tasksource/scinli · Datasets at Hugging Face

id stringlengths 7 12	sentence1 stringlengths 6 1.27k	sentence2 stringlengths 6 926	label stringclasses 4 values
train_20000	Naturally, these are not appropriate comparisons for the work reported here.	as is evident from the discussion above, previous spelling research does provide an important role in suggesting productive features to include in the decision tree.	contrasting
train_20001	Capitalization is his sole means of identifying names.	capitalization information is not available in closed captions.	contrasting
train_20002	More importantly, we investigate the relative role of each statistical and linguistic knowledge source in the proposed IR/NLP question-answering system.	to previous results, we find that statistical knowledge of word co-occurrences as computed by vector space models of IR can be used to quickly and accurately locate relevant documents in the restricted QA task.	contrasting
train_20003	In the cases we have investigated, this strategy works often enough to be useful for the surface object.	due to predicate coordination as well as relativization, such a strategy often fails to identify correctly the surface subject of bind (binds or bound) when more than one binding term precedes the verb.	contrasting
train_20004	Basically, the segmenter is based on frequency of individual nouns extracted from corpus.	the problem is that it is difficult to distinguish proper noun and common noun since there is no clue like capital letters in Korean.	contrasting
train_20005	Assume that 'sl' is a frequently appearing word in texts whereas 's2s3s4' is a rarely occurring sequence of syllables as a word.	'sis2' and 's3s4' occurs frequently but although they don't occur as frequently as 'sl'.	contrasting
train_20006	Actually, gyosaeng-hwal is not a word.	both hag-gyo and saeng-hwal are frequently occurring syllables, and actually they are all words.	contrasting
train_20007	It is important to note here that we would probably have obtained this improvement in recognition accuracy even with a manual revision of the grammars.	the main advantage in using our tool is the tremendous simplification of the whole process of revision for a grammar developer who now selects counter-examples with an interactive tool instead of manually revising the grammars.	contrasting
train_20008	We have successfully used the tool to constrain overgeneralizing grammars of speech understanding systems and obtained 20-30% higher recognition accuracy.	we believe the primary benefit of using our tool is the tremendously reduced effort for the grammar developer.	contrasting
train_20009	This is a language dependent approach.	we propose a framework of language independent morphological analysis system.	contrasting
train_20010	Disambiguation i s already language independent, since it does not process strings directly and therefore will not be taken up.	tokenization and dictionary look-up are language dependent and shall be explained more in this paper.	contrasting
train_20011	If a hyphenated segment such as "data-base," "F-16," or "MS-DOS" exists in the dictionary, it should be an independent lexeme.	if a hyphenated segment such as "55-years-old" does not exist in the dictionary, hyphens should be treated as independent tokens (Fox, 1992).	contrasting
train_20012	Then we pass it to the analyzer for non-segmented languages.	the analyzer may return the result as "They / 've / gone / to / school / to / get / her / ."	contrasting
train_20013	The set of the delimiters acting as boundaries: These act as boundaries of MFs.	these can not be independent MFs (can not start nor end a lexeme).	contrasting
train_20014	These problems are discussed by several authors including Mills (Mills, 1998) Webster & Kit (Webster and Kit, 1992).	their proposed solutions are not language independent.	contrasting
train_20015	Using case-sensitive rules is straightforward since generally only nouns (and proper names) are written in standard German with a capitalized initial letter (e.g., "das Unternehmen" -the enterprise vs. "wir unternehmen" -we undertake).	for disambiguation of word forms appearing at the beginning of the sentence local contextual filtering rules are applied.	contrasting
train_20016	company names) may appear in the text either with or without a designator, we use a dynamic lexicon to store recognized named entities without their designators (e.g., "Braun AG" vs. "Braun") in order to identify subsequent occurrences correctly.	a named entity, consisting solely of one word, may be also a valid word form (e.g., "Braun" -brown).	contrasting
train_20017	This means that although the exact attachment point of each individual PP is not known it is guaranteed that a PP can only be attached to phrases which are dominated by the main verb of the sentence (which is the root node of the clause's tree).	the exact point of attachment is a matter of domain-specific knowledge and hence should be defined as part of the domain knowledge of an application.	contrasting
train_20018	Only in the case of the MC-module the difference is about 5%.	the result for the isolated evaluation of the MC-module suggests that this is mainly due to errors caused by previous components.	contrasting
train_20019	The output document from constrained HMM contains MUC-standard NE.tags such as person, location and organization.	for a real information extraction system, the MUC-standard NE tag may not be enough and further detailed NE information might be necessary.	contrasting
train_20020	Further adding the morphological analysis capability that automatically analyzes unknown words (deriving additional morphological relationships and some semantic subsumption relationships) significantly improves that result (from 50.0% to 60.7%).	we found that adding the same semantic subsumption relationships to the commercial search engine, using its provided thesaurus capability degraded its results, and results were still degraded when we added only those facts that we knew would help find relevant documents.	contrasting
train_20021	Knowledge classifiers are part of almost any knowledge representation system.	the problem we face here is more difficult.	contrasting
train_20022	By default this concept is placed under policy.	in WordNet there is a hierarchy fiscal policy -IS-Aeconomic policy -IS-A -policy.	contrasting
train_20023	Higher interest rates are normally associated with weaker bond markets.	if interest rates go down, bonds go up, and your bond becomes more valuable.	contrasting
train_20024	Overgeneration will produce some name variants that have nothing to do with the company.	although overgeneration of variants routinely occurs, testing showed that such overgeneration has little adverse effect.	contrasting
train_20025	The prior work most closely related to this study is (Riloff, 1996), which, along with (Riloff, 1993), seeks automatic methods for filling slots in event templates.	the prior work differs from that presented here in several crucial respects; firstly, the prior work does not attempt to find entire events, after the fashion of MUC's highest-level scenario-template task.	contrasting
train_20026	Name lists provide an extremely efficient way of recognising names, as the only processing required is to match the name pattern in the list against the text and no expensive advanced processing such as full text parsing is required.	name lists are a naive method for recognising names.	contrasting
train_20027	We have also demonstrated that it is possible for the addition of corpus-derived lists to improve the performance of a NE recognition system based on gazetteers.	this is not guaranteed and it appears that adding too many names without any restriction may actually lead to poorer results, as happened when the LONG_TRAIN lists were applied.	contrasting
train_20028	HS also uses a frequency measure to provide an approximation of topic significance.	instead of counting frequency of stems or repetition of word sequences, this method counts frequency of a relatively easily identified grammatical element, heads of simplex noun phrases (SNPs).	contrasting
train_20029	Toward this end, we made several decisions about presentation of the data: 1.	threshold: So that no bias would be unintentionally introduced, we presented subjects with all terms output by each method, up to a specified cut-off poin-using lists of equal length for each method would have necessitated either omitting HSs and KWs or changing the definition of tts.	contrasting
train_20030	The graph in Figure 3 shows results for quality, not coverage.	figure 4, which shows the total number of terms rated at or below specified rankings, allows us to measure quality and coverage.	contrasting
train_20031	For example, the phrase 'Wall Street Journal' was included on the HS list only because it is specified as the document source.	four of the eight students assigned this term a high rating (1 or 2); this is puzzling because the document is about asbestos-related disease.	contrasting
train_20032	For example, the word exposure received an average rating of 2.2 when it appeared on the KW list, but a rating of only 2.75 on the HS list.	the more specific phrase racial quotas, which immediately followed quota on the HS list received a rating of 1.	contrasting
train_20033	One might imagine that a more carefully constructed lexicon could reduce the OOV rate for PERSONs while still staying within the 60,000 word limit.	even if a cleverly designed 60K lexicon succeeded in having the name coverage of the frequency-ordered 120K word lexicon (which contains roughly 40,000 more proper names than the 60K lexicon), it would reduce the PERSON OOV rate by only 4% absolute.	contrasting
train_20034	Once an example is in the training, IdentiFinder is able to extract it and use it in test.	when the test data is recognizer output, the rare names are less likely to appear in the test, either because they don't appear in the speech lexicon or they are poorly trained in the speech model and misrecognized.	contrasting
train_20035	Other GUI interface tools, such as a search field with command completion, can be envisioned that would provide direct access.	it is arguable that such an interface element belongs squarely to graphical user interfaces, but draws more on features of language.	contrasting
train_20036	In our case, the output of the dialog manager of a spoken dialog system provides the input to our sentence planner in the form of a single spoken dialog text plan for each of the turns.	(the dialog managers of most dialog systems today simply output completely formed utterances which are passed on to the TTS module.)	contrasting
train_20037	Perfect performance would have meant that there would be no significant difference.	the mean of BEST is 4.82 as compared with the mean of SPOT of 4.56, for a mean difference of 0.26 on a scale of 1 to 5.	contrasting
train_20038	Using hand-crafted evaluation metrics, they show that a genetic algorithm achieves good results in finding discourse trees.	they do not address clausecombining, and we do not use hand-crafted metrics.	contrasting
train_20039	The basic idea is that we view the labeling of new name ( l) as a transformation of the labeling of the old one ( l ).	we do not know l so we have to sum over all possible former labelings L .	contrasting
train_20040	In a similar way, we marked both "Dr." and "Sir" as honorifics in our test data.	the Wall Street Journal treats them very differently from "Mr." in that the former tend to be included even in the first mention of a name, while the latter is not.	contrasting
train_20041	Unfortunately, because "Mikio" does not appear elsewhere, the program is at a loss to decide which label to give it.	because Yotaro is assumed to be a first name, the program makes "Mikio Suzuki" coreferent with "Yotaro Suzuki" by labeling "Mikio" descriptor.	contrasting
train_20042	Naive Baseline The probabilistic model described in Section 2 explicitly takes adjective/adverb and verb co-occurrences into account.	one could derive meanings for polysemous adjective-noun combinations by solely concentrating on verb-noun relations, ignoring thus the adjective/adverb and verb dependencies.	contrasting
train_20043	More recently, (Gonzalo et al., 2000) also discusses potential usefulness of systematic polysemy for clustering word senses for IR.	extracting systematic relations from large sense inventories is a di cult task.	contrasting
train_20044	As a note, our lexicon is similar to CORELEX (Buitelaar, 1998) (or CORELEX-II presented in (Buitelaar, 2000)), in that both lexicons share the same motivation.	our lexicon di ers from CORELEX in that CORELEX looks at all senses of a w ord and groups words that have the same sense distribution pattern, whereas our lexicon groups \drop", \circle", \intersection", \dig", \crossing", \bull's eye" ARTIFACT-GROUP STRUCTURE, PEOPLE] \house", \convent", \market", \center" ARTIFACT-SUBSTANCE FABRIC, CHEMICAL COMPOUND] \acetate", \nylon", \acrylic", \polyester" COMMUNICATION-PERSON VOICE, SINGER] \soprano", \alto", \tenor", \baritone" WRITING, RELIGIOUS PERSON] \John", \Matthew", \Jonah", \Joshua", \Jeremiah" word senses that have the same systematic relation.	contrasting
train_20045	These two languages were chosen as the development set because they are represented by the most complete vocabularies and share the largest number of cognates.	as it turned out later, they are also the most closely related among the four Algonquian languages, according to all measures of phonetic similarity.	contrasting
train_20046	Most of the improvement comes from detecting entries that have matching glosses.	the contribution of WordNet is small.	contrasting
train_20047	In particular, decision trees can be converted into implicational rules that an expert could inspect and can in principle be compiled back into finite-state machines (Sproat and Riley, 1996), although that would re-introduce the original efficiency problems.	finite-state transducers have the advantage of being invertible, which can be exploited e. g. for testing hand-crafted rule sets.	contrasting
train_20048	This points to a fundamental shortcoming of the usual two-step procedure, which we followed here: the goodness of an alignment performed in the first step should be determined by the impact it has on producing an optimal classifier that is induced in the second step.	there is no provision for feedback from the second step to the first step.	contrasting
train_20049	For example, Core and Schubert [8] point to counterexamples such as "have the engine take the oranges to Elmira, um, I mean, take them to Corning" where the antecedent of "them" is found in the EDITED words.	we believe that the assumption is so close to true that the number of errors introduced by this assumption is small compared to the total number of errors made by the system.	contrasting
train_20050	All words are assigned one of the two possible labels, EDITED or not.	in our evaluation we report the accuracy of only words other than punctuation and filled pauses.	contrasting
train_20051	For example, [13] uses a subsection of the ATIS corpus, takes as input the actual speech signal (and thus has access to silence duration but not to words), and uses as its evaluation metric the percentage of time the program identifies the start of the interregnum (see Section 2.2).	[9,10] use an internally developed corpus of sentences, work from a transcript enhanced with information from the speech signal (and thus use words), but do use a metric that seems to be similar to ours.	contrasting
train_20052	Important technical problems involving leftrecursive and unit productions are examined and overcome in (Stolcke, 1995).	these complications do not add any further machinery to the parsing algorithm per se beyond the grammar rules and the dot-moving conventions: in particular, there are no heuristic parsing principles or intermediate structures that are later destroyed.	contrasting
train_20053	Previous studies in unsupervised methods for parsing have concentrated on the use of inside-outside algorithm (Lari and Young, 1990;Carroll and Rooth, 1998).	there are several limitations of the inside-outside algorithm for unsupervised parsing, see (Marcken, 1995) for some experiments that draw out the mismatch between minimizing error rate and iteratively increasing the likelihood of the corpus.	contrasting
train_20054	Other approaches have tried to move away from phrase structural representations into dependency style parsing (Lafferty et al., 1992;Fong and Wu, 1996).	there are still inherent computational limitations due to the vast search space (see (Pietra et al., 1994) for discussion).	contrasting
train_20055	Yet we illustrated that induced semantics can help overcome some of these errors.	we have since observed that induced semantics can give rise to different kinds of problems.	contrasting
train_20056	Despite the semantic, orthographic, and syntactic components of the algorithm, there are still valid PPMVs, (X<Y), that may seem unrelated due to corpus choice or weak distributional properties.	x and Y may appear as members of other valid PPMVs such as (x<Z) and (Z<Y) containing variants (Z, in this case) which are either semantically or syntactically related to both of the other words.	contrasting
train_20057	We will incorporate the variable context length model into our system.°C onsidering more predictable bound In our experiments, we introduce new types of voting methods which stem from the theorems of SVMs -VC bound and Leave-One-Out bound.	chapelle and Vapnik introduce an alternative and more predictable bound for the risk and report their proposed bound is quite useful for selecting the kernel function and soft margin parameter (chapelle and Vapnik, 2000).	contrasting
train_20058	These approaches have produced interesting results, including applications involving real world dialogue systems.	reinforcement learning suffers from the fact that it is state based.	contrasting
train_20059	Many of these take a usercentric approach based on Wizard of Oz studies and iterative design (Bernsen et al., 1998).	there are still no precise guidelines about when to use specific techniques such as mixed-initiative.	contrasting
train_20060	A typical method used for NE extraction of Japanese texts is a cascade of morphological analysis, POS tagging and chunking.	there are some cases where segmentation granularity contradicts the results of morphological analysis and the building units of NEs, so that extraction of some NEs are inherently impossible in this setting.	contrasting
train_20061	Then, the shorter units tend to be hidden behind the longer unit words.	introducing the shorter unit words is more necessary to named entity extraction to generalize the model, because the shorter units are shared by many compound words.	contrasting
train_20062	Our work presents a novel knowledge-lean algorithm that uses multiple-sequence alignment (MSA) to learn to generate sentence-level paraphrases essentially from unannotated corpus data alone.	to previous work using MSA for generation (Barzilay and Lee, several versions of their component sentences.	contrasting
train_20063	Previous approaches to paraphrase acquisition focused on certain rigid types of paraphrases, for instance, limiting the number of arguments.	our method is not limited to a set of a priori-specified paraphrase types.	contrasting
train_20064	For example, (Clarke et al., 2003) and (Lin et al., 2003) employ techniques for utilizing both unstruc-tured text and structured databases for question answering.	the approaches taken by both these systems differ from ours in that they enforce an order between the two strategies by attempting to locate answers in structured databases first for select question types and falling back to unstructured text when the former fails, while we explore both options in parallel and combine the results from multiple answering agents.	contrasting
train_20065	Such knowledge would include both the spelling and pronunciation of a new word, as well as an understanding of its usage in the language (e.g., a semantic category).	this is a difficult task to carry out effectively, challenging both with regard to the automatic acquisition of the sound-to-letter mapping from typically telephone-quality speech, and the system level aspect of integrating the usually off-line activities of system upgrade while seamlessly continuing the conversation with the user.	contrasting
train_20066	Deletions occur when part of the spelled portion is mistakenly identified as part of the unknown word or insertions arise when the end of a spoken word is confused for a spelled letter.	the multi-stage system produces a marked improvement if we compare it with the single-stage letter recognizer as a baseline.	contrasting
train_20067	Our final selection process is based only on the proposed spellings obtained from the pronounced word, after feeding information from the spelled part into the second stage.	performance may improve if we apply a strict constraint during the search, explicitly allowing only paths where the spoken and spelled part of the waveforms agree on the name spelling.	contrasting
train_20068	These cases are of central interest in language and speech processing.	automata with other kinds of weights can also be defined.	contrasting
train_20069	We have found no simple necessary and sufficient condition on (K, ⊗) that guarantees a globally consistent set of choices to exist.	we have given a useful nec-essary condition (greedy factorization), and we now give a useful sufficient condition.	contrasting
train_20070	SWAP any two non-overlapping regions .	if we limit the size of the swapped regions to a constant I and their distance to a constant P , we can reduce the number of swaps performed to a linear function of the input length.	contrasting
train_20071	The complexity of calculating the alignment probability globally (that is, over the entire alignment) is ¢ ¡£ § .	since there is a constant upper bound 3 on the size of local contexts, needs to be performed only once for the initial gloss, therafter, recalculation of only those probabilities affected by each change ( ) suffices.	contrasting
train_20072	For the Arabic test data from the same evaluation, we obtained a similar shape (although with a roughly level plateau).	the 'bumpiness' of the surface raises the question as to which of these differences are statistically We are aware of several ways to determine the statistical significance of BLEU score differences.	contrasting
train_20073	For example, both relations ¡ apartment 1; woman 1; No¢ and ¡ hand 1; woman 1; Yes¢ are mapped into the more general type ¡ entity 1; entity 1; Yes/No¢ .	the first example is negative (a POSSESSION relation), while the second one is a positive example.	contrasting
train_20074	For the examples described above, the procedure eliminates the ambiguity through specialization of the semantic classes into two new ones: whole -causal agent, and respectively part -causal agent.	if the training corpus contained the examples ¡ leg 2; insect 1; Yes¢ and ¡ world 7; insect 1; No¢ , the procedure specializes them in the ambiguous example ¡ part 7; organism 1; Yes/No¢ and the ambiguity still remains.	contrasting
train_20075	Word sense disambiguation could, therefore, also independently be performed using the semantic coherence scoring described herein as an additional application of our approach.	that has not been investigated thoroughly yet.	contrasting
train_20076	There were 209 added words (about 26%).	128 words (or 61% of missing words) were not actually missing, but rather not linked into the set of clusters evaluated by a particular annotator.	contrasting
train_20077	Schuetze (1998) proposed a method for dividing occurrences of a word into classes, each of which consists of contextually similar occurrences.	it does not produce definitions of senses such as sets of synonyms and sets of translation equivalents.	contrasting
train_20078	Generally speaking, the problem becomes more difficult in translingual distributional word clustering, since the sparseness of data in two languages is multiplied.	the sense-vs.-clue correlation matrix calculation method overcomes this difficulty; it calculates the correlations between senses and clues iteratively to smooth out the sparse data.	contrasting
train_20079	4) Cluster senses by using a hierarchical agglomerative clustering method, e.g., the group-average method.	this naive method is not effective because some senses usually have duplicated definitions in step 1) despite the fact that the sense-vs.-clue correlation matrix calculation algorithm presupposes a set of senses without duplicated definitions.	contrasting
train_20080	To situate our results, the FOMs used by (Caraballo and Charniak, 1998) require 10K edges to parse 96% of these sentences, while BF requires only 6K edges.	the more complex, tuned FOM in is able to parse all of these sentences using around 2K edges, while BF requires 7K edges.	contrasting
train_20081	We first observe that the complete set of word alignments generated by the ATTM (ATTM-C) is relatively poor.	when we consider only those word alignments generated by actual alignment templates (ATTM-A) (and discard the alignments generated by the dummy templates introduced as described in Section 3.1), we obtain very high alignment precision.	contrasting
train_20082	These components can be refined and improved by changing the corresponding transducers without requiring changes to the overall search procedure.	some of the modeling assumptions are extremely strong.	contrasting
train_20083	The Document Understanding Conference (DUC 2002) run by the National Institute of Standards and Technology (NIST) sets out to address this problem by providing annual large scale common evaluations in text summarization.	these evaluations involve human judges and hence are subject to variability (Rath et al.	contrasting
train_20084	It consistently correlated highly with human assessments and had high recall and precision in significance test with manual evaluation results.	the weighted average of variable length n-gram matches derived from IBM BLEU did not always give good correlation and high recall and precision.	contrasting
train_20085	For their purposes, they mainly need to ensure the correctness of consensus among different translations, so that different constituent orderings in input sentences do not pose a serious prob- lem.	we want to ensure the correctness of all paths represented by the FSAs, and direct application of MSA in the presence of different constituent orderings can be problematic.	contrasting
train_20086	From the curves, one can see that as the order increases, classification accuracy increases and testing entropy decreases, presumably because the longer context better captures the regularities of the text.	at some point accu- Figure 2: The entropy of different n-gram models racy begins to decrease and entropy begins to increase as the sparse data problems begin to set in.	contrasting
train_20087	Even if all f -structures are transferred from the same linguistically rich source, and all generated strings are grammatical, a reduction in error rate of around 50% relative to the upper bound can be achieved by stochastic selection.	a comparison between transfer runs with and without perfect disambiguation of the original string shows a decrease of about 5% in F-score, and of only .1 points for summarization quality when transferring from packed parses instead of from the manually selected parse.	contrasting
train_20088	In our case, combining the counts in this way yielded a half a point, perhaps because of the in-domain tuning of the smoothing parameters.	when we optimize α empirically on the held-out corpus, we can get nearly a full point improvement.	contrasting
train_20089	Furthermore, they are trained to minimize some function related to labeling error, leading to smaller error in practice if enough training data are available.	generative models are trained to maximize the joint probability of the training data, which is not as closely tied to the accuracy metrics of interest if the actual data was not generated by the model, as is always the case in practice.	contrasting
train_20090	In contrast, generative models are trained to maximize the joint probability of the training data, which is not as closely tied to the accuracy metrics of interest if the actual data was not generated by the model, as is always the case in practice.	since sequential classifiers are trained to make the best local decision, unlike generative models they cannot trade off decisions at different positions against each other.	contrasting
train_20091	The ap-proximated diagonal term H f for feature f has the form If this approximation is semidefinite, which is trivial to check, its inverse is an excellent preconditioner for early iterations of CG training.	when the model is close to the maximum, the approximation becomes unstable, which is not surprising since it is based on feature independence assumptions that become invalid as the weights of interaction features move away from zero.	contrasting
train_20092	Mixed CG training converges slightly more slowly than preconditioned CG.	cG without preconditioner converges much more slowly than both preconditioned cG and mixed cG training.	contrasting
train_20093	On the other hand, CG without preconditioner converges much more slowly than both preconditioned CG and mixed CG training.	it is still much faster than GIS.	contrasting
train_20094	On the application side, (log-)linear parsing models have the potential to supplant the currently dominant lexicalized PCFG models for parsing by allowing much richer feature sets and simpler smoothing, while avoiding the label bias problem that may have hindered earlier classifier-based parsers (Ratnaparkhi, 1997).	work in that direction has so far addressed only parse reranking (Collins and Duffy, 2002;Riezler et al., 2002).	contrasting
train_20095	In order to use more information, we might imagine using values of ec directly, rather than thresholding.	this quickly leads to data sparsity problems.	contrasting
train_20096	Various linguistics studies have also shown how intertwined syntax and discourse are (Maynard, 1998).	to our knowledge, this is the first paper that empirically shows that the connection between syntax and discourse can be computationally exploited at high levels of accuracy on open domain, newspaper text.	contrasting
train_20097	Little improvement comes from using manually built syntactic parse trees instead of automatically derived trees.	experiments show that there is much to be gained if better discourse segmentation algorithms are found; 83% accuracy on this task is not sufficient for building highly accurate discourse trees.	contrasting
train_20098	Current state-of-the-art statistical parsers (Collins, 1999;Charniak, 2000) are trained on large annotated corpora such as the Penn Treebank (Marcus et al., 1993).	the production of such corpora is expensive and labor-intensive.	contrasting
train_20099	One such criterion is the accuracy of the labeled examples, which may be estimated by the teacher parser's confidence in its labels.	the examples that the teacher correctly labeled may not be those that the student needs.	contrasting

train_20000

Naturally, these are not appropriate comparisons for the work reported here.

as is evident from the discussion above, previous spelling research does provide an important role in suggesting productive features to include in the decision tree.

contrasting

train_20001

Capitalization is his sole means of identifying names.

capitalization information is not available in closed captions.

contrasting

train_20002

More importantly, we investigate the relative role of each statistical and linguistic knowledge source in the proposed IR/NLP question-answering system.

to previous results, we find that statistical knowledge of word co-occurrences as computed by vector space models of IR can be used to quickly and accurately locate relevant documents in the restricted QA task.

contrasting

train_20003

In the cases we have investigated, this strategy works often enough to be useful for the surface object.

due to predicate coordination as well as relativization, such a strategy often fails to identify correctly the surface subject of bind (binds or bound) when more than one binding term precedes the verb.

contrasting

train_20004

Basically, the segmenter is based on frequency of individual nouns extracted from corpus.

the problem is that it is difficult to distinguish proper noun and common noun since there is no clue like capital letters in Korean.

contrasting

train_20005

Assume that 'sl' is a frequently appearing word in texts whereas 's2s3s4' is a rarely occurring sequence of syllables as a word.

'sis2' and 's3s4' occurs frequently but although they don't occur as frequently as 'sl'.

contrasting

train_20006

Actually, gyosaeng-hwal is not a word.

both hag-gyo and saeng-hwal are frequently occurring syllables, and actually they are all words.

contrasting

train_20007

It is important to note here that we would probably have obtained this improvement in recognition accuracy even with a manual revision of the grammars.

the main advantage in using our tool is the tremendous simplification of the whole process of revision for a grammar developer who now selects counter-examples with an interactive tool instead of manually revising the grammars.

contrasting

train_20008

We have successfully used the tool to constrain overgeneralizing grammars of speech understanding systems and obtained 20-30% higher recognition accuracy.

we believe the primary benefit of using our tool is the tremendously reduced effort for the grammar developer.

contrasting

train_20009

This is a language dependent approach.

we propose a framework of language independent morphological analysis system.

contrasting

train_20010

Disambiguation i s already language independent, since it does not process strings directly and therefore will not be taken up.

tokenization and dictionary look-up are language dependent and shall be explained more in this paper.

contrasting

train_20011

If a hyphenated segment such as "data-base," "F-16," or "MS-DOS" exists in the dictionary, it should be an independent lexeme.

if a hyphenated segment such as "55-years-old" does not exist in the dictionary, hyphens should be treated as independent tokens (Fox, 1992).

contrasting

train_20012

Then we pass it to the analyzer for non-segmented languages.

the analyzer may return the result as "They / 've / gone / to / school / to / get / her / ."

contrasting

train_20013

The set of the delimiters acting as boundaries: These act as boundaries of MFs.

these can not be independent MFs (can not start nor end a lexeme).

contrasting

train_20014

These problems are discussed by several authors including Mills (Mills, 1998) Webster & Kit (Webster and Kit, 1992).

their proposed solutions are not language independent.

contrasting

train_20015

Using case-sensitive rules is straightforward since generally only nouns (and proper names) are written in standard German with a capitalized initial letter (e.g., "das Unternehmen" -the enterprise vs. "wir unternehmen" -we undertake).

for disambiguation of word forms appearing at the beginning of the sentence local contextual filtering rules are applied.

contrasting

train_20016

company names) may appear in the text either with or without a designator, we use a dynamic lexicon to store recognized named entities without their designators (e.g., "Braun AG" vs. "Braun") in order to identify subsequent occurrences correctly.

a named entity, consisting solely of one word, may be also a valid word form (e.g., "Braun" -brown).

contrasting

train_20017

This means that although the exact attachment point of each individual PP is not known it is guaranteed that a PP can only be attached to phrases which are dominated by the main verb of the sentence (which is the root node of the clause's tree).

the exact point of attachment is a matter of domain-specific knowledge and hence should be defined as part of the domain knowledge of an application.

contrasting

train_20018

Only in the case of the MC-module the difference is about 5%.

the result for the isolated evaluation of the MC-module suggests that this is mainly due to errors caused by previous components.

contrasting

train_20019

The output document from constrained HMM contains MUC-standard NE.tags such as person, location and organization.

for a real information extraction system, the MUC-standard NE tag may not be enough and further detailed NE information might be necessary.

contrasting

train_20020

Further adding the morphological analysis capability that automatically analyzes unknown words (deriving additional morphological relationships and some semantic subsumption relationships) significantly improves that result (from 50.0% to 60.7%).

we found that adding the same semantic subsumption relationships to the commercial search engine, using its provided thesaurus capability degraded its results, and results were still degraded when we added only those facts that we knew would help find relevant documents.

contrasting

train_20021

Knowledge classifiers are part of almost any knowledge representation system.

the problem we face here is more difficult.

contrasting

train_20022

By default this concept is placed under policy.

in WordNet there is a hierarchy fiscal policy -IS-Aeconomic policy -IS-A -policy.

contrasting

train_20023

Higher interest rates are normally associated with weaker bond markets.

if interest rates go down, bonds go up, and your bond becomes more valuable.

contrasting

train_20024

Overgeneration will produce some name variants that have nothing to do with the company.

although overgeneration of variants routinely occurs, testing showed that such overgeneration has little adverse effect.

contrasting

train_20025

The prior work most closely related to this study is (Riloff, 1996), which, along with (Riloff, 1993), seeks automatic methods for filling slots in event templates.

the prior work differs from that presented here in several crucial respects; firstly, the prior work does not attempt to find entire events, after the fashion of MUC's highest-level scenario-template task.

contrasting

train_20026

Name lists provide an extremely efficient way of recognising names, as the only processing required is to match the name pattern in the list against the text and no expensive advanced processing such as full text parsing is required.

name lists are a naive method for recognising names.

contrasting

train_20027

We have also demonstrated that it is possible for the addition of corpus-derived lists to improve the performance of a NE recognition system based on gazetteers.

this is not guaranteed and it appears that adding too many names without any restriction may actually lead to poorer results, as happened when the LONG_TRAIN lists were applied.

contrasting

train_20028

HS also uses a frequency measure to provide an approximation of topic significance.

instead of counting frequency of stems or repetition of word sequences, this method counts frequency of a relatively easily identified grammatical element, heads of simplex noun phrases (SNPs).

contrasting

train_20029

Toward this end, we made several decisions about presentation of the data: 1.

threshold: So that no bias would be unintentionally introduced, we presented subjects with all terms output by each method, up to a specified cut-off poin-using lists of equal length for each method would have necessitated either omitting HSs and KWs or changing the definition of tts.

contrasting

train_20030

The graph in Figure 3 shows results for quality, not coverage.

figure 4, which shows the total number of terms rated at or below specified rankings, allows us to measure quality and coverage.

contrasting

train_20031

For example, the phrase 'Wall Street Journal' was included on the HS list only because it is specified as the document source.

four of the eight students assigned this term a high rating (1 or 2); this is puzzling because the document is about asbestos-related disease.

contrasting

train_20032

For example, the word exposure received an average rating of 2.2 when it appeared on the KW list, but a rating of only 2.75 on the HS list.

the more specific phrase racial quotas, which immediately followed quota on the HS list received a rating of 1.

contrasting

train_20033

One might imagine that a more carefully constructed lexicon could reduce the OOV rate for PERSONs while still staying within the 60,000 word limit.

even if a cleverly designed 60K lexicon succeeded in having the name coverage of the frequency-ordered 120K word lexicon (which contains roughly 40,000 more proper names than the 60K lexicon), it would reduce the PERSON OOV rate by only 4% absolute.

contrasting

train_20034

Once an example is in the training, IdentiFinder is able to extract it and use it in test.

when the test data is recognizer output, the rare names are less likely to appear in the test, either because they don't appear in the speech lexicon or they are poorly trained in the speech model and misrecognized.

contrasting

train_20035

Other GUI interface tools, such as a search field with command completion, can be envisioned that would provide direct access.

it is arguable that such an interface element belongs squarely to graphical user interfaces, but draws more on features of language.

contrasting

train_20036

In our case, the output of the dialog manager of a spoken dialog system provides the input to our sentence planner in the form of a single spoken dialog text plan for each of the turns.

(the dialog managers of most dialog systems today simply output completely formed utterances which are passed on to the TTS module.)

contrasting

train_20037

Perfect performance would have meant that there would be no significant difference.

the mean of BEST is 4.82 as compared with the mean of SPOT of 4.56, for a mean difference of 0.26 on a scale of 1 to 5.

contrasting

train_20038

Using hand-crafted evaluation metrics, they show that a genetic algorithm achieves good results in finding discourse trees.

they do not address clausecombining, and we do not use hand-crafted metrics.

contrasting

train_20039

The basic idea is that we view the labeling of new name ( l) as a transformation of the labeling of the old one ( l ).

we do not know l so we have to sum over all possible former labelings L .

contrasting

train_20040

In a similar way, we marked both "Dr." and "Sir" as honorifics in our test data.

the Wall Street Journal treats them very differently from "Mr." in that the former tend to be included even in the first mention of a name, while the latter is not.

contrasting

train_20041

Unfortunately, because "Mikio" does not appear elsewhere, the program is at a loss to decide which label to give it.

because Yotaro is assumed to be a first name, the program makes "Mikio Suzuki" coreferent with "Yotaro Suzuki" by labeling "Mikio" descriptor.

contrasting

train_20042

Naive Baseline The probabilistic model described in Section 2 explicitly takes adjective/adverb and verb co-occurrences into account.

one could derive meanings for polysemous adjective-noun combinations by solely concentrating on verb-noun relations, ignoring thus the adjective/adverb and verb dependencies.

contrasting

train_20043

More recently, (Gonzalo et al., 2000) also discusses potential usefulness of systematic polysemy for clustering word senses for IR.

extracting systematic relations from large sense inventories is a di cult task.

contrasting

train_20044

As a note, our lexicon is similar to CORELEX (Buitelaar, 1998) (or CORELEX-II presented in (Buitelaar, 2000)), in that both lexicons share the same motivation.

our lexicon di ers from CORELEX in that CORELEX looks at all senses of a w ord and groups words that have the same sense distribution pattern, whereas our lexicon groups \drop", \circle", \intersection", \dig", \crossing", \bull's eye" ARTIFACT-GROUP STRUCTURE, PEOPLE] \house", \convent", \market", \center" ARTIFACT-SUBSTANCE FABRIC, CHEMICAL COMPOUND] \acetate", \nylon", \acrylic", \polyester" COMMUNICATION-PERSON VOICE, SINGER] \soprano", \alto", \tenor", \baritone" WRITING, RELIGIOUS PERSON] \John", \Matthew", \Jonah", \Joshua", \Jeremiah" word senses that have the same systematic relation.

contrasting

train_20045

These two languages were chosen as the development set because they are represented by the most complete vocabularies and share the largest number of cognates.

as it turned out later, they are also the most closely related among the four Algonquian languages, according to all measures of phonetic similarity.

contrasting

train_20046

Most of the improvement comes from detecting entries that have matching glosses.

the contribution of WordNet is small.

contrasting

train_20047

In particular, decision trees can be converted into implicational rules that an expert could inspect and can in principle be compiled back into finite-state machines (Sproat and Riley, 1996), although that would re-introduce the original efficiency problems.

finite-state transducers have the advantage of being invertible, which can be exploited e. g. for testing hand-crafted rule sets.

contrasting

train_20048

This points to a fundamental shortcoming of the usual two-step procedure, which we followed here: the goodness of an alignment performed in the first step should be determined by the impact it has on producing an optimal classifier that is induced in the second step.

there is no provision for feedback from the second step to the first step.

contrasting

train_20049

For example, Core and Schubert [8] point to counterexamples such as "have the engine take the oranges to Elmira, um, I mean, take them to Corning" where the antecedent of "them" is found in the EDITED words.

we believe that the assumption is so close to true that the number of errors introduced by this assumption is small compared to the total number of errors made by the system.

contrasting

train_20050

All words are assigned one of the two possible labels, EDITED or not.

in our evaluation we report the accuracy of only words other than punctuation and filled pauses.

contrasting

train_20051

For example, [13] uses a subsection of the ATIS corpus, takes as input the actual speech signal (and thus has access to silence duration but not to words), and uses as its evaluation metric the percentage of time the program identifies the start of the interregnum (see Section 2.2).

[9,10] use an internally developed corpus of sentences, work from a transcript enhanced with information from the speech signal (and thus use words), but do use a metric that seems to be similar to ours.

contrasting

train_20052

Important technical problems involving leftrecursive and unit productions are examined and overcome in (Stolcke, 1995).

these complications do not add any further machinery to the parsing algorithm per se beyond the grammar rules and the dot-moving conventions: in particular, there are no heuristic parsing principles or intermediate structures that are later destroyed.

contrasting

train_20053

Previous studies in unsupervised methods for parsing have concentrated on the use of inside-outside algorithm (Lari and Young, 1990;Carroll and Rooth, 1998).

there are several limitations of the inside-outside algorithm for unsupervised parsing, see (Marcken, 1995) for some experiments that draw out the mismatch between minimizing error rate and iteratively increasing the likelihood of the corpus.

contrasting

train_20054

Other approaches have tried to move away from phrase structural representations into dependency style parsing (Lafferty et al., 1992;Fong and Wu, 1996).

there are still inherent computational limitations due to the vast search space (see (Pietra et al., 1994) for discussion).

contrasting

train_20055

Yet we illustrated that induced semantics can help overcome some of these errors.

we have since observed that induced semantics can give rise to different kinds of problems.

contrasting

train_20056

Despite the semantic, orthographic, and syntactic components of the algorithm, there are still valid PPMVs, (X<Y), that may seem unrelated due to corpus choice or weak distributional properties.

x and Y may appear as members of other valid PPMVs such as (x<Z) and (Z<Y) containing variants (Z, in this case) which are either semantically or syntactically related to both of the other words.

contrasting

train_20057

We will incorporate the variable context length model into our system.°C onsidering more predictable bound In our experiments, we introduce new types of voting methods which stem from the theorems of SVMs -VC bound and Leave-One-Out bound.

chapelle and Vapnik introduce an alternative and more predictable bound for the risk and report their proposed bound is quite useful for selecting the kernel function and soft margin parameter (chapelle and Vapnik, 2000).

contrasting

train_20058

These approaches have produced interesting results, including applications involving real world dialogue systems.

reinforcement learning suffers from the fact that it is state based.

contrasting

train_20059

Many of these take a usercentric approach based on Wizard of Oz studies and iterative design (Bernsen et al., 1998).

there are still no precise guidelines about when to use specific techniques such as mixed-initiative.

contrasting

train_20060

A typical method used for NE extraction of Japanese texts is a cascade of morphological analysis, POS tagging and chunking.

there are some cases where segmentation granularity contradicts the results of morphological analysis and the building units of NEs, so that extraction of some NEs are inherently impossible in this setting.

contrasting

train_20061

Then, the shorter units tend to be hidden behind the longer unit words.

introducing the shorter unit words is more necessary to named entity extraction to generalize the model, because the shorter units are shared by many compound words.

contrasting

train_20062

Our work presents a novel knowledge-lean algorithm that uses multiple-sequence alignment (MSA) to learn to generate sentence-level paraphrases essentially from unannotated corpus data alone.

to previous work using MSA for generation (Barzilay and Lee, several versions of their component sentences.

contrasting

train_20063

Previous approaches to paraphrase acquisition focused on certain rigid types of paraphrases, for instance, limiting the number of arguments.

our method is not limited to a set of a priori-specified paraphrase types.

contrasting

train_20064

For example, (Clarke et al., 2003) and (Lin et al., 2003) employ techniques for utilizing both unstruc-tured text and structured databases for question answering.

the approaches taken by both these systems differ from ours in that they enforce an order between the two strategies by attempting to locate answers in structured databases first for select question types and falling back to unstructured text when the former fails, while we explore both options in parallel and combine the results from multiple answering agents.

contrasting

train_20065

Such knowledge would include both the spelling and pronunciation of a new word, as well as an understanding of its usage in the language (e.g., a semantic category).

this is a difficult task to carry out effectively, challenging both with regard to the automatic acquisition of the sound-to-letter mapping from typically telephone-quality speech, and the system level aspect of integrating the usually off-line activities of system upgrade while seamlessly continuing the conversation with the user.

contrasting

train_20066

Deletions occur when part of the spelled portion is mistakenly identified as part of the unknown word or insertions arise when the end of a spoken word is confused for a spelled letter.

the multi-stage system produces a marked improvement if we compare it with the single-stage letter recognizer as a baseline.

contrasting

train_20067

Our final selection process is based only on the proposed spellings obtained from the pronounced word, after feeding information from the spelled part into the second stage.

performance may improve if we apply a strict constraint during the search, explicitly allowing only paths where the spoken and spelled part of the waveforms agree on the name spelling.

contrasting

train_20068

These cases are of central interest in language and speech processing.

automata with other kinds of weights can also be defined.

contrasting

train_20069

We have found no simple necessary and sufficient condition on (K, ⊗) that guarantees a globally consistent set of choices to exist.

we have given a useful nec-essary condition (greedy factorization), and we now give a useful sufficient condition.

contrasting

train_20070

SWAP any two non-overlapping regions .

if we limit the size of the swapped regions to a constant I and their distance to a constant P , we can reduce the number of swaps performed to a linear function of the input length.

contrasting

train_20071

The complexity of calculating the alignment probability globally (that is, over the entire alignment) is ¢ ¡£ § .

since there is a constant upper bound 3 on the size of local contexts, needs to be performed only once for the initial gloss, therafter, recalculation of only those probabilities affected by each change ( ) suffices.

contrasting

train_20072

For the Arabic test data from the same evaluation, we obtained a similar shape (although with a roughly level plateau).

the 'bumpiness' of the surface raises the question as to which of these differences are statistically We are aware of several ways to determine the statistical significance of BLEU score differences.

contrasting

train_20073

For example, both relations ¡ apartment 1; woman 1; No¢ and ¡ hand 1; woman 1; Yes¢ are mapped into the more general type ¡ entity 1; entity 1; Yes/No¢ .

the first example is negative (a POSSESSION relation), while the second one is a positive example.

contrasting

train_20074

For the examples described above, the procedure eliminates the ambiguity through specialization of the semantic classes into two new ones: whole -causal agent, and respectively part -causal agent.

if the training corpus contained the examples ¡ leg 2; insect 1; Yes¢ and ¡ world 7; insect 1; No¢ , the procedure specializes them in the ambiguous example ¡ part 7; organism 1; Yes/No¢ and the ambiguity still remains.

contrasting

train_20075

Word sense disambiguation could, therefore, also independently be performed using the semantic coherence scoring described herein as an additional application of our approach.

that has not been investigated thoroughly yet.

contrasting

train_20076

There were 209 added words (about 26%).

128 words (or 61% of missing words) were not actually missing, but rather not linked into the set of clusters evaluated by a particular annotator.

contrasting

train_20077

Schuetze (1998) proposed a method for dividing occurrences of a word into classes, each of which consists of contextually similar occurrences.

it does not produce definitions of senses such as sets of synonyms and sets of translation equivalents.

contrasting

train_20078

Generally speaking, the problem becomes more difficult in translingual distributional word clustering, since the sparseness of data in two languages is multiplied.

the sense-vs.-clue correlation matrix calculation method overcomes this difficulty; it calculates the correlations between senses and clues iteratively to smooth out the sparse data.

contrasting

train_20079

4) Cluster senses by using a hierarchical agglomerative clustering method, e.g., the group-average method.

this naive method is not effective because some senses usually have duplicated definitions in step 1) despite the fact that the sense-vs.-clue correlation matrix calculation algorithm presupposes a set of senses without duplicated definitions.

contrasting

train_20080

To situate our results, the FOMs used by (Caraballo and Charniak, 1998) require 10K edges to parse 96% of these sentences, while BF requires only 6K edges.

the more complex, tuned FOM in is able to parse all of these sentences using around 2K edges, while BF requires 7K edges.

contrasting

train_20081

We first observe that the complete set of word alignments generated by the ATTM (ATTM-C) is relatively poor.

when we consider only those word alignments generated by actual alignment templates (ATTM-A) (and discard the alignments generated by the dummy templates introduced as described in Section 3.1), we obtain very high alignment precision.

contrasting

train_20082

These components can be refined and improved by changing the corresponding transducers without requiring changes to the overall search procedure.

some of the modeling assumptions are extremely strong.

contrasting

train_20083

The Document Understanding Conference (DUC 2002) run by the National Institute of Standards and Technology (NIST) sets out to address this problem by providing annual large scale common evaluations in text summarization.

these evaluations involve human judges and hence are subject to variability (Rath et al.

contrasting

train_20084

It consistently correlated highly with human assessments and had high recall and precision in significance test with manual evaluation results.

the weighted average of variable length n-gram matches derived from IBM BLEU did not always give good correlation and high recall and precision.

contrasting

train_20085

For their purposes, they mainly need to ensure the correctness of consensus among different translations, so that different constituent orderings in input sentences do not pose a serious prob- lem.

we want to ensure the correctness of all paths represented by the FSAs, and direct application of MSA in the presence of different constituent orderings can be problematic.

contrasting

train_20086

From the curves, one can see that as the order increases, classification accuracy increases and testing entropy decreases, presumably because the longer context better captures the regularities of the text.

at some point accu- Figure 2: The entropy of different n-gram models racy begins to decrease and entropy begins to increase as the sparse data problems begin to set in.

contrasting

train_20087

Even if all f -structures are transferred from the same linguistically rich source, and all generated strings are grammatical, a reduction in error rate of around 50% relative to the upper bound can be achieved by stochastic selection.

a comparison between transfer runs with and without perfect disambiguation of the original string shows a decrease of about 5% in F-score, and of only .1 points for summarization quality when transferring from packed parses instead of from the manually selected parse.

contrasting

train_20088

In our case, combining the counts in this way yielded a half a point, perhaps because of the in-domain tuning of the smoothing parameters.

when we optimize α empirically on the held-out corpus, we can get nearly a full point improvement.

contrasting

train_20089

Furthermore, they are trained to minimize some function related to labeling error, leading to smaller error in practice if enough training data are available.

generative models are trained to maximize the joint probability of the training data, which is not as closely tied to the accuracy metrics of interest if the actual data was not generated by the model, as is always the case in practice.

contrasting

train_20090

In contrast, generative models are trained to maximize the joint probability of the training data, which is not as closely tied to the accuracy metrics of interest if the actual data was not generated by the model, as is always the case in practice.

since sequential classifiers are trained to make the best local decision, unlike generative models they cannot trade off decisions at different positions against each other.

contrasting

train_20091

The ap-proximated diagonal term H f for feature f has the form If this approximation is semidefinite, which is trivial to check, its inverse is an excellent preconditioner for early iterations of CG training.

when the model is close to the maximum, the approximation becomes unstable, which is not surprising since it is based on feature independence assumptions that become invalid as the weights of interaction features move away from zero.

contrasting

train_20092

Mixed CG training converges slightly more slowly than preconditioned CG.

cG without preconditioner converges much more slowly than both preconditioned cG and mixed cG training.

contrasting

train_20093

On the other hand, CG without preconditioner converges much more slowly than both preconditioned CG and mixed CG training.

it is still much faster than GIS.

contrasting

train_20094

On the application side, (log-)linear parsing models have the potential to supplant the currently dominant lexicalized PCFG models for parsing by allowing much richer feature sets and simpler smoothing, while avoiding the label bias problem that may have hindered earlier classifier-based parsers (Ratnaparkhi, 1997).

work in that direction has so far addressed only parse reranking (Collins and Duffy, 2002;Riezler et al., 2002).

contrasting

train_20095

In order to use more information, we might imagine using values of ec directly, rather than thresholding.

this quickly leads to data sparsity problems.

contrasting

train_20096

Various linguistics studies have also shown how intertwined syntax and discourse are (Maynard, 1998).

to our knowledge, this is the first paper that empirically shows that the connection between syntax and discourse can be computationally exploited at high levels of accuracy on open domain, newspaper text.

contrasting

train_20097

Little improvement comes from using manually built syntactic parse trees instead of automatically derived trees.

experiments show that there is much to be gained if better discourse segmentation algorithms are found; 83% accuracy on this task is not sufficient for building highly accurate discourse trees.

contrasting

train_20098

Current state-of-the-art statistical parsers (Collins, 1999;Charniak, 2000) are trained on large annotated corpora such as the Penn Treebank (Marcus et al., 1993).

the production of such corpora is expensive and labor-intensive.

contrasting

train_20099

One such criterion is the accuracy of the labeled examples, which may be estimated by the teacher parser's confidence in its labels.

the examples that the teacher correctly labeled may not be those that the student needs.

contrasting