. All other sentences are marked with 0s. Table 2 shows a portion of two separate annotation instances of N03-1003 corresponding to two of its references. Our annotation has resulted in 203 annotation instances each corresponding to one paper-reference pair. The goal of this work is to automatically identify all context sentences, which are marked as “1”. 3.1.1 Inter-judge Agreement We also asked a neutral annotator5 to annotate two of our datasets that are marked with ∗in Table 1. For each paper-reference pair, the annotator was provided with a vector in which explicit cita5Someone not involved with the paper but an expert in NLP. ACL-ID vector size # Annotations κ N07-1025∗ 153 21 0.889 ± 0.30 N03-2016∗ 92 8 0.853 ± 0.35 Table 3: Average κ coefficient as inter-judge agreement for annotations of two sets tions were already marked with Cs. The annotation guidelines instructed the annotator to look at each explicit citation sentence, and read up to 15 sentences before and after, then mark context sentences around that sentence with 1s. Next, the 29 annotation instances done by the external annotator were compared with the corresponding annotations that we did, and the Kappa coefficient (κ) was calculated. The κ statistic is formulated as κ = Pr(a) −Pr(e) 1 −Pr(e) where Pr(a) is the relative observed agreement among raters, and Pr(e) is the probability that annotators agree by chance if each annotator is randomly assigning categories. To calculate κ, we ignored all explicit citations (since they were provided to the external annotator) and used the binary categories (i.e., 1 for context sentences, and 0 otherwise) for all other sentences. Table 3 shows the annotation vector size (i.e., number of sentences), number of annotation instances (i.e., number of references), and average κ for each set. The average κ is above 0.85 in both cases, suggesting that the annotation process has a low degree of subjectivity and can be considered reliable. 3.2 Analysis In this section we describe our analysis. First, we look at the number of explicit citations each reference has received in a paper. Figure 1 (a) shows the histogram corresponding to this distribution. It indicates that the majority of references get cited in only 1 sentence in a scientific article, while the maximum being 9 in our collected dataset with only 1 instance (i.e., there is only 1 reference that gets cited 9 times in a paper). Moreover, the data exhibits a highly positive-skewed distribution. This is illustrated on a log-log scale in Figure 1 (b). This highly skewed distribution indicates that the majority of references get cited only once in a citing paper. The very small number of citing sentences can not make a full inventory of the contributions of the cited paper, and therefore, extracting explicit citations alone without context 557 gap size 0 1 2 4 9 10 15 16 instance 273 14 2 1 2 1 1 1 Table 4: The distribution of gaps in the annotated data sentences may result in information loss about the contributions of the cited paper. 1 2 3 4 5 6 7 8 9 0 20 40 60 80 100 120 140 cit 10 0 10 1 10 −3 10 −2 10 −1 10 0 cit p(cit) alpha = 3.13; D=0.02 a b Figure 1: (a) Histogram of the number of different citations to each reference in a paper. (b) The distribution observed for the number of different citations on a log-log scale. Next, we investigate the distance between context sentences and the closest citations. For each context sentence, we find its distance to the closets context sentence or explicit citation. Formally, we define the gap to be the number of sentences between a context sentence (marked with 1) and the closest context sentence or explicit citation (marked with either C or 1) to it. For example, the second column of Table 2 shows that there is a gap of size 1 in the 9th sentence in the set of context and citation sentences about Shinyama et al. (2002). Table 4 shows the distribution of gap sizes in the annotated data. This observation suggests that the majority of context sentences directly occur after or before a citation or another context sentence. However, it shows that gaps between sentences describing a cited paper actually exist, and a proposed method should have the capability to capture them. 4 Proposed Method In this section we propose our methodology that enables us to identify the context information of a cited paper. Particularly, the task is to assign a binary label XC to each sentence Si from a paper S, where XC = 1 shows a context sentence related to a given cited paper, C. To solve this problem we propose a systematic way to model the network level relationship between consecutive sentences. In summary, each sentence is represented with a node and is given two scores (context, noncontext), and we update these scores to be in harmony with the neighbors’ scores. A particular class of graphical models known as Markov Random Fields (MRFs) are suited for solving inference problems with uncertainty in observed data. The data is modeled as an undirected graph with two types of nodes: hidden and observed. Observed nodes represent values that are known from the data. Each hidden node xu, corresponding to an observed node yu, represents the true state underlying the observed value. The state of a hidden node is related to the value of its corresponding observed node as well as the states of its neighboring hidden nodes. The local Markov property of an MRF indicates that a variable is conditionally independent on all other variables given its neighbors: xv ⊥ ⊥xV \cl(v)|xne(v), where ne(v) is the set of neighbors of v, and cl(v) = {v} ∪ne(v) is the closed neighborhood of v. Thus, the state of a node is assumed to statistically depend only upon its hidden node and each of its neighbors, and independent of any other node in the graph given its neighbors. Dependencies in an MRF are represented using two functions: Compatibility function (ψ) and Potential function (φ). ψuv(xc, xd) shows the edge potential of an edge between two nodes u, v of classes xc and xd. Large values of ψuv would indicate a strong association between xc and xd at nodes u, v. The Potential function, φi(xc, yc), shows the statistical dependency between xc and yc at each node i assumed by the MRF model. In order to find the marginal probabilities of xis in a MRF we can use Belief Propagation (BP) (Yedidia et al., 2003). If we assume the yis are fixed and show φi(xi, yi) by φi(xi), we can find the joint probability distribution for unknown variables xi as p({x}) = 1 Z Y ij ψij(xi, xj) Y i φi(xi) In the BP algorithm a set of new variables m is introduced where mij(xj) is the message passed from i to j about what state xj should be in. Each message, mij(xj), is a vector with the same dimensionality of xj in which each dimension shows i’s opinion about j being in the corresponding class. Therefore each message could be considered as a probability distribution and its components should sum up to 1. The final belief at a 558 Figure 2: The illustration of the message updating rule. Elements that make up the message from a node i to another node j: messages from i’s neighbors, local evidence at i, and propagation function between i, j summed over all possible states of node i. node i, in the BP algorithm, is also a vector with the same dimensionality of messages, and is proportional to the local evidence as well as all messages from the node’s neighbors: bi(xi) ←kφi(xi) Y j∈ne(i) mji(xi) (1) where k is the normalization factor of the beliefs about different classes. The message passed from i to j is proportional to the propagation function between i, j, the local evidence at i, and all messages sent to i from its neighbors except j: mij(xj) ← X xi φi(xi)ψij(xi, xj) Y k∈ne(i)\j mki(xi) (2) Figure 2 illustrates the message update rule. Convergence can be determined based on a variety of criteria. It can occur when the maximum change of any message between iteration steps is less than some threshold. Convergence is guaranteed for trees but not for general graphs. However, it typically occurs in practice (McGlohon et al., 2009). Upon convergence, belief scores are determined by Equation 1. 4.1 MRF construction To find the sentences from a paper that form the context information of a given cited paper, we build an MRF in which a hidden node xi and an observed node yi correspond to each sentence. The structure of the graph associated with the MRF is dependent upon the validity of a basic assumption. This assumption indicates that the generation of a sentence (in form of its words) only (a) (b) Figure 3: The structure of the MRF constructed based on the independence of non-adjacent sentences; (a) left, each sentence is independent on all other sentences given its immediate neighbors. (b) right, sentences have dependency relationship with each other regardless of their position. depends on its surrounding sentences. Said differently, each sentence is written independently of all other sentences given a number of its neighbors. This local dependence assumption can result in a number of different MRFs, each built assuming a dependency between a sentence and all sentences within a particular distance. Figure 3 shows the structure of the two MRFs at either extreme of the local dependence assumption. In Figure 3 a, each sentence only depends on one following and one preceding sentence, while Figure 3 b shows an MRF in which sentences are dependent on each other regardless of their position. We refer to the former by BP1, and to the latter by BPn. Generally, we use BPi to denote an MRF in which each sentence is connected to i sentences before and after. ψij(xc, xd) xd = 0 xd = 1 xc = 0 0.5 0.5 xc = 1 1 −Sij Sij Table 5: The compatibility function ψ between any two nodes in the MRFs from the sentences in scientific papers 4.2 Compatibility Function The compatibility function of an MRF represents the association between the hidden node classes. A node’s belief to be in class 1 is its probability to be included in the context. The belief of a node i, about its neighbor j to be in either classes is assumed to be 0.5 if i is in class 0. In other words, if a node is not part of the context itself, we assume 559 it has no effect on its neighbors’ classes. In contrast, if i is in class 1 its belief about its neighbor j is determined by their mutual lexical similarity. If this similarity is close to 1 it indicates a stronger tie between i, j. However, if i, j are not similar, i’s probability of being in class 1, should not affect that of j’s. To formalize this assumption we use the sigmoid of the cosine similarity of two sentences to build ψ. More formally, we define S to be Sij = 1 1 + e−cosine(i,j) The sigmoid function obtains a value of 0.5 for a cosine of 0 indicating that there is no bias in the association of the two sentences. The matrix in Table 5 shows the compatibility function built based on the above arguments. 4.3 Potential Function The node potential function of an MRF can incorporate some other features observable from data. Here, the goal is to find all sentences that are about a specific cited paper, without having explicit citations. To build the node potential function of the observed nodes, we use some sentence level features. First, we use the explicit citation as an important feature of a sentence. This feature can affect the belief of the corresponding hidden node, which can in turn affect its neighbors’ beliefs. For a given paper-reference pair, we flag (with a 1) each sentence that has an explicit citation to the reference. The second set of features that we are interested in are discourse-based features. In particular we match each sentence with specific patterns and flag those that match. The first pattern is a bigram in which the first term matches any of “this; that; those; these; his; her; their; such; previous”, and the second term matches any of “work; approach; system; method; technique; result; example”. The second pattern includes all sentences that start with “this; such”. Finally, the similarity of each sentence to the reference is observable from the data and can be used as a sentence-level feature. Intuitively, if a sentence has higher similarity with the reference paper, it should have a higher potential of being in class 1 or C. The flag of each sentence here is a value between 0 and 1 and is determined by its cosine similarity to the reference. Once the flags for each sentence, Si are determined, we calculate normalized fi as the unweighted linear combination of individual features. Based on fis, we compute the potential function, φ, as shown in Table 6. φi(xc, yc) xc = 0 xc = 1 1 −fi fi Table 6: The node potential function φ for each node in the MRFs from the sentences in scientific papers is built using the sentences’ flags computed using sentence level features. 5 Experiments The intrinsic evaluation of our methodology means to directly compare the output of our method with the gold standards obtained from the annotated data. Our methodology finds the sentences that cite a reference implicitly. Therefore the output of the inference method is a vector, υ, of 1’s and 0’s, whereby a 1 at element i means that sentence i in the source document is a context sentence about the reference while a 0 means an explicit citation or neither. The gold standard for each paper-reference pair, ω (obtained from the annotated vectors in Section 3.1 by changing all Cs to 0s), is also a vector of the same format and dimensionality. Precision, recall, and Fβ for this task can be defined as p = υ · ω υ · 1 ; r = υ · ω ω · 1 ; Fβ = (1 + β2)p · r β2p + r (3) where 1 is a vector of 1’s with the same dimensionality and β is a non-negative real number. 5.1 Baseline Methods The first baseline that we use is an IR-based method. This baseline, B1, takes explicit citations as an input but use them to find context sentences. Given a paper-reference pair, for each explicit citation sentence, marked with C, B1 picks its preceding and following sentences if their similarities to that sentence is greater than a cutoff (the median of all such similarities), and repeats this for neighboring sentences of newly marked sentences. Intuitively, B1 tries to find the best chain (window) around citing sentences. As the second baseline, we use the hand-crafted discourse based features used in MRF’s potential function. Particularly, this baseline, B2, marks 560 paper B1 B2 SVM BP1 BP4 BPn P08-2026 0.441 0.237 0.249 0.470 0.613 0.285 N07-1025 0.388 0.102 0.124 0.313 0.466 0.138 N07-3002 0.521 0.339 0.232 0.742 0.627 0.315 P06-1101 0.125 0.388 0.127 0.649 0.889 0.193 P06-1116 0.283 0.104 0.100 0.307 0.341 0.130 W06-2933 0.313 0.100 0.176 0.338 0.413 0.160 P05-1044 0.225 0.100 0.060 0.172 0.586 0.094 P05-1073 0.144 0.100 0.144 0.433 0.518 0.171 N03-1003 0.245 0.249 0.126 0.523 0.466 0.125 N03-2016 0.100 0.181 0.224 0.439 0.482 0.185 Table 7: Average Fβ=3 for similarity based baseline (B1), discourse-based baseline (B2), a supervised method (SVM) and three MRF-based methods. each sentence that is within a particular distance (4 in our experiments) of an explicit citation and matches one of the two patterns mentioned in Section 4.3. After marking all such sentences, B2 also marks all sentences between them and the closest explicit citation, which is no farther than 4 sentences away. This baseline helps us understand how effectively this sentence level feature can work in the absence of other features and the network structure. Finally, we use a supervised method, SVM, to classify sentences as context/non-context. We use 4 features to train the SVM model. These 4 features comprise the 3 sentence level features used in MRF’s potential function (i.e., similarity to reference, explicit citation, matching certain regular-expressions) and a network level feature: distance to the closes explicit citation. For each source paper, P, we use all other source papers and their source-reference annotation instances to train a model. We then use this model to classify all instances in P. Although the number of references and thus source-reference pairs are different for different papers, this can be considered similar to a 10-fold cross validation scheme, since for each source paper the model is built using all source-reference pairs of all other 9 papers. We compare these baselines with 3 MRF-based systems each with a different assumption about independence of sentences. BP1 denotes an MRF in which each sentence is only connected to 1 sentence before and after. In BP4 locality is more relaxed and each sentence is connected to 4 sentences on each sides. BPn denotes an MRF in which all sentences are connected to each other regardless of their position in the paper. Table 7 shows Fβ=3 for our experiments and shows how BP4 outperforms the other methods on average. The value 4 may suggest the fact that although sentences might be independent of distant sentences, they depend on more than one sentence on each side. The final experiment we do to intrinsically evaluate the MRF-base method is to compare different sentence-level features. The first feature used to build the potential function is explicit citations. This feature does not directly affect context sentences (i.e., it affects the marginal probability of context sentences through the MRF network connections). Therefore, we do not alter this feature in comparing different features. However, we look at the effect of the second and the third features: hand-crafted regular expression-based features and similarity to the reference. For each paper, we use BP4 to perform 3 experiments: two in absence of each feature and one including all features. Figure 4 shows the average Fβ=3 for each experiment. This plot shows that the features lead to better results when used together. 6 Impact on Survey Generation We also performed an extrinsic evaluation of our context extraction methodology. Here we show how context sentences add important surveyworthy information to explicit citations. Previous work that generate surveys of scientific topics use the text of citation sentences alone (Mohammad et al., 2009; Qazvinian and Radev, 2008). Here, we show how the surveys generated using citations and their context sentences are better than those generated using citation sentences alone. We use the data from (Mohammad et al., 2009) 561 ... Naturally, our current work on question answering for the reading comprehension task is most related to those of (Hirschman et al. , 1999; Charniak et al. , 2000; Riloffand Thelen, 2000 ; Wang et al. , 2000). In fact, all of this body of work as well as ours are evaluated on the same set of test stories, and are developed (or trained) on the same development set of stories. The work of (Hirschman et al. , 1999) initiated this series of work, and it reported an accuracy of 36.3% on answering the questions in the test stories. Subsequently, the work of (Riloffand Thelen , 2000) and (Chaxniak et al. , 2000) improved the accuracy further to 39.7% and 41%, respectively. However, all of these three systems used handcrafted, deterministic rules and algorithms... ...The cross-model comparison showed that the performance ranking of these models was: U-SVM > PatternM > S-SVM > Retrieval-M. Compared with retrieval-based [Yang et al. 2003], pattern-based [Ravichandran et al. 2002 and Soubbotin et al. 2002], and deep NLP-based [Moldovan et al. 2002, Hovy et al. 2001; and Pasca et al. 2001] answer selection, machine learning techniques are more effective in constructing QA components from scratch. These techniques suffer, however, from the problem of requiring an adequate number of handtagged question-answer training pairs. It is too expensive and labor intensive to collect such training pairs for supervised machine learning techniques ... ... As expected, the definition and person-bio answer types are covered well by these resources. The web has been employed for pattern acquisition (Ravichandran et al. , 2003), document retrieval (Dumais et al. , 2002), query expansion (Yang et al. , 2003), structured information extraction, and answer validation (Magnini et al. , 2002). Some of these approaches enhance existing QA systems, while others simplify the question answering task, allowing a less complex approach to find correct answers ... Table 8: A portion of the QA survey generated by LexRank using the context information. Figure 4: Average Fβ=3 for BP4 employing different features. that contains two sets of cited papers and corresponding citing sentences, one on Question Answering (QA) with 10 papers and the other on Dependency Parsing (DP) with 16 papers. The QA set contains two different sets of nuggets extracted by experts respectively from paper abstracts and citation sentences. The DP set includes nuggets extracted only from citation sentences. We use these nugget sets, which are provided in form of regular expressions, to evaluate automatically generated summaries. To perform this experiment we needed to build a new corpus that includes context sentences. For each citation sentence, BP4 is used on the citing paper to extract the proper context. Here, we limit the context size to be 4 on each side. That is, we attach to a citing sentence any of its 4 preceding and following sentences if citation survey context survey QA CT nuggets 0.416 0.634 AB nuggets 0.397 0.594 DP CT nuggets 0.324 0.379 Table 9: Pyramid Fβ=3 scores of automatic surveys of QA and DP data. The QA surveys are evaluated using nuggets drawn from citation texts (CT), or abstracts (AB), and DP surveys are evaluated using nuggets from citation texts (CT). BP4 marks them as context sentences. Therefore, we build a new corpus in which each explicit citation sentence is replaced with the same sentence attached to at most 4 sentence on each side. After building the context corpus, we use LexRank (Erkan and Radev, 2004) to generate 2 QA and 2 DP surveys using the citation sentences only, and the new context corpus explained above. LexRank is a multidocument summarization system, which first builds a cosine similarity graph of all the candidate sentences. Once the network is built, the system finds the most central sentences by performing a random walk on the graph. We limit these surveys to be of a maximum length of 1000 words. Table 8 shows a portion of the survey generated from the QA context corpus. This example shows how context sentences add meaningful and survey-worthy information along with citation sentences. Table 9 shows the Pyramid Fβ=3 score of automatic surveys of QA and DP 562 data. The QA surveys are evaluated using nuggets drawn from citation texts (CT), or abstracts (AB), and DP surveys are evaluated using nuggets from citation texts (CT). In all evaluation instances the surveys generated with the context corpora excel at covering nuggets drawn from abstracts or citation sentences. 7 Conclusion In this paper we proposed a framework based on probabilistic inference to extract sentences that appear in the scientific literature, and which are about a secondary source, but which do not contain explicit citations to that secondary source. Our methodology is based on inference in an MRF built using the similarity of sentences and their lexical features. We show, by numerical experiments, that an MRF in which each sentence is connected to only a few adjacent sentences properly fits this problem. We also investigate the usefulness of such sentences in generating surveys of scientific literature. Our experiments on generating surveys for Question Answering and Dependency Parsing show how surveys generated using such context information along with citation sentences have higher quality than those built using citations alone. Generating fluent scientific surveys is difficult in absence of sufficient background information. Our future goal is to combine summarization and bibliometric techniques towards building automatic surveys that employ context information as an important part of the generated surveys. 8 Acknowledgments The authors would like to thank Arzucan ¨Ozg¨ur from University of Michigan for annotations. This paper is based upon work supported by the National Science Foundation grant ”iOPENER: A Flexible Framework to Support Rapid Learning in Unfamiliar Research Domains”, jointly awarded to University of Michigan and University of Maryland as IIS 0705832. Any opinions, findings, and conclusions or recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of the National Science Foundation. References Shannon Bradshaw. 2002. Reference Directed Indexing: Indexing Scientific Literature in the Context of Its Use. Ph.D. thesis, Northwestern University. Shannon Bradshaw. 2003. Reference directed indexing: Redeeming relevance for subject search in citation indexes. In Proceedings of the 7th European Conference on Research and Advanced Technology for Digital Libraries. Aaron Elkiss, Siwei Shen, Anthony Fader, G¨unes¸ Erkan, David States, and Dragomir R. Radev. 2008. Blind men and elephants: What do citation summaries tell us about a research article? Journal of the American Society for Information Science and Technology, 59(1):51–62. G¨unes¸ Erkan and Dragomir R. Radev. 2004. Lexrank: Graph-based centrality as salience in text summarization. Journal of Artificial Intelligence Research (JAIR). Dain Kaplan, Ryu Iida, and Takenobu Tokunaga. 2009. Automatic extraction of citation contexts for research paper summarization: A coreference-chain based approach. In Proceedings of the 2009 Workshop on Text and Citation Analysis for Scholarly Digital Libraries, pages 88–95, Suntec City, Singapore, August. Association for Computational Linguistics. Mary McGlohon, Stephen Bay, Markus G. Anderle, David M. Steier, and Christos Faloutsos. 2009. Snare: a link analytic system for graph labeling and risk detection. In KDD ’09: Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1265–1274. Qiaozhu Mei and ChengXiang Zhai. 2008. Generating impact-based summaries for scientific literature. In Proceedings of ACL ’08, pages 816–824. Donald Metzler and W. Bruce Croft. 2005. A markov random field model for term dependencies. In SIGIR ’05: Proceedings of the 28th annual international ACM SIGIR conference on Research and development in information retrieval, pages 472–479. Donald Metzler and W. Bruce Croft. 2007. Latent concept expansion using markov random fields. In SIGIR ’07: Proceedings of the 30th annual international ACM SIGIR conference on Research and development in information retrieval, pages 311–318. Donald A. Metzler. 2007. Automatic feature selection in the markov random field model for information retrieval. In CIKM ’07: Proceedings of the sixteenth ACM conference on Conference on information and knowledge management, pages 253–262. Saif Mohammad, Bonnie Dorr, Melissa Egan, Ahmed Hassan, Pradeep Muthukrishan, Vahed Qazvinian, Dragomir Radev, and David Zajic. 2009. Using citations to generate surveys of scientific paradigms. 563 In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 584–592, Boulder, Colorado, June. Association for Computational Linguistics. Hidetsugu Nanba and Manabu Okumura. 1999. Towards multi-paper summarization using reference information. In IJCAI1999, pages 926–931. Hidetsugu Nanba, Takeshi Abekawa, Manabu Okumura, and Suguru Saito. 2004a. Bilingual presri: Integration of multiple research paper databases. In Proceedings of RIAO 2004, pages 195–211, Avignon, France. Hidetsugu Nanba, Noriko Kando, and Manabu Okumura. 2004b. Classification of research papers using citation links and citation types: Towards automatic review article generation. In Proceedings of the 11th SIG Classification Research Workshop, pages 117–134, Chicago, USA. Mark E. J. Newman. 2001. The structure of scientific collaboration networks. PNAS, 98(2):404–409. Vahed Qazvinian and Dragomir R. Radev. 2008. Scientific paper summarization using citation summary networks. In COLING 2008, Manchester, UK. Dragomir R. Radev, Pradeep Muthukrishnan, and Vahed Qazvinian. 2009. The ACL anthology network corpus. In ACL workshop on Natural Language Processing and Information Retrieval for Digital Libraries. Matteo Romanello, Federico Boschetti, and Gregory Crane. 2009. Citations in the digital library of classics: Extracting canonical references by using conditional random fields. In Proceedings of the 2009 Workshop on Text and Citation Analysis for Scholarly Digital Libraries, pages 80–87, Suntec City, Singapore, August. Association for Computational Linguistics. Advaith Siddharthan and Simone Teufel. 2007. Whose idea was this, and why does it matter? attributing scientific work to citations. In Proceedings of NAACL/HLT-07. Simone Teufel and Marc Moens. 2002. Summarizing scientific articles: experiments with relevance and rhetorical status. Comput. Linguist., 28(4):409–445. Simone Teufel, Advaith Siddharthan, and Dan Tidhar. 2006. Automatic classification of citation function. In Proceedings of the EMNLP, Sydney, Australia, July. Simone Teufel. 2005. Argumentative Zoning for Improved Citation Indexing. Computing Attitude and Affect in Text: Theory and Applications, pages 159– 170. Jonathan S. Yedidia, William T. Freeman, and Yair Weiss. 2003. Understanding belief propagation and its generalizations. pages 239–269. 564
|
2010
|
57
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 565–574, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Automatic Generation of Story Highlights Kristian Woodsend and Mirella Lapata School of Informatics, University of Edinburgh Edinburgh EH8 9AB, United Kingdom [email protected], [email protected] Abstract In this paper we present a joint content selection and compression model for single-document summarization. The model operates over a phrase-based representation of the source document which we obtain by merging information from PCFG parse trees and dependency graphs. Using an integer linear programming formulation, the model learns to select and combine phrases subject to length, coverage and grammar constraints. We evaluate the approach on the task of generating “story highlights”—a small number of brief, self-contained sentences that allow readers to quickly gather information on news stories. Experimental results show that the model’s output is comparable to human-written highlights in terms of both grammaticality and content. 1 Introduction Summarization is the process of condensing a source text into a shorter version while preserving its information content. Humans summarize on a daily basis and effortlessly, but producing high quality summaries automatically remains a challenge. The difficulty lies primarily in the nature of the task which is complex, must satisfy many constraints (e.g., summary length, informativeness, coherence, grammaticality) and ultimately requires wide-coverage text understanding. Since the latter is beyond the capabilities of current NLP technology, most work today focuses on extractive summarization, where a summary is created simply by identifying and subsequently concatenating the most important sentences in a document. Without a great deal of linguistic analysis, it is possible to create summaries for a wide range of documents. Unfortunately, extracts are often documents of low readability and text quality and contain much redundant information. This is in marked contrast with hand-written summaries which often combine several pieces of information from the original document (Jing, 2002) and exhibit many rewrite operations such as substitutions, insertions, deletions, or reorderings. Sentence compression is often regarded as a promising first step towards ameliorating some of the problems associated with extractive summarization. The task is commonly expressed as a word deletion problem. It involves creating a short grammatical summary of a single sentence, by removing elements that are considered extraneous, while retaining the most important information (Knight and Marcu, 2002). Interfacing extractive summarization with a sentence compression module could improve the conciseness of the generated summaries and render them more informative (Jing, 2000; Lin, 2003; Zajic et al., 2007). Despite the bulk of work on sentence compression and summarization (see Clarke and Lapata 2008 and Mani 2001 for overviews) only a handful of approaches attempt to do both in a joint model (Daum´e III and Marcu, 2002; Daum´e III, 2006; Lin, 2003; Martins and Smith, 2009). One reason for this might be the performance of sentence compression systems which falls short of attaining grammaticality levels of human output. For example, Clarke and Lapata (2008) evaluate a range of state-of-the-art compression systems across different domains and show that machine generated compressions are consistently perceived as worse than the human gold standard. Another reason is the summarization objective itself. If our goal is to summarize news articles, then we may be better off selecting the first n sentences of the document. This “lead” baseline may err on the side of verbosity but at least will be grammatical, and it has indeed proved extremely hard to outperform by more sophisticated methods (Nenkova, 2005). In this paper we propose a model for sum565 marization that incorporates compression into the task. A key insight in our approach is to formulate summarization as a phrase rather than sentence extraction problem. Compression falls naturally out of this formulation as only phrases deemed important should appear in the summary. Obviously, our output summaries must meet additional requirements such as sentence length, overall length, topic coverage and, importantly, grammaticality. We combine phrase and dependency information into a single data structure, which allows us to express grammaticality as constraints across phrase dependencies. We encode these constraints through the use of integer linear programming (ILP), a well-studied optimization framework that is able to search the entire solution space efficiently. We apply our model to the task of generating highlights for a single document. Examples of CNN news articles with human-authored highlights are shown in Table 1. Highlights give a brief overview of the article to allow readers to quickly gather information on stories, and usually appear as bullet points. Importantly, they represent the gist of the entire document and thus often differ substantially from the first n sentences in the article (Svore et al., 2007). They are also highly compressed, written in a telegraphic style and thus provide an excellent testbed for models that generate compressed summaries. Experimental results show that our model’s output is comparable to hand-written highlights both in terms of grammaticality and informativeness. 2 Related work Much effort in automatic summarization has been devoted to sentence extraction which is often formalized as a classification task (Kupiec et al., 1995). Given appropriately annotated training data, a binary classifier learns to predict for each document sentence if it is worth extracting. Surface-level features are typically used to single out important sentences. These include the presence of certain key phrases, the position of a sentence in the original document, the sentence length, the words in the title, the presence of proper nouns, etc. (Mani, 2001; Sparck Jones, 1999). Relatively little work has focused on extraction methods for units smaller than sentences. Jing and McKeown (2000) first extract sentences, then remove redundant phrases, and use (manual) recombination rules to produce coherent output. Wan and Paris (2008) segment sentences heuristically into clauses before extraction takes place, and show that this improves summarization quality. In the context of multiple-document summarization, heuristics have also been used to remove parenthetical information (Conroy et al., 2004; Siddharthan et al., 2004). Witten et al. (1999) (among others) extract keyphrases to capture the gist of the document, without however attempting to reconstruct sentences or generate summaries. A few previous approaches have attempted to interface sentence compression with summarization. A straightforward way to achieve this is by adopting a two-stage architecture (e.g., Lin 2003) where the sentences are first extracted and then compressed or the other way round. Other work implements a joint model where words and sentences are deleted simultaneously from a document. Using a noisy-channel model, Daum´e III and Marcu (2002) exploit the discourse structure of a document and the syntactic structure of its sentences in order to decide which constituents to drop but also which discourse units are unimportant. Martins and Smith (2009) formulate a joint sentence extraction and summarization model as an ILP. The latter optimizes an objective function consisting of two parts: an extraction component, essentially a non-greedy variant of maximal marginal relevance (McDonald, 2007), and a sentence compression component, a more compact reformulation of Clarke and Lapata (2008) based on the output of a dependency parser. Compression and extraction models are trained separately in a max-margin framework and then interpolated. In the context of multi-document summarization, Daum´e III’s (2006) vine-growth model creates summaries incrementally, either by starting a new sentence or by growing already existing ones. Our own work is closest to Martins and Smith (2009). We also develop an ILP-based compression and summarization model, however, several key differences set our approach apart. Firstly, content selection is performed at the phrase rather than sentence level. Secondly, the combination of phrase and dependency information into a single data structure is new, and important in allowing us to express grammaticality as constraints across phrase dependencies, rather than resorting to a lan566 Most blacks say MLK’s vision fulfilled, poll finds WASHINGTON (CNN) – More than two-thirds of AfricanAmericans believe Martin Luther King Jr.’s vision for race relations has been fulfilled, a CNN poll found – a figure up sharply from a survey in early 2008. The CNN-Opinion Research Corp. survey was released Monday, a federal holiday honoring the slain civil rights leader and a day before Barack Obama is to be sworn in as the first black U.S. president. The poll found 69 percent of blacks said King’s vision has been fulfilled in the more than 45 years since his 1963 ’I have a dream’ speech – roughly double the 34 percent who agreed with that assessment in a similar poll taken last March. But whites remain less optimistic, the survey found. • 69 percent of blacks polled say Martin Luther King Jr’s vision realized. • Slim majority of whites say King’s vision not fulfilled. • King gave his “I have a dream” speech in 1963. 9/11 billboard draws flak from Florida Democrats, GOP (CNN) – A Florida man is using billboards with an image of the burning World Trade Center to encourage votes for a Republican presidential candidate, drawing criticism for politicizing the 9/11 attacks. ‘Please Don’t Vote for a Democrat’ reads the type over the picture of the twin towers after hijacked airliners hit them on September, 11, 2001. Mike Meehan, a St. Cloud, Florida, businessman who paid to post the billboards in the Orlando area, said former President Clinton should have put a stop to Osama bin Laden and al Qaeda before 9/11. He said a Republican president would have done so. • Billboards use image from 9/11 to encourage GOP votes. • 9/11 image wrong for ad, say Florida political parties. • Floridian praises President Bush, says ex-President Clinton failed to stop al Qaeda. Table 1: Two example CNN news articles, showing the title and the first few paragraphs, and below, the original highlights that accompanied each story. guage model. Lastly, our model is more compact, has fewer parameters, and does not require two training procedures. Our approach bears some resemblance to headline generation (Dorr et al., 2003; Banko et al., 2000), although we output several sentences rather than a single one. Headline generation models typically extract individual words from a document to produce a very short summary, whereas we extract phrases and ensure that they are combined into grammatical sentences through our ILP constraints. Svore et al. (2007) were the first to foreground the highlight generation task which we adopt as an evaluation testbed for our model. Their approach is however a purely extractive one. Using an algorithm based on neural networks and third-party resources (e.g., news query logs and Wikipedia entries) they rank sentences and select the three highest scoring ones as story highlights. In contrast, we aim to generate rather than extract highlights. As a first step we focus on deleting extraneous material, but other more sophisticated rewrite operations (e.g., Cohn and Lapata 2009) could be incorporated into our framework. 3 The Task Given a document, we aim to produce three or four short sentences covering its main topics, much like the “Story Highlights” accompanying the (online) CNN news articles. CNN highlights are written by humans; we aim to do this automatically. Documents Highlights Sentences 37.2 ± 39.6 3.5 ± 0.5 Tokens 795.0 ± 744.8 47.0 ± 9.6 Tokens/sentence 22.4 ± 4.2 13.3 ± 1.7 Table 2: Overview statistics on the corpus of documents and highlights (mean and standard deviation). A minority of documents are transcripts of interviews and speeches, and can be very long; this accounts for the very large standard deviation. Two examples of a news story and its associated highlights, are shown in Table 1. As can be seen, the highlights are written in a compressed, almost telegraphic manner. Articles, auxiliaries and forms of the verb be are often deleted. Compression is also achieved through paraphrasing, e.g., substitutions and reorderings. For example, the document sentence “The poll found 69 percent of blacks said King’s vision has been fulfilled.” is rephrased in the highlight as “69 percent of blacks polled say Martin Luther King Jr’s vision realized.”. In general, there is a fair amount of lexical overlap between document sentences and highlights (42.44%) but the correspondence between document sentences and highlights is not always one-to-one. In the first example in Table 1, the second paragraph gives rise to two highlights. Also note that the highlights need not form a coherent summary, each of them is relatively stand-alone, and there is little co-referencing between them. 567 (a) S S CC But NP NNS whites VP VBP remain ADJP RBR less JJ optimistic , , NP DT the NN survey VP VBD found . . (b) TOP found optimistic whites nsubj remain cop less advmod ccomp survey the det nsubj Figure 1: An example phrase structure (a) and dependency (b) tree for the sentence “But whites remain less optimistic, the survey found.”. In order to train and evaluate the model presented in the following sections we created a corpus of document-highlight pairs (approximately 9,000) which we downloaded from the CNN.com website.1 The articles were randomly sampled from the years 2007–2009 and covered a wide range of topics such as business, crime, health, politics, showbiz, etc. The majority were news articles, but the set also contained a mixture of editorials, commentary, interviews and reviews. Some overview statistics of the corpus are shown in Table 2. Overall, we observe a high degree of compression both at the document and sentence level. The highlights summary tends to be ten times shorter than the corresponding article. Furthermore, individual highlights have almost half the length of document sentences. 4 Modeling The objective of our model is to create the most informative story highlights possible, subject to constraints relating to sentence length, overall summary length, topic coverage, and grammaticality. These constraints are global in their scope, and cannot be adequately satisfied by optimizing each one of them individually. Our approach therefore uses an ILP formulation which will provide a globally optimal solution, and which can be efficiently solved using standard optimization tools. Specifically, the model selects phrases from which to form the highlights, and each highlight is created from a single sentence through phrase deletion. The model operates on parse trees augmented with 1The corpus is available from http://homepages.inf. ed.ac.uk/mlap/resources/index.html. dependency labels. We first describe how we obtain this representation and then move on to discuss the model in more detail. Sentence Representation We obtain syntactic information by parsing every sentence twice, once with a phrase structure parser and once with a dependency parser. The phrase structure and dependency-based representations for the sentence “But whites remain less optimistic, the survey found.” (from Table 1) are shown in Figures 1(a) and 1(b), respectively. We then combine the output from the two parsers, by mapping the dependencies to the edges of the phrase structure tree in a greedy fashion, shown in Figure 2(a). Starting at the top node of the dependency graph, we choose a node i and a dependency arc to node j. We locate the corresponding words i and j on the phrase structure tree, and locate their nearest shared ancestor p. We assign the label of the dependency i →j to the first unlabeled edge from p to j in the phrase structure tree. Edges assigned with dependency labels are shown as dashed lines. These edges are important to our formulation, as they will be represented by binary decision variables in the ILP. Further edges from p to j, and all the edges from p to i, are marked as fixed and shown as solid lines. In this way we keep the correct ordering of leaf nodes. Finally, leaf nodes are merged into parent phrases, until each phrase node contains a minimum of two tokens, shown in Figure 2(b). Because of this minimum length rule, it is possible for a merged node to be a clause rather than a phrase, but in the subsequent description we will use the term phrase rather loosely to describe any merged leaf node. 568 (a) S S CC But NP NNS whites nsubj VP VBP remain cop ADJP RBR less advmod JJ optimistic ccomp , , NP DT the det NN survey nsubj VP VBD found . . (b) S S But whites remain less optimistic ccomp , , NP the survey nsubj VBD found . Figure 2: Dependencies are mapped onto phrase structure tree (a) and leaf nodes are merged with parent phrases (b). ILP model The merged phrase structure tree, such as shown in Figure 2(b), is the actual input to our model. Each phrase in the document is given a salience score. We obtain these scores from the output of a supervised machine learning algorithm that predicts for each phrase whether it should be included in the highlights or not (see Section 5 for details). Let S be the set of sentences in a document, P be the set of phrases, and Ps ⊂P be the set of phrases in each sentence s ∈S. T is the set of words with the highest tf.idf scores, and Pt ⊂P is the set of phrases containing the token t ∈T . Let fi denote the salience score for phrase i, determined by the machine learning algorithm, and li is its length in tokens. We use a vector of binary variables x ∈{0,1}|P| to indicate if each phrase is to be within a highlight. These are either top-level nodes in our merged tree representation, or nodes whose edge to the parent has a dependency label (the dashed lines). Referring to our example in Figure 2(b), binary variables would be allocated to the top-level S node, the child S node and the NP node. The vector of auxiliary binary variables y ∈{0,1}|S| indicates from which sentences the chosen phrases come (see Equations (1i) and (1j)). Let the sets Di ⊂P, ∀i ∈P capture the phrase dependency information for each phrase i, where each set Di contains the phrases that depend on the presence of i. Our objective function function is given in Equation (1a): it is the sum of the salience scores of all the phrases chosen to form the highlights of a given document, subject to the constraints in Equations (1b)–(1j). The latter provide a natural way of describing the requirements the output must meet. max x ∑ i∈P fixi (1a) s.t. ∑ i∈P lixi ≤LT (1b) ∑ i∈Ps lixi ≤LMys ∀s ∈S (1c) ∑ i∈Ps lixi ≥Lmys ∀s ∈S (1d) ∑ i∈Pt xi ≥1 ∀t ∈T (1e) xj →xi ∀i ∈P, j ∈Di (1f) xi →ys ∀s ∈S,i ∈Ps (1g) ∑ s∈S ys ≤NS (1h) xi ∈{0,1} ∀i ∈P (1i) ys ∈{0,1} ∀s ∈S. (1j) Constraint (1b) ensures that the generated highlights do not exceed a total budget of LT tokens. This constraint may vary depending on the application or task at hand. Highlights on a small screen device would presumably be shorter than highlights for news articles on the web. It is also possible to set the length of each highlight to be within the range [Lm,LM]. Constraints (1c) and (1d) enforce this requirement. In particular, these constraints stop highlights formed from sentences at the beginning of the document (which tend to have 569 high salience scores) from being too long. Equation (1e) is a set-covering constraint, requiring that each of the words in T appears at least once in the highlights. We assume that words with high tf.idf scores reveal to a certain extent what the document is about. Constraint (1e) ensures that some of these words will be present in the highlights. We enforce grammatical correctness through constraint (1f) which ensures that the phrase dependencies are respected. Phrases that depend on phrase i are contained in the set Di. Variable xi is true, and therefore phrase i will be included, if any of its dependents xj ∈Di are true. The phrase dependency constraints, contained in the set Di and enforced by (1f), are the result of two rules based on the typed dependency information: 1. Any child node j of the current node i, whose connecting edge i →j is of type nsubj (nominal subject), nsubjpass (passive nominal subject), dobj (direct object), pobj (preposition object), infmod (infinitival modifier), ccomp (clausal complement), xcomp (open clausal complement), measure (measure phrase modifier) and num (numeric modifier) must be included if node i is included. 2. The parent node p of the current node i must always be included if i is, unless the edge p →i is of type ccomp (clausal complement) or advcl (adverbial clause), in which case it is possible to include i without including p. Consider again the example in Figure 2(b). There are only two possible outputs from this sentence. If the phrase “the survey” is chosen, then the parent node “found” will be included, and from our first rule the ccomp phrase must also be included, which results in the output: “But whites remain less optimistic, the survey found.” If, on the other hand, the clause “But whites remain less optimistic” is chosen, then due to our second rule there is no constraint that forces the parent phrase “found” to be included in the highlights. Without other factors influencing the decision, this would give the output: “But whites remain less optimistic.” We can see from this example that encoding the possible outputs as decisions on branches of the phrase structure tree provides a more compact representation of many options than would be possible with an explicit enumeration of all possible compressions. Which output is chosen (if any) depends on the scores of the phrases involved, and the influence of the other constraints. Constraint (1g) tells the ILP to create a highlight if one of its constituent phrases is chosen. Finally, note that a maximum number of highlights NS can be set beforehand, and (1h) limits the highlights to this maximum. 5 Experimental Set-up Training We obtained phrase-based salience scores using a supervised machine learning algorithm. 210 document-highlight pairs were chosen randomly from our corpus (see Section 3). Two annotators manually aligned the highlights and document sentences. Specifically, each sentence in the document was assigned one of three alignment labels: must be in the summary (1), could be in the summary (2), and is not in the summary (3). The annotators were asked to label document sentences whose content was identical to the highlights as “must be in the summary”, sentences with partially overlapping content as “could be in the summary” and the remainder as “should not be in the summary”. Inter-annotator agreement was .82 (p < 0.01, using Spearman’s ρ rank correlation). The mapping of sentence labels to phrase labels was unsupervised: if the phrase came from a sentence labeled (1), and there was a unigram overlap (excluding stop words) between the phrase and any of the original highlights, we marked this phrase with a positive label. All other phrases were marked negative. Our feature set comprised surface features such as sentence and paragraph position information, POS tags, unigram and bigram overlap with the title, and whether high-scoring tf.idf words were present in the phrase (66 features in total). The 210 documents produced a training set of 42,684 phrases (3,334 positive and 39,350 negative). We learned the feature weights with a linear SVM, using the software SVM-OOPS (Woodsend and Gondzio, 2009). This tool gave us directly the feature weights as well as support vector values, and it allowed different penalties to be applied to positive and negative misclassifications, enabling us to compensate for the unbalanced data set. The penalty hyper-parameters chosen were the ones that gave the best F-scores, using 10-fold validation. Highlight generation We generated highlights for a test set of 600 documents. We created and 570 solved an ILP for each document. Sentences were first tokenized to separate words and punctuation, then parsed to obtain phrases and dependencies as described in Section 4 using the Stanford parser (Klein and Manning, 2003). For each phrase, features were extracted and salience scores calculated from the feature weights determined through SVM training. The distance from the SVM hyperplane represents the salience score. The ILP model (see Equation (1)) was parametrized as follows: the maximum number of highlights NS was 4, the overall limit on length LT was 75 tokens, the length of each highlight was in the range of [8,28] tokens, and the topic coverage set T contained the top 5 tf.idf words. These parameters were chosen to capture the properties seen in the majority of the training set; they were also relaxed enough to allow a feasible solution of the ILP model (with hard constraints) for all the documents in the test set. To solve the ILP model we used the ZIB Optimization Suite software (Achterberg, 2007; Koch, 2004; Wunderling, 1996). The solution was converted into highlights by concatenating the chosen leaf nodes in order. The ILP problems we created had on average 290 binary variables and 380 constraints. The mean solve time was 0.03 seconds. Summarization In order to examine the generality of our model and compare with previous work, we also evaluated our system on a vanilla summarization task. Specifically, we used the same model (trained on the CNN corpus) to generate summaries for the DUC-2002 corpus2. We report results on the entire dataset and on a subset containing 140 documents. This is the same partition used by Martins and Smith (2009) to evaluate their ILP model.3 Baselines We compared the output of our model to two baselines. The first one simply selects the “leading” three sentences from each document (without any compression). The second baseline is the output of a sentence-based ILP model, similar to our own, but simpler. The model is given in (2). The binary decision variables x ∈{0,1}|S| now represent sentences, and fi the salience score for each sentence. The objective again is to maximize the total score, but now subject only to tf.idf coverage (2b) and a limit on the number of 2http://www-nlpir.nist.gov/projects/duc/ guidelines/2002.html 3We are grateful to Andr´e Martins for providing us with details of their testing partition. highlights (2c) which we set to 3. There are no sentence length or grammaticality constraints, as there is no sentence compression. max x ∑ i∈S fixi (2a) s.t. ∑ i∈St xi ≥1 ∀t ∈T (2b) ∑ i∈S xi ≤NS (2c) xi ∈{0,1} ∀i ∈S. (2d) The SVM was trained with the same features used to obtain phrase-based salience scores, but with sentence-level labels (labels (1) and (2) positive, (3) negative). Evaluation We evaluated summarization quality using ROUGE (Lin and Hovy, 2003). For the highlight generation task, the original CNN highlights were used as the reference. We report unigram overlap (ROUGE-1) as a means of assessing informativeness and the longest common subsequence (ROUGE-L) as a means of assessing fluency. In addition, we evaluated the generated highlights by eliciting human judgments. Participants were presented with a news article and its corresponding highlights and were asked to rate the latter along three dimensions: informativeness (do the highlights represent the article’s main topics?), grammaticality (are they fluent?), and verbosity (are they overly wordy and repetitive?). The subjects used a seven point rating scale. An ideal system would receive high numbers for grammaticality and informativeness and a low number for verbosity. We randomly selected nine documents from the test set and generated highlights with our model and the sentence-based ILP baseline. We also included the original highlights as a gold standard. We thus obtained ratings for 27 (9 × 3) document-highlights pairs.4 The study was conducted over the Internet using WebExp (Keller et al., 2009) and was completed by 34 volunteers, all self reported native English speakers. With regard to the summarization task, following Martins and Smith (2009), we used ROUGE-1 and ROUGE-2 to evaluate our system’s output. We also report results with ROUGE-L. Each document in the DUC-2002 dataset is paired with 4A Latin square design ensured that subjects did not see two different highlights of the same document. 571 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Recall Precision Rouge-1 F-score Recall Precision Rouge-L F-score Score Leading-3 ILP sentence ILP phrase Figure 3: ROUGE-1 and ROUGE-L results for phrase-based ILP model and two baselines, with error bars showing 95% confidence levels. a human-authored summary (approximately 100 words) which we used as reference. 6 Results We report results on the highlight generation task in Figure 3 with ROUGE-1 and ROUGE-L (error bars indicate the 95% confidence interval). In both measures, the ILP sentence baseline has the best recall, while the ILP phrase model has the best precision (the differences are statistically significant). F-score is higher for the phrase-based system but not significantly. This can be attributed to the fact that the longer output of the sentence-based model makes the recall task easier. Average highlight lengths are shown in Table 3, and the compression rates they represent. Our phrase model achieves the highest compression rates, whereas the sentence-based model tends to select long sentences even in comparison to the lead baseline. The sentence ILP model outperforms the lead baseline with respect to recall but not precision or F-score. The phrase ILP achieves a significantly better F-score over the lead baseline with both ROUGE-1 and ROUGE-L. The results of our human evaluation study are summarized in Table 4. There was no statistically significant difference in the grammaticality between the highlights generated by the phrase ILP system and the original CNN highlights (means differences were compared using a Post-hoc Tukey test). The grammaticality of the sentence ILP was significantly higher overall as no compression took place (α < 0.05). All three s toks/s C.R. Articles 36.5 22.2 ± 4.0 100% CNN highlights 3.5 13.3 ± 1.7 5.8% ILP phrase 3.8 18.0 ± 2.9 8.4% Leading-3 3.0 25.1 ± 7.4 9.3% ILP sentence 3.0 31.3 ± 7.9 11.6% Table 3: Comparison of output lengths: number of sentences, tokens per sentence, and compression rate, for CNN articles, their highlights, the ILP phrase model, and two baselines. Model Grammar Importance Verbosity CNN highlights 4.85 4.88 3.14 ILP sentence 6.41 5.47 3.97 ILP phrase 5.53 5.05 3.38 Table 4: Average human ratings for original CNN highlights, and two ILP models. systems performed on a similar level with respect to importance (differences in the means were not significant). The highlights created by the sentence ILP were considered significantly more verbose (α < 0.05) than those created by the phrasebased system and the CNN abstractors. Overall, the highlights generated by the phrase ILP model were not significantly different from those written by humans. They capture the same content as the full sentences, albeit in a more succinct manner. Table 5 shows the output of the phrase-based system for the documents in Table 1. Our results on the complete DUC-2002 corpus are shown in Table 6. Despite the fact that our model has not been optimized for the original task of generating 100-word summaries—instead it is trained on the CNN corpus, and generates highlights—the results are comparable with the best of the original participants5 in each of the ROUGE measures. Our model is also significantly better than the lead sentences baseline. Table 7 presents our results on the same DUC-2002 partition (140 documents) used by Martins and Smith (2009). The phrase ILP model achieves a significantly better F-score (for both ROUGE-1 and ROUGE-2) over the lead baseline, the sentence ILP model, and Martins and Smith. We should point out that the latter model is not a straw man. It significantly outperforms a pipeline 5The list of participants is on page 12 of the slides available from http://duc.nist.gov/pubs/2002slides/ overview.02.pdf. 572 • More than two-thirds of African-Americans believe Martin Luther King Jr.’s vision for race relations has been fulfilled. • 69 percent of blacks said King’s vision has been fulfilled in the more than 45 years since his 1963 ‘I have a dream’ speech. • But whites remain less optimistic, the survey found. • A Florida man is using billboards with an image of the burning World Trade Center to encourage votes for a Republican presidential candidate, drawing criticism. • ‘Please Don’t Vote for a Democrat’ reads the type over the picture of the twin towers. • Mike Meehan said former President Clinton should have put a stop to Osama bin Laden and al Qaeda before 9/11. Table 5: Generated highlights for the stories in Table 1 using the phrase ILP model. Participant ROUGE-1 ROUGE-2 ROUGE-L 28 0.464 0.222 0.432 19 0.459 0.221 0.431 21 0.458 0.216 0.426 29 0.449 0.208 0.419 27 0.445 0.209 0.417 Leading-3 0.416 0.200 0.390 ILP phrase 0.454 0.213 0.428 Table 6: ROUGE results on the complete DUC-2002 corpus, including the top 5 original participants. For all results, the 95% confidence interval is ±0.008. approach that first creates extracts and then compresses them. Furthermore, as a standalone sentence compression system it yields state of the art performance, comparable to McDonald’s (2006) discriminative model and superior to Hedge Trimmer (Zajic et al., 2007), a less sophisticated deterministic system. 7 Conclusions In this paper we proposed a joint content selection and compression model for single-document summarization. A key aspect of our approach is the representation of content by phrases rather than entire sentences. Salient phrases are selected to form the summary. Grammaticality, length and coverage requirements are encoded as constraints in an integer linear program. Applying the model to the generation of “story highlights” (and single document summaries) shows that it is a viable alternative to extraction-based systems. Both ROUGE scores and the results of our human study ROUGE-1 ROUGE-2 ROUGE-L Leading-3 .400 ± .018 .184 ± .015 .374 ± .017 M&S (2009) .403 ± .076 .180 ± .076 — ILP sentence .430 ± .014 .191 ± .015 .401 ± .014 ILP phrase .445 ± .014 .200 ± .014 .419 ± .014 Table 7: ROUGE results on DUC-2002 corpus (140 documents). —: only ROUGE-1 and ROUGE-2 results are given in Martins and Smith (2009). confirm that our system manages to create summaries at a high compression rate and yet maintain the informativeness and grammaticality of a competitive extractive system. The model itself is relatively simple and knowledge-lean, and achieves good performance without reference to any resources outside the corpus collection. Future extensions are many and varied. An obvious next step is to examine how the model generalizes to other domains and text genres. Although coherence is not so much of an issue for highlights, it certainly plays a role when generating standard summaries. The ILP model can be straightforwardly augmented with discourse constraints similar to those proposed in Clarke and Lapata (2007). We would also like to generalize the model to arbitrary rewrite operations, as our results indicate that compression rates are likely to improve with more sophisticated paraphrasing. Acknowledgments We would like to thank Andreas Grothey and members of ICCS at the School of Informatics for the valuable discussions and comments throughout this work. We acknowledge the support of EPSRC through project grants EP/F055765/1 and GR/T04540/01. References Achterberg, Tobias. 2007. Constraint Integer Programming. Ph.D. thesis, Technische Universit¨at Berlin. Banko, Michele, Vibhu O. Mittal, and Michael J. Witbrock. 2000. Headline generation based on statistical translation. In Proceedings of the 38th ACL. Hong Kong, pages 318– 325. Clarke, James and Mirella Lapata. 2007. Modelling compression with discourse constraints. In Proceedings of EMNLP-CoNLL. Prague, Czech Republic, pages 1–11. Clarke, James and Mirella Lapata. 2008. Global inference for sentence compression: An integer linear programming approach. Journal of Artificial Intelligence Research 31:399–429. Cohn, Trevor and Mirella Lapata. 2009. Sentence compression as tree transduction. Journal of Artificial Intelligence Research 34:637–674. 573 Conroy, J. M., J. D. Schlesinger, J. Goldstein, and D. P. O’Leary. 2004. Left-brain/right-brain multi-document summarization. In DUC 2004 Conference Proceedings. Daum´e III, Hal. 2006. Practical Structured Learning Techniques for Natural Language Processing. Ph.D. thesis, University of Southern California. Daum´e III, Hal and Daniel Marcu. 2002. A noisy-channel model for document compression. In Proceedings of the 40th ACL. Philadelphia, PA, pages 449–456. Dorr, Bonnie, David Zajic, and Richard Schwartz. 2003. Hedge trimmer: A parse-and-trim approach to headline generation. In Proceedings of the HLT-NAACL 2003 Workshop on Text Summarization. pages 1–8. Jing, Hongyan. 2000. Sentence reduction for automatic text summarization. In Proceedings of the 6th ANLP. Seattle, WA, pages 310–315. Jing, Hongyan. 2002. Using hidden Markov modeling to decompose human-written summaries. Computational Linguistics 28(4):527–544. Jing, Hongyan and Kathleen McKeown. 2000. Cut and paste summarization. In Proceedings of the 1st NAACL. Seattle, WA, pages 178–185. Keller, Frank, Subahshini Gunasekharan, Neil Mayo, and Martin Corley. 2009. Timing accuracy of web experiments: A case study using the WebExp software package. Behavior Research Methods 41(1):1–12. Klein, Dan and Christopher D. Manning. 2003. Accurate unlexicalized parsing. In Proceedings of the 41st ACL. Sapporo, Japan, pages 423–430. Knight, Kevin and Daniel Marcu. 2002. Summarization beyond sentence extraction: a probabilistic approach to sentence compression. Artificial Intelligence 139(1):91–107. Koch, Thorsten. 2004. Rapid Mathematical Prototyping. Ph.D. thesis, Technische Universit¨at Berlin. Kupiec, Julian, Jan O. Pedersen, and Francine Chen. 1995. A trainable document summarizer. In Proceedings of SIGIR95. Seattle, WA, pages 68–73. Lin, Chin-Yew. 2003. Improving summarization performance by sentence compression — a pilot study. In Proceedings of the 6th International Workshop on Information Retrieval with Asian Languages. Sapporo, Japan, pages 1–8. Lin, Chin-Yew and Eduard H. Hovy. 2003. Automatic evaluation of summaries using n-gram co-occurrence statistics. In Proceedings of HLT NAACL. Edmonton, Canada, pages 71–78. Mani, Inderjeet. 2001. Automatic Summarization. John Benjamins Pub Co. Martins, Andr´e and Noah A. Smith. 2009. Summarization with a joint model for sentence extraction and compression. In Proceedings of the Workshop on Integer Linear Programming for Natural Language Processing. Boulder, Colorado, pages 1–9. McDonald, Ryan. 2006. Discriminative sentence compression with soft syntactic constraints. In Proceedings of the 11th EACL. Trento, Italy. McDonald, Ryan. 2007. A study of global inference algorithms in multi-document summarization. In Proceedings of the 29th ECIR. Rome, Italy. Nenkova, Ani. 2005. Automatic text summarization of newswire: Lessons learned from the Document Understanding Conference. In Proceedings of the 20th AAAI. Pittsburgh, PA, pages 1436–1441. Siddharthan, Advaith, Ani Nenkova, and Kathleen McKeown. 2004. Syntactic simplification for improving content selection in multi-document summarization. In Proceedings of the 20th International Conference on Computational Linguistics (COLING 2004). pages 896–902. Sparck Jones, Karen. 1999. Automatic summarizing: Factors and directions. In Inderjeet Mani and Mark T. Maybury, editors, Advances in Automatic Text Summarization, MIT Press, Cambridge, pages 1–33. Svore, Krysta, Lucy Vanderwende, and Christopher Burges. 2007. Enhancing single-document summarization by combining RankNet and third-party sources. In Proceedings of EMNLP-CoNLL. Prague, Czech Republic, pages 448–457. Wan, Stephen and C´ecile Paris. 2008. Experimenting with clause segmentation for text summarization. In Proceedings of the 1st TAC. Gaithersburg, MD. Witten, Ian H., Gordon Paynter, Eibe Frank, Carl Gutwin, and Craig G. Nevill-Manning. 1999. KEA: Practical automatic keyphrase extraction. In Proceedings of the 4th ACM International Conference on Digital Libraries. Berkeley, CA, pages 254–255. Woodsend, Kristian and Jacek Gondzio. 2009. Exploiting separability in large-scale linear support vector machine training. Computational Optimization and Applications . Wunderling, Roland. 1996. Paralleler und objektorientierter Simplex-Algorithmus. Ph.D. thesis, Technische Universit¨at Berlin. Zajic, David, Bonnie J. Door, Jimmy Lin, and Richard Schwartz. 2007. Multi-candidate reduction: Sentence compression as a tool for document summarization tasks. Information Processing Management Special Issue on Summarization 43(6):1549–1570. 574
|
2010
|
58
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 575–584, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Sentence and Expression Level Annotation of Opinions in User-Generated Discourse Cigdem Toprak and Niklas Jakob and Iryna Gurevych Ubiquitous Knowledge Processing Lab Computer Science Department, Technische Universit¨at Darmstadt, Hochschulstraße 10 D-64289 Darmstadt, Germany www.ukp.tu-darmstadt.de Abstract In this paper, we introduce a corpus of consumer reviews from the rateitall and the eopinions websites annotated with opinion-related information. We present a two-level annotation scheme. In the first stage, the reviews are analyzed at the sentence level for (i) relevancy to a given topic, and (ii) expressing an evaluation about the topic. In the second stage, on-topic sentences containing evaluations about the topic are further investigated at the expression level for pinpointing the properties (semantic orientation, intensity), and the functional components of the evaluations (opinion terms, targets and holders). We discuss the annotation scheme, the inter-annotator agreement for different subtasks and our observations. 1 Introduction There has been a huge interest in the automatic identification and extraction of opinions from free text in recent years. Opinion mining spans a variety of subtasks including: creating opinion word lexicons (Esuli and Sebastiani, 2006; Ding et al., 2008), identifying opinion expressions (Riloff and Wiebe, 2003; Fahrni and Klenner, 2008), identifying polarities of opinions in context (Breck et al., 2007; Wilson et al., 2005), extracting opinion targets (Hu and Liu, 2004; Zhuang et al., 2006; Cheng and Xu, 2008) and opinion holders (Kim and Hovy, 2006; Choi et al., 2005). Data-driven approaches for extracting opinion expressions, their holders and targets require reliably annotated data at the expression level. In previous research, expression level annotation of opinions was extensively investigated on newspaper articles (Wiebe et al., 2005; Wilson and Wiebe, 2005; Wilson, 2008b) and on meeting dialogs (Somasundaran et al., 2008; Wilson, 2008a). Compared to the newspaper and meeting dialog genres, little corpus-based work has been carried out for interpreting the opinions and evaluations in user-generated discourse. Due to the high popularity of Web 2.0 communities1, the amount of usergenerated discourse and the interest in the analysis of such discourse has increased over the last years. To the best of our knowledge, there are two corpora of user-generated discourse which are annotated for opinion related information at the expression level: The corpus of Hu & Liu (2004) consists of customer reviews about consumer electronics, and the corpus of Zhuang et al. (2006) consists of movie reviews. Both corpora are tailored for application specific needs, therefore, do not contain certain related information explicitly annotated in the discourse, which we consider important (see Section 2). Furthermore, none of these works provide inter-annotator agreement studies. Our goal is to create sentence and expression level annotated corpus of customer reviews which fulfills the following requirements: (1) It filters individual sentences regarding their topic relevancy and the existence of an opinion or factual information which implies an evaluation. (2) It identifies opinion expressions including the respective opinion target, opinion holder, modifiers, and anaphoric expressions if applicable. (3) The semantic orientation of the opinion expression is identified while considering negation, and the opinion expression is linked to the respective holder and target in the discourse. Such a resource would (i) enable novel applications of opinion mining such as a fine-grained identification of opinion properties, e.g. opinion modification detection including negation, and (ii) enhance opinion target extraction and the polarity assignment by linking the opinion expression with its target 1http://blog.nielsen.com/nielsenwire/ wp-content/uploads/2008/10/press_ release24.pdf 575 and providing anaphoric resolutions in discourse. We present an annotation scheme which fulfills the mentioned requirements, an inter-annotator agreement study, and discuss our observations. The rest of this paper is structured as follows: Section 2 presents the related work. In Sections 3, we describe the annotation scheme. Section 4 presents the data and the annotation study, while Section 5 summarizes the main conclusions. 2 Previous Opinion Annotated Corpora 2.1 Newspaper Articles and Meeting Dialogs Most prominent work concerning the expression level annotation of opinions is the MultiPerspective Question Answering (MPQA) corpus2 (Wiebe et al., 2005). It was extended several times over the last years, either by adding new documents or annotating new types of opinion related information (Wilson and Wiebe, 2005; Stoyanov and Cardie, 2008; Wilson, 2008b). The MPQA annotation scheme builds upon the private state notion (Quirk et al., 1985) which describes mental states including opinions, emotions, speculations and beliefs among others. The annotation scheme strives to represent the private states in terms of their functional components (i.e. experiencer holding an attitude towards a target). It consists of frames (direct subjective, expressive subjective element, objective speech event, agent, attitude, and target frames) with slots representing various attributes and properties (e.g.intensity, nested source) of the private states. Wilson (2008a) adapts and extends the concepts from the MPQA scheme to annotate subjective content in meetings (AMI corpus), and creates the AMIDA scheme. Besides subjective utterances, the AMIDA scheme contains objective polar utterances which annotates evaluations without expressing explicit opinion expressions. Somasundaran et al. (2008) proposes opinion frames for representing discourse level associations in meeting dialogs. The annotation scheme focuses on two types of opinions, sentiment and arguing. It annotates the opinion expression and target spans. The link and link type attributes associate the target with other targets in the discourse through same or alternative relations. The opinion frames are built based on the links between targets. Somasundaran et al. (2008) show that opinion frames enable a coherent interpretation of the 2http://www.cs.pitt.edu/mpqa/ opinions in discourse and discover implicit evaluations through link transitivity. Similar to Somasundaran et al. (2008), Asher et al. (2008) performs discourse level analysis of opinions. They propose a scheme which first identifies and assigns categories to the opinion segments as reporting, judgment, advice, or sentiment; and then links the opinion segments with each other via rhetorical relations including contrast, correction, support, result, or continuation. However, in contrast to our scheme and other schemes, instead of marking expression boundaries without any restriction they annotate an opinion segment only if it contains an opinion word from their lexicon, or if it has a rhetorical relation to another opinion segment. 2.2 User-generated Discourse The two annotated corpora of user-generated content and their corresponding annotation schemes are far less complex. Hu & Liu (2004) present a dataset of customer reviews for consumer electronics crawled from amazon.com. The following example shows two annotations taken from the corpus of Hu & Liu (2004): camera[+2]##This is my first digital camera and what a toy it is... size[+2][u]##it is small enough to fit easily in a coat pocket or purse. The corpus provides only target and polarity annotations, and do not contain opinion expression or opinion modifier annotations which lead to these polarity scores. The annotation scheme allows the annotation of implicit features (indicated with the the attribute [u]). Implicit features are not resolved to any actual product feature instances in discourse. In fact, the actual positions of the product features (or any anaphoric references to them) are not explicitly marked in the discourse, i.e, it is unclear to which mention of the feature the opinion refers to. In their paper on movie review mining and summarization, Zhuang et al. (2006) introduce an annotated corpus of movie reviews from the Internet Movie Database. The corpus is annotated regarding movie features and corresponding opinions. The following example shows an annotated sentence: ⟨Sentence⟩I have never encountered a movie whose supporting cast was so perfectly realized.⟨FO Fword=“supporting cast” Ftype=“PAC” Oword=“perfect” Otype=“PRO”/⟩⟨/Sentence⟩ 576 The movie features (Fword) are attributed to one of 20 predefined categories (Ftype). The opinion words (Oword) and their semantic orientations (Otype) are identified. Possible negations are directly reflected by the semantic orientation, but not explicitly labeled in the sentence. (PD) in the following example indicates that the movie feature is referenced by anaphora: ⟨Sentence⟩It is utter nonsense and insulting to my intelligence and sense of history. ⟨FO Fword=“film(PD)” Ftype=“OA” Oword=“nonsense, insulting” Otype=“CON”/⟩⟨/Sentence⟩ However, similar to the corpus of Hu & Liu (2004) the referring pronouns are not explicitly marked in discourse. It is therefore neither possible to automatically determine which pronoun creates the link if there are more than one in a sentence, nor it is denoted which antecedent, i.e. the actual mention of the feature in the discourse it relates to. 3 Annotation Scheme 3.1 Opinion versus Polar Facts The goal of the annotation scheme is to capture the evaluations regarding the topics being discussed in the consumer reviews. The evaluations in consumer reviews are either explicit expressions of opinions, or facts which imply evaluations as discussed below. Explicit expressions of opinions: Opinions are private states (Wiebe et al., 2005; Quirk et al., 1985) which are not open to objective observation or verification. In this study, we focus on the opinions stating the quality or value of an entity, experience or a proposition from one’s perspective. (1) illustrates an example of an explicit expression of an opinion. Similar to Wiebe et al. (2005), we view opinions in terms of their functional components, as opinion holders, e.g., the author in (1), holding attitudes (polarity), e.g., negative attitude indicated with the word nightmare, towards possible targets, e.g., Capella University. (1) I had a nightmare with Capella University.3 Facts implying evaluations: Besides opinions, there are facts which can be objectively verified, but still imply an evaluation of the quality or value of an entity or a proposition. For instance, consider the snippet below: 3We use authentic examples from the corpus without correcting grammatical or spelling errors. (2) In a 6-week class, I counted 3 comments from the professors directly to me and two directed to my team. (3) I found that I spent most of my time learning from my fellow students. (4) A standard response from my professors would be that of a sentence fragment. The example above provides an evaluation about the professors without stating any explicit expressions of opinions. We call such objectively verifiable, but evaluative sentences polar facts. Explicit expressions of opinions typically contain specific cues, i.e. opinion words, loaded with a positive or negative connotation (e.g., nightmare). Even when they are taken out of the context in which they appear, they evoke an evaluation. However, evaluations in polar facts can only be inferred within the context of the review. For instance, the targets of the implied evalution in the polar facts (2), (3) and (4) are the professors. However, (3) may have been perceived as a positive statement if the review was explaining how good the fellow students were or how the course enforced team work etc. The annotation scheme consists of two levels. First, the sentence level scheme analyses each sentence in terms of (i) its relevancy to the overall topic of the review, and (ii) whether it contains an evaluation (an opinion or a polar fact) about the topic. Once the on-topic sentences containing evaluations are identified, the expression level scheme first focuses either on marking the text spans of the opinion expressions (if the sentence contains an explicit expression of an opinion) or marking the targets of the polar facts (if the sentence is a polar fact). Upon marking an opinion expression span, the target and holder of the opinion is marked and linked to the marked opinion expression. Furthermore, the expression level scheme allows assigning polarities to the marked opinion expression spans and targets of the polar facts. The following subsections introduce the sentence and the expression level annotation schemes in detail with examples. 3.2 Sentence Level Annotation The sentence annotation strives to identify the sentences containing evaluations about the topic. In consumer reviews people occasionally drift off the actual topic being reviewed. For instance, as in (5) taken from a review about an online university, they tend to provide information about their background or other experiences. (5) I am very fortunate and almost right out of high school 577 Figure 1: The sentence level annotation scheme with a very average GPA and only 20; I already make above $45,000 a year as a programmer with a large health care company for over a year and have had 3 promotions up in the first year and a half. Such sentences do not provide information about the actual topic, but typically serve for justifying the user’s point of view or provide a better understanding about her circumstances. However, they are not valuable for an application aiming to extract opinions about a specific topic. Reviews given to the annotators contain meta information stating the topic, for instance, the name of the university or the service being reviewed. A markable (i.e. an annotation unit) is created for each sentence prior to the annotation process. At this level, the annotation process is therefore a sentence labeling task. The annotators are able to see the whole review, and instructed to label sentences in the context of the whole review. Figure 1 presents the sentence level scheme. Attribute names are marked with oval circles and the possible values are given in parenthesis. The following attributes are used: topic relevant attribute is labeled as yes if the sentence discusses the given topic itself or its aspects, properties or features as in examples (1)(4). Other possible values for this attribute include none given which can be chosen in the absence of meta data, or no if the sentence drifted off the topic as in example (5). opinionated attribute is labeled as yes if the sentence contains any explicit expressions of opinions about the given topic. This attribute is presented if the topic relevant attribute has been labeled as none given or yes. In other words, only the on-topic sentences are considered in this step. Examples (6)-(8) illustrate examples labeled as topic relevant=yes and opinionated=yes. (6) Many people are knocking Devry but I have seen them to be a very great school. [Topic: Devry University] (7) University of Phoenix was a surprising disappointment. [Topic: University of Phoenix] (8) Assignments were passed down, but when asked to clarify the assignment because the syllabus had contradicting, poorly worded, information, my professors regularly responded....”refer to the syllabus”....but wait, the syllabus IS the question. [Topic: University of Phoenix] polar fact attribute is labeled as yes if the sentence is a polar fact. This attribute is presented if the opinionated attribute has been labeled as no. Examples (2)-(4) demonstrate sentences labeled as topic relevant=yes, opinionated=no and polar fact=yes. polar fact polarity attribute represents the polarity of the evaluation in a polar fact sentence. The possible values for this attribute include positive, negative, both. The value both is intended for the polar fact sentences containing more than one evaluation with contradicting polarities. At the expression level analysis, the targets of the contradicting polar fact evaluations are identified distinctly and assigned polarities of positive or negative later on. Examples (9)-(11) demonstrate examples of polar fact sentences with different values of the attribute polar fact polarity. (9) There are students in the first programming class and after taking this class twice they cannot write a single line of code. [polar fact polarity=negative] (10) The same class (i.e. computer class) being teach at Ivy League schools are being offered at Devry. [polar fact polarity=positive] (11) The lectures are interactive and recorded, but you need a consent from the instructor each time. [polar fact polarity=both] 3.3 Expression Level Annotation At the expression level, we focus on the topic relevant sentences containing evaluations, i.e., sentences labeled as topic relevant=yes, opinionated=yes or topic relevant=yes, opinionated=no, polar fact=yes. If the sentence is a polar fact, then the aim is to mark the target and label the polarity of the evaluation. If the sentence is opinionated, then, the aim is to mark the opinion expression span, and label its polarity and strength (i.e. intensity), and to link it to the target and the holder. Figure 2 presents the expression level scheme. At this stage, annotators mark text spans, and are allowed to assign one of the five labels to the marked span: The polar target is used to label the targets of the evaluations implied by polar facts. The isReference attribute labels polar targets which are anaphoric references. The polar target polarity 578 Figure 2: The expression level annotation scheme attribute is used to label the polarity as positive or negative. If the isReference attribute is labeled as true, then the referent attribute appears which enables the annotator to resolve the reference to its antecedent. Consider the example sentences (12) and (13) below. The polar target in (13), written bold, is labeled as isReference=true, polar target polarity=negative. To resolve the reference, annotator first creates another polar target markable for the antecedent, namely the bold text span in (12), then, links the antecedent to the referent attribute of the polar target in (13). (12) Since classes already started, CTU told me they would extend me so that I could complete the classes and get credit once I got back. (13) What they didn’t tell me is in order to extend, I also had to be enrolled in the next semester. The target annotation represents what the opinion is about. Both polar targets and targets can be the topic of the review or different aspects, i.e. features of the topic. Similar to the polar targets, the isReference attribute allows the identification of the targets which are anaphoric references and the referent attribute links them to their antecedents in the discourse. Bold span in (14) shows an example of a target in an opinionated sentence. (14) Capella U has incredible faculty in the Harold Abel School of Psychology. The holder type represents the holder of an opinion in the discourse and is labeled in the same manner as the targets and polar targets. In consumer reviews, holders are most of the time the authors of the reviews. To ease the annotation process, the holder is not labeled when this is the author. The modifier annotation labels the lexical items, such as not, very, hardly etc., which affect the strength of an opinion or shift its polarity. Upon creation of a modifier markable, annotators are asked to choose between negation, increase, decrease for identifying the influence of the modifier on the opinion. For instance, the marked span in (15) is labeled as modifier=increase as it gives the impression that the author is really offended by the negative comments about her university. (15) I am quite honestly appauled by some of the negative comments given for Capella University on this website. The opinionexpression annotation is used to label the opinion terms in the sentence. This markable type has five attributes, three of which, i.e., modifier, holder, and target are pointer attributes to the previously defined markable types. The polarity attribute assesses the semantic orientation of the attitude, where the strength attribute marks the intensity of this attitude. The polarity and strength attributes focus solely on the marked opinionexpression span, not the whole evaluation implied in the sentence. For instance, the opinionexpression span in (16) is labeled as polarity=negative, strength=average. We infer the polarity of the evaluation only after considering the modifier, polarity and the strength attributes together. In (16), the evaluation about the target is strongly negative after considering all three attributes of the opinionexpression annotation. In (17), the polarity of the opinionexpression1 itself (complaints) is labeled as negative. It is linked to the modifier1 which is labeled as negation. Target1 (PhD journey) is linked to the opinionexpression1. The overall evaluation regarding the target1 is positive after applying the affect of the modifier1 to the polarity of the opinionexpression1, i.e., after negating the negative polarity. (16) I am quite honestly[modifier] appauled by[opinionexpression] some of the negative comments given for Capella University on this website[target]. (17) I have no[modifier1] complaints[opinionexpression1] about the entire PhD journey[target1] and highly[modifier2] recommend[opinionexpression2] this school[target2]. Finally, Figure 3 demonstrates all expression level markables created for an opinionated sentence and how they relate to each other. 579 Figure 3: Expression level annotation example 4 Annotation Study Each review has been annotated by two annotators independently according to the annotation scheme introduced above. We used the freely available MMAX24 annotation tool capable of stand-off multi-level annotations. Annotators were native speaker linguistic students. They were trained on 15 reviews after reading the annotation manual.5 In the training stage, the annotators discussed with each other if different decisions have been made and were allowed to ask questions to clarify their understanding of the scheme. Annotators had access to the review text as a whole while making their decisions. 4.1 Data The corpus consists of consumer reviews collected from the review portals rateitall6 and eopinions7. It contains reviews from two domains including online universities, e.g., Capella University, Pheonix, University of Maryland University College etc. and online services, e.g., PayPal, egroups, eTrade, eCircles etc. These two domains were selected with the project-relevant, domainspecific research goals in mind. We selected a specific topic, e.g. Pheonix, if there were more than 3 reviews written about it. Table 1 shows descriptive statistics regarding the data. We used 118 reviews containing 1151 sentences from the university domain for measuring the sentence and expression level agreements. In the following subsections, we report the inter-annotator agreement (IAA) at each level. 4http://mmax2.sourceforge.net/ 5http://www.ukp.tu-darmstadt.de/ research/data/sentiment-analysis 6http://www.rateitall.com 7http://www.epinions.com University Service All Reviews 240 234 474 Sentences 2786 6091 8877 Words 49624 102676 152300 Avg sent./rev. 11.6 26 18.7 Std. dev. sent./rev. 8.2 16 14.6 Avg. words/rev. 206.7 438.7 321.3 Std. dev. words/rev. 159.2 232.1 229.8 Table 1: Descriptive statistics about the corpus 4.2 Sentence Level Agreement Sentence level markables were already created automatically prior to the annotation, i.e., the set of annotation units were the same for both annotators. We use Cohen’s kappa (κ) (Cohen, 1960) for measuring the IAA. The sentence level annotation scheme has a hierarchical structure. A new attribute is presented based on the decision made for the previous attribute, for instance, opinionated attribute is only presented if the topic relevant attribute is labeled as yes or none given; polar fact attribute is only presented if the opinionated attribute is labeled as no etc. We calculate κ for each attribute considering only the markables which were labeled the same by both annotators in the previously required step. Table 2 shows the κ values for each attribute, the size of the markable set on which the value was calculated, and the percentage agreement. Attribute Markables Agr. κ topic relevant 1151 0.89 0.73 opinionated 682 0.80 0.61 polar fact 258 0.77 0.56 polar fact polarity 103 0.96 0.92 Table 2: Sentence level inter-annotator agreement The agreement for topic relevancy shows that it is possible to label this attribute reliably. The sentences labeled as topic relevant by both annotators correspond to 59% of all sentences, suggesting that people often drift off the topic in consumer reviews. This is usually the case when they provide information about their backgrounds or alternatives to the given topic. On the other hand, we obtain moderate agreement levels for the opinionated and polar fact attributes. 62% of the topic relevant sentences were labeled as opinionated by at least one annotator, and the rest 38% constitute the topic relevant sentences labeled as not opinionated by both annotators. Nonetheless, they still contain evaluations (polar facts), as 15% of the topic relevant sen580 tences were labeled as polar facts by both annotators. When we merge the attributes opinionated and polar fact into a single category, we obtain κ of 0.75 and a percentage agreement of 87%. Thus, we conclude that opinion-relevant sentences, either in the form of an explicit expression of opinion or a polar fact, can be labeled reliably in consumer reviews. However, there is a thin border between polar facts and explicit expressions of opinions. To the best of our knowledge, similar annotation efforts on consumer or movie reviews do not provide any agreement figures for direct comparison. However, Wiebe et al. (2005) present an annotation study where they mark textual spans for subjective expressions in a newspaper corpus. They report pairwise κ values for three annotators ranging between 0.72 - 0.84 for the sentence level subjective/objective judgments. Wiebe et al. (2005) mark subjective spans, and do not explicitly perform the sentence level labeling task. They calculate the sentence level κ values based on the existence of a subjective expression span in the sentence. Although the task definitions, approaches and the corpora have quite disparate characteristics in both studies, we obtain comparable results when we merge opinionated and polar fact categories. 4.3 Expression Level Agreement At the expression level, annotators focus only on the sentences which were labeled as opinionated or polar fact by both annotators. Annotators were instructed to mark text spans, and then, assign them the annotation types such as polar target, opinionexpression etc. (see Figure 2). For calculating the text span agreement, we use the agreement metric presented by Wiebe et al. (2005) and Somasundaran et al. (2008). This metric corresponds to the precision (P) and recall (R) metrics in information retrieval where the decisions of one annotator are treated as the system; the decisions of the other annotator are treated as the gold standard; and the overlapping spans correspond to the correctly retrieved documents. Somasundaran et al. (2008) present a discourse level annotation study in which opinion and target spans are marked and linked with each other in a meeting transcript corpus. Following Somasundaran et al. (2008), we compute three different measures for the text span agreement: (i) exact matching in which the text spans should perfectly match; (ii) lenient (relaxed) matching in which the overlap between spans is considered as a match, and (iii) subset matching in which a span has to be contained in another span in order to be considered as a match.8 Agreement naturally increases as we relax the matching constraints. However, there were no differences between the lenient and the subset agreement values. Therefore, we report only the exact and lenient matching agreement results for each annotation type in Table 3. The same agreement results for the lenient and subset matching indicates that inexact matches are still very similar to each other, i.e., at least one span is totally contained in the other. Somasundaran et al. (2008) do not report any F-measure. However, they report span agreement results in terms of precision and recall ranging between 0.44 - 0.87 for opinion spans and between 0.74 - 0.90 for the target spans. Wiebe et al. (2005) use the lenient matching approach for reporting text span agreements ranging between 0.59 - 0.81 for subjective expressions. We obtain higher agreement values for both opinion expression and target spans. We attribute this to the fact that the annotators look for opinion expression and target spans within the opinionated sentences which they agreed upon. Sentence level analysis indeed increases the reliability at the expression level. Compared to the high agreement on marking target spans, we obtain lower agreement values on marking polar target spans. We observe that it is easier to attribute explicit expressions of evaluations to topic relevant entities compared to attributing evaluations implied by experiences to specific topic relevant entities in the reviews. We calculated the agreement on identifying anaphoric references using the method introduced in (Passonneau, 2004) which utilizes Krippendorf’s α (Krippendorff, 2004) for computing reliability for coreference annotation. We considered the overlapping target and polar target spans together in this calculation, and obtained an α value of 0.29. Compared to Passonneau (α values from 0.46 to 0.74), we obtain a much lower agreement value. This may be due to the different definitions and organizations of the annotation tasks. Passonneau requires prior marking of all noun phrases (or instances which needs to be processed by the an8An example of subset matching: waste of time vs. total waste of time 581 Span Exact Lenient P R F P R F opinionexpression 0.70 0.80 0.75 0.82 0.93 0.87 modifier 0.80 0.82 0.81 0.86 0.86 0.86 target 0.80 0.81 0.80 0.91 0.90 0.91 holder 0.75 0.72 0.73 0.93 0.88 0.91 polar target 0.67 0.42 0.51 0.75 0.49 0.59 Table 3: Inter-annotator agreement on text spans at the expression level notator). Annotator’s task is to identify whether an instance refers to another marked entity in the discourse, and then, to identify corefering entity chains. However, in our annotation process annotators were tasked to identify only one entity as the referent, and was free to choose it from anywhere in the discourse. In other words, our chains contain only one entity. It is possible that both annotators performed correct resolutions, but still did not overlap with each other, as they resolve to different instances of the same entity in the discourse. We plan to further investigate reference resolution annotation discrepancies and perform corrections in the future. Some annotation types require additional attributes to be labeled after marking the span. For instance, upon marking a text span as a polar target or an opinionexpression, one has to label the polarity and strength. We consider the overlapping spans for each annotation type and use κ for reporting the agreement on these attributes. Table 4 shows the κ values. Attribute Markables Agr. κ polarity 329 0.97 0.94 strength 329 0.74 0.55 modifier 136 0.88 0.77 polar target polarity 63 0.80 0.67 Table 4: Inter-annotator agreement at the expression level We observe that the strength of the opinionexpression and the polar target polarity cannot be labeled as reliably as the polarity of the opinionexpression. 61% of the agreed upon polar targets were labeled as negative by both annotators. On the other hand, only 35% of the agreed upon opinionexpressions were labeled as negative by both annotators. There were no neutral instances. This indicates that reviewers tend to report negative experiences using polar facts, probably objectively describing what has happened, but report positive experiences with explicit opinion expressions. Distribution of the strength attribute was as follows: weak 6%, average 54%, and strong 40%. The majority of the modifiers were annotated as intensifiers (70%), while 20% of the modifiers were labeled as negation. 4.4 Discussion We analyzed the discrepancies in the annotations to gain insights about the challenges involved in various opinion related labeling tasks. At the sentence level, there were several trivial cases of disagreement, for instance, failing to recognize topic relevancy when the topic was not mentioned or referenced explicitly in the sentence, as in (18). Occasionally, annotators disagreed about whether a sentence that was written as a reaction to the other reviewers, as in (19), should be considered as topic relevant or not. Another source of disagreement included sentences similar to (20) and (21). One annotator interpreted them as universally true statements regardless of the topic, while the other attributed them to the discussed topic. (18) Go to a state university if you know whats good for you! (19) Those with sour grapes couldnt cut it, have an ax to grind, and are devoting their time to smearing the school. (20) As far as learning, you really have to WANT to learn the material. (21) On an aside, this type of education is not for the undisciplined learner. Annotators easily distinguished the evaluations at the sentence level. However, they had difficulties distinguishing between a polar fact and an opinion. For instance, both annotators agreed that the sentences (22) and (23) contain evaluations regarding the topic of the review. However, one annotator interpreted both sentences as objectively verifiable facts giving a positive impression about the school, while the other one treated them as opinions. (22) All this work in the first 2 Years! (23) The school has a reputation for making students work really hard. Sentence level annotation increases the reliability of the expression level annotation in terms of marking text spans. However, annotators often had disagreements on labeling the strength attribute. For instance, one annotator labeled the 582 opinion expression in (24) as strong, while the other one labeled it as average. We observe that it is not easy to identify trivial causes of disagreements regarding strength as its perception by each individual is highly subjective. However, most of the disagreements occurred between weak and average cases. (24) the experience that i have when i visit student finance is much like going to the dentist, except when i leave, nothing is ever fixed. We did not apply any consolidation steps during our agreement studies. However, a final version of the corpus will be produced by the third judge (one of the co-authors) by consolidating the judgements of the two annotators. 5 Conclusions We presented a corpus of consumer reviews from the rateitall and eopinions websites annotated with opinion related information. Existing opinion annotated user-generated corpora suffer from several limitations which result in difficulties for interpreting the experimental results and for performing error analysis. To name a few, they do not explicitly link the functional components of the opinions like targets, holders, or modifiers with the opinion expression; some of them do not mark opinion expression spans, none of them resolves anaphoric references in discourse. Therefore, we introduced a two level annotation scheme consisting of the sentence and expression levels, which overcomes the limitations of the existing review corpora. The sentence level annotation labels sentences for (i) relevancy to a given topic, and (ii) expressing an evaluation about the topic. Similar to (Wilson, 2008a), our annotation scheme allows capturing evaluations made with factual (objective) sentences. The expression level annotation further investigates on-topic sentences containing evaluations for pinpointing the properties (polarity, strength), and marking the functional components of the evaluations (opinion terms, modifiers, targets and holders), and linking them within a discourse. We applied the annotation scheme to the consumer review genre and presented an extensive inter-annotator study providing insights to the challenges involved in various opinion related labeling tasks in consumer reviews. Similar to the MPQA scheme, which is successfully applied to the newspaper genre, the annotation scheme treats opinions and evaluations as a composition of functional components and it is easily extendable. Therefore, we hypothesize that the scheme can also be applied to other genres with minor extensions or as it is. Finally, the corpus and the annotation manual will be made available at http://www.ukp.tu-darmstadt.de/ research/data/sentiment-analysis. Acknowledgements This research was funded partially by the German Federal Ministry of Economy and Technology under grant 01MQ07012 and partially by the German Research Foundation (DFG) as part of the Research Training Group on Feedback Based Quality Management in eLearning under grant 1223. We are very grateful to Sandra K¨ubler for her help in organizing the annotators, and to Lizhen Qu for his programming support in harvesting the data. References Nicholas Asher, Farah Benamara, and Yvette Yannick Mathieu. 2008. Distilling opinion in discourse: A preliminary study. In Coling 2008: Companion volume: Posters, pages 7–10, Manchester, UK. Eric Breck, Yejin Choi, and Claire Cardie. 2007. Identifying expressions of opinion in context. In Proceedings of the Twentieth International Joint Conference on Artificial Intelligence (IJCAI-2007), pages 2683–2688, Hyderabad, India. Xiwen Cheng and Feiyu Xu. 2008. Fine-grained opinion topic and polarity identification. In Proceedings of the 6th International Conference on Language Resources and Evaluation, pages 2710–2714, Marrekech, Morocco. Yejin Choi, Claire Cardie, Ellen Riloff, and Siddharth Patwardhan. 2005. Identifying sources of opinions with conditional random fields and extraction patterns. In HLT ’05: Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, pages 355–362, Morristown, NJ, USA. Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement, 20(1):37–46. Xiaowen Ding, Bing Liu, and Philip S. Yu. 2008. A holistic lexicon-based approach to opinion mining. In Proceedings of the International Conference on Web Search and Web Data Mining, WSDM 2008, pages 231–240, Palo Alto, California, USA. Andrea Esuli and Fabrizio Sebastiani. 2006. SentiWordNet: A publicly available lexical resource for opinion mining. In Proceedings of the 5th International Conference on Language Resources and Evaluation, pages 417–422, Genova, Italy. 583 Angela Fahrni and Manfred Klenner. 2008. Old wine or warm beer: Target-specific sentiment analysis of adjectives. In Proceedings of the Symposium on Affective Language in Human and Machine, AISB 2008 Convention, pages 60 – 63, Aberdeen, Scotland. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In KDD’04: Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 168–177, Seattle, Washington. Soo-Min Kim and Eduard Hovy. 2006. Extracting opinions, opinion holders, and topics expressed in online news media text. In Proceedings of the Workshop on Sentiment and Subjectivity in Text at the joint COLING-ACL Conference, pages 1–8, Sydney, Australia. Klaus Krippendorff. 2004. Content Analysis: An Introduction to Its Methology. Sage Publications, Thousand Oaks, California. Rebecca J. Passonneau. 2004. Computing reliability for coreference. In Proceedings of LREC, volume 4, pages 1503–1506, Lisbon. Randolph Quirk, Sidney Greenbaum, Geoffrey Leech, and Jan Svartvik. 1985. A Comprehensive Grammar of the English Language. Longman, New York. Ellen Riloff and Janyce Wiebe. 2003. Learning extraction patterns for subjective expressions. In EMNLP03: Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 105–112. Swapna Somasundaran, Josef Ruppenhofer, and Janyce Wiebe. 2008. Discourse level opinion relations: An annotation study. In In Proceedings of SIGdial Workshop on Discourse and Dialogue, pages 129– 137, Columbus, Ohio. Veselin Stoyanov and Claire Cardie. 2008. Topic identification for fine-grained opinion analysis. In Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 817–824, Manchester, UK. Janyce Wiebe, Theresa Wilson, and Claire Cardie. 2005. Annotating expressions of opinions and emotions in language. Language Resources and Evaluation, 39:165–210. Theresa Wilson and Janyce Wiebe. 2005. Annotating attributions and private states. In Proceedings of the Workshop on Frontiers in Corpus Annotations II: Pie in the Sky, pages 53–60, Ann Arbor, Michigan. Theresa Wilson, Janyce Wiebe, and Paul Hoffmann. 2005. Recognizing contextual polarity in phraselevel sentiment analysis. In HLT ’05: Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, pages 347–354, Vancouver, British Columbia, Canada. Theresa Wilson. 2008a. Annotating subjective content in meetings. In Proceedings of the Sixth International Language Resources and Evaluation (LREC’08), Marrakech, Morocco. Theresa Ann Wilson. 2008b. Fine-grained Subjectivity and Sentiment Analysis: Recognizing the Intensity, Polarity, and Attitudes of Private States. Ph.D. thesis, University of Pittsburgh. Li Zhuang, Feng Jing, and Xiao-Yan Zhu. 2006. Movie review mining and summarization. In CIKM ’06: Proceedings of the 15th ACM international conference on Information and knowledge management, pages 43–50, Arlington, Virginia, USA. 584
|
2010
|
59
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 50–59, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Structural Semantic Relatedness: A Knowledge-Based Method to Named Entity Disambiguation Xianpei Han Jun Zhao∗ National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences Beijing 100190, China {xphan,jzhao}@nlpr.ia.ac.cn ∗ Corresponding author Abstract Name ambiguity problem has raised urgent demands for efficient, high-quality named entity disambiguation methods. In recent years, the increasing availability of large-scale, rich semantic knowledge sources (such as Wikipedia and WordNet) creates new opportunities to enhance the named entity disambiguation by developing algorithms which can exploit these knowledge sources at best. The problem is that these knowledge sources are heterogeneous and most of the semantic knowledge within them is embedded in complex structures, such as graphs and networks. This paper proposes a knowledge-based method, called Structural Semantic Relatedness (SSR), which can enhance the named entity disambiguation by capturing and leveraging the structural semantic knowledge in multiple knowledge sources. Empirical results show that, in comparison with the classical BOW based methods and social network based methods, our method can significantly improve the disambiguation performance by respectively 8.7% and 14.7%. 1 Introduction Name ambiguity problem is common on the Web. For example, the name “Michael Jordan” represents more than ten persons in the Google search results. Some of them are shown below: Michael (Jeffrey) Jordan, Basketball Player Michael (I.) Jordan, Professor of Berkeley Michael (B.) Jordan, American Actor The name ambiguity has raised serious problems in many relevant areas, such as web person search, data integration, link analysis and knowledge base population. For example, in response to a person query, search engine returns a long, flat list of results containing web pages about several namesakes. The users are then forced either to refine their query by adding terms, or to browse through the search results to find the person they are seeking. Besides, an ever-increasing number of question answering and information extraction systems are coming to rely on data from multi-sources, where name ambiguity will lead to wrong answers and poor results. For example, in order to extract the birth date of the Berkeley professor Michael Jordan, a system may return the birth date of his popular namesakes, e.g., the basketball player Michael Jordan. So there is an urgent demand for efficient, high-quality named entity disambiguation methods. Currently, the common methods for named entity disambiguation include name observation clustering (Bagga and Baldwin, 1998) and entity linking with knowledge base (McNamee and Dang, 2009). In this paper, we focus on the method of name observation clustering. Given a set of observations O = {o1, o2, …, on} of the target name to be disambiguated, a named entity disambiguation system should group them into a set of clusters C = {c1, c2, …, cm}, with each resulting cluster corresponding to one specific entity. For example, consider the following four observations of Michael Jordan: 1) Michael Jordan is a researcher in Computer Science. 2) Michael Jordan plays basketball in Chicago Bulls. 3) Michael Jordan wins NBA MVP. 4) Learning in Graphical Models: Michael Jordan. A named entity disambiguation system should group the 1st and 4th Michael Jordan observations into one cluster for they both refer to the Berke50 ley professor Michael Jordan, meanwhile group the other two Michael Jordan into another cluster as they refer to another person, the Basketball Player Michael Jordan. To a human, named entity disambiguation is usually not a difficult task as he can make decisions depending on not only contextual clues, but also the prior background knowledge. For example, as shown in Figure 1, with the background knowledge that both Learning and Graphical models are the topics related to Machine learning, while Machine learning is the sub domain of Computer science, a human can easily determine that the two Michael Jordan in the 1st and 4th observations represent the same person. In the same way, a human can also easily identify that the two Michael Jordan in the 2nd and 3rd observations represent the same person. Figure 1. The exploitation of knowledge in human named entity disambiguation The development of systems which could replicate the human disambiguation ability, however, is not a trivial task because it is difficult to capture and leverage the semantic knowledge as humankind. Conventionally, the named entity disambiguation methods measure the similarity between name observations using the bag of words (BOW) model (Bagga and Baldwin (1998); Mann and Yarowsky (2006); Fleischman and Hovy (2004); Pedersen et al. (2005)), where a name observation is represented as a feature vector consisting of the contextual terms. This model measures similarity based on only the cooccurrence statistics of terms, without considering all the semantic relations like social relatedness between named entities, associative relatedness between concepts, and lexical relatedness (e.g., acronyms, synonyms) between key terms. Figure 2. Part of the link structure of Wikipedia Fortunately, in recent years, due to the evolution of Web (e.g., the Web 2.0 and the Semantic Web) and many research efforts for the construction of knowledge bases, there is an increasing availability of large-scale knowledge sources, such as Wikipedia and WordNet. These largescale knowledge sources create new opportunities for knowledge-based named entity disambiguation methods as they contain rich semantic knowledge. For example, as shown in Figure 2, the link structure of Wikipedia contains rich semantic relations between concepts. And we believe that the disambiguation performance can be greatly improved by designing algorithms which can exploit these knowledge sources at best. The problem of these knowledge sources is that they are heterogeneous (e.g., they contain different types of semantic relations and different types of concepts) and most of the semantic knowledge within them is embedded in complex structures, such as graphs and networks. For example, as shown in Figure 2, the semantic relation between Graphical Model and Computer Science is embedded in the link structure of the Wikipedia. In recent years, some research has investigated to exploit some specific semantic knowledge, such as the social connection between named entities in the Web (Kalashnikov et al. (2008), Wan et al. (2005) and Lu et al. (2007)), the ontology connection in DBLP (Hassell et al., 2006) and the semantic relations in Wikipedia (Cucerzan (2007), Han and Zhao (2009)). These knowledge-based methods, however, usually are specialized to the knowledge sources they used, so they often have the knowledge coverage problem. Furthermore, these methods can only exploit the semantic knowledge to a limited extent because they cannot take the structural semantic knowledge into consideration. To overcome the deficiencies of previous methods, this paper proposes a knowledge-based method, called Structural Semantic Relatedness (SSR), which can enhance the named entity disambiguation by capturing and leveraging the structural semantic knowledge from multiple knowledge sources. The key point of our method is a reliable semantic relatedness measure between concepts (including WordNet concepts, NEs and Wikipedia concepts), called Structural Semantic Relatedness, which can capture both the explicit semantic relations between concepts and the implicit semantic knowledge embedded in graphs and networks. In particular, we first extract the semantic relations between two concepts from a variety of knowledge sources and Computer Science Machine learning Statistics Graphical model Learning Mathematic Probability Theory 2) Michael Jordan plays basketball in Chicago Bulls. 1) Michael Jordan is a researcher in Computer Science. 4) Learning in Graphical Models: Michael Jordan 3) Michael Jordan wins NBA MVP. Machine learning 51 represent them using a graph-based model, semantic-graph. Then based on the principle that “two concepts are semantic related if they are both semantic related to the neighbor concepts of each other”, we construct our Structural Semantic Relatedness measure. In the end, we leverage the structural semantic relatedness measure for named entity disambiguation and evaluate the performance on the standard WePS data sets. The experimental results show that our SSR method can significantly outperform the traditional methods. This paper is organized as follows. Section 2 describes how to construct the structural semantic relatedness measure. Next in Section 3 we describe how to leverage the captured knowledge for named entity disambiguation. Experimental results are demonstrated in Sections 4. Section 5 briefly reviews the related work. Section 6 concludes this paper and discusses the future work. 2 The Structural Semantic Relatedness Measure In this section, we demonstrate the structural semantic relatedness measure, which can capture the structural semantic knowledge in multiple knowledge sources. Totally, there are two problems we need to address: 1) How to extract and represent the semantic relations between concepts, since there are many types of semantic relations and they may exist as different patterns (the semantic knowledge may exist as explicit semantic relations or be embedded in complex structures). 2) How to capture all the extracted semantic relations between concepts in our semantic relatedness measure. To address the above two problems, in following we first introduce how to extract the semantic relations from multiple knowledge sources; then we represent the extracted semantic relations using the semantic-graph model; finally we build our structural semantic relatedness measure. 2.1 Knowledge Sources We extract three types of semantic relations (semantic relatedness between Wikipedia concepts, lexical relatedness between WordNet concepts and social relatedness between NEs) correspondingly from three knowledge sources: Wikipedia, WordNet and NE Co-occurrence Corpus. 1. Wikipedia1, a large-scale online encyclopedia, its English version includes more than 3,000,000 concepts and new articles are added quickly and up-to-date. Wikipedia contains rich semantic knowledge in the form of hyperlinks between Wikipedia articles, such as Polysemy (disambiguation pages), Synonym (redirect pages) and Associative relation (hyperlinks between Wikipedia articles). In this paper, we extract the semantic relatedness sr between Wikipedia concepts using the method described in Milne and Witten(2008): log(max( )) log( ) ( , ) 1 log( ) log(min( , )) A B A B sr a b W A B − = − − ∩ ， where a and b are the two concepts of interest, A and B are the sets of all the concepts that are respectively linked to a and b, and W is the entire Wikipedia. For demonstration, we show the semantic relatedness between four selected concepts in Table 1. Statistics Basketball Machine learning 0.58 0.00 MVP 0.00 0.45 Table 1. The semantic relatedness table of four selected Wikipedia concepts 2. WordNet 3.02 (Fellbaum et al., 1998), a lexical knowledge source includes over 110,000 WordNet concepts (word senses about English words). Various lexical relations are recorded between WordNet concepts, such as hyponyms, holonym and synonym. The lexical relatedness lr between two WordNet concepts are measured using the Lin (1998)’s WordNet semantic similarity measure. Table 2 shows some examples of the lexical relatedness. school science university 0.67 0.10 research 0.54 0.39 Table 2. The lexical relatedness table of four selected WordNet concepts 3. NE Co-occurrence Corpus, a corpus of documents for capturing the social relatedness between named entities. According to the fuzzy set theory (Baeza-Yates et al., 1999), the degree of named entities co-occurrence in a corpus is a measure of the relatedness between them. For example, in Google search results, the “Chicago Bulls” co-occurs with “NBA” in more than 1 http://www.wikipedia.org/ 2 http:// wordnet.princeton.edu/ 52 7,900,000 web pages, while only co-occurs with “EMNLP” in less than 1,000 web pages. So the co-occurrence statistics can be used to measure the social relatedness between named entities. In this paper, given a NE Co-occurrence Corpus D, the social relatedness scr between two named entities ne1 and ne2 is measured using the Google Similarity Distance (Cilibrasi and Vitanyi, 2007): 1 2 1 2 1 2 1 2 log(max( , )) log( ) ( , ) 1 log( ) log(min( , )) D D D D scr ne ne D D D − = − − ∩ where D1 and D2 are the document sets correspondingly containing ne1 and ne2. An example of social relatedness is shown in Table 3, which is computed using the Web corpus through Google. ACL NBA EMNLP 0.61 0.00 Chicago Bulls 0.19 0.55 Table 3. The social relatedness table of four selected named entities 2.2 The Semantic-Graph Model In this section we present a graph-based representation, called semantic-graph, to model the extracted semantic relations as a graph within which the semantic relations are interconnected and transitive. Concretely, the semantic-graph is defined as follows: A semantic-graph is a weighted graph G = (V, E), where each node represents a distinct concept; and each edge between a pair of nodes represents the semantic relation between the two concepts corresponding to these nodes, with the edge weight indicating the strength of the semantic relation. For demonstration, Figure 3 shows a semanticgraph which models the semantic knowledge extracted from Wikipedia for the Michael Jordan observations in Section 1. Figure 3. An example of semantic-graph Given a set of name observations, the construction of semantic-graph takes two steps: concept extraction and concept connection. In the following we respectively describe each step. 1) Concept Extraction. In this step we extract all the concepts in the contexts of name observations and represent them as the nodes in the semantic-graph. We first gather all the N-grams (up to 8 words) and identify whether they correspond to semantically meaningful concepts: if a N-gram is contained in the WordNet, we identify it as a WordNet concept, and use its primary word sense as its semantic meaning; to find whether a N-gram is a named entity, we match it to the named entity list extracted using the openCalais API3, which contains more than 30 types of named entities, such as Person, Organization and Award; to find whether a N-gram is a Wikipedia concept, we match it to the Wikipedia anchor dictionary, then find its corresponding Wikipedia concept using the method described in (Medelyan et al, 2008). After concept identification, we filter out all the N-grams which do not correspond to the semantic meaningful concepts, such as the N-grams “learning in” and “wins NBA MVP”. The retained N-grams are identified as concepts, corresponding with their semantic meanings (a concept may have multiple semantic meaning explanation, e.g., the “MVP” has three semantic meaning, as “most valuable player, MVP” in WordNet, as the “Most Valuable Player” in Wikipedia and as a named entity of Award type). 2) Concept Connection. In this step we represent the semantic relations as the edges between nodes. That is, for each pair of extracted concepts, we identify whether there are semantic relations between them: 1) If there is only one semantic relation between them, we connect these two concepts with an edge, where the edge weight is the strength of the semantic relation; 2) If there is more than one semantic relations between them, we choose the most reliable semantic relation, i.e., we choose the semantic relation in the knowledge sources according to the order of WordNet, Wikipedia and NE Co-concurrence corpus (Suchanek et al., 2007). For example, if both Wikipedia and WordNet provide the semantic relation between MVP and NBA, we choose the semantic relation provided by WordNet. 3 http://www.opencalais.com/ Researcher Graphical Model Learning NBA MVP Basketball Chicago Bulls Computer Science 0.32 0.28 0.48 0.41 0.58 0.76 0.45 0.71 0.71 0.57 53 2.3 The Structural Semantic Relatedness Measure In this section, we describe how to capture the semantic relations between the concepts in semantic-graph using a semantic relatedness measure. Totally, the semantic knowledge between concepts is modeled in two forms: 1) The edges of semantic-graph. The edges model the direct semantic relations between concepts. We call this form of semantic knowledge as explicit semantic knowledge. 2) The structure of semantic-graph. Except for the edges, the structure of the semanticgraph also models the semantic knowledge of concepts. For example, the neighbors of a concept represent all the concepts which are explicitly semantic-related to this concept; and the paths between two concepts represent all the explicit and implicit semantic relations between them. We call this form of semantic knowledge as structural semantic knowledge, or implicit semantic knowledge. Therefore, in order to deduce a reliable semantic relatedness measure, we must take both the edges and the structure of semantic-graph into consideration. Under the semantic-graph model, the measurement of semantic relatedness between concepts equals to quantifying the similarity between nodes in a weighted graph. To simplify the description, we assign each node in semantic-graph an integer index from 1 to |V| and use this index to represent the node, then we can write the adjacency matrix of the semantic-graph G as A, where A[i,j] or Aij is the edge weight between node i and node j. The problem of quantifying the relatedness between nodes in a graph is not a new problem, e.g., the structural equivalence and structural similarity (the SimRank in Jeh and Widom (2002) and the similarity measure in Leicht et al. (2006)). However, these similarity measures are not suitable for our task, because all of them assume that the edges are uniform so that they cannot take edge weight into consideration. In order to take both the graph structure and the edge weight into account, we design the structural semantic relatedness measure by extending the measure introduced in Leicht et al. (2006). The fundamental principle behind our measure is “a node u is semantically related to another node v if its immediate neighbors are semantically related to v”. This definition is natural, for example, as shown in Figure 3, the concept Basketball and its neighbors NBA and Chicago Bulls are all semantically related to MVP. This definition is recursive, and the starting point we choose is the semantic relatedness in the edge. Thus our structural semantic relatedness has two components: the neighbor term of the previous recursive phase which captures the graph structure and the semantic relatedness which captures the edge information. Thus, the recursive form of the structural semantic relatedness Sij between the node i and the node j can be written as: i il ij lj ij l N i A S S A d λ μ ∈ = + ∑ where λ and μ control the relative importance of the two components and Ni={j | Aij > 0} is the set of the immediate neighbors of node i; j Ni d Aij i ∈ ∑ = is the degree of node i. In order to solve this formula, we introduce the following two notations: T: The relatedness transition matrix, where T[i,j]=Aij/di, indicating the transition rate of relatedness from node j to its neighbor i. S: The structural semantic relatedness matrix, where S[i,j]=Sij. Now we can turn our first form of structural semantic relatedness into the matrix form: S TS A λ μ = + By solving this equation, we can get: 1 ( ) S I T A μ λ − = − where I is the identity matrix. Since μ is a parameter which only contributes an overall scale factor to the relatedness value, we can ignore it and get the final form of the structural semantic relatedness as: 1 ( ) S I T A λ − = − Because the S is asymmetric, the finally relatedness between node i and node j is the average of Sij and Sji. The meaning of λ : The last question of our structural semantic relatedness measure is how to set the free parameter λ . To understand the meaning of λ , let us expand the similarity as a power series thus: 2 2 ( ... ...) k k S I T T T A λ λ λ = + + + + + Noting that the [Tk]ij element is the relatedness transition rate from node i to node j with path length k, we can view the λ as a penalty factor for the transition path length: by setting the λ with a value within (0, 1), a longer graph path will contribute less to the final relatedness value. The optimal value of λ is 0.6 through a learning 54 process shown in Section 4. For demonstration, Table 4 shows some structural semantic relatedness values of the Semantic-graph in Figure 3 (CS represents computer science and GM represents Graphical model). From Table 4, we can see that the structural semantic relatedness can successfully capture the semantic knowledge embedded in the structure of semantic-graph, such as the implicit semantic relation between Researcher and Learning. Researcher CS GM Learning Researcher --- 0.50 0.27 0.31 CS 0.50 --- 0.62 0.73 GM 0.27 0.62 --- 0.80 Learning 0.31 0.73 0.80 --- Table 4. The structural semantic relatedness of the semantic-graph shown in Figure 3 3 Named Entity Disambiguation by Leveraging Semantic Knowledge In this section we describe how to leverage the semantic knowledge captured in the structural semantic relatedness measure for named entity disambiguation. Because the key problem of named entity disambiguation is to measure the similarity between name observations, we integrate the structural semantic relatedness in the similarity measure, so that it can better reflect the actual similarity between name observations. Concretely, our named entity disambiguation system works as follows: 1) Measuring the similarity between name observations; 2) Grouping name observations using the clustering algorithm. In the following we describe each step in detail. 3.1 Measuring the Similarity between Name Observations Intuitively, if two observations of the target name represent the same entity, it is highly possible that the concepts in their contexts are closely related, i.e., the named entities in their contexts are socially related and the Wikipedia concepts in their contexts are semantically related. In contrast, if two name observations represent different entities, the concepts within their contexts will not be closely related. Therefore we can measure the similarity between two name observations by summarizing all the semantic relatedness between the concepts in their contexts. To measure the similarity between name observations, we represent each name observation as a weighted vector of concepts (including named entities, Wikipedia concepts and WordNet concepts), where the concepts are extracted using the same method described in Section 2.2, so they are just the same concepts within the semantic-graph. Using the same concept index as the semantic-graph, a name observation oi is then represented as 1 2 { , ,..., } i i i in o w w w = , where wik is the kth concept’s weight in observation oi, computed using the standard TFIDF weight model, where the DF is computed using the Google Web1T 5-gram corpus4. Given the concept vector representation of two name observations oi and oj, their similarity is computed as: ( , ) i j il jk lk il jk l k l k SIM o o w w S w w =∑∑ ∑∑ which is the weighted average of all the structural semantic relatedness between the concepts in the contexts of the two name observations. 3.2 Grouping Name Observations through Hierarchical Agglomerative Clustering Given the computed similarities, name observations are disambiguated by grouping them according to their represented entities. In this paper, we group name observations using the hierarchical agglomerative clustering(HAC) algorithm, which is widely used in prior disambiguation research and evaluation task (WePS1 and WePS2). The HAC produce clusters in a bottomup way as follows: Initially, each name observation is an individual cluster; then we iteratively merge the two clusters with the largest similarity value to form a new cluster until this similarity value is smaller than a preset merging threshold or all the observations reside in one common cluster. The merging threshold can be determined through cross-validation. We employ the single-link method to compute the similarity between two clusters, which has been applied widely in prior research (Bagga and Baldwin (1998); Mann and Yarowsky (2003)). 4 Experiments To assess the performance of our method and compare it with traditional methods, we conduct a series of experiments. In the experiments, we evaluate the proposed SSR method on the task of personal name disambiguation, which is the most common type of named entity disambiguation. In the following, we first explain the general experimental settings in Section 4.1, 4.2 and 4.3; then evaluate and discuss the performance of our method in Section 4.4. 4 www.ldc.upenn.edu/Catalog/docs/LDC2006T13/ 55 4.1 Disambiguation Data Sets We adopted the standard data sets used in the First Web People Search Clustering Task (WePS1) (Artiles et al., 2007) and the Second Web People Search Clustering Task (WePS2) (Artiles et al., 2009). The three data sets we used are WePS1_training data set, WePS1_test data set, and WePS2_test data set. Each of the three data sets consists of a set of ambiguous personal names (totally 109 personal names); and for each name, we need to disambiguate its observations in the web pages of the top N (100 for WePS1 and 150 for WePS2) Yahoo! search results. The experiment made the standard “one person per document” assumption, which is widely used in the participated systems in WePS1 and WePS2, i.e., all the observations of the same name in a document are assumed to represent the same entity. Based on this assumption, the features within the entire web page are used to disambiguate personal names. 4.2 Knowledge Sources There were three knowledge sources we used for our experiments: the WordNet 3.0; the Sep. 9, 2007 English version of Wikipedia; and the Web pages of each ambiguous name in WePS datasets as the NE Co-occurrence Corpus. 4.3 Evaluation Criteria We adopted the measures used in WePS1 to evaluate the performance of name disambiguation. These measures are: Purity (Pur): measures the homogeneity of name observations in the same cluster; Inverse purity (Inv_Pur): measures the completeness of a cluster; F-Measure (F): the harmonic mean of purity and inverse purity. The detailed definitions of these measures can be found in Amigo, et al. (2008). We use Fmeasure as the primary measure just liking WePS1 and WePS2. 4.4 Experimental Results We compared our method with four baselines: (1) BOW: The first one is the traditional Bag of Words model (BOW) based methods: hierarchical agglomerative clustering (HAC) over term vector similarity, where the features including single words and NEs, and all the features are weighted using TFIDF. This baseline is also the state-of-art method in WePS1 and WePS2. (2) SocialNetwork: The second one is the social network based methods, which is the same as the method described in Malin et al. (2005): HAC over the similarity obtained through random walk over the social network built from the web pages of the top N search results. (3)SSRNoKnowledge: The third one is used as a baseline for evaluating the efficiency of semantic knowledge: HAC over the similarity computed on semantic-graph with no knowledge integrated, i.e., the similarity is computed as: ( , ) i j il jl il jk l l k SIM o o w w w w =∑ ∑∑ (4) SSR-NoStructure: The fourth one is used as a baseline for evaluating the efficiency of the semantic knowledge embedded in complex structures: HAC over the similarity computed by only integrating the explicit semantic relations, i.e., the similarity is computed as: ( , ) i j il jk lk il jk l k l k SIM o o w w A w w =∑∑ ∑∑ 4.4.1 Overall Performance We conducted several experiments on all the three WePS data sets: the four baselines, the proposed SSR method and the proposed SSR method with only one special type knowledge added, respectively SSR-NE, SSR-WordNet and SSRWikipedia. All the optimal merging thresholds used in HAC were selected by applying leaveone-out cross validation. The overall performance is shown in Table 5. Method WePS1_training Pur Inv_Pur F BOW 0.71 0.88 0.78 SocialNetwork 0.66 0.98 0.76 SSR-NoKnowledge 0.79 0.89 0.81 SSR-NoStructure 0.87 0.83 0.83 SSR-NE 0.80 0.86 0.82 SSR-WordNet 0.80 0.91 0.83 SSR-Wikipedia 0.82 0.90 0.84 SSR 0.82 0.92 0.85 WePS1_test Pur Inv_Pur F BOW 0.74 0.87 0.74 SocialNetwork 0.83 0.63 0.65 SSR-NoKnowledge 0.80 0.74 0.75 SSR-NoStructure 0.80 0.78 0.78 SSR-NE 0.73 0.80 0.74 SSR-WordNet 0.81 0.77 0.77 SSR-Wikipedia 0.88 0.77 0.81 SSR 0.85 0.83 0.84 WePS2_test Pur Inv_Pur F BOW 0.80 0.80 0.77 SocialNetwork 0.62 0.93 0.70 SSR-NoKnowledge 0.84 0.80 0.80 SSR-NoStructure 0.84 0.83 0.81 SSR-NE 0.78 0.88 0.80 SSR-WordNet 0.85 0.82 0.83 SSR-Wikipedia 0.84 0.81 0.82 SSR 0.89 0.84 0.86 Table 5. Performance results of baselines and SSR methods 56 From the performance results in Table 5, we can see that: 1) The semantic knowledge can greatly improve the disambiguation performance: compared with the BOW and the SocialNetwork baselines, SSR respectively gets 8.7% and 14.7% improvement on average on the three data sets. 2) By leveraging the semantic knowledge from multiple knowledge sources, we can obtain a better named entity disambiguation performance: compared with the SSR-NE’s 0% improvement, the SSR-WordNet’s 2.3% improvement and the SSR-Wikipedia’s 3.7% improvement, the SSR gets 6.3% improvement over the SSR-NoKnowledge baseline, which is larger than all the SSR methods with only one type of semantic knowledge integrated. 3) The exploitation of the structural semantic knowledge can further improve the disambiguation performance: compared with SSRNoStructure, our SSR method achieves 4.3% improvement. Figure 4. The F-Measure vs. λ on three data sets 4.4.2 Optimizing Parameters There is only one parameter λ needed to be configured, which is the penalty factor for the relatedness transition path length in the structural semantic relatedness measure. Usually a smaller λ will make the structural semantic knowledge contribute less in the resulting relatedness value. Figure 4 plots the performance of our method corresponding to the special λ settings. As shown in Figure 4, the SSR method is not very sensitive to the λ and can achieve its best average performance when the value of λ is 0.6. 4.4.3 Detailed Analysis To better understand the reasons why our SSR method works well and how the exploitation of structural semantic knowledge can improve performance, we analyze the results in detail. The Exploitation of Semantic Knowledge. The primary advantage of our method is the exploitation of semantic knowledge. Our method exploits the semantic knowledge in two directions: 1) The Integration of Multiple Semantic Knowledge Sources. Using the semantic-graph model, our method can integrate the semantic knowledge extracted from multiple knowledge sources, while most traditional knowledge-based methods are usually specialized to one type of knowledge. By integrating multiple semantic knowledge sources, our method can improve the semantic knowledge coverage. 2) The exploitation of Semantic Knowledge embedded in complex structures. Using the structural semantic relatedness measure, our method can exploit the implicit semantic knowledge embedded in complex structures; while traditional knowledge-based methods usually lack this ability. The Rich Meaningful Features. One another advantage of our method is the rich meaningful features, which is brought by the multiple semantic knowledge sources. With more meaningful features, our method can better describe the name observations with less information loss. Furthermore, unlike the traditional N-gram features, the features enriched by semantic knowledge sources are all semantically meaningful units themselves, so little noisy features will be added. The effect of rich meaningful features can also be shown in Table 5: by adding these features, the SSR-NoKnowledge respectively achieves 2.3% and 9.7% improvement over the BOW and the SocialNetwork baseline. 5 Related Work In this section, we briefly review the related work. Totally, the traditional named entity disambiguation methods can be classified into two categories: the shallow methods and the knowledge-based methods. Most of previous named entity disambiguation researches adopt the shallow methods, which are mostly the natural extension of the bag of words (BOW) model. Bagga and Baldwin (1998) represented a name as a vector of its contextual words, then two names were predicted to be the same entity if their cosine similarity is above a threshold. Mann and Yarowsky (2003) and Niu et al. (2004) extended the vector representation with extracted biographic facts. Pedersen et al. (2005) employed significant bigrams to represent 57 a name observation. Chen and Martin (2007) explored a range of syntactic and semantic features. In recent years some research has investigated employing knowledge sources to enhance the named entity disambiguation. Bunescu and Pasca (2006) disambiguated the names using the category information in Wikipedia. Cucerzan (2007) disambiguated the names by combining the BOW model with the Wikipedia category information. Han and Zhao (2009) leveraged the Wikipedia semantic knowledge for computing the similarity between name observations. Bekkerman and McCallum (2005) disambiguated names based on the link structure of the Web pages between a set of socially related persons. Kalashnikov et al. (2008) and Lu et al. (2007) used the cooccurrence statistics between named entities in the Web. The social network was also exploited for named entity disambiguation, where similarity is computed through random walking, such as the work introduced in Malin (2005), Malin and Airoldi (2005), Yang et al.(2006) and Minkov et al. (2006). Hassell et al. (2006) used the relationships from DBLP to disambiguate names in research domain. 6 Conclusions and Future Works In this paper we demonstrate how to enhance the named entity disambiguation by capturing and exploiting the semantic knowledge existed in multiple knowledge sources. In particular, we propose a semantic relatedness measure, Structural Semantic Relatedness, which can capture both the explicit semantic relations and the implicit structural semantic knowledge. The experimental results on the WePS data sets demonstrate the efficiency of the proposed method. For future work, we want to develop a framework which can uniformly model the semantic knowledge and the contextual clues for named entity disambiguation. Acknowledgments The work is supported by the National Natural Science Foundation of China under Grants no. 60875041 and 60673042, and the National High Technology Development 863 Program of China under Grants no. 2006AA01Z144. References Amigo, E., Gonzalo, J., Artiles, J. and Verdejo, F. 2008. A comparison of extrinsic clustering evaluation metrics based on formal constraints. Information Retrieval. Artiles, J., Gonzalo, J. & Sekine, S. 2007. The SemEval-2007 WePS Evaluation: Establishing a benchmark for the Web People Search Task. In SemEval. Artiles, J., Gonzalo, J. and Sekine, S. 2009. WePS2 Evaluation Campaign: Overview of the Web People Search Clustering Task. In WePS2, WWW 2009. Baeza-Yates, R., Ribeiro-Neto, B., et al. 1999. Modern information retrieval. Addison-Wesley Reading, MA. Bagga, A. & Baldwin, B. 1998. Entity-based crossdocument coreferencing using the vector space model. Proceedings of the 17th international conference on Computational linguistics-Volume 1, pp. 79-85. Bekkerman, R. & McCallum, A. 2005. Disambiguating web appearances of people in a social network. Proceedings of the 14th international conference on World Wide Web, pp. 463-470. Bunescu, R. & Pasca, M. 2006. Using encyclopedic knowledge for named entity disambiguation. Proceedings of EACL, vol. 6. Chen, Y. & Martin, J. 2007. Towards robust unsupervised personal name disambiguation. Proceedings of EMNLP and CoNLL, pp. 190-198. Cilibrasi, R. L., Vitanyi, P. M. & CWI, A. 2007. The google similarity distance, IEEE Transactions on knowledge and data engineering, vol. 19, no. 3, pp. 370-383. Cucerzan, S. 2007, Large-scale named entity disambiguation based on Wikipedia data. Proceedings of EMNLP-CoNLL, pp. 708-716. Fellbaum, C., et al. 1998. WordNet: An electronic lexical database. MIT press Cambridge, MA. Fleischman, M. B. & Hovy, E. 2004. Multi-document person name resolution. Proceedings of ACL, Reference Resolution Workshop. Han, X. & Zhao, J. 2009. Named entity disambiguation by leveraging Wikipedia semantic knowledge. Proceeding of the 18th ACM conference on Information and knowledge management, pp. 215-224. Hassell, J., Aleman-Meza, B. & Arpinar, I. 2006. Ontology-driven automatic entity disambiguation in unstructured text. Proceedings of The 2006 ISWC, pp. 44-57. Jeh, G. & Widom, J. 2002. SimRank: A measure of structural-context similarity, Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, p. 543. 58 Kalashnikov, D. V., Nuray-Turan, R. & Mehrotra, S. 2008. Towards Breaking the Quality Curse. A Web-Querying Approach to Web People Search. In Proc. of SIGIR. Leicht, E. A., Petter Holme, & M. E. J. Newman. 2006. Vertex similarity in networks. Physical Review E , vol. 73, no. 2, p. 26120. Lin., D. 1998. An information-theoretic definition of similarity. In Proc. of ICML. Lu, Y. & Nie , Z. et al. 2007. Name Disambiguation Using Web Connection. In Proc. of AAAI. Malin, B. 2005. Unsupervised name disambiguation via social network similarity. SIAM SDM Workshop on Link Analysis, Counterterrorism and Security. Malin, B., Airoldi, E. & Carley, K. M. 2005. A network analysis model for disambiguation of names in lists. Computational & Mathematical Organization Theory, vol. 11, no. 2, pp. 119-139. Mann, G. S. & Yarowsky, D. 2003. Unsupervised personal name disambiguation, Proceedings of the seventh conference on Natural language learning at HLT-NAACL 2003-Volume 4, p. 40. McNamee, P. & Dang, H. Overview of the TAC 2009 Knowledge Base Population Track. In Proceedings of Text Analysis Conference (TAC-2009), 2009. Medelyan, O., Witten, I. H. & Milne, D. 2008. Topic indexing with Wikipedia. Proceedings of the AAAI WikiAI workshop. Milne, D., Medelyan, O. & Witten, I. H. 2006. Mining domain-specific thesauri from wikipedia: A case study. IEEE/WIC/ACM International Conference on Web Intelligence, pp. 442-448. Minkov, E., Cohen, W. W. & Ng, A. Y. 2006. Contextual search and name disambiguation in email using graphs, Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pp. 27-34. Niu C., Li W. and Srihari, R. K. 2004. Weakly Supervised Learning for Cross-document Person Name Disambiguation Supported by Information Extraction. Proceedings of ACL, pp. 598-605. Pedersen, T., Purandare, A. & Kulkarni, A. 2005. Name discrimination by clustering similar contexts. Computational Linguistics and Intelligent Text Processing, pp. 226-237. Strube, M. & Ponzetto, S. P. 2006. WikiRelate! Computing semantic relatedness using Wikipedia, Proceedings of the National Conference on Artificial Intelligence, vol. 21, no. 2, p. 1419. Suchanek, F. M., Kasneci, G. & Weikum, G. 2007. Yago: a core of semantic knowledge, Proceedings of the 16th international conference on World Wide Web, p. 706. Wan, X., Gao, J., Li, M. & Ding, B. 2005. Person resolution in person search results: Webhawk. Proceedings of the 14th ACM international conference on Information and knowledge management, p. 170. Witten, D. M. & Milne, D. 2008. An effective, lowcost measure of semantic relatedness obtained from Wikipedia links. Proceeding of AAAI Workshop on Wikipedia and Artificial Intelligence: an Evolving Synergy, AAAI Press, Chicago, USA, pp. 2530. Yang, K. H., Chiou, K. Y., Lee, H. M. & Ho, J. M. 2006. Web appearance disambiguation of personal names based on network motif. Proceedings of the 2006 IEEE/WIC/ACM International Conference on Web Intelligence, pp. 386-389. 59
|
2010
|
6
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 585–594, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Generating Focused Topic-specific Sentiment Lexicons Valentin Jijkoun Maarten de Rijke Wouter Weerkamp ISLA, University of Amsterdam, The Netherlands jijkoun,derijke,[email protected] Abstract We present a method for automatically generating focused and accurate topicspecific subjectivity lexicons from a general purpose polarity lexicon that allow users to pin-point subjective on-topic information in a set of relevant documents. We motivate the need for such lexicons in the field of media analysis, describe a bootstrapping method for generating a topic-specific lexicon from a general purpose polarity lexicon, and evaluate the quality of the generated lexicons both manually and using a TREC Blog track test set for opinionated blog post retrieval. Although the generated lexicons can be an order of magnitude more selective than the general purpose lexicon, they maintain, or even improve, the performance of an opinion retrieval system. 1 Introduction In the area of media analysis, one of the key tasks is collecting detailed information about opinions and attitudes toward specific topics from various sources, both offline (traditional newspapers, archives) and online (news sites, blogs, forums). Specifically, media analysis concerns the following system task: given a topic and list of documents (discussing the topic), find all instances of attitudes toward the topic (e.g., positive/negative sentiments, or, if the topic is an organization or person, support/criticism of this entity). For every such instance, one should identify the source of the sentiment, the polarity and, possibly, subtopics that this attitude relates to (e.g., specific targets of criticism or support). Subsequently, a (human) media analyst must be able to aggregate the extracted information by source, polarity or subtopics, allowing him to build support/criticism networks etc. (Altheide, 1996). Recent advances in language technology, especially in sentiment analysis, promise to (partially) automate this task. Sentiment analysis is often considered in the context of the following two tasks: • sentiment extraction: given a set of textual documents, identify phrases, clauses, sentences or entire documents that express attitudes, and determine the polarity of these attitudes (Kim and Hovy, 2004); and • sentiment retrieval: given a topic (and possibly, a list of documents relevant to the topic), identify documents that express attitudes toward this topic (Ounis et al., 2007). How can technology developed for sentiment analysis be applied to media analysis? In order to use a sentiment extraction system for a media analysis problem, a system would have to be able to determine which of the extracted sentiments are actually relevant, i.e., it would not only have to identify specific targets of all extracted sentiments, but also decide which of the targets are relevant for the topic at hand. This is a difficult task, as the relation between a topic (e.g., a movie) and specific targets of sentiments (e.g., acting or special effects in the movie) is not always straightforward, in the face of ubiquitous complex linguistic phenomena such as referential expressions (“... this beautifully shot documentary”) or bridging anaphora (“the director did an excellent jobs”). In sentiment retrieval, on the other hand, the topic is initially present in the task definition, but it is left to the user to identify sources and targets of sentiments, as systems typically return a list of documents ranked by relevance and opinionatedness. To use a traditional sentiment retrieval system in media analysis, one would still have to manually go through ranked lists of documents returned by the system. 585 To be able to support media analysis, we need to combine the specificity of (phrase- or word-level) sentiment analysis with the topicality provided by sentiment retrieval. Moreover, we should be able to identify sources and specific targets of opinions. Another important issue in the media analysis context is evidence for a system’s decision. If the output of a system is to be used to inform actions, the system should present evidence, e.g., highlighting words or phrases that indicate a specific attitude. Most modern approaches to sentiment analysis, however, use various flavors of classification, where decisions (typically) come with confidence scores, but without explicit support. In order to move towards the requirements of media analysis, in this paper we focus on two of the problems identified above: (1) pinpointing evidence for a system’s decisions about the presence of sentiment in text, and (2) identifying specific targets of sentiment. We address these problems by introducing a special type of lexical resource: a topic-specific subjectivity lexicon that indicates specific relevant targets for which sentiments may be expressed; for a given topic, such a lexicon consists of pairs (syntactic clue, target). We present a method for automatically generating a topic-specific lexicon for a given topic and query-biased set of documents. We evaluate the quality of the lexicon both manually and in the setting of an opinionated blog post retrieval task. We demonstrate that such a lexicon is highly focused, allowing one to effectively pinpoint evidence for sentiment, while being competetive with traditional subjectivity lexicons consisting of (a large number of) clue words. Unlike other methods for topic-specific sentiment analysis, we do not expand a seed lexicon. Instead, we make an existing lexicon more focused, so that it can be used to actually pin-point subjectivity in documents relevant to a given topic. 2 Related Work Much work has been done in sentiment analysis. We discuss related work in four parts: sentiment analysis in general, domain- and targetspecific sentiment analysis, product review mining and sentiment retrieval. 2.1 Sentiment analysis Sentiment analysis is often seen as two separate steps for determining subjectivity and polarity. Most approaches first try to identify subjective units (documents, sentences), and for each of these determine whether it is positive or negative. Kim and Hovy (2004) select candidate sentiment sentences and use word-based sentiment classifiers to classify unseen words into a negative or positive class. First, the lexicon is constructed from WordNet: from several seed words, the structure of WordNet is used to expand this seed to a full lexicon. Next, this lexicon is used to measure the distance between unseen words and words in the positive and negative classes. Based on word sentiments, a decision is made at the sentence level. A similar approach is taken by Wilson et al. (2005): a classifier is learnt that distinguishes between polar and neutral sentences, based on a prior polarity lexicon and an annotated corpus. Among the features used are syntactic features. After this initial step, the sentiment sentences are classified as negative or positive; again, a prior polarity lexicon and syntactic features are used. The authors later explored the difference between prior and contextual polarity (Wilson et al., 2009): words that lose polarity in context, or whose polarity is reversed because of context. Riloff and Wiebe (2003) describe a bootstrapping method to learn subjective extraction patterns that match specific syntactic templates, using a high-precision sentence-level subjectivity classifier and a large unannotated corpus. In our method, we bootstrap from a subjectivity lexicion rather than a classifier, and perform a topicspecific analysis, learning indicators of subjectivity toward a specific topic. 2.2 Domain- and target-specific sentiment The way authors express their attitudes varies with the domain: An unpredictable movie can be positive, but unpredictable politicians are usually something negative. Since it is unrealistic to construct sentiment lexicons, or manually annotate text for learning, for every imaginable domain or topic, automatic methods have been developed. Godbole et al. (2007) aim at measuring overall subjectivity or polarity towards a certain entity; they identify sentiments using domain-specific lexicons. The lexicons are generated from manually selected seeds for a broad domain such as Health or Business, following an approach similar to (Kim and Hovy, 2004). All named entites in a sentence containing a clue from a lexicon are 586 considered targets of sentiment for counting. Because of the data volume, no expensive linguistic processing is performed. Choi et al. (2009) advocate a joint topicsentiment analysis. They identify “sentiment topics,” noun phrases assumed to be linked to a sentiment clue in the same expression. They address two tasks: identifying sentiment clues, and classifying sentences into positive, negative, or neutral. They start by selecting initial clues from SentiWordNet, based on sentences with known polarity. Next, the sentiment topics are identified, and based on these sentiment topics and the current list of clues, new potential clues are extracted. The clues can be used to classifiy sentences. Fahrni and Klenner (2008) identify potential targets in a given domain, and create a targetspecific polarity adjective lexicon. To this end, they find targets using Wikipedia, and associated adjectives. Next, the target-specific polarity of adjectives is detemined using Hearst-like patterns. Kanayama and Nasukawa (2006) introduce polar atoms: minimal human-understandable syntactic structures that specify polarity of clauses. The goal is to learn new domain-specific polar atoms, but these are not target-specific. They use manually-created syntactic patterns to identify atoms and coherency to determine polarity. In contrast to much of the work in the literature, we need to specialize subjectivity lexicons not for a domain and target, but for “topics.” 2.3 Product features and opinions Much work has been carried out for the task of mining product reviews, where the goal is to identify features of specific products (such as picture, zoom, size, weight for digital cameras) and opinions about these specific features in user reviews. Liu et al. (2005) describe a system that identifies such features via rules learned from a manually annotated corpus of reviews; opinions on features are extracted from the structure of reviews (which explicitly separate positive and negative opinions). Popescu and Etzioni (2005) present a method that identifies product features for using corpus statistics, WordNet relations and morphological cues. Opinions about the features are extracted using a hand-crafted set of syntactic rules. Targets extracted in our method for a topic are similar to features extracted in review mining for products. However, topics in our setting go beyond concrete products, and the diversity and generality of possible topics makes it difficult to apply such supervised or thesaurus-based methods to identify opinion targets. Moreover, in our method we directly use associations between targets and opinions to extract both. 2.4 Sentiment retrieval At TREC, the Text REtrieval Conference, there has been interest in a specific type of sentiment analysis: opinion retrieval. This interest materialized in 2006 (Ounis et al., 2007), with the opinionated blog post retrieval task. Finding blog posts that are not just about a topic, but also contain an opinion on the topic, proves to be a difficult task. Performance on the opinion-finding task is dominated by performance on the underlying document retrieval task (the topical baseline). Opinion finding is often approached as a twostage problem: (1) identify documents relevant to the query, (2) identify opinions. In stage (2) one commonly uses either a binary classifier to distinguish between opinionated and non-opinionated documents or applies reranking of the initial result list using some opinion score. Opinion add-ons show only slight improvements over relevanceonly baselines. The best performing opinion finding system at TREC 2008 is a two-stage approach using reranking in stage (2) (Lee et al., 2008). The authors use SentiWordNet and a corpus-derived lexicon to construct an opinion score for each post in an initial ranking of blog posts. This opinion score is combined with the relevance score, and posts are reranked according to this new score. We detail this approach in Section 6. Later, the authors use domain-specific opinion indicators (Na et al., 2009), like “interesting story” (movie review), and “light” (notebook review). This domain-specific lexicon is constructed using feedback-style learning: retrieve an initial list of documents and use the top documents as training data to learn an opinion lexicon. Opinion scores per document are then computed as an average of opinion scores over all its words. Results show slight improvements (+3%) on mean average precision. 3 Generating Topic-Specific Lexicons In this section we describe how we generate a lexicon of subjectivity clues and targets for a given topic and a list of relevant documents (e.g., re587 Extract all syntactic contexts of clue words Background corpus Topic-independent subjectivity lexicon Relevant docs Topic For each clue word, select D contexts with highest entropy List of syntactic clues: (clue word, syn. context) Extract all occurrences endpoints of syntactic clues Extract all occurrences endpoints of syntactic clues Potential targets in background corpus Potential targets in relevant doc. list Compare frequencies using chi-square; select top T targets List of T targets For each target, find syn. clues it co-occurs with Topic-specific lexicon of tuples: (syntactic clue, target) Step 1 Step 2 Step 3 Figure 1: Our method for learning a topicdependent subjectivity lexicon. trieved by a search engine for the topic). As an additional resource, we use a large background corpus of text documents of a similar style but with diverse subjects; we assume that the relevant documents are part of this corpus as well. As the background corpus, we used the set of documents from the assessment pools of TREC 2006–2008 opinion retrieval tasks (described in detail in section 4). We use the Stanford lexicalized parser1 to extract labeled dependency triples (head, label, modifier). In the extracted triples, all words indicate their category (noun, adjective, verb, adverb, etc.) and are normalized to lemmas. Figure 1 provides an overview of our method; below we describe it in more detail. 3.1 Step 1: Extracting syntactic contexts We start with a general domain-independent prior polarity lexicon of 8,821 clue words (Wilson et al., 2005). First, we identify syntactic contexts in which specific clue words can be used to express 1http://nlp.stanford.edu/software/ lex-parser.shtml attitude: we try to find how a clue word can be syntactically linked to targets of sentiments. We take a simple definition of the syntactic context: a single labeled directed dependency relation. For every clue word, we extract all syntactic contexts, i.e., all dependencies, in which the word is involved (as head or as modifier) in the background corpus, along with their endpoints. Table 1 shows examples of clue words and contexts that indicate sentiments. For every clue, we only select those contexts that exhibit a high entropy among the lemmas at the other endpoint of the dependencies. E.g., in our background corpus, the verb to like occurs 97,179 times with a nominal subject and 52,904 times with a direct object; however, the entropy of lemmas of the subjects is 4.33, compared to 9.56 for the direct objects. In other words, subjects of like are more “predictable.” Indeed, the pronoun I accounts for 50% of subjects, followed by you (14%), they (4%), we (4%) and people (2%). The most frequent objects of like are it (12%), what (4%), idea (2%), they (2%). Thus, objects of to like will be preferred by the method. Our entropy-driven selection of syntactic contexts of a clue word is based on the following assumption: Assumption 1: In text, targets of sentiments are more diverse than sources of sentiments or other accompanying attributes such as location, time, manner, etc. Therefore targets exhibit higher entropy than other attributes. For every clue word, we select the top D syntactic contexts whose entropy is at least half of the maximum entropy for this clue. To summarize, at the end of Step 1 of our method, we have extracted a list of pairs (clue word, syntactic context) such that for occurrences of the clue word, the words at the endpoint of the syntactic dependency are likely to be targets of sentiments. We call such a pair a syntactic clue. 3.2 Step 2: Selecting potential targets Here, we use the extracted syntantic clues to identify words that are likely to serve as specific targets for opinions about the topic in the relevant documents. In this work we only consider individual words as potential targets and leave exploring other options (e.g., NPs and VPs as targets) for future work. In extracting targets, we rely on the following assumption: 588 Clue word Syntactic context Target Example to like has direct object u2 I do still like U2 very much to like has clausal complement criticize I don’t like to criticize our intelligence services to like has about-modifier olympics That’s what I like about Winter Olympics terrible is adjectival modifier of idea it’s a terrible idea to recall judges for... terrible has nominal subject shirt And Neil, that shirt is terrible! terrible has clausal complement can It is terrible that a small group of extremists can . . . Table 1: Examples of subjective syntactic contexts of clue words (based on Stanford dependencies). Assumption 2: The list of relevant documents contains a substantial number of documents on the topic which, moreover, contain sentiments about the topic. We extract all endpoints of all occurrences of the syntactic clues in the relevant documents, as well as in the background corpus. To identify potential attitude targets in the relevant documents, we compare their frequency in the relevant documents to the frequency in the background corpus using the standard χ2 statistics. This technique is based on the following assumption: Assumption 3: Sentiment targets related to the topic occur more often in subjective context in the set of relevant documents, than in the background corpus. In other words, while the background corpus contains sentiments towards very diverse subjects, the relevant documents tend to express attitudes related to the topic. For every potential target, we compute the χ2score and select the top T highest scoring targets. As the result of Steps 1 and 2, as candidate targets for a given topic, we only select words that occur in subjective contexts, and that do so more often than we would normally expect. Table 2 shows examples of extracted targets for three TREC topics (see below for a description of our experimental data). 3.3 Step 3: Generating topic-specific lexicons In the last step of the method, we combine clues and targets. For each target identified in Step 2, we take all syntactic clues extracted in Step 1 that co-occur with the target in the relevant documents. The resulting list of triples (clue word, syntactic context, target) constitute the lexicon. We conjecture that an occurrence of a lexicon entry in a text indicates, with reasonable confidence, a subjective attitude towards the target. Topic “Relationship between Abramoff and Bush” abramoff lobbyist scandal fundraiser bush fund-raiser republican prosecutor tribe swirl corrupt corruption norquist democrat lobbying investigation scanlon reid lawmaker dealings president Topic “MacBook Pro” macbook laptop powerbook connector mac processor notebook fw800 spec firewire imac pro machine apple powerbooks ibook ghz g4 ata binary keynote drive modem Topic: “Super Bowl ads” ad bowl commercial fridge caveman xl endorsement advertising spot advertiser game super essential celebrity payoff marketing publicity brand advertise watch viewer tv football venue Table 2: Examples of targets extracted at Step 2. 4 Data and Experimental Setup We consider two types of evaluation. In the next section, we examine the quality of the lexicons we generate. In the section after that we evaluate lexicons quantitatively using the TREC Blog track benchmark. For extrinsic evaluation we apply our lexicon generation method to a collection of documents containing opinionated utterances: blog posts. The Blogs06 collection (Macdonald and Ounis, 2006) is a crawl of blog posts from 100,649 blogs over a period of 11 weeks (06/12/2005– 21/02/2006), with 3,215,171 posts in total. Before indexing the collection, we perform two preprocessing steps: (i) when extracting plain text from HTML, we only keep block-level elements longer than 15 words (to remove boilerplate material), and (ii) we remove non-English posts using TextCat2 for language detection. This leaves us with 2,574,356 posts with 506 words per post on average. We index the collection using Indri,3 version 2.10. TREC 2006–2008 came with the task of opinionated blog post retrieval (Ounis et al., 2007). For each year a set of 50 topics was created, giv2http://odur.let.rug.nl/∼vannoord/ TextCat/ 3http://www.lemurproject.org/indri/ 589 ing us 150 topics in total. Every topic comes with a set of relevance judgments: Given a topic, a blog post can be either (i) nonrelevant, (ii) relevant, but not opinionated, or (iii) relevant and opinionated. TREC topics consist of three fields (title, description, and narrative), of which we only use the title field: a query of 1–3 keywords. We use standard TREC evaluation measures for opinion retrieval: MAP (mean average precision), R-precision (precision within the top R retrieved documents, where R is the number of known relevant documents in the collection), MRR (mean reciprocal rank), P@10 and P@100 (precision within the top 10 and 100 retrieved documents). In the context of media analysis, recall-oriented measures such as MAP and R-precision are more meaningful than the other, early precision-oriented measures. Note that for the opinion retrieval task a document is considered relevant if it is on topic and contains opinions or sentiments towards the topic. Throughout Section 6 below, we test for significant differences using a two-tailed paired t-test, and report on significant differences for α = 0.01 (▲and ▼), and α = 0.05 (△and ▽). For the quantative experiments in Section 6 we need a topical baseline: a set of blog posts potentially relevant to each topic. For this, we use the Indri retrieval engine, and apply the Markov Random Fields to model term dependencies in the query (Metzler and Croft, 2005) to improve topical retrieval. We retrieve the top 1,000 posts for each query. 5 Qualitative Analysis of Lexicons Lexicon size (the number of entries) and selectivity (how often entries match in text) of the generated lexicons vary depending on the parameters D and T introduced above. The two rightmost columns of Table 4 show the lexicon size and the average number of matches per topic. Because our topic-specific lexicons consist of triples (clue word, syntactic context, target), they actually contain more words than topic-independent lexicons of the same size, but topic-specific entries are more selective, which makes the lexicon more focused. Table 3 compares the application of topic-independent and topic-specific lexicons to on-topic blog text. We manually performed an explorative error analysis on a small number of documents, annoThere are some tragic moments like eggs freezing , and predators snatching the females and little ones-you know the whole NATURE thing ... but this movie is awesome There are some tragic moments l ike eggs freezing , and predators snatching the females and little ones-you know the whole NATURE thing ... but this movie is awesome Saturday was more errands, then spent the evening with Dad and Stepmum, and finally was able to see March of the Penguins, which was wonderful. Christmas Day was lovely, surrounded by family, good food and drink, and little L to play with. Saturday was more errands, then spent the evening with Dad and Stepmum, and finally was able to see March of the Penguins, which was wonderful. Christmas Day was lovely, surrounded by family, good food and drink, and little L to play with. Table 3: Posts with highlighted targets (bold) and subjectivity clues (blue) using topic-independent (left) and topic-specific (right) lexicons. tated using the smallest lexicon in Table 4 for the topic “March of the Pinguins.” We assigned 186 matches of lexicon entries in 30 documents into four classes: • REL: sentiment towards a relevant target; • CONTEXT: sentiment towards a target that is irrelevant to the topic due to context (e.g., opinion about a target “film”, but refering to a film different from the topic); • IRREL: sentiment towards irrelevant target (e.g., “game” for a topic about a movie); • NOSENT: no sentiment at all In total only 8% of matches were manually classified as REL, with 62% classified as NOSENT, 23% as CONTEXT, and 6% as IRREL. On the other hand, among documents assessed as opionionated by TREC assessors, only 13% did not contain matches of the lexicon entries, compared to 27% of non-opinionated documents, which does indicate that our lexicon does attempt to separate non-opinionated documents from opinionated. 6 Quantitative Evaluation of Lexicons In this section we assess the quality of the generated topic-specific lexicons numerically and extrinsically. To this end we deploy our lexicons to the task of opinionated blog post retrieval (Ounis et al., 2007). A commonly used approach to this task works in two stages: (1) identify topically relevant blog posts, and (2) classify these posts as being opinionated or not. In stage 2 the standard 590 approach is to rerank the results from stage 1, instead of doing actual binary classification. We take this approach, as it has shown good performance in the past TREC editions (Ounis et al., 2007) and is fairly straightforward to implement. We also explore another way of using the lexicon: as a source for query expansion (i.e., adding new terms to the original query) in Section 6.2. For all experiments we use the collection described in Section 4. Our experiments have two goals: to compare the use of topic-independent and topic-specific lexicons for the opinionated post retrieval task, and to examine how different settings for the parameters of the lexicon generation affect the empirical quality. 6.1 Reranking using a lexicon To rerank a list of posts retrieved for a given topic, we opt to use the method that showed best performance at TREC 2008. The approach taken by Lee et al. (2008) linearly combines a (topical) relevance score with an opinion score for each post. For the opinion score, terms from a (topic-independent) lexicon are matched against the post content, and weighted with the probability of term’s subjectivity. Finally, the sum is normalized using the Okapi BM25 framework. The final opinion score Sop is computed as in Eq. 1: Sop(D) = Opinion(D) · (k1 + 1) Opinion(D) + k1 · (1 −b + b·|D| avgdl) , (1) where k1, and b are Okapi parameters (set to their default values k1 = 2.0, and b = 0.75), |D| is the length of document D, and avgdl is the average document length in the collection. The opinion score Opinion(D) is calculated using Eq. 2: Opinion(D) = X w∈O P(sub|w) · n(w, D), (2) where O is the set of terms in the sentiment lexicon, P(sub|w) indicates the probability of term w being subjective, and n(w, D) is the number of times term w occurs in document D. The opinion scoring can weigh lexicon terms differently, using P(sub|w); it normalizes scores to cancel out the effect of varying document sizes. In our experiments we use the method described above, and plug in the MPQA polarity lexicon.4 We compare the results of using this 4http://www.cs.pitt.edu/mpqa/ topic-independent lexicon to the topic-dependent lexicons our method generates, which are also plugged into the reranking of Lee et al. (2008). In addition to using Okapi BM25 for opinion scoring, we also consider a simpler method. As we observed in Section 5, our topic-specific lexicons are more selective than the topic-independent lexicon, and a simple number of lexicon matches can give a good indication of opinionatedness of a document: Sop(D) = min(n(O, D), 10)/10, (3) where n(O, D) is the number of matches of the term of sentiment lexicon O in document D. 6.1.1 Results and observations There are several parameters that we can vary when generating a topic-specific lexicon and when using it for reranking: D: the number of syntactic contexts per clue T: the number of extracted targets Sop(D): the opinion scoring function. α: the weight of the opinion score in the linear combination with the relevance score. Note that α does not affect the lexicon creation, but only how the lexicon is used in reranking. Since we want to assess the quality of lexicons, not in the opinionated retrieval performance as such, we factor out α by selecting the best setting for each lexicon (including the topic-independent) and each evaluation measure. In Table 4 we present the results of evaluation of several lexicons in the context of opinionated blog post retrieval. First, we note that reranking using all lexicons in Table 4 significantly improves over the relevance-only baseline for all evaluation measures. When comparing topic-specific lexicons to the topic-independent one, most of the differences are not statistically significant, which is surprising given the fact that most topic-specific lexicons we evaluated are substantially smaller (see the two rightmost columns in the table). The smallest lexicon in Table 4 is seven times more selective than the general one, in terms of the number of lexicon matches per document. The only evaluation measure where the topicindependent lexicon consistently outperforms topic-specific ones, is Mean Reciprocal Rank that depends on a single relevant opinionated document high in a ranking. A possible explanation 591 Lexicon MAP R-prec MRR P@10 P@100 |lexicon| hits per doc no reranking 0.2966 0.3556 0.6750 0.4820 0.3666 — — topic-independent 0.3182 0.3776 0.7714 0.5607 0.3980 8,221 36.17 D T Sop 3 50 count 0.3191 0.3769 0.7276▽ 0.5547 0.3963 2,327 5.02 3 100 count 0.3191 0.3777 0.7416 0.5573 0.3971 3,977 8.58 5 50 count 0.3178 0.3775 0.7246▽ 0.5560 0.3931 2,784 5.73 5 100 count 0.3178 0.3784 0.7316▽ 0.5513 0.3961 4,910 10.06 all 50 count 0.3167 0.3753 0.7264▽ 0.5520 0.3957 4,505 9.34 all 100 count 0.3146 0.3761 0.7283▽ 0.5347▽ 0.3955 8,217 16.72 all 50 okapi 0.3129 0.3713 0.7247▼ 0.5333▽ 0.3833▽ 4,505 9.34 all 100 okapi 0.3189 0.3755 0.7162▼ 0.5473 0.3921 8,217 16.72 all 200 okapi 0.3229▲ 0.3803 0.7389 0.5547 0.3987 14,581 29.14 Table 4: Evaluation of topic-specific lexicons applied to the opinion retrieval task, compared to the topicindependent lexicon. The two rightmost columns show the number of lexicon entries (average per topic) and the number of matches of lexicon entries in blog posts (average for top 1,000 posts). is that the large general lexicon easily finds a few “obviously subjective” posts (those with heavily used subjective words), but is not better at detecting less obvious ones, as indicated by the recalloriented MAP and R-precision. Interestingly, increasing the number of syntactic contexts considered for a clue word (parameter D) and the number of selected targets (parameter T) leads to substantially larger lexicons, but only gives marginal improvements when lexicons are used for opinion retrieval. This shows that our bootstrapping method is effective at filtering out non-relevant sentiment targets and syntactic clues. The evaluation results also show that the choice of opinion scoring function (Okapi or raw counts) depends on the lexicon size: for smaller, more focused lexicons unnormalized counts are more effective. This also confirms our intuition that for small, focused lexicons simple presence of a sentiment clue in text is a good indication of subjectivity, while for larger lexicons an overall subjectivity scoring of texts has to be used, which can be hard to interpret for (media analysis) users. 6.2 Query expansion with lexicons In this section we evaluate the quality of targets extracted as part of the lexicons by using them for query expansion. Query expansion is a commonly used technique in information retrieval, aimed at getting a better representation of the user’s information need by adding terms to the original retrieval query; for user-generated content, selective query expansion has proved very beneficial (Weerkamp et al., 2009). We hypothesize that if our method manages to identify targets that correspond to issues, subtopics or features associated Run MAP P@10 MRR Topical blog post retrieval Baseline 0.4086 0.7053 0.7984 Rel. models 0.4017▽ 0.6867 0.7383▼ Subj. targets 0.4190△ 0.7373△ 0.8470△ Opinion retrieval Baseline 0.2966 0.4820 0.6750 Rel. models 0.2841▼ 0.4467▼ 0.5479▼ Subj. targets 0.3075 0.5227▲ 0.7196 Table 5: Query expansion using relevance models and topic-specific subjectivity targets. Significance tested against the baseline. with the topic, the extracted targets should be good candidates for query expansion. The experiments described below test this hypothesis. For every test topic, we select the 20 top-scoring targets as expansion terms, and use Indri to return 1,000 most relevant documents for the expanded query. We evaluate the resulting ranking using both topical retrieval and opinionated retrieval measures. For the sake of comparison, we also implemented a well-known query expansion method based on Relevance Models (Lavrenko and Croft, 2001): this method has been shown to work well in many settings. Table 5 shows evaluation results for these two query expansion methods, compared to the baseline retrieval run. The results show that on topical retrieval query expansion using targets significantly improves retrieval performance, while using relevance models actually hurts all evaluation measures. The failure of the latter expansion method can be attributed to the relatively large amount of noise in user-generated content, such as boilerplate 592 material, timestamps of blog posts, comments etc. (Weerkamp and de Rijke, 2008). Our method uses full syntactic parsing of the retrieved documents, which might substantially reduce the amount of noise since only (relatively) wellformed English sentences are used in lexicon generation. For opinionated retrieval, target-based expansion also improves over the baseline, although the differences are only significant for P@10. The consistent improvement for topical retrieval suggests that a topic-specific lexicon can be used both for query expansion (as described in this section) and for opinion reranking (as described in Section 6.1). We leave this combination for future work. 7 Conclusions and Future Work We have described a bootstrapping method for deriving a topic-specific lexicon from a general purpose polarity lexicon. We have evaluated the quality of generated lexicons both manually and using a TREC Blog track test set for opinionated blog post retrieval. Although the generated lexicons can be an order of magnitude more selective, they maintain, or even improve, the performance of an opinion retrieval system. As to future work, we intend to combine our method with known methods for topic-specific lexicon expansion (our method is rather concerned with lexicon “restriction”). Existing sentenceor phrase-level (trained) sentiment classifiers can also be used easily: when collecting/counting targets we can weigh them by “prior” score provided by such classifiers. We also want to look at more complex syntactic patterns: Choi et al. (2009) report that many errors are due to exclusive use of unigrams. We would also like to extend potential opinion targets to include multi-word phrases (NPs and VPs), in addition to individual words. Finally, we do not identify polarity yet: this can be partially inherited from the initial lexicon and refined automatically via bootstrapping. Acknowledgements This research was supported by the European Union’s ICT Policy Support Programme as part of the Competitiveness and Innovation Framework Programme, CIP ICT-PSP under grant agreement nr 250430, by the DuOMAn project carried out within the STEVIN programme which is funded by the Dutch and Flemish Governments under project nr STE-09-12, and by the Netherlands Organisation for Scientific Research (NWO) under project nrs 612.066.512, 612.061.814, 612.061.815, 640.004.802. References Altheide, D. (1996). Qualitative Media Analysis. Sage. Choi, Y., Kim, Y., and Myaeng, S.-H. (2009). Domainspecific sentiment analysis using contextual feature generation. In TSA ’09: Proceeding of the 1st international CIKM workshop on Topic-sentiment analysis for mass opinion, pages 37–44, New York, NY, USA. ACM. Fahrni, A. and Klenner, M. (2008). Old Wine or Warm Beer: Target-Specific Sentiment Analysis of Adjectives. In Proc.of the Symposium on Affective Language in Human and Machine, AISB 2008 Convention, 1st-2nd April 2008. University of Aberdeen, Aberdeen, Scotland, pages 60 – 63. Godbole, N., Srinivasaiah, M., and Skiena, S. (2007). Largescale sentiment analysis for news and blogs. In Proceedings of the International Conference on Weblogs and Social Media (ICWSM). Kanayama, H. and Nasukawa, T. (2006). Fully automatic lexicon expansion for domain-oriented sentiment analysis. In EMNLP ’06: Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing, pages 355–363, Morristown, NJ, USA. Association for Computational Linguistics. Kim, S. and Hovy, E. (2004). Determining the sentiment of opinions. In Proceedings of COLING 2004. Lavrenko, V. and Croft, B. (2001). Relevance-based language models. In SIGIR ’01: Proceedings of the 24th annual international ACM SIGIR conference on research and development in information retrieval. Lee, Y., Na, S.-H., Kim, J., Nam, S.-H., Jung, H.-Y., and Lee, J.-H. (2008). KLE at TREC 2008 Blog Track: Blog Post and Feed Retrieval. In Proceedings of TREC 2008. Liu, B., Hu, M., and Cheng, J. (2005). Opinion observer: analyzing and comparing opinions on the web. In Proceedings of the 14th international conference on World Wide Web. Macdonald, C. and Ounis, I. (2006). The TREC Blogs06 collection: Creating and analysing a blog test collection. Technical Report TR-2006-224, Department of Computer Science, University of Glasgow. Metzler, D. and Croft, W. B. (2005). A markov random feld model for term dependencies. In SIGIR ’05: Proceedings of the 28th annual international ACM SIGIR conference on research and development in information retrieval, pages 472–479, New York, NY, USA. ACM Press. Na, S.-H., Lee, Y., Nam, S.-H., and Lee, J.-H. (2009). Improving opinion retrieval based on query-specific sentiment lexicon. In ECIR ’09: Proceedings of the 31th European Conference on IR Research on Advances in Information Retrieval, pages 734–738, Berlin, Heidelberg. Springer-Verlag. Ounis, I., Macdonald, C., de Rijke, M., Mishne, G., and Soboroff, I. (2007). Overview of the TREC 2006 blog track. In The Fifteenth Text REtrieval Conference (TREC 2006). NIST. Popescu, A.-M. and Etzioni, O. (2005). Extracting product features and opinions from reviews. In Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP). Riloff, E. and Wiebe, J. (2003). Learning extraction patterns 593 for subjective expressions. In Proceedings of the 2003 Conference on Empirical methods in Natural Language Processing (EMNLP). Weerkamp, W., Balog, K., and de Rijke, M. (2009). A generative blog post retrieval model that uses query expansion based on external collections. In Joint conference of the 47th Annual Meeting of the Association for Computational Linguistics and the 4th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing (ACL-ICNLP 2009), Singapore. Weerkamp, W. and de Rijke, M. (2008). Credibility improves topical blog post retrieval. In Proceedings of ACL08: HLT, page 923931, Columbus, Ohio. Association for Computational Linguistics, Association for Computational Linguistics. Wilson, T., Wiebe, J., and Hoffmann, P. (2005). Recognizing contextual polarity in phrase-level sentiment analysis. In HLT ’05: Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, pages 347–354, Morristown, NJ, USA. Association for Computational Linguistics. Wilson, T., Wiebe, J., and Hoffmann, P. (2009). Recognizing contextual polarity: an exploration of features for phrase-level sentiment analysis. Computational Linguistics, 35(3):399–433. 594
|
2010
|
60
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 595–603, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Evaluating Multilanguage-Comparability of Subjectivity Analysis Systems Jungi Kim, Jin-Ji Li and Jong-Hyeok Lee Division of Electrical and Computer Engineering Pohang University of Science and Technology, Pohang, Republic of Korea {yangpa,ljj,jhlee}@postech.ac.kr Abstract Subjectivity analysis is a rapidly growing field of study. Along with its applications to various NLP tasks, much work have put efforts into multilingual subjectivity learning from existing resources. Multilingual subjectivity analysis requires language-independent criteria for comparable outcomes across languages. This paper proposes to measure the multilanguage-comparability of subjectivity analysis tools, and provides meaningful comparisons of multilingual subjectivity analysis from various points of view. 1 Introduction The field of NLP has seen a recent surge in the amount of research on subjectivity analysis. Along with its applications to various NLP tasks, there have been efforts made to extend the resources and tools created for the English language to other languages. These endeavors have been successful in constructing lexicons, annotated corpora, and tools for subjectivity analysis in multiple languages. There are multilingual subjectivity analysis systems available that have been built to monitor and analyze various concerns and opinions on the Internet; among the better known are OASYS from the University of Maryland that analyzes opinions on topics from news article searches in multiple languages (Cesarano et al., 2007)1 and TextMap, an entity search engine developed by Stony Brook University for sentiment analysis along with other functionalities (Bautin et al., 2008).2 Though these systems currently rely on English analysis tools and a machine translation (MT) technology to 1http://oasys.umiacs.umd.edu/oasysnew/ 2http://www.textmap.com/ translate other languages into English, up-to-date research provides various ways to analyze subjectivity in multilingual environments. Given sentiment analysis systems in different languages, there are many situations when the analysis outcomes need to be multilanguagecomparable. For example, it has been common these days for the Internet users across the world to share their views and opinions on various topics including music, books, movies, and global affairs and incidents, and also multinational companies such as Apple and Samsung need to analyze customer feedbacks for their products and services from many countries in different languages. Governments may also be interested in monitoring terrorist web forums or its global reputation. Surveying these opinions and sentiments in various languages involves merging the analysis outcomes into a single database, thereby objectively comparing the result across languages. If there exists an ideal subjectivity analysis system for each language, evaluating the multilanguage-comparability would be unnecessary because the analysis in each language would correctly identify the exact meanings of all input texts regardless of the language. However, this requirement is not fulfilled with current technology, thus the need for defining and measuring the multilanguage-comparability of subjectivity analysis systems is evident. This paper proposes to evaluate the multilanguage-comparability of multilingual subjectivity analysis systems. We build a number of subjectivity classifiers that distinguishes subjective texts from objective ones, and measure the multilanguage-comparability according to our proposed evaluation method. Since subjectivity analysis tools in languages other than English are not readily available, we focus our experiments on comparing different methods to build multilingual analysis systems from the resources and systems 595 created for English. These approaches enable us to extend a monolingual system to many languages with a number of freely available NLP resources and tools. 2 Related Work Much research have been put into developing methods for multilingual subjectivity analysis recently. With the high availability of subjectivity resources and tools in English, an easy and straightforward approach would be to employ a machine translation (MT) system to translate input texts in target languages into English then carry out the analyses using an existing subjectivity analysis tool (Kim and Hovy, 2006; Bautin et al., 2008; Banea et al., 2008). Mihalcea et al. (2007) and Banea et al. (2008) proposed a number of approaches exploiting a bilingual dictionary, a parallel corpus, and an MT system to port the resources and systems available in English to languages with limited resources. For subjectivity lexicons translation, Mihalcea et al. (2007) and Wan (2008) used the first sense in a bilingual dictionary, Kim and Hovy (2006) used a parallel corpus and a word alignment tool to extract translation pairs, and Kim et al. (2009) used a dictionary to translate and a link analysis algorithm to refine the matching intensity. To overcome the shortcomings of available resources and to take advantage of ensemble systems, Wan (2008) and Wan (2009) explored methods for developing a hybrid system for Chinese using English and Chinese sentiment analyzers. Abbasi et al. (2008) and Boiy and Moens (2009) have created manually annotated gold standards in target languages and studied various feature selection and learning techniques in machine learning approaches to analyze sentiments in multilingual web documents. For learning multilingual subjectivity, the literature tentatively concludes that translating lexicon is less dependable in terms of preserving subjectivity than corpus translation (Mihalcea et al., 2007; Wan, 2008), and though corpus translation results in modest performance degradation, it provides a viable approach because no manual labor is required (Banea et al., 2008; Brooke et al., 2009). Based on the observation that the performances of subjectivity analysis systems in comparable experimental settings for two languages differ, Texts with an identical negative sentiment: * The iPad could cannibalize the e-reader market. * 아이패드가(iPad) 전자책 시장을(e-reader market) 위축시킬 수 있다(could cannibalize). Texts with different strengths of positive sentiments: * Samsung cell phones have excellent battery life. * 삼성(Samsung) 휴대전화(cell phone) 배터리는 (battery) 그럭저럭(somehow or other) 오래간다(last long). Figure 1: Examples of sentiments in multilingual text Banea et al. (2008) have attributed the variations in the difficulty level of subjectivity learning to the differences in language construction. Bautin et al. (2008)’s system analyzes the sentiment scores of entities in multilingual news and blogs and adjusted the sentiment scores using entity sentiment probabilities of languages. 3 Multilanguage-Comparability 3.1 Motivation The quality of a subjectivity analysis tool is measured by its ability to distinguish subjectivity from objectivity and/or positive sentiments from negative sentiments. Additionally, a multilingual subjectivity analysis system is required to generate unbiased analysis results across languages; the system should base its outcome solely on the subjective meanings of input texts irrespective of the language, and the equalities and inequalities of subjectivity labels and intensities must be useful within and throughout the languages. Let us consider two cases where the pairs of multilingual inputs in English and Korean have identical and different subjectivity meanings (Figure 1). The first pair of texts carry a negative sentiment about how the release of a new electronics device might affect an emerging business market. When a multilanguage-comparable system is inputted with such a pair, its output should appropriately reflect the negative sentiment, and be identical for both texts. The second pair of texts share a similar positive sentiment about a mobile device’s battery capacity but with different strengths. A good multilingual system must be able to identify the positive sentiments and distinguish the differences in their intensities. However, these kinds of conditions cannot be measured with performance evaluations indepen596 dently carried out on each language; A system with a dissimilar ability to analyze subjective expressions from one language to another may deliver opposite labels or biased scores on texts with an identical subjective meaning, and vice versa, but still might produce similar performances on the evaluation data. Macro evaluations on individual languages cannot provide any conclusions on the system’s multilanguage-comparability capability. To measure how much of a system’s judgment principles are preserved across languages, an evaluation from a different perspective is necessary. 3.2 Evaluation Approach An evaluation of multilanguage-comparability may be done in two ways: measuring agreements in the outcomes of a pair of multilingual texts with an identical subjective meaning, or measuring the consistencies in the label and/or accordance in the order of intensity of a pair of texts with different subjectivities. There are advantages and disadvantages to each approaches. The first approach requires multilingual texts aligned at the level of specificity, for instance, document, sentence and phrase, that the subjectivity analysis system works. Text corpora for MT evaluation such as newspapers, books, technical manuals, and government official records provide a wide variety of parallel texts, typically at the sentence level. Annotating these types of corpus can be efficient; as parallel texts must have identical semantic meanings, subjectivity–related annotations for one language can be projected into other languages without much loss of accuracy. The latter approach accepts any pair of multilingual texts as long as they are annotated with labels and/or intensity. In this case, evaluating the label consistency of a multilingual system is only as difficult as evaluating that of a monolingual system; we can produce all possible pairs of texts from test corpora annotated with labels for each language. Evaluating with intensity is not easy for the latter approach; if test corpora already exist with intensity annotations for both languages, normalizing the intensity scores to a comparable scale is necessary (yet is uncertain unless every pair is checked manually), otherwise every pair of multilingual texts needs a manual annotation with its relative order of intensity. In this paper, we utilize the first approach because it provides a more rational means; we can reasonably hypothesize that text translated into another language by a skilled translator carries an identical semantic meaning and thereby conveys identical subjectivity. Therefore the required resource is more easily attained in relatively inexpensive ways. For evaluation, we measure the consistency in the subjectivity labels and the correlation of subjectivity intensity scores of parallel texts. Section 5.1 describes the details of evaluation metrics. 4 Multilingual Subjectivity System We create a number of multilingual systems consisting of multiple subsystems each processing a language, where one system analyzes English, and the other systems analyze the Korean, Chinese, and Japanese languages. We try to reproduce a set of systems using diverse methods in order to compare the systems and find out which methods are more suitable for multilanguage-comparability. 4.1 Source Language System We adopt the three systems described below as our source language systems: a state-of-the-art subjectivity classifier, a corpus-based, and a lexiconbased systems. The resources needed for developing the systems or the system itself are readily available for research purposes. In addition, these systems cover the general spectrum of current approaches to subjectivity analysis. State-of-the-art (S-SA): OpinionFinder is a publicly-available NLP tool for subjectivity analysis (Wiebe and Riloff, 2005; Wilson et al., 2005).3 The software and its resources have been widely used in the field of subjectivity analysis, and it has been the de facto standard system against which new systems are validated. We use a highcoverage classifier from the OpinionFinder’s two sentence-level subjectivity classifiers. This Naive Bayes classifier builds upon a corpus annotated by a high-precision classifier with the bootstrapping of the corpus and extraction patterns. The classifier assesses a sentence’s subjectivity with a label and a score for confidence in its judgment. Corpus-based (S-CB): The MPQA opinion corpus is a collection of 535 newspaper articles in English annotated with opinions and private states at 3http://www.cs.pitt.edu/mpqa/opinionfinderrelease/, version 1.5 597 the sub-sentence level (Wiebe et al., 2003).4 We retrieve the sentence level subjectivity labels for 11,111 sentences using the set of rules described in (Wiebe and Riloff, 2005). The corpus provides a relatively balanced corpus with 55% subjective sentences. We train an ML-based classifier using the corpus. Previous studies have found that, among several ML-based approaches, the SVM classifier generally performs well in many subjectivity analysis tasks (Pang et al., 2002; Banea et al., 2008). We use SVMLight with its default configurations,5 inputted with a sentence represented as a feature vector of word unigrams and their counts in the sentence. An SVM score (a margin or the distance from a learned decision boundary) with a positive value predicts the input as being subjective, and negative value as objective. Lexicon-based (S-LB): OpinionFinder contains a list of English subjectivity clue words with intensity labels (Wilson et al., 2005). The lexicon is compiled from several manually and automatically built resources and contains 6885 unique entries. Riloff and Wiebe (2003) constructed a highprecision classifier for contiguous sentences using the number of strong and weak subjective words in current and nearby sentences. Unlike previous work, we do not (or rather, cannot) maintain assumptions about the proximity of input text. Using the lexicon, we build a simple and highcoverage rule-based subjectivity classifier. Setting the scores of strong and weak subjective words as 1.0 and 0.5, we evaluate the subjectivity of a given sentence as the sum of subjectivity scores; above a threshold, the input is subjective, and otherwise objective. The threshold value is optimized for an F-measure using the MPQA corpus, and is set to 1.0 throughout our experiments. 4.2 Target Language System To construct a target language system leveraging on available resources in the source language, we consider three approaches from previous literature: 1. translating test sentences in target language into source language and inputting them into 4http://www.cs.pitt.edu/mpqa/databaserelease/, version 1.2 5http://svmlight.joachims.org/, version 6.02 a source language system (Kim and Hovy, 2006; Bautin et al., 2008; Banea et al., 2008) 2. translating a source language training corpus into target language and creating a corpusbased system in target language (Banea et al., 2008) 3. translating a subjectivity lexicon from source language to target language and creating a lexicon-based system in target language (Mihalcea et al., 2007) Each approach has its advantages and disadvantages. The advantage of the first approach is its simple architecture, clear separation of subjectivity and MT systems, and that it has only one subjectivity system, and is thus easier to maintain. Its disadvantage is that the time-consuming MT has to be executed for each text input. In the second and third approaches, a subjectivity system in the target language is constructed sharing corpora, rules, and/or features with the source language system. Later on, it may also include its own set of resources specifically engineered for the target language as a performance improvement. However, keeping the systems up-to-date would require as much effort as the number of languages. All three approaches use MT, and would suffer significantly if the translation results are poor. Using the first approach, we can easily adopt all three source language systems; • Target input translated into source, analyzed by source language system S-SA • Target input translated into source, analyzed by source language system S-CB • Target input translated into source, analyzed by source language system S-LB The second and the third approaches are carried out as follows: Corpus-based (T-CB): We translate the MPQA corpus into the target languages sentence by sentence using a web-based service.6 Using the same method for S-CB, we train an SVM model for each language with the translated training corpora. Lexicon-based (T-LB): This classifier is identical to S-LB, where the English lexicon is replaced by one of the target languages. We automatically translate the lexicon using free bilingual dictionaries.7 First, the entries in the lexicon are looked 6Google Translate (http://translate.google.com/) 7quick english-korean, quick eng-zh CN, and JMDict from StarDict (http://stardict.sourceforge.net/) licensed under GPL and EDRDG. 598 Table 1: Agreement on subjectivity (S for subjective, O objective) of 859 sentence chunks in Korean between two annotators (An. 1 and An. 2). An. 2 S O Total An. 1 S 371 93 464 O 23 372 395 Total 394 465 859 up in the dictionary, if they are found, we select the first word in the first sense of the definition. If the entry is not in the dictionary, we lemmatize it,8 then repeat the search. Our simple approach produces moderate-sized lexicons (3,808, 3,980, 3,027 for Korean, Chinese, and Japanese) compared to Mihalcea et al. (2007)’s complicated translation approach (4,983 Romanian words). The threshold values are optimized using the MPQA corpus translated into each target language.9 5 Experiment 5.1 Experimental Setup Test Corpus Our evaluation corpus consists of 50 parallel newspaper articles from the Donga Daily News Website.10 The website provides news articles in Korean and their human translations in English, Japanese, and Chinese. We selected articles that contain Editorial in its English title from a 30day period. Three human annotators who are fluent in the two languages manually annotated Nto-N sentence alignments for each language pairs (KR-EN, KR-CH, KR-JP). By keeping only the sentence chunks whose Korean chunk appears in all language pairs, we were left with 859 sentence chunk pairs. The corpus was preprocessed with NLP tools for each language,11 and the Korean, Chinese, and Japanese texts were translated into English with the same web-based service used to translate the training corpus in Section 4.2. Manual Annotation and Agreement Study 8JWI (http://projects.csail.mit.edu/jwi/) 9Korean 1.0, Chinese 1.0, and Japanese 0.5 10http://www.donga.com/ 11Stanford POS Tagger 1.5.1 and Stanford Chinese Word Segmenter 2008-05-21 (http://nlp.stanford.edu/software/), Chasen 2.4.4 (http://chasen-legacy.sourceforge.jp/), Korean Morphological Analyzer (KoMA) (http://kle.postech.ac.kr/) Table 2: Agreement on projection of subjectivity (S for subjective, O objective) from Korean (KR) to English (EN) by one annotator. EN S O Total KR S 458 6 464 O 12 383 395 Total 470 389 859 To assess the performance of our subjectivity analysis systems, the Korean sentence chunks were manually annotated by two native speakers of Korean with Subjective and Objective labels (Table 1). A proportion agreement of 0.86 and a kappa value of 0.73 indicate a substantial agreement between the two annotators. We set aside 743 sentence chunks that both annotators agreed on for the automatic evaluation of subjectivity analysis systems, thereby removing the borderline cases, which are difficult even for humans to assess. The corresponding sentence chunks for other languages were extracted and tagged with labels equivalent to Korean chunks. In addition, to verify how consistently the subjectivity of the original texts is projected to the translated, we carried out another manual annotation and agreement study with Korean and English sentence chunks (Table 2). Note that our cross-lingual agreement study is similar to the one carried out by Mihalcea et al. (2007), where two annotators labeled the sentence subjectivity of a parallel text in different languages. They reported that, similarly to monolingual annotations, most cases of disagreements on annotations are due to the differences in the annotators’ judgments on subjectivity, and the rest from subjective meanings lost in the translation process and figurative language such as irony. To avoid the role played by annotators’ private views from disagreements, the subjectivity of sentence chunks in English were manually annotated by one of the annotators for the Korean text. Judged by the same annotator, we speculate that the disagreement in the annotation should account only for the inconsistency in the subjectivity projection. By proportion, the agreement between the annotation of Korean and English is 0.97, and the kappa is 0.96, suggesting an almost perfect agreement. Only a small number of sentence chunk pairs have inconsistent labels; six chunks in Ko599 Implicit sentiment expressed through translation: * 시간이 갈수록(with time) 그 격차가(disparity/gap) 벌어지고 있다(widening). * Worse, the (economic) disparity (between South Korea and North Korea) is worsening with time. Sentiment lost in translation: * 인도의 타타 자동차회사는(India's Tata Motors) 2200달러짜리 자동차 나노를(2,200-dollar automobile Nano) 내놓아(presented) 주목을 끌었다 (drew attention). * India's Tata Motors has produced the 2,200-dollar subcompact Nano. Figure 2: Excerpts from Donga Daily News with differing sentiments between parallel texts rean lost subjectivity in translation, and implied subjective meanings in twelve chunks were expressed explicitly through interpretation. Excerpts from our corpus show two such cases (Figure 2). Evaluation Metrics To evaluate the multilanguage-comparability of subjectivity analysis systems, we measure 1) how consistently the system assigns subjectivity labels and 2) how closely numeric scores for systems’ confidences correlate with regard to parallel texts in different languages. In particular, we use Cohen’s kappa coefficient for the first and Pearson’s correlation coefficient for the latter. These widely used metrics provide useful comparability measures for categorical and quantitative data. Both coefficients are scaled from −1 to +1, indicating negative to positive correlations. Kappa measures are corrected for chance, thereby yielding better measurements than agreement by proportion. The characteristics of Pearson’s correlation coefficient that it measures linear relationships and is independent of change in origin, scale, and unit comply with our experiments. 5.2 Subjectivity Classification Our multilingual subjectivity analysis systems were evaluated on the test corpora described in Section 5.1 (Table 3). Due to the difference in testbeds, the performance of the state-of-the-art English system (SSA) on our corpus is lower by about 10% relatively than the performance reported on the MPQA corpus.12 However, it still performs sufficiently 12precision, recall, and F-measure of 79.4, 70.6, and 74.7. well and provides the most balanced results among the three source language systems; The corpusbased system (S-CB) classifies with a high precision, and the lexicon-based (S-LB) with a high recall. The source language systems (S-SA,-CB,LB) lose a small percentage in precision when inputted with translations, but the recalls are generally on a par or even higher in the target languages. For the systems created from target language resources, Corpus-based systems (T-CB) generally perform better than the ones with source language resource (S-CB), and lexicon-based systems (TLB) perform worse than (S-LB). Similarly to systems with source language resources, T-CB classifies with a high precision and T-LB with a high recall, but the gap is less. Among the target languages, Korean tends to have a higher precision, and Japanese a higher recall than other languages in most systems. Overall, S-SA provides easy accessibility when analyzing both the source and the target languages, with a balanced precision and recall performance. Among the other approaches, only T-CB is better in all measures than S-SA, and S-LB performs best on F-measure evaluations. 5.3 Multilanguage-Comparability The evaluation results on multilanguagecomparability are presented in Table 4. The subjectivity analysis systems are evaluated with all language pairs with kappa and Pearson’s correlation coefficients. Kappa and Pearson’s correlation values are consistent with each other; Pearson’s correlation between the two evaluation measures is 0.91. We observe a distinct contrast in performances between corpus-based systems (S-CB and T-CB) and lexicon-based systems (S-LB and T-LB); All corpus-based systems show moderate agreements while agreements on lexicon-based systems are only fair. Within corpus-based systems, S-CB performs better with language pairs that include English, and T-CB performs better with language pairs of the target languages. For lexicon-based systems, systems in the target languages (T-LB) performs the worst with only slight to fair agreements between languages. Lexicon-based systems and state-of-the-art systems in the source language (S-LB and S-SA) result in average performances. 600 Table 3: Performance of subjectivity analysis with precision (P), recall (R), and F-measure (F). S-SA,CB,-LB systems in Korean, Chinese, Japanese indicate English analysis systems inputted with translations of the target languages into English. English Korean Chinese Japanese P R F P R F P R F P R F S-SA 71.1 63.5 67.1 70.7 61.1 65.6 67.3 68.8 68.0 69.1 67.5 68.3 S-CB 74.4 53.9 62.5 74.5 52.2 61.4 71.1 63.3 67.0 72.9 65.3 68.9 S-LB 62.5 87.7 73.0 62.9 87.7 73.3 59.9 91.5 72.4 61.8 94.1 74.6 T-CB 72.4 67.5 69.8 75.0 66.2 70.3 72.5 70.3 71.4 T-LB 59.4 71.0 64.7 58.4 82.3 68.2 56.9 92.4 70.4 Table 4: Performance of multilanguage-comparability: kappa coefficient (κ) for measuring comparability of classification labels and Pearson’s correlation coefficient (ρ) for classification scores for English (EN), Korean (KR), Chinese (CH), and Japanese (JP). Evaluations of T-CB,-LB for language pairs including English are carried out with results from S-CB,-LB for English and T-CB,-LB for target languages. S-SA S-CB S-LB T-CB T-LB κ ρ κ ρ κ ρ κ ρ κ ρ EN & KR 0.41 0.55 0.45 0.60 0.37 0.59 0.42 0.60 0.25 0.41 EN & CH 0.39 0.54 0.41 0.62 0.33 0.52 0.39 0.57 0.22 0.38 EN & JP 0.39 0.53 0.43 0.65 0.30 0.59 0.40 0.59 0.15 0.33 KR & CH 0.36 0.54 0.39 0.59 0.28 0.57 0.46 0.64 0.23 0.37 KR & JP 0.37 0.60 0.44 0.69 0.50 0.69 0.63 0.76 0.18 0.38 CH & JP 0.37 0.53 0.49 0.66 0.29 0.57 0.46 0.63 0.22 0.46 Average 0.38 0.55 0.44 0.64 0.35 0.59 0.46 0.63 0.21 0.39 -100 -50 0 50 100 -100 -50 0 50 100 (a) S-SA -4 -3 -2 -1 0 1 2 3 4 -4 -3 -2 -1 0 1 2 3 4 (b) S-CB -10 -5 0 5 10 -10 -5 0 5 10 (c) S-LB -4 -3 -2 -1 0 1 2 3 4 -4 -3 -2 -1 0 1 2 3 4 (d) T-CB -10 -5 0 5 10 -10 -5 0 5 10 (e) T-LB Figure 3: Scatter plots of English (x-axis) and Korean (y-axis) subjectivity scores from state-of-the-art (S-SA), corpus-based (S-CB), and lexicon-based (S-LB) systems of the source language, and corpusbased with translated corpora (T-CB), and lexicon-based with translated lexicon (T-LB) systems. Slanted lines in figures are best-fit lines through the origins. 601 Figure 3 shows scatter plots of subjectivity scores of our English and Korean test corpora evaluated on different systems; the data points on the first and the third quadrants are occurrences of label agreements, and the second and the fourth are disagreements. Linearly scattered data points are more correlated regardless of the slope. Figure 3a shows a moderate correlation for multilingual results from the state-of-the-art system (S-SA). Agreements on objective instances are clustered together while agreements on subjective instances are diffused over a wide region. Agreements between the source language corpus-based system (S-CB) and the corpus-based system trained with translated resources (T-CB) are more distinctively correlated than the results for other pairs of systems (Figures 3b and 3d). We notice that S-CB seems to have a lower number of outliers than T-CB, but slightly more diffusive. Lexicon-based systems (S-LB, T-LB) generate noticeably uncorrelated scores (Figures 3c and 3e). We observe that the results from the English system with translated inputs (S-LB) is more correlated than those from systems with translated lexicons (T-LB), and that analysis results from both systems are biased toward subjective scores. 6 Discussion Which approach is most suitable for multilingual subjectivity analysis? In our experiments, the corpus-based systems trained on corpora translated from English to the target languages (T-CB) perform well for subjectivity classification and multilanguagecomparability measures on the whole. However, the methods we employed to expand the languages were naively carried out without much considerations for optimization. Further adjustments could improve the other systems for both classification and multilanguage-comparability performances. Is there a correlation between classification performance and multilanguage-comparability? Lexicon-based systems in the source language (S-LB) have good overall classification performances, especially on recall and F-measures. However, these systems performs worse on multilanguage-comparability than other systems with poorer classification performances. Intrigued by the observation, we tried to measure which criteria for classification performance influences multilanguage-comparability. We again employed Pearson’s correlation metrics to measure the correlations of precision (P), recall (R), and F-measures (F) to kappa (κ) and Pearson’s correlation (ρ) values. Specifically, we measure the correlations between the sums of P, the sums of R, and the sums of F to κ and ρ for all pairs of systems.13 The correlations of P with κ and ρ are 0.78 and 0.68, R −0.38 and −0.28, and F −0.20 and −0.05. These numbers strongly suggest that multilanguage-comparability correlates with the precisions of classifiers. However, we cannot always expect a highprecision multilingual subjectivity classifier to be multilanguage-comparable as well. For example, the S-SA system has a much higher precision than S-LB consistently over all languages, but their multilanguage-comparability performances differed only by small amounts. 7 Conclusion Multilanguage-comparability is an analysis system’s ability to retain its decision criteria across different languages. We implemented a number of previously proposed approaches to learning multilingual subjectivity, and evaluated the systems on multilanguage-comparability as well as classification performance. Our experimental results provide meaningful comparisons of the multilingual subjectivity analysis systems across various aspects. Also, we developed a multilingual subjectivity evaluation corpus from a parallel text, and studied inter-annotator, inter-language agreements on subjectivity, and observed persistent subjectivity projections from one language to another from a parallel text. For future work, we aim extend this work to constructing a multilingual sentiment analysis system and evaluate it with multilingual datasets such as product reviews collected from different countries. We also plan to resolve the lexiconbased classifiers’ classification bias towards subjective meanings with a list of objective words (Esuli and Sebastiani, 2006) and their multilingual expansion (Kim et al., 2009), and evaluate the multilanguage-comparability of systems constructed with resources from different sources. 13Pairs of values such as 71.1 + 70.7 and 0.41 for precisions and Kappa of S-SA for English and Korean. 602 Acknowledgement We thank the anonymous reviewers for valuable comments and helpful suggestions. This work is supported in part by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (MEST) (20090075211), and in part by the BK 21 project in 2010. References Ahmed Abbasi, Hsinchun Chen, and Arab Salem. 2008. Sentiment analysis in multiple languages: Feature selection for opinion classification in web forums. ACM Transactions on Information Systems, 26(3):1–34. Carmen Banea, Rada Mihalcea, Janyce Wiebe, and Samer Hassan. 2008. Multilingual subjectivity analysis using machine translation. In EMNLP ’08: Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 127– 135, Morristown, NJ, USA. Mikhail Bautin, Lohit Vijayarenu, and Steven Skiena. 2008. International sentiment analysis for news and blogs. In Proceedings of the International Conference on Weblogs and Social Media (ICWSM). Erik Boiy and Marie-Francine Moens. 2009. A machine learning approach to sentiment analysis in multlingual Web texts. Information Retrieval, 12:526–558. Julian Brooke, Milan Tofiloski, and Maite Taboada. 2009. Cross-linguistic sentiment analysis: From english to spanish. In Proceedings of RANLP 2009, Borovets, Bulgaria. Carmine Cesarano, Antonio Picariello, Diego Reforgiato, and V.S. Subrahmanian. 2007. The oasys 2.0 opinion analysis system: A demo. In Proceedings of the International Conference on Weblogs and Social Media (ICWSM). Andrea Esuli and Fabrizio Sebastiani. 2006. Sentiwordnet: A publicly available lexical resource for opinion mining. In Proceedings of the 5th Conference on Language Resources and Evaluation (LREC’06), pages 417–422, Geneva, IT. Soo-Min Kim and Eduard Hovy. 2006. Identifying and analyzing judgment opinions. In Proceedings of the Human Language Technology Conference of the NAACL (HLT/NAACL’06), pages 200–207, New York, USA. Jungi Kim, Hun-Young Jung, Sang-Hyob Nam, Yeha Lee, and Jong-Hyeok Lee. 2009. Found in translation: Conveying subjectivity of a lexicon of one language into another using a bilingual dictionary and a link analysis algorithm. In ICCPOL ’09: Proceedings of the 22nd International Conference on Computer Processing of Oriental Languages. Language Technology for the Knowledge-based Economy, pages 112–121, Berlin, Heidelberg. Rada Mihalcea, Carmen Banea, and Janyce Wiebe. 2007. Learning multilingual subjective language via cross-lingual projections. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics (ACL’07), pages 976–983, Prague, CZ. Bo Pang, Lillian Lee, and Shivakumar Vaithyanathan. 2002. Thumbs up? Sentiment classification using machine learning techniques. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 79–86. Ellen Riloff and Janyce Wiebe. 2003. Learning extraction patterns for subjective expressions. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP). Xiaojun Wan. 2008. Using bilingual knowledge and ensemble techniques for unsupervised Chinese sentiment analysis. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 553–561, Honolulu, Hawaii, October. Association for Computational Linguistics. Xiaojun Wan. 2009. Co-training for cross-lingual sentiment classification. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 235–243, Suntec, Singapore, August. Association for Computational Linguistics. Janyce Wiebe and Ellen Riloff. 2005. Creating subjective and objective sentence classifiers from unannotated texts. In Proceedings of the 6th International Conference on Intelligent Text Processing and Computational Linguistics (CICLing-2005), pages 486– 497, Mexico City, Mexico. Janyce Wiebe, E. Breck, Christopher Buckley, Claire Cardie, P. Davis, B. Fraser, Diane Litman, D. Pierce, Ellen Riloff, Theresa Wilson, D. Day, and Mark Maybury. 2003. Recognizing and organizing opinions expressed in the world press. In Proceedings of the 2003 AAAI Spring Symposium on New Directions in Question Answering. Theresa Wilson, Janyce Wiebe, and Paul Hoffmann. 2005. Recognizing contextual polarity in phraselevel sentiment analysis. In Proceedings of the Conference on Human Language Technology and Empirical Methods in Natural Language Processing (HLT-EMNLP’05), pages 347–354, Vancouver, CA. 603
|
2010
|
61
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 604–611, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Error Detection for Statistical Machine Translation Using Linguistic Features Deyi Xiong, Min Zhang, Haizhou Li Human Language Technology Institute for Infocomm Research 1 Fusionopolis Way, #21-01 Connexis, Singapore 138632. {dyxiong, mzhang, hli}@i2r.a-star.edu.sg Abstract Automatic error detection is desired in the post-processing to improve machine translation quality. The previous work is largely based on confidence estimation using system-based features, such as word posterior probabilities calculated from Nbest lists or word lattices. We propose to incorporate two groups of linguistic features, which convey information from outside machine translation systems, into error detection: lexical and syntactic features. We use a maximum entropy classifier to predict translation errors by integrating word posterior probability feature and linguistic features. The experimental results show that 1) linguistic features alone outperform word posterior probability based confidence estimation in error detection; and 2) linguistic features can further provide complementary information when combined with word confidence scores, which collectively reduce the classification error rate by 18.52% and improve the F measure by 16.37%. 1 Introduction Translation hypotheses generated by a statistical machine translation (SMT) system always contain both correct parts (e.g. words, n-grams, phrases matched with reference translations) and incorrect parts. Automatically distinguishing incorrect parts from correct parts is therefore very desirable not only for post-editing and interactive machine translation (Ueffing and Ney, 2007) but also for SMT itself: either by rescoring hypotheses in the N-best list using the probability of correctness calculated for each hypothesis (Zens and Ney, 2006) or by generating new hypotheses using Nbest lists from one SMT system or multiple systems (Akibay et al., 2004; Jayaraman and Lavie, 2005). In this paper we restrict the “parts” to words. That is, we detect errors at the word level for SMT. A common approach to SMT error detection at the word level is calculating the confidence at which a word is correct. The majority of word confidence estimation methods follows three steps: 1) Calculate features that express the correctness of words either based on SMT model (e.g. translation/language model) or based on SMT system output (e.g. N-best lists, word lattices) (Blatz et al., 2003; Ueffing and Ney, 2007). 2) Combine these features together with a classification model such as multi-layer perceptron (Blatz et al., 2003), Naive Bayes (Blatz et al., 2003; Sanchis et al., 2007), or loglinear model (Ueffing and Ney, 2007). 3) Divide words into two groups (correct translations and errors) by using a classification threshold optimized on a development set. Sometimes the step 2) is not necessary if only one effective feature is used (Ueffing and Ney, 2007); and sometimes the step 2) and 3) can be merged into a single step if we directly output predicting results from binary classifiers instead of making thresholding decision. Various features from different SMT models and system outputs are investigated (Blatz et al., 2003; Ueffing and Ney, 2007; Sanchis et al., 2007; Raybaud et al., 2009). Experimental results show that they are useful for error detection. However, it is not adequate to just use these features as discussed in (Shi and Zhou, 2005) because the information that they carry is either from the inner components of SMT systems or from system outputs. To some extent, it has already been considered by SMT systems. Hence finding external information 604 sources from outside SMT systems is desired for error detection. Linguistic knowledge is exactly such a good choice as an external information source. It has already been proven effective in error detection for speech recognition (Shi and Zhou, 2005). However, it is not widely used in SMT error detection. The reason is probably that people have yet to find effective linguistic features that outperform nonlinguistic features such as word posterior probability features (Blatz et al., 2003; Raybaud et al., 2009). In this paper, we would like to show an effective use of linguistic features in SMT error detection. We integrate two sets of linguistic features into a maximum entropy (MaxEnt) model and develop a MaxEnt-based binary classifier to predict the category (correct or incorrect) for each word in a generated target sentence. Our experimental results show that linguistic features substantially improve error detection and even outperform word posterior probability features. Further, they can produce additional improvements when combined with word posterior probability features. The rest of the paper is organized as follows. In Section 2, we review the previous work on wordlevel confidence estimation which is used for error detection. In Section 3, we introduce our linguistic features as well as the word posterior probability feature. In Section 4, we elaborate our MaxEntbased error detection model which combine linguistic features and word posterior probability feature together. In Section 5, we describe the SMT system which we use to generate translation hypotheses. We report our experimental results in Section 6 and conclude in Section 7. 2 Related Work In this section, we present an overview of confidence estimation (CE) for machine translation at the word level. As we are only interested in error detection, we focus on work that uses confidence estimation approaches to detect translation errors. Of course, confidence estimation is not limited to the application of error detection, it can also be used in other scenarios, such as translation prediction in an interactive environment (Grandrabur and Foster, 2003) . In a JHU workshop, Blatz et al. (2003) investigate using neural networks and a naive Bayes classifier to combine various confidence features for confidence estimation at the word level as well as at the sentence level. The features they use for word level CE include word posterior probabilities estimated from N-best lists, features based on SMT models, semantic features extracted from WordNet as well as simple syntactic features, i.e. parentheses and quotation mark check. Among all these features, the word posterior probability is the most effective feature, which is much better than linguistic features such as semantic features, according to their final results. Ueffing and Ney (2007) exhaustively explore various word-level confidence measures to label each word in a generated translation hypothesis as correct or incorrect. All their measures are based on word posterior probabilities, which are estimated from 1) system output, such as word lattices or N-best lists and 2) word or phrase translation table. Their experimental results show that word posterior probabilities directly estimated from phrase translation table are better than those from system output except for the Chinese-English language pair. Sanchis et al. (2007) adopt a smoothed naive Bayes model to combine different word posterior probability based confidence features which are estimated from N-best lists, similar to (Ueffing and Ney, 2007). Raybaud et al. (2009) study several confidence features based on mutual information between words and n-gram and backward n-gram language model for word-level and sentence-level CE. They also explore linguistic features using information from syntactic category, tense, gender and so on. Unfortunately, such linguistic features neither improve performance at the word level nor at the sentence level. Our work departs from the previous work in two major respects. • We exploit various linguistic features and show that they are able to produce larger improvements than widely used system-related features such as word posterior probabilities. This is in contrast to some previous work. Yet another advantage of using linguistic features is that they are system-independent, which therefore can be used across different systems. • We treat error detection as a complete binary classification problem. Hence we di605 rectly output prediction results from our discriminatively trained classifier without optimizing a classification threshold on a distinct development set beforehand.1 Most previous approaches make decisions based on a pretuned classification threshold τ as follows class = { correct, Φ(correct, θ) > τ incorrect, otherwise where Φ is a classifier or a confidence measure and θ is the parameter set of Φ. The performance of these approaches is strongly dependent on the classification threshold. 3 Features We explore two sets of linguistic features for each word in a machine generated translation hypothesis. The first set of linguistic features are simple lexical features. The second set of linguistic features are syntactic features which are extracted from link grammar parse. To compare with the previously widely used features, we also investigate features based on word posterior probabilities. 3.1 Lexical Features We use the following lexical features. • wd: word itself • pos: part-of-speech tag from a tagger trained on WSJ corpus. 2 For each word, we look at previous n words/tags and next n words/tags. They together form a word/tag sequence pattern. The basic idea of using these features is that words in rare patterns are more likely to be incorrect than words in frequently occurring patterns. To some extent, these two features have similar function to a target language model or pos-based target language model. 3.2 Syntactic Features High-level linguistic knowledge such as syntactic information about a word is a very natural and promising indicator to decide whether this word is syntactically correct or not. Words occurring in an 1This does not mean we do not need a development set. We do validate our feature selection and other experimental settings on the development set. 2Available via http://www-tsujii.is.s.u-tokyo.ac.jp/ ∼tsuruoka/postagger/ ungrammatical part of a target sentence are prone to be incorrect. The challenge of using syntactic knowledge for error detection is that machinegenerated hypotheses are rarely fully grammatical. They are mixed with grammatical and ungrammatical parts, which hence are not friendly to traditional parsers trained on grammatical sentences because ungrammatical parts of a machinegenerated sentence could lead to a parsing failure. To overcome this challenge, we select the Link Grammar (LG) parser 3 as our syntactic parser to generate syntactic features. The LG parser produces a set of labeled links which connect pairs of words with a link grammar (Sleator and Temperley, 1993). The main reason why we choose the LG parser is that it provides a robustness feature: null-link scheme. The null-link scheme allows the parser to parse a sentence even when the parser can not fully interpret the entire sentence (e.g. including ungrammatical parts). When the parser fail to parse the entire sentence, it ignores one word each time until it finds linkages for remaining words. After parsing, those ignored words are not connected to any other words. We call them null-linked words. Our hypothesis is that null-linked words are prone to be syntactically incorrect. We hence straightforwardly define a syntactic feature for a word w according to its links as follows link(w) = { yes, w has links no, otherwise In Figure 1 we show an example of a generated translation hypothesis with its link parse. Here links are denoted with dotted lines which are annotated with link types (e.g., Jp, Op). Bracketed words, namely “,” and “including”, are null-linked words. 3.3 Word Posterior Probability Features Our word posterior probability is calculated on Nbest list, which is first proposed by (Ueffing et al., 2003) and widely used in (Blatz et al., 2003; Ueffing and Ney, 2007; Sanchis et al., 2007). Given a source sentence f, let {en}N 1 be the Nbest list generated by an SMT system, and let ei n is the i-th word in en. The major work of calculating word posterior probabilities is to find the Levenshtein alignment (Levenshtein, 1966) between the best hypothesis e1 and its competing hypothesis 3Available at http://www.link.cs.cmu.edu/link/ 606 Figure 1: An example of Link Grammar parsing results. en in the N-best list {en}N 1 . We denote the alignment between them as ℓ(e1, en). The word in the hypothesis en which ei 1 is Levenshtein aligned to is denoted as ℓi(e1, en). The word posterior probability of ei 1 is then calculated by summing up the probabilities over all hypotheses containing ei 1 in a position which is Levenshtein aligned to ei 1. pwpp(ei 1) = ∑ en: ℓi(e1,en)=ei 1 p(en) ∑N 1 p(en) To use the word posterior probability in our error detection model, we need to make it discrete. We introduce a feature for a word w based on its word posterior probability as follows dwpp(w) = ⌊−log(pwpp(w))/df⌋ where df is the discrete factor which can be set to 1, 0.1, 0.01 and so on. “⌊⌋” is a rounding operator which takes the largest integer that does not exceed −log(pwpp(w))/df. We optimize the discrete factor on our development set and find the optimal value is 1. Therefore a feature “dwpp = 2” represents that the logarithm of the word posterior probability is between -3 and -2; 4 Error Detection with a Maximum Entropy Model As mentioned before, we consider error detection as a binary classification task. To formalize this task, we use a feature vector ψ to represent a word w in question, and a binary variable c to indicate whether this word is correct or not. In the feature vector, we look at 2 words before and 2 words after the current word position (w−2, w−1, w, w1, w2). We collect features {wd, pos, link, dwpp} for each word among these words and combine them into the feature vector ψ for w. As such, we want the feature vector to capture the contextual environment, e.g., pos sequence pattern, syntactic pattern, where the word w occurs. For classification, we employ the maximum entropy model (Berger et al., 1996) to predict whether a word w is correct or incorrect given its feature vector ψ. p(c|ψ) = exp(∑ i θifi(c, ψ)) ∑ c′ exp(∑ i θifi(c′, ψ)) where fi is a binary model feature defined on c and the feature vector ψ. θi is the weight of fi. Table 1 shows some examples of our binary model features. In order to learn the model feature weights θ for probability estimation, we need a training set of m samples {ψi, ci}m 1 . The challenge of collecting training instances is that the correctness of a word in a generated translation hypothesis is not intuitively clear (Ueffing and Ney, 2007). We will describe the method to determine the correctness of a word in Section 6.1, which is broadly adopted in previous work. We tune our model feature weights using an off-the-shelf MaxEnt toolkit (Zhang, 2004). To avoid overfitting, we optimize the Gaussian prior on the development set. During test, if the probability p(correct|ψ) is larger than p(incorrect|ψ) according the trained MaxEnt model, the word is labeled as correct otherwise incorrect. 5 SMT System To obtain machine-generated translation hypotheses for our error detection, we use a state-of-the-art phrase-based machine translation system MOSES (Koehn et al., 2003; Koehn et al., 2007). The translation task is on the official NIST Chineseto-English evaluation data. The training data consists of more than 4 million pairs of sentences (including 101.93M Chinese words and 112.78M English words) from LDC distributed corpora. Table 2 shows the corpora that we use for the translation task. We build a four-gram language model using the SRILM toolkit (Stolcke, 2002), which is trained 607 Feature Example wd f(c, ψ) = { 1, ψ.w.wd = ”.”, c = correct 0, otherwise pos f(c, ψ) = { 1, ψ.w2.pos = ”NN”, c = incorrect 0, otherwise link f(c, ψ) = { 1, ψ.w.link = no, c = incorrect 0, otherwise dwpp f(c, ψ) = { 1, ψ.w−2.dwpp = 2, c = correct 0, otherwise Table 1: Examples of model features. LDC ID Description LDC2004E12 United Nations LDC2004T08 Hong Kong News LDC2005T10 Sinorama Magazine LDC2003E14 FBIS LDC2002E18 Xinhua News V1 beta LDC2005T06 Chinese News Translation LDC2003E07 Chinese Treebank LDC2004T07 Multiple Translation Chinese Table 2: Training corpora for the translation task. on Xinhua section of the English Gigaword corpus (181.1M words). For minimum error rate tuning (Och, 2003), we use NIST MT-02 as the development set for the translation task. In order to calculate word posterior probabilities, we generate 10,000 best lists for NIST MT-02/03/05 respectively. The performance, in terms of BLEU (Papineni et al., 2002) score, is shown in Table 4. 6 Experiments We conducted our experiments at several levels. Starting with MaxEnt models with single linguistic feature or word posterior probability based feature, we incorporated additional features incrementally by combining features together. In doing so, we would like the experimental results not only to display the effectiveness of linguistic features for error detection but also to identify the additional contribution of each feature to the task. 6.1 Data Corpus For the error detection task, we use the best translation hypotheses of NIST MT-02/05/03 generated by MOSES as our training, development, and test corpus respectively. The statistics about these corpora is shown in Table 3. Each translation hypothesis has four reference translations. Corpus Sentences Words Training MT-02 878 24,225 Development MT-05 1082 31,321 Test MT-03 919 25,619 Table 3: Corpus statistics (number of sentences and words) for the error detection task. To obtain the linkage information, we run the LG parser on all translation hypotheses. We find that the LG parser can not fully parse 560 sentences (63.8%) in the training set (MT-02), 731 sentences (67.6%) in the development set (MT-05) and 660 sentences (71.8%) in the test set (MT-03). For these sentences, the LG parser will use the the null-link scheme to generate null-linked words. To determine the true class of a word in a generated translation hypothesis, we follow (Blatz et al., 2003) to use the word error rate (WER). We tag a word as correct if it is aligned to itself in the Levenshtein alignment between the hypothesis and the nearest reference translation that has minimum edit distance to the hypothesis among four reference translations. Figure 2 shows the Levenshtein alignment between a machine-generated hypothesis and its nearest reference translation. The “Class” row shows the label of each word according to the alignment, where “c” and “i” represent correct and incorrect respectively. There are several other metrics to tag single words in a translation hypothesis as correct or incorrect, such as PER where a word is tagged as correct if it occurs in one of reference translations with the same number of occurrences, Set which is a less strict variant of PER, ignoring the number of occurrences per word. In Figure 2, the two words “last year” in the hypothesis will be tagged as correct if we use the PER or Set metric since they do not consider the occurring positions of words. Our 608 C
h
i
n
a
U
n
i
c
o
m
n
e
t
p
r
o
f
i
t
r
o
s
e
u
p
3
8
%
l
a
s
t
y
e
a
r
C
h
i
n
a
U
n
i
c
o
m
n
e
t
p
r
o
f
i
t
r
o
s
e
u
p
3
8
%
l
a
s
t
y
e
a
r
H
y
p
o
t
h
e
s
i
s
R
e
f
e
r
e
n
c
e
C
h
i
n
a
/
c
U
n
i
c
o
m
/
c
n
e
t
/
c
p
r
o
f
i
t
/
c
r
o
s
e
/
c
u
p
/
c
3
8
%
/
c
l
a
s
t
/
i
y
e
a
r
/
i
C
l
a
s
s
Figure 2: Tagging a word as correct/incorrect according to the Levenshtein alignment. Corpus BLEU (%) RCW (%) MT-02 33.24 47.76 MT-05 32.03 47.85 MT-03 32.86 47.57 Table 4: Case-insensitive BLEU score and ratio of correct words (RCW) on the training, development and test corpus. metric corresponds to the m-WER used in (Ueffing and Ney, 2007), which is stricter than PER and Set. It is also stricter than normal WER metric which compares each hypothesis to all references, rather than the nearest reference. Table 4 shows the case-insensitive BLEU score and the percentage of words that are labeled as correct according to the method described above on the training, development and test corpus. 6.2 Evaluation Metrics To evaluate the overall performance of the error detection, we use the commonly used metric, classification error rate (CER) to evaluate our classifiers. CER is defined as the percentage of words that are wrongly tagged as follows CER = # of wrongly tagged words Total # of words The baseline CER is determined by assuming the most frequent class for all words. Since the ratio of correct words in both the development and test set is lower than 50%, the most frequent class is “incorrect”. Hence the baseline CER in our experiments is equal to the ratio of correct words as these words are wrongly tagged as incorrect. We also use precision and recall on errors to evaluate the performance of error detection. Let ng be the number of words of which the true class is incorrect, nt be the number of words which are tagged as incorrect by classifiers, and nm be the number of words tagged as incorrect that are indeed translation errors. The precision Pre is the percentage of words correctly tagged as translation errors. Pre = nm nt The recall Rec is the proportion of actual translation errors that are found by classifiers. Rec = nm ng F measure, the trade-off between precision and recall, is also used. F = 2 × Pre × Rec Pre + Rec 6.3 Experimental Results Table 5 shows the performance of our experiments on the error detection task. To compare with previous work using word posterior probabilities for confidence estimation, we carried out experiments using wpp estimated from N-best lists with the classification threshold τ, which was optimized on our development set to minimize CER. A relative improvement of 9.27% is achieved over the baseline CER, which reconfirms the effectiveness of word posterior probabilities for error detection. We conducted three groups of experiments using the MaxEnt based error detection model with various feature combinations. • The first group of experiments uses single feature, such as dwpp, pos. We find the most effective feature is pos, which achieves a 16.12% relative improvement over the baseline CER and 7.55% relative improvement over the CER of word posterior probability thresholding. Using discrete word posterior probabilities as features in the MaxEnt based error detection model is marginally better than word posterior probability thresholding in terms of CER, but obtains a 13.79% relative improvement in F measure. The syntactic feature link also improves the error detection in terms of CER and particularly recall. 609 Combination Features CER (%) Pre (%) Rec (%) F (%) Baseline 47.57 Thresholding wpp 43.16 58.98 58.07 58.52 MaxEnt (dwpp) 44 43.07 56.12 81.86 66.59 MaxEnt (wd) 19,164 41.57 58.25 73.11 64.84 MaxEnt (pos) 199 39.90 58.88 79.23 67.55 MaxEnt (link) 19 44.31 54.72 89.72 67.98 MaxEnt (wd + pos) 19,363 39.43 59.36 78.60 67.64 MaxEnt (wd + pos + link) 19,382 39.79 58.74 80.97 68.08 MaxEnt (dwpp + wd) 19,208 41.04 57.18 83.75 67.96 MaxEnt (dwpp + wd + pos) 19,407 38.88 59.87 78.38 67.88 MaxEnt (dwpp + wd + pos + link) 19,426 38.76 59.89 78.94 68.10 Table 5: Performance of the error detection task. • The second group of experiments concerns with the combination of linguistic features without word posterior probability feature. The combination of lexical features improves both CER and precision over single lexical feature (wd, pos). The addition of syntactic feature link marginally undermines CER but improves recall by a lot. • The last group of experiments concerns about the additional contribution of linguistic features to error detection with word posterior probability. We added linguistic features incrementally into the feature pool. The best performance was achieved by using all features, which has a relative of improvement of 18.52% over the baseline CER. The first two groups of experiments show that linguistic features, individually (except for link) or by combination, are able to produce much better performance than word posterior probability features in both CER and F measure. The best combination of linguistic features achieves a relative improvement of 8.64% and 15.58% in CER and F measure respectively over word posterior probability thresholding. The Table 5 also reveals how linguistic features improve error detection. The lexical features (pos, wd) improve precision when they are used. This suggests that lexical features can help the system find errors more accurately. Syntactic features (link), on the other hand, improve recall whenever they are used, which indicates that they can help the system find more errors. We also show the number of features in each combination in Table 5. Except for the wd feature, 0
200
400
600
800
1000
38.6
38.8
39.0
39.2
39.4
39.6
39.8
40.0
40.2
40.4
40.6
CER (%)
Number of Training Sentences
Figure 3: CER vs. the number of training sentences. the pos has the largest number of features, 199, which is a small set of features. This suggests that our error detection model can be learned from a rather small training set. Figure 3 shows CERs for the feature combination MaxEnt (dwpp + wd + pos + link) when the number of training sentences is enlarged incrementally. CERs drop significantly when the number of training sentences is increased from 100 to 500. After 500 sentences are used, CERs change marginally and tend to converge. 7 Conclusions and Future Work In this paper, we have presented a maximum entropy based approach to automatically detect errors in translation hypotheses generated by SMT 610 systems. We incorporate two sets of linguistic features together with word posterior probability based features into error detection. Our experiments validate that linguistic features are very useful for error detection: 1) they by themselves achieve a higher improvement in terms of both CER and F measure than word posterior probability features; 2) the performance is further improved when they are combined with word posterior probability features. The extracted linguistic features are quite compact, which can be learned from a small training set. Furthermore, The learned linguistic features are system-independent. Therefore our approach can be used for other machine translation systems, such as rule-based or example-based system, which generally do not produce N-best lists. Future work in this direction involve detecting particular error types such as incorrect positions, inappropriate/unnecessary words (Elliott, 2006) and automatically correcting errors. References Yasuhiro Akibay, Eiichiro Sumitay, Hiromi Nakaiway, Seiichi Yamamotoy, and Hiroshi G. Okunoz. 2004. Using a Mixture of N-best Lists from Multiple MT Systems in Rank-sum-based Confidence Measure for MT Outputs. In Proceedings of COLING. Adam L. Berger, Stephen A. Della Pietra andVincent J. Della Pietra. 1996. A Maximum Entropy Approach to Natural Language Processing. Computational Linguistics, 22(1): 39-71. John Blatz, Erin Fitzgerald, George Foster, Simona Gandrabur, Cyril Goutte, Alex Kulesza, Alberto Sanchis, Nicola Ueffing. 2003. Confidence estimation for machine translation. final report, jhu/clsp summer workshop. Debra Elliott. 2006 Corpus-based Machine Translation Evaluation via Automated Error Detection in Output Texts. Phd Thesis, University of Leeds. Simona Gandrabur and George Foster. 2003. Confidence Estimation for Translation Prediction. In Proceedings of HLT-NAACL. S. Jayaraman and A. Lavie. 2005. Multi-engine Machine Translation Guided by Explicit Word Matching. In Proceedings of EAMT. Philipp Koehn, Franz Joseph Och, and Daniel Marcu. 2003. Statistical Phrase-based Translation. In Proceedings of HLT-NAACL. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constrantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of ACL, Demonstration Session. V. I. Levenshtein. 1966. Binary Codes Capable of Correcting Deletions, Insertions and Reversals. Soviet Physics Doklady, Feb. Franz Josef Och. 2003. Minimum Error Rate Training in Statistical Machine Translation. In Proceedings of ACL 2003. Kishore Papineni, Salim Roukos, Todd Ward and WeiJing Zhu. 2002. BLEU: a Method for Automatically Evaluation of Machine Translation. In Proceedings of ACL 2002. Sylvain Raybaud, Caroline Lavecchia, David Langlois, Kamel Sma¨ıli. 2009. Word- and Sentence-level Confidence Measures for Machine Translation. In Proceedings of EAMT 2009. Alberto Sanchis, Alfons Juan and Enrique Vidal. 2007. Estimation of Confidence Measures for Machine Translation. In Procedings of Machine Translation Summit XI. Daniel Sleator and Davy Temperley. 1993. Parsing English with a Link Grammar. In Proceedings of Third International Workshop on Parsing Technologies. Yongmei Shi and Lina Zhou. 2005. Error Detection Using Linguistic Features. In Proceedings of HLT/EMNLP 2005. Andreas Stolcke. 2002. SRILM - an Extensible Language Modeling Toolkit. In Proceedings of International Conference on Spoken Language Processing, volume 2, pages 901-904. Nicola Ueffing, Klaus Macherey, and Hermann Ney. 2003. Confidence Measures for Statistical Machine Translation. In Proceedings. of MT Summit IX. Nicola Ueffing and Hermann Ney. 2007. WordLevel Confidence Estimation for Machine Translation. Computational Linguistics, 33(1):9-40. Richard Zens and Hermann Ney. 2006. N-gram Posterior Probabilities for Statistical Machine Translation. In HLT/NAACL: Proceedings of the Workshop on Statistical Machine Translation. Le Zhang. 2004. Maximum Entropy Modeling Tooklkit for Python and C++. Available at http://homepages.inf.ed.ac.uk/s0450736 /maxent toolkit.html. 611
|
2010
|
62
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 612–621, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics TrustRank: Inducing Trust in Automatic Translations via Ranking Radu Soricut Language Weaver, Inc. 6060 Center Drive, Suite 150 Los Angeles, CA 90045 [email protected] Abdessamad Echihabi Language Weaver, Inc. 6060 Center Drive, Suite 150 Los Angeles, CA 90045 [email protected] Abstract The adoption of Machine Translation technology for commercial applications is hampered by the lack of trust associated with machine-translated output. In this paper, we describe TrustRank, an MT system enhanced with a capability to rank the quality of translation outputs from good to bad. This enables the user to set a quality threshold, granting the user control over the quality of the translations. We quantify the gains we obtain in translation quality, and show that our solution works on a wide variety of domains and language pairs. 1 Introduction The accuracy of machine translation (MT) software has steadily increased over the last 20 years to achieve levels at which large-scale commercial applications of the technology have become feasible. However, widespread adoption of MT technology remains hampered by the lack of trust associated with machine-translated output. This lack of trust is a normal reaction to the erratic translation quality delivered by current state-of-theart MT systems. Unfortunately, the lack of predictable quality discourages the adoption of largescale automatic translation solutions. Consider the case of a commercial enterprise that hosts reviews written by travellers on its web site. These reviews contain useful information about hotels, restaurants, attractions, etc. There is a large and continuous stream of reviews posted on this site, and the large majority is written in English. In addition, there is a large set of potential customers who would prefer to have these reviews available in their (non-English) native languages. As such, this enterprise presents the perfect opportunity for the deployment of a large-volume MT solution. However, travel reviews present specific challenges: the reviews tend to have poor spelling, loose grammar, and broad topics of discussion. The result is unpredictable levels of MT quality. This is undesirable for the commercial enterprise, who is not content to simply reach a broad audience, but also wants to deliver a high-quality product to that audience. We propose the following solution. We develop TrustRank, an MT system enhanced with a capability to rank the quality of translation outputs from good to bad. This enables the user to set a quality threshold, granting the user control over the quality of the translations that it employs in its product. With this enhancement, MT adoption stops being a binary should-we-or-shouldn’twe question. Rather, each user can make a personal trade-off between the scope and the quality of their product. 2 Related Work Work on automatic MT evaluation started with the idea of comparing automatic translations against human-produced references. Such comparisons are done either at lexical level (Papineni et al., 2002; Doddington, 2002), or at linguisticallyricher levels using paraphrases (Zhou et al., 2006; Kauchak and Barzilay, 2006), WordNet (Lavie and Agarwal, 2007), or syntax (Liu and Gildea, 2005; Owczarzak et al., 2007; Yang et al., 2008; Amig´o et al., 2009). In contrast, we are interested in performing MT quality assessments on documents for which reference translations are not available. Reference-free approaches to automatic MT quality assessment, based on Machine Learning techniques such as classification (Kulesza and Shieber, 2004), regression (Albrecht and Hwa, 2007), and ranking (Ye et al., 2007; Duh, 2008), have a different focus compared to ours. Their approach, which uses a test set that is held constant and against which various MT systems are mea612 sured, focuses on evaluating system performance. Similar proposals exist outside the MT field, for instance in syntactic parsing (Ravi et al., 2008). In this case, the authors focus on estimating performance over entire test sets, which in turn is used for evaluating system performance. In contrast, we focus on evaluating the quality of the translations themselves, while the MT system is kept constant. A considerable amount of work has been done in the related area of confidence estimation for MT, for which Blatz et al. (2004) provide a good overview. The goal of this work is to identify small units of translated material (words and phrases) for which one can be confident in the quality of the translation. Related to this goal, and closest to our proposal, is the work of Gamon et al. (2005) and Specia et al. (2009). They describe Machine Learning approaches (classification and regression, respectively) aimed at predicting which sentences are likely to be well/poorly translated. Our work, however, departs from all these works in several important aspects. First, we want to make the quality predictions at document-level, as opposed to sentencelevel (Gamon et al., 2005; Specia et al., 2009), or word/phrase-level (Blatz et al., 2004; Ueffing and Ney, 2005). Document-level granularity is a requirement for large-scale commercial applications that use fully-automated translation solutions. For these applications, the need to make the distinction between “good translation” and “poor translation” must be done at document level. Otherwise, it is not actionable. In contrast, quality-prediction or confidence estimation at sentence- or word-level fits best a scenario in which automated translation is only a part of a larger pipeline. Such pipelines usually involve human post-editing, and are useful for translation productivity (Lagarda et al., 2009). Such solutions, however, suffer from the inherent volume bottleneck associated with human involvement. Our fully-automated solution targets large volume translation needs, on the order of 10,000 documents/day or more. Second, we use automatically generated training labels for the supervised Machine Learning approach. In the experiments presented in this paper, we use BLEU scores (Papineni et al., 2002) as training labels. However, they can be substituted with any of the proposed MT metrics that use human-produced references to automatically assess translation quality (Doddington, 2002; Lavie and Agarwal, 2007). In a similar manner, the work of (Specia et al., 2009) uses NIST scores, and the work of (Ravi et al., 2008) uses PARSEVAL scores. The main advantage of this approach is that we can generate quickly and cheaply as many learning examples as needed. Additionally, we can customize the prediction models on a large variety of genres and domains, and quickly scale to multiple language pairs. In contrast, solutions that require training labels produced manually by humans (Gamon et al., 2005; Albrecht and Hwa, 2007) have difficulties producing prediction models fast enough, trained on enough data, and customized for specific domains. Third, the main metric we use to assess the performance of our solution is targeted directly at measuring translation quality gains. We are interested in the extrinsic evaluation of the quantitative impact of the TrustRank solution, rather than in the intrinsic evaluation of prediction errors (Ravi et al., 2008; Specia et al., 2009). 3 Experimental Framework 3.1 Domains We are interested in measuring the impact of TrustRank on a variety of genres, domains, and language pairs. Therefore, we set up the experimental framework accordingly. We use three proprietary data sets, taken from the domains of Travel (consumer reviews), Consumer Electronics (customer support for computers, data storage, printers, etc.), and HighTech (customer support for high-tech components). All these data sets come in a variety of European and Asian language pairs. We also use the publicly available data set used in the WMT09 task (Koehn and Haddow, 2009) (a combination of European parliament and news data). Information regarding the sizes of these data sets is provided in Table 2. 3.2 Metrics We first present the experimental framework designed to answer the main question we want to address: can we automatically produce a ranking for document translations (for which no humanproduced references are available), such that the translation quality of the documents at the top of this ranking is higher than the average translation quality? To this end, we use several metrics that can gauge how well we answer this question. 613 The first metric is Ranking Accuracy (rAcc), see (Gunawardana and Shani, 2009). We are interested in ranking N documents and assigning them into n quantiles. The formula is: rAcc[n] = Avgn i=1 TPi N n = 1 N × Σn i=1TPi where TPi (True-Positivei) is the number of correctly-assigned documents in quantile i. Intuitively, this formula is an average of the ratio of documents correctly assigned in each quantile. The rAcc metric provides easy to understand lowerbounds and upperbounds. For example, with a method that assigns random ranks, when using 4 quantiles, the accuracy is 25% in any of the quantiles, hence an rAcc of 25%. With an oracle-based ranking, the accuracy is 100% in any of the quantiles, hence an rAcc of 100%. Therefore, the performance of any decent ranking method, when using 4 quantiles, can be expected to fall somewhere between these bounds. The second and main metric is the volumeweighted BLEU gain (vBLEU∆) metric. It measures the average BLEU gain when trading-off volume for accuracy on a predefined scale. The general formula, for n quantiles, is vBLEU∆[n] = Σn−1 i=1 wi × (BLEU1...i −BLEU) with wi = i n Σn−1 j=1 j n = i Σn−1 j=1 j = 2i n(n−1) where BLEU1...i is the BLEU score of the first i quantiles, and BLEU is the score over all the quantiles. Intuitively, this formula provides a volume-weighted average of the BLEU gain obtained while varying the threshold of acceptance from 1 to n-1. (A threshold of acceptance set to the n-th quantile means accepting all the translations and therefore ignore the rankings, so we do not include it in the average.) Without rankings (or with random ranks), the expected vBLEU∆[n] is zero, as the value BLEU1...i is expected to be the same as the overall BLEU for any i. With oracle ranking, the expected vBLEU∆[n] is a positive number representative of the upperbound on the quality of the translations that pass an acceptance threshold. We report the vBLEU∆[n] values as signed numbers, both within a domain and when computed as an average across domains. The choice regarding the number of quantiles is closely related to the choice of setting an acceptance quality threshold. Because we want the solution to stay unchanged while the acceptance quality threshold can vary, we cannot treat this as a classification problem. Instead, we need to provide a complete ranking over an input set of documents. As already mentioned, TrustRank uses a regression method that is trained on BLEU scores as training labels. The regression functions are then used to predict a BLEU-like number for each document in the input set. The rankings are derived trivially from the predicted BLEU numbers, by simply sorting from highest to lowest. Reference ranking is obtained similarly, using actual BLEU scores. Although we are mainly interested in the ranking problem here, it helps to look at the error produced by the regression models to arrive at a more complete picture. Besides the two metrics for ranking described above, we use the well-known regression metrics MAE (mean absolute error) and TE (test-level error): MAE = 1 N × ΣN k=1|predBLEUk −BLEUk| TE = predBLEU −BLEU where BLEUk is the BLEU score for document k, predBLEUk is the predicted BLEU value, and predBLEU is a weighted average of the predicted document-level BLEU numbers over the entire set of N documents. 3.3 Experimental conditions The MT system used by TrustRank (TrustRankMT) is a statistical phrase-based MT system similar to (Och and Ney, 2004). As a reference point regarding the performance of this system, we use the official WMT09 parallel data, monolingual data, and development tuning set (news-dev2009a) to train baseline TrustRank-MT systems for each of the ten WMT09 language pairs. Our system produces translations that are competitive with state-of-the-art systems. We show our baselinesystem BLEU scores on the official development test set (news-dev2009b) for the WMT09 task in Table 1, along with the BLEU scores reported for the baseline Moses system (Koehn and Haddow, 2009). For each of the domains we consider, we partition the data sets as follows. We first set aside 3000 documents, which we call the Regression set 1. The remaining data is called the training MT 1For parallel data for which we do not have document 614 From Eng Fra Spa Ger Cze Hun Moses 17.8 22.4 13.5 11.4 6.5 TrustRank-MT 21.3 22.8 14.3 9.1 8.5 Into Eng Fra Spa Ger Cze Hun Moses 21.2 22.5 16.6 16.9 8.8 TrustRank-MT 22.4 23.8 19.8 13.3 10.4 Table 1: BLEU scores (uncased) for the TrustRank-MT system compared to Moses (WMT09 data). set, on which the MT system is trained. From the Regression set, we set aside 1000 parallel documents to be used as a blind test set (called Regression Test) for our experiments. An additional set of 1000 parallel documents is used as a development set, and the rest of 1000 parallel documents is used as the regression-model training set. We have also performed learning-curve experiments using between 100 and 2000 documents for regression-model training. We do not go into the details of these experiments here for lack of space. The conclusion derived from these experiments is that 1000 documents is the point where the learning-curves level off. In Table 2, we provide a few data points with respect to the data size of these sets (tokenized word-count on the source side). We also report the BLEU performance of the TrustRank-MT system on the Regression Test set. Note that the differences between the BLEU scores reported in Table 1 and the BLEU scores under the WMT09 label in Table 2 reflect differences in the genres of these sets. The official development test set (news-dev2009b) for the WMT09 task is news only. The regression Test sets have the same distribution between Europarl data and news as the corresponding training data set for each language pair. 4 The ranking algorithm As mentioned before, TrustRank takes a supervised Machine Learning approach. We automatically generate the training labels by computing BLEU scores for every document in the Regression training set. boundaries, we simply simulate document boundaries after every 10 consecutive sentences. LP MT set Regression set Train Train Test BLEU WMT09 Eng-Spa 41Mw 277Kw 281Kw 41.0 Eng-Fra 41Mw 282Kw 283Kw 37.1 Eng-Ger 41Mw 282Kw 280Kw 23.7 Eng-Cze 1.2Mw 241Kw 242Kw 10.3 Eng-Hun 30Mw 209Kw 206Kw 14.5 Spa-Eng 42Mw 287Kw 293Kw 40.1 Fra-Eng 44Mw 305Kw 308Kw 37.9 Ger-Eng 39Mw 269Kw 267Kw 29.4 Cze-Eng 1.0Mw 218Kw 219Kw 19.7 Hun-Eng 26Mw 177Kw 176Kw 24.0 Travel Eng-Spa 4.3Mw 123Kw 121Kw 31.2 Eng-Fra 3.5Mw 132Kw 126Kw 27.8 Eng-Ita 3.4Mw 179Kw 183Kw 22.5 Eng-Por 13.1Mw 83Kw 83Kw 41.9 Eng-Ger 7.0Mw 69Kw 69Kw 27.6 Eng-Dut 0.7Mw 89Kw 84Kw 41.9 Electronics Eng-Spa 7.0Mw 150Kw 149Kw 65.2 Eng-Fra 6.5Mw 129Kw 129Kw 55.8 Eng-Ger 5.9Mw 139Kw 140Kw 42.1 Eng-Chi 7.1Mw 135Kw 136Kw 63.9 Eng-Por 2.0Mw 124Kw 115Kw 47.9 HiTech Eng-Spa 2.8Mw 143Kw 148Kw 59.0 Eng-Ger 5.1Mw 162Kw 155Kw 36.6 Eng-Chi 5.6Mw 131Kw 129Kw 60.6 Eng-Rus 2.8Mw 122Kw 117Kw 39.2 Eng-Kor 4.2Mw 129Kw 140Kw 49.4 Table 2: Data sizes and BLEU on Regression Test. 4.1 The learning method The results we report here are obtained using the freely-available Weka engine 2. We have compared and contrasted results using all the regression packages offered by Weka, including regression functions based on simple and multiple-feature Linear regression, Pace regression, RBF networks, Isotonic regression, Gaussian Processes, Support Vector Machines (with SMO optimization) with polynomial and RBF kernels, and regression trees such as REP trees and M5P trees. Due to lack of space and the tangential impact on the message of this paper, we do not report 2Weka software at http://www.cs.waikato.ac.nz/ml/weka/, version 3.6.1, June 2009. 615 these contrastive experiments here. The learning technique that consistently yields the best results is M5P regression trees (weka.classifiers.trees.M5P). Therefore, we report all the results in this paper using this learning method. As an additional advantage, the decision trees and the regression models produced in training are easy to read, understand, and interpret. One can get a good insight into what the impact of a certain feature on a final predicted value is by simply inspecting these trees. 4.2 The features In contrast to most of the work on confidence estimation (Blatz et al., 2004), the features we use are not internal features of the MT system. Therefore, TrustRank can be applied for a large variety of MT approaches, from statistical-based to rulebased approaches. The features we use can be divided into textbased, language-model–based, pseudo-reference– based, example-based, and training-data–based feature types. These feature types can be computed either on the source-side (input documents) or on the target-side (translated documents). Text-based features These features simply look at the length of the input in terms of (tokenized) number of words. They can be applied on the input, where they induce a correlation between the number of words in the input document and the expected BLEU score for that document size. They can also be applied on the produced output, and learn a similar correlation for the produced translation. Language-model–based features These features are among the ones that were first proposed as possible differentiators between good and bad translations (Gamon et al., 2005). They are a measure of how likely a collection of strings is under a language model trained on monolingual data (either on the source or target side). The language-model–based feature values we use here are computed as document-level perplexity numbers using a 5-gram language model trained on the MT training set. Pseudo-reference–based features Previous work has shown that, in the absence of human-produced references, automaticallyproduced ones are still helpful in differentiating between good and bad translations (Albrecht and Hwa, 2008). When computed on the target side, this type of features requires one or more secondary MT systems, used to generate translations starting from the same input. These pseudoreferences are useful in gauging translation convergence, using BLEU scores as feature values. In intuitive terms, their usefulness can be summarized as follows: “if system X produced a translation A and system Y produced a translation B starting from the same input, and A and B are similar, then A is probably a good translation”. An important property here is that systems X and Y need to be as different as possible from each other. This property ensures that a convergence on similar translations is not just an artifact, but a true indication that the translations are correct. The secondary systems we use here are still phrasebased, but equipped with linguistically-oriented modules similar with the ones proposed in (Collins et al., 2005; Xu et al., 2009). The source-side pseudo-reference–based feature type is of a slightly different nature. It still requires one or more secondary MT systems, but operating in the reverse direction. A translated document produced by the main MT system is fed to the secondary MT system(s), translated back into the original source language, and used as pseudoreference(s) when computing a BLEU score for the original input. In intuitive terms: “if system X takes document A and produces B, and system X−1 takes B and produces C, and A and C are similar, then B is probably a good translation”. Example-based features For example-based features, we use a development set of 1000 parallel documents, for which we produce translations and compute document-level BLEU scores. We set aside the top-100 BLEU scoring documents and bottom-100 BLEU scoring documents. They are used as positive examples (with better-than-average BLEU) and negative examples (with worse-than-average BLEU), respectively. We define a positive-example–based feature function as a geometric mean of 1-to-4–gram precision scores (i.e., BLEU score without length penalty) between a document (on either source or target side) and the positive examples used as references (similarly for negative-example–based features). The intuition behind these features can be summarized as follows: “if system X translated docu616 ment A well/poorly, and A and B are similar, then system X probably translates B well/poorly”. Training-data–based features If the main MT system is trained on a parallel corpus, the data in this corpus can be exploited towards assessing translation quality (Specia et al., 2009). In our context, the documents that make up this corpus can be used in a fashion similar with the positive examples. One type of training-data– based features operates by computing the number of out-of-vocabulary (OOV) tokens with respect to the training data (on either source or target side). A more powerful type of training-data–based features operates by computing a BLEU score between a document (source or target side) and the training-data documents used as references. Intuitively, we assess the coverage with respect to the training data and correlate it with a BLEU score: “if the n-grams of input document A are well covered by the source-side of the training data, the translation of A is probably good” (on the source side); “if the n-grams in the output translation B are well covered by the target-side of the parallel training data, then B is probably a good translation” (on the target side). 4.3 Results We are interested in the best performance for TrustRank using the features described above. In this section, we focus on reporting the results obtain for the English-Spanish language pair. In the next section, we report results obtained on all the language pairs we considered. Before we discuss the results of TrustRank, let us anchor the numerical values using some lowerand upper-bounds. As a baseline, we use a regression function that outputs a constant number for each document, equal to the BLEU score of the Regression Training set. As an upperbound, we use an oracle regression function that outputs a number for each document that is equal to the actual BLEU score of that document. In Table 4, we present the performance of these regression functions across all the domains considered. As already mentioned, the rAcc values are bounded by the 25% lowerbound and the 100% upperbound. The vBLEU∆values are bounded by 0 as lowerbound, and some positive BLEU gain value that varies among the domains we considered from +6.4 (Travel) to +13.5 (HiTech). The best performance obtained by TrustRank Domain rAcc vBLEU∆[4] MAE TE Baseline WMT09 25% 0 9.9 +0.4 Travel 25% 0 8.3 +2.0 Electr. 25% 0 12.2 +2.6 HiTech 25% 0 16.9 +2.4 Dom. avg. 25% 0 11.8 1.9 Oracle WMT09 100% +8.2 0 0 Travel 100% +6.4 0 0 Electr. 100% +9.2 0 0 HiTech 100% +13.5 0 0 Dom. avg. 100% +9.3 0 0 Table 4: Lower- and upper-bounds for ranking and regression accuracy (English-Spanish). for English-Spanish, using all the features described, is presented in Table 3. The ranking accuracy numbers on a per-quantile basis reveals an important property for the approach we advocate. The ranking accuracy on the first quantile Q1 (identifying the best 25% of the translations) is 52% on average across the domains. For the last quantile Q4 (identifying the worst 25% of the translations), it is 56%. This is much better than the ranking accuracy for the median-quality translations (35-37% accuracy for the two middle quantiles). This property fits well our scenario, in which we are interested in associating trust in the quality of the translations in the top quantile. The quality of the top quantile translations is quantifiable in terms of BLEU gain. The 250 document translations in Q1 for Travel have a BLEU score of 38.0, a +6.8 BLEU gain compared to the overall BLEU of 31.2 (Q1−4). The Q1 HiTech translations, with a BLEU of 77.9, have a +18.9 BLEU gain compared to the overall BLEU of 59.0. The TrustRank algorithm allows us to tradeoff quantity versus quality on any scale. The results under the BLEU heading in Table 3 represent an instantiation of this ability to a 3-point scale (Q1,Q1−2,Q1−3). The vBLEU∆numbers reflect an average of the BLEU gains for this instantiation (e.g., a +11.6 volume-weighted average BLEU gain for the HiTech domain). We are also interested in the best performance under more restricted conditions, such as time constraints. The assumption we make here is that the translation time dwarfs the time needed for fea617 Domain Ranking Accuracy Translation Accuracy MAE TE BLEU vBLEU∆[4] Q1 Q2 Q3 Q4 rAcc Q1 Q1−2 Q1−3 Q1−4 WMT09 34% 26% 29% 40% 32% 44.8 43.6 42.4 41.1 +2.1 9.6 -0.1 Travel 50% 26% 29% 41% 36% 38.0 35.1 33.0 31.2 +3.4 7.4 -1.9 Electronics 57% 38% 39% 68% 51% 76.1 72.7 69.6 65.2 +6.5 8.4 -2.6 HiTech 65% 48% 49% 75% 59% 77.9 72.7 66.7 59.0 +11.6 8.6 -2.1 Dom. avg. 52% 35% 37% 56% 45% +5.9 8.5 1.7 Table 3: Detailed performance using all features (English-Spanish). ture and regression value computation. Therefore, the most time-expensive feature is the source-side pseudo-reference–based feature, which effectively doubles the translation time required. Under the “time-constrained” condition, we exclude this feature and use all of the remaining features. Table 5 presents the results obtained for English-Spanish. Domain rAcc vBLEU∆[4] MAE TE “Time-constrained” condition WMT09 32% +2.1 9.6 -0.1 Travel 35% +3.2 7.4 -1.8 Electronics 50% +6.3 8.4 -2.2 HiTech 59% +11.6 8.9 -2.1 Dom. avg. 44% +5.8 8.6 1.6 Table 5: “Time-constrained” performance (English-Spanish). The results presented above allow us to draw a series of conclusions. Benefits vary by domain Even with oracle rankings (Table 4), the benefits vary from one domain to the next. For Travel, with an overall BLEU score in the low 30s (31.2), we stand to gain at most +6.4 BLEU points on average (+6.4 vBLEU∆upperbound). For a domain such as HiTech, even with a high overall BLEU score close to 60 (59.0), we stand to gain twice as much (+13.5 vBLEU∆upperbound). Performance varies by domain As the results in Table 3 show, the best performance we obtain also varies from one domain to the next. For instance, the ranking accuracy for the WMT09 domain is only 32%, while for the HiTech domain is 59%. Also, the BLEU gain for the WMT09 domain is only +2.1 vBLEU∆(compared to the upperbound vBLEU∆of +8.2, it is only 26% of the oracle performance). In contrast, the BLEU gain for the HiTech domain is +11.6 vBLEU∆(compared to the +13.5 vBLEU∆upperbound, it is 86% of the oracle performance). Positive feature synergy and overlap The features we described capture different information, and their combination achieves the best performance. For instance, in the Electronics domain, the best single feature is the target-side ngram coverage feature, with +5.3 vBLEU∆. The combination of all features gives a +6.5 vBLEU∆. The numbers in Table 3 also show that eliminating some of the features results in lower performance. The rAcc drops from 45% to 44% in under the “time-constraint” condition (Table 5). The difference in the rankings is statistically significant at p < 0.01 using the Wilcoxon test (Demˇsar, 2006). However, this drop is quantitatively small (1% rAcc drop, -0.1 in vBLEU∆, averaged across domains). This suggests that, even when eliminating features that by themselves have a good discriminatory power (the source-side pseudo-reference– based feature achieves a +5.0 vBLEU∆as a single feature in the Electronics domain), the other features compensate to a large degree. Poor regression performance By looking at the results of the regression metrics, we conclude that the predicted BLEU numbers are not accurate in absolute value. The aggregated Mean Absolute Error (MAE) is 8.5 when using all the features. This is less than the baseline MAE of 11.8, but it is too high to allow us to confidently use the document-level BLEU numbers as reliable indicators of translation accuracy. The Test Error (TE) numbers are not encouraging either, as the 1.7 TE of TrustRank is close to the baseline TE of 1.9 (see Table 4 for baseline numbers). 618 5 Large-scale experimental results In this section, we present the performance of TrustRank on a variety of language pairs (Table 6). We report the BLEU score obtained on our 1000document regression Test, as well as ranking and regression performance using the rAcc, vBLEU∆, MAE, and TE metrics. As the numbers for the ranking and regression metrics show, the same trends we observed for English-Spanish hold for many other language pairs as well. Some domains, such as HiTech, are easier to rank regardless of the language pair, and the quality gains are consistently high (+9.9 average vBLEU∆for the 5 language pairs considered). Other domains, such as WMT09 and Travel, are more difficult to rank. However, the WMT09 English-Hungarian data set appears to be better suited for ranking, as the vBLEU∆numbers are higher compared to the rest of the language pairs from this domain (+4.3 vBLEU∆for Eng-Hun, +7.1 vBLEU∆for Hun-Eng). For Travel, EnglishDutch is also an outlier in terms of quality gains (+12.9 vBLEU∆). Overall, the results indicate that TrustRank obtains consistent performance across a large variety of language pairs. Similar with the conclusion for English-Spanish, the regression performance is currently too poor to allow us to confidently use the absolute document-level predicted BLEU numbers as indicators of translation accuracy. 6 Examples and Illustrations As the experimental results in Table 6 show, the regression performance varies considerably across domains. Even within the same domain, the nature of the material used to perform the experiments can influence considerably the results we obtain. In Figure 1, we plot ⟨BLEU,predBLEU⟩points for three of our language pairs presented in Table 6: Travel Eng-Fra, Travel Eng-Dut, and HiTech EngRus. These plots illustrate the tendency of the predicted BLEU values to correlate with the actual BLEU scores. The amount of correlation visible in these plots matches the performance numbers provided in Table 6, with Travel Eng-Fra at a lower level of correlation compared to Travel Eng-Dut and HiTech Eng-Rus. The ⟨BLEU,predBLEU⟩points tend to align along a line at an angle smaller than 45◦, an indication of the fact that the BLEU predictions tend to be more conservative compared to the actual BLEU scores. For example, in the Domain BLEU rAcc vBLEU∆[4] MAE TE WMT09 Eng-Spa 41.0 35% +2.4 9.2 -0.3 Eng-Fra 37.1 37% +3.3 8.3 -0.5 Eng-Ger 23.7 32% +1.9 5.8 -0.7 Eng-Cze 10.3 38% +1.3 3.1 -0.6 Eng-Hun 14.5 55% +4.3 3.7 -1.1 Spa-Eng 40.1 37% +3.3 8.1 -0.2 Fra-Eng 37.9 39% +3.8 10.1 -0.6 Ger-Eng 29.4 36% +2.7 5.9 -0.9 Cze-Eng 19.7 40% +2.4 4.3 -0.6 Hun-Eng 24.0 61% +7.1 4.9 -1.8 Travel Eng-Spa 31.2 36% +3.4 7.4 -1.9 Eng-Fra 27.8 39% +2.7 6.2 -0.9 Eng-Ita 22.5 39% +2.4 5.1 +0.0 Eng-Por 41.9 51% +5.6 8.6 +1.1 Eng-Ger 27.6 37% +5.7 11.8 -0.4 Eng-Dut 41.9 52% +12.9 12.9 -0.7 Electronics Eng-Spa 65.2 51% +6.5 8.4 -2.6 Eng-Fra 55.8 49% +7.7 8.4 -2.3 Eng-Ger 42.1 57% +8.9 7.4 -1.6 Eng-Chi 63.9 48% +6.4 8.6 -0.8 Eng-Por 47.9 49% +6.9 9.0 -1.8 HiTech Eng-Spa 59.0 59% +11.6 8.6 -2.1 Eng-Ger 36.6 62% +9.2 7.1 -1.0 Eng-Chi 60.3 54% +7.5 8.4 -1.0 Eng-Rus 39.2 62% +10.7 8.7 -2.1 Eng-Kor 49.4 61% +10.5 9.7 -3.2 Table 6: Performance of TrustRank on a variety of domains and language pairs. Travel Eng-Fra case, the predicted BLEU numbers are spread across a narrower band (95% of the values are in the [19-35] interval), compared to the actual BLEU scores (95% of the values are in the [11-47] interval). These intervals are also useful for gauging the level of difficulty stemming from the nature of the material used to perform the experiments. In the case of Travel Eng-Fra, the actual BLEU scores are clustered in a narrower band (interval [11-47] covers 95% of the values), compared to the actual BLEU scores for Travel Eng-Dut (interval [11-92] covers 95% of the values) and HiTech Eng-Rus (interval [3-80] covers 95% of the values). This 619 Figure 1: Examples of BLEU versus predBLEU. means that the documents in the latter cases are easier to distinguish, compared to the documents in Travel Eng-Fra. To provide an intuitive feel for the difference between the level of translation performance between documents ranked close to the bottom and documents ranked close to the top, we present here two example translations. They are documents that we randomly picked from the bottom 10% and top 10% of the Travel Eng-Fra document set, and they correspond to points A and B in the first plot of Figure 1, respectively. The A-Fra and B-Fra entries below are produced by our Eng-Fra TrustRank-MT system, starting from A-Eng and B-Eng3, respectively. A-Eng This will be our 18th year,still love it. Same hotel, room, staff, even other guests from other countries, its lovely to see everyone that you have gotten to know over the years, even if ,you or they ,do not speak each others language. We love the Island some much that, hopefuly, that is where we are retiring to, we do keep looking for that affordable place. A-Fra Ce sera notre 18`eme ann´ee, adore. Mˆeme hˆotel, 3We preserved the original writing style of the documents in the source language. la chambre, le personnel, mˆeme d’autres clients dans d’autres pays, c’est tr`es agr´eable de voir que tout le monde vous aurais savoir au cours de ces derni`eres ann´ees, mˆeme si, ou bien ils vous, ne parlent pas chaque d’autres langues. Nous adorons l’ˆıle des que, hopefuly, c’est l’endroit o`u nous avons retiring, nous ne pour chercher un endroit abordable. B-Eng Stayed at the Intercontinental for 4 nights. It is in an excellent location, not far from the French Quarter. The rooms are large, clean, and comfortable. The staff is friendly and helpful. Parking is very expensive, around $29. 00 a day. There is a garage next door which is a little more reasonable. I certainly suggest this hotel to others. B-Fra J’ai s´ejourn´e `a l’Intercontinental pour 4 nuits. Il est tr`es bien situ´e, pas loin du Quartier Franc¸ais. Les chambres sont grandes, propres et confortables. Le personnel est sympa et serviable. Le parking est tr`es cher, autour de 29 $ par jour. Il y a un garage `a cˆot´e, ce qui est un peu plus raisonnable. Je conseille cet hˆotel `a d’autres. Document A-Fra is a poor translation, and is ranked in the bottom 10%, while document B-Fra is a nearly-perfect translation ranked in the top 10%, out of a total of 1000 documents. 7 Conclusions and Future Work Commercial adoption of MT technology requires trust in the translation quality. Rather than delay this adoption until MT attains a near-human level of sophistication, we propose an interim approach. We present a mechanism that allows MT users to trade quantity for quality, using automaticallydetermined translation quality rankings. The results we present in this paper show that document-level translation quality rankings provide quantitatively strong gains in translation quality, as measured by BLEU. A difference of +18.9 BLEU, like the one we obtain for the EnglishSpanish HiTech domain (Table 3), is persuasive evidence for inspiring trust in the quality of selected translations. This approach enables us to develop TrustRank, a complete MT solution that enhances automatic translation with the ability to identify document subsets containing translations that pass an acceptable quality threshold. When measuring the performance of our solution across several domains, it becomes clear that some domains allow for more accurate quality prediction than others. Given the immediate benefit that can be derived from increasing the ranking accuracy for translation quality, we plan to open up publicly available benchmark data that can be used to stimulate and rigorously monitor progress in this direction. 620 References Joshua Albrecht and Rebecca Hwa. 2007. Regression for sentence-level MT evaluation with pseudo references. In Proceedings of ACL. Joshua Albrecht and Rebecca Hwa. 2008. The role of pseudo references in MT evaluation. In Proceedings of ACL. Enrique Amig´o, Jes´us Gim´enez, Julio Gonzalo, and Felisa Verdejo. 2009. The contribution of linguistic features to automatic machine translation evaluation. In Proceedings of ACL. John Blatz, Erin Fitzgerald, GEorge Foster, Simona Gandrabur, Cyril Gouette, Alex Kulesza, Alberto Sanchis, and Nicola Ueffing. 2004. Confidence estimation for machine translation. In Proceedings of COLING. Michael Collins, Philipp Koehn, and Ivona Kucerova. 2005. Clause restructuring for statistical machine translation. In Proceedings of ACL. J. Demˇsar. 2006. Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research, 7. George Doddington. 2002. Automatic evaluation of machine translation quality using n-gram coocurrence statistics. In Proceedings of HLT. Kevin Duh. 2008. Ranking vs. regression in machine translation evaluation. In Proceedings of the ACL Third Workshop on Statistical Machine Translation. Michael Gamon, Anthony Aue, and Martine Smets. 2005. Sentence-level MT evaluation without reference translations: Beyond language modeling. In Proceedings of EAMT. Asela Gunawardana and Guy Shani. 2009. A survey of accuracy evaluation metrics of recommendation tasks. Journal of Machine Learning Research, 10:2935–2962. David Kauchak and Regina Barzilay. 2006. Paraphrasing for automatic evaluation. In Proceedings of HLT/NAACL. Philipp Koehn and Barry Haddow. 2009. Edinburgh’s submission to all tracks of the WMT2009 shared task with reordering and speed improvements to Moses. In Proceedings of EACL Workshop on Statistical Machine Translation. Alex Kulesza and Stuart M. Shieber. 2004. A learning approach to improving sentence-level MT evaluation. In Proceedings of the 10th International Conference on Theoretical and Methodological Issues in Machine Translation. A.-L. Lagarda, V. Alabau, F. Casacuberta, R. Silva, and E. D´ıaz de Lia˜no. 2009. Statistical post-editing of a rule-based machine translation system. In Proceedings of HLT/NAACL. A. Lavie and A. Agarwal. 2007. METEOR: An autoamtic metric for mt evaluation with high levels of correlation with human judgments. In Proceedings of ACL Workshop on Statistical Machine Translation. Ding Liu and Daniel Gildea. 2005. Syntactic features for evaluation of machine translations. In Proceedings of ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization. Franz Joseph Och and Hermann Ney. 2004. The alignment template approach to statistical machine translation. Computational Linguistics, 30(4):417–449. Karolina Owczarzak, Josef Genabith, and Andy Way. 2007. Evaluating machine translation with LFG dependencies. Machine Translation, 21(2):95–119, June. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In Proceedings of ACL. Sujith Ravi, Kevin Knight, and Radu Soricut. 2008. Automatic prediction of parsing accuracy. In Proceedings of EMNLP. Lucia Specia, Nicola Cancedda, Marc Dymetman, Marcho Turchi, and Nello Cristianini. 2009. Estimating the sentence-level quality of machine translation. In Proceedings of EAMT. Nicola Ueffing and Hermann Ney. 2005. Application of word-level confidence measures in interactive statistical machine translation. In Proceedings of EAMT. Peng Xu, Jaeho Kang, Michael Ringaard, and Franz Och. 2009. Using a dependency parser to improve SMT for Subject-Object-Verb languages. In Proceedings of ACL. Muyun Yang, Shuqi Sun, Jufeng Li, Sheng Li, and Zhao Tiejun. 2008. A linguistically motivated MT evaluation system based on SVM regression. In Proceedings of AMTA. Yang Ye, Ming Zhou, and Chin-Yew Lin. 2007. Sentence level machine translation evaluation as a ranking. In Proceedings of the ACL Second Workshop on Statistical Machine Translation. Liang Zhou, Chin-Yew Lin, and Eduard Hovy. 2006. Re-evaluating machine translation results with paraphrase support. In Proceedings of EMNLP. 621
|
2010
|
63
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 622–630, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Bridging SMT and TM with Translation Recommendation Yifan He Yanjun Ma Josef van Genabith Andy Way Centre for Next Generation Localisation School of Computing Dublin City University {yhe,yma,josef,away}@computing.dcu.ie Abstract We propose a translation recommendation framework to integrate Statistical Machine Translation (SMT) output with Translation Memory (TM) systems. The framework recommends SMT outputs to a TM user when it predicts that SMT outputs are more suitable for post-editing than the hits provided by the TM. We describe an implementation of this framework using an SVM binary classifier. We exploit methods to fine-tune the classifier and investigate a variety of features of different types. We rely on automatic MT evaluation metrics to approximate human judgements in our experiments. Experimental results show that our system can achieve 0.85 precision at 0.89 recall, excluding exact matches. Furthermore, it is possible for the end-user to achieve a desired balance between precision and recall by adjusting confidence levels. 1 Introduction Recent years have witnessed rapid developments in statistical machine translation (SMT), with considerable improvements in translation quality. For certain language pairs and applications, automated translations are now beginning to be considered acceptable, especially in domains where abundant parallel corpora exist. However, these advances are being adopted only slowly and somewhat reluctantly in professional localization and post-editing environments. Post-editors have long relied on translation memories (TMs) as the main technology assisting translation, and are understandably reluctant to give them up. There are several simple reasons for this: 1) TMs are useful; 2) TMs represent considerable effort and investment by a company or (even more so) an individual translator; 3) the fuzzy match score used in TMs offers a good approximation of post-editing effort, which is useful both for translators and translation cost estimation and, 4) current SMT translation confidence estimation measures are not as robust as TM fuzzy match scores and professional translators are thus not ready to replace fuzzy match scores with SMT internal quality measures. There has been some research to address this issue, see e.g. (Specia et al., 2009a) and (Specia et al., 2009b). However, to date most of the research has focused on better confidence measures for MT, e.g. based on training regression models to perform confidence estimation on scores assigned by post-editors (cf. Section 2). In this paper, we try to address the problem from a different perspective. Given that most postediting work is (still) based on TM output, we propose to recommend MT outputs which are better than TM hits to post-editors. In this framework, post-editors still work with the TM while benefiting from (better) SMT outputs; the assets in TMs are not wasted and TM fuzzy match scores can still be used to estimate (the upper bound of) postediting labor. There are three specific goals we need to achieve within this framework. Firstly, the recommendation should have high precision, otherwise it would be confusing for post-editors and may negatively affect the lower bound of the postediting effort. Secondly, although we have full access to the SMT system used in this paper, our method should be able to generalize to cases where SMT is treated as a black-box, which is of622 ten the case in the translation industry. Finally, post-editors should be able to easily adjust the recommendation threshold to particular requirements without having to retrain the model. In our framework, we recast translation recommendation as a binary classification (rather than regression) problem using SVMs, perform RBF kernel parameter optimization, employ posterior probability-based confidence estimation to support user-based tuning for precision and recall, experiment with feature sets involving MT-, TM- and system-independent features, and use automatic MT evaluation metrics to simulate post-editing effort. The rest of the paper is organized as follows: we first briefly introduce related research in Section 2, and review the classification SVMs in Section 3. We formulate the classification model in Section 4 and present experiments in Section 5. In Section 6, we analyze the post-editing effort approximated by the TER metric (Snover et al., 2006). Section 7 concludes the paper and points out avenues for future research. 2 Related Work Previous research relating to this work mainly focuses on predicting the MT quality. The first strand is confidence estimation for MT, initiated by (Ueffing et al., 2003), in which posterior probabilities on the word graph or N-best list are used to estimate the quality of MT outputs. The idea is explored more comprehensively in (Blatz et al., 2004). These estimations are often used to rerank the MT output and to optimize it directly. Extensions of this strand are presented in (Quirk, 2004) and (Ueffing and Ney, 2005). The former experimented with confidence estimation with several different learning algorithms; the latter uses word-level confidence measures to determine whether a particular translation choice should be accepted or rejected in an interactive translation system. The second strand of research focuses on combining TM information with an SMT system, so that the SMT system can produce better target language output when there is an exact or close match in the TM (Simard and Isabelle, 2009). This line of research is shown to help the performance of MT, but is less relevant to our task in this paper. A third strand of research tries to incorporate confidence measures into a post-editing environment. To the best of our knowledge, the first paper in this area is (Specia et al., 2009a). Instead of modeling on translation quality (often measured by automatic evaluation scores), this research uses regression on both the automatic scores and scores assigned by post-editors. The method is improved in (Specia et al., 2009b), which applies Inductive Confidence Machines and a larger set of features to model post-editors’ judgement of the translation quality between ‘good’ and ‘bad’, or among three levels of post-editing effort. Our research is more similar in spirit to the third strand. However, we use outputs and features from the TM explicitly; therefore instead of having to solve a regression problem, we only have to solve a much easier binary prediction problem which can be integrated into TMs in a straightforward manner. Because of this, the precision and recall scores reported in this paper are not directly comparable to those in (Specia et al., 2009b) as the latter are computed on a pure SMT system without a TM in the background. 3 Support Vector Machines for Translation Quality Estimation SVMs (Cortes and Vapnik, 1995) are binary classifiers that classify an input instance based on decision rules which minimize the regularized error function in (1): min w,b,ξ 1 2wT w + C l ∑ i=1 ξi s. t. yi(wT ϕ(xi) + b) ⩾1 −ξi ξi ⩾0 (1) where (xi, yi) ∈Rn × {+1, −1} are l training instances that are mapped by the function ϕ to a higher dimensional space. w is the weight vector, ξ is the relaxation variable and C > 0 is the penalty parameter. Solving SVMs is viable using the ‘kernel trick’: finding a kernel function K in (1) with K(xi, xj) = Φ(xi)T Φ(xj). We perform our experiments with the Radial Basis Function (RBF) kernel, as in (2): K(xi, xj) = exp(−γ||xi −xj||2), γ > 0 (2) When using SVMs with the RBF kernel, we have two free parameters to tune on: the cost parameter C in (1) and the radius parameter γ in (2). In each of our experimental settings, the parameters C and γ are optimized by a brute-force grid 623 search. The classification result of each set of parameters is evaluated by cross validation on the training set. 4 Translation Recommendation as Binary Classification We use an SVM binary classifier to predict the relative quality of the SMT output to make a recommendation. The SVM classifier uses features from the SMT system, the TM and additional linguistic features to estimate whether the SMT output is better than the hit from the TM. 4.1 Problem Formulation As we treat translation recommendation as a binary classification problem, we have a pair of outputs from TM and MT for each sentence. Ideally the classifier will recommend the output that needs less post-editing effort. As large-scale annotated data is not yet available for this task, we use automatic TER scores (Snover et al., 2006) as the measure for the required post-editing effort. In the future, we hope to train our system on HTER (TER with human targeted references) scores (Snover et al., 2006) once the necessary human annotations are in place. In the meantime we use TER, as TER is shown to have high correlation with HTER. We label the training examples as in (3): y = { +1 if TER(MT) < TER(TM) −1 if TER(MT) ≥TER(TM) (3) Each instance is associated with a set of features from both the MT and TM outputs, which are discussed in more detail in Section 4.3. 4.2 Recommendation Confidence Estimation In classical settings involving SVMs, confidence levels are represented as margins of binary predictions. However, these margins provide little insight for our application because the numbers are only meaningful when compared to each other. What is more preferable is a probabilistic confidence score (e.g. 90% confidence) which is better understood by post-editors and translators. We use the techniques proposed by (Platt, 1999) and improved by (Lin et al., 2007) to obtain the posterior probability of a classification, which is used as the confidence score in our system. Platt’s method estimates the posterior probability with a sigmod function, as in (4): Pr(y = 1|x) ≈PA,B(f) ≡ 1 1 + exp(Af + B) (4) where f = f(x) is the decision function of the estimated SVM. A and B are parameters that minimize the cross-entropy error function F on the training data, as in Eq. (5): min z=(A,B) F(z) = − l ∑ i=1 (tilog(pi) + (1 −ti)log(1 −pi)), where pi = PA,B(fi), and ti = { N++1 N++2 if yi = +1 1 N−+2 if yi = −1 (5) where z = (A, B) is a parameter setting, and N+ and N−are the numbers of observed positive and negative examples, respectively, for the label yi. These numbers are obtained using an internal cross-validation on the training set. 4.3 The Feature Set We use three types of features in classification: the MT system features, the TM feature and systemindependent features. 4.3.1 The MT System Features These features include those typically used in SMT, namely the phrase-translation model scores, the language model probability, the distance-based reordering score, the lexicalized reordering model scores, and the word penalty. 4.3.2 The TM Feature The TM feature is the fuzzy match (Sikes, 2007) cost of the TM hit. The calculation of fuzzy match score itself is one of the core technologies in TM systems and varies among different vendors. We compute fuzzy match cost as the minimum Edit Distance (Levenshtein, 1966) between the source and TM entry, normalized by the length of the source as in (6), as most of the current implementations are based on edit distance while allowing some additional flexible matching. hfm(t) = min e EditDistance(s, e) Len(s) (6) where s is the source side of t, the sentence to translate, and e is the source side of an entry in the TM. For fuzzy match scores F, this fuzzy match cost hfm roughly corresponds to 1−F. The difference in calculation does not influence classification, and allows direct comparison between a pure TM system and a translation recommendation system in Section 5.4.2. 624 4.3.3 System-Independent Features We use several features that are independent of the translation system, which are useful when a third-party translation service is used or the MT system is simply treated as a black-box. These features are source and target side LM scores, pseudo source fuzzy match scores and IBM model 1 scores. Source-Side Language Model Score and Perplexity. We compute the language model (LM) score and perplexity of the input source sentence on a LM trained on the source-side training data of the SMT system. The inputs that have lower perplexity or higher LM score are more similar to the dataset on which the SMT system is built. Target-Side Language Model Perplexity. We compute the LM probability and perplexity of the target side as a measure of fluency. Language model perplexity of the MT outputs are calculated, and LM probability is already part of the MT systems scores. LM scores on TM outputs are also computed, though they are not as informative as scores on the MT side, since TM outputs should be grammatically perfect. The Pseudo-Source Fuzzy Match Score. We translate the output back to obtain a pseudo source sentence. We compute the fuzzy match score between the original source sentence and this pseudo-source. If the MT/TM system performs well enough, these two sentences should be the same or very similar. Therefore, the fuzzy match score here gives an estimation of the confidence level of the output. We compute this score for both the MT output and the TM hit. The IBM Model 1 Score. The fuzzy match score does not measure whether the hit could be a correct translation, i.e. it does not take into account the correspondence between the source and target, but rather only the source-side information. For the TM hit, the IBM Model 1 score (Brown et al., 1993) serves as a rough estimation of how good a translation it is on the word level; for the MT output, on the other hand, it is a black-box feature to estimate translation quality when the information from the translation model is not available. We compute bidirectional (source-to-target and target-to-source) model 1 scores on both TM and MT outputs. 5 Experiments 5.1 Experimental Settings Our raw data set is an English–French translation memory with technical translation from Symantec, consisting of 51K sentence pairs. We randomly selected 43K to train an SMT system and translated the English side of the remaining 8K sentence pairs. The average sentence length of the training set is 13.5 words and the size of the training set is comparable to the (larger) TMs used in the industry. Note that we remove the exact matches in the TM from our dataset, because exact matches will be reused and not presented to the post-editor in a typical TM setting. As for the SMT system, we use a standard log-linear PB-SMT model (Och and Ney, 2002): GIZA++ implementation of IBM word alignment model 4,1 the refinement and phraseextraction heuristics described in (Koehn et al., 2003), minimum-error-rate training (Och, 2003), a 5-gram language model with Kneser-Ney smoothing (Kneser and Ney, 1995) trained with SRILM (Stolcke, 2002) on the English side of the training data, and Moses (Koehn et al., 2007) to decode. We train a system in the opposite direction using the same data to produce the pseudosource sentences. We train the SVM classifier using the libSVM (Chang and Lin, 2001) toolkit. The SVMtraining and testing is performed on the remaining 8K sentences with 4-fold cross validation. We also report 95% confidence intervals. The SVM hyper-parameters are tuned using the training data of the first fold in the 4-fold cross validation via a brute force grid search. More specifically, for parameter C in (1) we search in the range [2−5, 215], and for parameter γ (2) we search in the range [2−15, 23]. The step size is 2 on the exponent. 5.2 The Evaluation Metrics We measure the quality of the classification by precision and recall. Let A be the set of recommended MT outputs, and B be the set of MT outputs that have lower TER than TM hits. We standardly define precision P, recall R and F-value as in (7): 1More specifically, we performed 5 iterations of Model 1, 5 iterations of HMM, 3 iterations of Model 3, and 3 iterations of Model 4. 625 P = |A ∩B| |A| , R = |A ∩B| |B| and F = 2PR P + R (7) 5.3 Recommendation Results In Table 1, we report recommendation performance using MT and TM system features (SYS), system features plus system-independent features (ALL:SYS+SI), and system-independent features only (SI). Table 1: Recommendation Results Precision Recall F-Score SYS 82.53±1.17 96.44±0.68 88.95±.56 SI 82.56±1.46 95.83±0.52 88.70±.65 ALL 83.45±1.33 95.56±1.33 89.09±.24 From Table 1, we observe that MT and TM system-internal features are very useful for producing a stable (as indicated by the smaller confidence interval) recommendation system (SYS). Interestingly, only using some simple systemexternal features as described in Section 4.3.3 can also yield a system with reasonably good performance (SI). We expect that the performance can be further boosted by adding more syntactic and semantic features. Combining all the systeminternal and -external features leads to limited gains in Precision and F-score compared to using only system-internal features (SYS) only. This indicates that at the default confidence level, current system-external (resp. system-internal) features can only play a limited role in informing the system when current system-internal (resp. systemexternal) features are available. We show in Section 5.4.2 that combing both system-internal and external features can yield higher, more stable precision when adjusting the confidence levels of the classifier. Additionally, the performance of system SI is promising given the fact that we are using only a limited number of simple features, which demonstrates a good prospect of applying our recommendation system to MT systems where we do not have access to their internal features. 5.4 Further Improving Recommendation Precision Table 1 shows that classification recall is very high, which suggests that precision can still be improved, even though the F-score is not low. Considering that TM is the dominant technology used by post-editors, a recommendation to replace the hit from the TM would require more confidence, i.e. higher precision. Ideally our aim is to obtain a level of 0.9 precision at the cost of some recall, if necessary. We propose two methods to achieve this goal. 5.4.1 Classifier Margins We experiment with different margins on the training data to tune precision and recall in order to obtain a desired balance. In the basic case, the training example would be marked as in (3). If we label both the training and test sets with this rule, the accuracy of the prediction will be maximized. We try to achieve higher precision by enforcing a larger bias towards negative examples in the training set so that some borderline positive instances would actually be labeled as negative, and the classifier would have higher precision in the prediction stage as in (8). y = { +1 if TER(SMT) + b < TER(TM) −1 if TER(SMT) + b ⩾TER(TM) (8) We experiment with b in [0, 0.25] using MT system features and TM features. Results are reported in Table 2. Table 2: Classifier margins Precision Recall TER+0 83.45±1.33 95.56±1.33 TER+0.05 82.41±1.23 94.41±1.01 TER+0.10 84.53±0.98 88.81±0.89 TER+0.15 85.24±0.91 87.08±2.38 TER+0.20 87.59±0.57 75.86±2.70 TER+0.25 89.29±0.93 66.67±2.53 The highest accuracy and F-value is achieved by TER + 0, as all other settings are trained on biased margins. Except for a small drop in TER+0.05, other configurations all obtain higher precision than TER + 0. We note that we can obtain 0.85 precision without a big sacrifice in recall with b=0.15, but for larger improvements on precision, recall will drop more rapidly. When we use b beyond 0.25, the margin becomes less reliable, as the number of positive examples becomes too small. In particular, this causes the SVM parameters we tune on in the first fold to become less applicable to the other folds. This is one limitation of using biased margins to 626 obtain high precision. The method presented in Section 5.4.2 is less influenced by this limitation. 5.4.2 Adjusting Confidence Levels An alternative to using a biased margin is to output a confidence score during prediction and to threshold on the confidence score. It is also possible to add this method to the SVM model trained with a biased margin. We use the SVM confidence estimation techniques in Section 4.2 to obtain the confidence level of the recommendation, and change the confidence threshold for recommendation when necessary. This also allows us to compare directly against a simple baseline inspired by TM users. In a TM environment, some users simply ignore TM hits below a certain fuzzy match score F (usually from 0.7 to 0.8). This fuzzy match score reflects the confidence of recommending the TM hits. To obtain the confidence of recommending an SMT output, our baseline (FM) uses fuzzy match costs hFM ≈1−F (cf. Section 4.3.2) for the TM hits as the level of confidence. In other words, the higher the fuzzy match cost of the TM hit is (lower fuzzy match score), the higher the confidence of recommending the SMT output. We compare this baseline with the three settings in Section 5. 0.7 0.75 0.8 0.85 0.9 0.95 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Precision Confidence SI Sys All FM Figure 1: Precision Changes with Confidence Level Figure 1 shows that the precision curve of FM is low and flat when the fuzzy match costs are low (from 0 to 0.6), indicating that it is unwise to recommend an SMT output when the TM hit has a low fuzzy match cost (corresponding to higher fuzzy match score, from 0.4 to 1). We also observe that the precision of the recommendation receives a boost when the fuzzy match costs for the TM hits are above 0.7 (fuzzy match score lower than 0.3), indicating that SMT output should be recommended when the TM hit has a high fuzzy match cost (low fuzzy match score). With this boost, the precision of the baseline system can reach 0.85, demonstrating that a proper thresholding of fuzzy match scores can be used effectively to discriminate the recommendation of the TM hit from the recommendation of the SMT output. However, using the TM information only does not always find the easiest-to-edit translation. For example, an excellent SMT output should be recommended even if there exists a good TM hit (e.g. fuzzy match score is 0.7 or more). On the other hand, a misleading SMT output should not be recommended if there exists a poor but useful TM match (e.g. fuzzy match score is 0.2). Our system is able to tackle these complications as it incorporates features from the MT and the TM systems simultaneously. Figure 1 shows that both the SYS and the ALL setting consistently outperform FM, indicating that our classification scheme can better integrate the MT output into the TM system than this naive baseline. The SI feature set does not perform well when the confidence level is set above 0.85 (cf. the descending tail of the SI curve in Figure 1). This might indicate that this feature set is not reliable enough to extract the best translations. However, when the requirement on precision is not that high, and the MT-internal features are not available, it would still be desirable to obtain translation recommendations with these black-box features. The difference between SYS and ALL is generally small, but ALL performs steadily better in [0.5, 0,8]. Table 3: Recall at Fixed Precision Recall SYS @85PREC 88.12±1.32 SYS @90PREC 52.73±2.31 SI @85PREC 87.33±1.53 ALL @85PREC 88.57±1.95 ALL @90PREC 51.92±4.28 5.5 Precision Constraints In Table 3 we also present the recall scores at 0.85 and 0.9 precision for SYS, SI and ALL models to demonstrate our system’s performance when there is a hard constraint on precision. Note that our system will return the TM entry when there is an exact match, so the overall precision of the system 627 is above the precision score we set here in a mature TM environment, as a significant portion of the material to be translated will have a complete match in the TM system. In Table 3 for MODEL@K, the recall scores are achieved when the prediction precision is better than K with 0.95 confidence. For each model, precision at 0.85 can be obtained without a very big loss on recall. However, if we want to demand further recommendation precision (more conservative in recommending SMT output), the recall level will begin to drop more quickly. If we use only system-independent features (SI), we cannot achieve as high precision as with other models even if we sacrifice more recall. Based on these results, the users of the TM system can choose between precision and recall according to their own needs. As the threshold does not involve training of the SMT system or the SVM classifier, the user is able to determine this trade-off at runtime. Table 4: Contribution of Features Precision Recall F Score SYS 82.53±1.17 96.44±0.68 88.95±.56 +M1 82.87±1.26 96.23±0.53 89.05±.52 +LM 82.82±1.16 96.20±1.14 89.01±.23 +PS 83.21±1.33 96.61±0.44 89.41±.84 5.6 Contribution of Features In Section 4.3.3 we suggested three sets of system-independent features: features based on the source- and target-side language model (LM), the IBM Model 1 (M1) and the fuzzy match scores on pseudo-source (PS). We compare the contribution of these features in Table 4. In sum, all the three sets of system-independent features improve the precision and F-scores of the MT and TM system features. The improvement is not significant, but improvement on every set of system-independent features gives some credit to the capability of SI features, as does the fact that SI features perform close to SYS features in Table 1. 6 Analysis of Post-Editing Effort A natural question on the integration models is whether the classification reduces the effort of the translators and post-editors: after reading these recommendations, will they translate/edit less than they would otherwise have to? Ideally this question would be answered by human post-editors in a large-scale experimental setting. As we have not yet conducted a manual post-editing experiment, we conduct two sets of analyses, trying to show which type of edits will be required for different recommendation confidence levels. We also present possible methods for human evaluation at the end of this section. 6.1 Edit Statistics We provide the statistics of the number of edits for each sentence with 0.95 confidence intervals, sorted by TER edit types. Statistics of positive instances in classification (i.e. the instances in which MT output is recommended over the TM hit) are given in Table 5. When an MT output is recommended, its TM counterpart will require a larger average number of total edits than the MT output, as we expect. If we drill down, however, we also observe that many of the saved edits come from the Substitution category, which is the most costly operation from the post-editing perspective. In this case, the recommended MT output actually saves more effort for the editors than what is shown by the TER score. It reflects the fact that TM outputs are not actual translations, and might need heavier editing. Table 6 shows the statistics of negative instances in classification (i.e. the instances in which MT output is not recommended over the TM hit). In this case, the MT output requires considerably more edits than the TM hits in terms of all four TER edit types, i.e. insertion, substitution, deletion and shift. This reflects the fact that some high quality TM matches can be very useful as a translation. 6.2 Edit Statistics on Recommendations of Higher Confidence We present the edit statistics of recommendations with higher confidence in Table 7. Comparing Tables 5 and 7, we see that if recommended with higher confidence, the MT output will need substantially less edits than the TM output: e.g. 3.28 fewer substitutions on average. From the characteristics of the high confidence recommendations, we suspect that these mainly comprise harder to translate (i.e. different from the SMT training set/TM database) sentences, as indicated by the slightly increased edit operations 628 Table 5: Edit Statistics when Recommending MT Outputs in Classification, confidence=0.5 Insertion Substitution Deletion Shift MT 0.9849 ± 0.0408 2.2881 ± 0.0672 0.8686 ± 0.0370 1.2500 ± 0.0598 TM 0.7762 ± 0.0408 4.5841 ± 0.1036 3.1567 ± 0.1120 1.2096 ± 0.0554 Table 6: Edit Statistics when NOT Recommending MT Outputs in Classification, confidence=0.5 Insertion Substitution Deletion Shift MT 1.0830 ± 0.1167 2.2885 ± 0.1376 1.0964 ± 0.1137 1.5381 ± 0.1962 TM 0.7554 ± 0.0376 1.5527 ± 0.1584 1.0090 ± 0.1850 0.4731 ± 0.1083 Table 7: Edit Statistics when Recommending MT Outputs in Classification, confidence=0.85 Insertion Substitution Deletion Shift MT 1.1665 ± 0.0615 2.7334 ± 0.0969 1.0277 ± 0.0544 1.5549 ± 0.0899 TM 0.8894 ± 0.0594 6.0085 ± 0.1501 4.1770 ± 0.1719 1.6727 ± 0.0846 on the MT side. TM produces much worse editcandidates for such sentences, as indicated by the numbers in Table 7, since TM does not have the ability to automatically reconstruct an output through the combination of several segments. 6.3 Plan for Human Evaluation Evaluation with human post-editors is crucial to validate and improve translation recommendation. There are two possible avenues to pursue: • Test our system on professional post-editors. By providing them with the TM output, the MT output and the one recommended to edit, we can measure the true accuracy of our recommendation, as well as the post-editing time we save for the post-editors; • Apply the presented method on open domain data and evaluate it using crowdsourcing. It has been shown that crowdsourcing tools, such as the Amazon Mechanical Turk (Callison-Burch, 2009), can help developers to obtain good human judgements on MT output quality both cheaply and quickly. Given that our problem is related to MT quality estimation in nature, it can potentially benefit from such tools as well. 7 Conclusions and Future Work In this paper we present a classification model to integrate SMT into a TM system, in order to facilitate the work of post-editors. Insodoing we handle the problem of MT quality estimation as binary prediction instead of regression. From the posteditors’ perspective, they can continue to work in their familiar TM environment, use the same costestimation methods, and at the same time benefit from the power of state-of-the-art MT. We use SVMs to make these predictions, and use grid search to find better RBF kernel parameters. We explore features from inside the MT system, from the TM, as well as features that make no assumption on the translation model for the binary classification. With these features we make glass-box and black-box predictions. Experiments show that the models can achieve 0.85 precision at a level of 0.89 recall, and even higher precision if we sacrifice more recall. With this guarantee on precision, our method can be used in a TM environment without changing the upper-bound of the related cost estimation. Finally, we analyze the characteristics of the integrated outputs. We present results to show that, if measured by number, type and content of edits in TER, the recommended sentences produced by the classification model would bring about less post-editing effort than the TM outputs. This work can be extended in the following ways. Most importantly, it is useful to test the model in user studies, as proposed in Section 6.3. A user study can serve two purposes: 1) it can validate the effectiveness of the method by measuring the amount of edit effort it saves; and 2) the byproduct of the user study – post-edited sentences – can be used to generate HTER scores to train a better recommendation model. Furthermore, we want to experiment and improve on the adaptability of this method, as the current experiment is on a specific domain and language pair. 629 Acknowledgements This research is supported by the Science Foundation Ireland (Grant 07/CE/I1142) as part of the Centre for Next Generation Localisation (www.cngl.ie) at Dublin City University. We thank Symantec for providing the TM database and the anonymous reviewers for their insightful comments. References John Blatz, Erin Fitzgerald, George Foster, Simona Gandrabur, Cyril Goutte, Alex Kulesza, Alberto Sanchis, and Nicola Ueffing. 2004. Confidence estimation for machine translation. In The 20th International Conference on Computational Linguistics (Coling-2004), pages 315 – 321, Geneva, Switzerland. Peter F. Brown, Vincent J. Della Pietra, Stephen A. Della Pietra, and Robert L. Mercer. 1993. The mathematics of statistical machine translation: parameter estimation. Computational Linguistics, 19(2):263 – 311. Chris Callison-Burch. 2009. Fast, cheap, and creative: Evaluating translation quality using Amazon’s Mechanical Turk. In The 2009 Conference on Empirical Methods in Natural Language Processing (EMNLP-2009), pages 286 – 295, Singapore. Chih-Chung Chang and Chih-Jen Lin, 2001. LIBSVM: a library for support vector machines. Software available at http://www.csie.ntu.edu.tw/ ˜cjlin/libsvm. Corinna Cortes and Vladimir Vapnik. 1995. Support-vector networks. Machine learning, 20(3):273 – 297. R. Kneser and H. Ney. 1995. Improved backing-off for m-gram language modeling. In The 1995 International Conference on Acoustics, Speech, and Signal Processing (ICASSP-95), pages 181 – 184, Detroit, MI. Philipp. Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In The 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology (NAACL/HLT-2003), pages 48 – 54, Edmonton, Alberta, Canada. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In The 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions (ACL-2007), pages 177 – 180, Prague, Czech Republic. Vladimir Iosifovich Levenshtein. 1966. Binary codes capable of correcting deletions, insertions, and reversals. Soviet Physics Doklady, 10(8):707 – 710. Hsuan-Tien Lin, Chih-Jen Lin, and Ruby C. Weng. 2007. A note on platt’s probabilistic outputs for support vector machines. Machine Learning, 68(3):267 – 276. Franz Josef Och and Hermann Ney. 2002. Discriminative training and maximum entropy models for statistical machine translation. In Proceedings of 40th Annual Meeting of the Association for Computational Linguistics (ACL2002), pages 295 – 302, Philadelphia, PA. Franz Josef Och. 2003. Minimum error rate training in statistical machine translation. In The 41st Annual Meeting on Association for Computational Linguistics (ACL2003), pages 160 – 167. John C. Platt. 1999. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in Large Margin Classifiers, pages 61 – 74. Christopher B. Quirk. 2004. Training a sentence-level machine translation confidence measure. In The Fourth International Conference on Language Resources and Evaluation (LREC-2004), pages 825 – 828, Lisbon, Portugal. Richard Sikes. 2007. Fuzzy matching in theory and practice. Multilingual, 18(6):39 – 43. Michel Simard and Pierre Isabelle. 2009. Phrase-based machine translation in a computer-assisted translation environment. In The Twelfth Machine Translation Summit (MT Summit XII), pages 120 – 127, Ottawa, Ontario, Canada. Matthew Snover, Bonnie Dorr, Richard Schwartz, Linnea Micciulla, and John Makhoul. 2006. A study of translation edit rate with targeted human annotation. In The 2006 conference of the Association for Machine Translation in the Americas (AMTA-2006), pages 223 – 231, Cambridge, MA. Lucia Specia, Nicola Cancedda, Marc Dymetman, Marco Turchi, and Nello Cristianini. 2009a. Estimating the sentence-level quality of machine translation systems. In The 13th Annual Conference of the European Association for Machine Translation (EAMT-2009), pages 28 – 35, Barcelona, Spain. Lucia Specia, Craig Saunders, Marco Turchi, Zhuoran Wang, and John Shawe-Taylor. 2009b. Improving the confidence of machine translation quality estimates. In The Twelfth Machine Translation Summit (MT Summit XII), pages 136 – 143, Ottawa, Ontario, Canada. Andreas Stolcke. 2002. SRILM-an extensible language modeling toolkit. In The Seventh International Conference on Spoken Language Processing, volume 2, pages 901 – 904, Denver, CO. Nicola Ueffing and Hermann Ney. 2005. Application of word-level confidence measures in interactive statistical machine translation. In The Ninth Annual Conference of the European Association for Machine Translation (EAMT-2005), pages 262 – 270, Budapest, Hungary. Nicola Ueffing, Klaus Macherey, and Hermann Ney. 2003. Confidence measures for statistical machine translation. In The Ninth Machine Translation Summit (MT Summit IX), pages 394 – 401, New Orleans, LA. 630
|
2010
|
64
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 631–639, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics On Jointly Recognizing and Aligning Bilingual Named Entities Yufeng Chen, Chengqing Zong Institute of Automation, Chinese Academy of Sciences Beijing, China {chenyf,cqzong}@nlpr.ia.ac.cn Keh-Yih Su Behavior Design Corporation Hsinchu, Taiwan, R.O.C. [email protected] Abstract We observe that (1) how a given named entity (NE) is translated (i.e., either semantically or phonetically) depends greatly on its associated entity type, and (2) entities within an aligned pair should share the same type. Also, (3) those initially detected NEs are anchors, whose information should be used to give certainty scores when selecting candidates. From this basis, an integrated model is thus proposed in this paper to jointly identify and align bilingual named entities between Chinese and English. It adopts a new mapping type ratio feature (which is the proportion of NE internal tokens that are semantically translated), enforces an entity type consistency constraint, and utilizes additional monolingual candidate certainty factors (based on those NE anchors). The experiments show that this novel approach has substantially raised the type-sensitive F-score of identified NE-pairs from 68.4% to 81.7% (42.1% F-score imperfection reduction) in our Chinese-English NE alignment task. 1 Introduction In trans-lingual language processing tasks, such as machine translation and cross-lingual information retrieval, named entity (NE) translation is essential. Bilingual NE alignment, which links source NEs and target NEs, is the first step to train the NE translation model. Since NE alignment can only be conducted after its associated NEs have first been identified, the including-rate of the first recognition stage significantly limits the final alignment performance. To alleviate the above error accumulation problem, two strategies have been proposed in the literature. The first strategy (Al-Onaizan and Knight, 2002; Moore, 2003; Feng et al., 2004; Lee et al., 2006) identifies NEs only on the source side and then finds their corresponding NEs on the target side. In this way, it avoids the NE recognition errors which would otherwise be brought into the alignment stage from the target side; however, the NE errors from the source side still remain. To further reduce the errors from the source side, the second strategy (Huang et al., 2003) expands the NE candidate-sets in both languages before conducting the alignment, which is done by treating the original results as anchors, and then re-generating further candidates by enlarging or shrinking those anchors' boundaries. Of course, this strategy will be in vain if the NE anchor is missed in the initial detection stage. In our data-set, this strategy significantly raises the NE-pair type-insensitive including-rate 1 from 83.9% to 96.1%, and is thus adopted in this paper. Although the above expansion strategy has substantially alleviated the error accumulation problem, the final alignment accuracy is still not good (type-sensitive F-score only 68.4%, as indicated in Table 2 in Section 4.2). After having examined the data, we found that: (1) How a given NE is translated, either semantically (called translation) or phonetically (called transliteration), depends greatly on its associated entity type2. The mapping type ratio, which is the percentage of NE internal tokens which are translated semantically, can help with the recognition of the associated NE type; (2) Entities within an aligned pair should share the same type, and this restriction should be integrated into NE alignment as a constraint; (3) Those initially identified monolingual NEs can act as anchors to give monolingual candidate certainty scores 1 Which is the percentage of desired NE-pairs that are included in the expanded set, and is the upper bound on NE alignment performance (regardless of NE types). 2 The proportions of semantic translation, which denote the ratios of semantically translated words among all the associated NE words, for person names (PER), location names (LOC), and organization names (ORG) approximates 0%, 28.6%, and 74.8% respectively in Chinese-English name entity list (2005T34) released by the Linguistic Data Consortium (LDC). Since the title, such as “sir” and “chairman”, is not considered as a part of person names in this corpus, PERs are all transliterated there. 631 (preference weightings) for the re-generated candidates. Based on the above observation, a new joint model which adopts the mapping type ratio, enforces the entity type consistency constraint, and also utilizes the monolingual candidate certainty factors is proposed in this paper to jointly identify and align bilingual NEs under an integrated framework. This framework is decomposed into three subtasks: Initial Detection, Expansion, and Alignment&Re-identification. The Initial Detection subtask first locates the initial NEs and their associated NE types inside both the Chinese and English sides. Afterwards, the Expansion subtask re-generates the candidate-sets in both languages to recover those initial NE recognition errors. Finally, the Alignment&Re-identification subtask jointly recognizes and aligns bilingual NEs via the proposed joint model presented in Section 3. With this new approach, 41.8% imperfection reduction in type-sensitive F-score, from 68.4% to 81.6%, has been observed in our ChineseEnglish NE alignment task. 2 Motivation The problem of NE recognition requires both boundary identification and type classification. However, the complexity of these tasks varies with different languages. For example, Chinese NE boundaries are especially difficult to identify because Chinese is not a tokenized language. In contrast, English NE boundaries are easier to identify due to capitalization clues. On the other hand, classification of English NE types can be more challenging (Ji et al., 2006). Since alignment would force the linked NE pair to share the same semantic meaning, the NE that is more reliably identified in one language can be used to ensure its counterpart in another language. This benefits both the NE boundary identification and type classification processes, and it hints that alignment can help to re-identify those initially recognized NEs which had been less reliable. As shown in the following example, although the desired NE “北韩中央通信社” is recognized partially as “北韩中央” in the initial recognition stage, it would be more preferred if its English counterpart “North Korean's Central News Agency” is given. The reason for this is that “News Agency” would prefer to be linked to “通 信社”, rather than to be deleted (which would happen if “北韩中央” is chosen as the corresponding Chinese NE). (I) The initial NE detection in a Chinese sentence: 官方的 <ORG>北韩中央</ORG> 通信社引述海军... (II) The initial NE detection of its English counterpart: Official <ORG>North Korean's Central News Agency </ORG> quoted the navy's statement… (III) The word alignment between two NEs: (VI) The re-identified Chinese NE boundary after alignment: 官方的 <ORG>北韩中央通信社</ORG> 引述海军声明... As another example, the word “lake” in the English NE is linked to the Chinese character “湖” as illustrated below, and this mapping is found to be a translation and not a transliteration. Since translation rarely occurs for personal names (Chen et al., 2003), the desired NE type “LOC” would be preferred to be shared between the English NE “Lake Constance” and its corresponding Chinese NE “康斯坦茨湖”. As a result, the original incorrect type “PER” of the given English NE is fixed, and the necessity of using mapping type ratio and NE type consistency constraint becomes evident. (I) The initial NE detection result in a Chinese sentence: 在 <LOC>康斯坦茨湖</LOC> 工作的一艘渡船船长… (II) The initial NE detection of its English counterpart: The captain of a ferry boat who works on <PER>Lake Constance </PER>… (III) The word alignment between two NEs: (VI) The re-identified English NE type after alignment: The captain of a ferry boat who works on <LOC>Lake Constance</LOC>… 3 The Proposed Model As mentioned in the introduction section, given a Chinese-English sentence-pair ( , , with its initially recognized Chinese NEs ) CS ES 1 , , S i i i CNE CType S 1 1 [ , ] , T j j j ENE EType T and English NEs ( and 1 ie CTyp j Ety i CNE pe EN are original NE types assigned to and , respectively), we will first re-generate two NE candidate-sets from them by enlarging and shrinking the boundaries of those initially recognized NEs. Let j E 1 C K R and CNE 1 E K RENE C denote these two re-generated candidate sets for Chinese and English NEs respectively ( K and E K are their set-sizes), and min , K S T , then a total K pairs of final Chinese and English NEs will be picked up from the Cartesian product of 632 1 C K RCNE and 1 E K RENE ( , RCNE R [ ] k RENE RType RE i CNE , according to their associated linking score, which is defined as follows. Let denote the associated linking score for a given candidate-pair and , where and are the associated indexes of the re-generated Chinese and English NE candidates, respectively. Furthermore, let be the NE type to be reassigned and shared by RCNE and (as they possess the same meaning). Assume that and are derived from initially recognized and , respectively, and [ ] k re ENE k k [ ] k NE EN Sco k RCNE RCNE ) k k k j E [ ] k RENE[ ] k IC M denotes their internal component mapping, to be defined in Section 3.1, then is defined as follows: [ ] ( , k RENE [ ] , , k k IC k i i RENE M RType NE CType ) k NE Score RC ,max IC k M RType Score RCN P [ ] ( , ) , , , ,[ , ], k k j j E RCNE RENE C CS ENE EType ES | (1) Here, the “max” operator varies over each possible internal component mapping IC M and re-assigned type (PER, LOC, and ORG). For brevity, we will drop those associated subscripts from now on, if there is no confusion. The associated probability factors in the above linking score can be further derived as follows. , , , , , [ , ], , , , , , , IC IC CNE CType CS P M RType ENE EType ES P M RTyp ENE P RCNE CS RType P RENE E ES RType P RType Type EType , , , | , | , | , RCNE RENE e RCNE R CNE CType NE EType CNE ENE C (2) In the above equation, , , e RCNE | , , ENE C | ,CType | , NE EType IC P M RTyp RENE and are the Bilingual Alignment Factor and the Bilingual Type Re-assignment Factor respectively, to represent the bilingual related scores (Section 3.1). Also, and are Monolingual Candidate Certainty Factors (Section 3.2) used to assign preference to each selected and , based on the initially recognized NEs (which act as anchors). , P RType CNE Type EType , , P RCNE CNE CS RType , , P RENE E ES RType RENE RCNE 3.1 Bilingual Related Factors The bilingual alignment factor mainly represents the likelihood value of a specific internal component mapping IC M , given a pair of possible NE configurations RCNE and and their associated . Since Chinese word segmentation is problematic, especially for transliterated words, the bilingual alignment factor RENE RType , , CNE RE IC P M RType R NE in Eq (2) is derived to be conditioned on RE (i.e., starting from the English part). NE We define the internal component mapping IC M to be [ ] 1 [ , , ] , N IC n n n n M cpn ew Mtype [ ] [ , , ] n n n ew Mtype n cpn , where denotes a linked pair consisting of a Chinese component cpn [ ] n ew RCNE (which might contain several Chinese characters) and an English word within and respectively, with their internal mapping type RENE n Mtype TL N 2 [ , n ew to be either translation (abbreviated as TS) or transliteration (abbreviated as TL). In total, there are N component mappings, with translation mappings and transliteration mappings TS N cpn 1 1 [ ] [ , , TS N n n cpn ew TS 2 2 [ ] 1 , ] TL N n n TL 1 1 ]n TS TL N N N , so that . Moreover, since the mapping type distributions of various NE types deviate greatly from one another, as illustrated in the second footnote, the associated mapping type ratio / TS N N is thus an important feature, and is included in the internal component mapping configuration specified above. For example, the IC M between “康斯 坦茨湖” and “Constance Lake” is [康斯坦茨, Constance, TL] and [湖, Lake, TS], so its associated mapping type ratio will be “0.5” (i.e., 1/2). Therefore, the internal mapping is further deduced by introducing the internal mapping type ( | , IC P M RType RENE) n Mtype and the mapping type ratio as follows: [ ] 1 [ ] 1 [ ] ( | , ) ([ , , ] , | , ) ( | , , ) ( | , ) ( | ) IC N n n n n N n n n n n n P M RType RENE P cpn ew Mtype RType RENE P cpn Mtype ew RType P Mtype ew RType P RType (3) In the above equation, the mappings between internal components are trained from the syllable/word alignment of NE pairs of different NE types. In more detail ，for transliteration, the model adopted in (Huang et al., 2003), which first Romanizes Chinese characters and then transliterates them into English characters, is 633 used for . For translation, conditional probability is directly used for . [ ] ( | , , n n n P cpn TL ew RType [ ] ( | , , ) n n n TS ew RType ) P cpn Lastly, the bilingual type re-assignment factor proposed in Eq (2) is derived as follows: | , , , P RType CNE ENE CType EType | , , , | , P RType RCNE RENE CType EType P RType CType EType (4) As Eq (4) shows, both the Chinese initial NE type and English initial NE type are adopted to jointly identify their shared NE type RType . 3.2 Monolingual Candidate Certainty Factors On the other hand, the monolingual candidate certainty factors in Eq (2) indicate the likelihood that a re-generated NE candidate is the true NE given its originally detected NE. For Chinese, it is derived as follows: 1 1 ( , , , ) , , [ ] , , ( , , ) ( , , ) ( | , ) C C C M m m m P RCNE CNE CType CS RType P LeftD RightD Str RCNE Len CType RType P LeftD Len CType RType P RightD Len CType RType P cc cc RType | | | | (5) Where, the subscript C denotes Chinese, and is the length of the originally recognized Chinese NE CN . and denote the left and right distance (which are the numbers of Chinese characters) that R shrinks/enlarges from the left and right boundary of its anchor , respectively. As in the above example, assume that CN and are “北韩中央” and “韩中央通信社” respectively, Le and will be “-1” and “+3”. Also, stands for the associated Chinese string of , denotes the m-th Chinese character within that string, and C Len CNE RightD m cc E E LeftD R RightD CNE CNE ftD Str R R [ ] CNE CNE M denotes the total number of Chinese characters within . RCNE On the English side, following Eq (5), | , , , P RENE ENE EType ES RType ftD E RENE LeftD RightD m cc can be derived similarly, except that Le and will be measured in number of English words. For instance, with EN and as “Lake Constance” and “on Lake Constance” respectively, and will be “+1” and “0”. Also, the bigram unit of the Chinese NE string is replaced by the English word unit . RightD n ew All the bilingual and monolingual factors mentioned above, which are derived from Eq (1), are weighted differently according to their contributions. The corresponding weighting coefficients are obtained using the well-known Minimum Error Rate Training (Och, 2003; commonly abbreviated as MERT) algorithm by minimizing the number of associated errors in the development set. 3.3 Framework for the Proposed Model The above model is implemented with a threestage framework: (A) Initial NE Recognition; (B) NE-Candidate-Set Expansion; and (C) NE Alignment&Re-identification. The Following Diagram gives the details of this framework: For each given bilingual sentence-pair: (A) Initial NE Recognition: generates the initial NE anchors with off-the-self packages. (B) NE-Candidate-Set Expansion: For each initially detected NE, several NE candidates will be re-generated from the original NE by allowing its boundaries to be shrunk or enlarged within a pre-specified range. (B.1) Create both RCNE and RENE candidate-sets, which are expanded from those initial NEs identified in the previous stage. (B.2) Construct an NE-pair candidateset (named NE-Pair-CandidateSet), which is the Cartesian product of the RCNE and RENE candidate-sets created above. (C) NE Alignment&Re-identification: Rank each candidate in the NE-Pair-CandidateSet constructed above with the linking score specified in Eq (1). Afterwards, conduct a beam search process to select the top K non-overlapping NE-pairs from this set. Diagram 1. Steps to Generate the Final NE-Pairs It is our observation that, four Chinese characters for both shrinking and enlarging, two English words for shrinking and three for enlarging are enough in most cases. Under these conditions, the including-rates for NEs with correct boundaries are raised to 95.8% for Chinese and 97.4% for English; and even the NE-pair including rate is raised to 95.3%. Since the above range limitation setting has an including-rate only 0.8% lower than that can be obtained without any range limitation (which is 96.1%), it is adopted in this paper to greatly reduce the number of NEpair-candidates. 634 4 Experiments To evaluate the proposed joint approach, a prior work (Huang et al., 2003) is re-implemented in our environment as the baseline, in which the translation cost, transliteration cost and tagging cost are used. This model is selected for comparison because it not only adopts the same candidate-set expansion strategy as mentioned above, but also utilizes the monolingual information when selecting NE-pairs (however, only a simple bi-gram model is used as the tagging cost in their paper). Note that it enforces the same NE type only when the tagging cost is evaluated: 1 1 1 1 min [ log( ( | , )) log( ( | , ))] RType M tag m m m N n n n C P cc cc RType P ew ew RType . To give a fairer comparison, the same training-set and testing-set are adopted. The trainingset includes two parts. The first part consists of 90,412 aligned sentence-pairs newswire data from the Foreign Broadcast Information Service (FBIS), which is denoted as Training-Set-I. The second Part of the training set is the LDC2005T34 bilingual NE dictionary3, which is denoted as Training-Set-II. The required feature information is then manually labeled throughout the two training sets. In our experiments, for the baseline system, the translation cost and the transliteration cost are trained on Training-Set-II, while the tagging cost is trained on Training-Set-I. For the proposed approach, the monolingual candidate certainty factors are trained on Training-Set-I, and Training-Set-II is used to train the parameters relating to bilingual alignment factors. For the testing-set, 300 sentence pairs are randomly selected from the LDC Chinese-English News Text (LDC2005T06). The average length of the Chinese sentences is 59.4 characters, while the average length of the English sentences is 24.8 words. Afterwards, the answer keys for NE recognition and alignment were annotated manually, and used as the gold standard to calculate metrics of precision (P), recall (R), and F-score (F) for both NE recognition (NER) and NE alignment (NEA). In Total 765 Chinese NEs and 747 English NEs were manually labeled in the testing-set, within which there are only 718 NE pairs, including 214 PER, 371 LOC and 133 ORG NE-pairs. The number of NE pairs is less 3 The LDC2005T34 data-set consists of proofread bilingual entries: 73,352 person names, 76,460 location names and 68,960 organization names. than that of NEs, because not all those recognized NEs can be aligned. Besides, the development-set for MERT weight training is composed of 200 sentence pairs selected from the LDC2005T06 corpus, which includes 482 manually tagged NE pairs. There is no overlap between the training-sets, the development-set and the testing-set. 4.1 Baseline System Both the baseline and the proposed models share the same initial detection subtask, which adopts the Chinese NE recognizer reported by Wu et al. (2005), which is a hybrid statistical model incorporating multi-knowledge sources, and the English NE recognizer included in the publicly available Mallet toolkit4 to generate initial NEs. Initial Chinese NEs and English NEs are recognized by these two available packages respectively. NE-type P (%): C/E R (%): C/E F (%): C/E PER 80.2 / 79.2 87.7 / 85.3 83.8 / 82.1 LOC 89.8 / 85.9 87.3 / 81.5 88.5/ 83.6 ORG 78.6 / 82.9 82.8 / 79.6 80.6 / 81.2 ALL 83.4 / 82.1 86.0 / 82.6 84.7 / 82.3 Table 1. Initial Chinese/English NER Table 1 shows the initial NE recognition performances for both Chinese and English (the largest entry in each column is highlighted for visibility). From Table 1, it is observed that the F-score of ORG type is the lowest among all NE types for both English and Chinese. This is because many organization names are partially recognized or missed. Besides, not shown in the table, the location names or abbreviated organization names tend to be incorrectly recognized as person names. In general, the initial Chinese NER outperforms the initial English NER, as the NE type classification turns out to be a more difficult problem for this English NER system. When those initially identified NEs are directly used for baseline alignment, only 64.1% F score (regard of their name types) is obtained. Such a low performance is mainly due to those NE recognition errors which have been brought into the alignment stage. To diminish the effect of errors accumulating, which stems from the recognition stage, the baseline system also adopts the same expansion strategy described in Section 3.3 to enlarge the possi 4 http://mallet.cs.umass.edu/index.php/Main_Page 635 ble NE candidate set. However, only a slight improvement (68.4% type-sensitive F-score) is obtained, as shown in Table 2. Therefore, it is conjectured that the baseline alignment model is unable to achieve good performance if those features/factors proposed in this paper are not adopted. 4.2 The Recognition and Alignment Joint Model To show the individual effect of each factor in the joint model, a series of experiments, from Exp0 to Exp11, are conducted. Exp0 is the basic system, which ignores monolingual candidate certainty scores, and also disregards mapping type and NE type consistency constraint by ignoring and [ ] ( | , n n P Mtype ew RType) ( | ) P RType , and also replacing P with in Eq (3). [ ] , , n n n ew RType ( | cpn [ ] ( | ) n n P cpn ew ) ) ) ) ) n Mtype To show the effect of enforcing NE type consistency constraint on internal component mapping, Exp1 (named Exp0+RType) replaces in Exp0 with ; On the other hand, Exp2 (named Exp0+MappingType) shows the effect of introducing the component mapping type to Eq (3) by replacing in Exp0 by ; Then Exp3 (named Exp2+MappingTypeRatio) further adds [ ] ( | n n P cpn ew [ ] ( | n n P cpn ew ( | n n P cpn Mtype ( | P RTy , RType P c [ ] ,ew ) pe [ ] ( | n n pn ew ) ( n P Mtype e [ ] | n w to Exp2, to manifest the contribution from the mapping type ratio. In addition, Exp4 (named Exp0+RTypeReassignment) adds the NE type reassignment score, Eq (4), to Exp0 to show the effect of enforcing NE-type consistency. Furthermore, Exp5 (named All-BiFactors) shows the full power of the set of proposed bilingual factors by turning on all the options mentioned above. As the bilingual alignment factors would favor the candidates with shorter lengths, [ ] 1 ([ , , ] , | , ), N n n n n P cpn ew Mtype RType RENE Eq (3), is further normalized into the following form: 1 [ ] 1 [ ] ( | , , ) ( | ), ( | , ) N N n n n n n n P cpn Mtype ew RType P RType P Mtype ew RType and is shown by Exp6 (named All-N-BiFactors). To show the influence of additional information carried by those initially recognized NEs, Exp7 (named Exp6+LeftD/RightD) adds left and right distance information into Exp6, as that specified in Eq (5). To study the monolingual bigram capability, Exp8 (named Exp6+Bigram) adds the NEtype dependant bigram model of each language to Exp6. We use SRI Language Modeling Toolkit5 (SRILM) (Stolcke, 2002) to train various character/word based bi-gram models with different NE types. Similar to what we have done on the bilingual alignment factor above, Exp9 (named Exp6+N-Bigram) adds the normalized NEtype dependant bigram to Exp6 for removing the bias induced by having different NE lengths. The normalized Chinese NEtype dependant bigram score is defined as 1 1 1 [ ( | , ) M ]M m m m P cc cc RType . A Similar transformation is also applied to the English side. Lastly, Exp10 (named Fully-JointModel) shows the full power of the proposed Recognition and Alignment Joint Model by adopting all the normalized factors mentioned above. The result of a MERT weighted version is further shown by Exp11 (named Weighted-JointModel). Model P (%) R (%) F (%) Baseline 77.1 (67.1) 79.7 (69.8) 78.4 (68.4) Exp0 (Basic System) 67.9 (62.4) 70.3 (64.8) 69.1 (63.6) Exp1 (Exp0 + Rtype) 69.6 (65.7) 71.9 (68.0) 70.8 (66.8) Exp2 (Exp0 + MappingType) 70.5 (65.3) 73.0 (67.5) 71.7 (66.4) Exp3 (Exp2 + MappingTypeRatio) 72.0 (68.3) 74.5 (70.8) 73.2 (69.5) Exp4 (Exp0 + RTypeReassignment) 70.2 (66.7) 72.7 (69.2) 71.4 (67.9) Exp5 (All-BiFactors) 76.2 (72.3) 78.5 (74.6) 77.3 (73.4) Exp6 (All-N-BiFactors) 77.7 (73.5) 79.9 (75.7) 78.8 (74.6) Exp7 (Exp6 + LeftD/RightD) 83.5 (77.7) 85.8 (80.1) 84.6 (78.9) Exp8 (Exp6 + Bigram) 80.4 (75.5) 82.7 (77.9) 81.5 (76.7) Exp9 (Exp6 + N-Bigram) 82.7 (77.1) 85.1 (79.6) 83.9 (78.3) Exp10 (Fully-JointModel) 83.7 (78.1) 86.2 (80.7) 84.9 (79.4) Exp11 (Weighted-Joint Model) 85.9 (80.5) 88.4 (83.0) 87.1 (81.7) Table 2. NEA Type-Insensitive (Type-Sensitive) Performance Since most papers in the literature are evaluated only based on the boundaries of NEs, two kinds of performance are thus given here. The first one (named type-insensitive) only checks the scope of each NE without taking its associated NE type into consideration, and is reported 5 http://www.speech.sri.com/projects/srilm/ 636 as the main data at Table 2. The second one (named type-sensitive) would also evaluate the associated NE type of each NE, and is given within parentheses in Table 2. A large degradation is observed when NE type is also taken into account. The highlighted entries are those that are statistically better6 than that of the baseline system. 4.3 ME Approach with Primitive Features Although the proposed model has been derived above in a principled way, since all these proposed features can also be directly integrated with the well-known maximum entropy (ME) (Berger et al., 1996) framework without making any assumptions, one might wonder if it is still worth to deriving a model after all the related features have been proposed. To show that not only the features but also the adopted model contribute to the performance improvement, an ME approach is tested as follows for comparison. It directly adopts all those primitive features mentioned above as its inputs (including internal component mapping, initial and final NE type, NE bigram-based string, and left/right distance), without involving any related probability factors derived within the proposed model. This ME method is implemented with a public package YASMET7, and is tested under various training-set sizes (400, 4,000, 40,000, and 90,412 sentence-pairs). All those training-sets are extracted from the Training-Set-I mentioned above (a total of 298,302 NE pairs included are manually labeled). Since the ME approach is unable to utilize the bilingual NE dictionary (Training-SetII), for fair comparison, this dictionary was also not used to train our models here. Table 3 shows the performance (F-score) using the same testing-set. The data within parentheses are relative improvements. Model 400 4,000 40,000 90,412 ME framework 36.5 (0%) 50.4 (0%) 62.6 (0%) 67.9 (0%) Un-weighted- JointModel +4.6 (+12.6%) +4.5 (+8.9%) +4.3 (+6.9%) +4.1 (+6.0%) Weighted- JointModel +5.0 (+13.7%) +4.7 (+9.3%) +4.6 (+7.3%) +4.5 (+6.6%) Table 3. Comparison between ME Framework and Derived Model on the Testing-Set 6 Statistical significance test is measured on 95% confidence level on 1,000 re-sampling batches (Zhang et al., 2004) 7 http://www.fjoch.com/YASMET.html The improvement indicated in Table 3 clearly illustrates the benefit of deriving the model shown in Eq (2). Since a reasonably derived model not only shares the same training-set with the primitive ME version above, but also enjoys the additional knowledge introduced by the human (i.e., the assumptions/constraints implied by the model), it is not surprising to find out that a good model does help, and that it also becomes more noticeable as the training-set gets smaller. 5 Error Analysis and Discussion Although the proposed model has substantially improved the performance of both NE alignment and recognition, some errors still remain. Having examined those type-insensitive errors, we found that they can be classified into four categories: (A) Original NEs or their components are already not one-to-one mapped (23%). (B) NE components are one-to-one linked, but the associated NE anchors generated from the initial recognition stage are either missing or spurious (24%). Although increasing the number of output candidates generated from the initial recognition stage might cover the missing problem, possible side effects might also be expected (as the complexity of the alignment task would also be increased). (C) Mapping types are not assumed by the model (27%). For example, one NE is abbreviated while its counterpart is not; or some loanwords or out-of-vocabulary terms are translated neither semantically nor phonetically. (D) Wrong NE scopes are selected (26%). Errors of this type are uneasy to resolve, and their possible solutions are beyond the scope of this paper. Examples of above category (C) are interesting and are further illustrated as follows. As an instance of abbreviation errors, a Chinese NE “葛兰素制药厂 (GlaxoSmithKline Factory)” is tagged as “葛兰素/PRR 制药厂/n”, while its counterpart in the English side is simply abbreviated as “GSK” (or replaced by a pronoun “it” sometimes). Linking “葛兰素” to “GSK” (or to the pronoun “it”) is thus out of reach of our model. It seems an abbreviation table (or even anaphora analysis) is required to recover these kind of errors. As an example of errors resulting from loanwords; Japanese kanji “明仁” (the name of a Japanese emperor) is linked to the English word “Akihito”. Here the Japanese kanji “明仁” is directly adopted as the corresponding Chinese characters (as those characters were originally borrowed from Chinese), which would be pro 637 nounced as “Mingren” in Chinese and thus deviates greatly from the English pronunciation of “Akihito”. Therefore, it is translated neither semantically nor phonetically. Further extending the model to cover this new conversion type seems necessary; however, such a kind of extension is very likely to be language pair dependent. 6 Capability of the Proposed Model In addition to improving NE alignment, the proposed joint model can also boost the performance of NE recognition in both languages. The corresponding differences in performance (of the weighted version) when compared with the initial NER ( , and P R F ) are shown in Table 4. Again, those marked entries indicate that they are statistically better than that of the original NER. NEtype P (%): C/E R (%): C/E F (%): C/E PER +5.4 / +6.4 +2.2 / +2.6 +3.9 / +4.6 LOC +4.0 / +3.4 -0.2 / +2.7 +1.8 / +3.0 ORG +7.0 / +3.9 +5.6 / +9.1 +6.2 / +6.4 ALL +5.3 /+5.2 +2.4 / +4.0 +3.9 / +4.6 Table 4. Improvement in Chinese/English NER The result shows that the proposed joint model has a clear win over the initial NER for either Chinese or English NER. In particular, ORG seems to have yielded the greatest gain amongst NE types, which matches our previous observations that the boundaries of Chinese ORG are difficult to identify with the information only coming from the Chinese sentence, while the type of English ORG is uneasy to classify with the information only coming from the English sentence. Though not shown in the tables, it is also observed that the proposed approach achieves a 28.9% reduction on the spurious (false positive) and partial tags over the initial Chinese NER, as well as 16.1% relative error reduction compared with the initial English NER. In addition, total 27.2% wrong Chinese NEs and 40.7% wrong English NEs are corrected into right NE types. However, if the mapping type ratio is omitted, only 21.1% wrong Chinese NE types and 34.8% wrong English NE types can be corrected. This clearly indicates that the ratio is essential for identifying NE types. With the benefits shown above, the alignment model could thus be used to train the monolingual NE recognition model via semi-supervised learning. This advantage is important for updating the NER model from time to time, as various domains frequently have different sets of NEs and new NEs also emerge with time. Since the Chinese NE recognizer we use is not an open source toolkit, it cannot be used to carry out semi-supervised learning. Therefore, only the English NE recognizer and the alignment model are updated during training iterations. In our experiments, 50,412 sentence pairs are first extracted from Training-Set-I as unlabeled data. Various labeled data-sets are then extracted from the remaining data as different seed corpora (100, 400, 4,000 and 40,000 sentence-pairs). Table 5 shows the results of semi-supervised learning after convergence for adopting only the English NER model (NER-Only), the baseline alignment model (NER+Baseline), and our un-weighted joint model (NER+JointModel) respectively. The Initial-NER row indicates the initial performance of the NER model re-trained from different seed corpora. The data within parentheses are relative improvement over Initial-NER. Note that the testing set is still the same as before. As Table 5 shows, with the NER model alone, the performance may even deteriorate after convergence. This is due to the fact that maximizing likelihood does not imply minimizing the error rate. However, with additional mapping constraints from the aligned sentence of another language, the alignment module could guide the searching process to converge to a more desirable point in the parameter space; and these additional constraints become more effective as the seed-corpus gets smaller. Model 100 400 4,000 40,000 Initial-NER 36.7 (0%) 58.6 (0%) 71.4 (0%) 79.1 (0%) NER-Only -2.3 (-6.3%) -0.5 (-0.8%) -0.3 (-0.4%) -0.1 (-0.1%) NER+Baseline +4.9 (+13.4%) +3.4 (5.8%) +1.7 (2.4%) +0.7 (0.9%) NER+Joint Model +10.7 (+29.2%) +8.7 (+14.8%) +4.8 (+6.7%) +2.3 (+2.9%) Table 5. Testing-Set Performance for SemiSupervised Learning of English NE Recognition 7 Conclusion In summary, our experiments show that the new monolingual candidate certainty factors are more effective than the tagging cost (only bigram model) adopted in the baseline system. Moreover, both the mapping type ratio and the entity type consistency constraint are very helpful in identifying the associated NE boundaries and types. After having adopted the features and enforced 638 the constraint mentioned above, the proposed framework, which jointly recognizes and aligns bilingual named entities, achieves a remarkable 42.1% imperfection reduction on type-sensitive F-score (from 68.4% to 81.7%) in our ChineseEnglish NE alignment task. Although the experiments are conducted on the Chinese-English language pair, it is expected that the proposed approach can also be applied to other language pairs, as no language dependent linguistic feature (or knowledge) is adopted in the model/algorithm used. Acknowledgments The research work has been partially supported by the National Natural Science Foundation of China under Grants No. 60975053, 90820303, and 60736014, the National Key Technology R&D Program under Grant No. 2006BAH03B02, and also the Hi-Tech Research and Development Program (“863” Program) of China under Grant No. 2006AA010108-4. References Al-Onaizan, Yaser, and Kevin Knight. 2002. Translating Named Entities Using Monolingual and Bilingual resources. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), pages 400-408. Berger, Adam L., Stephen A. Della Pietra and Vincent J. Della Pietra. 1996. A Maximum Entropy Approach to Natural Language Processing. Computational Linguistics, 22(1):39-72, March. Chen, Hsin-His, Changhua Yang and Ying Lin. 2003. Learning Formulation and Transformation Rules for Multilingual Named Entities. In Proceedings of the ACL 2003 Workshop on Multilingual and Mixed-language Named Entity Recognition, pages 1-8. Feng, Donghui, Yajuan Lv and Ming Zhou. 2004. A New Approach for English-Chinese Named Entity Alignment. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP 2004), pages 372-379. Huang, Fei, Stephan Vogel and Alex Waibel. 2003. Automatic Extraction of Named Entity Translingual Equivalence Based on Multi-Feature Cost Minimization. In Proceedings of ACL’03, Workshop on Multilingual and Mixed-language Named Entity Recognition. Sappora, Japan. Ji, Heng and Ralph Grishman. 2006. Analysis and Repair of Name Tagger Errors. In Proceedings of COLING/ACL 06, Sydney, Australia. Lee, Chun-Jen, Jason S. Chang and Jyh-Shing R. Jang. 2006. Alignment of Bilingual Named Entities in Parallel Corpora Using Statistical Models and Multiple Knowledge Sources. ACM Transactions on Asian Language Information Processing (TALIP), 5(2): 121-145. Moore, R. C.. 2003. Learning Translations of NamedEntity Phrases from Parallel Corpora. In Proceedings of 10th Conference of the European Chapter of ACL, Budapest, Hungary. Och, Franz Josef. 2003. Minimum Error Rate Training in Statistical Machine Translation. In Proceedings of the 41st Annual Conference of the Association for Computational Linguistics (ACL). July 810, 2003. Sapporo, Japan. Pages: 160-167. Stolcke, A. 2002. SRILM -- An Extensible Language Modeling Toolkit. Proc. Intl. Conf. on Spoken Language Processing, vol. 2, pp. 901-904, Denver. Wu, Youzheng, Jun Zhao and Bo Xu. 2005. Chinese Named Entity Recognition Model Based on Multiple Features. In Proceedings of HLT/EMNLP 2005, pages 427-434. Zhang, Ying, Stephan Vogel, and Alex Waibel, 2004. Interpreting BLEU/NIST Scores: How Much Improvement Do We Need to Have a Better System? In Proceedings of the 4th International Conference on Language Resources and Evaluation, pages 2051--2054. 639
|
2010
|
65
|
Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 640–649, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics Generating Templates of Entity Summaries with an Entity-Aspect Model and Pattern Mining Peng Li1 and Jing Jiang2 and Yinglin Wang1 1Department of Computer Science and Engineering, Shanghai Jiao Tong University 2School of Information Systems, Singapore Management University {lipeng,ylwang}@sjtu.edu.cn [email protected] Abstract In this paper, we propose a novel approach to automatic generation of summary templates from given collections of summary articles. This kind of summary templates can be useful in various applications. We first develop an entity-aspect LDA model to simultaneously cluster both sentences and words into aspects. We then apply frequent subtree pattern mining on the dependency parse trees of the clustered and labeled sentences to discover sentence patterns that well represent the aspects. Key features of our method include automatic grouping of semantically related sentence patterns and automatic identification of template slots that need to be filled in. We apply our method on five Wikipedia entity categories and compare our method with two baseline methods. Both quantitative evaluation based on human judgment and qualitative comparison demonstrate the effectiveness and advantages of our method. 1 Introduction In this paper, we study the task of automatically generating templates for entity summaries. An entity summary is a short document that gives the most important facts about an entity. In Wikipedia, for instance, most articles have an introduction section that summarizes the subject entity before the table of contents and other elaborate sections. These introduction sections are examples of entity summaries we consider. Summaries of entities from the same category usually share some common structure. For example, biographies of physicists usually contain facts about the nationality, educational background, affiliation and major contributions of the physicist, whereas introductions of companies usually list information such as the industry, founder and headquarter of the company. Our goal is to automatically construct a summary template that outlines the most salient types of facts for an entity category, given a collection of entity summaries from this category. Such kind of summary templates can be very useful in many applications. First of all, they can uncover the underlying structures of summary articles and help better organize the information units, much in the same way as infoboxes do in Wikipedia. In fact, automatic template generation provides a solution to induction of infobox structures, which are still highly incomplete in Wikipedia (Wu and Weld, 2007). A template can also serve as a starting point for human editors to create new summary articles. Furthermore, with summary templates, we can potentially apply information retrieval and extraction techniques to construct summaries for new entities automatically on the fly, improving the user experience for search engine and question answering systems. Despite its usefulness, the problem has not been well studied. The most relevant work is by Filatova et al. (2006) on automatic creation of domain templates, where the defintion of a domain is similar to our notion of an entity category. Filatova et al. (2006) first identify the important verbs for a domain using corpus statistics, and then find frequent parse tree patterns from sentences containing these verbs to construct a domain template. There are two major limitations of their approach. First, the focus on verbs restricts the template patterns that can be found. Second, redundant or related patterns using different verbs to express the same or similar facts cannot be grouped together. For example, “won X award” and “received X prize” are considered two different patterns by this approach. We propose a method that can overcome these two limitations. Automatic template generation is also related to a number of other problems that have been studied before, in640 cluding unsupervised IE pattern discovery (Sudo et al., 2003; Shinyama and Sekine, 2006; Sekine, 2006; Yan et al., 2009) and automatic generation of Wikipedia articles (Sauper and Barzilay, 2009). We discuss the differences of our work from existing related work in Section 6. In this paper we propose a novel approach to the task of automatically generating entity summary templates. We first develop an entity-aspect model that extends standard LDA to identify clusters of words that can represent different aspects of facts that are salient in a given summary collection (Section 3). For example, the words “received,” “award,” “won” and “Nobel” may be clustered together from biographies of physicists to represent one aspect, even though they may appear in different sentences from different biographies. Simultaneously, the entity-aspect model separates words in each sentence into background words, document words and aspect words, and sentences likely about the same aspect are naturally clustered together. After this aspect identification step, we mine frequent subtree patterns from the dependency parse trees of the clustered sentences (Section 4). Different from previous work, we leverage the word labels assigned by the entity-aspect model to prune the patterns and to locate template slots to be filled in. We evaluate our method on five entity categories using Wikipedia articles (Section 5). Because the task is new and thus there is no standard evaluation criteria, we conduct both quantitative evaluation using our own human judgment and qualitative comparison. Our evaluation shows that our method can obtain better sentence patterns in terms of f1 measure compared with two baseline methods, and it can also achieve reasonably good quality of aspect clusters in terms of purity. Compared with standard LDA and K-means sentence clustering, the aspects identified by our method are also more meaningful. 2 The Task Given a collection of entity summaries from the same entity category, our task is to automatically construct a summary template that outlines the most important information one should include in a summary for this entity category. For example, given a collection of biographies of physicists, ideally the summary template should indicate that important facts about a physicist include his/her edAspect Pattern ENT received his phd from ? university 1 ENT studied ? under ? ENT earned his ? in physics from university of ? ENT was awarded the medal in ? 2 ENT won the ? award ENT received the nobel prize in physics in ? ENT was ? director 3 ENT was the head of ? ENT worked for ? ENT made contributions to ? 4 ENT is best known for work on ? ENT is noted for ? Table 1: Examples of some good template patterns and their aspects generated by our method. ucational background, affiliation, major contributions, awards received, etc. However, it is not clear what is the best representation of such templates. Should a template comprise a list of subtopic labels (e.g. “education” and “affiliation”) or a set of explicit questions? Here we define a template format based on the usage of the templates as well as our observations from Wikipedia entity summaries. First, since we expect that the templates can be used by human editors for creating new summaries, we use sentence patterns that are human readable as basic units of the templates. For example, we may have a sentence pattern “ENT graduated from ? University” for the entity category “physicist,” where ENT is a placeholder for the entity that the summary is about, and ‘?’ is a slot to be filled in. Second, we observe that information about entities of the same category can be grouped into subtopics. For example, the sentences “Bohr is a Nobel laureate” and “Einstein received the Nobel Prize” are paraphrases of the same type of facts, while the sentences “Taub earned his doctorate at Princeton University” and “he graduated from MIT” are slightly different but both describe a person’s educational background. Therefore, it makes sense to group sentence patterns based on the subtopics they pertain to. Here we call these subtopics the aspects of a summary template. Formally, we define a summary template to be a set of sentence patterns grouped into aspects. Each sentence pattern has a placeholder for the entity to be summarized and possibly one or more template slots to be filled in. Table 1 shows some sentence patterns our method has generated for the “physicist” category. 641 2.1 Overview of Our Method Our automatic template generation method consists of two steps: Aspect Identification: In this step, our goal is to automatically identify the different aspects or subtopics of the given summary collection. We simultaneously cluster sentences and words into aspects, using an entity-aspect model extended from the standard LDA model that is widely used in text mining (Blei et al., 2003). The output of this step are sentences clustered into aspects, with each word labeled as a stop word, a background word, a document word or an aspect word. Sentence Pattern Generation: In this step, we generate human-readable sentence patterns to represent each aspect. We use frequent subtree pattern mining to find the most representative sentence structures for each aspect. The fixed structure of a sentence pattern consists of aspect words, background words and stop words, while document words become template slots whose values can vary from summary to summary. 3 Aspect Identification At the aspect identification step, our goal is to discover the most salient aspects or subtopics contained in a summary collection. Here we propose a principled method based on a modified LDA model to simultaneously cluster both sentences and words to discover aspects. We first make the following observation. In entity summaries such as the introduction sections of Wikipedia articles, most sentences are talking about a single fact of the entity. If we look closely, there are a few different kinds of words in these sentences. First of all, there are stop words that occur frequently in any document collection. Second, for a given entity category, some words are generally used in all aspects of the collection. Third, some words are clearly associated with the aspects of the sentences they occur in. And finally, there are also words that are document or entity specific. For example, in Table 2 we show two sentences related to the “affiliation” aspect from the “physicist” summary collection. Stop words such as “is” and “the” are labeled with “S.” The word “physics” can be regarded as a background word for this collection. “Professor” and “university” are clearly related to the “affiliation” aspect. Finally words such as “Modena” and “Chicago” are specifically associated with the subject entities being discussed, that is, they are specific to the summary documents. To capture background words and documentspecific words, Chemudugunta et al. (2007) proposed to introduce a background topic and document-specific topics. Here we borrow their idea and also include a background topic as well as document-specific topics. To discover aspects that are local to one or a few adjacent sentences but may occur in many documents, Titov and McDonald (2008) proposed a multi-grain topic model, which relies on word co-occurrences within short paragraphs rather than documents in order to discover aspects. Inspired by their model, we rely on word co-occurrences within single sentences to identify aspects. 3.1 Entity-Aspect Model We now formally present our entity-aspect model. First, we assume that stop words can be identified using a standard stop word list. We then assume that for a given entity category there are three kinds of unigram language models (i.e. multinomial word distributions). There is a background model φB that generates words commonly used in all documents and all aspects. There are D document models ψd (1 ≤d ≤D), where D is the number of documents in the given summary collection, and there are A aspect models φa (1 ≤a ≤A), where A is the number of aspects. We assume that these word distributions have a uniform Dirichlet prior with parameter β. Since not all aspects are discussed equally frequently, we assume that there is a global aspect distribution θ that controls how often each aspect occurs in the collection. θ is sampled from another Dirichlet prior with parameter α. There is also a multinomial distribution π that controls in each sentence how often we encounter a background word, a document word, or an aspect word. π has a Dirichlet prior with parameter γ. Let Sd denote the number of sentences in document d, Nd,s denote the number of words (after stop word removal) in sentence s of document d, and wd,s,n denote the n’th word in this sentence. We introduce hidden variables zd,s for each sentence to indicate the aspect a sentence belongs to. We also introduce hidden variables yd,s,n for each word to indicate whether a word is generated from the background model, the document model, or the aspect model. Figure 1 shows the process of 642 Venturi/D is/S a/S professor/A of/S physics/B at/S the/S University/A of/S Modena/D ./S He/S was/S a/S professor/A of/S physics/B at/S the/S University/A of/S Chicago/D until/S 1982/D ./S Table 2: Two sentences on “affiliation” from the “physicist” entity category. S: stop word. B: background word. A: aspect word. D: document word. 1. Draw θ ∼Dir(α), φB ∼Dir(β), π ∼Dir(γ) 2. For each aspect a = 1, . . . , A, (a) draw φa ∼Dir(β) 3. For each document d = 1, . . . , D, (a) draw ψd ∼Dir(β) (b) for each sentence s = 1, . . . , Sd i. draw zd,s ∼Multi(θ) ii. for each word n = 1, . . . , Nd,s A. draw yd,s,n ∼Multi(π) B. draw wd,s,n ∼Multi(φB) if yd,s,n = 1, wd,s,n ∼Multi(ψd) if yd,s,n = 2, or wd,s,n ∼Multi(φzd,s) if yd,s,n = 3 Figure 1: The document generation process. y z θ π γ α ϕ φ A d S D s d N , B φ β w Figure 2: The entity-aspect model. generating the whole document collection. The plate notation of the model is shown in Figure 2. Note that the values of α, β and γ are fixed. The number of aspects A is also manually set. 3.2 Inference Given a summary collection, i.e. the set of all wd,s,n, our goal is to find the most likely assignment of zd,s and yd,s,n, that is, the assignment that maximizes p(z, y|w; α, β, γ), where z, y and w represent the set of all z, y and w variables, respectively. With the assignment, sentences are naturally clustered into aspects, and words are labeled as either a background word, a document word, or an aspect word. We approximate p(y, z|w; α, β, γ) by p(y, z|w; ˆφB, { ˆψd}D d=1, {ˆφa}A a=1, ˆθ, ˆπ), where ˆφB, { ˆψd}D d=1, {ˆφa}A a=1, ˆθ and ˆπ are estimated using Gibbs sampling, which is commonly used for inference for LDA models (Griffiths and Steyvers, 2004). Due to space limit, we give the formulas for the Gibbs sampler below without derivation. First, given sentence s in document d, we sample a value for zd,s given the values of all other z and y variables using the following formula: p(zd,s = a|z¬{d,s}, y, w) ∝ CA (a) + α CA (·) + Aα · QV v=1 QE(v) i=0 (Ca (v) + i + β) QE(·) i=0 (Ca (·) + i + V β) . In the formula above, z¬{d,s} is the current aspect assignment of all sentences excluding the current sentence. CA (a) is the number of sentences assigned to aspect a, and CA (·) is the total number of sentences. V is the vocabulary size. Ca (v) is the number of times word v has been assigned to aspect a. Ca (·) is the total number of words assigned to aspect a. All the counts above exclude the current sentence. E(v) is the number of times word v occurs in the current sentence and is assigned to be an aspect word, as indicated by y, and E(·) is the total number of words in the current sentence that are assigned to be an aspect word. We then sample a value for yd,s,n for each word in the current sentence using the following formulas: p(yd,s,n = 1|z, y¬{d,s,n}) ∝ Cπ (1) + γ Cπ (·) + 3γ · CB (wd,s,n) + β CB (·) + V β , p(yd,s,n = 2|z, y¬{d,s,n}) ∝ Cπ (2) + γ Cπ (·) + 3γ · Cd (wd,s,n) + β Cd (·) + V β , p(yd,s,n = 3|z, y¬{d,s,n}) ∝ Cπ (3) + γ Cπ (·) + 3γ · Ca (wd,s,n) + β Ca (·) + V β . In the formulas above, y¬{d,s,n} is the set of all y variables excluding yd,s,n. Cπ (1), Cπ (2) and Cπ (3) are the numbers of words assigned to be a background word, a document word, or an aspect word, respectively, and Cπ (·) is the total number of words. CB and Cd are counters similar to Ca but are for the background model and the document models. In all these counts, the current word is excluded. With one Gibbs sample, we can make the following estimation: 643 ˆφB v = CB (v) + β CB (·) + V β , ˆψd v = Cd (v) + β Cd (·) + V β , ˆφa v = Ca (v) + β Ca (·) + V β , ˆθa = CA (a) + α CA (·) + Aα, ˆπt = Cπ (t) + γ Cπ (·) + 3γ (1 ≤t ≤3). Here the counts include all sentences and all words. In our experiments, we set α = 5, β = 0.01 and γ = 20. We run 100 burn-in iterations through all documents in a collection to stabilize the distribution of z and y before collecting samples. We found that empirically 100 burn-in iterations were sufficient for our data set. We take 10 samples with a gap of 10 iterations between two samples, and average over these 10 samples to get the estimation for the parameters. After estimating ˆφB, { ˆψd}D d=1, {ˆφa}A a=1, ˆθ and ˆπ, we find the values of each zd,s and yd,s,n that maximize p(y, z|w; ˆφB, { ˆψd}D d=1, {ˆφa}A a=1, ˆθ, ˆπ). This assignment, together with the standard stop word list we use, gives us sentences clustered into A aspects, where each word is labeled as either a stop word, a background word, a document word or an aspect word. 3.3 Comparison with Other Models A major difference of our entity-aspect model from standard LDA model is that we assume each sentence belongs to a single aspect while in LDA words in the same sentence can be assigned to different topics. Our one-aspect-per-sentence assumption is important because our goal is to cluster sentences into aspects so that we can mine common sentence patterns for each aspect. To cluster sentences, we could have used a straightforward solution similar to document clustering, where sentences are represented as feature vectors using the vector space model, and a standard clustering algorithm such as K-means can be applied to group sentences together. However, there are some potential problems with directly applying this typical document clustering method. First, unlike documents, sentences are short, and the number of words in a sentence that imply its aspect is even smaller. Besides, we do not know the aspect-related words in advance. As a result, the cosine similarity between two sentences may not reflect whether they are about the same aspect. We can perform heuristic term weighting, but the method becomes less robust. Second, after sentence clustering, we may still want to identify the the aspect words in each sentence, which are useful in the next pattern mining step. Directly taking the most frequent words from each sentence cluster as aspect words may not work well even after stop word removal, because there can be background words commonly used in all aspects. 4 Sentence Pattern Generation At the pattern generation step, we want to identify human-readable sentence patterns that best represent each cluster. Following the basic idea from (Filatova et al., 2006), we start with the parse trees of sentences in each cluster, and apply a frequent subtree pattern mining algorithm to find sentence structures that have occurred at least K times in the cluster. Here we use dependency parse trees. However, different from (Filatova et al., 2006), the word labels (S, B, D and A) assigned by the entity-aspect model give us some advantages. Intuitively, a representative sentence pattern for an aspect should contain at least one aspect word. On the other hand, document words are entity-specific and therefore should not appear in the generic template patterns; instead, they correspond to template slots that need to be filled in. Furthermore, since we work on entity summaries, in each sentence there is usually a word or phrase that refers to the subject entity, and we should have a placeholder for the subject entity in each pattern. Based on the intuitions above, we have the following sentence pattern generation process. 1. Locate subject entities: In each sentence, we want to locate the word or phrase that refers to the subject entity. For example, in a biography, usually a pronoun “he” or “she” is used to refer to the subject person. We use the following heuristic to locate the subject entities: For each summary document, we first find the top 3 frequent base noun phrases that are subjects of sentences. For example, in a company introduction, the phrase “the company” is probably used frequently as a sentence subject. Then for each sentence, we first look for the title of the Wikipedia article. If it occurs, it is tagged as the subject entity. Otherwise, we check whether one of the top 3 subject base noun phrases occurs, and if so, it is tagged as the subject entity. Otherwise, we tag the subject of the sentence as the subject entity. Finally, for the identified subject entity word or phrase, we replace the label assigned by the entity-aspect model with a 644 professor_A is_S ENT a_S physics_B university_A ? the_S nsubj cop det prep_of det prep_at prep_of Figure 3: An example labeled dependency parse tree. new label E. 2. Generate labeled parse trees: We parse each sentence using the Stanford Parser1. After parsing, for each sentence we obtain a dependency parse tree where each node is a single word and each edge is labeled with a dependency relation. Each word is also labeled with one of {E, S, B, D, A}. We replace words labeled with E by a placeholder ENT, and replace words labeled with D by a question mark to indicate that these correspond to template slots. For the other words, we attach their labels to the tree nodes. Figure 3 shows an example labeled dependency parse tree. 3. Mine frequent subtree patterns: For the set of parse trees in each cluster, we use FREQT2, a software that implements the frequent subtree pattern mining algorithm proposed in (Zaki, 2002), to find all subtrees with a minimum support of K. 4. Prune patterns: We remove subtree patterns found by FREQT that do not contain ENT or any aspect word. We also remove small patterns that are contained in some other larger pattern in the same cluster. 5. Covert subtree patterns to sentence patterns: The remaining patterns are still represented as subtrees. To covert them back to human-readable sentence patterns, we map each pattern back to one of the sentences that contain the pattern to order the tree nodes according to their original order in the sentence. In the end, for each summary collection, we obtain A clusters of sentence patterns, where each cluster presumably corresponds to a single aspect or subtopic. 1http://nlp.stanford.edu/software/ lex-parser.shtml 2http://chasen.org/˜taku/software/ freqt/ Category D S Sd min max avg US Actress 407 1721 1 21 4 Physicist 697 4238 1 49 6 US CEO 179 1040 1 24 5 US Company 375 2477 1 36 6 Restaurant 152 1195 1 37 7 Table 3: The number of documents (D), total number of sentences (S) and minimum, maximum and average numbers of sentences per document (Sd) of the data set. 5 Evaluation Because we study a non-standard task, there is no existing annotated data set. We therefore created a small data set and made our own human judgment for quantitative evaluation purpose. 5.1 Data We downloaded five collections of Wikipedia articles from different entity categories. We took only the introduction sections of each article (before the tables of contents) as entity summaries. Some statistics of the data set are given in Table 3. 5.2 Quantitative Evaluation To quantitatively evaluate the summary templates, we want to check (1) whether our sentence patterns are meaningful and can represent the corresponding entity categories well, and (2) whether semantically related sentence patterns are grouped into the same aspect. It is hard to evaluate both together. We therefore separate these two criteria. 5.2.1 Quality of sentence patterns To judge the quality of sentence patterns without looking at aspect clusters, ideally we want to compute the precision and recall of our patterns, that is, the percentage of our sentence patterns that are meaningful, and the percentage of true meaningful sentence patterns of each category that our method can capture. The former is relatively easy to obtain because we can ask humans to judge the quality of our patterns. The latter is much harder to compute because we need human judges to find the set of true sentence patterns for each entity category, which can be very subjective. We adopt the following pooling strategy borrowed from information retrieval. Assume we want to compare a number of methods that each can generate a set of sentence patterns from a summary collection. We take the union of these sets 645 of patterns generated by the different methods and order them randomly. We then ask a human judge to decide whether each sentence pattern is meaningful for the given category. We can then treat the set of meaningful sentence patterns found by the human judge this way as the ground truth, and precision and recall of each method can be computed. If our goal is only to compare the different methods, this pooling strategy should suffice. We compare our method with the following two baseline methods. Baseline 1: In this baseline, we use the same subtree pattern mining algorithm to find sentence patterns from each summary collection. We also locate the subject entities and replace them with ENT. However, we do not have aspect words or document words in this case. Therefore we do not prune any pattern except to merge small patterns with the large ones that contain them. The patterns generated by this method do not have template slots. Baseline 2: In the second baseline, we apply a verb-based pruning on the patterns generated by the first baseline, similar to (Filatova et al., 2006). We first find the top-20 verbs using the scoring function below that is taken from (Filatova et al., 2006), and then prune patterns that do not contain any of the top-20 verbs. s(vi) = N(vi) P vj∈V N(vj) · M(vi) D , where N(vi) is the frequency of verb vi in the collection, V is the set of all verbs, D is the total number of documents in the collection, and M(vi) is the number of documents in the collection that contains vi. In Table 4, we show the precision, recall and f1 of the sentence patterns generated by our method and the two baseline methods for the five categories. For our method, we set the support of the subtree patterns K to 2, that is, each pattern has occurred in at least two sentences in the corresponding aspect cluster. For the two baseline methods, because sentences are not clustered, we use a larger support K of 3; otherwise, we find that there can be too many patterns. We can see that overall our method gives better f1 measures than the two baseline methods for most categories. Our method achieves a good balance between precision and recall. For BL-1, the precision is high but recall is low. Intuitively BL-1 should have a higher recall than our method because our method Category B Purity US Actress 4 0.626 Physicist 6 0.714 US CEO 4 0.674 US Company 4 0.614 Restaurant 3 0.587 Table 5: The true numbers of aspects as judged by the human annotator (B), and the purity of the clusters. does more pattern pruning than BL-1 using aspect words. Here it is not the case mainly because we used a higher frequency threshold (K = 3) to select frequent patterns in BL-1, giving overall fewer patterns than in our method. For BL-2, the precision is higher than BL-1 but recall is lower. It is expected because the patterns of BL-2 is a subset of that of BL-1. There are some advantages of our method that are not reflected in Table 4. First, many of our patterns contain template slots, which make the pattern more meaningful. In contrast the baseline patterns do not contain template slots. Because the human judge did not give preference over patterns with slots, both “ENT won the award” and “ENT won the ? award” were judged to be meaningful without any distinction, although the former one generated by our method is more meaningful. Second, compared with BL-2, our method can obtain patterns that do not contain a non-auxiliary verb, such as “ENT was ? director.” 5.2.2 Quality of aspect clusters We also want to judge the quality of the aspect clusters. To do so, we ask the human judge to group the ground truth sentence patterns of each category based on semantic relatedness. We then compute the purity of the automatically generated clusters against the human judged clusters using purity. The results are shown in Table 5. In our experiments, we set the number of clusters A used in the entity-aspect model to be 10. We can see from Table 5 that our generated aspect clusters can achieve reasonably good performance. 5.3 Qualitative evaluation We also conducted qualitative comparison between our entity-aspect model and standard LDA model as well as a K-means sentence clustering method. In Table 6, we show the top 5 frequent words of three sample aspects as found by our method, standard LDA, and K-means. Note that although we try to align the aspects, there is 646 Category Method US Actress Physicist US CEO US Company Restaurant BL-1 precision 0.714 0.695 0.778 0.622 0.706 recall 0.545 0.300 0.367 0.425 0.361 f1 0.618 0.419 0.499 0.505 0.478 BL-2 precision 0.845 0.767 0.829 0.809 1.000 recall 0.260 0.096 0.127 0.167 0.188 f1 0.397 0.17 0.220 0.276 0.316 Ours precision 0.544 0.607 0.586 0.450 0.560 recall 0.710 0.785 0.712 0.618 0.701 f1 0.616 0.684 0.643 0.520 0.624 Table 4: Quality of sentence patterns in terms of precision, recall and f1. Method Sample Aspects 1 2 3 Our university prize academy entityreceived nobel sciences aspect ph.d. physics member model college awarded national degree medal society Standard physics nobel physics LDA american prize institute professor physicist research received awarded member university john sciences K-means physics physicist physics university american academy institute physics sciences work university university research nobel new Table 6: Comparison of the top 5 words of three sample aspects using different methods. no correspondence between clusters numbered the same but generated by different methods. We can see that our method gives very meaningful aspect clusters. Standard LDA also gives meaningful words, but background words such as “physics” and “physicist” are mixed with aspect words. Entity-specific words such as “john” also appear mixed with aspect words. K-means clusters are much less meaningful, with too many background words mixed with aspect words. 6 Related Work The most related existing work is on domain template generation by Filatova et al. (2006). There are several differences between our work and theirs. First, their template patterns must contain a non-auxiliary verb whereas ours do not have this restriction. Second, their verb-centered patterns are independent of each other, whereas we group semantically related patterns into aspects, giving more meaningful templates. Third, in their work, named entities, numbers and general nouns are treated as template slots. In our method, we apply the entity-aspect model to automatically identify words that are document-specific, and treat these words as template slots, which can be potentially more robust as we do not rely on the quality of named entity recognition. Last but not least, their documents are event-centered while ours are entity-centered. Therefore we can use heuristics to anchor our patterns on the subject entities. Sauper and Barzilay (2009) proposed a framework to learn to automatically generate Wikipedia articles. There is a fundamental difference between their task and ours. The articles they generate are long, comprehensive documents consisting of several sections on different subtopics of the subject entity, and they focus on learning the topical structures from complete Wikipedia articles. We focus on learning sentence patterns of the short, concise introduction sections of Wikipedia articles. Our entity-aspect model is related to a number of previous extensions of LDA models. Chemudugunta et al. (2007) proposed to introduce a background topic and document-specific topics. Our background and document language models are similar to theirs. However, they still treat documents as bags of words rather than sets of sentences as in our model. Titov and McDonald (2008) exploited the idea that a short paragraph within a document is likely to be about the same aspect. Our one-aspect-per-sentence assumption is a stricter than theirs, but it is required in our model for the purpose of mining sentence patterns. The way we separate words into stop words, background words, document words and aspect words bears similarity to that used in (Daum´e III and Marcu, 2006; Haghighi and Vanderwende, 2009), but their task is multi-document summarization while ours is to induce summary templates. 647 7 Conclusions and Future Work In this paper, we studied the task of automatically generating templates for entity summaries. We proposed an entity-aspect model that can automatically cluster sentences and words into aspects. The model also labels words in sentences as either a stop word, a background word, a document word or an aspect word. We then applied frequent subtree pattern mining to generate sentence patterns that can represent the aspects. We took advantage of the labels generated by the entity-aspect model to prune patterns and to locate template slots. We conducted both quantitative and qualitative evaluation using five collections of Wikipedia entity summaries. We found that our method gave overall better template patterns than two baseline methods, and the aspect clusters generated by our method are reasonably good. There are a number of directions we plan to pursue in the future in order to improve our method. First, we can possibly apply linguistic knowledge to improve the quality of sentence patterns. Currently the method may generate similar sentence patterns that differ only slightly, e.g. change of a preposition. Also, the sentence patterns may not form complete, meaningful sentences. For example, a sentence pattern may contain an adjective but not the noun it modifies. We plan to study how to use linguistic knowledge to guide the construction of sentence patterns and make them more meaningful. Second, we have not quantitatively evaluated the quality of the template slots, because our judgment is only at the whole sentence pattern level. We plan to get more human judges and more rigorously judge the relevance and usefulness of both the sentence patterns and the template slots. It is also possible to introduce certain rules or constraints to selectively form template slots rather than treating all words labeled with D as template slots. Acknowledgments This work was done during Peng Li’s visit to the Singapore Management University. This work was partially supported by the National High-tech Research and Development Project of China (863) under the grant number 2009AA04Z106 and the National Science Foundation of China (NSFC) under the grant number 60773088. We thank the anonymous reviewers for their helpful comments. References David Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent dirichlet allocation. Journal of Machine Learning Research, 3:993–1022. Chaitanya Chemudugunta, Padhraic Smyth, and Mark Steyvers. 2007. Modeling general and specific aspects of documents with a probabilistic topic model. In Advances in Neural Information Processing Systems 19, pages 241–248. Hal Daum´e III and Daniel Marcu. 2006. Bayesian query-focused summarization. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 305–312. Elena Filatova, Vasileios Hatzivassiloglou, and Kathleen McKeown. 2006. Automatic creation of domain templates. In Proceedings of 21st International Conference on Computational Linguistics and the 44th Annual Meeting of the Association for Computational Linguistics, pages 207–214. Thomas L. Griffiths and Mark Steyvers. 2004. Finding scientific topics. Proceedings of the National Academy of Sciences of the United States of America, 101(Suppl. 1):5228–5235. Aria Haghighi and Lucy Vanderwende. 2009. Exploring content models for multi-document summarization. In Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics, pages 362–370. Christina Sauper and Regina Barzilay. 2009. Automatically generating Wikipedia articles: A structureaware approach. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 208–216. Satoshi Sekine. 2006. On-demand information extraction. In Proceedings of 21st International Conference on Computational Linguistics and the 44th Annual Meeting of the Association for Computational Linguistics, pages 731–738. Yusuke Shinyama and Satoshi Sekine. 2006. Preemptive information extraction using unrestricted relation discovery. In Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics, pages 304–311. Kiyoshi Sudo, Satoshi Sekine, and Ralph Grishman. 2003. An improved extraction pattern representation model for automatic IE pattern acquisition. In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, pages 224– 231. Ivan Titov and Ryan McDonald. 2008. Modeling online reviews with multi-grain topic models. In 648 Proceeding of the 17th International Conference on World Wide Web, pages 111–120. Fei Wu and Daniel S. Weld. 2007. Autonomously semantifying Wikipedia. In Proceedings of the 16th ACM Conference on Information and Knowledge Management, pages 41–50. Yulan Yan, Naoaki Okazaki, Yutaka Matsuo, Zhenglu Yang, and Mitsuru Ishizuka. 2009. Unsupervised relation extraction by mining Wikipedia texts using information from the Web. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 1021–1029. Mohammed J. Zaki. 2002. Efficiently mining frequent trees in a forest. In Proceedings of the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 71–80. 649
|
2010
|
66
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.