id
stringlengths 7
12
| sentence1
stringlengths 6
1.27k
| sentence2
stringlengths 6
926
| label
stringclasses 4
values |
|---|---|---|---|
train_11200 | RNNs trained specifically to perform the agreement task can achieve very good average performance on a corpus, with accuracy close to 99%. | error rates increase substantially on complex sentences (Linzen et al., 2016, suggesting that the syntactic knowledge acquired by the RNN is imperfect. | contrasting |
train_11201 | The relevant studies such as clustering models and similarity computation in context (Fu et al., 2016;Wang, 2010) mainly focus on the metaphoric sense of each individual noun or adjectival phrase because the analyses are highly dependent on contextual information. | metaphoric senses of verbs are less touched because it is difficult to define regularities of their contextual information. | contrasting |
train_11202 | (2016) point out that the challenge arises from the highly context-dependent property of homonymies since the relations of senses are not unsystematic. | the senses of a polysemy form a systematic system, and thus CDSM has a better chance to detect metaphoric senses (Gutiérrez et al. | contrasting |
train_11203 | Figure 1 shows that verbs having the same radical are relatively similar to each other compared to verbs which belong to different radical groups. | the grouping by radicals does not work well in the metaphoric senses, as shown in the lower graph. | contrasting |
train_11204 | The incorporation of all the 17 features does improve the classification model by 1.70% in F-score. | group 1 has the best performance, outperforming the result when all the 17 features are used. | contrasting |
train_11205 | Since each of the syntactic conditions links to a particular aspect of a conceptual event, its change is an informative indicator of which sense, literal or metaphoric, is in use. | the conditions in Group 3 do not contribute much to detecting senses. | contrasting |
train_11206 | The formulation of Strehl and Ghosh (2003) is identical to ours. | there are several important differences. | contrasting |
train_11207 | This is an ambiguous result: on the one hand, a difference of 1.5% on the CoNLL dataset is non-negligible. | even the MUC-based method outperforms each individual component. | contrasting |
train_11208 | Here, the best MUC results (F = 73.35) are achieved with the top nine systems. | this MUC improvement comes at a high cost in B 3 and CEAF, leading to low MELA values even with the optimal parametrization. | contrasting |
train_11209 | coreference and entity relatedness) that are unavailable to local methods, and have significantly outperformed the local approach on standard datasets (Guo and Barbosa, 2014;Pershina et al., 2015;Globerson et al., 2016). | global approaches are difficult to apply in domains where only short and noisy text is available, as often occurs in social media, questions and answers, and other short web documents. | contrasting |
train_11210 | We find that both using pre-initialized embedding vocabularies, and the corrupt-sampling method employed are very important for properly training our model. | the gap between results of all systems tested on both CoNLL-YAGO and WikilinksNED indicates that mentions with noisy context are indeed a challenging test. | contrasting |
train_11211 | Deep learning has been successfully applied to many recent QA approaches and related tasks (Bordes et al., 2015;Hermann et al., 2015;He and Golub, 2016;Dong et al., 2015;Tan et al., 2016, inter alia). | large quantities of data are needed to train the millions of parameters often contained in these models. | contrasting |
train_11212 | Babelfy (Moro et al., 2014)) to a corpus and then using word2vec to learn sense embeddings over the pre-disambiguated text. | in their approach words are replaced by their intended senses, consequently producing as output sense representations only. | contrasting |
train_11213 | One important advantage of region representations is that they can distinguish words with a broad meaning from those with a more narrow meaning, and should thus in principle be better suited for tasks such as hypernym detection and taxonomy learning. | it is currently not well understood how such region based representations can best be learned. | contrasting |
train_11214 | (2011) using a sentiment lexicon to enhance embeddings for sentiment classification. | learning word embeddings with a particular target makes the approach less generic, also implying that customized adaptation has to be made whenever a new knowledge source is considered. | contrasting |
train_11215 | Typically to encode the entire vocabulary, the depth of the tree falls in a manageable range around 15 to 18. | different from learning context words, to encode a regularizer as shown on the right part of Figure 1, using hierarchical softmax is intractable due to exponential space demand. | contrasting |
train_11216 | The way we construct regularization matrix may be lossy, risking losing information that is explicitly delivered in the lexicon. | it provides us effective encodings for words, and also yields better learning performance empirically in our experiments. | contrasting |
train_11217 | Consequently, it can be observed from our experiment that unannotated knowledge, i.e., topic distributions, is not an effective source as a good guidance. | ppDB, which is of high quality of semantic knowledge, outperforms other types of information in most cases. | contrasting |
train_11218 | Empirical results show that with only character embedding features, the performance of our model outperforms CNN-CNN-MoT, and is close to LSTM-MoT. | there is still a big gap between character embedding and word embedding models, which could come from the fact that we use pretrained word embeddings, which helps improve the performance. | contrasting |
train_11219 | Taghipour and Ng (2016) tried to attend over words on their one-layer LSTM model, but failed to beat the baseline model that employs mean-over-time pooling, because of that text-level model contains a quite long sequence of words, which may weaken the effect of attention. | sentence-level model contains relatively short sequences of words, which makes attention more effective. | contrasting |
train_11220 | If you previously recognized the sentiment to be neutral or even positive, it is possible that the addition of this new word would cause the sentiment to change to negative. | it is hard to imagine a set of words to which adding the word drive-thru would change the sentiment in any direction. | contrasting |
train_11221 | While motivated differently, contrastive estimation (Smith and Eisner, 2005) is also related to matching. | ive estimation, negative training examples are synthesized by perturbing positive instances. | contrasting |
train_11222 | This phenomena has already been studied in script learning works (Chatman, 1980;Chambers and Jurafsky, 2008b;Ferraro and Van Durme, 2016;Pichotta and Mooney, 2016a;Peng and Roth, 2016). | modeling actions is not sufficient; participants in actions and their emotions are also important. | contrasting |
train_11223 | Recent research in the area has focused on neural approaches to summarisation, which can be very data-hungry. | few large datasets exist and none for the traditionally popular domain of scientific publications, which opens up challenging research avenues centered on encoding large, complex documents. | contrasting |
train_11224 | The introduction tends to be longer (average length of 72.1 sentences) than other sections, but still of a relatively simple level compared to the method (average length of 41.6 sentences), thus has more potential sentences for an author to use in highlights, giving the high Copy/Paste score. | it would also have more sentences which are not good summaries and thus reduce the overall average ROUGE score of the introduction. | contrasting |
train_11225 | We design this model to automatically generate mention embeddings and mention-pair embeddings that are used to construct cluster features for entity linking. | even though this model's success in coreference resolution is not our final objective, its success directly correlates to the success of entity linking because of the similarity between these two tasks. | contrasting |
train_11226 | Question q 1 is relevant with respect to it as it asks the same thing, and so is q 2 , which asks how much one should tip in a specific situation. | q 9 and q 10 are irrelevant: the former asks about what to wear at business meetings, and the latter asks about how to tip a kind of person who does not normally receive tips. | contrasting |
train_11227 | We train the network by minimizing the negative log-probability of the gold labels: The network described so far learns the abstract features through multiple hidden layers that are discriminative for the classification task, i.e., similar vs. non-similar. | our goal is also to make these features invariant across languages. | contrasting |
train_11228 | First, since here we are training with only 50% of the original training data, both FNN and CLANN-unsup yield lower results compared to before, i.e., compared to Table 1; this is to be expected. | the unsupervised adaptation, i.e., using the CLANN-unsup model, still yields improvements over the FNN model by a sizable margin, according to all three evaluation measures. | contrasting |
train_11229 | Our method achieves stateof-the-art F1 on the Jobs dataset. | even without such domain-specific supervision, the parser performs reasonably well. | contrasting |
train_11230 | Our grammar formalism can be related to synchronous CFGs (SCFGs) (Aho and Ullman, 1972) where the semantics and syntax are generated simultaneously. | instead of modeling the joint probability of the logical form and natural language utterance p(x, y), we model the factorized probability p(x)p(y|x). | contrasting |
train_11231 | Considering embeddings only, Levy and Goldberg (2014) proposed to learn word representations that incorporate syntax from dependency-based contexts. | we inject syntactic information by means of TKs, which establish a hard match between tree fragments, while the soft match is enabled by the similarities of distributed representations. | contrasting |
train_11232 | An alternative approach is to train the context embedding using neural networks on a sense annotated corpus, which can remap the word embeddings in a supervised fashion. | since there are not enough large disambiguated corpora, we need to approximate the word senses with coarse-grained information, e.g., the category of the context. | contrasting |
train_11233 | This shows that a more sophisticated interaction layer can help. | the differences are not substantial, indicating that this extension does not offer any systematic advantage. | contrasting |
train_11234 | It shows that more than half or more than a third of all questions in SQuAD or NewsQA, respectively, are (partially) answerable by a very simple neural BoW baseline. | the gap to state-of-the-art systems is quite large (≈ 20%F1) which indicates that employing 77.9 69.5 FastQA w/ beam-size k = 5 77.1 68.4 FastQAExt k = 5 78.9 70.8 Tables 2 and 3 clearly demonstrate the strength of the FastQA system. | contrasting |
train_11235 | This is in contrast to earlier studies (Sahu and Anand, 2017;Collobert and Weston, 2008) where pretrained embeddings Model i2b2-2010 DDI extraction Precision Recall F1 score Precision Recall F1 score SVM (Rink et al., 2011) 67 usually improved model performances by 3-4%. | this result aligns with the observations made in (Johnson and Zhang, 2015) and supports the argument for one-hot LSTMs. | contrasting |
train_11236 | The existing tool for automatic PID computation, CPIDR (Brown et al., 2008), is based on counting POS tags. | we noticed that the propositional structure of a sentence is very similar to its dependency structure, see the first column in Table 1. | contrasting |
train_11237 | When the repeated ideas are not counted (DEPID-R), the difference between groups becomes non-significant. | we were curious about why the association between the lower PID values and the AD diagnosis cannot be observed on DementiaBank. | contrasting |
train_11238 | 4 On AMI data, the difference between group means is non-significant for both CPIDR and DE-PID values. | when the repeated ideas are excluded (DEPID-R), the mean PID for AD patients is significantly lower than for controls, as expected. | contrasting |
train_11239 | On AMI data, the SID performs surprisingly well, considering that the automatic ICUs were extracted from only 10 clusters and the number of clusters was not tuned to that dataset at all. | pID performs ca 5% better than SID in terms of all measures. | contrasting |
train_11240 | On DementiaBank dataset, cluster features alone do not perform too well and using cluster features together with PID and SID gives only minor improvements. | both the Table 8: Classification results on DementiaBank (DB) and AMI using cluster features (C) combined with PID and SID. | contrasting |
train_11241 | On the open-domain dataset we found that the PID was more predictive than SID as expected. | the small number of automatically extracted cluster features underlying the SID, modeling the broad discussion topics, led to even better results. | contrasting |
train_11242 | In these efforts, the costs scale linearly in the number of instances, requiring significant investments for large datasets. | schema querification can generate an enormous amount of data for a fraction of the cost by labeling at the relation level; as evidence, we were able to generate a dataset 300 times larger than Simple QA. | contrasting |
train_11243 | We showed that relation extraction can be reduced to a reading comprehension problem, allowing us to generalize to unseen relations that are defined on-the-fly in natural language. | the problem of zero-shot relation extraction is far from solved, and poses an interesting challenge to both the information extraction and machine reading communities. | contrasting |
train_11244 | Now, as we saw, M2 contrasts with M1. | it can be understood in comparison with it: We can understand abstract change in terms of physical or local change. | contrasting |
train_11245 | Combining source and target data together into a single dataset is a simple way to jointly train for both domains. | this approach might not work well in the crosslingual case, i.e. | contrasting |
train_11246 | As expected, higher values of β yield lower precisions but higher recalls. | f 1 increases until Table 2: Number of: "false anaphor" (fA, a non-anaphoric mention marked as anaphoric), "false new" (fN, an anaphoric mention marked as non-anaphoric), and "wrong link" (WL, an anaphoric mention is linked to a wrong antecedent) errors on the development set. | contrasting |
train_11247 | (2014a) in the sense that our resolvers incrementally add mentions to previously built clusters. | different from both Ma et al. | contrasting |
train_11248 | A disadvantage of using critics is that there is no guarantee that the critic converges to the true evaluation metric given finite training data. | our differentiable relaxations do not need to train, and the convergence is guaranteed as T → 0. | contrasting |
train_11249 | Yu and Jiang (2016) presented a convolutional NN that learns sentence embeddings using two auxiliary tasks (whether the sentence contains a positive or a negative domain independent sentiment word), purposely avoiding prediction with respect to a large set of pivot features. | to these works our model can learn useful cross-domain representations for any type of input example and in our cross-domain sentiment classification experiments it learns document level embeddings. | contrasting |
train_11250 | Our model is similar in structure to an autoencoder. | instead of reconstructing the input x from the hidden layer h(x), its reconstruction function r receives a low dimensional representation of the non-pivot features of the input (h(x np ), where x np is the non-pivot representation of x (Section 3)) and predicts whether each of the pivot features appears in this example or not. | contrasting |
train_11251 | The semantic role labeler achieves competitive performance on English, even without any kind of syntactic information and only using local inference. | when automatically predicted partof-speech tags are provided as input, it substantially outperforms all previous local models and approaches the best reported results on the English CoNLL-2009 dataset. | contrasting |
train_11252 | Semantic banks (e.g., PropBank (Palmer et al., 2005)) often represent arguments as syntactic constituents or, more generally, text spans (Baker et al., 1998). | coNLL-2008 and shared tasks (Surdeanu et al., 2008;Hajic et al., 2009) popularized dependency-based semantic role labeling where the goal is to identify syntactic heads of arguments rather than entire constituents. | contrasting |
train_11253 | Moreover, as syntactic parsers are not reliable when used out-of-domain, standard (i.e., syntactically-informed) dependency SRL models are crippled when applied to such data. | our syntax-agnostic model appears to be considerably more robust: we achieve the best result so far on the English and Czech out-of-domain test set (77.7% and 87.2% F 1 , respectively). | contrasting |
train_11254 | Though LSTMs are likely to be affected by a similar trend, their states may be able to capture the uncertainty about the structure and thus let the role classifier account for this uncertainty without the need to explicitly sum over potential syntactic analysis. | pathLSTM will have access only to the single (top scoring) parse tree and, thus, may be more brittle. | contrasting |
train_11255 | For example, if a word is a noun, it should take one of three tags from the case category: nominative (n), accusative (a), or genitive (g), while it should take "not applicable" (na) from the mood category since mood is not defined for nominals. | most of the previous approaches in Arabic did not utilize this information, applying one model for each task (Habash and Rambow, 2005;Pasha et al., 2014;Shahrour et al., 2015). | contrasting |
train_11256 | tional resources such as a morphological analyzer or a dependency parser, indicating the effectiveness of joint modeling of morphosyntactic categories. | the independent model gives an accuracy of 87.74%, which is 1.53% absolute worse than CamelParser. | contrasting |
train_11257 | (2016) proposed an approach in which they encode a sequence of possible morphosyntactic tags provided by a morphological analyzer using bi-directional LSTMs. | we provide an alternative way of encoding this information, as well as an analysis on the most influential categories in the encoded tag embeddings. | contrasting |
train_11258 | The rational for the separation is that different dialects have different affixes, make different lexical choices, and are influenced by different foreign languages. | performing reliable dialect identification to properly route text to the appropriate segmenter may be problematic, because conventional dialectal identification may lead to results that are lower than 90% . | contrasting |
train_11259 | Recently, it has become common to locally parametrize these models using rich features extracted by recurrent neural networks (such as LSTM), while enforcing consistent outputs through a simple linear-chain model, representing Markovian dependencies between successive labels. | the simple graphical model structure belies the often complex non-local constraints between output labels. | contrasting |
train_11260 | Outside of NLP, in tasks such as computer vision, certain classes of fully-connected graphical models over the output pixels have been used for multi-dimensional smoothing (Adams et al., 2010;Krähenbühl and Koltun, 2011), borrowing techniques for the graphics literature. | none of these models performs representation learning in the output space, as in the case of our proposed embedded latent-state model. | contrasting |
train_11261 | (Koo et al., 2010;Anzaroot et al., 2014;Vilnis et al., 2015), many based on dual decomposition (Rush and Collins, 2012). | the constraints are often fixed, and even when learned (Anzaroot et al., 2014;Vilnis et al., 2015), the learning is done simply on constraint weights generated from pre-made templates, the construction of which requires domain knowledge. | contrasting |
train_11262 | Table 1 shows that overall performance on the UMass Citation dataset using the embedded-state latent CRF (95.18) is marginally better than the baseline BiLSTM+CRF model (95.07). | examining the entities with the largest F1 score improvement in Table 2, we see that they are mostly within the VENUE section, which has longrange constraints with other sections, giving evidence of the model's ability to learn constraints from the citation dataset. | contrasting |
train_11263 | Recently, the RNN-based generators have shown improving results in tackling the NLG problems in task oriented-dialogue systems with varied proposed methods, such as HLSTM (Wen et al., 2015a), SCLSTM (Wen et al., 2015b), or espe-cially RNN Encoder-Decoder models integrating with attention mechanism, such as Enc-Dec (Wen et al., 2016b), and RALSTM . | such models have proved to work well only when providing a sufficient in-domain data since a modest dataset may harm the models' performance. | contrasting |
train_11264 | 9 is differentiable, we can jointly optimize the parameter θ and variational parameter φ using standard gradient ascent techniques. | the KL divergence loss tends to be significantly small during training (Bowman et al., 2015). | contrasting |
train_11265 | Take, for example, the scr100 scenario in which the CrossVAE model mostly outperformed all the previous strong baselines with regard to the BLEU and the slot error rate ERR scores. | the previous methods showed extremely impaired performances regarding low BLEU score and high slot error rate ERR when training the models from scratch with only 10% of in-domain data (scr10). | contrasting |
train_11266 | The representation-focused approach (Huang et al., 2013) independently learns a representation for each ranking element (e.g., query and document) and then employ a similarity function. | the interactionfocused models are designed based on the early interactions between the ranking pairs as the input of network. | contrasting |
train_11267 | Our work studies on the strength of the dynamic relatedness between entities, hence we focus more on Pearson index. | traditional correlation metrics do not consider the positions in the ranked list (correlations at the top or bottom are treated equally). | contrasting |
train_11268 | Locations of social media users are important to many applications such as rapid disaster response, targeted advertisement, and news recommendation. | many users do not share their exact geographical coordinates due to reasons such as privacy concerns. | contrasting |
train_11269 | User network model can correctly geolocate only 18.4% of users. | our experiments show that excluding user network information declines the performance of the final model in terms of accuracy by 5.1%. | contrasting |
train_11270 | As reported in Table 2, our proposed approach achieves quite low median errors over the TWIT-TERUS and WNUT datasets (i.e., 40.1km and 0km, respectively). | there are some cases with large error distances, which make the mean errors much larger than median errors. | contrasting |
train_11271 | Our ablation study demonstrates that the location field highly contributes to the geolocation per- Table 3: Performance breakdown for each component over WNUT dataset formance. | some prediction errors arise when location fields are incorrect. | contrasting |
train_11272 | To the best of our knowledge, there exists no prior work explicitly aiming at discovering thematic hierarchies in corpora. | the hierarchy-related effects are reported in some studies. | contrasting |
train_11273 | The pairwise ranking approach takes context into account. | some role pairs only co-occur rarely. | contrasting |
train_11274 | PropBank semantic role annotation and the corresponding SemLink reference are constituents-based. | uD is a dependency formalism, and we employ a number of heuristics to align original PropBank annotations with the CoNLL-2009 datasets (Hajič et al., 2009) to recover the head node positions. | contrasting |
train_11275 | Let s 1 , s 2 ∈ S be two sentences: this rule is violated if, according to the model, a sentence s 1 contradicts s 2 , but s 2 does not contradict s 1 . | if we just use the final decision made by the neural NLI model, we can simply check whether the rule is violated by two given sentences, without any information on the degree of such a violation. | contrasting |
train_11276 | We can see that, in the case of rule R 1 (reflexivity of entailment), DAM and ESIM make a relatively low number of violations -namely 0.09 and 1.00 %, respectively. | in the case of cBiLSTM, we can see that, each sentence s ∈ S in the SNLI training set, with a 23.76 % chance, s does not entail itself -which violates our background knowledge. | contrasting |
train_11277 | Similar to these studies, our approach is also based on a Bi-LSTM CRF architecture. | considering the limited contexts within each individual sequence, we design two attention mechanisms to further incorporate topically related contextual information on both the document-level and corpus-level. | contrasting |
train_11278 | 8 The impact on performance is fairly small indicating the robustness of both models. | after only two epochs of optimizing the selected words in the added sentence, the performance drops markedly under all variants of the sentence-level black-box attacks as displayed in the two bottom rows of and incorrect answers (AddQA) is most detrimental and causes both models to perform at or even below chance level. | contrasting |
train_11279 | They reported to achieve 63% and 45% accuracies for coarse and fine-grained answer-type detection, respectively under 5-fold cross validation setup. | we manually create 5, 535 codemixed questions and train a CNN model that shows 87.21% and 83.56% accuracies for coarse and fine answer types, respectively, for the 5-fold cross validation. | contrasting |
train_11280 | (2017)'s system was mainly developed to answer pure factoid questions based only on the named entities denoting person, location and organization. | the datasets used in this experiment have different types of answers beyond the basic factoid questions. | contrasting |
train_11281 | Machine translation (MT) techniques greatly boost the productivity of the translation agencies (Arenas, 2008). | despite the recent advances achieved in this field, MT systems are still far to be perfect and make errors. | contrasting |
train_11282 | The SMT system was unable to handle well such unseen sentences. | iNMT systems usually respond much better to the human feedback than interactive SMT systems (Knowles and Koehn, 2016;Peris et al., 2017). | contrasting |
train_11283 | I would like to switch :)", the intention is not churny for the brand this tweet is addressed to. | in "@MARKE Internet langsamer als gedrosseltes. | contrasting |
train_11284 | They trained their models on a Twitter sentiment analysis corpus (Pak and Paroubek, 2010) which is composed of 15M data points with labeled sentiment. | to the best of our knowledge, there is no work that uses churn detection in the context of chatbot conversations. | contrasting |
train_11285 | 2017), they weighted the representation of each word according to the position, and the words close to the aspect could be paid more attention. | this operation is not always reasonable and sometimes the adjunct word may be far away from the target word. | contrasting |
train_11286 | The most common approach to training NMT is to maximize the conditional log-probability of the correct translation given the source sentence. | as argued in Bengio et al. | contrasting |
train_11287 | Named-entity Recognition (NER) is an important task in the NLP field , and is widely used to solve many challenges. | in many scenarios, not all of the entities are explicitly mentioned in the text. | contrasting |
train_11288 | To an average human reader who is familiar with contemporary norms and trends, it is quite clear that Instagram app is discussed in the textual passage above. | it is not explicitly written, thus it is practically a latent entity. | contrasting |
train_11289 | Traditional approaches label each token in texts as part of named-entity, and achieve high performance (Ratinov and Roth, 2009;Passos et al., 2014;Chiu and Nichols, 2016). | these approaches are relying on the assumption that entities are necessarily mentioned in the text. | contrasting |
train_11290 | Therefore, a possible approach could possibly suggest that all of the entities should be predicted together as a single multi-task classification process. | this method is based on the assumption that all entities are necessarily related to one another (as presented in section 5.1). | contrasting |
train_11291 | Thus, there is a small difference between multitask and nonmultitask approaches in Table 2. | in Table 4 we examine the performance over the top-40 frequent entities, including very frequent entities (such ATP and ADP), and less frequent (such Oxygen, NADPH, NADP+ and water) as well. | contrasting |
train_11292 | The expected performance for Web applications is thus closer to the one observed on the Inspec and DUC2001 datasets, rather than on NUS. | on long documents, Multipartite outperforms all other methods. | contrasting |
train_11293 | As shown in Figure 1b, EmbedRank++ reduces the redundancy we faced with EmbedRank. | embedRank++ surprisingly results in a decrease of the F-score, as shown in Table 2. | contrasting |
train_11294 | The results obtained by GREEDY can be arbitrarily bad. | there are performance guarantees if the objective function f and the constraints I are "sufficiently nice" (Calinescu et al., 2011). | contrasting |
train_11295 | These models typically perform either "early" (input data are concatenated and pushed through a common model) or "late" (outputs of the last layer are combined together through linear or non-linear weighting) fusion. | our model does not fall into any of these categories directly as it is "iterative" in the sense that there are multiple fusions per decision, with an evolving belief state -the memory. | contrasting |
train_11296 | The very recent work (Zadeh et al., 2018a) also approaches multi-view learning with recourse to a system of recurrent encoders and attention mediated by global memory fusion. | fusion takes place at the encoder cell level, requires hard alignment, and is performed online in one sweep so it cannot be informed by upstream context. | contrasting |
train_11297 | Information about a specific semantic property can be stored latently over the dimensions of a semantic embedding model, such that the semantic prop-erty can be reliably decoded given an embedding vector, as tested in the previous section. | a stronger test of how closely an embedding model relates to human-elicited conceptual knowledge is to investigate whether the embedding dimensions encode interpretable, human-like semantic properties directly. | contrasting |
train_11298 | For instance, the Indo-European dataset has cognate cluster sizes ranging from 37 for meaning because to 1 for meaning name. | a non-parametric clustering method such as Chinese Restaurant Process (CRP; Gershman and Blei 2012) can form clusters directly from the data without the need for tuning the threshold. | contrasting |
train_11299 | The UPGMA algorithm is simple and yields reasonable results across various language families (List, 2012a). | uPGMA clustering algorithm is dependent on the threshold that needs to be tuned to obtain optimal performance (List et al., 2017b). | contrasting |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.