tasksource/scinli · Datasets at Hugging Face

id stringlengths 7 12	sentence1 stringlengths 6 1.27k	sentence2 stringlengths 6 926	label stringclasses 4 values
train_20300	Second, if we compare the best performance achieved in each feature subspace, we can see that for both classifiers, syntactic parse tree is the most effective feature space, while sequence and dependency tree are similar.	the difference in performance between the syntactic parse tree subspace and the other two subspaces is not very large.	contrasting
train_20301	The standard methods for inferring the parameters of probabilistic models in computational linguistics are based on the principle of maximum-likelihood estimation; for example, the parameters of Probabilistic Context-Free Grammars (PCFGs) are typically estimated from strings of terminals using the Inside-Outside (IO) algorithm, an instance of the Expectation Maximization (EM) procedure (Lari and Young, 1990).	much recent work in machine learning and statistics has turned away from maximum-likelihood in favor of Bayesian methods, and there is increasing interest in Bayesian methods in computational linguistics as well (Finkel et al., 2006).	contrasting
train_20302	We also experimented with a version of IO modified to perform Bayesian MAP estimation, where the Maximization step of the IO procedure is replaced with Bayesian inference using a Dirichlet prior, i.e., where the rule probabilities θ (k) at iteration k are estimated using: Clearly such an approach is very closely related to the Bayesian procedures presented in this article, and in some circumstances this may be a useful estimator.	in our experiments with the Sesotho data above we found that for the small values of α necessary to obtain a sparse solution,the expected rule count E[f r ] for many rules r was less than 1 − α.	contrasting
train_20303	Thus The spans counted by ℓ 1 and ℓ 2 correspond to continuous spans counted in computing the complexity of a chart parsing operation.	unlike in the diagrams in the earlier part of this paper, in our graph theoretic argument there is no requirement that T select only corresponding pairs of vertices from V 1 and V 2 .	contrasting
train_20304	Schone and Jurafsky's (2001) work represents one of the few attempts to address this inappropriate morpheme attachment problem, introducing a method that exploits the semantic relatedness between word pairs.	we propose two arguably simpler, yet effective techniques that rely on relative corpus frequency and suffix level similarity to solve the problem.	contrasting
train_20305	The random mode serves as an absolute baseline in this respect: it indicates how well a particular base generator performs on its own.	different base generators have different effects on the generation modes.	contrasting
train_20306	a statistical model that translates MRL to NL.	there has been little, if any, research on exploiting recent SMT methods for NLG.	contrasting
train_20307	players 2 3 7 5, or not learn any phrases for those predicates at all.	only 4.47% of the phrases generated during testing for GEOQUERY were discontiguous, so the advantage of WASP −1 ++ over PHARAOH++ was not as obvious.	contrasting
train_20308	The three next sentences display some advantages of our approach over the K&M model: here, the latter model performs deletion with too little lexicosyntactic information, and accidentally removes certain modifiers that are sometimes, but not always, good candidates for deletions (e.g., ADJP in Sentence 2, PP in sentences 3 and 4).	our model keeps these constituent intact.	contrasting
train_20309	Research on WSD has concentrated on using contextual information, which may be limited for infrequent unknown words.	chinese characters carry semantic information that is useful for predicting the semantic properties of the words containing them.	contrasting
train_20310	Rule-based models have not been used for this task before.	there are some regularities in the relationship between the semantic categories of unknown words and those of their component characters that can be captured in a more direct and effective way by linguistic rules than by statistical models.	contrasting
train_20311	Variant 2 performs better when only the best guess is considered, indicating that it is useful to model the effect of position on a character's contribution to word meaning in this case.	it is not helpful to be sensitive to character position when the best five guesses are considered.	contrasting
train_20312	These outputs could simply be replaced with the perfect SVM predictions (1 for the true class, -1 elsewhere) since the labels are known.	the second-pass learner might actually benefit from the information contained in the misclassifications.	contrasting
train_20313	The success of factoid QA systems, particularly in the NIST-sponsored TREC evaluations (Voorhees, 2003), has reinforced the perception about the superiority of QA systems over traditional IR engines.	is QA really better than IR?	contrasting
train_20314	that answers a particular question-and the development of such technology has been encouraged by the setup of the TREC QA tracks.	a study by shows that users actually prefer answers embedded within some sort of context, e.g., the sentence or the paragraph that the answer was found in.	contrasting
train_20315	Previous work (Shapiro and Taksa, 2003) in the web environment attempted to convert a user's natural language query into one suited for use with web search engines.	the focus was on merging the results from using different sub-queries, and not selection of a single sub-query.	contrasting
train_20316	The modified Powell's method has been previously used in optimizing the weights of a standard feature-based MT decoder in (Och, 2003) where a more efficient algorithm for log-linear models was proposed.	this is specific to log-linear models and cannot be easily extended for more complicated functions.	contrasting
train_20317	One option would be to use the output from the system with the best performance on some development set.	it was found that this approach did not always yield better combination output compared to the best single system on all evaluation metrics.	contrasting
train_20318	The phrase-level combination yields fairly good gains on the tuning sets.	the performance does not seem to generalize to the test sets used in this work.	contrasting
train_20319	There is a large literature on document classification and automated text categorization (Sebastiani, 2002).	that work assumes that the categories of interest are already known, and tries to assign documents to categories.	contrasting
train_20320	However, that work assumes that the categories of interest are already known, and tries to assign documents to categories.	in this paper we focus on the problem of determining the categories of interest.	contrasting
train_20321	In unsupervised learning, where no training takes place, one simply hopes that the unsupervised learner will work well on any unlabeled test collection.	when the variability in the data is large, such hope may be unrealistic; a tuning of the unsupervised algorithm may then be necessary in order to perform well on new test collections.	contrasting
train_20322	9 One might obtain better performance (across all methods being compared) by choosing a separate θ * for each of the 9 training sets.	to simulate real limited-data training conditions, one should then find the θ * for each {i, ..., j} using a separate cross-validation within {i, ..., j} only; this would slow down the experiments considerably.	contrasting
train_20323	Side information is sometimes encoded as "virtual examples" like our contrast examples or pseudoexamples.	past work generates these by automatically transforming the training examples in ways that are expected to preserve or alter the classification (Abu-Mostafa, 1995).	contrasting
train_20324	(2005) asked annotators to identify globally "relevant" features.	our approach does not force the annotator to evaluate the importance of features individually, nor in a global context outside any specific document, nor even to know the learner's feature space.	contrasting
train_20325	Our previous results indicated that Concept Repetition was the best feature to add to the Baseline 2 state-space model, but also that Percent Correctness and Frustration (when added to Baseline 2) offered an improvement over the Baseline MDP's.	these conclusions were based on a very qualitative approach for determining if a policy is reliable or not.	contrasting
train_20326	For example, a model can have an optimal policy with a very high ECR value, but have very wide confidence bounds reaching even into negative rewards.	another model can have a relatively lower ECR but if its bounds are tighter (and the lower bound is not negative), one can know that that policy is less affected by poor parameter estimates stemming from data sparsity issues.	contrasting
train_20327	The Concept Repetition model (which is composed of the Concept Repetition, Certainty and Correctness features) was especially useful, as it led to a policy that outperformed the Baseline 1 model, even when we take into account the uncertainty in the model estimates caused by data sparsity.	for the Baseline 2, Percent Correctness, and Frustration models, the estimates for the expected cumulative reward are much less reliable, and no conclusion can be reliably drawn about the usefulness of these features.	contrasting
train_20328	Our initial findings indicate that, as more data becomes available the bounds tighten for most parameters in the transition matrix.	for some of the parameters the bounds can remain wide, and that is enough to keep the confidence interval for the expected cumulative reward from converging.	contrasting
train_20329	We also tried learning the priming weight with an EM algorithm.	we found out that the learned priming weight performed worse than the empirical one in our experiments.	contrasting
train_20330	In late application, the S-Bigram1 model performed better than the trigram model (t = 2.11, p < 0.02, one-tailed), so did the S-Bigram2 model (t = 1.99, p < 0.025, onetailed).	compared to the bigram model, the S-Bigram1 model did not change the recognition performance significantly (t = 0.38, N.S., two-tailed) in late application, neither did the S-Bigram2 model (t = 0.50, N.S., two-tailed).	contrasting
train_20331	In the bigram case, the path "<s> a tree" has a higher language score (summation of bigram probabilities along the path) than "<s> sheet", and "a floor" has a higher language score than "a full".	these correct paths "<s> a tree" and "a floor" (not exactly correct, but better than "a full") do not appear in the best hypothesis in the resulting n-best list.	contrasting
train_20332	In this example, the user was talking about a picture entity picture bamboo.	this entity was not salient, only entities bedroom and chandelier 1 were salient.	contrasting
train_20333	The four methods above are based on context information.	our method exploits the internal structure of the semantic orientations of phrases.	contrasting
train_20334	Since we cannot expect a large labeled dataset to be available for each adjective, we use not the approximated leave-one-out error rate, but the magnetization-like criterion.	the magnetization above is defined for the Ising model.	contrasting
train_20335	There are still 3320 (=12066-8746) word pairs which could not be classified, because there are no entries for those words in the dictionary.	the main cause of this problem is word segmenta-tion, since many compound nouns and exceedinglysubdivided morphemes are not in dictionaries.	contrasting
train_20336	As Table 2 shows, we outperform the PRANK baseline in both cases.	on the consensus instances we achieve a relative reduction in error of 21.8% compared to only a 1.1% reduction for the other set.	contrasting
train_20337	The agreement model, when considered on its own, achieves classification accuracy of 67% on the test set, compared to a majority baseline of 58%.	those instances with high confidence \|a • x\| exhibit substantially higher classification accuracy.	contrasting
train_20338	Each classifier in Section 3.3 is also used to learn a sentence classifier.	we then must make a document-level prediction.	contrasting
train_20339	In fact, the proposed method outperforms simple WordNet based approaches such as Edge-Counting and Information Content measures.	considering the high correlation between human subjects (0.9), there is still room for improvement.	contrasting
train_20340	In word sense disambiguation, choosing the most frequent sense for an ambiguous word is a powerful heuristic.	its usefulness is restricted by the availability of sense-annotated data.	contrasting
train_20341	Nonetheless, the task is still difficult, even for human judges, as we will see in Section 6.4.	because the solution allows only one correct answer the accuracies are underestimated.	contrasting
train_20342	Transliteration between languages which share similar alphabets and sound systems is usually not difficult, because the majority of letters remain the same.	the task is significantly more difficult when the language pairs are considerably different, for example, English-Arabic, English-Chinese, and English-Japanese.	contrasting
train_20343	The second solution is to add a null letter in the grapheme sequence.	the null letter not only confuses the phoneme prediction model, but also complicates the the phoneme generation phase.	contrasting
train_20344	Arabic words consist of a stem surrounded by prefixes and suffixes, which are fairly successfully segmented out by Morfessor.	arabic also has templatic morphology, i.e., the stem is formed through the insertion of a vowel pattern into a "consonantal skeleton".	contrasting
train_20345	The approach is more suitable for dependency parsing since trees do not have non-terminal nodes, therefore revisions do not require adding/removing nodes.	the method applies to any parser since it only analyzes output trees.	contrasting
train_20346	Similarly to re-ranking our method aims at improving the accuracy of the base parser with an additional learner.	as in transformationbased learning, it avoids generating multiple parses and applies revisions to arcs in the tree which it considers incorrect.	contrasting
train_20347	(2006) that a hierarchically split PCFG could exceed the accuracy of lexicalized PCFGs (Collins, 1999;Charniak and Johnson, 2005).	many questions about inference with such split PCFGs remain open.	contrasting
train_20348	Adapting unlexicalized parsers appears to be equally difficult: Levy and Manning (2003) adapt the unlexicalized parser of Klein and Manning (2003) to Chinese, but even after significant efforts on choosing category splits, only modest performance gains are reported.	automatically learned grammars like the one of Matsuzaki et al.	contrasting
train_20349	On the one hand, segmentation is done automatically, so it is realistic that given a "real world" document, we can compute segment boundaries to help our classification judgments.	testing on unsegmented input allows us to compare more directly to the numbers from our previous section.	contrasting
train_20350	On the PDTB test data, performance using the segment-trained classifiers improves in two of three cases, with a mean improvement of 1.2%.	because of the small size of this set, this margin is not statistically significant.	contrasting
train_20351	Whether or not we use topic segmentation to constrain our training instances, our patterns rely on sentence boundaries and cue phrase anchors to demarcate the extents of the text spans which form our RSR instances.	an instance which matches such a pattern often contains some amount of text which is not relevant to the relation in question.	contrasting
train_20352	as "a drop in oil prices" and "weakness in the automotive sector."	the output of our instance mining process simply splits the string around the cue phrase "because of" and extracts the entire first and second parts of the sentence as the constituents.	contrasting
train_20353	We evaluate the impact of our syntactic heuristics on classification over the Auto and PDTB test sets using the same instance set of 400,000 training instances per relation.	each applies different filters to the instances I before computing the frequencies F (all other parameters use the same values; these are set slightly differently than the optimized values discussed earlier because of the smaller training sets).	contrasting
train_20354	Even with this reduced training set, the syntactic heuristic improves performance in two out of three cases on the PDTB test set, including a 2.7 percent improvement for the Cause vs NoRel classifier.	due to the small size of the PDTB test set, none of these differences is statistically significant.	contrasting
train_20355	Some local models also have trouble deciding which of a pair of related sentences ought to come first.	the global HMM model of Barzilay and Lee (2004) tries to track the predictable changes in topic between sentences.	contrasting
train_20356	The AIRPLANE documents have some advantages for coherence research: they are short (11.5 sentences on average) and quite formulaic, which makes it easy to find lexical and structural patterns.	they do have some oddities.	contrasting
train_20357	Domains with more natural writing styles will make lexical prediction a much more difficult problem.	the wider variety of grammatical constructions used may motivate more complex syntactic features, for instance as proposed by (Siddharthan et al., 2004) in sentence clustering.	contrasting
train_20358	Being heuristic in nature, this algorithm is not guaranteed to find an optimal solution.	its simplicity and time efficiency make it a decoding algorithm of choice for a wide range of NLP applications.	contrasting
train_20359	We constrain demand variables d v by: The right hand side sums the number of selected edges entering v, and will therefore be either 0 or 1.	next we add variables a v and b v constrained by the equations: note that a v sums over f values on all edges entering v. by the previous constraints those f -values can only be nonzero on the (at most one) selected edge entering v. So, a v is simply the fvalue on the selected edge entering v, if one exists, and 0 otherwise.	contrasting
train_20360	A static linear interpolation of predictions using the two approaches led to only minimal reductions of prediction error, likely because predictions from the poorer performing grammar-based classifier were always given the same weight.	with the confidence measures, predictions from the grammar-based classifier could be given more weight when the confidence measure was high, and less weight when the measure was low and the predictions were likely to be inaccurate.	contrasting
train_20361	The trends were similar for both sets of grammatical features.	the first set of features that included complex syntactic constructs led to better performance than the second set, which included only verb tenses, part of speech labels, and sentence length.	contrasting
train_20362	ated reference translation to explicitly bias the language model.	one has to take care not to bias towards the correct response so strongly that the student is allowed to make mistakes with impunity.	contrasting
train_20363	The transcriber was instructed to transcribe spontaneous speech effects, such as false starts and filled pauses.	tonal mispronunciations are completely ignored, and segmental errors are largely ignored to the extent that they do not result in a different syllable.	contrasting
train_20364	The use of the Europarl corpus is not an accurate simulation of the intended situation because it enables us to use a relatively large parallel corpus for direct training.	it is necessary to evaluate the performance of the pivot strategies against that of Direct SMT systems under controlled experiments in order to determine how much the pivot strategies can be improved.	contrasting
train_20365	of phrases in PhraseTrans system Recall was reasonably high.	the upper bound of recall was 100 percent because we used a multilingual corpus whose sentences were aligned to each other across all four languages, as described in Section 4.1.	contrasting
train_20366	They showed that this direct training method is superior to the sentence translation strategy (SntTrans1) in translating Catalan into English but is inferior to it in the opposite translation direction (in terms of the BLEU score).	we have shown that the phrase translation strategy consistently outperformed the sentence translation strategy in the controlled experiments.	contrasting
train_20367	MT users typically try to reduce the post-editing load by customizing their MT systems.	in Rule-based Machine Translation (RBMT), which still constitutes the bulk of the current commercial offering, customization is usually restricted to the development of "user dictionaries".	contrasting
train_20368	For how-questions, this corresponds to a numerical type.	retrieving all numerical entities will provide lower answer identification precision than a system that only provides those specified with the expected answer units.	contrasting
train_20369	The queryspecific oracle is able to achieve the highest performance because of perfect knowledge of the units appropriate to a given question.	its precision is only 42.2%.	contrasting
train_20370	Its precision runs at just under 5%, about a quarter of the lowest precision of any of our unit-list approaches at any recall value.	fixed-category typing does achieve high recall, roughly 96%, missing only numerical entities unrecognized by Minipar.	contrasting
train_20371	In response to questions about particular events of interest that can be enumerated in advance, it is possible to perform a deeper semantic analysis focusing on the entities, relations, and sub-events of interest.	the deeper analysis may be errorful and will also not always provide complete coverage of the information relevant to the query.	contrasting
train_20372	The collected Discovery video is a small but "long" OCR document set, which results the estimation of DF value unreliable.	a passage is more coherent than a long document, thus we replace the DF estimation with PF score.	contrasting
train_20373	This approach works surprisingly well, mainly because an explicit effort was made to use arguments Arg0 and Arg1 consistently across different verbs; and because those two argument labels account for 85% of all arguments.	this approach causes the system to conflate different argument types, especially with the highly overloaded arguments Arg2-Arg5.	contrasting
train_20374	For example, the Arg1 (underlined) in "John broke the window" is the same window that is annotated as the Arg1 in "The window broke", even though it is the syntactic subject in one sentence and the syntactic object in the other.	there is no guarantee that an argument label will be used consistently across different verbs.	contrasting
train_20375	Generally, the arguments are simply listed in the order of their prominence for each verb.	an explicit effort was made when PropBank was created to use Arg0 for arguments that fulfill Dowty's criteria for "prototypical agent," and Arg1 for arguments that fulfill the criteria for "prototypical patient."	contrasting
train_20376	From this figure, we can see that Arg0 maps to agent-like roles, such as "agent" and "experiencer," over 94% of the time; and Arg1 maps to patientlike roles, including "theme," "topic," and "patient," over 82% of the time.	arguments Arg2-5 get mapped to a much broader variety of roles.	contrasting
train_20377	For a system trained on WSJ, there is a fairly small drop in performance of the ID task when tested on Brown vs tested on WSJ (Fscore 92.7 vs 95.3).	in this same condition, the Classification task has a very large drop in performance (72.0% vs 86.1%).	contrasting
train_20378	We investigated this technique but discarded it due to subtle yet critical issues with pattern canonicalization that resulted in rejecting nearly all inferences.	we are investigating other ways of using Web corpora for this task.	contrasting
train_20379	There were 225 valid abbreviations already identified by the expertise taxonomy team.	we found many abbreviations that appeared in the skill statements but were not listed there.	contrasting
train_20380	The assumption for this shortest path approach is that the links in the taxonomy represent uniform distances.	in most taxonomies, sibling concepts deep in the taxonomy are usually more closely related than those higher up.	contrasting
train_20381	Since the system strategies are evaluated and adapted based on the analysis of these simulated dialog behaviors, we would expect that these behaviors are what we are going to see in the test phase when the systems interact with human users.	for automatically learning dialog strategies, it is not clear how realistic versus how exploratory (Singh et al., 2002) the training corpus should be.	contrasting
train_20382	EM1 measures exactly what RF1 tries to optimize; while EM2 measures exactly what RF2 tries to optimize.	we show the results evaluated by both EM1 and EM2 for all configurations since the two evaluation measures have their own practical values and can be deployed under different design requirements.	contrasting
train_20383	It does explore some states that are at the end of the paths, but since these are the infrequent states in the test phase, exploring these states does not actually improve the model's performance much.	while the student correctness rate in the real corpus is 60%, the RRM prevents itself from being trapped in the less-likely states on incorrect answer paths by keeping its correctness rate to be 50%.	contrasting
train_20384	To find the k-best matches for a particular substring s, we do what we would normally do for standard suffix arrays on lexicographic splits.	on popularity splits, we search the more popular half first and then we search the less popular half, if necessary.	contrasting
train_20385	In order to make the method computationally feasible, potential cognates were filtered based on word order, location in the document, frequency, and length.	we found that a faster and simpler procedure, which is described below, performed extremely well, eliminating the need for a more sophisticated approach.	contrasting
train_20386	Students then hear a segment of a lecture concerning the same topic.	the lecture contradicts and complements the information contained in the reading.	contrasting
train_20387	Vergyri and Kirchhoff (2004) follow an approach similar to ours in that they choose from the diacritizations proposed by BAMA.	they train a single tagger using unannotated data and EM, which necessarily leads to a lower performance.	contrasting
train_20388	It is possible to construct a combined system which uses a lexicon, but backs off to a Zitouni-style system for unknown words.	a large portion of the unknown words are in fact foreign words and names, and it is not clear whether the models learned handle such words well.	contrasting
train_20389	The bound is the same as for the reported algorithms for checking wellnestedness (Möhl, 2006).	the following theorem allows wellnestedness checking to be linear for projective trees, to be faster for random input, and to remain O(n 2 ).	contrasting
train_20390	These systems use supervised learning methods which only utilize annotated NL sentences.	it requires considerable human effort to annotate sentences.	contrasting
train_20391	However, it requires considerable human effort to annotate sentences.	unannotated NL sentences are usually easily available.	contrasting
train_20392	Recently some alignment evaluation metrics have been proposed which are more informative when the alignments are used to extract translation units (Fraser and Marcu, 2006;Ayan and Dorr, 2006).	these metrics assess translation quality very indirectly.	contrasting
train_20393	For the classification task, training with MT data is less effective than with non-native data.	for the ranking task, models trained on publicly available MT data generalize well, performing as well as those trained with a non-native corpus of size 10000.	contrasting
train_20394	Finally, we fuse the extracted binary relations into a ternary relation.	with our discriminative model, a statistical parsing based generative model (Shi et al., 2007) has been proposed for a related task on this data set where the NEs and their relations are extracted together and used to identify which NEs are relevant in a particular sentence.	contrasting
train_20395	In this paper we explored the use of rich syntactic features for the relation extraction task.	with the previously used set of syntactic features for this task, we use a large number of features originally proposed for the Semantic Role Labeling task.	contrasting
train_20396	To our knowledge, no research has yet been conducted on soundbite speaker name identification in Mandarin BN domain.	this work is related to some extent to speaker role identification, speaker diarization, and named entity recognition.	contrasting
train_20397	IceTagger makes 11.5% less errors than TnT.	when tag profile gaps in the lexicon, used by TnT, are filled with tags produced by IceMorphy, our morphological analyser, TnT's tagging accuracy increases to 91.18%.	contrasting
train_20398	New assignments of part-ofspeech (POS) to words is found by optimising the product of lexical probabilities (p(w i \|t j )) and contextual probabilities (p(t i \|t i−1 , t i−2 )) (where w i and t i are the i th word and tag, respectively).	to DDMs, LRBMs are developed with the purpose of tagging a specific language using a particular tagset.	contrasting
train_20399	The average tagging accuracy of Ice-Tagger is 91.54%, compared to 90.44% obtained by the TnT tagger.	we were able to obtain 91.18% accuracy using TnT along with the tag profile gap filling mechanism of IceMorphy.	contrasting

train_20300

Second, if we compare the best performance achieved in each feature subspace, we can see that for both classifiers, syntactic parse tree is the most effective feature space, while sequence and dependency tree are similar.

the difference in performance between the syntactic parse tree subspace and the other two subspaces is not very large.

contrasting

train_20301

The standard methods for inferring the parameters of probabilistic models in computational linguistics are based on the principle of maximum-likelihood estimation; for example, the parameters of Probabilistic Context-Free Grammars (PCFGs) are typically estimated from strings of terminals using the Inside-Outside (IO) algorithm, an instance of the Expectation Maximization (EM) procedure (Lari and Young, 1990).

much recent work in machine learning and statistics has turned away from maximum-likelihood in favor of Bayesian methods, and there is increasing interest in Bayesian methods in computational linguistics as well (Finkel et al., 2006).

contrasting

train_20302

We also experimented with a version of IO modified to perform Bayesian MAP estimation, where the Maximization step of the IO procedure is replaced with Bayesian inference using a Dirichlet prior, i.e., where the rule probabilities θ (k) at iteration k are estimated using: Clearly such an approach is very closely related to the Bayesian procedures presented in this article, and in some circumstances this may be a useful estimator.

in our experiments with the Sesotho data above we found that for the small values of α necessary to obtain a sparse solution,the expected rule count E[f r ] for many rules r was less than 1 − α.

contrasting

train_20303

Thus The spans counted by ℓ 1 and ℓ 2 correspond to continuous spans counted in computing the complexity of a chart parsing operation.

unlike in the diagrams in the earlier part of this paper, in our graph theoretic argument there is no requirement that T select only corresponding pairs of vertices from V 1 and V 2 .

contrasting

train_20304

Schone and Jurafsky's (2001) work represents one of the few attempts to address this inappropriate morpheme attachment problem, introducing a method that exploits the semantic relatedness between word pairs.

we propose two arguably simpler, yet effective techniques that rely on relative corpus frequency and suffix level similarity to solve the problem.

contrasting

train_20305

The random mode serves as an absolute baseline in this respect: it indicates how well a particular base generator performs on its own.

different base generators have different effects on the generation modes.

contrasting

train_20306

a statistical model that translates MRL to NL.

there has been little, if any, research on exploiting recent SMT methods for NLG.

contrasting

train_20307

players 2 3 7 5, or not learn any phrases for those predicates at all.

only 4.47% of the phrases generated during testing for GEOQUERY were discontiguous, so the advantage of WASP −1 ++ over PHARAOH++ was not as obvious.

contrasting

train_20308

The three next sentences display some advantages of our approach over the K&M model: here, the latter model performs deletion with too little lexicosyntactic information, and accidentally removes certain modifiers that are sometimes, but not always, good candidates for deletions (e.g., ADJP in Sentence 2, PP in sentences 3 and 4).

our model keeps these constituent intact.

contrasting

train_20309

Research on WSD has concentrated on using contextual information, which may be limited for infrequent unknown words.

chinese characters carry semantic information that is useful for predicting the semantic properties of the words containing them.

contrasting

train_20310

Rule-based models have not been used for this task before.

there are some regularities in the relationship between the semantic categories of unknown words and those of their component characters that can be captured in a more direct and effective way by linguistic rules than by statistical models.

contrasting

train_20311

Variant 2 performs better when only the best guess is considered, indicating that it is useful to model the effect of position on a character's contribution to word meaning in this case.

it is not helpful to be sensitive to character position when the best five guesses are considered.

contrasting

train_20312

These outputs could simply be replaced with the perfect SVM predictions (1 for the true class, -1 elsewhere) since the labels are known.

the second-pass learner might actually benefit from the information contained in the misclassifications.

contrasting

train_20313

The success of factoid QA systems, particularly in the NIST-sponsored TREC evaluations (Voorhees, 2003), has reinforced the perception about the superiority of QA systems over traditional IR engines.

is QA really better than IR?

contrasting

train_20314

that answers a particular question-and the development of such technology has been encouraged by the setup of the TREC QA tracks.

a study by shows that users actually prefer answers embedded within some sort of context, e.g., the sentence or the paragraph that the answer was found in.

contrasting

train_20315

Previous work (Shapiro and Taksa, 2003) in the web environment attempted to convert a user's natural language query into one suited for use with web search engines.

the focus was on merging the results from using different sub-queries, and not selection of a single sub-query.

contrasting

train_20316

The modified Powell's method has been previously used in optimizing the weights of a standard feature-based MT decoder in (Och, 2003) where a more efficient algorithm for log-linear models was proposed.

this is specific to log-linear models and cannot be easily extended for more complicated functions.

contrasting

train_20317

One option would be to use the output from the system with the best performance on some development set.

it was found that this approach did not always yield better combination output compared to the best single system on all evaluation metrics.

contrasting

train_20318

The phrase-level combination yields fairly good gains on the tuning sets.

the performance does not seem to generalize to the test sets used in this work.

contrasting

train_20319

There is a large literature on document classification and automated text categorization (Sebastiani, 2002).

that work assumes that the categories of interest are already known, and tries to assign documents to categories.

contrasting

train_20320

However, that work assumes that the categories of interest are already known, and tries to assign documents to categories.

in this paper we focus on the problem of determining the categories of interest.

contrasting

train_20321

In unsupervised learning, where no training takes place, one simply hopes that the unsupervised learner will work well on any unlabeled test collection.

when the variability in the data is large, such hope may be unrealistic; a tuning of the unsupervised algorithm may then be necessary in order to perform well on new test collections.

contrasting

train_20322

9 One might obtain better performance (across all methods being compared) by choosing a separate θ * for each of the 9 training sets.

to simulate real limited-data training conditions, one should then find the θ * for each {i, ..., j} using a separate cross-validation within {i, ..., j} only; this would slow down the experiments considerably.

contrasting

train_20323

Side information is sometimes encoded as "virtual examples" like our contrast examples or pseudoexamples.

past work generates these by automatically transforming the training examples in ways that are expected to preserve or alter the classification (Abu-Mostafa, 1995).

contrasting

train_20324

(2005) asked annotators to identify globally "relevant" features.

our approach does not force the annotator to evaluate the importance of features individually, nor in a global context outside any specific document, nor even to know the learner's feature space.

contrasting

train_20325

Our previous results indicated that Concept Repetition was the best feature to add to the Baseline 2 state-space model, but also that Percent Correctness and Frustration (when added to Baseline 2) offered an improvement over the Baseline MDP's.

these conclusions were based on a very qualitative approach for determining if a policy is reliable or not.

contrasting

train_20326

For example, a model can have an optimal policy with a very high ECR value, but have very wide confidence bounds reaching even into negative rewards.

another model can have a relatively lower ECR but if its bounds are tighter (and the lower bound is not negative), one can know that that policy is less affected by poor parameter estimates stemming from data sparsity issues.

contrasting

train_20327

The Concept Repetition model (which is composed of the Concept Repetition, Certainty and Correctness features) was especially useful, as it led to a policy that outperformed the Baseline 1 model, even when we take into account the uncertainty in the model estimates caused by data sparsity.

for the Baseline 2, Percent Correctness, and Frustration models, the estimates for the expected cumulative reward are much less reliable, and no conclusion can be reliably drawn about the usefulness of these features.

contrasting

train_20328

Our initial findings indicate that, as more data becomes available the bounds tighten for most parameters in the transition matrix.

for some of the parameters the bounds can remain wide, and that is enough to keep the confidence interval for the expected cumulative reward from converging.

contrasting

train_20329

We also tried learning the priming weight with an EM algorithm.

we found out that the learned priming weight performed worse than the empirical one in our experiments.

contrasting

train_20330

In late application, the S-Bigram1 model performed better than the trigram model (t = 2.11, p < 0.02, one-tailed), so did the S-Bigram2 model (t = 1.99, p < 0.025, onetailed).

compared to the bigram model, the S-Bigram1 model did not change the recognition performance significantly (t = 0.38, N.S., two-tailed) in late application, neither did the S-Bigram2 model (t = 0.50, N.S., two-tailed).

contrasting

train_20331

In the bigram case, the path "<s> a tree" has a higher language score (summation of bigram probabilities along the path) than "<s> sheet", and "a floor" has a higher language score than "a full".

these correct paths "<s> a tree" and "a floor" (not exactly correct, but better than "a full") do not appear in the best hypothesis in the resulting n-best list.

contrasting

train_20332

In this example, the user was talking about a picture entity picture bamboo.

this entity was not salient, only entities bedroom and chandelier 1 were salient.

contrasting

train_20333

The four methods above are based on context information.

our method exploits the internal structure of the semantic orientations of phrases.

contrasting

train_20334

Since we cannot expect a large labeled dataset to be available for each adjective, we use not the approximated leave-one-out error rate, but the magnetization-like criterion.

the magnetization above is defined for the Ising model.

contrasting

train_20335

There are still 3320 (=12066-8746) word pairs which could not be classified, because there are no entries for those words in the dictionary.

the main cause of this problem is word segmenta-tion, since many compound nouns and exceedinglysubdivided morphemes are not in dictionaries.

contrasting

train_20336

As Table 2 shows, we outperform the PRANK baseline in both cases.

on the consensus instances we achieve a relative reduction in error of 21.8% compared to only a 1.1% reduction for the other set.

contrasting

train_20337

The agreement model, when considered on its own, achieves classification accuracy of 67% on the test set, compared to a majority baseline of 58%.

those instances with high confidence |a • x| exhibit substantially higher classification accuracy.

contrasting

train_20338

Each classifier in Section 3.3 is also used to learn a sentence classifier.

we then must make a document-level prediction.

contrasting

train_20339

In fact, the proposed method outperforms simple WordNet based approaches such as Edge-Counting and Information Content measures.

considering the high correlation between human subjects (0.9), there is still room for improvement.

contrasting

train_20340

In word sense disambiguation, choosing the most frequent sense for an ambiguous word is a powerful heuristic.

its usefulness is restricted by the availability of sense-annotated data.

contrasting

train_20341

Nonetheless, the task is still difficult, even for human judges, as we will see in Section 6.4.

because the solution allows only one correct answer the accuracies are underestimated.

contrasting

train_20342

Transliteration between languages which share similar alphabets and sound systems is usually not difficult, because the majority of letters remain the same.

the task is significantly more difficult when the language pairs are considerably different, for example, English-Arabic, English-Chinese, and English-Japanese.

contrasting

train_20343

The second solution is to add a null letter in the grapheme sequence.

the null letter not only confuses the phoneme prediction model, but also complicates the the phoneme generation phase.

contrasting

train_20344

Arabic words consist of a stem surrounded by prefixes and suffixes, which are fairly successfully segmented out by Morfessor.

arabic also has templatic morphology, i.e., the stem is formed through the insertion of a vowel pattern into a "consonantal skeleton".

contrasting

train_20345

The approach is more suitable for dependency parsing since trees do not have non-terminal nodes, therefore revisions do not require adding/removing nodes.

the method applies to any parser since it only analyzes output trees.

contrasting

train_20346

Similarly to re-ranking our method aims at improving the accuracy of the base parser with an additional learner.

as in transformationbased learning, it avoids generating multiple parses and applies revisions to arcs in the tree which it considers incorrect.

contrasting

train_20347

(2006) that a hierarchically split PCFG could exceed the accuracy of lexicalized PCFGs (Collins, 1999;Charniak and Johnson, 2005).

many questions about inference with such split PCFGs remain open.

contrasting

train_20348

Adapting unlexicalized parsers appears to be equally difficult: Levy and Manning (2003) adapt the unlexicalized parser of Klein and Manning (2003) to Chinese, but even after significant efforts on choosing category splits, only modest performance gains are reported.

automatically learned grammars like the one of Matsuzaki et al.

contrasting

train_20349

On the one hand, segmentation is done automatically, so it is realistic that given a "real world" document, we can compute segment boundaries to help our classification judgments.

testing on unsegmented input allows us to compare more directly to the numbers from our previous section.

contrasting

train_20350

On the PDTB test data, performance using the segment-trained classifiers improves in two of three cases, with a mean improvement of 1.2%.

because of the small size of this set, this margin is not statistically significant.

contrasting

train_20351

Whether or not we use topic segmentation to constrain our training instances, our patterns rely on sentence boundaries and cue phrase anchors to demarcate the extents of the text spans which form our RSR instances.

an instance which matches such a pattern often contains some amount of text which is not relevant to the relation in question.

contrasting

train_20352

as "a drop in oil prices" and "weakness in the automotive sector."

the output of our instance mining process simply splits the string around the cue phrase "because of" and extracts the entire first and second parts of the sentence as the constituents.

contrasting

train_20353

We evaluate the impact of our syntactic heuristics on classification over the Auto and PDTB test sets using the same instance set of 400,000 training instances per relation.

each applies different filters to the instances I before computing the frequencies F (all other parameters use the same values; these are set slightly differently than the optimized values discussed earlier because of the smaller training sets).

contrasting

train_20354

Even with this reduced training set, the syntactic heuristic improves performance in two out of three cases on the PDTB test set, including a 2.7 percent improvement for the Cause vs NoRel classifier.

due to the small size of the PDTB test set, none of these differences is statistically significant.

contrasting

train_20355

Some local models also have trouble deciding which of a pair of related sentences ought to come first.

the global HMM model of Barzilay and Lee (2004) tries to track the predictable changes in topic between sentences.

contrasting

train_20356

The AIRPLANE documents have some advantages for coherence research: they are short (11.5 sentences on average) and quite formulaic, which makes it easy to find lexical and structural patterns.

they do have some oddities.

contrasting

train_20357

Domains with more natural writing styles will make lexical prediction a much more difficult problem.

the wider variety of grammatical constructions used may motivate more complex syntactic features, for instance as proposed by (Siddharthan et al., 2004) in sentence clustering.

contrasting

train_20358

Being heuristic in nature, this algorithm is not guaranteed to find an optimal solution.

its simplicity and time efficiency make it a decoding algorithm of choice for a wide range of NLP applications.

contrasting

train_20359

We constrain demand variables d v by: The right hand side sums the number of selected edges entering v, and will therefore be either 0 or 1.

next we add variables a v and b v constrained by the equations: note that a v sums over f values on all edges entering v. by the previous constraints those f -values can only be nonzero on the (at most one) selected edge entering v. So, a v is simply the fvalue on the selected edge entering v, if one exists, and 0 otherwise.

contrasting

train_20360

A static linear interpolation of predictions using the two approaches led to only minimal reductions of prediction error, likely because predictions from the poorer performing grammar-based classifier were always given the same weight.

with the confidence measures, predictions from the grammar-based classifier could be given more weight when the confidence measure was high, and less weight when the measure was low and the predictions were likely to be inaccurate.

contrasting

train_20361

The trends were similar for both sets of grammatical features.

the first set of features that included complex syntactic constructs led to better performance than the second set, which included only verb tenses, part of speech labels, and sentence length.

contrasting

train_20362

ated reference translation to explicitly bias the language model.

one has to take care not to bias towards the correct response so strongly that the student is allowed to make mistakes with impunity.

contrasting

train_20363

The transcriber was instructed to transcribe spontaneous speech effects, such as false starts and filled pauses.

tonal mispronunciations are completely ignored, and segmental errors are largely ignored to the extent that they do not result in a different syllable.

contrasting

train_20364

The use of the Europarl corpus is not an accurate simulation of the intended situation because it enables us to use a relatively large parallel corpus for direct training.

it is necessary to evaluate the performance of the pivot strategies against that of Direct SMT systems under controlled experiments in order to determine how much the pivot strategies can be improved.

contrasting

train_20365

of phrases in PhraseTrans system Recall was reasonably high.

the upper bound of recall was 100 percent because we used a multilingual corpus whose sentences were aligned to each other across all four languages, as described in Section 4.1.

contrasting

train_20366

They showed that this direct training method is superior to the sentence translation strategy (SntTrans1) in translating Catalan into English but is inferior to it in the opposite translation direction (in terms of the BLEU score).

we have shown that the phrase translation strategy consistently outperformed the sentence translation strategy in the controlled experiments.

contrasting

train_20367

MT users typically try to reduce the post-editing load by customizing their MT systems.

in Rule-based Machine Translation (RBMT), which still constitutes the bulk of the current commercial offering, customization is usually restricted to the development of "user dictionaries".

contrasting

train_20368

For how-questions, this corresponds to a numerical type.

retrieving all numerical entities will provide lower answer identification precision than a system that only provides those specified with the expected answer units.

contrasting

train_20369

The queryspecific oracle is able to achieve the highest performance because of perfect knowledge of the units appropriate to a given question.

its precision is only 42.2%.

contrasting

train_20370

Its precision runs at just under 5%, about a quarter of the lowest precision of any of our unit-list approaches at any recall value.

fixed-category typing does achieve high recall, roughly 96%, missing only numerical entities unrecognized by Minipar.

contrasting

train_20371

In response to questions about particular events of interest that can be enumerated in advance, it is possible to perform a deeper semantic analysis focusing on the entities, relations, and sub-events of interest.

the deeper analysis may be errorful and will also not always provide complete coverage of the information relevant to the query.

contrasting

train_20372

The collected Discovery video is a small but "long" OCR document set, which results the estimation of DF value unreliable.

a passage is more coherent than a long document, thus we replace the DF estimation with PF score.

contrasting

train_20373

This approach works surprisingly well, mainly because an explicit effort was made to use arguments Arg0 and Arg1 consistently across different verbs; and because those two argument labels account for 85% of all arguments.

this approach causes the system to conflate different argument types, especially with the highly overloaded arguments Arg2-Arg5.

contrasting

train_20374

For example, the Arg1 (underlined) in "John broke the window" is the same window that is annotated as the Arg1 in "The window broke", even though it is the syntactic subject in one sentence and the syntactic object in the other.

there is no guarantee that an argument label will be used consistently across different verbs.

contrasting

train_20375

Generally, the arguments are simply listed in the order of their prominence for each verb.

an explicit effort was made when PropBank was created to use Arg0 for arguments that fulfill Dowty's criteria for "prototypical agent," and Arg1 for arguments that fulfill the criteria for "prototypical patient."

contrasting

train_20376

From this figure, we can see that Arg0 maps to agent-like roles, such as "agent" and "experiencer," over 94% of the time; and Arg1 maps to patientlike roles, including "theme," "topic," and "patient," over 82% of the time.

arguments Arg2-5 get mapped to a much broader variety of roles.

contrasting

train_20377

For a system trained on WSJ, there is a fairly small drop in performance of the ID task when tested on Brown vs tested on WSJ (Fscore 92.7 vs 95.3).

in this same condition, the Classification task has a very large drop in performance (72.0% vs 86.1%).

contrasting

train_20378

We investigated this technique but discarded it due to subtle yet critical issues with pattern canonicalization that resulted in rejecting nearly all inferences.

we are investigating other ways of using Web corpora for this task.

contrasting

train_20379

There were 225 valid abbreviations already identified by the expertise taxonomy team.

we found many abbreviations that appeared in the skill statements but were not listed there.

contrasting

train_20380

The assumption for this shortest path approach is that the links in the taxonomy represent uniform distances.

in most taxonomies, sibling concepts deep in the taxonomy are usually more closely related than those higher up.

contrasting

train_20381

Since the system strategies are evaluated and adapted based on the analysis of these simulated dialog behaviors, we would expect that these behaviors are what we are going to see in the test phase when the systems interact with human users.

for automatically learning dialog strategies, it is not clear how realistic versus how exploratory (Singh et al., 2002) the training corpus should be.

contrasting

train_20382

EM1 measures exactly what RF1 tries to optimize; while EM2 measures exactly what RF2 tries to optimize.

we show the results evaluated by both EM1 and EM2 for all configurations since the two evaluation measures have their own practical values and can be deployed under different design requirements.

contrasting

train_20383

It does explore some states that are at the end of the paths, but since these are the infrequent states in the test phase, exploring these states does not actually improve the model's performance much.

while the student correctness rate in the real corpus is 60%, the RRM prevents itself from being trapped in the less-likely states on incorrect answer paths by keeping its correctness rate to be 50%.

contrasting

train_20384

To find the k-best matches for a particular substring s, we do what we would normally do for standard suffix arrays on lexicographic splits.

on popularity splits, we search the more popular half first and then we search the less popular half, if necessary.

contrasting

train_20385

In order to make the method computationally feasible, potential cognates were filtered based on word order, location in the document, frequency, and length.

we found that a faster and simpler procedure, which is described below, performed extremely well, eliminating the need for a more sophisticated approach.

contrasting

train_20386

Students then hear a segment of a lecture concerning the same topic.

the lecture contradicts and complements the information contained in the reading.

contrasting

train_20387

Vergyri and Kirchhoff (2004) follow an approach similar to ours in that they choose from the diacritizations proposed by BAMA.

they train a single tagger using unannotated data and EM, which necessarily leads to a lower performance.

contrasting

train_20388

It is possible to construct a combined system which uses a lexicon, but backs off to a Zitouni-style system for unknown words.

a large portion of the unknown words are in fact foreign words and names, and it is not clear whether the models learned handle such words well.

contrasting

train_20389

The bound is the same as for the reported algorithms for checking wellnestedness (Möhl, 2006).

the following theorem allows wellnestedness checking to be linear for projective trees, to be faster for random input, and to remain O(n 2 ).

contrasting

train_20390

These systems use supervised learning methods which only utilize annotated NL sentences.

it requires considerable human effort to annotate sentences.

contrasting

train_20391

However, it requires considerable human effort to annotate sentences.

unannotated NL sentences are usually easily available.

contrasting

train_20392

Recently some alignment evaluation metrics have been proposed which are more informative when the alignments are used to extract translation units (Fraser and Marcu, 2006;Ayan and Dorr, 2006).

these metrics assess translation quality very indirectly.

contrasting

train_20393

For the classification task, training with MT data is less effective than with non-native data.

for the ranking task, models trained on publicly available MT data generalize well, performing as well as those trained with a non-native corpus of size 10000.

contrasting

train_20394

Finally, we fuse the extracted binary relations into a ternary relation.

with our discriminative model, a statistical parsing based generative model (Shi et al., 2007) has been proposed for a related task on this data set where the NEs and their relations are extracted together and used to identify which NEs are relevant in a particular sentence.

contrasting

train_20395

In this paper we explored the use of rich syntactic features for the relation extraction task.

with the previously used set of syntactic features for this task, we use a large number of features originally proposed for the Semantic Role Labeling task.

contrasting

train_20396

To our knowledge, no research has yet been conducted on soundbite speaker name identification in Mandarin BN domain.

this work is related to some extent to speaker role identification, speaker diarization, and named entity recognition.

contrasting

train_20397

IceTagger makes 11.5% less errors than TnT.

when tag profile gaps in the lexicon, used by TnT, are filled with tags produced by IceMorphy, our morphological analyser, TnT's tagging accuracy increases to 91.18%.

contrasting

train_20398

New assignments of part-ofspeech (POS) to words is found by optimising the product of lexical probabilities (p(w i |t j )) and contextual probabilities (p(t i |t i−1 , t i−2 )) (where w i and t i are the i th word and tag, respectively).

to DDMs, LRBMs are developed with the purpose of tagging a specific language using a particular tagset.

contrasting

train_20399

The average tagging accuracy of Ice-Tagger is 91.54%, compared to 90.44% obtained by the TnT tagger.

we were able to obtain 91.18% accuracy using TnT along with the tag profile gap filling mechanism of IceMorphy.

contrasting