Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
6
9
sentence1
stringlengths
31
593
sentence2
stringlengths
29
609
label
stringclasses
4 values
test_1953
We also did not consider the gender ratio of our participants.
currently the total number male participants in our corpus is 70, and the total number of female participants is 151 (table 1).
reasoning
test_469
Not only is beam search usually more accurate than greedy search, but it also outputs a diverse set of decodings, enabling reranking approaches to further improve accuracy (Yee et al., 2019; Ng et al., 2019; Charniak and Johnson, 2005; Ge and Mooney, 2006).
it is challenging to optimize the performance of beam search for modern neural architectures.
contrasting
test_290
On the one hand, technical proposals as pre-trained embeddings, finetuning, and end-to-end modeling, have advanced NLP greatly.
neural advances often overlook MRL complexities, and disregard strategies that were proven useful for MRLs in the past.
contrasting
test_2104
This is clear from their equation (7) where the empirical distribution is defined over the first n samples, instead of the B samples that we use here
they claim, at least in the text, to use Fˆ n instead of FˆB for their estimator Vˆ n n .
entailment
test_2365
It is an open question whether these word-pieces capture relevant aspects of morphology.
it is unclear that the strategy of relying on chars or char-strings is adequate for encoding nonconcatenative phenomena that go beyond simple character sequencing, such as templatic morphology, substraction, reduplication, and more (Ackerman and Malouf, 2006; Blevins, 2016).
entailment
test_2578
During the evaluation of the argument pair extraction performance, we determine the argument pairing relation on span level with the sentence pairing results.
we perform binary classification on all sentence pairs enumerated from a candidate argument pair.
entailment
test_758
They collected explicit argument pairs with freely omissible discourse connectives which can be dropped independently of the context without changing the interpretation of the discourse relation.
sporleder and Lascarides (2008) argued training on explicit argument pairs was not a good strategy.
contrasting
test_2966
On the other hand, text classification results provided new insights on the most predictive features for distinguishing abusive and not-abusive swear words.
we found that a wide range of features can actually improve the models performance.
entailment
test_2644
As shown in Table 7, we decompose queries from several other datasets, using our decomposition model trained on only questions in HOTPOTQAand Common Crawl.
we generate subquestions for (1) questions in ComplexWebQuestions (Talmor and Berant, 2018), which are multihop questions about knowledge-bases, (2) questions in CLEVR (Johnson et al., 2017b), which are multi-hop questions about images, and (3) claims (statements) in fact-verification challenges, FEVER 1.0 (Thorne et a...
entailment
test_1925
The authors of the original paper mentioned that word n-grams and word embeddings are not suitable for cross-language classification.
the considered features are: baseline, domain features, POS n-grams and dependency n-grams.
reasoning
test_1766
In addition, from the perspective of EL solutions, it is observed that (1) EL on short text tends to require excessive hand-crafted features specific to a certain kind of application, which makes it not necessarily applicable to others; and (2) short-text oriented corpus finds itself inappropriate for evaluating the cl...
long-text oriented corpora are considered to be at least of equal, if not greater, significance to verifying the effectiveness and robustness of EL methods.
reasoning
test_2956
We describe a resource and first studies towards answering this question.
we create wikiHowToImprove, a collection of revision histories for about 2.7 million sentences from about 246 000 wikiHow articles.
entailment
test_2447
Figure 1 shows that masking achieves decent performance without hyperparameter search.
(i) a large initial sparsity removing most pretrained parameters, e.g., 95%, leads to bad performance for the four tasks.
entailment
test_3298
We hypothesize that these datasets contain many examples where their gold scores are easy to predict by either having similar structure and word choice and a high score or dissimilar structure and word choice and a low score.
this trick allows for the expectation under q to be approximated through sampling in a way that preserves backpropagation.
neutral
test_2160
We present a case study to showcase the power of contextualized weak supervision.
we investigate the differences between the expanded seed words in the plain corpus and contextualized corpus over iterations.
entailment
test_3781
This information is repeated for all the vocabulary modules, although there is no way to go between languages on the TUFS website.
translation sets that did not link directly are discussed further below.
neutral
test_2561
We proposed a framework for learning structured representation of entity names under low-resource settings.
we focus on a challenging scenario, where entity names are given as textual strings without context.
entailment
test_866
For both setups, we normalize labels to be conform with the PropBank (Palmer et al., 2005) notation (e.g., A1 becomes ARG1).
as shown in Figure 1, the experiments with the full label set have a slightly better accuracy than the ones with a simplified label set, so we will present only the results for the former.
contrasting
test_2595
In this paper, we want to shed some light on the impact of the discrepancies between regular and open NER, and provides some valuable insights into the construction of general NER models in a more effective and efficient way.
we want to answer the following question: Can pretrained supervised neural networks still generalize well on NER when either weaker name regularity, lower mention coverage or inadequate context diversity exists?
entailment
test_2040
We use vertices to represent utterance content, and edges to represent dialog transitions between utterances.
there are two types of vertices: (1) a what-vertex that contains a keyword, and (2) a how-vertex that contains a responding mechanism (from a multi-mapping based generator in Section 3.1) to capture rich variability of expressions.
entailment
test_2939
In this section we describe a methodology to categorise gold standards according to linguistic complexity, required reasoning and background knowledge, and their factual correctness.
we use those dimensions as highlevel categories of a qualitative annotation schema for annotating question, expected answer and the corresponding context.
entailment
test_1397
We make use of the heuristics that nearby sequences in the document contain the most important information to recover the masked words.
the challenging retrieval part can be replaced by soft-attention mechanism, making our model much easier to train.
reasoning
test_2991
Gradient judgments would account for the fact that bwick is typically judged to be a possible English word like blick but not as good
bwick is better than bnick but not as good as blick.
entailment
test_2610
Our goal is to obtain a fair comparison without the confounds that may result in performance differences on these two sets.
the examples from the two sets should be similar except for the presence of a feature that is biased in one set and anti-biased in the other.
entailment
test_2087
s and o are identified by their entity mentions, and p is identified by a unique ID.
two entities from different triples that have the same mentions will be regarded as the same node.
entailment
test_3392
After this initially automated training, we scheduled a 1hour long phone call with them to discuss our annotation instructions and annotation interface.
small-scale studies on manually annotated datasets have also been conducted (Morris and Picard, 2012; Lord et al., 2015).
neutral
test_652
TABLE-BERT is a BERT-base model that similar to our approach directly predicts the truth value of the statement.
the model does not use special embeddings to encode the table structure but relies on a template approach to format the table as natural language.
contrasting
test_1238
In NLP, dropped pronouns can cause loss of important information, such as the subject or object of the central predicate in a sentence, introducing ambiguity to applications such as machine translation (Nakaiwa and Shirai, 1996; Wang et al., 2016; Takeno et al., 2016), question answering (Choi et al., 2018; Reddy et al...
zero pronouns have recently received much research attention (Liu et al., 2017; Yin et al., 2018a,b)
reasoning
test_3339
Given that we did not find even the "good" subnetworks to be stable, or preferentially containing the heads that could have interpretable linguistic functions, the latter seems more likely.
the degree to which the "good" subnetworks overlap across tasks may be a useful way to characterize the tasks themselves.
neutral
test_2889
Therefore, our Hadith corpus relies on the source.
missing values or inconsistencies with the original book are dependent on Sunnah.com.
entailment
test_3161
While it is tempting to equate such information with the meaning of an utterance, a large body of literature in linguistics and psycholinguistics argues that an utterance conveys much more than a simple set of facts: it carries with it a halo of intimations arising from the speaker's choices, including considerations o...
Systems following instructions also require a means of segmenting continuous sensorimotor data and linking it to discrete linguistic categories (Regneri et al., 2013; Yagcioglu et al., 2018) (cf. the symbol grounding problem (Harnad, 1990)).
neutral
test_2723
In this work, the graph edges are labelled dependency relations, which are predicted as part of the actions of a transition-based dependency parser.
the Relation classifier uses the output embeddings of the top two elements on the stack and predicts the label of their dependency relation, conditioned on its direction.
entailment
test_3472
The research on TE dates back more than two decades and has made significant progress.
we often solve each task separately by first gathering task-specific training data and then tuning a machine learning system to learn the patterns in the data.
neutral
test_550
Longer texts offer the potential for discourse-level inferences, the addition of which should yield a dataset that is more difficult, more diverse, and less likely to contain trivial artifacts.
one might expect that asking annotators to read full paragraphs should increase the time required to create a single example; time which could potentially be better spent creating more examples.
contrasting
test_2413
Similarly, a predicate tagged with future tense may be more likely to have a goal argument.
these features are not mutually exclusive, and cannot be mutually exclusive.
entailment
test_931
They are considered improvements over word models, and their effectiveness is usually judged with benchmarks such as semantic similarity datasets.
most of these datasets are not designed for evaluating sense embeddings.
contrasting
test_2690
The interpreter is a parameter-free function that executes the editing action produced by the programmer.
the interpreter first checks if the action is the termination action.
entailment
test_3806
We used the BLEU score for evaluating our system performance.
we addressed this issue with BPE to make this whole process more efficient and reliable.
neutral
test_3804
In modern electronic text corpora, cuneiform text is represented in two distict latinizations: (1) sign-to-sign level graphemic transliteration and (2) phonemic transcription based on an approximation of the Akkadian language and its reconstructed sound system (Kouwenberg, 2011).
it would be useful to minimize ambiguity by first using a large dictionary lookup (e.g. consisting of the whole corpus of a given dialect in Oracc), and then trying to predict the correct phonological rendering only if the transcription is clearly ambiguous
neutral
test_3934
The inter annotator agreement is 0.708, based on Cohen's Kappa coefficient, which denotes a substantial agreement.
we found several difficult cases
neutral
test_1012
The main goal of our effort is to reduce the time needed by experts to produce training data for automatic CN generation.
the primary evaluation measure is the average time needed to obtain a proper pair.
reasoning
test_2708
To model the above ideas, in this paper, we propose a novel neural architecture named Edge-Enhanced Graph Convolutional Networks (EE-GCN), which explicitly takes advantage of the typed dependency labels with dynamic representations.
ee-GCN transforms a sentence to a graph by treating words and dependency labels as nodes and typed edges, respectively.
entailment
test_2807
Results obtained and analysis performed show how injecting syntax into the model helped achieve better results.
the joint learning of syntactic features allowed the model uncover complex syntactic patterns that could not be captured by simply fine-tuning BERT.
entailment
test_2243
In Table 1, we see that performance for both BERT-12 and BERT-24 steadily increases across all datasets with increasing N ; this trend holds for the other 7 pretrained models.
in the largest setting with N = 1M, the BERT-24 embeddings distilled from the best-performing layer for each dataset drastically outperform both Word2Vec and GloVe.
entailment
test_2992
Making use of data that are concept-aligned across the languages provides a certain amount of control (to the extent possible) of the influence of linguistic content on the forms that we are modeling.
these forms should be largely comparable across the languages in terms of how common they are in the active vocabulary of adult speakers.
entailment
test_1582
However, human ground-truth construction for summarization is time-consuming and laborintensive.
a more flexible summary generation framework could minimize manual labor and generate useful summaries more efficiently.
reasoning
test_142
Most of the previous AT and OT extraction methods formulate the task as a sequence tagging problem (Wang et al., 2016, 2017; Wang and Pan, 2018; Yu et al., 2019), specifically using a 5-class tag set: {BA (beginning of aspect), IA (inside of aspect), BP (beginning of opinion), IP (inside of opinion), O (others)}.
the sequence tagging methods suffer from a huge search space due to the compositionality of labels for extractive ABSA tasks, which has been proven in (Lee et al., 2017b;Hu et al., 2019).
contrasting
test_2779
Embedded in that neural machinery is latent knowledge about semantics, linguistic relations, and lexical features that are necessary to generate fluent text.
t5 has access to an additional source of knowledge that BERt does not.
entailment
test_1641
A natural way of verifying the instruction sets from Stage 1 is to have new workers follow them (Chen et al., 2019).
during Stage 2 Verification, a new worker is placed in the environment encountered by the Stage 1 worker and is provided with the NL instructions that were written by that Stage 1 worker.
reasoning
test_2978
We adopt and make key extensions to Grusky et al.’s (2018) methodology for the development of their Newsroom dataset to the Danish language.
our clarifications, extensions, and associated code presented here permit researchers to easily develop similar automatic summarisation datasets for other non-English languages.
entailment
test_2957
For simplicity, we focus on edits on the sentence level.
we consider all articles in wiki-How for which a revision history is available and examine each original sentence, henceforth base version, and how it is changed at subsequent points in time, henceforth revised versions.
entailment
test_326
Entity linking systems consider three sources of information: 1) similarity between mention strings and names for the KB entity; 2) comparison of the document context to information about the KB entity (e.g. entity description); 3) information contained in the KB, such as entity popularity or inter-entity relations
to the dense KBs in entity linking, concept linking uses sparse ontologies, which contain a unique identifier (CUI), title, and links to synonyms and related concepts, but rarely longform text.
contrasting
test_1230
However, algorithms also provide systematic ways to reduce bias, and some see the mitigation of bias in algorithm decisions as a potential opportunity to move the needle positively (Kleinberg et al., 2018).
we can apply frameworks of contemporaries in human behavior to machines (Rahwan et al., 2019), and perhaps benefit from a more scalable experimentation process.
reasoning
test_2534
We also plot the similarity within English, measured using two independently sampled batches.
gradients between two different languages are indeed less similar than those within the same language.
entailment
test_2251
In contrast, our method instead introduces a smaller rectifier network with ≈ 1000 additional parameters while still producing similar improvements
using trained constraints is computationally more efficient.
entailment
test_2877
In the domain of machine-readable lexicons, different projects have tackled the difficult issue of unifying features to describe lexical entries, allowing a meaningful semantic interoperability.
projects have aimed to unify the descriptors used to label specific forms in a lexical entry.
entailment
test_511
Very similar to our task, Kang et al. (2019) developed language models informed by discourse relations on the bridging task; given the first and last sentences, predicting the intermediate sentences (bidirectional flow).
they did not explicitly predict content words given context nor use them as a self-supervision signal in training.
contrasting
test_2006
We opted for a scale of 0-3, rather than a CE binary response, since it allows us to study various thresholds for better data selection.
the meanings of the scores are as follows: 0 is not suitable; 1 is suitable with small modifications, such as grammar or semantic; 2 is suitable; and 3 is extremely good as a CN.
entailment
test_3584
For all the experiments, we use finetune the pretrained model for 3 epoches with learning rate 2e-5 and batch size 32.
these results suggest that our method is more suitable for fine-tuning with smaller amounts of data, and that our approach to injecting the label information is at least not detrimental to the original pretrained model.
neutral
test_2540
TNT enhances the language learning by utilizing text normalization pre-training objective, inspired by misspelling correction.
tNt randomly manipulates tokens from the input text.
entailment
test_944
The only previous attempt at normalizing Italian social media data is from Weber and Zhekova (2016).
they have a different scope of the task, mostly focusing on readability, not on normalization on the lexical level.
contrasting
test_3945
Analor (Avanzi, Lacheret-Dujour, Victorri, 2008) is a semi-automatic segmentation tool developed within this framework.
this pilot corpus contains transcriptions of preparing a meal, meetings, conferences, radio transmissions, interviews, etc.
neutral
test_695
XLNet (Yang et al., 2019) also marginalize over all possible factorizations.
their work is focused on the conditional distribution p(y|x), and they do not marginalize over all possible factorizations of the joint distribution.
contrasting
test_2842
Most algorithms and tools in the field simply assume that the given text is free from errors such as the above and they fail if this is not the case.
this is true for basic problems such as POS-tagging, sentence parsing and entity recognition, as well as for more complex problems such as question answering (QA).
entailment
test_2817
While opinion mining is a well-established task with many standard datasets and well-defined methodologies, emotion mining has received less attention due to its complexity.
the annotated gold standard resources available are not enough.
entailment
test_2537
Our work is also related to transfer learning (Pan and Yang, 2010) and multi-task learning (Ruder, 2017).
prior work have observed (Rosenstein et al., 2005) and studied (Wang et al., 2019) negative transfer, such that transferring knowledge from source tasks can degrade the performance in the target task.
entailment
test_2021
Motivated by this, we propose P 2 BOT, a transmitter-receiver based framework with the aim of explicitly modeling understanding.
p 2 BOT incorporates mutual persona perception to enhance the quality of personalized dialogue generation.
entailment
test_101
This can result in different computational complexity.
since a typical Graphics Processing Unit (GPU) computes matrices in parallel, the actual difference in inference time is not that significant.
contrasting
test_2200
Considering the above problem in the framework of chart parsing, we would like to construct a data structure to efficiently access all chart items.
when partial information is provided, this data structure can quickly find all compatible chart items.
entailment
test_2762
To address this problem, we study automated techniques to improve datasets for training and testing.
we focus on paraphrase identification task, which aims to determine whether two given sentences are semantically equivalent.
entailment
test_2287
Our implementations are all based on PyTorch.
to implement our classification based and span-based model, we use pytorch-transformers (Wolf et al., 2019) 6 .
entailment
test_2794
At the same time, existing corpora with quotation annotation vary greatly in whether, how, and to what extent they cover the interpersonal structure (i.e., who communicates what to whom -see Section 2. for details).
we are not aware of any large-scale, publicly available corpora which mark both speakers and addressees for quotations in literary text.
entailment
test_2624
While the OWA is beneficial because it helps us assess KGE calibration under more realistic conditions, it is also challenging because it significantly changes the requirements for evaluation.
now we need a label for every triple considered, whereas with the CWA we only needed labels for a small group of positives.
entailment
test_2582
Finally, we ask the two annotators involved in Step 3 to judge whether each hyperbolic sentence obtained in the previous step is indeed hyperbolic.
if at least one of them thinks a sentence is not hyperbolic or does not truly reflect the meaning of its non-hyperbolic counterpart, we will delete it.
entailment
test_2384
Although the dependency connections might be less specific to the training data than the whole tree structures, the major limitation of the edge-based representation is that it only captures the pairwise (local) connections between the words and completely ignores the overall (global) importance of the words in the sen...
some words in a given sentence might involve more useful information for relation prediction in RE than the other words, and the dependency tree for this sentence can help to better identify those important words and assign higher importance scores for them (e.g., choosing the words along the shortest dependency paths ...
entailment
test_3361
For instance, in this work, if a propaganda class c is predicted by the multiple binary classifiers (indicates the sentence contains this propaganda technique), then the token-level predictions belonging to the propaganda class c should also exist.
neural networks are less interpretable and need to be trained with a large amount of data to make it possible to learn such implicit logic.
neutral
test_2826
While most research on emotion analysis focuses on emotion classification, including predicting the emotions of the writer as well as those of the reader of a text (Chang et al., 2015), a few studies so far have focused on identifying what might have triggered that emotion (Sailunaz et al., 2018).
the task has been framed as Emotion Cause Extraction (Lee et al., 2010a) from news and microblogs, where the cause of an emotion is usually a single clause (Chen et al., 2010) connected by a discourse relation to another clause that explicitly expresses a given emotion (Cheng et al., 2017), as in this example from Gui ...
entailment
test_2366
As a complementary area of investigation, a plausible direction would be to shift the focus from the decomposition of words into morphemes, to the organization of words as complete paradigms.
instead of relying on sub-word units, identify sets of words organized into morphological paradigms (Blevins, 2016).
entailment
test_1389
G module is implemented as a two-layer MLP (the number of neurons in the second layer is set as one).
the additional computing cost comes from the training of the two-layer MLP, which is of O(T * N * K * 1), where T is the number of training iterations, N number of training examples, K number of neurons in the first layer of MLP (without considering the compute cost of the activation function).
reasoning
test_2649
This was done using SUTime library (Chang and Manning, 2012) to tag time terms, which uses a regular expression based approach.
we focused on times that pinpointed hours within a day such as "two o'clock" or "noon".
entailment
test_83
Finally, research into multimodal or multi-view deep learning (Ngiam et al., 2011; Li et al., 2018) offers insights to effectively combine multiple data modalities or views on the same learning problem.
most work does not directly apply to our problem: i) the audio-text modality is significantly under-represented, ii) the models are typically not required to work online, and iii) most tasks are cast as document-level classification and not sequence labeling (Zadeh et al., 2018).
contrasting
test_372
As a result, the relationships between entities are not captured.
since KB is naturally a graph structure (nodes are entities and edges are relations between entities).
contrasting
test_3918
we describe two datasets providing disambiguations in the form of Wikipedia pages.
to gain insights on the features of each sense-annotated corpora, we provide a small analysis on the entropy and ambiguity levels.
neutral
test_1320
And the meaning of each character changes dramatically when the context changes.
a CSC system needs to recognize the semantics and aggregate the surrounding information for necessary modifications.
reasoning
test_2430
To obtain feature representations of entity-pairs containing their neighbors' information, our approach propagates attribute features of entity-pairs following these edges.
our approach uses a Graph Neural Network (GNN) to propagate the attribute features of entity-pairs over the PCG.
entailment
test_2609
In this paper, we propose a new strategy to enable the existing debiasing methods to be applicable in settings where there is minimum prior information about the biases.
models should automatically identify potentially biased examples without being pinpointed at a specific bias in advance.
entailment
test_2045
We observed that while there is still a gap between the systemgenerated and human-written Impressions, over 80% of our system-generated Impressions are as good 9 as the associated human-written Impressions.
73% (readability), and 71% (accuracy) of our system-generated Impressions ties with human-written Impressions, both achieving full-score of 3; nonetheless, this percentage is 62% for completeness metric.
entailment
test_2593
As there is no currently available dataset to investigate this problem, this paper proposes to conduct randomization test on standard benchmarks.
we erase name regularity, mention coverage and context diversity respectively from the benchmarks, in order to explore their impact on the generalization ability of models.
entailment
test_535
This indicates that unregularized training optimizes faster on certain examples, possibly due to the presence of biases.
self-debiased training maintains relatively less variability of losses throughout the training.
contrasting
test_3489
Latent state s ∈ S contains the current game information (e.g. locations of the player and items, the player’s inventory), which is only partially reflected in o.
we train this model for 10 epochs until the validation loss converges, unlike previous models which we train for 3 epochs.
neutral
test_2400
We use the GPT-2 architecture (Radford et al., 2019), which is trained using a causal language modeling loss (CLM), and includes a leftto-right decoder suitable for text generation.
we use the gpt2-medium model.
entailment
test_1356
For (i), the target probability of each word is set proportional to the square root of its frequency in the visited stream.
highly frequent words would take a smaller portion compared to reservoir sampling where the word distribution in the memory is linear to its frequency in the visited stream, leaving space for storing more diverse examples.
reasoning
test_2796
VGG consists of 99 texts for a total of more than 22,000 pages that were written in Italian during the period of the First World War or shortly afterwards.
they go from 1914 up to 1923, in order to cover not only the years of the war, but also the cultural and social environment leading to the war and the aftermath of the Great War.
entailment
test_3394
For instance, BEESL identifies the +REG-ULATION event anchored at "activated" in the following sentence: "Tax [...] maximally activated HTLV-I-LTR-CAT and kappa B-fos-CA" albeit the gold standard does not contain the event in this instance.
this shows that BEESL's performance is clearly affected, but that the system is relatively robust to noisy, nongold silver entities.
neutral
test_3086
Since NER is not the focus of this study, the readers can choose the public Chinese NER API 4 from Baidu for fast experiments.
interpretability is very important in the Ai-empowered healthcare studies.
neutral
test_271
Systems for supervised or semi-supervised paradigm completion are commonly being evaluated using word-level accuracy (Dreyer and Eisner, 2011; Cotterell et al., 2017).
this is not possible for our task because our system cannot access the gold data paradigm slot descriptions and, thus, does not necessarily produce one word for each ground-truth inflected form.
contrasting
test_2532
Similar to NER, POS is also a sequence labelling task but with a focus on synthetic knowledge.
we use the Universal Dependencies treebanks (Nivre et al., 2018).
entailment
test_3759
When English is the source language, and Japanese is the target language, there are only 5 pairs in the test data where the source and target words are identical, i.e., cases where the copy baseline is correct.
Details of the evaluation datasets are shown in Table 2.
neutral
test_192
Similar to results on Toxicity Comments, we find that both Weight and Supplement perform significantly better than Baseline in terms of IPTTS AUC and FPED, and the results of Weight and Supplement are comparable.
we notice that Weight and Supplement improve FNED slightly, while the differences are not statistically significant at confidence level 0.05.
contrasting
test_2636
To address comprehension-based generation like generative QA, PALM uses the pre-training objectives that are closely related to the downstream tasks.
it differs from existing generative pre-training methods in that PALM goes beyond the solely autoencoding/autoregressive methods and combines the merits of autoencoding and autoregression in a single framework.
entailment
End of preview. Expand in Data Studio
README.md exists but content is empty.
Downloads last month
24