tasksource/scinli · Datasets at Hugging Face

id stringlengths 7 12	sentence1 stringlengths 6 1.27k	sentence2 stringlengths 6 926	label stringclasses 4 values
train_21100	Both systems are rule-based, rather than data-driven, and do not train or test their systems with real-world portmanteaux.	to these approaches, this paper presents a data-driven model that accomplishes (2) by blending two given words into a portmanteau.	contrasting
train_21101	Since P (z, x, y) is relatively low compared to the step probabilities, this method prefers very short alignments-the reverse of the effect seen in the conditional method.	the model can also zero out the probabilities of unlikely aligments, so overall it learns fewer possible alignments between phonemes.	contrasting
train_21102	This "mixed" method, like the joint method, is more conservative in learning phoneme alignments.	like the conditional method, it has high alignment probabilities and prefers longer alignments.	contrasting
train_21103	Some hypotheses, like "architecology" and "japanglish," might even be considered superior to their gold counterparts.	some errors, like "js" and "mman," are clearly unacceptable system outputs.	contrasting
train_21104	Paraphrases are usually defined as "meaning equivalent" words or phrases.	many paraphrases, even while capturing the same meaning overall, display subtle differences which effect their substitutability (Gardiner and Dras, 2007).	contrasting
train_21105	A common theme across all these settings requires addressing two difficulties in linking decisions: matching the textual name mention to the form contained in the knowledge base, and using contextual clues to disambiguate similar entities.	all of these studies have focused on written language, while linking of spoken language remains untested.	contrasting
train_21106	al to identify explanatory text in candidate answers perform better than our approach, which relies solely on lexicalized alignment.	we expect that our two approaches are complementary, because they address different aspects of the QA task (structure vs. similarity).	contrasting
train_21107	These attempt to find parameter settings for which unnormalized model scores approximate the true label probability.	the theoretical properties of such techniques (and of self-normalization generally) have not been investigated.	contrasting
train_21108	This is somewhat surprising, since intuitively, ToBI should be capturing information very similar to what pauses and word durations capture, particularly when it is predicted based partially on these phenomena.	our learned ToBI predictor only gets roughly 50 F 1 on break prediction, so ToBI prediction is clearly a hard task even with sophisticated features.	contrasting
train_21109	The SIU-model has previously been applied to two datasets from the Pentomino domain , where the speaker's goal was to identify one out of a set of tetris-like (but consisting of five instead of four blocks) puzzle pieces.	in these datasets, the references were "one-shot" and not embedded in longer dialogues, as is the case in the REX corpus.	contrasting
train_21110	Traditionally, text categorization and NER approaches are based on textual information only.	information in visually rich formats such as PDF and HTML is often conveyed by a combination of textual and visual features.	contrasting
train_21111	As a result, the feature space of NER and text categorization involves purely textual features: word attributes and characteristics, their contexts and frequencies.	textual information in visually rich formats, such as PDF and HTML, is interlaced with typographic and other visually salient characteristics.	contrasting
train_21112	Flyers contain a single list-ing, which in turn has a single address.	broker information and space information are multivalue attributes.	contrasting
train_21113	Lastly, it should be noted that an overall system performance baseline is one that measures the average performance of data entry staff in commercial real estate listing services.	the terms and conditions of most listing services prohibit gathering and using data for such purposes.	contrasting
train_21114	If an English speaker was asked to simply choose one of the Chinese translations, they likely could not decide which is correct.	if they were additionally given English T2 translations corresponding to each of the Chinese translations, they could easily choose the third as the most natural, even without knowing a word of Chinese.	contrasting
train_21115	It can be noted that this formalism is a relatively simple expansion of standard SCFGs.	the additional targets require non-trivial modifications to the standard training and search processes, which we discuss in the following sections.	contrasting
train_21116	This limit is generally imposed by ordering rules by the phrase probability P (α 1 \|γ) and only using the top few (in our case, 10) for each source γ.	in the MSCFG case, this is not so simple.	contrasting
train_21117	One of the weaknesses of current supervised word sense disambiguation (WSD) systems is that they only treat a word as a discrete entity.	a continuous-space representation of words (word embeddings) can provide valuable information and thus improve generalization accuracy.	contrasting
train_21118	• Surrounding words: Additionally, the surrounding words of a target word (after removing stop words) are also used as features in IMS.	unlike POS tags, the words occurring in the immediately adjacent sentences are also included.	contrasting
train_21119	Thus the selected development sets are not proper representatives of the test sets and the tuning process results in overfitting the parameters (window sizes) to the development sets, with low generalization accuracy.	per-task tuning is relatively stable and performs better on the test sets.	contrasting
train_21120	For example, on SE2, the improvement achieved on verbs is much larger than the other two POS types and on SE3, adjectives benefited from word embeddings more than nouns and verbs.	this Finally, we evaluated the effect of word embeddings and the adaptation process.	contrasting
train_21121	For example, in the sentence "the winning goal came with less than a minute left to play 2 ", the sense tag for word 'goal' is 'goal%1:04:00::'.	the training data for IMS does not contain any sample with this sense tag and so it is impossible for IMS to assign this tag to any test instances.	contrasting
train_21122	Without a language model the greedy DBRNN decoding procedure loses relatively little in terms of CER as compared with the DBRNN+NN-3 model.	this 3% difference in CER translates to a 16% gap in WER on the full Eval2000 test set.	contrasting
train_21123	For transition-based dependency parsers, the feature context for a parsing state is represented by the neighboring elements of a word token in the stack containing the partial parse or the buffer containing unprocessed word tokens.	in our treeto graph parser, as already stated, buffers σ and β only specify which arc or node is to be examined next.	contrasting
train_21124	(7)) are similar to the LP-MERT constraints (5), although with the addition of slack variables and the ∆ function to handle infeasible solutions.	if a feasible solution is available for MIRA, then these extra quantities are unnecessary.	contrasting
train_21125	However, if we consider the normal fan as a whole we can clearly see that w ∈ N {h i } is the optimal point under the regularisation.	it is not obvious in the projected parameter space thatŵ is the better choice.	contrasting
train_21126	Some systems focus on syntactic generation (Bangalore and Rambow, 2000;Langkilde-Geary, 2002;Filippova and Strube, 2008) or linearization and inflection (Filippova and Strube, 2007;He et al., 2009;Wan et al., 2009;Guo et al., 2011a), and avoid thus the need to cope with this projection all together; some use a rule-based module to handle the projection between non-isomorphic structures (Knight and Hatzivassiloglou, 1995;Langkilde and Knight, 1998;Bohnet et al., 2011); and some adapt the meaning structures to be isomorphic with syntactic structures (Bohnet et al., 2010).	it is obvious that a "syntacticization" of meaning structures can be only a temporary workaround and that a rule-based module raises the usual questions of coverage, maintenance and portability.	contrasting
train_21127	Such research on temporal text analysis generally focuses on determining when events start and end or how they relate temporally to each other; specific goals include information extraction of time-dependent facts from news media (Ling and Weld, 2010;Talukdar et al., 2012), or extracting personal histories in social media (Wen et al., 2013).	our goal is to find the temporal orientation of people.	contrasting
train_21128	For instance, Twitterspecific tokens (e.g., retweets) are often removed during normalization, so examining the removal of these words as a group is warranted.	these tokens are never added, so different segmentation is appropriate when examining word addition.	contrasting
train_21129	Note that since the first measure is one of overall performance, smaller numbers reflect larger performance drops when removing a given type of edit, so that the smaller the number the more critical the need to perform the given type of normalization.	the latter judgment is one of error rate, and thus interpretation is reversed; the larger the error rate when it is removed, the more critical the normalization edit.	contrasting
train_21130	Perhaps unsurprisingly, failing to add subjects and verbs resulted in the largest issues, as the parser has little chance of identifying these dependencies if the terms simply do not appear in the sentence.	not all word additions proved crit- ical, as failing to add in a missing determiner generally had little impact on the overall performance.	contrasting
train_21131	Similar to those on dependency parsing, the results on speech synthesis suggest that a broad approach that considers several different types of normalization edit is necessary to produce results comparable to those seen on clean text.	at a high level there is a clear divide in importance between normalization types, where the greatest performance gains can be obtained by focusing on the comparatively small number of token removals.	contrasting
train_21132	While normalization for speech synthesis is primarily dependent on removing unknown tokens, normalization that targets name entity recognition would be better served focusing on replacing non-standard to-kens with their standard forms.	parsertargeted normalization must attend to both of the tasks, as well as the task of restoring dropped tokens.	contrasting
train_21133	Alternatively, you may argue in favor of the death penalty because it gives victims of the crimes closure.	you may argue against the death penalty because some innocent people will be wrongfully executed or because it is a cruel and unusual punishment.	contrasting
train_21134	For example, one author may believe that the death penalty is a cruel and unusual punishment while the other one attacks that position.	in order to attack that position they must be discussing the same facet.	contrasting
train_21135	For the first batch we randomly selected 500 pairs from our pairs dataset of 1131 pairs.	our subsequent impression was that the clustering had not filtered out enough of the unrelated pairs (score 0-1).	contrasting
train_21136	This biases the distribution of the training set to having a much larger set of more similar pairs, which has been a problem for previous work (Boltuzic andŠnajder, 2014), where the vast majority of pairs that were labelled were unrelated.	the AFS task is clearly different than STS, partly because the data is dialogic and partly because it is argumentative.	contrasting
train_21137	Another parallel may exist between work on nuclearity in RST and its use in summarization (Marcu, 1999).	our notion of a CENTRAL PROPOSITION is different than nuclearity in RST, since FACETS are derived from CEN-TRAL PROPOSITIONS that rise to the top of the pyramid across summarizers, and then (via AFS) across many dialogs on a topic, while RST nuclearity is only defined for a span of text by a single speaker.	contrasting
train_21138	Emotional information has been observed even in summaries of professional chats discussing technology (Zhou and Hovy, 2005).	the instructions to our Pyramid annotators were to not include information of this type in the pyramids.	contrasting
train_21139	Even though LwR provides huge benefits, providing both a label and a rationale is expected to take more time of the labeler than simply providing a label.	the improvements of LwR over Lw/oR is so huge that it might be worth spending the extra time in providing rationales.	contrasting
train_21140	Context representations are a key element in distributional models of word meaning.	to typical representations based on neighboring words, a recently proposed approach suggests to represent a context of a target word by a substitute vector, comprising the potential fillers for the target word slot in that context.	contrasting
train_21141	Finally, we rank the candidates using the scores in our in-context paraphrase vectors from Section 6.2.	this time we check the effect of injecting a stronger bias towards the given context c, by averaging only the top-m percent contexts most similar to c, for m ∈ {1%, 5%, 10%, 100%}, as described in Section 4.2.	contrasting
train_21142	Then, while processing all instances of the same word type u one after the other, only the substitute vectors in C u need to be loaded into memory.	in an 'online' mode, to be ready for any arbitrary word instance input, our model would need to keep in memory substitute vectors for all the word types in the vocabulary V .	contrasting
train_21143	(2013) use the Web as their corpus, and de Melo & Bansal use Google N-grams (Brants and Franz, 2006).	this results in a large number of instances where satisfied lexical patterns do not correspond to adjectives (e.g., sometimes but not always).	contrasting
train_21144	They make use of standard context vectors for clustering adjectives, where context for every adjective comprises of nouns it modifies across all sentences in a corpus.	recent work shows promise for context vectors embedded in a compressed semantic space that are derived using neural networks: Baroni et al.	contrasting
train_21145	Given a window size w, the CBOW model predicts the current word given the neighboring words as context.	the skip-gram model predicts the neighboring words given the current word.	contrasting
train_21146	Zero pronoun resolution, like general pronoun resolution, is almost universally approached as a problem of linking a pronoun to an overt noun phrase antecedent in the text.	while some zero pronouns do have overt noun phrase antecedents, many other zero pronouns do not.	contrasting
train_21147	3h 9) Teacher Ø 还在面试阶段吗 Are you still in the interview phase?	2v there are plenty of utterances (e.g., Table 1 lines 4 and 6) in which the English translation does not contain an overt subject.	contrasting
train_21148	In most previous work on KB completion to predict missing relation facts (Mintz et al., 2009;, the methods are evaluated on a subset of facts from a single KB snapshot, that are hidden while training.	given that the missing entries are usually selected randomly, the distribution of the selected unknown entries could be very different from the actual missing facts distribution.	contrasting
train_21149	One might expect that with the increased number of types, the embedding model would perform better than the classifier since they share parameters across types.	despite the recent popularity of embedding models in NLP, linear model still performs better in our task.	contrasting
train_21150	The pronunciation of Spanish words is recoverable from the spelling by applying a limited set of rules (Kominek and Black, 2006).	there is some ambiguity in the opposite direction; for example, the phoneme [b] can be expressed with either 'b' or 'v'.	contrasting
train_21151	For example, the word amputate is listed as monomorphemic, but in fact contains the suffix -ate.	amputee is analyzed as amputee = amputate − ate + ee.	contrasting
train_21152	The LCCT method achieves a relative high accuracy with 10% of the reviews labeled, better than SVM, TSVM and Self-learning with 100% of the reviews labeled.	when all the training data are labeled, LCCT is still significantly more accurate than all the competitors except Nguyen's method.	contrasting
train_21153	Various other researchers have tried to improve the performance of their paraphrase systems or vector space models by using diverse sources of information such as bilingual corpora (Bannard and Callison-Burch, 2005;Huang et al., 2012;Zou et al., 2013), 10 structured datasets (Yu and Dredze, 2014; or even tagged images (Bruni et al., 2012).	most previous work 11 did not adopt the general, simplifying view that all of these sources of data are just cooccurrence statistics coming from different sources with underlying latent factors.	contrasting
train_21154	We know that whenever P > 0.5, the error rate decreases (and therefore Acc increases) so the text is improved.	3 an increase in P , R or F alone does not necessarily imply an increase in Acc or WAcc, as illustrated in Table 8.	contrasting
train_21155	Therefore, we must accept that any metric used in such scenarios will not be perfect.	it is worth noting that this limitation does not extend to evaluation of error detection per se using such metrics.	contrasting
train_21156	Recording speech poses no particular problems, but retrieval of spoken content using spoken queries is presently available only for the approximately two dozen languages in which there is an established path to market; English, German, or Chinese, for example.	many of the mobile-only users who could benefit most from such systems speak only one of the several hundred other languages that each have at least a million speakers; 1 Balochi, Mossi or Quechua, for example.	contrasting
train_21157	Without multiple draws, confidence intervals on this value cannot be established.	we are confident that random baselines even as high as 0.1 for any of our measures would be surprising.	contrasting
train_21158	Thus, even if we somehow adapt ABA's product-of-marginals heuristic to such models, we run the risk of estimating highly inaccurate posteriors (specifically, zero-valued posteriors).	mIR extends to all IBm-style word alignment models and does not add heuristics.	contrasting
train_21159	On both language pairs, ABA, PostCAT and MIR outperform their respective EM baseline with comparable gains overall.	we noticed that ABA and MIR are not producing the same alignments.	contrasting
train_21160	For evaluating slot induction (AP and PR-AUC), the double-graph random walk (row (e)) performs better on both ASR and manual results, which implies that additionally integrating the lexical knowledge graph helps decide a more coherent and complete slot set since we can model the score propagation more precisely (not only slot-level but wordlevel information).	for SLU evaluation (WAP and AF), the single-graph random walk (row (c)) performs better, which may imply that the slots carrying the coherent relations from the row (e) may not have good semantic decoder performance so that the performance is decreased a little.	contrasting
train_21161	The difficulty lies in the manner of distinguishing paraphrases from expressions that stand in different semantic relations, e.g., antonyms and sibling words, using only the statistics estimated from such corpora.	highly accurate paraphrases can be extracted from parallel or comparable corpora, but their coverage is limited owing to the limited availability of such corpora for most languages.	contrasting
train_21162	(3) a. amendment of regulation ⇔ amending regulation b. amendment of X ⇔ amending X c. amendment of documents ⇔ amending documents Using that method, they were able to expand the seed lexicon by a large multiple (15 to 40 times), and the new paraphrase pairs were of reasonably good quality.	they introduced variables only for identical word forms shared by both sides of each pair and left corresponding pairs of lexical variants, e.g., ("amendment", "amending") in (3a), untouched.	contrasting
train_21163	However, these scores are reasonably high, considering that no use is made of rich language-specific resources 12 .	more grammatical errors occurred than with S Seed and S ID .	contrasting
train_21164	Training in machine learning often uses starting big which is to use up all training data at the same time.	elman (1993) suggests that in some cases, learning should start by training simple models on small data and then gradually increase the model complexity and add more difficult data.	contrasting
train_21165	MaxEnc strategy helps the system achieve the highest accuracy on short sentences (up to length 10).	it is less helpful than Mi-nEnc when performing on long sentences.	contrasting
train_21166	It shows that lexical semantics plays a decisive role in the performance of the system.	it is worth noting that, even without that knowledge (i.e., with the ∞-order generative model alone), the DDA of phase 1 is 2% higher than before being trained (66.89% vs 64.9%).	contrasting
train_21167	The two-registers system is not quite arc-decomposable (Goldberg and Nivre, 2013): if the wrong vertex is stored in a register then a later pair of crossed arcs might both be individually reachable but not jointly reachable.	there may be a "crossing-sensitive" variant of arcdecomposability that takes into account the vertices crossed arcs are incident to that would apply here.	contrasting
train_21168	Xiao and Guo (2013) induce word embeddings across multiple domains, and concatenate these representations into a single feature vector for labeled instances in each domain, following EasyAdapt (Daumé III, 2007).	they do not apply this idea to unsupervised domain adaptation, and do not work in the structured feature setting that we consider here.	contrasting
train_21169	Our formulation is related to multi-domain learning, particularly in the multiattribute setting.	rather than partitioning all instances into domains, the domain attribute formulation allows information to be shared across instances which share metadata attributes.	contrasting
train_21170	We were surprised that SG-EM+RETRO actually performed worse than SG-EM, given how poorly SG-EM performed in the other evaluations.	an analysis again revealed that this was due to the kind of similarity encouraged by WordNet rather than an inability of the model to learn useful vectors.	contrasting
train_21171	(2014) introduce an approach that chooses between multiple composition tensors (AdaMC-RNTN), which yields further gains with respect to RNTN performance.	to the lexicalized and highdimensional RNTN model, there are several lines of work that attempt to work in a more data-scarce setting.	contrasting
train_21172	Since Brown clusters are mostly syntactic/semantic in nature and do not automatically distinguish positive or negative sentiment, we additionally performed multiple experiments to use clusters while incorporating additional sentiment information: On one hand, we try to incorporate the judgements on the Amazon near-domain dataset more directly into the clusters by using the repeated bisecting K-Means algorithm as implemented in CLUTO (Zhao and Karypis, 2005), with previous/next word, part-of-speech tag, and the score of the containing review as features.	we split the Brown clusters according to the sentiment value that they have in a particular sentiment lexicon (e.g.	contrasting
train_21173	The similarity is computed based on statistics such as co-occurrence which are unable to accommodate the subtlety that whether two words labeled as similar are truly similar depends on which topic they appear in, as explained by the aforementioned examples.	ideally, the knowledge would be word A and B are similar under topic C. in reality, we only know two words are similar, but not under which topic.	contrasting
train_21174	DF-LDA and Quad-LDA proposed to use word correlations to enhance the coherence of learned topics.	they improperly enforce words labeled as similar to have similar probabilities in all topics, which violates the fact that whether two words are similar depend on which topic they appear in.	contrasting
train_21175	Topic models provide insights into document collections, and their supervised extensions also capture associated document-level metadata such as sentiment.	inferring such models from data is often slow and cannot scale to big data.	contrasting
train_21176	2 Typically, the topics are discovered through a process of probabilistic inference, either variational EM (Wang et al., 2009) or Gibbs sampling (Boyd-Graber and Resnik, 2010).	these methods scale poorly to large datasets.	contrasting
train_21177	While these advancements improve the scalability of max-margin supervised topic models, the improvement is limited by the fact that the sampling algorithm grows with the number of tokens.	this paper explores a different vein of research that focuses on using efficient representations of summary statistics to estimate statistical models.	contrasting
train_21178	Traditional information extraction would be content with extracting two binary relation instances (NEG-REG,BCL,RFLAT) and (NEG-REG,IL-10,RFLAT), where NEG-REG represents a negative regulation (i.e., inhibition).	the sentence also discloses important contextual information, i.e., BCL regulates RFLAT by stimulating the inhibitive effect of IL-10, and likewise the inhibition of RFLAT by IL-10 is controlled by BCL.	contrasting
train_21179	Complex knowledge extraction can be naturally framed as a semantic parsing problem, with the event structure represented by a semantic parse; see Figure 2.	annotating example sentences is expensive and time-consuming.	contrasting
train_21180	(2014) took an important step toward this direction, by learning a semantic parser based on combinatorial categorial grammar (CCG) from Freebase and web sentences.	krishnamurthy and Mitchell (2012) still learned from binary relations, using only simple sentences (of length ten or less).	contrasting
train_21181	Automatic sentiment analysis of text, especially social media posts, has a number of applications in commerce, public health, and public policy development.	a vast majority of prior research on automatic sentiment analysis has been on English texts.	contrasting
train_21182	Thus, instead of building source-language specific sentiment analysis systems, one can translate the texts into English and use an English sentiment analysis system.	it is widely believed that aspects of sentiment may be lost in translation, especially in automatic translation.	contrasting
train_21183	This suggests that certain attributes of automatically translated text mislead humans with regards to the true sentiment of the source text.	these same attributes do not seem to affect the automatic sentiment analysis system as much.	contrasting
train_21184	The LM score features are extracted from large amounts of news article data, and are good representation of the general importance of bigrams for the test domain.	word-Net information is collected from a more general aspect, which may not be a very good choice for this task.	contrasting
train_21185	By the constraints of the algorithm, a head word x h must combine with each of its left and right dependents.	the order of combination can lead to different tree structures (as illustrated in Figure 2).	contrasting
train_21186	The scoring function requires specifying a set of parse features f which, in theory, could be directly adapted from existing lexicalized c-parsers.	the structure of the dependency parse greatly limits the number of decisions that need to be made, and allows for a smaller set of features.	contrasting
train_21187	This task effectively amounts to disambiguating the sense of discourse connective, which can be done with high accuracy (Pitler et al., 2008).	in the absence of an explicit discourse connective, inferring the sense of a discourse relation has proved to a very challenging task (Park and Cardie, 2012;Rutherford and Xue, 2014).	contrasting
train_21188	The common approach has been to inject knowledge as features.	these pieces of knowledge provide relatively strong evidence that loses impact in standard training due to sparsity.	contrasting
train_21189	As hard coreference problems are rare in standard coreference datasets, we do not have significant performance improvement.	these results show that our additional Predicate Schemas do not harm the predictions for regular mentions.	contrasting
train_21190	With the development of machine learning based models (Connolly et al., 1994;Soon et al., 2001b;Ng and Cardie, 2002a), attention shifted to solving standard coreference resolution problems.	many hard coreference problems involve pronouns.	contrasting
train_21191	Mnih and Teh (2012) fix these parameters to 1 and obtain the same perplexities, thereby circumventing the need for explicit normalisation.	this method does not provide any guarantees that the models are normalised at test time.	contrasting
train_21192	One option for speeding up factored models is using a GPU to perform the vector-matrix operations.	gPU integration is architecture specific and thus against our goal of making our language modelling toolkit usable by everyone.	contrasting
train_21193	We obtain a slightly better BLEU score with stochastic gradient descent, but this is likely to be just noise from tuning the translation system with MERT.	noise contrastive training reduces training time by a factor of 7.	contrasting
train_21194	As a first attempt to automate the test, we only experiment with basic linguistic features.	we believe that the task itself offers an opportunity for the development of-and subsequent evaluation of-rich linguistic features that may be better equipped for determining the aboutness of conversations.	contrasting
train_21195	A portion of user utterances refer to general Java knowledge, and in these cases semantic interpretation can be accomplished by mapping to a domain-specific ontology (e.g., Dzikovska et al., 2007).	many utterances refer to concrete entities within the dynamically changing, user-created programing artifact.	contrasting
train_21196	Importantly, this value does not test generalization to unseen questions, since KNOWBOT has held dialogs on these questions.	it does show that our system can effectively learn about its domain: a poor dialog extraction system will fail to extract any helpful knowledge from users during a training dialog.	contrasting
train_21197	Our open system acquires relations from a wide variety of user explanations without the bottleneck of a hand-built dialog model, but the tradeoff is that we use relatively simple, templated system prompts.	our collected corpus of real human-system dialogs can be used to improve our system in further iterations.	contrasting
train_21198	This is to be expected, as the speech in the control group has more complete sentences and fewer disfluencies.	it is interesting to note that performance on the PNFA and SD groups is not much worse.	contrasting
train_21199	All the features relating to Yngve depth and height of the parse trees are significantly different (in at least one of the three clinical groups).	of the eight primary syntactic units calculated by Lu's SCA, six show no significant difference when measured on the automatically segmented transcripts.	contrasting

train_21100

Both systems are rule-based, rather than data-driven, and do not train or test their systems with real-world portmanteaux.

to these approaches, this paper presents a data-driven model that accomplishes (2) by blending two given words into a portmanteau.

contrasting

train_21101

Since P (z, x, y) is relatively low compared to the step probabilities, this method prefers very short alignments-the reverse of the effect seen in the conditional method.

the model can also zero out the probabilities of unlikely aligments, so overall it learns fewer possible alignments between phonemes.

contrasting

train_21102

This "mixed" method, like the joint method, is more conservative in learning phoneme alignments.

like the conditional method, it has high alignment probabilities and prefers longer alignments.

contrasting

train_21103

Some hypotheses, like "architecology" and "japanglish," might even be considered superior to their gold counterparts.

some errors, like "js" and "mman," are clearly unacceptable system outputs.

contrasting

train_21104

Paraphrases are usually defined as "meaning equivalent" words or phrases.

many paraphrases, even while capturing the same meaning overall, display subtle differences which effect their substitutability (Gardiner and Dras, 2007).

contrasting

train_21105

A common theme across all these settings requires addressing two difficulties in linking decisions: matching the textual name mention to the form contained in the knowledge base, and using contextual clues to disambiguate similar entities.

all of these studies have focused on written language, while linking of spoken language remains untested.

contrasting

train_21106

al to identify explanatory text in candidate answers perform better than our approach, which relies solely on lexicalized alignment.

we expect that our two approaches are complementary, because they address different aspects of the QA task (structure vs. similarity).

contrasting

train_21107

These attempt to find parameter settings for which unnormalized model scores approximate the true label probability.

the theoretical properties of such techniques (and of self-normalization generally) have not been investigated.

contrasting

train_21108

This is somewhat surprising, since intuitively, ToBI should be capturing information very similar to what pauses and word durations capture, particularly when it is predicted based partially on these phenomena.

our learned ToBI predictor only gets roughly 50 F 1 on break prediction, so ToBI prediction is clearly a hard task even with sophisticated features.

contrasting

train_21109

The SIU-model has previously been applied to two datasets from the Pentomino domain , where the speaker's goal was to identify one out of a set of tetris-like (but consisting of five instead of four blocks) puzzle pieces.

in these datasets, the references were "one-shot" and not embedded in longer dialogues, as is the case in the REX corpus.

contrasting

train_21110

Traditionally, text categorization and NER approaches are based on textual information only.

information in visually rich formats such as PDF and HTML is often conveyed by a combination of textual and visual features.

contrasting

train_21111

As a result, the feature space of NER and text categorization involves purely textual features: word attributes and characteristics, their contexts and frequencies.

textual information in visually rich formats, such as PDF and HTML, is interlaced with typographic and other visually salient characteristics.

contrasting

train_21112

Flyers contain a single list-ing, which in turn has a single address.

broker information and space information are multivalue attributes.

contrasting

train_21113

Lastly, it should be noted that an overall system performance baseline is one that measures the average performance of data entry staff in commercial real estate listing services.

the terms and conditions of most listing services prohibit gathering and using data for such purposes.

contrasting

train_21114

If an English speaker was asked to simply choose one of the Chinese translations, they likely could not decide which is correct.

if they were additionally given English T2 translations corresponding to each of the Chinese translations, they could easily choose the third as the most natural, even without knowing a word of Chinese.

contrasting

train_21115

It can be noted that this formalism is a relatively simple expansion of standard SCFGs.

the additional targets require non-trivial modifications to the standard training and search processes, which we discuss in the following sections.

contrasting

train_21116

This limit is generally imposed by ordering rules by the phrase probability P (α 1 |γ) and only using the top few (in our case, 10) for each source γ.

in the MSCFG case, this is not so simple.

contrasting

train_21117

One of the weaknesses of current supervised word sense disambiguation (WSD) systems is that they only treat a word as a discrete entity.

a continuous-space representation of words (word embeddings) can provide valuable information and thus improve generalization accuracy.

contrasting

train_21118

• Surrounding words: Additionally, the surrounding words of a target word (after removing stop words) are also used as features in IMS.

unlike POS tags, the words occurring in the immediately adjacent sentences are also included.

contrasting

train_21119

Thus the selected development sets are not proper representatives of the test sets and the tuning process results in overfitting the parameters (window sizes) to the development sets, with low generalization accuracy.

per-task tuning is relatively stable and performs better on the test sets.

contrasting

train_21120

For example, on SE2, the improvement achieved on verbs is much larger than the other two POS types and on SE3, adjectives benefited from word embeddings more than nouns and verbs.

this Finally, we evaluated the effect of word embeddings and the adaptation process.

contrasting

train_21121

For example, in the sentence "the winning goal came with less than a minute left to play 2 ", the sense tag for word 'goal' is 'goal%1:04:00::'.

the training data for IMS does not contain any sample with this sense tag and so it is impossible for IMS to assign this tag to any test instances.

contrasting

train_21122

Without a language model the greedy DBRNN decoding procedure loses relatively little in terms of CER as compared with the DBRNN+NN-3 model.

this 3% difference in CER translates to a 16% gap in WER on the full Eval2000 test set.

contrasting

train_21123

For transition-based dependency parsers, the feature context for a parsing state is represented by the neighboring elements of a word token in the stack containing the partial parse or the buffer containing unprocessed word tokens.

in our treeto graph parser, as already stated, buffers σ and β only specify which arc or node is to be examined next.

contrasting

train_21124

(7)) are similar to the LP-MERT constraints (5), although with the addition of slack variables and the ∆ function to handle infeasible solutions.

if a feasible solution is available for MIRA, then these extra quantities are unnecessary.

contrasting

train_21125

However, if we consider the normal fan as a whole we can clearly see that w ∈ N {h i } is the optimal point under the regularisation.

it is not obvious in the projected parameter space thatŵ is the better choice.

contrasting

train_21126

Some systems focus on syntactic generation (Bangalore and Rambow, 2000;Langkilde-Geary, 2002;Filippova and Strube, 2008) or linearization and inflection (Filippova and Strube, 2007;He et al., 2009;Wan et al., 2009;Guo et al., 2011a), and avoid thus the need to cope with this projection all together; some use a rule-based module to handle the projection between non-isomorphic structures (Knight and Hatzivassiloglou, 1995;Langkilde and Knight, 1998;Bohnet et al., 2011); and some adapt the meaning structures to be isomorphic with syntactic structures (Bohnet et al., 2010).

it is obvious that a "syntacticization" of meaning structures can be only a temporary workaround and that a rule-based module raises the usual questions of coverage, maintenance and portability.

contrasting

train_21127

Such research on temporal text analysis generally focuses on determining when events start and end or how they relate temporally to each other; specific goals include information extraction of time-dependent facts from news media (Ling and Weld, 2010;Talukdar et al., 2012), or extracting personal histories in social media (Wen et al., 2013).

our goal is to find the temporal orientation of people.

contrasting

train_21128

For instance, Twitterspecific tokens (e.g., retweets) are often removed during normalization, so examining the removal of these words as a group is warranted.

these tokens are never added, so different segmentation is appropriate when examining word addition.

contrasting

train_21129

Note that since the first measure is one of overall performance, smaller numbers reflect larger performance drops when removing a given type of edit, so that the smaller the number the more critical the need to perform the given type of normalization.

the latter judgment is one of error rate, and thus interpretation is reversed; the larger the error rate when it is removed, the more critical the normalization edit.

contrasting

train_21130

Perhaps unsurprisingly, failing to add subjects and verbs resulted in the largest issues, as the parser has little chance of identifying these dependencies if the terms simply do not appear in the sentence.

not all word additions proved crit- ical, as failing to add in a missing determiner generally had little impact on the overall performance.

contrasting

train_21131

Similar to those on dependency parsing, the results on speech synthesis suggest that a broad approach that considers several different types of normalization edit is necessary to produce results comparable to those seen on clean text.

at a high level there is a clear divide in importance between normalization types, where the greatest performance gains can be obtained by focusing on the comparatively small number of token removals.

contrasting

train_21132

While normalization for speech synthesis is primarily dependent on removing unknown tokens, normalization that targets name entity recognition would be better served focusing on replacing non-standard to-kens with their standard forms.

parsertargeted normalization must attend to both of the tasks, as well as the task of restoring dropped tokens.

contrasting

train_21133

Alternatively, you may argue in favor of the death penalty because it gives victims of the crimes closure.

you may argue against the death penalty because some innocent people will be wrongfully executed or because it is a cruel and unusual punishment.

contrasting

train_21134

For example, one author may believe that the death penalty is a cruel and unusual punishment while the other one attacks that position.

in order to attack that position they must be discussing the same facet.

contrasting

train_21135

For the first batch we randomly selected 500 pairs from our pairs dataset of 1131 pairs.

our subsequent impression was that the clustering had not filtered out enough of the unrelated pairs (score 0-1).

contrasting

train_21136

This biases the distribution of the training set to having a much larger set of more similar pairs, which has been a problem for previous work (Boltuzic andŠnajder, 2014), where the vast majority of pairs that were labelled were unrelated.

the AFS task is clearly different than STS, partly because the data is dialogic and partly because it is argumentative.

contrasting

train_21137

Another parallel may exist between work on nuclearity in RST and its use in summarization (Marcu, 1999).

our notion of a CENTRAL PROPOSITION is different than nuclearity in RST, since FACETS are derived from CEN-TRAL PROPOSITIONS that rise to the top of the pyramid across summarizers, and then (via AFS) across many dialogs on a topic, while RST nuclearity is only defined for a span of text by a single speaker.

contrasting

train_21138

Emotional information has been observed even in summaries of professional chats discussing technology (Zhou and Hovy, 2005).

the instructions to our Pyramid annotators were to not include information of this type in the pyramids.

contrasting

train_21139

Even though LwR provides huge benefits, providing both a label and a rationale is expected to take more time of the labeler than simply providing a label.

the improvements of LwR over Lw/oR is so huge that it might be worth spending the extra time in providing rationales.

contrasting

train_21140

Context representations are a key element in distributional models of word meaning.

to typical representations based on neighboring words, a recently proposed approach suggests to represent a context of a target word by a substitute vector, comprising the potential fillers for the target word slot in that context.

contrasting

train_21141

Finally, we rank the candidates using the scores in our in-context paraphrase vectors from Section 6.2.

this time we check the effect of injecting a stronger bias towards the given context c, by averaging only the top-m percent contexts most similar to c, for m ∈ {1%, 5%, 10%, 100%}, as described in Section 4.2.

contrasting

train_21142

Then, while processing all instances of the same word type u one after the other, only the substitute vectors in C u need to be loaded into memory.

in an 'online' mode, to be ready for any arbitrary word instance input, our model would need to keep in memory substitute vectors for all the word types in the vocabulary V .

contrasting

train_21143

(2013) use the Web as their corpus, and de Melo & Bansal use Google N-grams (Brants and Franz, 2006).

this results in a large number of instances where satisfied lexical patterns do not correspond to adjectives (e.g., sometimes but not always).

contrasting

train_21144

They make use of standard context vectors for clustering adjectives, where context for every adjective comprises of nouns it modifies across all sentences in a corpus.

recent work shows promise for context vectors embedded in a compressed semantic space that are derived using neural networks: Baroni et al.

contrasting

train_21145

Given a window size w, the CBOW model predicts the current word given the neighboring words as context.

the skip-gram model predicts the neighboring words given the current word.

contrasting

train_21146

Zero pronoun resolution, like general pronoun resolution, is almost universally approached as a problem of linking a pronoun to an overt noun phrase antecedent in the text.

while some zero pronouns do have overt noun phrase antecedents, many other zero pronouns do not.

contrasting

train_21147

3h 9) Teacher Ø 还在面试阶段吗 Are you still in the interview phase?

2v there are plenty of utterances (e.g., Table 1 lines 4 and 6) in which the English translation does not contain an overt subject.

contrasting

train_21148

In most previous work on KB completion to predict missing relation facts (Mintz et al., 2009;, the methods are evaluated on a subset of facts from a single KB snapshot, that are hidden while training.

given that the missing entries are usually selected randomly, the distribution of the selected unknown entries could be very different from the actual missing facts distribution.

contrasting

train_21149

One might expect that with the increased number of types, the embedding model would perform better than the classifier since they share parameters across types.

despite the recent popularity of embedding models in NLP, linear model still performs better in our task.

contrasting

train_21150

The pronunciation of Spanish words is recoverable from the spelling by applying a limited set of rules (Kominek and Black, 2006).

there is some ambiguity in the opposite direction; for example, the phoneme [b] can be expressed with either 'b' or 'v'.

contrasting

train_21151

For example, the word amputate is listed as monomorphemic, but in fact contains the suffix -ate.

amputee is analyzed as amputee = amputate − ate + ee.

contrasting

train_21152

The LCCT method achieves a relative high accuracy with 10% of the reviews labeled, better than SVM, TSVM and Self-learning with 100% of the reviews labeled.

when all the training data are labeled, LCCT is still significantly more accurate than all the competitors except Nguyen's method.

contrasting

train_21153

Various other researchers have tried to improve the performance of their paraphrase systems or vector space models by using diverse sources of information such as bilingual corpora (Bannard and Callison-Burch, 2005;Huang et al., 2012;Zou et al., 2013), 10 structured datasets (Yu and Dredze, 2014; or even tagged images (Bruni et al., 2012).

most previous work 11 did not adopt the general, simplifying view that all of these sources of data are just cooccurrence statistics coming from different sources with underlying latent factors.

contrasting

train_21154

We know that whenever P > 0.5, the error rate decreases (and therefore Acc increases) so the text is improved.

3 an increase in P , R or F alone does not necessarily imply an increase in Acc or WAcc, as illustrated in Table 8.

contrasting

train_21155

Therefore, we must accept that any metric used in such scenarios will not be perfect.

it is worth noting that this limitation does not extend to evaluation of error detection per se using such metrics.

contrasting

train_21156

Recording speech poses no particular problems, but retrieval of spoken content using spoken queries is presently available only for the approximately two dozen languages in which there is an established path to market; English, German, or Chinese, for example.

many of the mobile-only users who could benefit most from such systems speak only one of the several hundred other languages that each have at least a million speakers; 1 Balochi, Mossi or Quechua, for example.

contrasting

train_21157

Without multiple draws, confidence intervals on this value cannot be established.

we are confident that random baselines even as high as 0.1 for any of our measures would be surprising.

contrasting

train_21158

Thus, even if we somehow adapt ABA's product-of-marginals heuristic to such models, we run the risk of estimating highly inaccurate posteriors (specifically, zero-valued posteriors).

mIR extends to all IBm-style word alignment models and does not add heuristics.

contrasting

train_21159

On both language pairs, ABA, PostCAT and MIR outperform their respective EM baseline with comparable gains overall.

we noticed that ABA and MIR are not producing the same alignments.

contrasting

train_21160

For evaluating slot induction (AP and PR-AUC), the double-graph random walk (row (e)) performs better on both ASR and manual results, which implies that additionally integrating the lexical knowledge graph helps decide a more coherent and complete slot set since we can model the score propagation more precisely (not only slot-level but wordlevel information).

for SLU evaluation (WAP and AF), the single-graph random walk (row (c)) performs better, which may imply that the slots carrying the coherent relations from the row (e) may not have good semantic decoder performance so that the performance is decreased a little.

contrasting

train_21161

The difficulty lies in the manner of distinguishing paraphrases from expressions that stand in different semantic relations, e.g., antonyms and sibling words, using only the statistics estimated from such corpora.

highly accurate paraphrases can be extracted from parallel or comparable corpora, but their coverage is limited owing to the limited availability of such corpora for most languages.

contrasting

train_21162

(3) a. amendment of regulation ⇔ amending regulation b. amendment of X ⇔ amending X c. amendment of documents ⇔ amending documents Using that method, they were able to expand the seed lexicon by a large multiple (15 to 40 times), and the new paraphrase pairs were of reasonably good quality.

they introduced variables only for identical word forms shared by both sides of each pair and left corresponding pairs of lexical variants, e.g., ("amendment", "amending") in (3a), untouched.

contrasting

train_21163

However, these scores are reasonably high, considering that no use is made of rich language-specific resources 12 .

more grammatical errors occurred than with S Seed and S ID .

contrasting

train_21164

Training in machine learning often uses starting big which is to use up all training data at the same time.

elman (1993) suggests that in some cases, learning should start by training simple models on small data and then gradually increase the model complexity and add more difficult data.

contrasting

train_21165

MaxEnc strategy helps the system achieve the highest accuracy on short sentences (up to length 10).

it is less helpful than Mi-nEnc when performing on long sentences.

contrasting

train_21166

It shows that lexical semantics plays a decisive role in the performance of the system.

it is worth noting that, even without that knowledge (i.e., with the ∞-order generative model alone), the DDA of phase 1 is 2% higher than before being trained (66.89% vs 64.9%).

contrasting

train_21167

The two-registers system is not quite arc-decomposable (Goldberg and Nivre, 2013): if the wrong vertex is stored in a register then a later pair of crossed arcs might both be individually reachable but not jointly reachable.

there may be a "crossing-sensitive" variant of arcdecomposability that takes into account the vertices crossed arcs are incident to that would apply here.

contrasting

train_21168

Xiao and Guo (2013) induce word embeddings across multiple domains, and concatenate these representations into a single feature vector for labeled instances in each domain, following EasyAdapt (Daumé III, 2007).

they do not apply this idea to unsupervised domain adaptation, and do not work in the structured feature setting that we consider here.

contrasting

train_21169

Our formulation is related to multi-domain learning, particularly in the multiattribute setting.

rather than partitioning all instances into domains, the domain attribute formulation allows information to be shared across instances which share metadata attributes.

contrasting

train_21170

We were surprised that SG-EM+RETRO actually performed worse than SG-EM, given how poorly SG-EM performed in the other evaluations.

an analysis again revealed that this was due to the kind of similarity encouraged by WordNet rather than an inability of the model to learn useful vectors.

contrasting

train_21171

(2014) introduce an approach that chooses between multiple composition tensors (AdaMC-RNTN), which yields further gains with respect to RNTN performance.

to the lexicalized and highdimensional RNTN model, there are several lines of work that attempt to work in a more data-scarce setting.

contrasting

train_21172

Since Brown clusters are mostly syntactic/semantic in nature and do not automatically distinguish positive or negative sentiment, we additionally performed multiple experiments to use clusters while incorporating additional sentiment information: On one hand, we try to incorporate the judgements on the Amazon near-domain dataset more directly into the clusters by using the repeated bisecting K-Means algorithm as implemented in CLUTO (Zhao and Karypis, 2005), with previous/next word, part-of-speech tag, and the score of the containing review as features.

we split the Brown clusters according to the sentiment value that they have in a particular sentiment lexicon (e.g.

contrasting

train_21173

The similarity is computed based on statistics such as co-occurrence which are unable to accommodate the subtlety that whether two words labeled as similar are truly similar depends on which topic they appear in, as explained by the aforementioned examples.

ideally, the knowledge would be word A and B are similar under topic C. in reality, we only know two words are similar, but not under which topic.

contrasting

train_21174

DF-LDA and Quad-LDA proposed to use word correlations to enhance the coherence of learned topics.

they improperly enforce words labeled as similar to have similar probabilities in all topics, which violates the fact that whether two words are similar depend on which topic they appear in.

contrasting

train_21175

Topic models provide insights into document collections, and their supervised extensions also capture associated document-level metadata such as sentiment.

inferring such models from data is often slow and cannot scale to big data.

contrasting

train_21176

2 Typically, the topics are discovered through a process of probabilistic inference, either variational EM (Wang et al., 2009) or Gibbs sampling (Boyd-Graber and Resnik, 2010).

these methods scale poorly to large datasets.

contrasting

train_21177

While these advancements improve the scalability of max-margin supervised topic models, the improvement is limited by the fact that the sampling algorithm grows with the number of tokens.

this paper explores a different vein of research that focuses on using efficient representations of summary statistics to estimate statistical models.

contrasting

train_21178

Traditional information extraction would be content with extracting two binary relation instances (NEG-REG,BCL,RFLAT) and (NEG-REG,IL-10,RFLAT), where NEG-REG represents a negative regulation (i.e., inhibition).

the sentence also discloses important contextual information, i.e., BCL regulates RFLAT by stimulating the inhibitive effect of IL-10, and likewise the inhibition of RFLAT by IL-10 is controlled by BCL.

contrasting

train_21179

Complex knowledge extraction can be naturally framed as a semantic parsing problem, with the event structure represented by a semantic parse; see Figure 2.

annotating example sentences is expensive and time-consuming.

contrasting

train_21180

(2014) took an important step toward this direction, by learning a semantic parser based on combinatorial categorial grammar (CCG) from Freebase and web sentences.

krishnamurthy and Mitchell (2012) still learned from binary relations, using only simple sentences (of length ten or less).

contrasting

train_21181

Automatic sentiment analysis of text, especially social media posts, has a number of applications in commerce, public health, and public policy development.

a vast majority of prior research on automatic sentiment analysis has been on English texts.

contrasting

train_21182

Thus, instead of building source-language specific sentiment analysis systems, one can translate the texts into English and use an English sentiment analysis system.

it is widely believed that aspects of sentiment may be lost in translation, especially in automatic translation.

contrasting

train_21183

This suggests that certain attributes of automatically translated text mislead humans with regards to the true sentiment of the source text.

these same attributes do not seem to affect the automatic sentiment analysis system as much.

contrasting

train_21184

The LM score features are extracted from large amounts of news article data, and are good representation of the general importance of bigrams for the test domain.

word-Net information is collected from a more general aspect, which may not be a very good choice for this task.

contrasting

train_21185

By the constraints of the algorithm, a head word x h must combine with each of its left and right dependents.

the order of combination can lead to different tree structures (as illustrated in Figure 2).

contrasting

train_21186

The scoring function requires specifying a set of parse features f which, in theory, could be directly adapted from existing lexicalized c-parsers.

the structure of the dependency parse greatly limits the number of decisions that need to be made, and allows for a smaller set of features.

contrasting

train_21187

This task effectively amounts to disambiguating the sense of discourse connective, which can be done with high accuracy (Pitler et al., 2008).

in the absence of an explicit discourse connective, inferring the sense of a discourse relation has proved to a very challenging task (Park and Cardie, 2012;Rutherford and Xue, 2014).

contrasting

train_21188

The common approach has been to inject knowledge as features.

these pieces of knowledge provide relatively strong evidence that loses impact in standard training due to sparsity.

contrasting

train_21189

As hard coreference problems are rare in standard coreference datasets, we do not have significant performance improvement.

these results show that our additional Predicate Schemas do not harm the predictions for regular mentions.

contrasting

train_21190

With the development of machine learning based models (Connolly et al., 1994;Soon et al., 2001b;Ng and Cardie, 2002a), attention shifted to solving standard coreference resolution problems.

many hard coreference problems involve pronouns.

contrasting

train_21191

Mnih and Teh (2012) fix these parameters to 1 and obtain the same perplexities, thereby circumventing the need for explicit normalisation.

this method does not provide any guarantees that the models are normalised at test time.

contrasting

train_21192

One option for speeding up factored models is using a GPU to perform the vector-matrix operations.

gPU integration is architecture specific and thus against our goal of making our language modelling toolkit usable by everyone.

contrasting

train_21193

We obtain a slightly better BLEU score with stochastic gradient descent, but this is likely to be just noise from tuning the translation system with MERT.

noise contrastive training reduces training time by a factor of 7.

contrasting

train_21194

As a first attempt to automate the test, we only experiment with basic linguistic features.

we believe that the task itself offers an opportunity for the development of-and subsequent evaluation of-rich linguistic features that may be better equipped for determining the aboutness of conversations.

contrasting

train_21195

A portion of user utterances refer to general Java knowledge, and in these cases semantic interpretation can be accomplished by mapping to a domain-specific ontology (e.g., Dzikovska et al., 2007).

many utterances refer to concrete entities within the dynamically changing, user-created programing artifact.

contrasting

train_21196

Importantly, this value does not test generalization to unseen questions, since KNOWBOT has held dialogs on these questions.

it does show that our system can effectively learn about its domain: a poor dialog extraction system will fail to extract any helpful knowledge from users during a training dialog.

contrasting

train_21197

Our open system acquires relations from a wide variety of user explanations without the bottleneck of a hand-built dialog model, but the tradeoff is that we use relatively simple, templated system prompts.

our collected corpus of real human-system dialogs can be used to improve our system in further iterations.

contrasting

train_21198

This is to be expected, as the speech in the control group has more complete sentences and fewer disfluencies.

it is interesting to note that performance on the PNFA and SD groups is not much worse.

contrasting

train_21199

All the features relating to Yngve depth and height of the parse trees are significantly different (in at least one of the three clinical groups).

of the eight primary syntactic units calculated by Lu's SCA, six show no significant difference when measured on the automatically segmented transcripts.

contrasting