tasksource/scinli · Datasets at Hugging Face

id stringlengths 7 12	sentence1 stringlengths 6 1.27k	sentence2 stringlengths 6 926	label stringclasses 4 values
train_2600	A very interesting work of Mohammad and Turney (2010) uses Mechanical Turk in order to build the lexicon of emotions evoked by words.	we present an automatic approach that infers the general connotation of words.	contrasting
train_2601	We presented an Egyptian to English MT system.	to previous work, we used an automatic conversion method to map Egyptian close to MSA.	contrasting
train_2602	Recent work (Zhao and Gildea, 2010) described an extension to the HMM with a fertility model, using MCMC techniques for parameter estimation.	they do not have a efficient means of MAP inference, which is necessary in many applications such as machine translation.	contrasting
train_2603	The structure of the fertility model violates the Markov assumptions used in this dynamic programming method.	we may empirically estimate the posterior distribution using Markov chain Monte Carlo methods such as Gibbs sampling (Zhao and Gildea, 2010).	contrasting
train_2604	1 Once Upon a Time... For years, the standard way to do statistical machine translation parameter tuning has been to use minimum error-rate training, or MERT (Och, 2003).	as researchers started using models with thousands of parameters, new scalable optimization algorithms such as MIRA (Watanabe et al., 2007;Chiang et al., 2008) and PRO (Hopkins and May, 2011) have emerged.	contrasting
train_2605	Ideally, the learning algorithm should be able to recover from overshooting.	once monsters are encountered, they quickly start dominating, with no chance for PRO to recover since it accumulates n-best lists, and thus also monsters, over iterations.	contrasting
train_2606	We can see in Table 3 that all selection mechanisms considerably improve BLEU compared to the baseline PRO, by 2-3 BLEU points.	not every selection alternative gets rid of monsters, which can be seen by the large lengths and low BLEU+1 for the negative examples (in bold).	contrasting
train_2607	Large variance between the results obtained with MIRA has also been reported (Simianer et al., 2012).	none of this work has focused on monsters.	contrasting
train_2608	For the first day ( ), the model functions the same as a standard DPM, i.e., all the topics use the same base measure, ( ).	for later days ( ), besides the base measure, ( ), we make use of topics learned from previous days as priors.	contrasting
train_2609	When available, using in-project training data proves significantly more successful than using out-of-project data.	we find that when using out-of-project data, a dataset based on more words than code performs consistently better.	contrasting
train_2610	as information source to produce or derive interpretations based on them.	existing uncertainty cues are ineffective in social media context because of its specific characteristics.	contrasting
train_2611	Only 'hat' refers unambigu-ously.	the language and scenario facilitate scalar implicature (Horn, 1972;Harnish, 1979;Gazdar, 1979).	contrasting
train_2612	show that their agents' linguistic behavior is broadly Gricean.	their agents' language is too simple to reveal implicatures.	contrasting
train_2613	We are not the first researchers to use lexicalized features for coreference resolution.	pre- vious work has evaluated the benefit of lexical features only for broad-coverage data sets.	contrasting
train_2614	It is typically applicable in the text generation field, both for concept-to-text generation and text-totext generation (Lapata, 2003), such as multiple document summarization (MDS), question answering and so on.	ordering a set of sentences into a coherent text is still a hard and challenging problem for computers.	contrasting
train_2615	Nahnsen (2009) employed features which were based on discourse entities, shallow syntactic analysis, and temporal precedence relations retrieved from VerbOcean.	the model does not perform well on datasets describing the consequences of events.	contrasting
train_2616	Like sentence (3) and (4) in Figure 1, they have a highest cosine similarity of 0.2240 and most overlap words of "Israel" and "nuclear".	the similarity or overlap of the two sen-tences does not help to decide which sentence should be arranged before another.	contrasting
train_2617	This can be attributed to two factors: (i) the PCTB annotates non-local dependencies using traces, and (ii) Chinese syntax generates more traces than English syntax (Guo et al., 2007).	for parsers that do not return traces they are a benign error.	contrasting
train_2618	These methods usually rely heavily on the manually annotated treebanks for training the dependency models.	annotating syntac- Figure 1: Different grammar formalisms of syntactic structures between CTB (upper) and CDT (below).	contrasting
train_2619	They also justified that the popular "early update" (Collins and Roark, 2004) is valid for the systems that do not exhibit spurious ambiguity 4 .	for the easy-first algorithm or more generally, systems that exhibit spurious ambiguity, even "early update" could fail to ensure validity of update (see the example in figure 1).	contrasting
train_2620	Work on opinion and sentiment tends to focus on explicit expressions of opinions.	many attitudes are conveyed implicitly, and benefactive/malefactive events are important for inferring implicit attitudes.	contrasting
train_2621	We report the results on 2011 only, because even when the same team participated in more than one year, the metrics submitted were different and the 2011 results represent the best effort of these teams.	as we saw in Table 1, in 2011 there were very few significant differences between the top summarization systems.	contrasting
train_2622	Because of its dependence on strings, it performs better with larger sets of model summaries.	to ROUGE, pyramid scoring is robust with as few as four or five model summaries (Nenkova and Passonneau, 2004).	contrasting
train_2623	The TDRM approach captures topics that are centered around a specific information need, often with a limited vocabulary, which favors word co-occurrence.	topics learned on entire collections are coarser than ours, which leads to lower coherence scores.	contrasting
train_2624	For that purpose, numerous procedures have been proposed (Milligan and Cooper, 1985).	none of the listed methods were effective or adaptable to our specific problem.	contrasting
train_2625	Ideally, each query subtopic should be represented by a unique cluster containing all the relevant Web pages inside.	this task is far from being achievable.	contrasting
train_2626	First results are provided in Table 1 and evidence that the best configurations for different < p, K, S(W ik , W jl ) > tuples are obtained for high values of p, K ranging from 4 to 6 clusters and P M I steadily improving over SCP .	such a fuzzy configuration is not satisfactory.	contrasting
train_2627	Although supervised CWS models (Xue, 2003;Zhao et al., 2006;Zhang and Clark, 2007;Sun, 2011) proposed in the past years showed some reasonably accurate results, the outstanding problem is that they rely heavily on a large amount of labeled da-ta.	the production of segmented Chinese texts is time-consuming and expensive, since hand-labeling individual words and word boundaries is very hard (Jiao et al., 2006).	contrasting
train_2628	We note that when the number of texts changes from 0 to 50,000, the f-score and OOV both are improved.	when unlabeled data changes to 200,000, the performance is a bit decreased, while still better than not using unlabeled data.	contrasting
train_2629	These supervised methods show good results, however, are unable to incorporate information from new domain, where OOV problem is a big challenge for the research community.	unsupervised word segmentation Peng and Schuurmans (2001); Goldwater et al.	contrasting
train_2630	Their approach falls into an offline approach, which focuses on creating dictionaries by extracting new words from large corpora separately before WS.	offline approaches have limitation unless the lexicon is constantly updated.	contrasting
train_2631	This is why the result of +LM-S+LM-P is not shown for Chinese.	replacing LM-S with LM-P improved the performance significantly.	contrasting
train_2632	We found positive changes such as * 欧麦/尔萨利赫 oumai/ersalihe to 欧麦尔/萨利赫 oumaier/salihe "Umar Saleh" and * 领导/人曼德拉 lingdao/renmandela to 领导人/曼德拉 lingdaoren/mandela"Leader Mandela".	considering the overall F-measure increase and proper noun F-measure decrease suggests that the effect of LM projection is not limited to proper nouns but also promoted finer granularity because we observed proper noun recall increase.	contrasting
train_2633	For the automated processing of spoken communication in these scenarios, a speech recognition system must be able to handle code switches.	the components of speech recognition systems are usually trained on monolingual data.	contrasting
train_2634	Furthermore, the integration of additional features as input is rather straightforward due to their structure.	factored language models (FLMs) have been used successfully for languages with rich morphology due to their ability to process syntactical features, such as word stems or part-of-speech tags (Bilmes and Kirchhoff, 2003;El-Desoky et al., 2010).	contrasting
train_2635	For the FLM, it leads to no improvement to add the language identifier as feature.	clustering the words into their languages on the output layer of the RNNLM leads to lower perplexities.	contrasting
train_2636	For all k and k , k-th column vector in WS are aligned k -th column vector in WT .	we can not measure similarity between the topic proportions because we do not have any language resources such as dictionary.	contrasting
train_2637	To mimimize the objective, gradient descend can be used.	but that is not convex function, we only obtained local optimal.	contrasting
train_2638	This is reasonable because we can assume that both languages share the same latent concept.	we cannot quantify the similarity between the topics because we do not have any external language resources such as a dictionary.	contrasting
train_2639	(2011) carefully explored review-related features based on content and sentiment, training a semi-supervised classifier for opinion spam detection.	the disadvantages of standard supervised learning methods are obvious.	contrasting
train_2640	Note that this is only a limitation of our inference procedure, not the model, and future work will look at other ways (e.g., Gibbs sampling) to perform inference.	generating Y and Z given X, such that the joke is funny, is still a formidable challenge that a lot of humans are not able to perform successfully (cf.	contrasting
train_2641	However, intermediate distances beyond the n-gram model limits can be very useful and should not be discarded.	distant-bigram models and distance-dependent trigger models make use of both, distance and co-occurrence, information up to window sizes of ten to twenty.	contrasting
train_2642	If this is not the case, then the text is likely to make little sense.	if this is the case, then the taboo meaning is potentially expanded to the phrase level.	contrasting
train_2643	Several other studies report results only on light verb constructions formed with certain light verbs (Stevenson et al., 2004;Tan et al., 2006;Tu and Roth, 2011).	we aimed to identify all kinds of LVCs, i.e.	contrasting
train_2644	In case of a large number of participating systems each assessor ranks only a subset of MT outputs.	a fair overall ranking cannot be always derived from such partial rankings (Callison-Burch et al., 2012).	contrasting
train_2645	(2010) proposed a methodology for classifying more details than was possible in the study by Sammons et al..	these studies were based on only English data sets.	contrasting
train_2646	The above studies rely on the high coverage of the original bilingual knowledge and a specific data source together with the translation vocabularies, co-occurrence information and language links.	the severest problem is that they cannot understand semantic information.	contrasting
train_2647	There is a large body of work around WSD and translation selection.	many of these approaches require lexical resources or large bilingual corpora with rich information fields and annotations, as reviewed in section 2.	contrasting
train_2648	However, aligned corpora can be difficult to obtain for under-resourced language pairs, and are expensive to construct.	documents in a comparable corpus comprise bilingual or multilingual text of a similar nature, and need not even be exact translations of each other.	contrasting
train_2649	For the precision metric, wiki-lsi scored 0.650 when all 80 input sentences are tested, while the base-freq baseline scored 0.550. goog-tr has the highest precision at 0.797.	if only the Chinese and Malay inputs -which has less presence on the Internet and 'less resource-rich' than English -were tested (since goog-tr cannot accept Iban inputs), wiki-lsi and goog-tr actually performs equally well at 0.690 precision.	contrasting
train_2650	An example is given in Figure 2, where the English word seems align to two Hindi words hE and lagawA.	from the small amount of labeled training data (i.e., a set of handcorrected tree pairs), we can learn what kind of source words are likely to align to multiple target words, and which target word is likely to the head.	contrasting
train_2651	Sign languages (SL), the vernaculars of deaf people, are complete, rich, standalone communication systems which have evolved in parallel with oral languages (Valli and Lucas, 2000).	in contrast to the last ones, research in automatic SL processing has not yet managed to build a complete, formal definition oriented to their automatic recognition (Cuxac and Dalle, 2007).	contrasting
train_2652	As the table shows, increasing the number of folds results in higher BLEU scores.	doing such will generally lead to higher variance among base learners.	contrasting
train_2653	Although the translation is correct in this situation, translating the Chinese word "lingyu" to "waters" appears very few times since the common translations are "areas" or "fields".	simply filtering out this kind of sentence pairs may lead to some loss of native English expressions, thereby the trans-lation performance is unstable since both nonparallel sentence pairs and non-literal but parallel sentence pairs are filtered.	contrasting
train_2654	Although the grammar may have rules to translate these two phrases, they can be safely pruned for this particular sentence pair.	to chart pruning for monolingual parsing, our pruning decisions are based on the source context, its target translation and the mapping between the two.	contrasting
train_2655	Moreover, for the frame "攻占" in the input sentence, the MEANT-tuned system has correctly translated the ARG0 "哈玛斯好战份子" into "Hamas militants" and the ARG1 "加萨走廊" into "Gaza".	the TER-tuned system has dropped the predicate "施行" so that the corresponding arguments "The Palestinian Authority" and "into a state of emergency" have all been incorrectly associated with the predicate "攻占 /seized".	contrasting
train_2656	In answer ranking phase, TI-QA considers the probabilities of different answer types as well: On one hand, TD-QA can achieve relative high ranking precision, as using a unique answer type greatly reduces the size of the candidate list for ranking.	as the answer-typing model is far from perfect, if prediction errors happen, TD-QA can no longer give correct answers at all.	contrasting
train_2657	However, as the answer-typing model is far from perfect, if prediction errors happen, TD-QA can no longer give correct answers at all.	tI-QA can provide higher answer coverage, as it can extract answer candidates with multiple answer types.	contrasting
train_2658	Since question-answer pairs are usually short, the word mismatching problem is especially important.	due to the lexical gap between questions and answers as well as spam typically existing in user-generated content, filtering and ranking answers is very challenging.	contrasting
train_2659	For example, they convert a 1, 000, 000-dimensional vector of word space into a 1000 × 1000 matrix.	in our model, a document is still represented by a vector.	contrasting
train_2660	For instance, in the British National Corpus, time and see are more frequent than thing or may and man is more frequent than part.	it seems intuitively right to say that time, see and man are more 'precise' concepts than thing, may and part respectively.	contrasting
train_2661	1 In Section 2, we defined semantic content as a notion encompassing various referential properties, including a basic concept of extension in cases where it is applicable.	we do not know of a dataset providing human judgements over the general informativeness of lexical items.	contrasting
train_2662	In biased-SVM, it is necessary to run SVM many times, as we searched "c" and "j".	mCDC does not require such parameter tuning.	contrasting
train_2663	Antonio Roque (2012) has challenged Ramsay's claim, and certainly there has been successful work done in the computational analysis and modeling of narratives, as we will review in the next section.	we believe that most previous work (except possibly (Elsner, 2012)) has failed to directly address the root cause of Ramsay's skepticism: can computers extract the emotions encoded in a narrative?	contrasting
train_2664	Not surprisingly, Claudius draws the most negative sentiment from Hamlet, receiving a score of -27.	gertrude is very well liked by Hamlet (+24), which is unexpected (at least to us) since Hamlet suspects that his mother was involved in murdering King Hamlet.	contrasting
train_2665	In the case of deep neural embeddings, for example, training time can number in days.	learned embeddings are becoming more abundant, as much research and computing effort is being invested in learning word representations using large-scale deep architectures trained on web-scale corpora.	contrasting
train_2666	min (φ iφj − φ i φ j ) 2 ∀i, j (where, with some abuse of notation, φ andφ are the source and target embeddings respectively).	this objective is no longer convex in the embeddings.	contrasting
train_2667	Adj + NP It is common practice to extract any NP modified by a sentiment adjective.	this simple extraction rule suffers from precision problems.	contrasting
train_2668	Quotes are used in news articles as evidence of a person's opinion, and thus are a useful target for opinion mining.	labelling each quote with a polarity score directed at a textually-anchored target can ignore the broader issue that the speaker is commenting on.	contrasting
train_2669	In sentiment analysis over product reviews, polarity labels are commonly used because the target, the product, is clearly identified.	for quotes on topics of debate, the target and meaning of polarity labels is less clear.	contrasting
train_2670	These approaches assume that each document has a single source (the document's author), whose communicative goal is to evaluate a well-defined target, such as a product or a movie.	this does not hold in news articles, where the goal of the journalist is to present the viewpoints of potentially many people.	contrasting
train_2671	Bag-of-words (BOW) is now the most popular way to model text in machine learning based sentiment classification.	the performance of such approach sometimes remains rather limited due to some fundamental deficiencies of the BOW model.	contrasting
train_2672	2012;Shi et al., 2010;Prettenhofer and Stein 2010).	the binary classification task is very different from the regression task studied in this paper, and the proposed methods in the above previous works cannot be directly applied.	contrasting
train_2673	REG_STC: It combines REG_S and REG_T by averaging their prediction values.	the above regression methods do not perform very well due to the unsatisfactory machine translation quality and the various language expressions.	contrasting
train_2674	Thus, each example in the unlabeled set is required to be checked by training a new regression model utilizing the example.	the model training process is usually very time-consuming for many regression algorithms, which significantly limits the use of the work in (Zhou and Li, 2005).	contrasting
train_2675	Suppose we would have sense tagged data, p(S\|w, c) could have been computed as: But since the sense tagged corpus is not available, we cannot find #(S, w, c) from the corpus directly.	we can estimate it using the comparable corpus in other language.	contrasting
train_2676	Therefore, combining additional information with the LMs could reduce recognition errors.	direct integration of such information in the decoder is difficult.	contrasting
train_2677	An available treebank is a major resource for syntactic parsing.	it is often a key bottleneck to acquire credible treebanks.	contrasting
train_2678	Future systems might use the same or a similar feature set to ours, but in an architecture that does not include any generative parser.	some systems might indeed incorporate this generative model's score.	contrasting
train_2679	But, Φ phrase+deps+gen is significantly better than Φ phrase+gen only on F 1 , but not on UAS or LAS.	on the out-of-domain BROWN tests, we find that adding Φ deps always adds considerably, and in a statistically significant way, to both LAS and UAS.	contrasting
train_2680	This result is perhaps counter-intuitive, in the sense that one might have supposed that higher-order dependency features, being highly specific by nature, might only have only served to over-fit the training material.	this result shows otherwise.	contrasting
train_2681	While closed-form solutions have been developed for some specialized components (Martins et al., 2011), this problem is in general more difficult than the one arising in the subgradient algorithm.	the following result, proved in Martins et al.	contrasting
train_2682	We present an unsupervised approach to part-of-speech tagging based on projections of tags in a word-aligned bilingual parallel corpus.	to the existing state-of-the-art approach of Das and Petrov, we have developed a substantially simpler method by automatically identifying "good" training sentences from the parallel corpus and applying self-training.	contrasting
train_2683	(2012) build supervised POS taggers for 22 languages using the TNT tagger (Brants, 2000), with an average accuracy of 95.2%.	many widelyspoken languages -including Bengali, Javanese, and Lahnda -have little data manually labelled for POS, limiting supervised approaches to POS tagging for these languages.	contrasting
train_2684	Algorithm 2 describes this process of self training and revision, and assumes that the parallel source-target corpus has been word aligned, with many-to-one alignments removed, and that the sentences are sorted by alignment score.	to Algorithm 1, all sentences are used, not just the 60k sentences with the highest alignment scores.	contrasting
train_2685	Finally, the keywords S ⊆ t, with \|S\| ≤ k, are chosen by maximizing the cumulative reward function over all the topics, formulated as follows: Since R(S) is submodular, the greedy algorithm for maximizing R(S) is shown as Algorithm 1 on the next page, with r {w},z being similar to r S,z with S = {w}.	if λ = 1, the reward function is linear and only measures the topical similarity of words with the main topics of t. when 0 < λ < 1, as soon as a word is selected from a topic, other words from the same topic start having diminishing gains.	contrasting
train_2686	They find that these features only contribute less than 2% to precision.	in our approach linguistic features are quite useful.	contrasting
train_2687	Relatively low perplexity has made modified Kneser-Ney smoothing (Kneser and Ney, 1995;Chen and Goodman, 1998) a popular choice for language modeling.	existing estimation methods require either large amounts of RAM (Stolcke, 2002) or machines (Brants et al., 2007).	contrasting
train_2688	We have found that, even with mostlysequential access, memory mapping is slower because the kernel does not explicitly know where to read ahead or write behind.	we use dedicated threads for reading and writing.	contrasting
train_2689	We note that interpolation (Equation 2) used the different backoff b(w n−1 1 ) and so b(w n 1 ) is not immediately available.	the backoff values were saved in suffix order ( §3.3) and interpolation produces probabilities in suffix order.	contrasting
train_2690	To avoid the need for hard decisions about domain membership, some have used topic modeling to improve SMT performance, e.g., using latent semantic analysis (Tam et al., 2007) or 'biTAM' (Zhao and Xing, 2006).	to our source language approach, these authors use both source and target information.	contrasting
train_2691	For instance, in the task of tagging mail addresses, a feature of "5 consecutive digits" is highly indicative of a POSTCODE.	in the alignment model, it does not make sense to design features based on a hard-coded state, say, a feature of "source word lemma matching target word lemma" fires for state index 6.	contrasting
train_2692	While the participants in that study were younger (7 to 9+ years old), the study is relevant because the challenges those young participants face, are faced again when readers of any age encounter new and complicated texts that present words they do not know, and ideas they have never considered.	there is ample work on the basic algorithm to place a sequence of words in a typesetting area with a certain width, commonly known as the optimal line breaking problem (e.g., , Knuth and Plass (1981)).	contrasting
train_2693	It should be noted that when a text is typeset into an area of width of a certain number of characters, an erroneous break need not necessarily lead to an actual break in the final output, that is an error may not be too bad.	a missed break while not hurting the readability of the text may actually lead to a long segment that may eventually worsen raggedness in the final typesetting.	contrasting
train_2694	Each example has more than one annotation and we need to determine a single sense tag for each example.	if we assign senses by majority voting, we need a backoff strategy in case of ties.	contrasting
train_2695	Columns V and B show respectively whether the underspecified senses are a result of majority voting or backoff.	to volunteers, turkers disprefer the underspecified option and most of the English underspecified senses are assigned by backoff.	contrasting
train_2696	Syntax-based vector spaces are used widely in lexical semantics and are more versatile than word-based spaces (Baroni and Lenci, 2010).	they are also sparse, with resulting reliability and coverage problems.	contrasting
train_2697	The sentence Karen threw her arms round my neck, spilling champagne everywhere contains the LU throw.v evoking the frame BODY MOVEMENT.	throw.v is ambiguous and may also evoke CAUSE MOTION.	contrasting
train_2698	The evaluation of whole-sentence semantic structures plays an important role in semantic parsing and large-scale semantic structure annotation.	there is no widely-used metric to evaluate wholesentence semantic structures.	contrasting
train_2699	Evaluating such structures is necessary for semantic parsing tasks, as well as semantic annotation tasks which create linguistic resources for semantic parsing.	there is no widely-used evaluation method for whole-sentence semantic structures.	contrasting

train_2600

A very interesting work of Mohammad and Turney (2010) uses Mechanical Turk in order to build the lexicon of emotions evoked by words.

we present an automatic approach that infers the general connotation of words.

contrasting

train_2601

We presented an Egyptian to English MT system.

to previous work, we used an automatic conversion method to map Egyptian close to MSA.

contrasting

train_2602

Recent work (Zhao and Gildea, 2010) described an extension to the HMM with a fertility model, using MCMC techniques for parameter estimation.

they do not have a efficient means of MAP inference, which is necessary in many applications such as machine translation.

contrasting

train_2603

The structure of the fertility model violates the Markov assumptions used in this dynamic programming method.

we may empirically estimate the posterior distribution using Markov chain Monte Carlo methods such as Gibbs sampling (Zhao and Gildea, 2010).

contrasting

train_2604

1 Once Upon a Time... For years, the standard way to do statistical machine translation parameter tuning has been to use minimum error-rate training, or MERT (Och, 2003).

as researchers started using models with thousands of parameters, new scalable optimization algorithms such as MIRA (Watanabe et al., 2007;Chiang et al., 2008) and PRO (Hopkins and May, 2011) have emerged.

contrasting

train_2605

Ideally, the learning algorithm should be able to recover from overshooting.

once monsters are encountered, they quickly start dominating, with no chance for PRO to recover since it accumulates n-best lists, and thus also monsters, over iterations.

contrasting

train_2606

We can see in Table 3 that all selection mechanisms considerably improve BLEU compared to the baseline PRO, by 2-3 BLEU points.

not every selection alternative gets rid of monsters, which can be seen by the large lengths and low BLEU+1 for the negative examples (in bold).

contrasting

train_2607

Large variance between the results obtained with MIRA has also been reported (Simianer et al., 2012).

none of this work has focused on monsters.

contrasting

train_2608

For the first day ( ), the model functions the same as a standard DPM, i.e., all the topics use the same base measure, ( ).

for later days ( ), besides the base measure, ( ), we make use of topics learned from previous days as priors.

contrasting

train_2609

When available, using in-project training data proves significantly more successful than using out-of-project data.

we find that when using out-of-project data, a dataset based on more words than code performs consistently better.

contrasting

train_2610

as information source to produce or derive interpretations based on them.

existing uncertainty cues are ineffective in social media context because of its specific characteristics.

contrasting

train_2611

Only 'hat' refers unambigu-ously.

the language and scenario facilitate scalar implicature (Horn, 1972;Harnish, 1979;Gazdar, 1979).

contrasting

train_2612

show that their agents' linguistic behavior is broadly Gricean.

their agents' language is too simple to reveal implicatures.

contrasting

train_2613

We are not the first researchers to use lexicalized features for coreference resolution.

pre- vious work has evaluated the benefit of lexical features only for broad-coverage data sets.

contrasting

train_2614

It is typically applicable in the text generation field, both for concept-to-text generation and text-totext generation (Lapata, 2003), such as multiple document summarization (MDS), question answering and so on.

ordering a set of sentences into a coherent text is still a hard and challenging problem for computers.

contrasting

train_2615

Nahnsen (2009) employed features which were based on discourse entities, shallow syntactic analysis, and temporal precedence relations retrieved from VerbOcean.

the model does not perform well on datasets describing the consequences of events.

contrasting

train_2616

Like sentence (3) and (4) in Figure 1, they have a highest cosine similarity of 0.2240 and most overlap words of "Israel" and "nuclear".

the similarity or overlap of the two sen-tences does not help to decide which sentence should be arranged before another.

contrasting

train_2617

This can be attributed to two factors: (i) the PCTB annotates non-local dependencies using traces, and (ii) Chinese syntax generates more traces than English syntax (Guo et al., 2007).

for parsers that do not return traces they are a benign error.

contrasting

train_2618

These methods usually rely heavily on the manually annotated treebanks for training the dependency models.

annotating syntac- Figure 1: Different grammar formalisms of syntactic structures between CTB (upper) and CDT (below).

contrasting

train_2619

They also justified that the popular "early update" (Collins and Roark, 2004) is valid for the systems that do not exhibit spurious ambiguity 4 .

for the easy-first algorithm or more generally, systems that exhibit spurious ambiguity, even "early update" could fail to ensure validity of update (see the example in figure 1).

contrasting

train_2620

Work on opinion and sentiment tends to focus on explicit expressions of opinions.

many attitudes are conveyed implicitly, and benefactive/malefactive events are important for inferring implicit attitudes.

contrasting

train_2621

We report the results on 2011 only, because even when the same team participated in more than one year, the metrics submitted were different and the 2011 results represent the best effort of these teams.

as we saw in Table 1, in 2011 there were very few significant differences between the top summarization systems.

contrasting

train_2622

Because of its dependence on strings, it performs better with larger sets of model summaries.

to ROUGE, pyramid scoring is robust with as few as four or five model summaries (Nenkova and Passonneau, 2004).

contrasting

train_2623

The TDRM approach captures topics that are centered around a specific information need, often with a limited vocabulary, which favors word co-occurrence.

topics learned on entire collections are coarser than ours, which leads to lower coherence scores.

contrasting

train_2624

For that purpose, numerous procedures have been proposed (Milligan and Cooper, 1985).

none of the listed methods were effective or adaptable to our specific problem.

contrasting

train_2625

Ideally, each query subtopic should be represented by a unique cluster containing all the relevant Web pages inside.

this task is far from being achievable.

contrasting

train_2626

First results are provided in Table 1 and evidence that the best configurations for different < p, K, S(W ik , W jl ) > tuples are obtained for high values of p, K ranging from 4 to 6 clusters and P M I steadily improving over SCP .

such a fuzzy configuration is not satisfactory.

contrasting

train_2627

Although supervised CWS models (Xue, 2003;Zhao et al., 2006;Zhang and Clark, 2007;Sun, 2011) proposed in the past years showed some reasonably accurate results, the outstanding problem is that they rely heavily on a large amount of labeled da-ta.

the production of segmented Chinese texts is time-consuming and expensive, since hand-labeling individual words and word boundaries is very hard (Jiao et al., 2006).

contrasting

train_2628

We note that when the number of texts changes from 0 to 50,000, the f-score and OOV both are improved.

when unlabeled data changes to 200,000, the performance is a bit decreased, while still better than not using unlabeled data.

contrasting

train_2629

These supervised methods show good results, however, are unable to incorporate information from new domain, where OOV problem is a big challenge for the research community.

unsupervised word segmentation Peng and Schuurmans (2001); Goldwater et al.

contrasting

train_2630

Their approach falls into an offline approach, which focuses on creating dictionaries by extracting new words from large corpora separately before WS.

offline approaches have limitation unless the lexicon is constantly updated.

contrasting

train_2631

This is why the result of +LM-S+LM-P is not shown for Chinese.

replacing LM-S with LM-P improved the performance significantly.

contrasting

train_2632

We found positive changes such as * 欧麦/尔萨利赫 oumai/ersalihe to 欧麦尔/萨利赫 oumaier/salihe "Umar Saleh" and * 领导/人曼德拉 lingdao/renmandela to 领导人/曼德拉 lingdaoren/mandela"Leader Mandela".

considering the overall F-measure increase and proper noun F-measure decrease suggests that the effect of LM projection is not limited to proper nouns but also promoted finer granularity because we observed proper noun recall increase.

contrasting

train_2633

For the automated processing of spoken communication in these scenarios, a speech recognition system must be able to handle code switches.

the components of speech recognition systems are usually trained on monolingual data.

contrasting

train_2634

Furthermore, the integration of additional features as input is rather straightforward due to their structure.

factored language models (FLMs) have been used successfully for languages with rich morphology due to their ability to process syntactical features, such as word stems or part-of-speech tags (Bilmes and Kirchhoff, 2003;El-Desoky et al., 2010).

contrasting

train_2635

For the FLM, it leads to no improvement to add the language identifier as feature.

clustering the words into their languages on the output layer of the RNNLM leads to lower perplexities.

contrasting

train_2636

For all k and k , k-th column vector in WS are aligned k -th column vector in WT .

we can not measure similarity between the topic proportions because we do not have any language resources such as dictionary.

contrasting

train_2637

To mimimize the objective, gradient descend can be used.

but that is not convex function, we only obtained local optimal.

contrasting

train_2638

This is reasonable because we can assume that both languages share the same latent concept.

we cannot quantify the similarity between the topics because we do not have any external language resources such as a dictionary.

contrasting

train_2639

(2011) carefully explored review-related features based on content and sentiment, training a semi-supervised classifier for opinion spam detection.

the disadvantages of standard supervised learning methods are obvious.

contrasting

train_2640

Note that this is only a limitation of our inference procedure, not the model, and future work will look at other ways (e.g., Gibbs sampling) to perform inference.

generating Y and Z given X, such that the joke is funny, is still a formidable challenge that a lot of humans are not able to perform successfully (cf.

contrasting

train_2641

However, intermediate distances beyond the n-gram model limits can be very useful and should not be discarded.

distant-bigram models and distance-dependent trigger models make use of both, distance and co-occurrence, information up to window sizes of ten to twenty.

contrasting

train_2642

If this is not the case, then the text is likely to make little sense.

if this is the case, then the taboo meaning is potentially expanded to the phrase level.

contrasting

train_2643

Several other studies report results only on light verb constructions formed with certain light verbs (Stevenson et al., 2004;Tan et al., 2006;Tu and Roth, 2011).

we aimed to identify all kinds of LVCs, i.e.

contrasting

train_2644

In case of a large number of participating systems each assessor ranks only a subset of MT outputs.

a fair overall ranking cannot be always derived from such partial rankings (Callison-Burch et al., 2012).

contrasting

train_2645

(2010) proposed a methodology for classifying more details than was possible in the study by Sammons et al..

these studies were based on only English data sets.

contrasting

train_2646

The above studies rely on the high coverage of the original bilingual knowledge and a specific data source together with the translation vocabularies, co-occurrence information and language links.

the severest problem is that they cannot understand semantic information.

contrasting

train_2647

There is a large body of work around WSD and translation selection.

many of these approaches require lexical resources or large bilingual corpora with rich information fields and annotations, as reviewed in section 2.

contrasting

train_2648

However, aligned corpora can be difficult to obtain for under-resourced language pairs, and are expensive to construct.

documents in a comparable corpus comprise bilingual or multilingual text of a similar nature, and need not even be exact translations of each other.

contrasting

train_2649

For the precision metric, wiki-lsi scored 0.650 when all 80 input sentences are tested, while the base-freq baseline scored 0.550. goog-tr has the highest precision at 0.797.

if only the Chinese and Malay inputs -which has less presence on the Internet and 'less resource-rich' than English -were tested (since goog-tr cannot accept Iban inputs), wiki-lsi and goog-tr actually performs equally well at 0.690 precision.

contrasting

train_2650

An example is given in Figure 2, where the English word seems align to two Hindi words hE and lagawA.

from the small amount of labeled training data (i.e., a set of handcorrected tree pairs), we can learn what kind of source words are likely to align to multiple target words, and which target word is likely to the head.

contrasting

train_2651

Sign languages (SL), the vernaculars of deaf people, are complete, rich, standalone communication systems which have evolved in parallel with oral languages (Valli and Lucas, 2000).

in contrast to the last ones, research in automatic SL processing has not yet managed to build a complete, formal definition oriented to their automatic recognition (Cuxac and Dalle, 2007).

contrasting

train_2652

As the table shows, increasing the number of folds results in higher BLEU scores.

doing such will generally lead to higher variance among base learners.

contrasting

train_2653

Although the translation is correct in this situation, translating the Chinese word "lingyu" to "waters" appears very few times since the common translations are "areas" or "fields".

simply filtering out this kind of sentence pairs may lead to some loss of native English expressions, thereby the trans-lation performance is unstable since both nonparallel sentence pairs and non-literal but parallel sentence pairs are filtered.

contrasting

train_2654

Although the grammar may have rules to translate these two phrases, they can be safely pruned for this particular sentence pair.

to chart pruning for monolingual parsing, our pruning decisions are based on the source context, its target translation and the mapping between the two.

contrasting

train_2655

Moreover, for the frame "攻占" in the input sentence, the MEANT-tuned system has correctly translated the ARG0 "哈玛斯好战份子" into "Hamas militants" and the ARG1 "加萨走廊" into "Gaza".

the TER-tuned system has dropped the predicate "施行" so that the corresponding arguments "The Palestinian Authority" and "into a state of emergency" have all been incorrectly associated with the predicate "攻占 /seized".

contrasting

train_2656

In answer ranking phase, TI-QA considers the probabilities of different answer types as well: On one hand, TD-QA can achieve relative high ranking precision, as using a unique answer type greatly reduces the size of the candidate list for ranking.

as the answer-typing model is far from perfect, if prediction errors happen, TD-QA can no longer give correct answers at all.

contrasting

train_2657

However, as the answer-typing model is far from perfect, if prediction errors happen, TD-QA can no longer give correct answers at all.

tI-QA can provide higher answer coverage, as it can extract answer candidates with multiple answer types.

contrasting

train_2658

Since question-answer pairs are usually short, the word mismatching problem is especially important.

due to the lexical gap between questions and answers as well as spam typically existing in user-generated content, filtering and ranking answers is very challenging.

contrasting

train_2659

For example, they convert a 1, 000, 000-dimensional vector of word space into a 1000 × 1000 matrix.

in our model, a document is still represented by a vector.

contrasting

train_2660

For instance, in the British National Corpus, time and see are more frequent than thing or may and man is more frequent than part.

it seems intuitively right to say that time, see and man are more 'precise' concepts than thing, may and part respectively.

contrasting

train_2661

1 In Section 2, we defined semantic content as a notion encompassing various referential properties, including a basic concept of extension in cases where it is applicable.

we do not know of a dataset providing human judgements over the general informativeness of lexical items.

contrasting

train_2662

In biased-SVM, it is necessary to run SVM many times, as we searched "c" and "j".

mCDC does not require such parameter tuning.

contrasting

train_2663

Antonio Roque (2012) has challenged Ramsay's claim, and certainly there has been successful work done in the computational analysis and modeling of narratives, as we will review in the next section.

we believe that most previous work (except possibly (Elsner, 2012)) has failed to directly address the root cause of Ramsay's skepticism: can computers extract the emotions encoded in a narrative?

contrasting

train_2664

Not surprisingly, Claudius draws the most negative sentiment from Hamlet, receiving a score of -27.

gertrude is very well liked by Hamlet (+24), which is unexpected (at least to us) since Hamlet suspects that his mother was involved in murdering King Hamlet.

contrasting

train_2665

In the case of deep neural embeddings, for example, training time can number in days.

learned embeddings are becoming more abundant, as much research and computing effort is being invested in learning word representations using large-scale deep architectures trained on web-scale corpora.

contrasting

train_2666

min (φ iφj − φ i φ j ) 2 ∀i, j (where, with some abuse of notation, φ andφ are the source and target embeddings respectively).

this objective is no longer convex in the embeddings.

contrasting

train_2667

Adj + NP It is common practice to extract any NP modified by a sentiment adjective.

this simple extraction rule suffers from precision problems.

contrasting

train_2668

Quotes are used in news articles as evidence of a person's opinion, and thus are a useful target for opinion mining.

labelling each quote with a polarity score directed at a textually-anchored target can ignore the broader issue that the speaker is commenting on.

contrasting

train_2669

In sentiment analysis over product reviews, polarity labels are commonly used because the target, the product, is clearly identified.

for quotes on topics of debate, the target and meaning of polarity labels is less clear.

contrasting

train_2670

These approaches assume that each document has a single source (the document's author), whose communicative goal is to evaluate a well-defined target, such as a product or a movie.

this does not hold in news articles, where the goal of the journalist is to present the viewpoints of potentially many people.

contrasting

train_2671

Bag-of-words (BOW) is now the most popular way to model text in machine learning based sentiment classification.

the performance of such approach sometimes remains rather limited due to some fundamental deficiencies of the BOW model.

contrasting

train_2672

2012;Shi et al., 2010;Prettenhofer and Stein 2010).

the binary classification task is very different from the regression task studied in this paper, and the proposed methods in the above previous works cannot be directly applied.

contrasting

train_2673

REG_STC: It combines REG_S and REG_T by averaging their prediction values.

the above regression methods do not perform very well due to the unsatisfactory machine translation quality and the various language expressions.

contrasting

train_2674

Thus, each example in the unlabeled set is required to be checked by training a new regression model utilizing the example.

the model training process is usually very time-consuming for many regression algorithms, which significantly limits the use of the work in (Zhou and Li, 2005).

contrasting

train_2675

Suppose we would have sense tagged data, p(S|w, c) could have been computed as: But since the sense tagged corpus is not available, we cannot find #(S, w, c) from the corpus directly.

we can estimate it using the comparable corpus in other language.

contrasting

train_2676

Therefore, combining additional information with the LMs could reduce recognition errors.

direct integration of such information in the decoder is difficult.

contrasting

train_2677

An available treebank is a major resource for syntactic parsing.

it is often a key bottleneck to acquire credible treebanks.

contrasting

train_2678

Future systems might use the same or a similar feature set to ours, but in an architecture that does not include any generative parser.

some systems might indeed incorporate this generative model's score.

contrasting

train_2679

But, Φ phrase+deps+gen is significantly better than Φ phrase+gen only on F 1 , but not on UAS or LAS.

on the out-of-domain BROWN tests, we find that adding Φ deps always adds considerably, and in a statistically significant way, to both LAS and UAS.

contrasting

train_2680

This result is perhaps counter-intuitive, in the sense that one might have supposed that higher-order dependency features, being highly specific by nature, might only have only served to over-fit the training material.

this result shows otherwise.

contrasting

train_2681

While closed-form solutions have been developed for some specialized components (Martins et al., 2011), this problem is in general more difficult than the one arising in the subgradient algorithm.

the following result, proved in Martins et al.

contrasting

train_2682

We present an unsupervised approach to part-of-speech tagging based on projections of tags in a word-aligned bilingual parallel corpus.

to the existing state-of-the-art approach of Das and Petrov, we have developed a substantially simpler method by automatically identifying "good" training sentences from the parallel corpus and applying self-training.

contrasting

train_2683

(2012) build supervised POS taggers for 22 languages using the TNT tagger (Brants, 2000), with an average accuracy of 95.2%.

many widelyspoken languages -including Bengali, Javanese, and Lahnda -have little data manually labelled for POS, limiting supervised approaches to POS tagging for these languages.

contrasting

train_2684

Algorithm 2 describes this process of self training and revision, and assumes that the parallel source-target corpus has been word aligned, with many-to-one alignments removed, and that the sentences are sorted by alignment score.

to Algorithm 1, all sentences are used, not just the 60k sentences with the highest alignment scores.

contrasting

train_2685

Finally, the keywords S ⊆ t, with |S| ≤ k, are chosen by maximizing the cumulative reward function over all the topics, formulated as follows: Since R(S) is submodular, the greedy algorithm for maximizing R(S) is shown as Algorithm 1 on the next page, with r {w},z being similar to r S,z with S = {w}.

if λ = 1, the reward function is linear and only measures the topical similarity of words with the main topics of t. when 0 < λ < 1, as soon as a word is selected from a topic, other words from the same topic start having diminishing gains.

contrasting

train_2686

They find that these features only contribute less than 2% to precision.

in our approach linguistic features are quite useful.

contrasting

train_2687

Relatively low perplexity has made modified Kneser-Ney smoothing (Kneser and Ney, 1995;Chen and Goodman, 1998) a popular choice for language modeling.

existing estimation methods require either large amounts of RAM (Stolcke, 2002) or machines (Brants et al., 2007).

contrasting

train_2688

We have found that, even with mostlysequential access, memory mapping is slower because the kernel does not explicitly know where to read ahead or write behind.

we use dedicated threads for reading and writing.

contrasting

train_2689

We note that interpolation (Equation 2) used the different backoff b(w n−1 1 ) and so b(w n 1 ) is not immediately available.

the backoff values were saved in suffix order ( §3.3) and interpolation produces probabilities in suffix order.

contrasting

train_2690

To avoid the need for hard decisions about domain membership, some have used topic modeling to improve SMT performance, e.g., using latent semantic analysis (Tam et al., 2007) or 'biTAM' (Zhao and Xing, 2006).

to our source language approach, these authors use both source and target information.

contrasting

train_2691

For instance, in the task of tagging mail addresses, a feature of "5 consecutive digits" is highly indicative of a POSTCODE.

in the alignment model, it does not make sense to design features based on a hard-coded state, say, a feature of "source word lemma matching target word lemma" fires for state index 6.

contrasting

train_2692

While the participants in that study were younger (7 to 9+ years old), the study is relevant because the challenges those young participants face, are faced again when readers of any age encounter new and complicated texts that present words they do not know, and ideas they have never considered.

there is ample work on the basic algorithm to place a sequence of words in a typesetting area with a certain width, commonly known as the optimal line breaking problem (e.g., , Knuth and Plass (1981)).

contrasting

train_2693

It should be noted that when a text is typeset into an area of width of a certain number of characters, an erroneous break need not necessarily lead to an actual break in the final output, that is an error may not be too bad.

a missed break while not hurting the readability of the text may actually lead to a long segment that may eventually worsen raggedness in the final typesetting.

contrasting

train_2694

Each example has more than one annotation and we need to determine a single sense tag for each example.

if we assign senses by majority voting, we need a backoff strategy in case of ties.

contrasting

train_2695

Columns V and B show respectively whether the underspecified senses are a result of majority voting or backoff.

to volunteers, turkers disprefer the underspecified option and most of the English underspecified senses are assigned by backoff.

contrasting

train_2696

Syntax-based vector spaces are used widely in lexical semantics and are more versatile than word-based spaces (Baroni and Lenci, 2010).

they are also sparse, with resulting reliability and coverage problems.

contrasting

train_2697

The sentence Karen threw her arms round my neck, spilling champagne everywhere contains the LU throw.v evoking the frame BODY MOVEMENT.

throw.v is ambiguous and may also evoke CAUSE MOTION.

contrasting

train_2698

The evaluation of whole-sentence semantic structures plays an important role in semantic parsing and large-scale semantic structure annotation.

there is no widely-used metric to evaluate wholesentence semantic structures.

contrasting

train_2699

Evaluating such structures is necessary for semantic parsing tasks, as well as semantic annotation tasks which create linguistic resources for semantic parsing.

there is no widely-used evaluation method for whole-sentence semantic structures.

contrasting