tasksource/scinli · Datasets at Hugging Face

id stringlengths 7 12	sentence1 stringlengths 6 1.27k	sentence2 stringlengths 6 926	label stringclasses 4 values
train_20700	This performance drop is statistically significant out of domain.	the difference between the Eisner and Attardi algorithms is not statistically significant out of domain.	contrasting
train_20701	Recent work has shown that the combination of base parsers at learning time, e.g., through stacking, yields considerable benefits (Nivre and McDonald, 2008;Attardi and Dell'Orletta, 2009).	it is unclear how these approaches compare against the simpler ensemble models, which combine parsers only at runtime.	contrasting
train_20702	SITGs have proven to be a powerful tool in Syntax Machine Translation.	the algorithms have been proposed do not explore all the possible parse trees.	contrasting
train_20703	Using the reranker features distributed with the Brown reranker (Charniak and Johnson, 2005), which we call the "standard" set below, we obtained no overall improvement in f-score when either reranking the Berkeley parser n-best lists alone, or when the Berkeley parses were combined with the Brown parses.	it is possible that these results reflect the fact that the features used by the reranker were chosen because they improve the Brown parser, i.e., they are the result of feature selection based on reranking the Brown parser's n-best lists.	contrasting
train_20704	It would lead to deemphasizing those unusual types of information that might be being discussed as part of a post.	one might expect that those things that are unusual types of information might actually be more likely to be the in-focus information within an initiation that responses may be likely to refer to.	contrasting
train_20705	Conversational Agents have been shown to be effective tutors in a wide range of educational domains.	these agents are often ignored and abused in collaborative learning scenarios involving multiple students.	contrasting
train_20706	In all three conditions, students go through the same task plan.	the degree of social performance is varied from minimal (Task) to ideal (Human).	contrasting
train_20707	On the right side of Table 2, we notice that the human tutors (H) were rated higher on being part of the team (Integration), being more liked, being friendlier and keeping the group more socially comfortable (T.Releasing).	the social tutors (S) were rated to be friendlier and were only marginally better at being seen as part of the team.	contrasting
train_20708	Language identification has traditionally been approached with character-level language models.	the language model approach crucially depends on the length of the text in question.	contrasting
train_20709	Both approaches benefit from their ability to consider large, flexible spans of source context when making transduction decisions.	they encode this context in different ways, providing their respective models with different information.	contrasting
train_20710	One approach to string transduction is to view it as a tagging problem where the input characters are tagged with the output characters.	since sounds are often represented by multicharacter units, the relationship between the input and output characters is often complex.	contrasting
train_20711	In the first 2 passes, we use a standard bi-gram LM to generate lattices, followed by a standard tri-gram LM rescoring of lattices.	in the third pass, we generate both lattices and N-best lists based on the same bi-gram LM.	contrasting
train_20712	provement of 0.8% (about 4.8% relative) compared to the standard tri-gram lattice rescoring.	we have 0.6% absolute improvement (about 3.7% relative) compared to the standard 4gram lattice rescoring.	contrasting
train_20713	(Zitouni et al., 2005) used Arabic morphologically segmented data and claimed to have very competitive results in ACE 2003 and ACE 2004 data.	(Benajiba et al., 2008) report good results for Arabic NER on ACE 2003, 2004and 2005 segmentation.	contrasting
train_20714	This is because the token itself becomes insignificant information to the classifier.	when only punctuation separation is performed (W ord s ), the data is significantly sparse and the obtained results achieves high F-measure (77.1) only when outputs of other classifiers are used.	contrasting
train_20715	Hence, we keep the edge to indicate that the child can be used as a preferred substitute for the parent.	the edge is removed if the ratio is small (less than a threshold t, see Fig.	contrasting
train_20716	Using aggressive pruning, the list size and number of redun-dant phrase patterns are greatly reduced.	the classification accuracy does not decrease.	contrasting
train_20717	The detection of pitch accents and phrase boundaries has received significantly more research attention than the classification of accent types and phrase ending behavior.	one technique that has been used in a number of research efforts is to simultaneously detect and classify pitch accent.	contrasting
train_20718	In both systems, the frame-based log posterior vector of P (phone\|acoustics) over all phones is decorrelated using the Karhunen-Loeve (KL) transform; unlike MLPs, CRFs take into account the entire label sequence when computing local posteriors.	posterior estimates from the CRF tend to be overconfident compared to MLP posteriors (Morris and Fosler-Lussier, 2009).	contrasting
train_20719	Calculating these is computationally inexpensive for many simple tasks (such as classification and regression).	marginal and MAP inference tends to be expensive for complex structured prediction models (such as the joint information extraction models of ), making semisupervised learning intractable.	contrasting
train_20720	The softmax-margin approach offers (1) a convex objective, (2) the ability to incorporate task-specific cost functions, and (3) a probabilistic interpretation (which supports, e.g., hidden-variable learning and computation of posteriors).	max-margin training and MIRA do not provide (3); risk and JRB do not provide (1); and CLL does not support (2).	contrasting
train_20721	A drawback of this approach is that it is extremely local: while decisions can be based on complex structures on the left, they can look only at a few words to the right.	our algorithm builds a dependency tree by iteratively selecting the best pair of neighbours to connect at each parsing step.	contrasting
train_20722	Each performed action changes the partial struc-tures and with it the extracted features and the computed scores.	these changes are limited to a fixed local context around the attachment point of the action.	contrasting
train_20723	Shorter edges are arguably easier to predict, and our parses builds them early in time.	it is also capable of producing long dependencies at later stages in the parsing process.	contrasting
train_20724	(2009b) study dependency treebanks for nine languages and find that all dependency structures meet the mildly ill-nested condition in the dependency treebanks for some gap degree.	they do not report the maximum gap degree or parsing complexity.	contrasting
train_20725	This is certainly true when the graph is of high quality and all paths trustworthy.	in a graph constructed from web cooccurrence statistics, this is rarely the case.	contrasting
train_20726	Ideally we would see an order on such phrases, e.g., "more brittle" has a larger negative polarity than "brittle", which in turn has a larger negative polarity than "less brittle".	this is rarely the case and usually the adjective has the highest polarity magnitude.	contrasting
train_20727	In the approach, a subjective sentence is represented as a set of words in the sentence, ignoring word order and head-modifier relation between words.	sentiment classification is different from traditional topic-based text classification.	contrasting
train_20728	Recently, several methods have been proposed to cope with the problem (Zaenen, 2004;Ikeda et al., 2008).	these methods are based on flat bag-of-features representation, and do not consider syntactic structures which seem essential to infer the polarity of a whole sentence.	contrasting
train_20729	Titov and McDonald (2008b) underline the need for unsupervised methods for aspect detection.	according to the authors, existing topic models, such as standard Latent Dirichlet Allocation (LDA) (Blei et al., 2003), are not suited to the task of aspect detection in reviews, because they tend to capture global topics in the data, rather than rateable aspects pertinent to the review.	contrasting
train_20730	For example, for many verbs, nsubj tends to start a cause path and dobj a theme path.	for "bind" that signifies a Binding event, both lead to themes, as in "A binds B".	contrasting
train_20731	Several studies have been performed on identifying PICO elements in abstracts (Demner-Fushman and Lin, 2007;Hansen et al., 2008;Chung, 2009).	all of them are reporting coarsegrain (sentence-level) tagging methods that have not yet been shown to be sufficient for the purpose of IR.	contrasting
train_20732	Moreover, there is currently no standard test collection of questions in PICO structure available for evaluation.	the most critical aspect in IR is term weighting.	contrasting
train_20733	These approaches follow the assumption that the user knows where the most relevant information is located.	(Kamps et al., 2005) showed that it is preferable to use structure as a search hint, and not as a strict search requirement The second approach consists in integrating the document structure at the indexing step by introducing a structure weighting scheme (Wilkinson, 1994).	contrasting
train_20734	There has been several studies that cover the PICO extraction problem.	as far as we know, none of them analyses and uses the positional ditribution of these elements within the documents for the purpose of IR.	contrasting
train_20735	It could also be caused by the limited number of query in our test collection.	we can determine reasonable weights by tuning each part weight separately.	contrasting
train_20736	A straightforward idea is to detect PICO elements in documents and use the elements in the retrieval process.	this approach does not work well because of the difficulty to arrive at a consistent tagging of these elements.	contrasting
train_20737	Given a query, search engines retrieve relevant pictures by analyzing the image caption (if it exists), textual descriptions found adjacent to the image, and other text-related factors such as the file name of the image.	since they do not analyze the actual content of the images, search engines cannot be used to retrieve pictures from unannotated collections.	contrasting
train_20738	The probability of a document d in a corpus is defined as: Computing the posterior distribution P(θ, z\|d, α, β) of the hidden variables given a document is intractable in general.	a variety of approximate inference algorithms have been proposed in the literature including variational inference which our model adopts .	contrasting
train_20739	All are almost surprisingly intuitive, but this is not terribly surprising since Chinese and English have very similar large-scale structures (both are head initial, both have adjectives and quantifiers that precede nouns).	we see two entries in the list (starred) that correspond to an En- glish word order that is ungrammatical in Chinese: PP modifiers in Chinese typically precede the VPs they modify, and CPs (relative clauses) also typically precede the nouns they modify.	contrasting
train_20740	4aThe VSO basic word order is evident: early in the sentence, there is a strong tendency towards right movement around arguments after covering the verb.	right movement is increasingly penalized at the end of the sentence.	contrasting
train_20741	The binarization in Figure 1 is unfortunate because the obtained production cannot be factorized such that only two nonterminals occur in each rule.	the binarization of Figure 4 allows the factorization into S(U , NP ) ↔ S(U , NP ) and U : @(NP , V ) ↔ @(V , NP ), which are fully binarized productions.	contrasting
train_20742	However, the binarization of Figure 4 allows the factorization into S(U , NP ) ↔ S(U , NP ) and U : @(NP , V ) ↔ @(V , NP ), which are fully binarized productions.	in general, STSGs (or SCFGs or extended tree transducers) cannot be fully binarized as shown in Aho and Ullman (1972).	contrasting
train_20743	Consequently, we cannot compute forward or backward applications for arbitrary MBOT.	if the MBOT is equivalent to an STSG (for example, because it was constructed by the method presented before Theorem 3), then forward and backward application can be computed essentially as for STSG.	contrasting
train_20744	MBOT can efficiently be used (with computational benefits) as an alternative representation for transformations computed by STSG (or compositions of STSG).	mBOT can also compute transformations, of which the domain or range cannot be represented by a TSG.	contrasting
train_20745	Both methods were trained and tested on data from the Ziff-Davis corpus (Knight and Marcu, 2002), and they achieved very similar grammaticality and meaning preservation scores, with no statistically significant difference.	their compression rates (counted in words) were very different: 70.37% for the noisy-channel method and 57.19% for the C4.5-based one.	contrasting
train_20746	Our method is similar to Nomoto's, in that it uses two stages, one that chops the source dependency tree generating candidate compressions, and one that ranks the candidates.	we experimented with more elaborate ranking models, and our method does not employ any manually crafted rules.	contrasting
train_20747	guage model trained on a large background corpus.	language models tend to assign smaller probabilities to longer sentences; therefore they favor short sentences, but not necessarily the most appropriate compressions.	contrasting
train_20748	Extrinsic evaluations have also shown that, while extractive summaries may be less coherent than human abstracts, users still find them to be valuable tools for browsing documents (He et al., 1999;.	these same evaluations also indicate that concise abstracts are generally preferred by users and lead to higher objective task scores.	contrasting
train_20749	The system described thus far may appear extractive in nature, as the transformation step is identifying informative sentences in the conversation.	these selected sentences correspond to < participant, relation, entity > triples in the ontology, for which we can subsequently generate novel text by creating linguistic annotations of the conversation ontology (Galanis and Androutsopolous, 2007).	contrasting
train_20750	This is due to the fact that the leading sentences for these two domains do not indicate any significance, hence the Lead system just behaves like Random.	for the scientific and newswire domains, the leading sentences do have importance so the Lead system consistently outperforms Random.	contrasting
train_20751	f MMR , is not guaranteed everywhere monotone.	our theoretical results still holds for f MMR with high probability in practice.	contrasting
train_20752	We assume that there is not enough high-quality data to build a monolingual selectional preference model for the source language (shown by dotted lines).	we can use a bilingual vector space, that is, a semantic space in which words of both the source and the target language are represented, to translate each source language word s into the target language by identifying its nearest (most similar) target word tr(s): Now we can use a target language selectional preference model to obtain plausibilities for source triples: where the superscript indicates the language.	contrasting
train_20753	Data Our experiments were carried out on the CoNLL 2008 (Surdeanu et al., 2008) training dataset which contains both verbal and nominal predicates.	we focused solely on verbal predicates, following most previous work on semantic role labeling (Màrquez et al., 2008).	contrasting
train_20754	Since different similarity functions can be used within this framework, one may wish to select the one that is the most appropriate or relevant to the task considered.	a crucial requirement for this choice to be realistic is to ensure that for the family of similarity functions considered the expected similarity maximization is efficiently computable.	contrasting
train_20755	As mentioned earlier, we leave this question to future work.	we can offer a brief look at how one could tackle this question.	contrasting
train_20756	The empirical adequacy of 2-SCFG models would presumably be lower with automatically-aligned texts and if the study also included non-European languages.	phrase-based systems can properly handle inside-out alignments when used with a reasonably large distortion limit, and all configurations in Fig.	contrasting
train_20757	6 Likewise, the procedure could be applied to statistical systems that only generate k-best lists.	we would not expect the same strong performance from model combination in these constrained settings.	contrasting
train_20758	Additionally, the experiments were performed over a large corpus of messages that are not available for use by other researchers.	we use messages from the widely-available Enron email corpus (Klimt and Yang, 2004) for our own experiments.	contrasting
train_20759	This strongly suggests that without zoning, the classifier is not learning features from the training set at a useful level of generality.	once we add the zoning classifier, the top-10 unigrams and bigrams appear to correspond much better with linguistic intuitions about the language of requests.	contrasting
train_20760	There is much disagreement about the units and elementary relations of discourse structure, but they agree that the structures are hierarchical, most commonly trees (Marcu, 2000), while others have argued for directed acyclic graphs (Danlos, 2004), or general graphs (Wolf and Gibson, 2004).	most of the segmentation research to date has focused on linear segmentation, in which segments are non-overlapping and sequential, and it has been argued that this sequence model is sufficient for many purposes (Hearst, 1994).	contrasting
train_20761	The mean scores for the BIN baseline are over 50% on the encyclopedia data.	the mean score for BIN on the Choi standard data (Fig.	contrasting
train_20762	For example, SkillSum (Williams and Reiter, 2008) and ICONOCLAST (Power et al., 2003) are two contemporary generation systems that allow for specifying aspects of style such as choice of discourse marker, clause order, repetition and sentence and paragraph lengths in the form of constraints that can be optimised.	to date, these systems do not consider syntactic reformulations of the type we are interested in.	contrasting
train_20763	This pulls the average for reformulated sentences down.	on average 2 out of 7 reformulations score quite high.	contrasting
train_20764	In general, one would expect that short transformation sequences to provide good evidence of true entailments.	to account for the grandparent-child relationship in the hypothesis, TED would produce a fairly long sequence, relabeling nearby to be near, deleting the two nodes for Rossville Blvd, and then reinserting those nodes under near.	contrasting
train_20765	Settings of 0.1, 0.2, 0.3, and 0.4 led to 10-fold cross-validation The main difference between our kernel and the CTK is that we sum over all pairs of subtrees (Equation 3).	the CTK only considers only one pair of subtrees.	contrasting
train_20766	9 In accuracy values that were not significantly different from each other.	we did observe that increased search failure ( §3.4) resulted from settings above 0.5.	contrasting
train_20767	Chawathe and Garcia-Molina (1997) describe a tree edit algorithm for detecting changes in structured documents that incorporates edits for moving subtrees and reordering children.	they make assumptions unsuitable for natural language, such as the absence of re-cursive syntactic rewrite rules.	contrasting
train_20768	Lexical-syntactic rules can be automatically extracted from plain corpora (e.g., (Lin and Pantel, 2001;Szpektor and Dagan, 2008)) but the quality (also in terms of little noise) and the coverage is low.	rules written at the semantic level are more accurate but their automatic design is difficult and so they are typically hand-coded for the specific phenomena.	contrasting
train_20769	by using FrameNet semantics (e.g., like in (Burchardt et al., 2007)), it is possible to encode a lexical-syntactic rule using the KILLING and the DEATH frames, i.e.	: to use this model, specific rules and a semantic role labeler on the specific corpora are needed.	contrasting
train_20770	Again, the syntactic rules (with variables) which this kernel BNC WN WIKI RTE2 0.55 0.42 0.83 RTE3 0.54 0.41 0.83 RTE5 0.45 0.34 0.82 Table 3: Coverage of the different resources for the words of the three datasets can provide are not enough general for RTE3.	maxSSTK+WOK improves WOK on all datasets thanks to its generalization ability.	contrasting
train_20771	This may be due to the fact that, as shown in Table 1, the overall differences between partners in mixed-gender pairs are quite low, and so neither partner may be doing much turn-byturn matching.	as we expected, entrainment is least prevalent among male-male pairs.	contrasting
train_20772	This reflects the intuition that someone overly eager to be liked may be perceived as annoying and socially inept.	similarity-attraction theory states that similarity promotes attraction, and someone might therefore entrain in order to obtain his partner's social approval.	contrasting
train_20773	As we expected, giving encouragement is correlated with entrainment for all three gender groups, and trying to be liked is correlated with entrainment for male-male and female-male groups.	trying to dominate is not correlated with entrainment on any feature, and conversation awkward is actually positively correlated with entrainment on jitter.	contrasting
train_20774	A high number is negative in that it is the sign of an inefficient dialogue, one which takes many turn exchanges to accomplish the objective.	it may also be the sign of easy, flowing dialogue between the partners.	contrasting
train_20775	Besides shell language, there were other annotations relevant to essay scoring.	we ignored them for this study because they are not directly relevant to the task of shell language detection.	contrasting
train_20776	For the annotated essay test set ( §3.2), the percentage of tokens tagged as shell was 14.0% (11.6% were labeled as shell by the first annotator).	the percentage of tokens tagged as shell was 4.2% for Lincoln-Douglas, 5.4% for Kennedy-Nixon, 4.6% for Gore-Bush, and 4.8% for Obama-McCain.	contrasting
train_20777	It is not completely clear whether the smaller percentages tagged as shell are due to a lack of coverage by the shell detector or more substantial differences in the domain.	it seems that these debates genuinely include less shell.	contrasting
train_20778	Also, Bush employs a somewhat atypical sentence structure here: "It's not what I think and its not my intentions and not my plan."	the system also incorrectly tagged sequences as shell, particularly in short sentences (e.g., "Are we as strong as we should be?").	contrasting
train_20779	Discourse markers, however, are typically only single words or short phrases that express a limited number of relationships.	shell can capture longer sequences that express more complex relationships between the components of an argumentative discourse (e.g., "But let's get back to the core issue here" signals that the following point is more important than the previous one).	contrasting
train_20780	In shell detection, we focus on the lexico-syntactic level, aiming to identify the bold words as shell.	work on argumentation schemes focuses at a higher level of abstraction, aiming to classify the sentence as an attempt to persuade by appealing to an external authority.	contrasting
train_20781	There have been many attempts over the last decade to develop model-based approaches to the phrase alignment problem (Marcu and Wong, 2002;Birch et al., 2006;DeNero et al., 2008;Blunsom et al., 2009).	most of these have met with limited success compared to the simpler heuristic method.	contrasting
train_20782	e 3 e 4 e 5 e 6 e 7 Because there is a sure link at a 48 , σ f 8 = [4, 4] does not include the possible link at a 38 .	f 7 only has possible links, so σ f 7 = [5, 6] is the span containing those.	contrasting
train_20783	DeNero and Klein 2010implicitly included these constraints in their representation: instead of sets of variables, they used a structured representation that only encodes triples (a, π, σ) satisfying both the mapping π = π(a) and the structural constraint that a can be generated by a block ITG grammar.	our inference procedure, BP, requires that we represent (a, π, σ) as an assignment of values to a set of variables.	contrasting
train_20784	The BP ITG model performs comparably to the Viterbi ITG model.	because posterior decoding permits explicit tradeoffs between precision and recall, it can do much better in the recallbiased measures, even though the Viterbi ITG model was explicitly trained to maximize F 5 (DeNero and Klein, 2010).	contrasting
train_20785	One practical solution is to restrict the output vocabulary to a short-list composed of the most frequent words (Schwenk, 2007).	the usual size of the short-list is under 20k, which does not seem sufficient to faithfully represent the translation models of section 2.	contrasting
train_20786	One solution would be full-blown transliteration (Knight and Graehl, 1998), followed by application of Jaro-Winkler.	transliteration systems are complex and require significant training resources.	contrasting
train_20787	11 Because a similar corpus did not exist for development, we split the evaluation corpus into development and test sections.	the usual method of splitting by document would not confine all mentions of each entity to one side of the split.	contrasting
train_20788	First, the model learns that high edit distance is predictive of a mismatch.	singleton strings that do not match often have a lower edit distance than longer strings that do match.	contrasting
train_20789	To our knowledge, Baron and Freedman (2008) reported the only previous results on the ACE2008 data set.	they only gave gold results for English, and clustered the entire evaluation corpus (test+development).	contrasting
train_20790	Within this work, specific affect definitions vary slightly with the intention of being coherent within the application and domain and being relevant to the specific adaptation goal (Martalo et al., 2008).	affective systems researchers generally agree that disengaged users show little involvement in the interaction, and often display facial, gestural and linguistic signals such as gaze avoidance, finger tapping, humming, sarcasm, et cetera.	contrasting
train_20791	All turns are used in the disengagement detection experiments described next.	only the training problem dialogues (360, 5 per student, 6044 student turns) are used for the performance analyses in Sections 6-7, because the final test problem was given after the instruments measuring performance (survey and posttest).	contrasting
train_20792	Compared to the results in Table 2, we find that the VSM and LSA methods are very robust to recognition errors, and we only observe slight correlation decreases on these features.	the decrease for the PMI-based method is quite large.	contrasting
train_20793	Our experimental results showed that all the features obtained good correlations with human proficiency scores if there are no recognition errors in the text transcripts, with the PMI-based method performing the best over three similarity measures.	if we used ASR transcripts, we observed a marked performance drop for the PMI-based method.	contrasting
train_20794	In our experiments, we noticed that shorter conversations suffer from poor classification.	the results from the above section appear to contradict this assertion, as a 30-word window can give very good performance.	contrasting
train_20795	Though the performance gain over the random ranker has shrunk considerably, there is still some utility in using the opening of a conversation to determine its ultimate duration.	it is clear predicting duration via conversation opening is a much more difficult task overall.	contrasting
train_20796	show that on certain synthetic-data problems, this frequentist training regimen significantly reduced test-data loss compared to approximate maximum likelihood estimation (MLE).	this method has not been evaluated on real-world problems until now.	contrasting
train_20797	Finally, we hypothesized that sum-product inference may produce more accurate results in certain cases as it allows more information about different parts of the model to be exchanged.	our results show that for these three problems, sum-product and max-product inference yield statistically indistinguishable results.	contrasting
train_20798	These computations are tractable for HMMs, since the distribution q(t) = p θ (t \| w) that is optimal at the E-step (which makes the inequality tight) can be represented as a lattice (a certain kind of weighted DFA), and this makes the M-step tractable via the forward-backward algorithm.	there are many extensions such as factorial HMMs and Bayesian HMMs in which an expectation under p θ (t \| w) involves an intractable sum.	contrasting
train_20799	The other view is that the corpus loglikelihood is a sum over many terms of the form (2), one for each training sentence w, and we bound each summand individually using a different q φ .	neither view leads to a practical implementation in our setting.	contrasting

train_20700

This performance drop is statistically significant out of domain.

the difference between the Eisner and Attardi algorithms is not statistically significant out of domain.

contrasting

train_20701

Recent work has shown that the combination of base parsers at learning time, e.g., through stacking, yields considerable benefits (Nivre and McDonald, 2008;Attardi and Dell'Orletta, 2009).

it is unclear how these approaches compare against the simpler ensemble models, which combine parsers only at runtime.

contrasting

train_20702

SITGs have proven to be a powerful tool in Syntax Machine Translation.

the algorithms have been proposed do not explore all the possible parse trees.

contrasting

train_20703

Using the reranker features distributed with the Brown reranker (Charniak and Johnson, 2005), which we call the "standard" set below, we obtained no overall improvement in f-score when either reranking the Berkeley parser n-best lists alone, or when the Berkeley parses were combined with the Brown parses.

it is possible that these results reflect the fact that the features used by the reranker were chosen because they improve the Brown parser, i.e., they are the result of feature selection based on reranking the Brown parser's n-best lists.

contrasting

train_20704

It would lead to deemphasizing those unusual types of information that might be being discussed as part of a post.

one might expect that those things that are unusual types of information might actually be more likely to be the in-focus information within an initiation that responses may be likely to refer to.

contrasting

train_20705

Conversational Agents have been shown to be effective tutors in a wide range of educational domains.

these agents are often ignored and abused in collaborative learning scenarios involving multiple students.

contrasting

train_20706

In all three conditions, students go through the same task plan.

the degree of social performance is varied from minimal (Task) to ideal (Human).

contrasting

train_20707

On the right side of Table 2, we notice that the human tutors (H) were rated higher on being part of the team (Integration), being more liked, being friendlier and keeping the group more socially comfortable (T.Releasing).

the social tutors (S) were rated to be friendlier and were only marginally better at being seen as part of the team.

contrasting

train_20708

Language identification has traditionally been approached with character-level language models.

the language model approach crucially depends on the length of the text in question.

contrasting

train_20709

Both approaches benefit from their ability to consider large, flexible spans of source context when making transduction decisions.

they encode this context in different ways, providing their respective models with different information.

contrasting

train_20710

One approach to string transduction is to view it as a tagging problem where the input characters are tagged with the output characters.

since sounds are often represented by multicharacter units, the relationship between the input and output characters is often complex.

contrasting

train_20711

In the first 2 passes, we use a standard bi-gram LM to generate lattices, followed by a standard tri-gram LM rescoring of lattices.

in the third pass, we generate both lattices and N-best lists based on the same bi-gram LM.

contrasting

train_20712

provement of 0.8% (about 4.8% relative) compared to the standard tri-gram lattice rescoring.

we have 0.6% absolute improvement (about 3.7% relative) compared to the standard 4gram lattice rescoring.

contrasting

train_20713

(Zitouni et al., 2005) used Arabic morphologically segmented data and claimed to have very competitive results in ACE 2003 and ACE 2004 data.

(Benajiba et al., 2008) report good results for Arabic NER on ACE 2003, 2004and 2005 segmentation.

contrasting

train_20714

This is because the token itself becomes insignificant information to the classifier.

when only punctuation separation is performed (W ord s ), the data is significantly sparse and the obtained results achieves high F-measure (77.1) only when outputs of other classifiers are used.

contrasting

train_20715

Hence, we keep the edge to indicate that the child can be used as a preferred substitute for the parent.

the edge is removed if the ratio is small (less than a threshold t, see Fig.

contrasting

train_20716

Using aggressive pruning, the list size and number of redun-dant phrase patterns are greatly reduced.

the classification accuracy does not decrease.

contrasting

train_20717

The detection of pitch accents and phrase boundaries has received significantly more research attention than the classification of accent types and phrase ending behavior.

one technique that has been used in a number of research efforts is to simultaneously detect and classify pitch accent.

contrasting

train_20718

In both systems, the frame-based log posterior vector of P (phone|acoustics) over all phones is decorrelated using the Karhunen-Loeve (KL) transform; unlike MLPs, CRFs take into account the entire label sequence when computing local posteriors.

posterior estimates from the CRF tend to be overconfident compared to MLP posteriors (Morris and Fosler-Lussier, 2009).

contrasting

train_20719

Calculating these is computationally inexpensive for many simple tasks (such as classification and regression).

marginal and MAP inference tends to be expensive for complex structured prediction models (such as the joint information extraction models of ), making semisupervised learning intractable.

contrasting

train_20720

The softmax-margin approach offers (1) a convex objective, (2) the ability to incorporate task-specific cost functions, and (3) a probabilistic interpretation (which supports, e.g., hidden-variable learning and computation of posteriors).

max-margin training and MIRA do not provide (3); risk and JRB do not provide (1); and CLL does not support (2).

contrasting

train_20721

A drawback of this approach is that it is extremely local: while decisions can be based on complex structures on the left, they can look only at a few words to the right.

our algorithm builds a dependency tree by iteratively selecting the best pair of neighbours to connect at each parsing step.

contrasting

train_20722

Each performed action changes the partial struc-tures and with it the extracted features and the computed scores.

these changes are limited to a fixed local context around the attachment point of the action.

contrasting

train_20723

Shorter edges are arguably easier to predict, and our parses builds them early in time.

it is also capable of producing long dependencies at later stages in the parsing process.

contrasting

train_20724

(2009b) study dependency treebanks for nine languages and find that all dependency structures meet the mildly ill-nested condition in the dependency treebanks for some gap degree.

they do not report the maximum gap degree or parsing complexity.

contrasting

train_20725

This is certainly true when the graph is of high quality and all paths trustworthy.

in a graph constructed from web cooccurrence statistics, this is rarely the case.

contrasting

train_20726

Ideally we would see an order on such phrases, e.g., "more brittle" has a larger negative polarity than "brittle", which in turn has a larger negative polarity than "less brittle".

this is rarely the case and usually the adjective has the highest polarity magnitude.

contrasting

train_20727

In the approach, a subjective sentence is represented as a set of words in the sentence, ignoring word order and head-modifier relation between words.

sentiment classification is different from traditional topic-based text classification.

contrasting

train_20728

Recently, several methods have been proposed to cope with the problem (Zaenen, 2004;Ikeda et al., 2008).

these methods are based on flat bag-of-features representation, and do not consider syntactic structures which seem essential to infer the polarity of a whole sentence.

contrasting

train_20729

Titov and McDonald (2008b) underline the need for unsupervised methods for aspect detection.

according to the authors, existing topic models, such as standard Latent Dirichlet Allocation (LDA) (Blei et al., 2003), are not suited to the task of aspect detection in reviews, because they tend to capture global topics in the data, rather than rateable aspects pertinent to the review.

contrasting

train_20730

For example, for many verbs, nsubj tends to start a cause path and dobj a theme path.

for "bind" that signifies a Binding event, both lead to themes, as in "A binds B".

contrasting

train_20731

Several studies have been performed on identifying PICO elements in abstracts (Demner-Fushman and Lin, 2007;Hansen et al., 2008;Chung, 2009).

all of them are reporting coarsegrain (sentence-level) tagging methods that have not yet been shown to be sufficient for the purpose of IR.

contrasting

train_20732

Moreover, there is currently no standard test collection of questions in PICO structure available for evaluation.

the most critical aspect in IR is term weighting.

contrasting

train_20733

These approaches follow the assumption that the user knows where the most relevant information is located.

(Kamps et al., 2005) showed that it is preferable to use structure as a search hint, and not as a strict search requirement The second approach consists in integrating the document structure at the indexing step by introducing a structure weighting scheme (Wilkinson, 1994).

contrasting

train_20734

There has been several studies that cover the PICO extraction problem.

as far as we know, none of them analyses and uses the positional ditribution of these elements within the documents for the purpose of IR.

contrasting

train_20735

It could also be caused by the limited number of query in our test collection.

we can determine reasonable weights by tuning each part weight separately.

contrasting

train_20736

A straightforward idea is to detect PICO elements in documents and use the elements in the retrieval process.

this approach does not work well because of the difficulty to arrive at a consistent tagging of these elements.

contrasting

train_20737

Given a query, search engines retrieve relevant pictures by analyzing the image caption (if it exists), textual descriptions found adjacent to the image, and other text-related factors such as the file name of the image.

since they do not analyze the actual content of the images, search engines cannot be used to retrieve pictures from unannotated collections.

contrasting

train_20738

The probability of a document d in a corpus is defined as: Computing the posterior distribution P(θ, z|d, α, β) of the hidden variables given a document is intractable in general.

a variety of approximate inference algorithms have been proposed in the literature including variational inference which our model adopts .

contrasting

train_20739

All are almost surprisingly intuitive, but this is not terribly surprising since Chinese and English have very similar large-scale structures (both are head initial, both have adjectives and quantifiers that precede nouns).

we see two entries in the list (starred) that correspond to an En- glish word order that is ungrammatical in Chinese: PP modifiers in Chinese typically precede the VPs they modify, and CPs (relative clauses) also typically precede the nouns they modify.

contrasting

train_20740

4aThe VSO basic word order is evident: early in the sentence, there is a strong tendency towards right movement around arguments after covering the verb.

right movement is increasingly penalized at the end of the sentence.

contrasting

train_20741

The binarization in Figure 1 is unfortunate because the obtained production cannot be factorized such that only two nonterminals occur in each rule.

the binarization of Figure 4 allows the factorization into S(U , NP ) ↔ S(U , NP ) and U : @(NP , V ) ↔ @(V , NP ), which are fully binarized productions.

contrasting

train_20742

However, the binarization of Figure 4 allows the factorization into S(U , NP ) ↔ S(U , NP ) and U : @(NP , V ) ↔ @(V , NP ), which are fully binarized productions.

in general, STSGs (or SCFGs or extended tree transducers) cannot be fully binarized as shown in Aho and Ullman (1972).

contrasting

train_20743

Consequently, we cannot compute forward or backward applications for arbitrary MBOT.

if the MBOT is equivalent to an STSG (for example, because it was constructed by the method presented before Theorem 3), then forward and backward application can be computed essentially as for STSG.

contrasting

train_20744

MBOT can efficiently be used (with computational benefits) as an alternative representation for transformations computed by STSG (or compositions of STSG).

mBOT can also compute transformations, of which the domain or range cannot be represented by a TSG.

contrasting

train_20745

Both methods were trained and tested on data from the Ziff-Davis corpus (Knight and Marcu, 2002), and they achieved very similar grammaticality and meaning preservation scores, with no statistically significant difference.

their compression rates (counted in words) were very different: 70.37% for the noisy-channel method and 57.19% for the C4.5-based one.

contrasting

train_20746

Our method is similar to Nomoto's, in that it uses two stages, one that chops the source dependency tree generating candidate compressions, and one that ranks the candidates.

we experimented with more elaborate ranking models, and our method does not employ any manually crafted rules.

contrasting

train_20747

guage model trained on a large background corpus.

language models tend to assign smaller probabilities to longer sentences; therefore they favor short sentences, but not necessarily the most appropriate compressions.

contrasting

train_20748

Extrinsic evaluations have also shown that, while extractive summaries may be less coherent than human abstracts, users still find them to be valuable tools for browsing documents (He et al., 1999;.

these same evaluations also indicate that concise abstracts are generally preferred by users and lead to higher objective task scores.

contrasting

train_20749

The system described thus far may appear extractive in nature, as the transformation step is identifying informative sentences in the conversation.

these selected sentences correspond to < participant, relation, entity > triples in the ontology, for which we can subsequently generate novel text by creating linguistic annotations of the conversation ontology (Galanis and Androutsopolous, 2007).

contrasting

train_20750

This is due to the fact that the leading sentences for these two domains do not indicate any significance, hence the Lead system just behaves like Random.

for the scientific and newswire domains, the leading sentences do have importance so the Lead system consistently outperforms Random.

contrasting

train_20751

f MMR , is not guaranteed everywhere monotone.

our theoretical results still holds for f MMR with high probability in practice.

contrasting

train_20752

We assume that there is not enough high-quality data to build a monolingual selectional preference model for the source language (shown by dotted lines).

we can use a bilingual vector space, that is, a semantic space in which words of both the source and the target language are represented, to translate each source language word s into the target language by identifying its nearest (most similar) target word tr(s): Now we can use a target language selectional preference model to obtain plausibilities for source triples: where the superscript indicates the language.

contrasting

train_20753

Data Our experiments were carried out on the CoNLL 2008 (Surdeanu et al., 2008) training dataset which contains both verbal and nominal predicates.

we focused solely on verbal predicates, following most previous work on semantic role labeling (Màrquez et al., 2008).

contrasting

train_20754

Since different similarity functions can be used within this framework, one may wish to select the one that is the most appropriate or relevant to the task considered.

a crucial requirement for this choice to be realistic is to ensure that for the family of similarity functions considered the expected similarity maximization is efficiently computable.

contrasting

train_20755

As mentioned earlier, we leave this question to future work.

we can offer a brief look at how one could tackle this question.

contrasting

train_20756

The empirical adequacy of 2-SCFG models would presumably be lower with automatically-aligned texts and if the study also included non-European languages.

phrase-based systems can properly handle inside-out alignments when used with a reasonably large distortion limit, and all configurations in Fig.

contrasting

train_20757

6 Likewise, the procedure could be applied to statistical systems that only generate k-best lists.

we would not expect the same strong performance from model combination in these constrained settings.

contrasting

train_20758

Additionally, the experiments were performed over a large corpus of messages that are not available for use by other researchers.

we use messages from the widely-available Enron email corpus (Klimt and Yang, 2004) for our own experiments.

contrasting

train_20759

This strongly suggests that without zoning, the classifier is not learning features from the training set at a useful level of generality.

once we add the zoning classifier, the top-10 unigrams and bigrams appear to correspond much better with linguistic intuitions about the language of requests.

contrasting

train_20760

There is much disagreement about the units and elementary relations of discourse structure, but they agree that the structures are hierarchical, most commonly trees (Marcu, 2000), while others have argued for directed acyclic graphs (Danlos, 2004), or general graphs (Wolf and Gibson, 2004).

most of the segmentation research to date has focused on linear segmentation, in which segments are non-overlapping and sequential, and it has been argued that this sequence model is sufficient for many purposes (Hearst, 1994).

contrasting

train_20761

The mean scores for the BIN baseline are over 50% on the encyclopedia data.

the mean score for BIN on the Choi standard data (Fig.

contrasting

train_20762

For example, SkillSum (Williams and Reiter, 2008) and ICONOCLAST (Power et al., 2003) are two contemporary generation systems that allow for specifying aspects of style such as choice of discourse marker, clause order, repetition and sentence and paragraph lengths in the form of constraints that can be optimised.

to date, these systems do not consider syntactic reformulations of the type we are interested in.

contrasting

train_20763

This pulls the average for reformulated sentences down.

on average 2 out of 7 reformulations score quite high.

contrasting

train_20764

In general, one would expect that short transformation sequences to provide good evidence of true entailments.

to account for the grandparent-child relationship in the hypothesis, TED would produce a fairly long sequence, relabeling nearby to be near, deleting the two nodes for Rossville Blvd, and then reinserting those nodes under near.

contrasting

train_20765

Settings of 0.1, 0.2, 0.3, and 0.4 led to 10-fold cross-validation The main difference between our kernel and the CTK is that we sum over all pairs of subtrees (Equation 3).

the CTK only considers only one pair of subtrees.

contrasting

train_20766

9 In accuracy values that were not significantly different from each other.

we did observe that increased search failure ( §3.4) resulted from settings above 0.5.

contrasting

train_20767

Chawathe and Garcia-Molina (1997) describe a tree edit algorithm for detecting changes in structured documents that incorporates edits for moving subtrees and reordering children.

they make assumptions unsuitable for natural language, such as the absence of re-cursive syntactic rewrite rules.

contrasting

train_20768

Lexical-syntactic rules can be automatically extracted from plain corpora (e.g., (Lin and Pantel, 2001;Szpektor and Dagan, 2008)) but the quality (also in terms of little noise) and the coverage is low.

rules written at the semantic level are more accurate but their automatic design is difficult and so they are typically hand-coded for the specific phenomena.

contrasting

train_20769

by using FrameNet semantics (e.g., like in (Burchardt et al., 2007)), it is possible to encode a lexical-syntactic rule using the KILLING and the DEATH frames, i.e.

: to use this model, specific rules and a semantic role labeler on the specific corpora are needed.

contrasting

train_20770

Again, the syntactic rules (with variables) which this kernel BNC WN WIKI RTE2 0.55 0.42 0.83 RTE3 0.54 0.41 0.83 RTE5 0.45 0.34 0.82 Table 3: Coverage of the different resources for the words of the three datasets can provide are not enough general for RTE3.

maxSSTK+WOK improves WOK on all datasets thanks to its generalization ability.

contrasting

train_20771

This may be due to the fact that, as shown in Table 1, the overall differences between partners in mixed-gender pairs are quite low, and so neither partner may be doing much turn-byturn matching.

as we expected, entrainment is least prevalent among male-male pairs.

contrasting

train_20772

This reflects the intuition that someone overly eager to be liked may be perceived as annoying and socially inept.

similarity-attraction theory states that similarity promotes attraction, and someone might therefore entrain in order to obtain his partner's social approval.

contrasting

train_20773

As we expected, giving encouragement is correlated with entrainment for all three gender groups, and trying to be liked is correlated with entrainment for male-male and female-male groups.

trying to dominate is not correlated with entrainment on any feature, and conversation awkward is actually positively correlated with entrainment on jitter.

contrasting

train_20774

A high number is negative in that it is the sign of an inefficient dialogue, one which takes many turn exchanges to accomplish the objective.

it may also be the sign of easy, flowing dialogue between the partners.

contrasting

train_20775

Besides shell language, there were other annotations relevant to essay scoring.

we ignored them for this study because they are not directly relevant to the task of shell language detection.

contrasting

train_20776

For the annotated essay test set ( §3.2), the percentage of tokens tagged as shell was 14.0% (11.6% were labeled as shell by the first annotator).

the percentage of tokens tagged as shell was 4.2% for Lincoln-Douglas, 5.4% for Kennedy-Nixon, 4.6% for Gore-Bush, and 4.8% for Obama-McCain.

contrasting

train_20777

It is not completely clear whether the smaller percentages tagged as shell are due to a lack of coverage by the shell detector or more substantial differences in the domain.

it seems that these debates genuinely include less shell.

contrasting

train_20778

Also, Bush employs a somewhat atypical sentence structure here: "It's not what I think and its not my intentions and not my plan."

the system also incorrectly tagged sequences as shell, particularly in short sentences (e.g., "Are we as strong as we should be?").

contrasting

train_20779

Discourse markers, however, are typically only single words or short phrases that express a limited number of relationships.

shell can capture longer sequences that express more complex relationships between the components of an argumentative discourse (e.g., "But let's get back to the core issue here" signals that the following point is more important than the previous one).

contrasting

train_20780

In shell detection, we focus on the lexico-syntactic level, aiming to identify the bold words as shell.

work on argumentation schemes focuses at a higher level of abstraction, aiming to classify the sentence as an attempt to persuade by appealing to an external authority.

contrasting

train_20781

There have been many attempts over the last decade to develop model-based approaches to the phrase alignment problem (Marcu and Wong, 2002;Birch et al., 2006;DeNero et al., 2008;Blunsom et al., 2009).

most of these have met with limited success compared to the simpler heuristic method.

contrasting

train_20782

e 3 e 4 e 5 e 6 e 7 Because there is a sure link at a 48 , σ f 8 = [4, 4] does not include the possible link at a 38 .

f 7 only has possible links, so σ f 7 = [5, 6] is the span containing those.

contrasting

train_20783

DeNero and Klein 2010implicitly included these constraints in their representation: instead of sets of variables, they used a structured representation that only encodes triples (a, π, σ) satisfying both the mapping π = π(a) and the structural constraint that a can be generated by a block ITG grammar.

our inference procedure, BP, requires that we represent (a, π, σ) as an assignment of values to a set of variables.

contrasting

train_20784

The BP ITG model performs comparably to the Viterbi ITG model.

because posterior decoding permits explicit tradeoffs between precision and recall, it can do much better in the recallbiased measures, even though the Viterbi ITG model was explicitly trained to maximize F 5 (DeNero and Klein, 2010).

contrasting

train_20785

One practical solution is to restrict the output vocabulary to a short-list composed of the most frequent words (Schwenk, 2007).

the usual size of the short-list is under 20k, which does not seem sufficient to faithfully represent the translation models of section 2.

contrasting

train_20786

One solution would be full-blown transliteration (Knight and Graehl, 1998), followed by application of Jaro-Winkler.

transliteration systems are complex and require significant training resources.

contrasting

train_20787

11 Because a similar corpus did not exist for development, we split the evaluation corpus into development and test sections.

the usual method of splitting by document would not confine all mentions of each entity to one side of the split.

contrasting

train_20788

First, the model learns that high edit distance is predictive of a mismatch.

singleton strings that do not match often have a lower edit distance than longer strings that do match.

contrasting

train_20789

To our knowledge, Baron and Freedman (2008) reported the only previous results on the ACE2008 data set.

they only gave gold results for English, and clustered the entire evaluation corpus (test+development).

contrasting

train_20790

Within this work, specific affect definitions vary slightly with the intention of being coherent within the application and domain and being relevant to the specific adaptation goal (Martalo et al., 2008).

affective systems researchers generally agree that disengaged users show little involvement in the interaction, and often display facial, gestural and linguistic signals such as gaze avoidance, finger tapping, humming, sarcasm, et cetera.

contrasting

train_20791

All turns are used in the disengagement detection experiments described next.

only the training problem dialogues (360, 5 per student, 6044 student turns) are used for the performance analyses in Sections 6-7, because the final test problem was given after the instruments measuring performance (survey and posttest).

contrasting

train_20792

Compared to the results in Table 2, we find that the VSM and LSA methods are very robust to recognition errors, and we only observe slight correlation decreases on these features.

the decrease for the PMI-based method is quite large.

contrasting

train_20793

Our experimental results showed that all the features obtained good correlations with human proficiency scores if there are no recognition errors in the text transcripts, with the PMI-based method performing the best over three similarity measures.

if we used ASR transcripts, we observed a marked performance drop for the PMI-based method.

contrasting

train_20794

In our experiments, we noticed that shorter conversations suffer from poor classification.

the results from the above section appear to contradict this assertion, as a 30-word window can give very good performance.

contrasting

train_20795

Though the performance gain over the random ranker has shrunk considerably, there is still some utility in using the opening of a conversation to determine its ultimate duration.

it is clear predicting duration via conversation opening is a much more difficult task overall.

contrasting

train_20796

show that on certain synthetic-data problems, this frequentist training regimen significantly reduced test-data loss compared to approximate maximum likelihood estimation (MLE).

this method has not been evaluated on real-world problems until now.

contrasting

train_20797

Finally, we hypothesized that sum-product inference may produce more accurate results in certain cases as it allows more information about different parts of the model to be exchanged.

our results show that for these three problems, sum-product and max-product inference yield statistically indistinguishable results.

contrasting

train_20798

These computations are tractable for HMMs, since the distribution q(t) = p θ (t | w) that is optimal at the E-step (which makes the inequality tight) can be represented as a lattice (a certain kind of weighted DFA), and this makes the M-step tractable via the forward-backward algorithm.

there are many extensions such as factorial HMMs and Bayesian HMMs in which an expectation under p θ (t | w) involves an intractable sum.

contrasting

train_20799

The other view is that the corpus loglikelihood is a sum over many terms of the form (2), one for each training sentence w, and we bound each summand individually using a different q φ .

neither view leads to a practical implementation in our setting.

contrasting