tasksource/scinli · Datasets at Hugging Face

id stringlengths 7 12	sentence1 stringlengths 6 1.27k	sentence2 stringlengths 6 926	label stringclasses 4 values
train_8600	Therefore, the number of big PASs used in PAS(BTG) reduces.	the child PAS in PAS+BTG is only a choice but not essential for translating its father PAS.	contrasting
train_8601	Furthermore, in the second example, our PASbased method successfully recognizes the [AM-TMP] argument "2005年" and move it to the end of sentence.	the BTG system only performs translation without any reordering.	contrasting
train_8602	Parse trees are indispensable to the existing tree-based translation models.	there exist two major challenges in utilizing parse trees: 1) For most language pairs, it is hard to get parse trees due to the lack of syntactic resources for training.	contrasting
train_8603	The parse tree is usually generated by a linguistic parser which is trained on a manually annotated corpus, such as Treebank.	the manually annotated corpus is always too inadequate to fully display the strengths of tree-based models.	contrasting
train_8604	Theoretically, in a sub-sentence pair, all included words cannot align to words outside it.	since many words are wrongly aligned via the automatic word alignment, numerous correct aligned sub-sentence pairs are often excluded under this restriction.	contrasting
train_8605	Bilingual sentence segmentation leads to a great space reduction for constructing packed forests.	even after sentence segmentation, the generation space of tree structures would be still very large, especially when the sub-sentence is very long.	contrasting
train_8606	The reason is that as a sequence labeling model, CRF gives a most probable abbreviation character sequence by analyzing local information for each character.	for Chinese abbreviations, local information alone is not adequate.	contrasting
train_8607	The reason is that CRF+AEPW tries to extract information between well-formed full forms and also well-formed abbreviations.	in the current Chinese abbreviation generation process, some ill-formed candidates may be generated, like include illegal terms and common phrases which are in fact substrings of the full form.	contrasting
train_8608	They also reported their results on CTB5.1 data set.	we didn't make a comparison as they employed a different conversion method for CS-to-DS.	contrasting
train_8609	These knowledge bases have been shown to form a valuable component for many natural language processing tasks such as knowledge base population , text classification (Wang and Domeniconi, 2008), and cross-document coreference (Finin et al., 2009).	to be able to utilize or enrich these KB resources, the applications usually require linking the mentions of entities in text to their corresponding entries in the knowledge bases, which is called entity linking task and has been proposed and studied in Text Analysis Conference (TAC) since 2009 (McNamee and Dang, 2009).	contrasting
train_8610	Given the query, the previous systems are static at evaluation time and their training does not depend on the input query.	figure 2 illustrates the distribution of the labeled instances related to the three names in a feature space (Bag of Words, Named Entities and Edit Distance, the popular features used in previous work).	contrasting
train_8611	Thus, in this benchmark, the set of company names in the training and test corpora are different.	the lazy learning approach proposed in this paper demonstrates that it is feasible to train separate system for each company, and the system can immediately react to any company name without manually labeling new corpora.	contrasting
train_8612	However, for some queried names, it is hard to find a sufficient number of unambiguous synonyms or the related documents containing these synonyms.	the total number of available manually labeled instances M for other irrelevant names is relatively large.	contrasting
train_8613	Then, we decompose the weight vector u l for problem l into two parts: one part that models the distribution knowledge specific to each problem l and one part that models the common predictive structure, where w l and v l are weight vectors specific to each prediction problem l. Then, the parameters Θ, w l and v l can be learned by joint empirical risk minimization, i.e., by minimizing the joint empirical loss of the predictors for the K problems on the training instances as Eq.	5, It shows that w l and v l are estimated on n (l) training instances of problem l. Θ is estimated on all the training instances of the K problems.	contrasting
train_8614	As previous entity linking systems do not consider the information of the queried name, they usually use all the instances in M without any difference to train the linker.	in data set M, some instances related with some particular names may share more predictive information with the queried name than other instances.	contrasting
train_8615	Traditional approaches usually need an additional classification step to resolve this problem (Zheng et al., 2010;Lehmann et al., 2010).	our approach seamlessly takes into account the NIL prediction problem.	contrasting
train_8616	Hence, comparing with previous work, it increases the response time of the system for each query.	many of the entity linking applications such as knowledge base population do not require real-time user interaction, and therefore they are time-insensitive applications.	contrasting
train_8617	, it is obvious that the term "charge" is the direct neighbour of term "capacitor", hence, the dependency relation path length d r_path_l en equals to 1 as shown in Figure 1(b) (1).	"charge" is a bit farther away from the term "farad" as the d r_path_l en between term "charge" and "farad" equals to 2 as shown in Figure 1(b) (3).	contrasting
train_8618	(Ming et al., 2010) utilized three domain specific metrics to explore term weights and integrated them into existing IR models.	most of these term weighting based retrieval models ignore the dependency relations between term pairs.	contrasting
train_8619	In previous work on morphological learning, such as (Chan, 2008), (Zhao and Marcus, 2011) and (Goldwater et al., 2006(Goldwater et al., , 2011, the straight lines on logistic scales are usually interpreted as following Zipf's law (Zipf, 1949).	lognormal distributions with large variance also yield straight lines on both the log-log rank-frequency plot and the CCDF plot.	contrasting
train_8620	The rule-based bootstrapping algorithm utilizes type frequencies only, no matter what form of input is given.	as shown in (Goldwater et al., 2006), when Dirichlet-multinomial model is assumed, the option of utilizing token frequency in generative model doesn't help, and in the contrast it hurts the inference of the generative model.	contrasting
train_8621	For example, as shown in Section 3.1, the bootstrapping algorithm measures contextual diversity with type frequency only.	natural text data typically use most types of words more than once.	contrasting
train_8622	We have shown the acquisition outputs of the bootstrapping algorithm in Table 1, upon which we can build a rule-based segmentation model.	to the learning progress of generative models, which will converge to a relatively steady state, we stop the acquisition process after 20 bootstrapping iterations following previous experiments.	contrasting
train_8623	Search log sessions contain a large number of paraphrases contributed by users during query rewriting.	it is a big challenge to distinguish paraphrases from the simply related queries in the sessions.	contrasting
train_8624	On the one hand, related queries can be used for query suggestion and recommendation, using which the users can extend their search interest and get some related information.	the related queries are not suitable for direct query rewriting with the purpose of retrieving more and better results exactly reflecting the user's requirement, since the related queries often have different meanings from the original user query.	contrasting
train_8625	Our work is close to the third group.	what we learn are paraphrase patterns rather than related queries or patterns.	contrasting
train_8626	The third class, i.e., spelling correction was not conventionally regarded as a type of paraphrase.	pattern pairs of this class actually convey the same requirement of users and are useful in applications such as query correction.	contrasting
train_8627	Question retrieval in CQA archives aims to find historical questions that are semantically equivalent or relevant to the queried questions.	question retrieval is challenging partly due to the word ambiguity and lexical gap between the queried questions and the historical questions in the archives.	contrasting
train_8628	Therefore, it is a meaningful task to retrieve the semantically equivalent or relevant questions to the queried questions.	question retrieval is challenging partly due to the word ambiguity and lexical gap between the queried questions and the historical questions in the archives.	contrasting
train_8629	(2010) enriched the original word-based representation with a concept-based representation, thereby proposing the translation of the original word language to a concept language.	their translation models are based solely on the use of translation at the lexical level (e.g., word-to-concept), and thus their method is very different from our context-dependent style of translation.	contrasting
train_8630	Na and Ng (2011) also applied automatic translation for monolingual retrieval.	they used the expected frequency of a word computed from all possible translated representations, while we use the state-of-the-art commercial machine translation service (e.g., Google Tranlate), which is much simpler than their translation strategies.	contrasting
train_8631	Moreover, these studies performed translation without taking into account the context information of an original word (Chen and Gey, 2004;Kraaij et al., 2003).	our approach is contextdependent and thus produces different translated words depending on the context of a word in original language.	contrasting
train_8632	In dependency and CCG parsing, shift-reduce parsing is among the best-performing algorithms (Huang and Sagae, 2010;Zhang and Clark, 2011).	compared to commonly-used statistical parsers available on the web such as Charniak-Johnson (Charniak and Johnson, 2005) and Petrov-Klein (Petrov and Klein, 2007), shift-reduce constituency parsers still have room left for further improvements on parsing accuracy.	contrasting
train_8633	that, on Chinese parsing our parser outperforms Bikel (2004) and Charniak (2000) by 0.6% and 0.4%, respectively.	our parser lags behind Petrov and Klein (2007) and the reranking parser (Charniak and Johnson, 2005).	contrasting
train_8634	Supervised computational systems can be trained to mark up the AZ structure of a text automatically (see Section 2); the output of such systems has been shown to aid summarisation and human browsing of the scientific literature (Teufel and Moens, 2002;Guo et al., 2011a;Contractor et al., 2012).	supervised systems require manually annotated training data that must be created anew for each discipline (and language) before they can be deployed, while large quantities of unannotated text are often available.	contrasting
train_8635	Moreover, to solve this issue using a supervised learner, one needs the gold standard of coreference at least on the target side of the bitext.	given such data, the typological differences in languages can be exploited to aid a CR system to perform better than if CR is performed independently for each language.	contrasting
train_8636	The alignment described in the previous section is sufficiently accurate for content words, such as verbs, nouns, and adjectives.	errors become more frequent as we move to pronouns.	contrasting
train_8637	It shows that more than 55% of English personal pronouns are dropped from the surface representation of the Czech sentence, though still present in its deep structure.	english pleonastic pronouns are not present even there.	contrasting
train_8638	The CR system we use definitely does not aim to compete with current state-of-the-art systems.	for the purpose of research on crosslingual CR, it can be employed as a reasonable baseline.	contrasting
train_8639	Related studies have focused on detecting noteworthiness from meeting transcripts (Banerjee and Rudnicky, 2009).	very little work has been done to date to identify this kind of information in other types of human communication, such as spontaneous phone conversations.	contrasting
train_8640	The advent of online social networks has produced a crescent interest on the task of sentiment analysis for short text messages (Go et al., 2009;Barbosa and Feng, 2010;Nakov et al., 2013).	sentiment analysis of short texts such as single sentences and and microblogging posts, like Twitter messages, is challenging because of the limited amount of contextual data in this type of text.	contrasting
train_8641	In our experiments we focus in sentiment prediction of complete sentences.	we show the impact of training with sentences and phrases instead of only sentences.	contrasting
train_8642	In our experiments, we check whether using examples that are single phrases, in addition to complete sentences, can provide useful information for training the proposed NN.	in our experiments the test set always includes only complete sentences.	contrasting
train_8643	Previous work in NLP on sentiment analysis has mainly focused on explicit sentiments.	as noted in , many opinions are expressed implicitly, as shown by this example: Ex(1) The reform would lower health care costs, which would be a tremendous positive change across the entire health-care system.	contrasting
train_8644	There is an explicit positive sentiment toward the event of "reform lower costs".	in expressing this sentiment, the writer also implies he is negative toward the "costs", since he's happy to see the costs being decreased.	contrasting
train_8645	Different from their work, which do not cover all cases relevant to gfbf events, defines a generalized set of implicature rules and proposes a graph-based model to achieve sentiment propagation between the agents and themes of gfbf events.	that system requires all of the gfbf information (Q1)-(Q4) to be input from the manual annotations; the only ambiguity it resolves is sentiments toward entities.	contrasting
train_8646	Similar comments apply to Equation (7).	the writer has opposite sentiments toward entities in a bf relation.	contrasting
train_8647	That is, the system may know businesses in Santa Fe, but not whether they sell roasted chiles.	collectively people themselves know this type of information, and they frequently mention it in social media.	contrasting
train_8648	The recently proposed methods, such as CBOW and Skip-gram, have demonstrated their effectiveness in learning word embeddings based on context information such that the obtained word embeddings can capture both semantic and syntactic relationships between words.	it is quite challenging to produce high-quality word representations for rare or unknown words due to their insufficient context information.	contrasting
train_8649	Suitable initialization will help increase the embedding quality which works like training with multi-epochs.	as there are two matrix M and M in our network structure, the initialization of both of them are more sensible.	contrasting
train_8650	Some recent studies attempted to train multi-prototype word embeddings through clustering context window features of the word.	due to a large number of parameters to train, these methods yield limited scalability and are inefficient to be trained with big data.	contrasting
train_8651	Compared with traditional single prototype model, these models have demonstrated significant improvements in many semantic natural language processing (NLP) tasks.	they suffer from a crucial restriction in terms of scalability when facing exploding training text corpus, mainly due to the deep layers and huge amounts of parameters in the neural networks in these models.	contrasting
train_8652	Compared with conventional neural network language models which usually set up a multi-layer neural network, Word2Vec merely leverages a threelayer neural network to learn word embeddings, resulting in greatly decreased number of parameters and largely increased scalability.	similar to most of existing word embedding models, Word2Vec also assumes one embedding for one word.	contrasting
train_8653	We also conduct a comparison on the number of parameters between the new EM algorithm and the stateof-the-art multi-prototype model proposed in (Huang et al., 2012), which can illustrate the efficiency superior of our algorithm.	to the conventional ways of using context words to predict the next word or the central word, the Skip-Gram model (Mikolov et al., 2013b) aims to leverage the central word to predict its context words.	contrasting
train_8654	Now imagine that in the corpus used to train the parser none of these nouns have been observed, then it is unlikely that these attachments can be resolved correctly.	if an accurate noun-adjective bilexical operator were available most of the uncertainty could be resolved.	contrasting
train_8655	The advantage of this approach is that as long as we can estimate the distribution of contexts of words we can compute the value of the bilexical operator.	this approach has a clear limitation: to design a bilinear operator for a target linguistic relation we must design the appropriate distributional representation.	contrasting
train_8656	Traditionally, grammar-based methods have been used for CSL, but more recently machine learning approaches to semantic structure computation have been shown to yield higher accuracy.	most previous work did not exploit syntactic/semantic structures of the utterances, and the state-of-the-art is represented by conditional models for sequence labeling, such as Conditional Random Fields (Lafferty et al., 2001) trained with simple morphological and lexical features.	contrasting
train_8657	Hashtags have been proven to be useful for many applications, including microblog retrieval (Efron, 2010), query expansion (A. Bandyopadhyay et al., 2011), sentiment analysis (Davidov et al., 2010;Wang et al., 2011).	only a few microblogs contain hashtags provided by their authors.	contrasting
train_8658	Our motivation to test multiple classifiers stemmed also from related works which mostly test more than one classifier.	the choice between state-of-the-art linear classifiers might not be much of importance, as the most important is the feature engineering.	contrasting
train_8659	Czech is a highly flective language and uses a lot of diacritics.	some Czech users type only the unaccented characters.	contrasting
train_8660	We adapted all their preprocessing pipelines.	as the number of combinations would be too large, we report only the settings with better performance.	contrasting
train_8661	occurrence > 5) + extended pointedness; FS2: POS word-shape + pattern + POS characteristics + emoticons + word-case.	imbalanced distribution data in the real world do not necessarily resemble the balanced distribution.	contrasting
train_8662	Thus, it means that learning from different views should be softly regularized towards a common latent structure.	it is not easy to directly formulate it in a probabilistic framework, because weak consensus modeling can not be separated from a joint higher task, i.e., recovering spare rating matrix, in our case.	contrasting
train_8663	The generative process is described as follows: , and set the item latent vector as -draw the response , univariate Gauss distribution, where c ij is a confidence parameter for rating r ij , a > b. c ij = a (higher confidence), if r ij = 1, and c ij = b, if r ij = 0.	cTR does not take the complex social network information, which is available and crucial in many real-world applications, into consideration.	contrasting
train_8664	In addition, we can see that CTR-smf (Purushotham et al., 2012) is sensitive to the quality of graph (SMF-1 with low quality and SMF-2 with high quality as shown in Figure 2).	we can use the low quality noisy graph (SMF-1) to improve the overall performance by this transformation process.	contrasting
train_8665	OmegaWiki is a freely editable online dictionary like Wiktionary.	instead of distinct language editions, OmegaWiki contains language-independent concepts ("Defined Meanings") which carry lexicalizations in different languages.	contrasting
train_8666	Wiktionary-Wikipedia (English): No evaluation dataset (let alone a full alignment) has been reported for this resource pair yet.	as the datasets for WordNet-Wiktionary (Meyer and and WordNet-Wikipedia (Niemann and Gurevych, 2011) are lexically overlapping, we were able to automatically create a gold standard for Wiktionary-Wikipedia by exploiting the transitivity of the alignment relation, i.e.	contrasting
train_8667	We were able to create a gold standard in a novel way by exploiting the fact that many German Wiktionary senses contain links to the corresponding Wikipedia articles, inducing a sense alignment between the two LSRs manually validated by the Wiktionary community.	we were unable to extract such an alignment for English, as Wikipedia articles are attached to the lexical entry page in this version and not to a specific sense.	contrasting
train_8668	(Bond and Foster, 2013)) and overlap of domain labels (Wikipedia, Wiktionary, Word-Net, OmegaWiki).	for none of these features we could observe any significant 1 impact on the results, mostly due to sparsity of the respective features.	contrasting
train_8669	For conversion from PS to DS, a head-table approach (Magerman, 1994;Collins, 2003;Yamada and Matsumoto, 2003;Sun and Jurafsky, 2004;Nivre, 2006;Johansson and Nugues, 2007;Duan et al., 2007;Zhang and Clark, 2008) is widely used.	the reliability of head tables has been questioned (Xue, 2007).	contrasting
train_8670	Secondly, we detect and annotate clause pairs in a document that hold logical discourse relations.	since this is too complicated to assign as one task using crowdsourcing, we divide the task into two steps: determining the existence of logical discourse relations and annotating the type of relation.	contrasting
train_8671	For the examples in Table 5, we confirmed that the discourse relation types of the top four examples were surely correct.	we judged the type (Contrast) of the bottom example as incorrect.	contrasting
train_8672	On the one hand, using words as the reordering units will reduce the number of candidates generated.	word segmentation results will affect the performance of WOE detection and correction.	contrasting
train_8673	If a segment without WOEs is misjudged to be erroneous, the word order still has a chance to be kept by the WOE correction models.	if a segment with WOEs is misjudged to be correct, words in the misjudged segment will not be reordered in the correction part because the error correction module is not triggered.	contrasting
train_8674	Clearly, the recall at the segment level and the correctable rate at the sentence level are 1 by the all-tag-B baseline.	its accuracy at the segment and the sentence levels are low.	contrasting
train_8675	It means the correct candidates are ranked 3.7 on the average.	the MRR by using the C sys dataset is 0.208.	contrasting
train_8676	We address the task of automatic terminology acquisition (ATA), the task of finding technical terms in texts without reliance on existing resources that list terms of the domain.	to this stands automatic terminology recognition (ATR), which we define as finding known terms and their variants (Jacquemin and Bourigault, 2003).	contrasting
train_8677	This is similar to distant supervision (Mintz et al., 2009) which also uses pre-existing resources such as gazetteers for, e.g., relation extraction.	our method is applied to ATA for the technological domain and does not rely on precompiled resources -we make use of figure references, which are an inherent part of patents.	contrasting
train_8678	Our method can be characterized as training data identification: we exploit given conditions in patents for our search of training data.	training data recognition methods need precompiled resources as input and search for instances of resource elements in texts.	contrasting
train_8679	Like our first baseline, it needs no training data.	to our first baseline, it was specifically designed for terminology acquisition.	contrasting
train_8680	We would also like to evaluate candidate classification on gold boundaries (manually verified boundaries of term candidates); this allows us to quantify by how much performance can be improved if candidate identification is perfect.	since gold boundary annotation is expensive, we instead approximated it: (i) We run automatic term candidate identification.	contrasting
train_8681	All observations hold for T l dev and T l test .	numbers are higher for T l dev because the ratio of FRTCs to candidates is higher than for T l test (38% vs. 27%) which improves classification performance on T l dev -this holds for ATAS as well as for the baselines.	contrasting
train_8682	In such cases the figure reference serves as a disambiguator.	in other positions they are non-terms, e.g., "They include braces, collars, splints and other similar apparatus".	contrasting
train_8683	We consider that selecting a value of n grater than 4 could lead to find few n-grams, so that many web pages could be under-represented.	previous experiments using also bigrams showed that they are not suitable for this approach.	contrasting
train_8684	According to the model, one of the most prominent distinctions is that amateurs tend to refer to the opponent as he, whereas experts use white and black more frequently.	it is of course not universally true, which leads to the misclassification of some experts as amateurs.	contrasting
train_8685	This seems counterintuitive at first, as we may expect lower-rated players to be less familiar with such terms.	it appears that they are frequently overused by weaker players.	contrasting
train_8686	(2003) presented the theories and approaches to calculate the frequencies of Tibetan characters, pieces, syllables and words based on a large scale Tibetan corpus including about 40, 000, 000 syllables (Lu et al., 2003).	a large part of the corpus they used are Buddhist literatures and the work can't be done well without a pragmatic Tibetan word segmentation tool (Chen et al., 2003a;Chen et al., 2003b;Jiang, 2006;Jiang and Kong, 2006;Sun et al., 2009;Sun et al., 2010;Lu and Shi, 2011;Liu et al., 2012a).	contrasting
train_8687	The sources and scales in different units are shown in It's a heavy task to manually classify those document into domains.	we still can get the domain information for a certain subsets of the corpus.	contrasting
train_8688	maturity, background of information etc."	we do not know of any such investigations for Bangla text readability that have investigate the way background of a reader affect the readability of text.	contrasting
train_8689	This group of models moves beyond the surface features of a text and try to measure objectively the different cognitive indicators associated with text and the reader.	it has been observed that, many situations, some traditional indicators perform as well as the newer and more difficult versions (Crossley et al., 2007).	contrasting
train_8690	For adult data, it can be seen that this feature has a strong and significant correlation, which not true for the user data of group 2 for separate texts.	for the common texts this feature was found to have high significant correlation with both the reader groups.	contrasting
train_8691	The correlations are also significant.	the correlations with $(noun phrase), $(verb phrase) $(postpositions), #(postpositions), #(adjective) were found to be insignificant.	contrasting
train_8692	Different with normal practice in WSI work, there is no feature engineering in our model.	our BNP model outperformed all the systems on supervised evaluation.	contrasting
train_8693	Those previous works on combining TM and SMT can be classified into four categories: (1) selecting the better translation sentence from TM and SMT (He et al., 2010a;2010b;Dara et al., 2013); (2) incorporating TM matched sub-segments into SMT in a pipelined manner (Koehn and Senellart, 2010;; (3) only enhancing the SMT phrase table with new TM phrase-pairs (Biç ici and Dymetman, 2008;Simard and Isabelle, 2009); and (4) incorporating the associated TM information with each source phrase to guide the SMT decoding .	all previous works mentioned above only focus on the case in which the TM database and the SMT training set share the same data-set.	contrasting
train_8694	Therefore, the factor ( \| ) from the test set will possess a different probability distribution in comparison with that from the training set.	the development set is not big enough (only a few hundreds sentence-pairs at each interval) to re-train all TM factors of the proposed model.	contrasting
train_8695	For example, at interval [0.9, 1.0), those added TM phrase-pairs significantly improve the SMT system from 63.65 to 73.55, and Model-III from 80.69 to 86.40.	if Model-III + is compared with Model-III, the improvements from merging the TM phrase-pairs get less when the fuzzy match score decreases, because the matched TM parts are fewer at low fuzzy match intervals.	contrasting
train_8696	However, it is still worse than the TM in TER (13.32 vs. 10.42).	although Model-III has greatly exceeded the SMT at each interval, Model-III + still significantly outperforms Model-III at most intervals.	contrasting
train_8697	In real environments, the SMT training set and the TM database could be the same before translation projects starts.	the TM database will gradually deviate from the SMT training set while the translation task progresses.	contrasting
train_8698	The useful reordering pattern learned through this example is: Ich kann umstellen → I can rearrange which is memorized through the operation sequence: Generate(Ich, I) -Generate(kann, can) -Insert Gap -Generate(umstellen, rearrange) It can generalize to the test sentence shown in Figure 1(a).	it fails to generalize to the sentences in Figure 1(b) and (c) although the underlying reordering pattern is the same.	contrasting
train_8699	All three measures showed divergence in message similarity both between individuals, and in the community as a whole, across time.	the common theme with all these techniques is that although they can effectively measure adaptation of linguistic feature use within and between dialogs, they fail to capture the precise direction of convergence or divergence between individuals (i.e., do both interactants within a conversational pair accommodate their language use to the same extent?)	contrasting

train_8600

Therefore, the number of big PASs used in PAS(BTG) reduces.

the child PAS in PAS+BTG is only a choice but not essential for translating its father PAS.

contrasting

train_8601

Furthermore, in the second example, our PASbased method successfully recognizes the [AM-TMP] argument "2005年" and move it to the end of sentence.

the BTG system only performs translation without any reordering.

contrasting

train_8602

Parse trees are indispensable to the existing tree-based translation models.

there exist two major challenges in utilizing parse trees: 1) For most language pairs, it is hard to get parse trees due to the lack of syntactic resources for training.

contrasting

train_8603

The parse tree is usually generated by a linguistic parser which is trained on a manually annotated corpus, such as Treebank.

the manually annotated corpus is always too inadequate to fully display the strengths of tree-based models.

contrasting

train_8604

Theoretically, in a sub-sentence pair, all included words cannot align to words outside it.

since many words are wrongly aligned via the automatic word alignment, numerous correct aligned sub-sentence pairs are often excluded under this restriction.

contrasting

train_8605

Bilingual sentence segmentation leads to a great space reduction for constructing packed forests.

even after sentence segmentation, the generation space of tree structures would be still very large, especially when the sub-sentence is very long.

contrasting

train_8606

The reason is that as a sequence labeling model, CRF gives a most probable abbreviation character sequence by analyzing local information for each character.

for Chinese abbreviations, local information alone is not adequate.

contrasting

train_8607

The reason is that CRF+AEPW tries to extract information between well-formed full forms and also well-formed abbreviations.

in the current Chinese abbreviation generation process, some ill-formed candidates may be generated, like include illegal terms and common phrases which are in fact substrings of the full form.

contrasting

train_8608

They also reported their results on CTB5.1 data set.

we didn't make a comparison as they employed a different conversion method for CS-to-DS.

contrasting

train_8609

These knowledge bases have been shown to form a valuable component for many natural language processing tasks such as knowledge base population , text classification (Wang and Domeniconi, 2008), and cross-document coreference (Finin et al., 2009).

to be able to utilize or enrich these KB resources, the applications usually require linking the mentions of entities in text to their corresponding entries in the knowledge bases, which is called entity linking task and has been proposed and studied in Text Analysis Conference (TAC) since 2009 (McNamee and Dang, 2009).

contrasting

train_8610

Given the query, the previous systems are static at evaluation time and their training does not depend on the input query.

figure 2 illustrates the distribution of the labeled instances related to the three names in a feature space (Bag of Words, Named Entities and Edit Distance, the popular features used in previous work).

contrasting

train_8611

Thus, in this benchmark, the set of company names in the training and test corpora are different.

the lazy learning approach proposed in this paper demonstrates that it is feasible to train separate system for each company, and the system can immediately react to any company name without manually labeling new corpora.

contrasting

train_8612

However, for some queried names, it is hard to find a sufficient number of unambiguous synonyms or the related documents containing these synonyms.

the total number of available manually labeled instances M for other irrelevant names is relatively large.

contrasting

train_8613

Then, we decompose the weight vector u l for problem l into two parts: one part that models the distribution knowledge specific to each problem l and one part that models the common predictive structure, where w l and v l are weight vectors specific to each prediction problem l. Then, the parameters Θ, w l and v l can be learned by joint empirical risk minimization, i.e., by minimizing the joint empirical loss of the predictors for the K problems on the training instances as Eq.

5, It shows that w l and v l are estimated on n (l) training instances of problem l. Θ is estimated on all the training instances of the K problems.

contrasting

train_8614

As previous entity linking systems do not consider the information of the queried name, they usually use all the instances in M without any difference to train the linker.

in data set M, some instances related with some particular names may share more predictive information with the queried name than other instances.

contrasting

train_8615

Traditional approaches usually need an additional classification step to resolve this problem (Zheng et al., 2010;Lehmann et al., 2010).

our approach seamlessly takes into account the NIL prediction problem.

contrasting

train_8616

Hence, comparing with previous work, it increases the response time of the system for each query.

many of the entity linking applications such as knowledge base population do not require real-time user interaction, and therefore they are time-insensitive applications.

contrasting

train_8617

, it is obvious that the term "charge" is the direct neighbour of term "capacitor", hence, the dependency relation path length d r_path_l en equals to 1 as shown in Figure 1(b) (1).

"charge" is a bit farther away from the term "farad" as the d r_path_l en between term "charge" and "farad" equals to 2 as shown in Figure 1(b) (3).

contrasting

train_8618

(Ming et al., 2010) utilized three domain specific metrics to explore term weights and integrated them into existing IR models.

most of these term weighting based retrieval models ignore the dependency relations between term pairs.

contrasting

train_8619

In previous work on morphological learning, such as (Chan, 2008), (Zhao and Marcus, 2011) and (Goldwater et al., 2006(Goldwater et al., , 2011, the straight lines on logistic scales are usually interpreted as following Zipf's law (Zipf, 1949).

lognormal distributions with large variance also yield straight lines on both the log-log rank-frequency plot and the CCDF plot.

contrasting

train_8620

The rule-based bootstrapping algorithm utilizes type frequencies only, no matter what form of input is given.

as shown in (Goldwater et al., 2006), when Dirichlet-multinomial model is assumed, the option of utilizing token frequency in generative model doesn't help, and in the contrast it hurts the inference of the generative model.

contrasting

train_8621

For example, as shown in Section 3.1, the bootstrapping algorithm measures contextual diversity with type frequency only.

natural text data typically use most types of words more than once.

contrasting

train_8622

We have shown the acquisition outputs of the bootstrapping algorithm in Table 1, upon which we can build a rule-based segmentation model.

to the learning progress of generative models, which will converge to a relatively steady state, we stop the acquisition process after 20 bootstrapping iterations following previous experiments.

contrasting

train_8623

Search log sessions contain a large number of paraphrases contributed by users during query rewriting.

it is a big challenge to distinguish paraphrases from the simply related queries in the sessions.

contrasting

train_8624

On the one hand, related queries can be used for query suggestion and recommendation, using which the users can extend their search interest and get some related information.

the related queries are not suitable for direct query rewriting with the purpose of retrieving more and better results exactly reflecting the user's requirement, since the related queries often have different meanings from the original user query.

contrasting

train_8625

Our work is close to the third group.

what we learn are paraphrase patterns rather than related queries or patterns.

contrasting

train_8626

The third class, i.e., spelling correction was not conventionally regarded as a type of paraphrase.

pattern pairs of this class actually convey the same requirement of users and are useful in applications such as query correction.

contrasting

train_8627

Question retrieval in CQA archives aims to find historical questions that are semantically equivalent or relevant to the queried questions.

question retrieval is challenging partly due to the word ambiguity and lexical gap between the queried questions and the historical questions in the archives.

contrasting

train_8628

Therefore, it is a meaningful task to retrieve the semantically equivalent or relevant questions to the queried questions.

question retrieval is challenging partly due to the word ambiguity and lexical gap between the queried questions and the historical questions in the archives.

contrasting

train_8629

(2010) enriched the original word-based representation with a concept-based representation, thereby proposing the translation of the original word language to a concept language.

their translation models are based solely on the use of translation at the lexical level (e.g., word-to-concept), and thus their method is very different from our context-dependent style of translation.

contrasting

train_8630

Na and Ng (2011) also applied automatic translation for monolingual retrieval.

they used the expected frequency of a word computed from all possible translated representations, while we use the state-of-the-art commercial machine translation service (e.g., Google Tranlate), which is much simpler than their translation strategies.

contrasting

train_8631

Moreover, these studies performed translation without taking into account the context information of an original word (Chen and Gey, 2004;Kraaij et al., 2003).

our approach is contextdependent and thus produces different translated words depending on the context of a word in original language.

contrasting

train_8632

In dependency and CCG parsing, shift-reduce parsing is among the best-performing algorithms (Huang and Sagae, 2010;Zhang and Clark, 2011).

compared to commonly-used statistical parsers available on the web such as Charniak-Johnson (Charniak and Johnson, 2005) and Petrov-Klein (Petrov and Klein, 2007), shift-reduce constituency parsers still have room left for further improvements on parsing accuracy.

contrasting

train_8633

that, on Chinese parsing our parser outperforms Bikel (2004) and Charniak (2000) by 0.6% and 0.4%, respectively.

our parser lags behind Petrov and Klein (2007) and the reranking parser (Charniak and Johnson, 2005).

contrasting

train_8634

Supervised computational systems can be trained to mark up the AZ structure of a text automatically (see Section 2); the output of such systems has been shown to aid summarisation and human browsing of the scientific literature (Teufel and Moens, 2002;Guo et al., 2011a;Contractor et al., 2012).

supervised systems require manually annotated training data that must be created anew for each discipline (and language) before they can be deployed, while large quantities of unannotated text are often available.

contrasting

train_8635

Moreover, to solve this issue using a supervised learner, one needs the gold standard of coreference at least on the target side of the bitext.

given such data, the typological differences in languages can be exploited to aid a CR system to perform better than if CR is performed independently for each language.

contrasting

train_8636

The alignment described in the previous section is sufficiently accurate for content words, such as verbs, nouns, and adjectives.

errors become more frequent as we move to pronouns.

contrasting

train_8637

It shows that more than 55% of English personal pronouns are dropped from the surface representation of the Czech sentence, though still present in its deep structure.

english pleonastic pronouns are not present even there.

contrasting

train_8638

The CR system we use definitely does not aim to compete with current state-of-the-art systems.

for the purpose of research on crosslingual CR, it can be employed as a reasonable baseline.

contrasting

train_8639

Related studies have focused on detecting noteworthiness from meeting transcripts (Banerjee and Rudnicky, 2009).

very little work has been done to date to identify this kind of information in other types of human communication, such as spontaneous phone conversations.

contrasting

train_8640

The advent of online social networks has produced a crescent interest on the task of sentiment analysis for short text messages (Go et al., 2009;Barbosa and Feng, 2010;Nakov et al., 2013).

sentiment analysis of short texts such as single sentences and and microblogging posts, like Twitter messages, is challenging because of the limited amount of contextual data in this type of text.

contrasting

train_8641

In our experiments we focus in sentiment prediction of complete sentences.

we show the impact of training with sentences and phrases instead of only sentences.

contrasting

train_8642

In our experiments, we check whether using examples that are single phrases, in addition to complete sentences, can provide useful information for training the proposed NN.

in our experiments the test set always includes only complete sentences.

contrasting

train_8643

Previous work in NLP on sentiment analysis has mainly focused on explicit sentiments.

as noted in , many opinions are expressed implicitly, as shown by this example: Ex(1) The reform would lower health care costs, which would be a tremendous positive change across the entire health-care system.

contrasting

train_8644

There is an explicit positive sentiment toward the event of "reform lower costs".

in expressing this sentiment, the writer also implies he is negative toward the "costs", since he's happy to see the costs being decreased.

contrasting

train_8645

Different from their work, which do not cover all cases relevant to gfbf events, defines a generalized set of implicature rules and proposes a graph-based model to achieve sentiment propagation between the agents and themes of gfbf events.

that system requires all of the gfbf information (Q1)-(Q4) to be input from the manual annotations; the only ambiguity it resolves is sentiments toward entities.

contrasting

train_8646

Similar comments apply to Equation (7).

the writer has opposite sentiments toward entities in a bf relation.

contrasting

train_8647

That is, the system may know businesses in Santa Fe, but not whether they sell roasted chiles.

collectively people themselves know this type of information, and they frequently mention it in social media.

contrasting

train_8648

The recently proposed methods, such as CBOW and Skip-gram, have demonstrated their effectiveness in learning word embeddings based on context information such that the obtained word embeddings can capture both semantic and syntactic relationships between words.

it is quite challenging to produce high-quality word representations for rare or unknown words due to their insufficient context information.

contrasting

train_8649

Suitable initialization will help increase the embedding quality which works like training with multi-epochs.

as there are two matrix M and M in our network structure, the initialization of both of them are more sensible.

contrasting

train_8650

Some recent studies attempted to train multi-prototype word embeddings through clustering context window features of the word.

due to a large number of parameters to train, these methods yield limited scalability and are inefficient to be trained with big data.

contrasting

train_8651

Compared with traditional single prototype model, these models have demonstrated significant improvements in many semantic natural language processing (NLP) tasks.

they suffer from a crucial restriction in terms of scalability when facing exploding training text corpus, mainly due to the deep layers and huge amounts of parameters in the neural networks in these models.

contrasting

train_8652

Compared with conventional neural network language models which usually set up a multi-layer neural network, Word2Vec merely leverages a threelayer neural network to learn word embeddings, resulting in greatly decreased number of parameters and largely increased scalability.

similar to most of existing word embedding models, Word2Vec also assumes one embedding for one word.

contrasting

train_8653

We also conduct a comparison on the number of parameters between the new EM algorithm and the stateof-the-art multi-prototype model proposed in (Huang et al., 2012), which can illustrate the efficiency superior of our algorithm.

to the conventional ways of using context words to predict the next word or the central word, the Skip-Gram model (Mikolov et al., 2013b) aims to leverage the central word to predict its context words.

contrasting

train_8654

Now imagine that in the corpus used to train the parser none of these nouns have been observed, then it is unlikely that these attachments can be resolved correctly.

if an accurate noun-adjective bilexical operator were available most of the uncertainty could be resolved.

contrasting

train_8655

The advantage of this approach is that as long as we can estimate the distribution of contexts of words we can compute the value of the bilexical operator.

this approach has a clear limitation: to design a bilinear operator for a target linguistic relation we must design the appropriate distributional representation.

contrasting

train_8656

Traditionally, grammar-based methods have been used for CSL, but more recently machine learning approaches to semantic structure computation have been shown to yield higher accuracy.

most previous work did not exploit syntactic/semantic structures of the utterances, and the state-of-the-art is represented by conditional models for sequence labeling, such as Conditional Random Fields (Lafferty et al., 2001) trained with simple morphological and lexical features.

contrasting

train_8657

Hashtags have been proven to be useful for many applications, including microblog retrieval (Efron, 2010), query expansion (A. Bandyopadhyay et al., 2011), sentiment analysis (Davidov et al., 2010;Wang et al., 2011).

only a few microblogs contain hashtags provided by their authors.

contrasting

train_8658

Our motivation to test multiple classifiers stemmed also from related works which mostly test more than one classifier.

the choice between state-of-the-art linear classifiers might not be much of importance, as the most important is the feature engineering.

contrasting

train_8659

Czech is a highly flective language and uses a lot of diacritics.

some Czech users type only the unaccented characters.

contrasting

train_8660

We adapted all their preprocessing pipelines.

as the number of combinations would be too large, we report only the settings with better performance.

contrasting

train_8661

occurrence > 5) + extended pointedness; FS2: POS word-shape + pattern + POS characteristics + emoticons + word-case.

imbalanced distribution data in the real world do not necessarily resemble the balanced distribution.

contrasting

train_8662

Thus, it means that learning from different views should be softly regularized towards a common latent structure.

it is not easy to directly formulate it in a probabilistic framework, because weak consensus modeling can not be separated from a joint higher task, i.e., recovering spare rating matrix, in our case.

contrasting

train_8663

The generative process is described as follows: , and set the item latent vector as -draw the response , univariate Gauss distribution, where c ij is a confidence parameter for rating r ij , a > b. c ij = a (higher confidence), if r ij = 1, and c ij = b, if r ij = 0.

cTR does not take the complex social network information, which is available and crucial in many real-world applications, into consideration.

contrasting

train_8664

In addition, we can see that CTR-smf (Purushotham et al., 2012) is sensitive to the quality of graph (SMF-1 with low quality and SMF-2 with high quality as shown in Figure 2).

we can use the low quality noisy graph (SMF-1) to improve the overall performance by this transformation process.

contrasting

train_8665

OmegaWiki is a freely editable online dictionary like Wiktionary.

instead of distinct language editions, OmegaWiki contains language-independent concepts ("Defined Meanings") which carry lexicalizations in different languages.

contrasting

train_8666

Wiktionary-Wikipedia (English): No evaluation dataset (let alone a full alignment) has been reported for this resource pair yet.

as the datasets for WordNet-Wiktionary (Meyer and and WordNet-Wikipedia (Niemann and Gurevych, 2011) are lexically overlapping, we were able to automatically create a gold standard for Wiktionary-Wikipedia by exploiting the transitivity of the alignment relation, i.e.

contrasting

train_8667

We were able to create a gold standard in a novel way by exploiting the fact that many German Wiktionary senses contain links to the corresponding Wikipedia articles, inducing a sense alignment between the two LSRs manually validated by the Wiktionary community.

we were unable to extract such an alignment for English, as Wikipedia articles are attached to the lexical entry page in this version and not to a specific sense.

contrasting

train_8668

(Bond and Foster, 2013)) and overlap of domain labels (Wikipedia, Wiktionary, Word-Net, OmegaWiki).

for none of these features we could observe any significant 1 impact on the results, mostly due to sparsity of the respective features.

contrasting

train_8669

For conversion from PS to DS, a head-table approach (Magerman, 1994;Collins, 2003;Yamada and Matsumoto, 2003;Sun and Jurafsky, 2004;Nivre, 2006;Johansson and Nugues, 2007;Duan et al., 2007;Zhang and Clark, 2008) is widely used.

the reliability of head tables has been questioned (Xue, 2007).

contrasting

train_8670

Secondly, we detect and annotate clause pairs in a document that hold logical discourse relations.

since this is too complicated to assign as one task using crowdsourcing, we divide the task into two steps: determining the existence of logical discourse relations and annotating the type of relation.

contrasting

train_8671

For the examples in Table 5, we confirmed that the discourse relation types of the top four examples were surely correct.

we judged the type (Contrast) of the bottom example as incorrect.

contrasting

train_8672

On the one hand, using words as the reordering units will reduce the number of candidates generated.

word segmentation results will affect the performance of WOE detection and correction.

contrasting

train_8673

If a segment without WOEs is misjudged to be erroneous, the word order still has a chance to be kept by the WOE correction models.

if a segment with WOEs is misjudged to be correct, words in the misjudged segment will not be reordered in the correction part because the error correction module is not triggered.

contrasting

train_8674

Clearly, the recall at the segment level and the correctable rate at the sentence level are 1 by the all-tag-B baseline.

its accuracy at the segment and the sentence levels are low.

contrasting

train_8675

It means the correct candidates are ranked 3.7 on the average.

the MRR by using the C sys dataset is 0.208.

contrasting

train_8676

We address the task of automatic terminology acquisition (ATA), the task of finding technical terms in texts without reliance on existing resources that list terms of the domain.

to this stands automatic terminology recognition (ATR), which we define as finding known terms and their variants (Jacquemin and Bourigault, 2003).

contrasting

train_8677

This is similar to distant supervision (Mintz et al., 2009) which also uses pre-existing resources such as gazetteers for, e.g., relation extraction.

our method is applied to ATA for the technological domain and does not rely on precompiled resources -we make use of figure references, which are an inherent part of patents.

contrasting

train_8678

Our method can be characterized as training data identification: we exploit given conditions in patents for our search of training data.

training data recognition methods need precompiled resources as input and search for instances of resource elements in texts.

contrasting

train_8679

Like our first baseline, it needs no training data.

to our first baseline, it was specifically designed for terminology acquisition.

contrasting

train_8680

We would also like to evaluate candidate classification on gold boundaries (manually verified boundaries of term candidates); this allows us to quantify by how much performance can be improved if candidate identification is perfect.

since gold boundary annotation is expensive, we instead approximated it: (i) We run automatic term candidate identification.

contrasting

train_8681

All observations hold for T l dev and T l test .

numbers are higher for T l dev because the ratio of FRTCs to candidates is higher than for T l test (38% vs. 27%) which improves classification performance on T l dev -this holds for ATAS as well as for the baselines.

contrasting

train_8682

In such cases the figure reference serves as a disambiguator.

in other positions they are non-terms, e.g., "They include braces, collars, splints and other similar apparatus".

contrasting

train_8683

We consider that selecting a value of n grater than 4 could lead to find few n-grams, so that many web pages could be under-represented.

previous experiments using also bigrams showed that they are not suitable for this approach.

contrasting

train_8684

According to the model, one of the most prominent distinctions is that amateurs tend to refer to the opponent as he, whereas experts use white and black more frequently.

it is of course not universally true, which leads to the misclassification of some experts as amateurs.

contrasting

train_8685

This seems counterintuitive at first, as we may expect lower-rated players to be less familiar with such terms.

it appears that they are frequently overused by weaker players.

contrasting

train_8686

(2003) presented the theories and approaches to calculate the frequencies of Tibetan characters, pieces, syllables and words based on a large scale Tibetan corpus including about 40, 000, 000 syllables (Lu et al., 2003).

a large part of the corpus they used are Buddhist literatures and the work can't be done well without a pragmatic Tibetan word segmentation tool (Chen et al., 2003a;Chen et al., 2003b;Jiang, 2006;Jiang and Kong, 2006;Sun et al., 2009;Sun et al., 2010;Lu and Shi, 2011;Liu et al., 2012a).

contrasting

train_8687

The sources and scales in different units are shown in It's a heavy task to manually classify those document into domains.

we still can get the domain information for a certain subsets of the corpus.

contrasting

train_8688

maturity, background of information etc."

we do not know of any such investigations for Bangla text readability that have investigate the way background of a reader affect the readability of text.

contrasting

train_8689

This group of models moves beyond the surface features of a text and try to measure objectively the different cognitive indicators associated with text and the reader.

it has been observed that, many situations, some traditional indicators perform as well as the newer and more difficult versions (Crossley et al., 2007).

contrasting

train_8690

For adult data, it can be seen that this feature has a strong and significant correlation, which not true for the user data of group 2 for separate texts.

for the common texts this feature was found to have high significant correlation with both the reader groups.

contrasting

train_8691

The correlations are also significant.

the correlations with $(noun phrase), $(verb phrase) $(postpositions), #(postpositions), #(adjective) were found to be insignificant.

contrasting

train_8692

Different with normal practice in WSI work, there is no feature engineering in our model.

our BNP model outperformed all the systems on supervised evaluation.

contrasting

train_8693

Those previous works on combining TM and SMT can be classified into four categories: (1) selecting the better translation sentence from TM and SMT (He et al., 2010a;2010b;Dara et al., 2013); (2) incorporating TM matched sub-segments into SMT in a pipelined manner (Koehn and Senellart, 2010;; (3) only enhancing the SMT phrase table with new TM phrase-pairs (Biç ici and Dymetman, 2008;Simard and Isabelle, 2009); and (4) incorporating the associated TM information with each source phrase to guide the SMT decoding .

all previous works mentioned above only focus on the case in which the TM database and the SMT training set share the same data-set.

contrasting

train_8694

Therefore, the factor ( | ) from the test set will possess a different probability distribution in comparison with that from the training set.

the development set is not big enough (only a few hundreds sentence-pairs at each interval) to re-train all TM factors of the proposed model.

contrasting

train_8695

For example, at interval [0.9, 1.0), those added TM phrase-pairs significantly improve the SMT system from 63.65 to 73.55, and Model-III from 80.69 to 86.40.

if Model-III + is compared with Model-III, the improvements from merging the TM phrase-pairs get less when the fuzzy match score decreases, because the matched TM parts are fewer at low fuzzy match intervals.

contrasting

train_8696

However, it is still worse than the TM in TER (13.32 vs. 10.42).

although Model-III has greatly exceeded the SMT at each interval, Model-III + still significantly outperforms Model-III at most intervals.

contrasting

train_8697

In real environments, the SMT training set and the TM database could be the same before translation projects starts.

the TM database will gradually deviate from the SMT training set while the translation task progresses.

contrasting

train_8698

The useful reordering pattern learned through this example is: Ich kann umstellen → I can rearrange which is memorized through the operation sequence: Generate(Ich, I) -Generate(kann, can) -Insert Gap -Generate(umstellen, rearrange) It can generalize to the test sentence shown in Figure 1(a).

it fails to generalize to the sentences in Figure 1(b) and (c) although the underlying reordering pattern is the same.

contrasting

train_8699

All three measures showed divergence in message similarity both between individuals, and in the community as a whole, across time.

the common theme with all these techniques is that although they can effectively measure adaptation of linguistic feature use within and between dialogs, they fail to capture the precise direction of convergence or divergence between individuals (i.e., do both interactants within a conversational pair accommodate their language use to the same extent?)

contrasting