tasksource/scinli · Datasets at Hugging Face

id stringlengths 7 12	sentence1 stringlengths 6 1.27k	sentence2 stringlengths 6 926	label stringclasses 4 values
train_20100	For the same parser performance level, it selects the fewest number of sentences for a human to check and requires the human to make the least number of corrections.	as we have seen in the earlier experiment, very few sentences in the unlabeled pool satisfy its stringent criteria, so it ran out of data before the parser was trained to convergence.	contrasting
train_20101	Alternatively, the number of documents that contain the pair can also be used.	the nature of the language tests in this work make it impractical to be applied.	contrasting
train_20102	In a sequence will to fight, that trend indicates that will should be a noun rather than a modal verb.	that effect is completely lost in a CMM like (a): P(t will \|will, star t ) prefers the modal tagging, and P(TO\|to, t will ) is roughly 1 regardless of t will .	contrasting
train_20103	to-left model, fight will receive its more common noun tagging by symmetric reasoning.	the bidirectional model (c) discussed in the next section makes both directions available for conditioning at all locations, using replicated models of P(t 0 \|t −1 , t +1 , w 0 ), and will be able to get this example correct.	contrasting
train_20104	Since this model has roughly twice as many tag-tag features, the fact that it outperforms the unidirectional models is not by itself compelling evidence for using bidirectional networks.	it also outperforms model L+L 2 which adds the t 0 , t −2 secondprevious word features instead of next word features, which gives only 96.05% (and R+R 2 gives 95.25%).	contrasting
train_20105	When an information seeker can readily think up a suitable term or linguistic expression to represent the information need, direct searching of text by user-generated terms is faster and more effective than browsing.	when users do not know (or can't remember) the exact expression used in relevant documents, they necessarily struggle to find relevant information in full-text search systems.	contrasting
train_20106	The extracts generated are in size shorter than the original texts.	the number of sentences that E states emit cannot be predetermined.	contrasting
train_20107	In the LNK task, incorrectly flagging two stories as being on the same event is considered a false alarm.	in the NED task, incorrectly flagging two stories as being on the same event will cause a true first story to be missed.	contrasting
train_20108	A LNK system should minimize false alarms by identifying only linked stories, which results in high precision for LNK.	a NED system will minimize false alarms by identifying all stories that are linked, which translates to high recall for LNK.	contrasting
train_20109	A corpus-based approach is able to quickly build a machine translation system for a new domain if a bilingual corpus of that domain is available.	if only a small-sized corpus is available, a low translation quality is obtained.	contrasting
train_20110	These methods evaluate the quality of the translation by measuring the similarity between machine translation results and translations done by humans (called references).	the accuracy increases when multiple references are applied because one source sentence can be translated into multiple target expressions.	contrasting
train_20111	However, no nodes contain only the links 'get', and 'me' in the lower sentence.	focusing on the upper phrase "get me a taxi," it contains four word links that correspond to the lower phrase "get a taxi for me", and they have the same syntactic category.	contrasting
train_20112	The hope is that the same techniques will work for extracting prefixes.	even that will not handle the complex combi-nations of infixes that are possible in agglutinative languages like Turkish or polysynthetic languages like Inuktitut.	contrasting
train_20113	In naturally occurring text, the more frequent sense for the two-sense distinction is reported to occur 92% of the time on average; this result has been found both on the CACM collection and on the WordNet SEMCOR sensetagged corpus (Sanderson & van Rijsbergen, 1999).	the challenge for WSD programs is to work on the harder cases, and the artificially constructed SENSE-VAL1 corpus has more evenly distributed senses (Gaustad, 2001).	contrasting
train_20114	In a manually built corpus, a coreference chain can include pronouns and common nouns that refer to the person.	these forms could not be automatically identified, so coreference chains in our corpus only include noun phrases that contain at least one word from the name.	contrasting
train_20115	In this paper we describe CarmelTC , a novel automatic essay grading approach using a hybrid text classification technique for analyzing essay answers to qualitative physics questions inside the Why2 tutorial dialogue system (VanLehn et al., 2002).	to many previous approaches to automated essay grading (Burstein et al., 1998;Larkey, 1998), our goal is not to assign a letter grade to student essays.	contrasting
train_20116	In general, the performance of existing speech recognition systems, whose designs are predicated on relatively noise-free conditions, degrades rapidly in the presence of a high level of adverse conditions.	a recognizer can provide good performance even in very noisy background conditions if the exact testing condition is used to provide the training material from which the reference patterns of the vocabulary are obtained, which is practically not always the case.	contrasting
train_20117	Thus, a system could map both verbs wipe and remove onto the same action scheme.	the apparently equivalent transformations from (1a) to (1b) and from (2a) to (2b) show otherwise.	contrasting
train_20118	One may say that it is not semantically ambiguous.	simple algorithms such as maximal matching [6,9] and longest matching [6] may not be able to discriminate this kind of ambiguity.	contrasting
train_20119	Within this framework, any type of feature can be used, enabling the system designer to experiment with interesting feature types, rather than worry about specific feature interactions.	in a rule based system, the system designer would have to consider how, for instance, a WordNet (Miller, 1995) derived information for a particular example interacts with a part-of-speech-based information and chunking information.	contrasting
train_20120	Since the documents were selected by subject, one may argue that the task of clustering entities will be much easier if the entities are clearly from different genres.	if this is true, then it may account for about 85% of the entities in the person-x corpus that occur only in one domain subject.	contrasting
train_20121	The form of the salience metric, and the choice of features that factor into it, is governed by our knowledge about the way speech and gesture work.	the penalty function also requires parameters that weigh the importance of each factor.	contrasting
train_20122	This is an uncommon phenomenon, and as such, was penalized highly.	anecdotally it appears that the presence of a disfluency makes this phenomenon more likely.	contrasting
train_20123	Their approach of seeking to maximize this probability is similar to the saliencemaximizing approach that we have described.	instead of using a parametric salience function, they learn a set of conditional probability distributions directly from the data.	contrasting
train_20124	However, scalability of such systems is a bottleneck due to the heavy cost of authoring and maintenance of rule sets and inevitable brittleness due to lack of coverage in the rule sets.	data-driven approaches are robust and the procedure for model building is usually simple.	contrasting
train_20125	Data-driven approaches are robust and provide a simple process of developing applications given the data from the application domain.	the reliance on domain-specific data is also one of the significant bottlenecks of data-driven approaches.	contrasting
train_20126	For a system at ability level 0, the odds increase by another factor of 2.72 to 7.39, giving a probability of .88.	a system with an ability of -3, would have the even odds decrease by a factor of 2.72 to .369, yielding P sq = .27.	contrasting
train_20127	Second, for obvious reasons the raw-score estimates based on the Easy sets are considerably higher than those based on the Hard sets.	table 2 also shows that the standard deviations of the number correct estimates obtained for the Easy sets exceed those of the Hard sets as well (sometimes by over 100%).	contrasting
train_20128	There are many multi-word expressions whose hypernyms are their suffixes, and if some expressions share a common suffix, it is likely to be their hypernym.	if a hypernym candidate appears in a position other than as a suffix of a hyponym candidate, the hypernym candidate is likely to be an erroneous one.	contrasting
train_20129	Sentences are generally demarcated by a major fall (or rise) in f0, lengthening of the final syllable, and following pauses.	the usefulness of prosodic information in sentenceinternal parsing is less clear.	contrasting
train_20130	Simple statistical tests show that there is in fact a significant correlation between the location of opening and closing phrase boundaries and all of the prosodic pseudo-punctuation symbols described above, so there is no doubt that these do convey information about syntactic structure.	adding the prosodic pseudo-punctuation symbols uniformly decreased parsing accuracy relative to input with no prosodic information.	contrasting
train_20131	(2003) consider selection of individual AL methods at run-time.	their AL methods are only ever based on single model approaches.	contrasting
train_20132	For example, the categories of the Model 3 Collins parser distinguish between heads, arguments, and adjuncts and they mark some longdistance dependency paths; these distinctions can guide application-specific postprocessors in extracting important semantic relations.	state-of-the-art parsing systems based on deep grammars mark explicitly and in much more detail a wider variety of syntactic and semantic dependencies and should therefore provide even better support for meaning-sensitive applications.	contrasting
train_20133	Moreover, a surprising variety of problems are attackable with FSTs, from partof-speech tagging to letter-to-sound conversion to name transliteration.	language problems like machine translation break this mold, because they involve massive reordering of symbols, and because the transformation processes seem sensitive to hierarchical tree structure.	contrasting
train_20134	Section 1 informally described the root-to-frontier transducer class R. We saw that R allows, by use of states, finite lookahead and arbitrary rearrangement of nonsibling input subtrees removed by a finite distance.	it is often easier to write rules that explicitly represent such lookahead and movement, relieving the burden on the user to produce the requisite intermediary rules and states.	contrasting
train_20135	In arbitrary document collections, such patterns might be too variable to be easily detected by statistical means.	research has shown that texts from the same domain tend to exhibit high similarity (Wray, 2002).	contrasting
train_20136	We conjecture that this difference in performance stems from the ability of content models to capture global document structure.	the other two algorithms are local, taking into account only the relationships between adjacent word pairs and adjacent sentence pairs, respectively.	contrasting
train_20137	Keller and Lapata's (2003) results suggest that webbased frequencies can be a viable alternative to bigram frequencies obtained from smaller corpora or recreated using smoothing.	they do not demonstrate that realistic NLP tasks can benefit from web counts.	contrasting
train_20138	Phone recognition is known to be less accurate than word recognition.	the second method can only generate phone strings that are substrings of the pronunciations of in-vocabulary word strings.	contrasting
train_20139	We have catalogued a variety of such relationships, and note here that we believe it could prove useful to address semantic interdependencies among SCUS in future work that would involve adding a new annotation layer.	4 in our approach, SCUs are treated as independent annotation values, which has the advantage of affording a rigorous analysis of interannotator reliability (see following section).	contrasting
train_20140	This makes sense in light of the fact that a score is dominated by the higher weight SCUS that appear in a summary.	we wanted to study more precisely at what point scores become independent of the choice of models that populate the pyramid.	contrasting
train_20141	First, an SCU is a set of contributors that are largely similar in meaning, thus SCUs differ from each other in both meaning and weight (number of contributors).	factoids are semi-formal expressions in a FOPL-style semantics, which are compositionally interpreted.	contrasting
train_20142	The model in (6) is simplistic in that the relationships between the features across the clauses are not captured directly.	if two values of these features for the main and subordinate clauses co-occur frequently with a particular marker, then the conditional probability of these features on that marker will approximate the right biases.	contrasting
train_20143	The majority of the main clauses in our data are sentence intitial (80.8%).	there are differences among individual markers.	contrasting
train_20144	Knowing the linear precedence of the two clauses is highly predictive of their type: 80.8% of the main clauses are sentence initial.	this type of positional information is typically not known when fragments are synthesised into a meaningful sentence.	contrasting
train_20145	Given that different Machine Translation (MT) evaluation metrics are useful for capturing different aspects of translation quality, it becomes desirable to create MT systems tuned with respect to each individual criterion.	the maximum likelihood techniques that underlie the decision processes of most current MT systems do not take into account these application specific goals.	contrasting
train_20146	The two hypothesis translations are very similar at the word level and therefore the BLEU score, PER and the WER are identical.	we observe that the sentences differ substantially in their syntactic structure (as seen from Parse-Trees in Figure 3), and to a lesser extent in their word-to-word alignments (Figure 1) to the source sentence.	contrasting
train_20147	In (Och and Weber, 1998;Och et al., 1999), a two-level alignment model was employed to utilize shallow phrase structures: alignment between templates was used to handle phrase reordering, and word alignments within a template were used to handle phrase to phrase translation.	phrase level alignment cannot handle long distance reordering effectively.	contrasting
train_20148	One is the PRank algorithm, a variant of the perceptron algorithm, that uses multiple biases to represent the boundaries between every two consecutive ranks (Crammer and Singer, 2001;Harrington, 2003).	as we will show in section 3.7, the PRank algorithm does not work on the reranking tasks due to the introduction of global ranks.	contrasting
train_20149	In addition, we only want to maintain the order of two candidates if their ranks are far away from each other.	we do not care the order of two translations whose ranks are very close, e.g.	contrasting
train_20150	As a result, we cannot use the PRank algorithm in the reranking task, since there are no global ranks or boundaries for all the samples.	the approach of using pairwise samples does work.	contrasting
train_20151	It achieves a BLEU score of 31.7% on the Baseline, 32.8% on the Best Feature, but only 32.6% on the Top Twenty features.	it is within the range of 95% confidence.	contrasting
train_20152	Much of this research has used databases of speech read by actors or native speakers as training data (often with semantically neutral content) (Oudeyer, 2002;Polzin and Waibel, 1998;Liscombe et al., 2003).	such prototypical emotional speech does not necessarily reflect natural speech (Batliner et al., 2003), such as found in tutoring dialogues.	contrasting
train_20153	Although much past work predicts only two classes (e.g., negative/non-negative) (Batliner et al., 2003;Ang et al., 2002;Lee et al., 2001), our experiments produced the best predictions using our three-way distinction.	to (Lee et al., 2001), our classifications are contextrelative (relative to other turns in the dialogue), and taskrelative (relative to tutoring), because like (Ang et al., 2002), we are interested in detecting emotional changes across our dialogues.	contrasting
train_20154	Comparing the results of these combined (speech+text) feature sets with the speech versus text results in Table 1, we find that for autotext+speech-ident and all +ident feature sets, the combined feature set slightly decreases predictive accuracy when compared to the corresponding text-only feature set.	there is no significant difference between the best results in each table (all-text+speech+ident vs. alltext+ident).	contrasting
train_20155	Comparing these results with the results in Tables 1 and 2, we find that while overall the performance of contextual non-combined feature sets shows a small performance increase over most non-contextual combined or non-combined feature sets, there is again a slight decrease in performance across the best results in each table.	there is no significant difference between these best results (alltext+glob-ident vs. all-text+speech+ident vs. alltext+ident).	contrasting
train_20156	Ideally, by training acoustic models on target non-native speech, one would capture its specific characteristics just as training on native speech does.	collecting amounts of non-native speech that are large enough to fully train speaker-independent models is a hard and often impractical task.	contrasting
train_20157	Yet, this is acceptable given the small amount of training data for the language model and the conversational nature of the speech.	performance degrades significantly for non-native speakers, with a word error rate of 52.0%.	contrasting
train_20158	For example, when referring to bus stops by street intersections, all native speakers in our training set simply used "A and B", hence the word "intersection" was not in the language model.	many non-native speakers used the full expression "the intersection of A and B".	contrasting
train_20159	In fact, this is not only true for non-native speakers and lexical entrainment is often described as a negotiation process between the speakers (Clark and Wilkes-Gibbs, 1986).	while it is possible for limited-domain system designers to establish a set of words and constructions that are widely used among native speakers, the variable nature of the expressions mastered by non-native speakers make adaptation a desirable feature of the system.	contrasting
train_20160	Partial Path -For the argument identification task, path is the most salient feature.	it is also the most data sparse feature.	contrasting
train_20161	On performing the search, we found that the overall performance improvement was not much different than that obtained by resolving overlaps as mentioned earlier.	we found that there was an improvement in the CORE ARGUMENT accuracy on the combined task of identifying and assigning semantic arguments, given hand-corrected parses, whereas the accuracy of the ADJUNCTIVE ARGUMENTS slightly deteriorated.	contrasting
train_20162	Although, there was an increase in F 1 score when the language model probabilities were jointly estimated over all the predicates, this improvement is not statistically significant.	estimating the same using specific predicate lemmas, showed a significant improvement in accuracy.	contrasting
train_20163	Speech form is familiar to humans, and can convey information effectively (Nadamoto et al., 2001;Hayashi et al., 1999).	little electronic information is provided in speech form so far.	contrasting
train_20164	However, little electronic information is provided in speech form so far.	there is a lot of information in text form, and it can be transformed into speech by a speech synthesis.	contrasting
train_20165	There are several studies that compare two corpora which have different styles, for example, written and spoken corpora or British and American English corpora, and try to find expressions unique to either of the styles (Kilgarriff, 2001).	those studies did not deal with paraphrases.	contrasting
train_20166	This paper has focused only on paraphrasing predicates.	there are other kinds of paraphrasing which are necessary in order to paraphrase written language text into spoken language.	contrasting
train_20167	On the one hand, we have three different variants of the single-word based model IBM4.	we have two phrase-based systems, namely the alignment templates and the one described in this work.	contrasting
train_20168	This method yielded word classes that offered more robust count approximations for their member words.	both methods yielded similar results when embedded in the larger system, and so we will report on the results of using Good-Turing so as to remain more directly comparable to Dagan et al.	contrasting
train_20169	The MaxEnt model selected the temerity as the antecedent of its (salience value: 0.30), preferring it to the correct antecedent the endowment (salience value: 0.10).	altaVista found no occurrences of temerity's money or its variants on the web, and thus the unnormalized and normalized counts were 0.	contrasting
train_20170	However, AltaVista found no occurrences of temerity's money or its variants on the web, and thus the unnormalized and normalized counts were 0.	endowment's money and its variants had unnormalized and normalized statistics of 1583 and 1.47 × 10 −3 respectively.	contrasting
train_20171	In the cases in which statistics reinforced a wrong answer, no (reasonable) manipulation of statistical features or filters can rescue the prediction.	for the cases in which statistics could help, their successful use will depend on the existence of a formula that can capture these cases without changing the predictions for examples that the model currently classifies correctly.	contrasting
train_20172	Since pronouns carry little semantics of their own, resolving them depends almost entirely on context.	even though context can be helpful for resolving definite NPs, context can be trumped by the semantics of the nouns themselves.	contrasting
train_20173	For example, sources in a language that is translated to English will consistently use the same terminology, resulting in greater similarity between linked documents with the same native language.	sources from radio broadcasts may be transcribed much less consistently than text sources due to recognition errors, so that the expected similarity of a radio broadcast and a text source is less than that of two text sources.	contrasting
train_20174	We refer to the statistics characterizing story pairs with the same source types as source-pair specific information.	to the source-specific thresholds used by CMU, we normalize the similarity measures based on the sourcepair specific information, simultaneously with combining different similarity measures.	contrasting
train_20175	SVM-based systems, such as that described in (Joachims, 1998), are typically among the best performers for the categorization task.	attempts to directly apply SVMs to TDT tasks such as tracking and link detection have not been successful; this has been attributed in part to the lack of enough data for training the SVM 1 .	contrasting
train_20176	The cosine distance between the word distribution for documents ¡ X and ¡ is computed as: This measure has been found to perform well and was used by all the TDT 2002 link detection systems (unpublished presentations at the TDT2002 workshop).	to the Euclidean distance based cosine measure, the Hellinger measure is a probabilistic measure.	contrasting
train_20177	That is, they use patterns to independently discover semantic relationships of words.	for infrequent words, these patterns do not match or, worse yet, generate incorrect relationships.	contrasting
train_20178	Given a typical annotation rate of 5,000 words per hour, we estimated that setting up a name finder for a new problem would take four person days of annotation work -a period we considered reasonable.	this user's problems were too dynamic for that much setup time.	contrasting
train_20179	Capitalization models presented in most previous approaches are monolingual because the models are estimated only from monolingual texts.	for capitalizing machine translation outputs, using only monolingual capitalization models is not enough.	contrasting
train_20180	As one can see, our features are "coarse-grained" (e.g., the language model feature).	kim and Woodland (2004) and Roark et al.	contrasting
train_20181	Unlike a real phrase aligner, the NPA need not wait for the training of the translation model to finish, making it possible for parallelization of translation model training and capitalization model training.	we believe that a real phrase aligner may make phrase alignment quality higher.	contrasting
train_20182	Given that storing all of these phrases leads to very large phrase tables, many research systems simply limit the phrases gathered to those that could possibly influence some test set.	this is not feasible for true production MT systems, since the data to be translated is unknown.	contrasting
train_20183	Table 5 gives the percentage of these which have translations in each of the three training corpora, if we do not use paraphrasing.	after expanding the phrase table using the translations of paraphrases, the coverage of the unique test set phrases goes up dramatically (shown in Table 6).	contrasting
train_20184	The computational cost of training the DTs on large quantities of data is comparable to that of training phrase tables on the same data -large but manageable -and increases linearly with the amount of training data.	currently there is a major problem with DT training: the low proportion of Chinese-English sentence pairs that can be fully segment-aligned and thus be used for DT training (about 27%).	contrasting
train_20185	The way to show that one graph element does not follow from another is to make the cost of aligning them high.	since we are embedded in a search for the lowest cost alignment, this will just cause the system to choose an alternate alignment rather than recognizing a non-entailment.	contrasting
train_20186	Such a sentence is not observed frequently in corpora, and will not be used as clues to generate rules in practice.	we frequently observe sentences of the second type in corpora, and our method generates the paraphrases from the verb-verb cooccurrences taken from such sentences.	contrasting
train_20187	S-V V (n, vcon, vpre, arg, arg ) = Parg(n, vcon)P arg (n, vpre)/P (n) 2 S-N V (n, vcon, vpre) = P coord (vcon, vpre) M I(n, vcon, vpre) = P coord (vcon, vpre)/(P (vcon)P (vpre)) Cond(n, vcon, vpre, arg, arg ) = P coord (vcon, vpre, arg, arg )Parg(n\|vcon)P arg (n\|vpre) /(P arg (n, vpre)P (n)) Rand(n, vcon, vpre, arg, arg ) = random number S-V V was obtained by approximating the probabilities of coordinated sentences, as in the case of BasicS.	we assumed the occurrences of two verbs were independent.	contrasting
train_20188	These methods were discussed in (Yang and Pedersen, 1997).	with any of the three feature ranking criteria, cross validation showed that selecting all features gave the best average validation performance.	contrasting
train_20189	In general, when the test data is similar to the training data, IG (or CHI) is advantageous over F (Yang and Pedersen, 1997).	in this case when the test domain is different from the training domains, F shows advantages for adaptation.	contrasting
train_20190	Moreover, it is often difficult to find experts in these languages both for the expensive annotation effort and even for language specific clues.	comparable multilingual data (such as multilingual news streams) are increasingly available (see section 4).	contrasting
train_20191	Generally, the ensemble of classifiers is generated by training on different subsets of data, rather than different features.	there is some literature within unstructured classified on combining models trained on feature subsets.	contrasting
train_20192	A recent attempt to combine outputs of different alignments views the combination problem as a classifier ensemble in the neural network framework (Ayan et al., 2005).	this method is subject to the unpredictability of random network initialization, whereas ACME is guaranteed to find the model that maximizes the likelihood of training data.	contrasting
train_20193	Also, when we want to combine two models for prediction, finding the Viterbi alignment argmax z p 1 (z \| x)p 2 (z \| x) is intractable for HMM models (by a reduction from quadratic assignment), and a hard intersection argmax z 1 p 1 (z 1 \| x) ∩ argmax z 2 p 2 (z 2 \| x) might be too sparse.	we can threshold the product of two edge posteriors quite easily: z = {z ij = 1 : We noticed a 5.8% relative reduction in AER (for our best model) by using posterior decoding with a validation-set optimized threshold δ instead of using hard intersection of Viterbi alignments.	contrasting
train_20194	Initializing the HMM with model 1 parameters alleviates this problem.	if we jointly train two HMMs starting from a uniform initialization, the HMMs converge to a surprisingly good solution.	contrasting
train_20195	One of the most useful features for the basic matching model is, of course, the set of predictions of IBM model 4.	computing these features is very expensive and we would like to build a competitive model that doesn't require them.	contrasting
train_20196	The example in Figure 4 shows another interesting phenomenon: the multi-fertile alignments for not and député are learned even without lexical fertility features (Figure 4b), because the Dice coefficients of those words with their two alignees are both high.	the surface association of aurait with have is much higher than with would.	contrasting
train_20197	We have avoided using expensive-to-compute features like IBM model 4 predictions up to this point.	if these are available, our model can improve further.	contrasting
train_20198	For example, the antecedent for the pronoun subject they in the first example of work in Table 3 should be ringers, an agent subject that is typical for Sense 1 (exert oneself in an activity).	the feature exaction module found the wrong antecedent changes that is an unlikely fit for the intended verb sense.	contrasting
train_20199	This may be due to the indiscriminately high confidence scores; or it could indicate that classifiers, which are geared at distinguishing between known classes rather than detecting objects that differ from all seen data, are not optimally suited to the task.	one further disadvantage of this approach is that, as mentioned above, it can only be applied to lemmas with more than one annotated sense.	contrasting

train_20100

For the same parser performance level, it selects the fewest number of sentences for a human to check and requires the human to make the least number of corrections.

as we have seen in the earlier experiment, very few sentences in the unlabeled pool satisfy its stringent criteria, so it ran out of data before the parser was trained to convergence.

contrasting

train_20101

Alternatively, the number of documents that contain the pair can also be used.

the nature of the language tests in this work make it impractical to be applied.

contrasting

train_20102

In a sequence will to fight, that trend indicates that will should be a noun rather than a modal verb.

that effect is completely lost in a CMM like (a): P(t will |will, star t ) prefers the modal tagging, and P(TO|to, t will ) is roughly 1 regardless of t will .

contrasting

train_20103

to-left model, fight will receive its more common noun tagging by symmetric reasoning.

the bidirectional model (c) discussed in the next section makes both directions available for conditioning at all locations, using replicated models of P(t 0 |t −1 , t +1 , w 0 ), and will be able to get this example correct.

contrasting

train_20104

Since this model has roughly twice as many tag-tag features, the fact that it outperforms the unidirectional models is not by itself compelling evidence for using bidirectional networks.

it also outperforms model L+L 2 which adds the t 0 , t −2 secondprevious word features instead of next word features, which gives only 96.05% (and R+R 2 gives 95.25%).

contrasting

train_20105

When an information seeker can readily think up a suitable term or linguistic expression to represent the information need, direct searching of text by user-generated terms is faster and more effective than browsing.

when users do not know (or can't remember) the exact expression used in relevant documents, they necessarily struggle to find relevant information in full-text search systems.

contrasting

train_20106

The extracts generated are in size shorter than the original texts.

the number of sentences that E states emit cannot be predetermined.

contrasting

train_20107

In the LNK task, incorrectly flagging two stories as being on the same event is considered a false alarm.

in the NED task, incorrectly flagging two stories as being on the same event will cause a true first story to be missed.

contrasting

train_20108

A LNK system should minimize false alarms by identifying only linked stories, which results in high precision for LNK.

a NED system will minimize false alarms by identifying all stories that are linked, which translates to high recall for LNK.

contrasting

train_20109

A corpus-based approach is able to quickly build a machine translation system for a new domain if a bilingual corpus of that domain is available.

if only a small-sized corpus is available, a low translation quality is obtained.

contrasting

train_20110

These methods evaluate the quality of the translation by measuring the similarity between machine translation results and translations done by humans (called references).

the accuracy increases when multiple references are applied because one source sentence can be translated into multiple target expressions.

contrasting

train_20111

However, no nodes contain only the links 'get', and 'me' in the lower sentence.

focusing on the upper phrase "get me a taxi," it contains four word links that correspond to the lower phrase "get a taxi for me", and they have the same syntactic category.

contrasting

train_20112

The hope is that the same techniques will work for extracting prefixes.

even that will not handle the complex combi-nations of infixes that are possible in agglutinative languages like Turkish or polysynthetic languages like Inuktitut.

contrasting

train_20113

In naturally occurring text, the more frequent sense for the two-sense distinction is reported to occur 92% of the time on average; this result has been found both on the CACM collection and on the WordNet SEMCOR sensetagged corpus (Sanderson & van Rijsbergen, 1999).

the challenge for WSD programs is to work on the harder cases, and the artificially constructed SENSE-VAL1 corpus has more evenly distributed senses (Gaustad, 2001).

contrasting

train_20114

In a manually built corpus, a coreference chain can include pronouns and common nouns that refer to the person.

these forms could not be automatically identified, so coreference chains in our corpus only include noun phrases that contain at least one word from the name.

contrasting

train_20115

In this paper we describe CarmelTC , a novel automatic essay grading approach using a hybrid text classification technique for analyzing essay answers to qualitative physics questions inside the Why2 tutorial dialogue system (VanLehn et al., 2002).

to many previous approaches to automated essay grading (Burstein et al., 1998;Larkey, 1998), our goal is not to assign a letter grade to student essays.

contrasting

train_20116

In general, the performance of existing speech recognition systems, whose designs are predicated on relatively noise-free conditions, degrades rapidly in the presence of a high level of adverse conditions.

a recognizer can provide good performance even in very noisy background conditions if the exact testing condition is used to provide the training material from which the reference patterns of the vocabulary are obtained, which is practically not always the case.

contrasting

train_20117

Thus, a system could map both verbs wipe and remove onto the same action scheme.

the apparently equivalent transformations from (1a) to (1b) and from (2a) to (2b) show otherwise.

contrasting

train_20118

One may say that it is not semantically ambiguous.

simple algorithms such as maximal matching [6,9] and longest matching [6] may not be able to discriminate this kind of ambiguity.

contrasting

train_20119

Within this framework, any type of feature can be used, enabling the system designer to experiment with interesting feature types, rather than worry about specific feature interactions.

in a rule based system, the system designer would have to consider how, for instance, a WordNet (Miller, 1995) derived information for a particular example interacts with a part-of-speech-based information and chunking information.

contrasting

train_20120

Since the documents were selected by subject, one may argue that the task of clustering entities will be much easier if the entities are clearly from different genres.

if this is true, then it may account for about 85% of the entities in the person-x corpus that occur only in one domain subject.

contrasting

train_20121

The form of the salience metric, and the choice of features that factor into it, is governed by our knowledge about the way speech and gesture work.

the penalty function also requires parameters that weigh the importance of each factor.

contrasting

train_20122

This is an uncommon phenomenon, and as such, was penalized highly.

anecdotally it appears that the presence of a disfluency makes this phenomenon more likely.

contrasting

train_20123

Their approach of seeking to maximize this probability is similar to the saliencemaximizing approach that we have described.

instead of using a parametric salience function, they learn a set of conditional probability distributions directly from the data.

contrasting

train_20124

However, scalability of such systems is a bottleneck due to the heavy cost of authoring and maintenance of rule sets and inevitable brittleness due to lack of coverage in the rule sets.

data-driven approaches are robust and the procedure for model building is usually simple.

contrasting

train_20125

Data-driven approaches are robust and provide a simple process of developing applications given the data from the application domain.

the reliance on domain-specific data is also one of the significant bottlenecks of data-driven approaches.

contrasting

train_20126

For a system at ability level 0, the odds increase by another factor of 2.72 to 7.39, giving a probability of .88.

a system with an ability of -3, would have the even odds decrease by a factor of 2.72 to .369, yielding P sq = .27.

contrasting

train_20127

Second, for obvious reasons the raw-score estimates based on the Easy sets are considerably higher than those based on the Hard sets.

table 2 also shows that the standard deviations of the number correct estimates obtained for the Easy sets exceed those of the Hard sets as well (sometimes by over 100%).

contrasting

train_20128

There are many multi-word expressions whose hypernyms are their suffixes, and if some expressions share a common suffix, it is likely to be their hypernym.

if a hypernym candidate appears in a position other than as a suffix of a hyponym candidate, the hypernym candidate is likely to be an erroneous one.

contrasting

train_20129

Sentences are generally demarcated by a major fall (or rise) in f0, lengthening of the final syllable, and following pauses.

the usefulness of prosodic information in sentenceinternal parsing is less clear.

contrasting

train_20130

Simple statistical tests show that there is in fact a significant correlation between the location of opening and closing phrase boundaries and all of the prosodic pseudo-punctuation symbols described above, so there is no doubt that these do convey information about syntactic structure.

adding the prosodic pseudo-punctuation symbols uniformly decreased parsing accuracy relative to input with no prosodic information.

contrasting

train_20131

(2003) consider selection of individual AL methods at run-time.

their AL methods are only ever based on single model approaches.

contrasting

train_20132

For example, the categories of the Model 3 Collins parser distinguish between heads, arguments, and adjuncts and they mark some longdistance dependency paths; these distinctions can guide application-specific postprocessors in extracting important semantic relations.

state-of-the-art parsing systems based on deep grammars mark explicitly and in much more detail a wider variety of syntactic and semantic dependencies and should therefore provide even better support for meaning-sensitive applications.

contrasting

train_20133

Moreover, a surprising variety of problems are attackable with FSTs, from partof-speech tagging to letter-to-sound conversion to name transliteration.

language problems like machine translation break this mold, because they involve massive reordering of symbols, and because the transformation processes seem sensitive to hierarchical tree structure.

contrasting

train_20134

Section 1 informally described the root-to-frontier transducer class R. We saw that R allows, by use of states, finite lookahead and arbitrary rearrangement of nonsibling input subtrees removed by a finite distance.

it is often easier to write rules that explicitly represent such lookahead and movement, relieving the burden on the user to produce the requisite intermediary rules and states.

contrasting

train_20135

In arbitrary document collections, such patterns might be too variable to be easily detected by statistical means.

research has shown that texts from the same domain tend to exhibit high similarity (Wray, 2002).

contrasting

train_20136

We conjecture that this difference in performance stems from the ability of content models to capture global document structure.

the other two algorithms are local, taking into account only the relationships between adjacent word pairs and adjacent sentence pairs, respectively.

contrasting

train_20137

Keller and Lapata's (2003) results suggest that webbased frequencies can be a viable alternative to bigram frequencies obtained from smaller corpora or recreated using smoothing.

they do not demonstrate that realistic NLP tasks can benefit from web counts.

contrasting

train_20138

Phone recognition is known to be less accurate than word recognition.

the second method can only generate phone strings that are substrings of the pronunciations of in-vocabulary word strings.

contrasting

train_20139

We have catalogued a variety of such relationships, and note here that we believe it could prove useful to address semantic interdependencies among SCUS in future work that would involve adding a new annotation layer.

4 in our approach, SCUs are treated as independent annotation values, which has the advantage of affording a rigorous analysis of interannotator reliability (see following section).

contrasting

train_20140

This makes sense in light of the fact that a score is dominated by the higher weight SCUS that appear in a summary.

we wanted to study more precisely at what point scores become independent of the choice of models that populate the pyramid.

contrasting

train_20141

First, an SCU is a set of contributors that are largely similar in meaning, thus SCUs differ from each other in both meaning and weight (number of contributors).

factoids are semi-formal expressions in a FOPL-style semantics, which are compositionally interpreted.

contrasting

train_20142

The model in (6) is simplistic in that the relationships between the features across the clauses are not captured directly.

if two values of these features for the main and subordinate clauses co-occur frequently with a particular marker, then the conditional probability of these features on that marker will approximate the right biases.

contrasting

train_20143

The majority of the main clauses in our data are sentence intitial (80.8%).

there are differences among individual markers.

contrasting

train_20144

Knowing the linear precedence of the two clauses is highly predictive of their type: 80.8% of the main clauses are sentence initial.

this type of positional information is typically not known when fragments are synthesised into a meaningful sentence.

contrasting

train_20145

Given that different Machine Translation (MT) evaluation metrics are useful for capturing different aspects of translation quality, it becomes desirable to create MT systems tuned with respect to each individual criterion.

the maximum likelihood techniques that underlie the decision processes of most current MT systems do not take into account these application specific goals.

contrasting

train_20146

The two hypothesis translations are very similar at the word level and therefore the BLEU score, PER and the WER are identical.

we observe that the sentences differ substantially in their syntactic structure (as seen from Parse-Trees in Figure 3), and to a lesser extent in their word-to-word alignments (Figure 1) to the source sentence.

contrasting

train_20147

In (Och and Weber, 1998;Och et al., 1999), a two-level alignment model was employed to utilize shallow phrase structures: alignment between templates was used to handle phrase reordering, and word alignments within a template were used to handle phrase to phrase translation.

phrase level alignment cannot handle long distance reordering effectively.

contrasting

train_20148

One is the PRank algorithm, a variant of the perceptron algorithm, that uses multiple biases to represent the boundaries between every two consecutive ranks (Crammer and Singer, 2001;Harrington, 2003).

as we will show in section 3.7, the PRank algorithm does not work on the reranking tasks due to the introduction of global ranks.

contrasting

train_20149

In addition, we only want to maintain the order of two candidates if their ranks are far away from each other.

we do not care the order of two translations whose ranks are very close, e.g.

contrasting

train_20150

As a result, we cannot use the PRank algorithm in the reranking task, since there are no global ranks or boundaries for all the samples.

the approach of using pairwise samples does work.

contrasting

train_20151

It achieves a BLEU score of 31.7% on the Baseline, 32.8% on the Best Feature, but only 32.6% on the Top Twenty features.

it is within the range of 95% confidence.

contrasting

train_20152

Much of this research has used databases of speech read by actors or native speakers as training data (often with semantically neutral content) (Oudeyer, 2002;Polzin and Waibel, 1998;Liscombe et al., 2003).

such prototypical emotional speech does not necessarily reflect natural speech (Batliner et al., 2003), such as found in tutoring dialogues.

contrasting

train_20153

Although much past work predicts only two classes (e.g., negative/non-negative) (Batliner et al., 2003;Ang et al., 2002;Lee et al., 2001), our experiments produced the best predictions using our three-way distinction.

to (Lee et al., 2001), our classifications are contextrelative (relative to other turns in the dialogue), and taskrelative (relative to tutoring), because like (Ang et al., 2002), we are interested in detecting emotional changes across our dialogues.

contrasting

train_20154

Comparing the results of these combined (speech+text) feature sets with the speech versus text results in Table 1, we find that for autotext+speech-ident and all +ident feature sets, the combined feature set slightly decreases predictive accuracy when compared to the corresponding text-only feature set.

there is no significant difference between the best results in each table (all-text+speech+ident vs. alltext+ident).

contrasting

train_20155

Comparing these results with the results in Tables 1 and 2, we find that while overall the performance of contextual non-combined feature sets shows a small performance increase over most non-contextual combined or non-combined feature sets, there is again a slight decrease in performance across the best results in each table.

there is no significant difference between these best results (alltext+glob-ident vs. all-text+speech+ident vs. alltext+ident).

contrasting

train_20156

Ideally, by training acoustic models on target non-native speech, one would capture its specific characteristics just as training on native speech does.

collecting amounts of non-native speech that are large enough to fully train speaker-independent models is a hard and often impractical task.

contrasting

train_20157

Yet, this is acceptable given the small amount of training data for the language model and the conversational nature of the speech.

performance degrades significantly for non-native speakers, with a word error rate of 52.0%.

contrasting

train_20158

For example, when referring to bus stops by street intersections, all native speakers in our training set simply used "A and B", hence the word "intersection" was not in the language model.

many non-native speakers used the full expression "the intersection of A and B".

contrasting

train_20159

In fact, this is not only true for non-native speakers and lexical entrainment is often described as a negotiation process between the speakers (Clark and Wilkes-Gibbs, 1986).

while it is possible for limited-domain system designers to establish a set of words and constructions that are widely used among native speakers, the variable nature of the expressions mastered by non-native speakers make adaptation a desirable feature of the system.

contrasting

train_20160

Partial Path -For the argument identification task, path is the most salient feature.

it is also the most data sparse feature.

contrasting

train_20161

On performing the search, we found that the overall performance improvement was not much different than that obtained by resolving overlaps as mentioned earlier.

we found that there was an improvement in the CORE ARGUMENT accuracy on the combined task of identifying and assigning semantic arguments, given hand-corrected parses, whereas the accuracy of the ADJUNCTIVE ARGUMENTS slightly deteriorated.

contrasting

train_20162

Although, there was an increase in F 1 score when the language model probabilities were jointly estimated over all the predicates, this improvement is not statistically significant.

estimating the same using specific predicate lemmas, showed a significant improvement in accuracy.

contrasting

train_20163

Speech form is familiar to humans, and can convey information effectively (Nadamoto et al., 2001;Hayashi et al., 1999).

little electronic information is provided in speech form so far.

contrasting

train_20164

However, little electronic information is provided in speech form so far.

there is a lot of information in text form, and it can be transformed into speech by a speech synthesis.

contrasting

train_20165

There are several studies that compare two corpora which have different styles, for example, written and spoken corpora or British and American English corpora, and try to find expressions unique to either of the styles (Kilgarriff, 2001).

those studies did not deal with paraphrases.

contrasting

train_20166

This paper has focused only on paraphrasing predicates.

there are other kinds of paraphrasing which are necessary in order to paraphrase written language text into spoken language.

contrasting

train_20167

On the one hand, we have three different variants of the single-word based model IBM4.

we have two phrase-based systems, namely the alignment templates and the one described in this work.

contrasting

train_20168

This method yielded word classes that offered more robust count approximations for their member words.

both methods yielded similar results when embedded in the larger system, and so we will report on the results of using Good-Turing so as to remain more directly comparable to Dagan et al.

contrasting

train_20169

The MaxEnt model selected the temerity as the antecedent of its (salience value: 0.30), preferring it to the correct antecedent the endowment (salience value: 0.10).

altaVista found no occurrences of temerity's money or its variants on the web, and thus the unnormalized and normalized counts were 0.

contrasting

train_20170

However, AltaVista found no occurrences of temerity's money or its variants on the web, and thus the unnormalized and normalized counts were 0.

endowment's money and its variants had unnormalized and normalized statistics of 1583 and 1.47 × 10 −3 respectively.

contrasting

train_20171

In the cases in which statistics reinforced a wrong answer, no (reasonable) manipulation of statistical features or filters can rescue the prediction.

for the cases in which statistics could help, their successful use will depend on the existence of a formula that can capture these cases without changing the predictions for examples that the model currently classifies correctly.

contrasting

train_20172

Since pronouns carry little semantics of their own, resolving them depends almost entirely on context.

even though context can be helpful for resolving definite NPs, context can be trumped by the semantics of the nouns themselves.

contrasting

train_20173

For example, sources in a language that is translated to English will consistently use the same terminology, resulting in greater similarity between linked documents with the same native language.

sources from radio broadcasts may be transcribed much less consistently than text sources due to recognition errors, so that the expected similarity of a radio broadcast and a text source is less than that of two text sources.

contrasting

train_20174

We refer to the statistics characterizing story pairs with the same source types as source-pair specific information.

to the source-specific thresholds used by CMU, we normalize the similarity measures based on the sourcepair specific information, simultaneously with combining different similarity measures.

contrasting

train_20175

SVM-based systems, such as that described in (Joachims, 1998), are typically among the best performers for the categorization task.

attempts to directly apply SVMs to TDT tasks such as tracking and link detection have not been successful; this has been attributed in part to the lack of enough data for training the SVM 1 .

contrasting

train_20176

The cosine distance between the word distribution for documents ¡ X and ¡ is computed as: This measure has been found to perform well and was used by all the TDT 2002 link detection systems (unpublished presentations at the TDT2002 workshop).

to the Euclidean distance based cosine measure, the Hellinger measure is a probabilistic measure.

contrasting

train_20177

That is, they use patterns to independently discover semantic relationships of words.

for infrequent words, these patterns do not match or, worse yet, generate incorrect relationships.

contrasting

train_20178

Given a typical annotation rate of 5,000 words per hour, we estimated that setting up a name finder for a new problem would take four person days of annotation work -a period we considered reasonable.

this user's problems were too dynamic for that much setup time.

contrasting

train_20179

Capitalization models presented in most previous approaches are monolingual because the models are estimated only from monolingual texts.

for capitalizing machine translation outputs, using only monolingual capitalization models is not enough.

contrasting

train_20180

As one can see, our features are "coarse-grained" (e.g., the language model feature).

kim and Woodland (2004) and Roark et al.

contrasting

train_20181

Unlike a real phrase aligner, the NPA need not wait for the training of the translation model to finish, making it possible for parallelization of translation model training and capitalization model training.

we believe that a real phrase aligner may make phrase alignment quality higher.

contrasting

train_20182

Given that storing all of these phrases leads to very large phrase tables, many research systems simply limit the phrases gathered to those that could possibly influence some test set.

this is not feasible for true production MT systems, since the data to be translated is unknown.

contrasting

train_20183

Table 5 gives the percentage of these which have translations in each of the three training corpora, if we do not use paraphrasing.

after expanding the phrase table using the translations of paraphrases, the coverage of the unique test set phrases goes up dramatically (shown in Table 6).

contrasting

train_20184

The computational cost of training the DTs on large quantities of data is comparable to that of training phrase tables on the same data -large but manageable -and increases linearly with the amount of training data.

currently there is a major problem with DT training: the low proportion of Chinese-English sentence pairs that can be fully segment-aligned and thus be used for DT training (about 27%).

contrasting

train_20185

The way to show that one graph element does not follow from another is to make the cost of aligning them high.

since we are embedded in a search for the lowest cost alignment, this will just cause the system to choose an alternate alignment rather than recognizing a non-entailment.

contrasting

train_20186

Such a sentence is not observed frequently in corpora, and will not be used as clues to generate rules in practice.

we frequently observe sentences of the second type in corpora, and our method generates the paraphrases from the verb-verb cooccurrences taken from such sentences.

contrasting

train_20187

S-V V (n, vcon, vpre, arg, arg ) = Parg(n, vcon)P arg (n, vpre)/P (n) 2 S-N V (n, vcon, vpre) = P coord (vcon, vpre) M I(n, vcon, vpre) = P coord (vcon, vpre)/(P (vcon)P (vpre)) Cond(n, vcon, vpre, arg, arg ) = P coord (vcon, vpre, arg, arg )Parg(n|vcon)P arg (n|vpre) /(P arg (n, vpre)P (n)) Rand(n, vcon, vpre, arg, arg ) = random number S-V V was obtained by approximating the probabilities of coordinated sentences, as in the case of BasicS.

we assumed the occurrences of two verbs were independent.

contrasting

train_20188

These methods were discussed in (Yang and Pedersen, 1997).

with any of the three feature ranking criteria, cross validation showed that selecting all features gave the best average validation performance.

contrasting

train_20189

In general, when the test data is similar to the training data, IG (or CHI) is advantageous over F (Yang and Pedersen, 1997).

in this case when the test domain is different from the training domains, F shows advantages for adaptation.

contrasting

train_20190

Moreover, it is often difficult to find experts in these languages both for the expensive annotation effort and even for language specific clues.

comparable multilingual data (such as multilingual news streams) are increasingly available (see section 4).

contrasting

train_20191

Generally, the ensemble of classifiers is generated by training on different subsets of data, rather than different features.

there is some literature within unstructured classified on combining models trained on feature subsets.

contrasting

train_20192

A recent attempt to combine outputs of different alignments views the combination problem as a classifier ensemble in the neural network framework (Ayan et al., 2005).

this method is subject to the unpredictability of random network initialization, whereas ACME is guaranteed to find the model that maximizes the likelihood of training data.

contrasting

train_20193

Also, when we want to combine two models for prediction, finding the Viterbi alignment argmax z p 1 (z | x)p 2 (z | x) is intractable for HMM models (by a reduction from quadratic assignment), and a hard intersection argmax z 1 p 1 (z 1 | x) ∩ argmax z 2 p 2 (z 2 | x) might be too sparse.

we can threshold the product of two edge posteriors quite easily: z = {z ij = 1 : We noticed a 5.8% relative reduction in AER (for our best model) by using posterior decoding with a validation-set optimized threshold δ instead of using hard intersection of Viterbi alignments.

contrasting

train_20194

Initializing the HMM with model 1 parameters alleviates this problem.

if we jointly train two HMMs starting from a uniform initialization, the HMMs converge to a surprisingly good solution.

contrasting

train_20195

One of the most useful features for the basic matching model is, of course, the set of predictions of IBM model 4.

computing these features is very expensive and we would like to build a competitive model that doesn't require them.

contrasting

train_20196

The example in Figure 4 shows another interesting phenomenon: the multi-fertile alignments for not and député are learned even without lexical fertility features (Figure 4b), because the Dice coefficients of those words with their two alignees are both high.

the surface association of aurait with have is much higher than with would.

contrasting

train_20197

We have avoided using expensive-to-compute features like IBM model 4 predictions up to this point.

if these are available, our model can improve further.

contrasting

train_20198

For example, the antecedent for the pronoun subject they in the first example of work in Table 3 should be ringers, an agent subject that is typical for Sense 1 (exert oneself in an activity).

the feature exaction module found the wrong antecedent changes that is an unlikely fit for the intended verb sense.

contrasting

train_20199

This may be due to the indiscriminately high confidence scores; or it could indicate that classifiers, which are geared at distinguishing between known classes rather than detecting objects that differ from all seen data, are not optimally suited to the task.

one further disadvantage of this approach is that, as mentioned above, it can only be applied to lemmas with more than one annotated sense.

contrasting