id
stringlengths 7
12
| sentence1
stringlengths 6
1.27k
| sentence2
stringlengths 6
926
| label
stringclasses 4
values |
|---|---|---|---|
train_18400 | (Collobert et al., 2011) proposed a multi-task learning system using deep learning methods for various natural language processing tasks. | the system with window approach cannot be jointly trained with that using sentence window approach. | contrasting |
train_18401 | With the power of deep learning, there are emerging better approaches of LU (Hakkani-Tür et al., 2016;Chen et al., 2016b,a;. | most of above work focused on single-turn interactions, where each utterance is treated independently. | contrasting |
train_18402 | Table 1 shows the similar trend as LU results, where applying either rolebased contextual models or intermediate guidance brings advantages for both semantics-encoded and NL-encoded history. | to NL, semantic labels (intent-attribute pairs) can be seen as more explicit and concise information for modeling the history, which indeed gains more in our experiments for both LU and dialogue policy learning. | contrasting |
train_18403 | Neural conversation systems, typically using sequence-to-sequence (seq2seq) models, are showing promising progress recently. | traditional seq2seq suffer from a severe weakness: during beam search decoding, they tend to rank universal replies at the top of the candidate list, resulting in the lack of diversity among candidate replies. | contrasting |
train_18404 | This is why summary writing is sometimes used together with essays to assess university-level abilities in L2. | assessing L2 summaries is highly demanding, especially if analytic rubrics are involved, as they require raters' expertise and much concentration when assessing language proficiency at various levels (e.g., lexis, syntax, discourse). | contrasting |
train_18405 | Analytic scoring based on different rubrics (e.g., accuracy, cohesion) is therefore particularly appropriate when assessing summaries in L2 as it offers more informative feedback (Bernhardt, 2010) and better captures different facets of L2 writing competence than holistic assessment (Weigle, 2002). | analytic scoring is often exceptionally demanding for raters, especially in the case of longer texts and more than four or five scoring categories (CEFR), which motivates the use of automated assessment. | contrasting |
train_18406 | Our method outperforms the previous studies (TextRank (Mihalcea and Tarau, 2005), L2R (Henß et al, 2015), Regression (Henß et al, 2015), RL-full (Henß et al, 2015) ) on ACL-ARC and WIKI data but achieves lower performance than the previous studies on DUCs. | the differences of ROUGE scores between the models are relatively small. | contrasting |
train_18407 | The PEAK (Yang et al., 2016) method for evaluation automates the extraction part of the SCUs, and they use the ADW (Align, Disambiguate and Walk) algorithm (Pilehvar et al., 2013) to compute semantic similarity. | their approach fails to model contradiction, paraphrase identification and other features like natural language inference. | contrasting |
train_18408 | Since SSAS performs similarity assessment using deep semantic analysis, it does take significantly large amount of execution time compared to other methods. | ssAs computations for multiple summaries can be easily parallelized. | contrasting |
train_18409 | They have also been developed for the update summarization task, with work such as (Li et al., 2015) about the weighting of concepts. | the most obvious weakness of such methods, particularly the one proposed by Gillick and Favre (2009), is their implicit way of modeling information redundancy. | contrasting |
train_18410 | The improvement over ICSISumm 2009, which has the same settings as our system, confirms the interest of handling redundancy explicitly in the update summarization task. | the improvement over ICSISumm-BG-DOWN-1&2 also shows that basic methods for performing this handling are not efficient in ILP models, contrary to our sentence semantic clustering approach. | contrasting |
train_18411 | (2016) used the word embeddings in order to perform query expansion. | previous word embedding-based probabilistic IR methods have no theoretical validity since the word embeddings were heuristically introduced. | contrasting |
train_18412 | Clearly, the local window of contexts is filled with irrelevant words, meanwhile, it fails to capture important temporal status indicators. | as shown in figure 1, the event "launch" is actually a high order indirect governor word of the target event word "protest" and is only two words away from the target event in the dependency tree. | contrasting |
train_18413 | Other approaches include shortest path kernels , graph kernel (Airola et al., 2008), composite kernel (Miwa et al., 2009), subsequence kernels (Kim et al., 2010), and tree kernels (Eom et al., 2006;Qian and Zhou, 2012). | engineering features from different sources may not lead to optimal results. | contrasting |
train_18414 | When using randomly initialized embeddings, RNN exhibits similar performance as other NN models. | by taking advantage of pre-trained embeddings, RNN further advances Fscores by 7% and 13% on AIMed and BioInfer, respectively. | contrasting |
train_18415 | used the result of analyzing social support in OHCs for identifying influential users. | these works address emotional support in general and do not focus on identifying empathetic messages. | contrasting |
train_18416 | Note that firstly, LSTM without profile does not perform better than CNN-Wang. | other studies show that when attention model is incorporated, LSTM generally outperforms that of CNN model (Chen et al., 2016;Yang et al., 2016) which will be shown later. | contrasting |
train_18417 | We can also simply apply an encoder-decoder model to text normalization tasks. | it is well-known that encoderdecoder models often fail to perform better than conventional methods when the availability of training data is insufficient. | contrasting |
train_18418 | Other text normalization task In this study, we evaluated our methods with Japanese dialect data. | these methods are not limited to Japanese dialects because they do not use dialogspecific information. | contrasting |
train_18419 | General purpose language identification tools such as TextCat (Cavnar and Trenkle, 1994) and langid.py (Lui and Baldwin, 2012) can identify 50-100 languages with accuracy of 86-99%. | these tools have not considered discrimination between closely related language varieties. | contrasting |
train_18420 | In all four languages, the top linked language is same as a target language. | their percentages vary from 78.08-94.26%, indicating differences in the amount of texts in different languages. | contrasting |
train_18421 | With the rapid development of research on Neural Machine Translation (NMT), translation quality has been improved significantly compared with traditional statistical based method (Bahdanau et al., 2014;Cho et al., 2015;Zhou et al., 2016;Sennrich et al., 2015). | training efficiency is one of the main challenges for both academia and industry. | contrasting |
train_18422 | As it can be seen, boost outperforms the default method by +1.49 (BLEU) at the cost of using 10% of additional training data. | the system converges faster than any other system as best performance is achieved at epoch 14, while others need to achieve the best after epoch 16. | contrasting |
train_18423 | We reported that the majority vote did not consistently improve the BLEU score. | this study develops a voice classifier using the following seven features. | contrasting |
train_18424 | SrcVoice represents the voice of the source side. | unlike English, it is difficult to formulate simple rules to obtain the voice of Japanese 3 Experiments Here we explain the voice controlling method proposed by Yamagishi et al. | contrasting |
train_18425 | Subword units like character (Vilar et al., 2007;Tiedemann, 2009), orthographic syllables (Kunchukuttan and Bhattacharyya, 2016b) and byte pair encoded units (Kunchukuttan and Bhattacharyya, 2017) have been used with varying degrees of success. | if no parallel corpus is available between two languages, pivot-based SMT (Gispert and Marino, 2006;Utiyama and Isahara, 2007) provides a systematic way of using an intermediate language, called the pivot language, to build the source-target translation system. | contrasting |
train_18426 | Therefore we only extract top-level functions since they are usually small and relatively self-contained, thus we conjecture that they constitute meaningful units as individual training examples. | in order to support research on project-level code documentation and code generation, we annotate each sample with metadata (repository owner, repository name, file name and line number), enabling users to reconstruct dependency graphs and exploit contextual information. | contrasting |
train_18427 | In contrast, methods using radial-basis function (RBF) kernels do not provide weight vectors, via which we cannot obtain each feature's importance. | we used SVR-RBF, SVR with a radial-basis function (RBF) kernel, GPR-RBF, GPR with an RBF kernel, and JGPR-RBF, GPR with an RBF kernel and joint prediction ( §3) since these models can take into account combinations of features using the RBF kernels, which are useful for combining both domain and semantic features. | contrasting |
train_18428 | As shown in figure 2, in the case of CCA and Word2vec, bio-markers are distributed evenly around COPD. | in the case of GloVe, COPD is found to be far apart. | contrasting |
train_18429 | However, in the case of GloVe, COPD is found to be far apart. | in all three models, bio-markers located close to COPD are CC-16 and CEA. | contrasting |
train_18430 | This shows that the CCA does not play a significant role in this problem. | word2vec is a very stable methodology because all the BEST marker values show higher values than the wORST cases, unless the dimension is reduced by t-SNE. | contrasting |
train_18431 | On the other hand, Word2vec is a very stable methodology because all the BEST marker values show higher values than the WORST cases, unless the dimension is reduced by t-SNE. | gloVe showed a stable appearance as the number of dimensions increased, while it didn't in case of 2and 5-dimension. | contrasting |
train_18432 | Table 2, we observe that the SPLITEM-BEDDING method, which is the embodiment of Scheme 1 (given in Section 2.1) where σ(w i , l j ), the semantic similarity of word w i with the label l j , is computed using embedding, yields the highest precision across all the methods. | sPLITMULTIEMBEDDING, a variant of scheme 2 (given in section 2.2) where the embeddings of section header and body are independently computed, and a weighted combination of the embeddings is used to retrieve the representative words of the section to compare with the embeddings of the labels, delivers the highest recall and Fscore values, as well as, the highest overall accuracy. | contrasting |
train_18433 | This also applies for CVs that have separate section headers for identification, such as Name, Email etc. | for sections that are intuitively more complex, show some (meaningful) confusions across classes. | contrasting |
train_18434 | (2016) investigated zero anaphoric resolution on news text in Japanese and they considered only the zero arguments of a predicate. | in math problems, the arguments of unsaturated nouns are often omitted. | contrasting |
train_18435 | From this result, we can say that using only Per-ceivedInfo as system utterances is not an effective method. | since there may be a difference among the types of PerceivedInfo, we further investigated the evaluation scores in each type of PerceivedInfo. | contrasting |
train_18436 | If there is fully labelled data available, e.g., labelled documents, our model will account for the full supervision from document labels during the generative process, where each document can associate with a single class label or multiple class labels. | if the dataset contains both labelled and unlabelled data, our model will account for the available labels during the generative process as well as incorporate the labelled features as above to constrain the Dirichlet prior. | contrasting |
train_18437 | For instance, the two topics under Performance+ suggests that some people feel the system performs better and app runs faster, whereas Performance-seems to show highly contrastive opinion that people have bad experience after upgrade, e.g., app crashes or freezes, mac becomes slow. | it is still impossible to accurately interpret the extracted topics solely based on its multinomial distribution, especially when one is unfamiliar with the topic domain. | contrasting |
train_18438 | (Mencap, 2002;PlainLanguage, 2011;Freyhoff et al., 1998). | manual production of texts from scratch for each target population separately cannot keep up with the amount of information which should be accessible for everyone. | contrasting |
train_18439 | Around 90% of CPs have been selected by at least two annotators (see Table 1). | when we separate the selections made by native and nonnative speakers, we see that: (1) the percentage of multiple selected CPs by native speakers and non-native speakers decreases; (2) the percentage of multiply selected CPs by non-native speakers is always lower (83%-85%) than the percentage of multiply selected CPs by native speakers (84%-86%), regardless of the text genre; and (3) the percentage of CPs selected by at least one native and one non-native annotator is lower for the NEWS genre (70%) than for the WIKINEWS and WIKE-PEDIA genres (77%). | contrasting |
train_18440 | Inspired by the success of the data-driven approach, one can prepare a corpus of conversations for every possible style and feed it to a seq2seq. | in order to obtain a stylistically consistent DRG system through a vanilla seq2seq, this method requires millions of training instances for each target style, which is prohibitively expensive. | contrasting |
train_18441 | In the past two decades, the majority of NLP research focused on developing tools for the Standard English on newswire data. | the nonstandard part of the language is not well-studied in the community, even though it becomes more and more important in the era of social media. | contrasting |
train_18442 | Recently, summarization has also been considered as a submodular function maximization (Lin andBilmes, 2010, 2011;Dasgupta et al., 2013) where greedy algorithms were adopted to achieve near optimal summaries. | the main drawback of all the extractive approaches is that they can not avoid the inclusion of insignificant information which degrades the summary quality. | contrasting |
train_18443 | However, the main drawback of all the extractive approaches is that they can not avoid the inclusion of insignificant information which degrades the summary quality. | the abstractive approach in a multi-document setting aims at generating summaries by deeply understanding the contents of the document set and rewriting the most relevant information in natural language. | contrasting |
train_18444 | relevant (P ∪ U) and irrelevant (N ), as was done in SemEval. | by ignoring the difference in utility between paraphrases and useful questions during training, a binary classification approach is likely to underperform a ranking approach that is trained on all the ranking triples implied by the original 3 categories of questions. | contrasting |
train_18445 | To also use the subject, one could simply concatenate the subject and the body (Body + Subj) and apply the same SLA architecture from Figure 1. | as shown in Table 6, this actually hurt performance, likely because the system did not know where the question body started in each input sequence. | contrasting |
train_18446 | (1) Initialize matching with singletons relations (u, v). | relates a child v of e to d by considering the set U that aligns to v. For each u ∈ U , we increment the score if its parent is d and give additional points if the parent's POS tag and edge label is the same as those of e. these are normalized by |U |. | contrasting |
train_18447 | While monolingual insights like paraphrases have potential applications in semantic textual similarity (Agirre et al., 2012), there exist bigger corpora for those tasks, such as PPDB (Ganitkevitch et al., 2013). | as the Bible is often the only significant parallel text for many of the world's languages, improved 27-way consensus English POS tags Before corpus alignments: TIME: NN (1.00) SECRET: NN (0.54), JJ (0.46) With corpus alignments: TIME: NN (0.94), NNS (0.05) . | contrasting |
train_18448 | 1 BabelNet has been applied in many natural language processing tasks, such as multilingual lexicon extraction, crosslingual word-sense disambiguation, annotation, and information extraction, all with good performance (Elbedweihy et al., 2013;Jadidinejad, 2013;Navigli et al., 2013;Ehrmann et al., 2014;Moro et al., 2014). | to the best of our knowledge, to date there is no comprehensive work applying BabelNet knowledge to SMT tasks. | contrasting |
train_18449 | The EN-PL BabelNet dictionary contains 6,199,888 bilingual entries. | the raw data contains a lot of noise, so we performed some pre-processing of the dictionary, including: • if East Asian characters are included in either the English or Polish side, we remove this pair; • if the English side contains symbols which are neither letters nor digits, then we remove this pair; • if either side contains punctuation, then we remove this pair; • if the English side is the same as the Polish side, then we remove this pair; • if the ratio of the word-level entry lengths between the English side and the Polish side is less than 0.5 or greater than 2, then we remove it. | contrasting |
train_18450 | We can see that the Ba-belNet API method did not beat the best domain-adaptation method on all tasks in terms of BLEU and TER. | it does improve system performance compared to the baselines, which shows that using BabelNet can alleviate the issue of unknown words to some extent. | contrasting |
train_18451 | They are clearly useful for indicating certain overarching trends, but say little about actual improvements for translation buyers or post-editors. | these metrics are commonly referenced when discussing pricing and models, both with translation buyers and service providers. | contrasting |
train_18452 | Using several of them in parallel helps to validate the individual scores and thus increase confidence in the results. | while they are very useful for indicating quality trends (for example system B with a BLEU of 79 can quite safely be expected to be better than old system A with a BLEU of 60), their interpretation remains cryptic to many users: "What does it mean if my English to Polish MT system gets a BLEU of 50, is this good or bad? | contrasting |
train_18453 | It provides a segment-level score. | the Edit Distance score by itself does not take into account the actual time spent on a given translation unit and provides an absolute score, whereby the maximum number of edits possible per segment would be the total number of characters of the reference translation. | contrasting |
train_18454 | This percentage is also frequently quoted in the context of the post-editor recruitment process or as explanation to translation buyers, as MT output, reference translation and expected effort are shown side-by-side. | it still does not consider the actual time spent making the edits. | contrasting |
train_18455 | Sentiment analysis systems, including the best performing ones such as the NRC-Canada system Kiritchenko et al., 2014a;Zhu et al., 2014), use sentiment lexicons to obtain significant improvements (Wilson et al., 2013;Pontiki et al., 2014;Rosenthal et al., 2015;Mohammad et al., 2016a). | much of the past work has focused on English texts and English sentiment lexicons. | contrasting |
train_18456 | As shown above, lexicons created by translating existing ones in other languages can be beneficial for automatic sentiment analysis. | the above experiments do not explicitly quantify the extent to which such translated entries are appropriate, and how translation alters the sentiment of the source word. | contrasting |
train_18457 | A similar happening takes place in community 19 between the Primary Sentiment level and topic 3 level although in this case it was found 4 users with exactly the same sentiment in topic 1. | there are communities that are only present in some topics, such as communities 3, 7, 8, 11, 12 or 19. | contrasting |
train_18458 | It has been used in many different research works over the past years. | we have also used the Bing Liu's sentiment lexicon (Hu and Liu, 2004). | contrasting |
train_18459 | General domain datasets, like the whole Wikipedia data or News dataset from online newspapers, capture very well general syntactic and semantic regularities. | to capture in-domain word polarities smaller domain focused dataset might work better (García-Pablos et al., 2015). | contrasting |
train_18460 | The restaurant-adjectives-test-set contains 119 positive adjectives and 81 negatives adjectives, while laptops-adjectives-test-set contains 127 positives and 73 negatives 11 . | we have used the SemEval 2015 task 12 datasets 12 . | contrasting |
train_18461 | Furthermore, as has been mentioned in many studies (Brody, 1985), (Hall et al., 2000) the gender difference in emotional expression has been detected during several psychological investigations. | to the very specific nature of speaker-adaptive ER, gender-adaptive ER might be more general. | contrasting |
train_18462 | The results may be even better if the SI component performs more accurately (compare E and G systems in Figure 1). | the performances of the Sep system dropped on Emo-DB, LEGO, and Ruslana. | contrasting |
train_18463 | Detailed analysis reveals that it is extremely difficult (for both models) to distinguish Slovak adverbs from nouns. | prepositions are moderately difficult with the c7 model (P=48%, R=75%) but they are practically solved with the Czech model (P=R=99%). | contrasting |
train_18464 | The presence of anger was always signaled by the shouting or use of offensive language. | there are other ways of user's to express their (dis)satisfaction towards the system. | contrasting |
train_18465 | the length of the utterances in the MapTask corpus). | when the classifier is trained on the combination of the MapTask corpus with other corpora, the performance of the classifier improves. | contrasting |
train_18466 | Moreover, the clear separation of dialogue dimensions and communicative functions, coupled with the hierarchical organization of the latter, allows for classification at different levels of granularity. | re-annotating existing corpora with the new scheme might require significant effort. | contrasting |
train_18467 | Yet, also for nouns in general, adjectives and verbs the findings indicate that OCE may have a potential to improve the learning of their semantic similarity-or perhaps relatedness-properties. | we also found significant disadvantages of the OCE-model used here, especially on test sets of similarity judgments. | contrasting |
train_18468 | Corpora annotated with coreference are arguably a valuable resource for such studies. | they are mostly monolingual. | contrasting |
train_18469 | The corpus, sized over 600 sentences, was manually annotated with full-fledged coreference chains in both languages. | it is not publicly available. | contrasting |
train_18470 | In PCEDT, we did our best to keep this convention. | in some cases coreferential relations between nominal groups without definite determiners are so obvious that they were annotated even in Ontonotes, in spite of negative instructions (as in Example (7)). | contrasting |
train_18471 | Thus, in some cases it is difficult to decide whether a given nominal group should be annotated for coreference. | generic nouns may be used anaphorically and with a determiner (as in Example (8)) inciting annotators to mark a coreferential relation, regardless the annotation instructions. | contrasting |
train_18472 | As the occurrence of entities with no counterparts may be, among other reasons, attributed to grammatical differences, such cases can hardly disappear. | around 6% of the entities with the 1:N mapping must be a result of an error, either in alignment or coreference. | contrasting |
train_18473 | Competence errors are attributable to a lack of knowledge in the language, while performance errors are due to external factors, such as lack of attention, stress or fatigue (Corder, 1967). | researchers have highlighted the fact that even though this distinction is theoretically relevant, it is practically impossible to distinguish competence errors from performance errors (Thouësny, 2011). | contrasting |
train_18474 | In the case of missing articles, we notice that results in the two comparison corpora are similar, and are much lower than in the requirement corpus. | determination errors in general account for 24.1 % (student essays) and 15.8 % (scientific papers) of all errors in the two comparison corpora, indicating that authors produce a more varied range of determination errors in these types of writing than in requirements, where determination errors other than missing articles are marginal to non-existent. | contrasting |
train_18475 | English and Norwegian for the past 2-3 decades, resources for this kind of studies have been largely lacking for L2 Swedish. | researchers of L2 vocabulary and grammar acquisition are in great demand of digitized L2 corpora of Swedish, that can help verify hypotheses generated by experimental studies and/or smaller scaled empirical studies. | contrasting |
train_18476 | • not to correct author's mistakes. | in dubious cases, we applied a principle of positive assumption, i.e. | contrasting |
train_18477 | Kelly is a frequency-based wordlist generated from web corpora, and translated into and compared between nine languages for identification of core vocabulary across these languages (Kilgarriff et al., 2014). | kelly has shortcomings, namely that (1) frequency statistics are collected from web texts aimed at L1 speakers of Swedish, which can be misleading since the vocabulary used for L1 speakers may differ from what beginner L2 speakers need to concentrate on; (2) the division into the CEFR levels is based on frequency and L1 text coverage, which needs explicit validation to confirm its relevance for a CEFR-based curriculum; and (3) kelly lacks some vocabulary useful in the L2 context, such as table, alphabet, toothpaste -i.e. | contrasting |
train_18478 | In the context of second language learning it means that a learner who has acquired the knowledge of these words can read and understand most of the modern Swedish texts. | as was the case for the coverage approach of Laufer and Ravenhorst-Kalovski (2010), this list only describes a global learning goal and does not contain any indications for appropriateness at different levels of learner proficiency. | contrasting |
train_18479 | The methodology applied has the great advantage to be able to assign different difficulty levels to the different senses of a word. | it is also a long and costly process to repeat for other languages. | contrasting |
train_18480 | As a result, the lack of available corpora calls for the use of word frequency lists in the lexical substitution task (Devlin and Tait, 1998;Shardlow, 2014). | as the commonly used frequency lists are mainly L1-focused, we believe that they do not fit the L2 context very well. | contrasting |
train_18481 | On the one hand, we automatically annotated each lexical unit in the global corpus with its level of difficulty based on FLELex (see Section 2.2.1.). | we asked a group of Dutch-speaking learners of French to manually annotate the 51-text sample according to their vocabulary knowledge (see Section 2.2.2.). | contrasting |
train_18482 | In order to detect a significant lexical simplification effect, we compared each pair of original and simplified texts with respect to the number of lexical units that were attributed to each of the six difficulty levels in FLELex (Table 3). | as the texts included in both the Tales and the Wiki corpus differed greatly in terms of text length, we normalised the counts to a length of 1,000. | contrasting |
train_18483 | The core vocabulary per level includes those words that are observed in most textbooks. | the peripheral vocabulary per level includes those words that tend to appear in only few textbooks and which are hence not indicative of the vocabulary commonly targeted at a given proficiency level. | contrasting |
train_18484 | In contrast, the peripheral vocabulary per level includes those words that tend to appear in only few textbooks and which are hence not indicative of the vocabulary commonly targeted at a given proficiency level. | we could collect more extensive learner annotations for all six proficiency levels in order to automatically learn a new transformation rule that defines vocabulary knowledge similarly to the learners' annotations. | contrasting |
train_18485 | On the one hand, we observed that defining the word's CEFR level on the basis of its first occurrence in FLELex enabled us to detect a significant simplification effect in manually simplified texts. | we observed that an expert model predicting the learner's actual vocabulary knowledge based on the word's CEFR level was relatively accurate, but that it presented some errors with respect to the recall of unknown words. | contrasting |
train_18486 | Since the baseline score on the F metric is harder to beat, its value is correspondingly less informative. | the D metric, which measures discriminative ability rather than error rate, works equally well for both groups of systems. | contrasting |
train_18487 | This is a counterintuitive result, since it would suggest that the system is less useful to students producing higher proportions of correct responses; in fact, the results presented in chapter 7 of (Baur, 2015) suggest the opposite pattern. | the D metric returns the same value irrespective of the balance between correct and incorrect answers, as long as the relative reject rate on each group stays the same. | contrasting |
train_18488 | If the RALL systems cover a wide range of topics, learners can train their conversational skill in L2 on the topics of their interest (Wilcock & Yamamoto, 2015). | aSR of L2 speech is still a challenge because the speech is often made with poor pronunciation and contains grammatical errors. | contrasting |
train_18489 | Spearman's scores showed very low correlation between OSMAN scores of the two Arabic versions (0.035) which indicates the importance of the presence of diacritics, which plays a vital role in determining the text ease of reading. | the OS-MAN scores (0.329) for the diacriticised Arabic showed a higher positive correlation with the English Flesch scores. | contrasting |
train_18490 | Spearman's scores showed very low correlation between OSMAN scores of the two Arabic versions which indicates the importance of the presence of diacritics, which plays a vital role in determining the text ease of reading. | the OSMAN scores for the diacriticised Arabic showed a higher positive correlation with the English Flesch scores. | contrasting |
train_18491 | Enabling users of intelligent systems to enhance the system performance by providing feedback on their errors is an important need. | the ability of systems to learn from user feedback is difficult to evaluate in an objective and comparative way. | contrasting |
train_18492 | Still another one is interactive information retrieval, where the user provides feedback on the relevance of retrieved documents to improve the overall search results, be it for textual (Salton and Buckley, 1990) or multimedia documents (Nguyen and Worring, 2008). | in general, evaluating the ability of systems to learn from user feedback in an objective and comparative way is difficult. | contrasting |
train_18493 | As mentioned above, the goal of system adaptation is to maximize the system performance while minimizing the amount of feedback information needed to get the improved performance. | since the feedback provided, and in particular the amount of information therein, is under the control of the user, if the task under study does not naturally lend itself to have this amount of information also controlled by the evaluator, no meaningful comparison between systems can be drawn. | contrasting |
train_18494 | Furthermore, if the oracle is designed using knowledge of the internal functioning of the system with which it interacts, the evaluation can be biased. | these issues Error rate (E) Figure 2: Computation of a Minimum Supervision Rate (MSR): Starting from the initial system error rate (E T0 ), some amount of information is provided as feedback (F T1 ), leading the system to update its models and yield a new error rate (E T1 ). | contrasting |
train_18495 | In such situations, which are common in information retrieval and entity recognition, metrics like precision and recall are typically used to describe performance. | precision and recall fail to describe the differences between sets of objects selected by different decision strategies, instead just describing the proportional amount of correct and incorrect objects selected. | contrasting |
train_18496 | If one feature group's removal keeps overall performance similar, but selects completely different entities, this indicates a potentially interesting analysis candidate. | this is not possible using the conventional measures. | contrasting |
train_18497 | They are based on precision and recall, thus using conventional solutions to evaluation problems that come from the typical skew toward negatives. | they also take into account complementarity, giving more insight into differences than F-score, precision or recall. | contrasting |
train_18498 | Also, from the high complementary recall that ANNIE offers over OpenNLP, adding ANNIE's results is likely to offer strong recall improvements. | from P Comp (OpenN LP, AN N IE) being only 20.00 we can see that ANNIE and OpenNLP select quite different entities, and that precision gains are not likely to be made. | contrasting |
train_18499 | That topic modelling does well at identifying differences in topic is perhaps not so surprising, and this could be seen as consistent with the good performance of topic modelling on the nyt june KSC, where news articles from different time periods would be expected to be on somewhat different topics. | (this does not provide an account for why perplexity does so well for nyt 5678 where we would also expect differences in topic.) | contrasting |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.