id
stringlengths 7
12
| sentence1
stringlengths 6
1.27k
| sentence2
stringlengths 6
926
| label
stringclasses 4
values |
|---|---|---|---|
train_95200 | This paper deals with building linguistic resources for Gulf Arabic, one of the Arabic variations, for sentiment analysis task using machine learning. | we started with building linguistic resources for this dialect, i.e. | neutral |
train_95201 | The class indicates the part of speech (POS) tag and the polarity of that word. | here, P and N denote the sets of positive and negative sentences of the corpus, respectively. | neutral |
train_95202 | Then, our goal is to design a conversational strategy that makes it possible to maintain some control over the selected response of the system. | 92% of the utterances of the system are unique in a dialogue on average. | neutral |
train_95203 | We use the Weka Machine Learning Library's naive bayes classifier in our experiments (Hall et al., 2009). | that corpus contained a much lower percentage of guessable clues than did the human generated clues in the corpus collected by (Paetzel et al., 2014). | neutral |
train_95204 | (2014) leveraged the existence of bilingual subtitles as a source of parallel data for the Chinese-English language pair to improve the MT systems in the movie domain. | to calculate the term weights according to the appearance of a term within the document collection, we apply term frequency-inverse document frequency (tF-IDF) (Ramos, 2003) as one term-weighting model. | neutral |
train_95205 | A number of successful approaches, therefore, combine different types information. | using contextual information with the aforementioned enhancements helps to distinguish Machine Translations (MTs) that are different from the human translation and still essentially correct from those that share a high number of words with the reference but alter the meaning of the sentence due to translation errors. | neutral |
train_95206 | In addition to the three MT results, we prepared two kinds of human translations. | if it is used as an automatic interpreter on smartphones, it is a totally different type of matter. | neutral |
train_95207 | Additionally, this study improves on previous works by using actor model, which enables us to take advantage of all available cores by highly parallelizing the corpus building process. | it is therefore worth investigating how early deduplication, e.g. | neutral |
train_95208 | This can either be the consequence of a linguistic ambiguity or of a tagger error. | • The rarest word in the sentence should have a rank of about 32000. | neutral |
train_95209 | We declare an element Y of an accepted frame X optional if X \ Y is also an accepted frame. | perhaps the most surprising result evident from this table is that the short vector and the long vector conditions are quite competitive: this lends support to a recent finding of (Levy et al. | neutral |
train_95210 | We address the task of automatically correcting preposition errors in learners' Dutch by modelling preposition usage in native language. | as we can see from the tables, results show that the system is very accurate for some prepositions (van, in, te, op), while not so much for others (als, bij, door). | neutral |
train_95211 | The best performance on DUC 2004 is reached by the composite kernel CK according to both ROUGE-1 and ROUGE-2 scores. | this is hardly true for many cases. | neutral |
train_95212 | In other 13% of the cases, both the verb and the particle are translated independently of each other, yielding an incorrect literal translation such as tourner sur for 'turn on'. | for English, we created a small corpus of verb-particle split constructions from scratch. | neutral |
train_95213 | In the automatic alignments of parallel corpora, most of the particles align to NULL. | approximately 4.5 million sentences were used for training, combining Europarlv7, CommonCrawl and News data. | neutral |
train_95214 | NULL alignments such as those illustrated in (7) are potentially detrimental for MT quality since they add weight to incorrect translations. | table 5: Summary of English-to-French word-level alignment of verb-particle split constructions. | neutral |
train_95215 | In the example cases, if the clue (here: the quantification) is present, the respective noun groups (e.g. | for "bad requirement" the true positive value is 94, there are 2 false positives and 0 false negatives. | neutral |
train_95216 | In this paper we present recent works contributing to transformation of the initial PolNet, a Polish wordnet developed at the Adam Mickiewicz University, into a Lexicon Grammar of Polish. | in all such cases we keep the corresponding synsets related by the transformational relations which describe the differences among their morpho-syntactic properties. | neutral |
train_95217 | In (applied) linguistics and (second) language acquisition, we find more awareness of coding schemes and reproducibility of studies. | there are no agreed-upon coding schemes. | neutral |
train_95218 | They display results of the searches as concordances. | as for the 'focus' words (the new information in the sentence), the frequency distribution of their functor attribute was a bit different. | neutral |
train_95219 | This work has been partially funded by the SpeDial project supported by the EU FP7 with grant num. | two evaluation metrics were used: two-class (binary) classification accuracy and Pearson correlation coefficient with respect to human ratings for the Greek and the English language. | neutral |
train_95220 | As for Hungarian databases, to the best of our knowledge, only one Hungarian sentiment corpus has been made so far, the OpinHuBank corpus (Miháltz, 2013). | moreover, it is also important to point out that negations, irreals and decreasers (Intensifierminus) are more frequent in the negative fragments than in the positive ones; e.g. | neutral |
train_95221 | We develop a systematic account of the inventory of possible functor types and their interrelationships in Section 5. | for instance, the functor embedding a state of possession in Table 1 is to be read as follows. | neutral |
train_95222 | We construct the utterances for the survey items by varying the stimuli along the dimensions shown in Table 11. | the remainder of this paper is structured as follows. | neutral |
train_95223 | 3 shows an example of an extraction rule derived from the sentence Die EU profitiert davon, dass sie andere Länder zwingt, auf Wachstum zu verzichten. | neither could we expect to have a complete set of extraction rules, given that we started with 120 sentences forming their basis. | neutral |
train_95224 | For instance in The minister does not force the president to cheat nothing happens at all, neither minister nor president receive an effect. | only if a tree was correct wrt. | neutral |
train_95225 | Future work includes the expansion of the lexicon, and experimenting with other types of datasets. | the combined lexicon offers a slight improvement in the Egyptian dialect when carrying out 10 FCV, but decreases the accuracy in the Saudi dialect dataset for the same experiment. | neutral |
train_95226 | As expected, the introduction of the neutral class causes the accuracy drops dramatically for both datasets. | manual additions and re-validations have been conducted over the past two years. | neutral |
train_95227 | In the case of hybrid dependency patterns+CRF method, CRF part was evaluated in a 10-fold cross-validation scenario on the review dataset. | • For the method based on shallow parser, a JSON or SOAP format defined according to multiservice specification (Ogrodniczuk and Lenart, 2012). | neutral |
train_95228 | This treebank is a result of parsing 20000 Polish sentences with the syntactic parserŚwigra. | any method of sentiment identification can be used with the tool. | neutral |
train_95229 | These results underline the difficulty of this first step, extracting the correct aspect terms is a challenging task and it would definitely benefit from additional optimization experiments. | data statistics for both corpora are presented in Table 2 and in the next sections we will have a closer look at the actual annotations. | neutral |
train_95230 | In building the SRL model, the following features are adopted: Features related to the current node: the syntactic node label, the extracted functional word, the extracted syntactic labels which are extracted from the subtree node during flattening, the linguistic features (such as voice, tense, aspect, etc. | finally, section 5 concludes the paper. | neutral |
train_95231 | Currently, FrameNet and PropBank are the two most commonly used semantic resources in Semantic Role Labeling (SRL) and SMT tasks. | as an example, the bilingual tree pairs in Figure 3 and Figure 4 have the same semantic meaning. | neutral |
train_95232 | Besides, it is also found that the work of those early stages (e.g., syntactic tree and word alignment annotation) notably influences our work. | the steps (4) to (6) will be repeated iteratively until all sentence-pairs are annotated with their case labels. | neutral |
train_95233 | With this iterative process, we incrementally improve our SRL model and save annotator's effort via providing more accurate pre-labeled case trees. | inspired by the work of Su et al. | neutral |
train_95234 | In analytic languages, such as English, one word usually equals one morpheme. | morpheme alignment annotation is manually performed by LDC annotators. | neutral |
train_95235 | We first align articles on the same topic in Wikipedia via the interlanguage links ((1) in Figure 1). | the second line of table 4 shows the percentage of infinite scores of "+NN-ASPEC" for the character and word based NN features, respectively. | neutral |
train_95236 | Intuitively, because the NN features give a measure of the bilingual similarity of a sentence pair, they could be helpful for this task. | as comparable corpora are far more available, many studies have been conducted to extract parallel sentences from them for MT. | neutral |
train_95237 | It is most surprising in the case of Galician, because being co-official with Spanish it was expected that there would be at least regional government related multilingual accounts. | the boxplot shows a similar error distribution between ca-es and eu-es language pairs, and in fact, the average error rate per account was 21% for eu-es and 18% for ca-es. | neutral |
train_95238 | The error rate in C dev is 3% (ca-es) and 33% (eu-es). | it does result in high quality parallel corpora when such Twitter accounts are available, which can be exploited to adapt an MT system to the new domain presented by Twitter. | neutral |
train_95239 | Researchers make available new findings and knowledge in biology and medicine through the publication of research articles. | regarding percentages of titles and documents in the training data, table 1 shows that more abstracts are available in comparison to titles. | neutral |
train_95240 | ACL WMT'16) so the research community can experiment with additional translation methods. | only titles are available in more than one language in PubMed (Wu et al., 2011). | neutral |
train_95241 | In addition to this, it is worth noting that many websites are available in several languages and this translated content is another useful source of parallel data. | to deal with this limitation, we propose a new method that focuses on crawling top-level domains (tLD) for multilingual data, and then detects parallel data inside the crawled data. | neutral |
train_95242 | Our analysis indicated that while there were 3 comments made about accuracy out of 744 people with correct predictions (0.40% commented), and 13 noting incorrect answers out of 201 with incorrect predictions (6.5% commented). | it is unknown whether the up-and down-voting was motivated by the accuracy of the quiz, but researchers making interactive prediction sites may discover that inaccuracy is in fact more engaging in some regards. | neutral |
train_95243 | The word embeddings are again the best of the individual feature sets in terms of recall, F1 score and accuracy. | wikipedia A dump of English wikipedia from 1 September 2015. | neutral |
train_95244 | A popular approach used by recommendation systems is collaborative filtering for identifying users similar to a target user based on their purchase history, and then recommending products that similar users have already bought but the target user has not (Schafer et al., 1999;Sarwar et al., 2000). | figure 1 shows an example of a portion of the tweet timeline of a Twitter user who wants an iPhone. | neutral |
train_95245 | To determine whether monthly aggregation is changing the rate at which prevalence estimates converage, for each sampling level we compute how close an estimates' correlation with GFT is to the correlation's maximum value. | they also find that twitter's 1% sample is similar to the Firehose, however, indicating that most of the bias they observed is introduced by the Streaming API. | neutral |
train_95246 | for 10-CV evaluation), training/testing splits were defined with respect to the overall and per class proportions of tweets indicated in Table 4 (%All column). | we defined four communication categories to classify tweets in our work: 1. | neutral |
train_95247 | In the former, words occurrences are assigned sense labels from a predefined sense inventory. | we then asked the annotators to cluster these words into sense clusters, giving them as much freedom as possible, and allowing for one word to appear in multiple clusters (soft clustering). | neutral |
train_95248 | Both resources are large, manually constructed networks intended to describe the language in general rather than any specific domain. | each entry except a special root is connected to other entries, its semantic descriptors. | neutral |
train_95249 | From a machine translation (MT) perspective, the deeper the processing of utterances, the less language-specific differences will remain between the representations of the meaning of the source and target texts. | afterwards, improvements over the baseline system have been applied, adapting and replacing some of the original sieves (Soraluze et al., 2015), taking into account that morphosyntactic features are crucial in the design of the sieves for agglutinative languages like Basque. | neutral |
train_95250 | In case of ambiguity, the algorithm picks up the most "popular" article. | this paper presents the first release of such corpora, which includes NERC, NED, WSD and coreferencelevel annotation for these six languages. | neutral |
train_95251 | The changes we propose for the original VerbNet role inventory in Figure 1 are kept as small as possible. | we employ webAnno (Yimam et al., 2014). | neutral |
train_95252 | Hindi vocabulary has a number of foreign language words, mostly taken from English and these find a place in Hindi WordNet as well. | f-score of WfS baseline on Health, Tourism and News domains was found to be 60%, 65% and 55% respectively. | neutral |
train_95253 | It was seen that most of the time the literal senses of a word were placed in ranks above the metaphorical or figurative uses. | hindi WordNet 1 (hWN) is developed for capturing the fine grained senses of hindi language. | neutral |
train_95254 | These results are as a rule not sorted by their relation to different ambiguous query senses and returned in the order of their relevance ranking, which for many of them seems to be almost equal. | in Section 2. we introduce our approach and put it into the context of the previous work in the field. | neutral |
train_95255 | A small smart car driving in the city. | this type of multi-modal data offers an interesting basis for vision and language research but most existing datasets use crowdsourced text, which removes the images from their original context. | neutral |
train_95256 | A more comprehensive analysis of the cases where the system and the human disagreed on classification is reported in the conference presentation. | in this paper we present a multiple instance learning-based segmentation system that accurately labels 91.27% of the video frames of 500 continuous utterances (including 7 different subjects) from the publicly accessible NCSLGR corpus <http://secrets.rutgers.edu/dai/queryPages/> (Neidle and Vogler, 2012). | neutral |
train_95257 | (It involves a kind of acting out of rolling dice; the linguistic properties are not typical of lexical signs.) | the total feature vector length is 14; and each frame would have approximately from 1 to 6 feature vectors corresponding to the high local motion regions. | neutral |
train_95258 | In step 2, we extract features related to the shape and velocity of the moving regions (e.g., hands). | there is no fixed fingerspelled vocabulary: many fingerspelled productions would not be included in any ASL dictionary. | neutral |
train_95259 | The best performing algorithm performed global optimization over the input sentence set. | a supervised approach with lightweight and efficient features improves over the lead-based baseline. | neutral |
train_95260 | The second kind, neoclassical compounding, combines some elements of Greek or Latin etymological origin, such as hydro + logy = hydrology. | some prefixes are very close to the neoclassical roots, compare prefix biwith neoclassical root uniaccording to (Béchade, 1992). | neutral |
train_95261 | The UniMorph Schema includes the features necessary to distinguish all these categories, which are marked by surface contrasts in each language and are not decomposed further in any natural language. | we present methods inspired by linguistic fieldwork for gathering inflectional paradigm data in a machine-readable, interoperable format from remotely-located speakers of any language. | neutral |
train_95262 | In other word, user can mentally create his own virtual app on top of existing ones. | there was no effort to conceal the wizard arrangement from the participant. | neutral |
train_95263 | The wizard was instructed to respond directly to participant's goal-directed requests and to not accept out-of-domain inputs. | sometimes users produce utterances which may involve several apps, e.g., "Boost my phone so I can play [game] spiderman" requires CLEANMASTER to clear the RAM and the game SPIDERMAN. | neutral |
train_95264 | 8 The inter-annotator agreement before reconciliation was 0.89 and after reconciliation became 0.95. | 3 http://ii.tudelft.nl/genius/ 4 https://github.com/vaskonov/tutorial We used a frame-based semantic representation, where one or more composite semantic labels are used to represent each utterance in a dialogue. | neutral |
train_95265 | Hu, Walker, Neff, & Fox Tree, 2015;Tolins, Liu, Wang, Fox Tree, Neff, & Walker, 2013, Hu, Dick, Chang, Bowden, Neff, Fox Tree, & Walker, 2016. | participants in lab-based behavioral experiments often experience a task that is already somewhat routine. | neutral |
train_95266 | entities expressing frequency or quantity). | term variants in the UMLS do not always instantiate this pattern: e.g., concept C0003611 is designated by the term appendicectomie, but not opération de l'appendice. | neutral |
train_95267 | 1 Moreover, both of these corpora have additional shortcomings. | we would like to acknowledge Kellen Maicher who created the virtual environment and Bruce wilcox who authored ChatScript and customized the software for this project. | neutral |
train_95268 | On the one hand, given the objective of the study, a fine grained study of the form -function relationship requires a fine grained functional analysis, too. | the campaign was realised on a duration of 2 months for most raters. | neutral |
train_95269 | Feedback utterances are among the most frequent in dialogue. | with the suggested annotation schema, a good number of annotators and high quality recordings, it is possible to achieve good inter-annotator agreement on large collections of instances. | neutral |
train_95270 | In this paper we discuss the design of the application and the influence of metadata and various forms of feedback. | but this was the closest we could get to a simple score. | neutral |
train_95271 | Judgements of pronunciation quality were obtained from experts (Burgos et al., 2013;Burgos et al., 2014). | almost three times as many women participated and they transcribed 1.5 times more items on average than men, which is a significant difference (t' (130.328) = 2.203, p =.029). | neutral |
train_95272 | National tests have been collected since 1996. | the main purpose of the corpus is to facilitate the systematic and quantitative empirical study of the writings of various student groups based on gender, geographic area, age, grade awarded or a combination of these, synchronically or diachronically. | neutral |
train_95273 | Students are aged between 18 and 40 years (80% aged 18-30) and represent 14 different mother tongues (we provide in brackets the number of texts per first language): Chinese (323), English (142), Spanish (139), German (76), Russian (70), French (43), Japanese (50), Italian (34), Dutch (15), Tetum (22), Arabic (13), Polish (22), Korean (9) and Romanian (8). | we established a set of 7 fields that were required for the productions of the informant to be included in COPLE2: name, age, nationality, gender, mother tongue, knowledge of other second or foreign languages, period of time studying Portuguese. | neutral |
train_95274 | We chose the realization of the velar nasal consonant /N/. | the women were aged between 22 and 47 years, the men between 30 and 45. | neutral |
train_95275 | GG speakers are used to read in German some of them privileged a fast reading in order to finish quickly others chose to read aloud carefully in order to interpret the stories. | this result could be explained by the written input. | neutral |
train_95276 | Precision is calculated on the level of a set of candidates where each set containing the right word is considered correct and the set not containing the correct word as incorrect. | text samples produced in a digital form can be identified in the corpus as they are marked as such and can be analyzed separately. | neutral |
train_95277 | In case of both presented technologies, the optimal number of candidates turns out to be 3 to 4. | every text sample was annotated by only one annotator and no inter-annotator agreement was measured. | neutral |
train_95278 | These ways and techniques are usually mastered by linguists through their experiences in a field and from a limited amount of knowledge gathered from books or lectures. | in our three projects we have confirmed that using a common application among linguists is an unrealistic solution and using a shared scheme such a standard proposed as multi-link-path model, is also unrealistic or at least unsuitable for linguists. | neutral |
train_95279 | A training programme must take this variability into account, in order to deliver appropriate measures for the different conditions and needs of languages with respect to their digital usability. | finally, DLDP will deliver a number of recommendations specifically addressed at language stakeholders and policy makers, the Roadmap for Digital Language Diversity. | neutral |
train_95280 | The digital readiness of a language is inextricably linked to its digital presence: whenever a language is technologically supported and thus widely digitally usable, its digital representation expands. | provision of state-of-the-art language-based applications, which would enable and foster their use over digital media and devices, is severely limited (Mariani, 2015). | neutral |
train_95281 | The darkness of the color in the map reflects the relative frequency of the selected variant in each location. | for example, distribution maps showing the distribution in the Tuscan dialect area of a specific dialectal form (selected via the ALT-Web website) are easily obtainable. | neutral |
train_95282 | We also use crowdsourcing to obtain a portion of the translation for some languages, although not all languages have sufficient numbers of crowd workers with the required skills to perform translation. | the first set of LORELEI language packs is expected to appear in LDC's catalog in late 2016. | neutral |
train_95283 | According to the Ethnologue (Lewis et al., 2015), the three dialects to the South of the Danube were developed between the 5th and the 10th century, while according to Rosetti (1966), this process took place after the 10th century. | we report the number of correctly classified and misclassified instances. | neutral |
train_95284 | The higher the possibility of the translation pair candidate being selected correctly is (determined by the structure of the transgraph), the lower is the cost to be paid of adding any new edge to it. | an arrow line denotes derivation of word from proto-word and a dashed edge denotes high possibility of shared meanings between two words, and thus denotes the possibility of missed meanings in the transgraph. | neutral |
train_95285 | Consequently, when w A 2 also shares the same meaning with pivot w B 1 , there is a high possibility that proto-words w P 1 and w P 2 also share the same meaning. | the rest of this paper is organized as follows: In Section 2, we will briefly discuss closely related languages and methods in comparative linguistics. | neutral |
train_95286 | The new translation pair candidate induced from two new edges like candidate 8 (w A 3 -w B 2 -w C 1 ) in Figure 4 is too strong to be considered. | there is a high possibility that w C 2 also share the same meaning with pivot w B 1 . | neutral |
train_95287 | The geolocations were further manipulated: Multi-part expressions were split as described for the concept list in 3.1 and additionally extended by plural and genitive endings. | for smaller places, we manually checked whether the place reference could contribute to re-identification of any third person mentioned. | neutral |
train_95288 | However, some experiments showed that completely blackening that timespan invalidates the whole sentence for further linguistic analysis as suprasegmental signals are disturbed. | it is obvious that any method working on the translation of the language data instead of on the original data will be skewed. | neutral |
train_95289 | Statistics for the crowdsourced dataset are presented in Table 1. | in this section, we present the methodology of crowdsourcing the entity salience corpus as well as its specific features. | neutral |
train_95290 | In this paper we present a gold standard dataset for Entity Linking (EL) in the Music Domain. | the ELVIS code is available at https://github.com/sergiooramas/elvis. | neutral |
train_95291 | tagging 'Yellow Submarine' as dbpedia.org/page/Yellow Submarine (song). | this information is extracted from the whole LASt.FM corpus for those entities falling in one of the four musical categories previously defined. | neutral |
train_95292 | Our method relies on the argumentum ad populum intuition, i.e. | we decided to select for the final ELMD dataset only those URIs stemming from a typeequivalent setting where agreement score is equal or greater to 1. | neutral |
train_95293 | Instead of expressing [ɛ] with ARABIC LETTER AE, they create the appropriate on-screen letter form by typing ARABIC LET-TER HEH and then, when necessary, preventing it from joining with the following letter using the special invisible character ZERO WIDTH NON-JOINER U+200C. | an entirely digital complication in processing Sorani text regards the Unicode expression of the very frequent short vowel /ɛ/ (also transcribed as /ae/), which is expressed by a special variant of the arabic letter heh . | neutral |
train_95294 | If only lung (ANATOMY) or cancer (PROBLEM) were recognized, the evaluation would yield only 0.5 true positives. | the type of relation that can hold between the top-level entity and the nested entity is defined broadly and includes specialization and attribute, among other types of semantic modification. | neutral |
train_95295 | Where the German DBpedia links were not available, the owl:sameAs links were taken from the English DBpedia entities, as such links were available for most of the entities from the dataset. | since NEL or KBP evaluation tasks might require a new corpus or at least a new gold standard, and the creation of such resources requires significant effort, there is a desire to automate steps in the corpus creation process. | neutral |
train_95296 | During the initial stages of the project, the main focus was the development of the annotated data sets for each of the languages. | most location and organisation named entities in the conjunctive languages require a nominal prefix that is followed by a capitalised character, e.g. | neutral |
train_95297 | In the light of the above, it appears that NE resources are quite well developed, benefiting from two decades of work. | knowledge base population, anonymisation and information extraction at large) (Fort et al., 2009). | neutral |
train_95298 | This strategy involves the misclassification of some mentions of the same entity in a single document. | due to the document length (≈ 380k tokens), some mentions belonging to person entities, with many occurrences in different contexts, were frequently labeled as locations and organizations (and even as miscellaneous entities, not shown in the table). | neutral |
train_95299 | A mention is selected for analysis if (i) it is not the first mention of the text, and (ii) it is the first mention of the entity it belongs to. | the heuristics also select the best named entity class for each entity, thus correcting previous NER errors. | neutral |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.