tasksource/scinli · Datasets at Hugging Face

id stringlengths 7 12	sentence1 stringlengths 6 1.27k	sentence2 stringlengths 6 926	label stringclasses 4 values
train_94100	To amortize both costs and annotators' effort at answering questions, each HIT requires the participants to annotate three stories after answering demographic questions.	my favorite set was all the art that was made out of stylish trash .	neutral
train_94101	S α : the flowers were in full bloom.	if annotators are asked to write their own text given a reference sentence, they may simply produce arbitrarily paraphrased output which does not exhibit a stylistic diversity.	neutral
train_94102	The TA shows that the annotators rarely indicate the same scenes, even if they are asked to annotate an event in the screenplay that is described by a specific synopsis sentence.	it still remains competitive against the baselines, indicating that tracking TPs in screenplays fully automatically is feasible.	neutral
train_94103	Traditional suicide risk assessment like Suicide Probability Scale (Bagge and Osman, 1998), Adult Suicide Ideation Questionnaire (Fu et al., 2007), Suicidal Affect-Behavior-Cognition Scale (Harris et al., 2015), etc.	the experimental result showed that by combining all of three emotion traits together, the proposed model could generate more discriminative suicidal prediction performance.	neutral
train_94104	Although highly beneficial, those models' use may not be sufficient for the case of dialogue as response generation for goal-oriented dialogue from extremely limited data requires specialized tools.	the base optimization objective is as follows: where x usr is user's query, x sys is the system's response, c is the dialogue context, and F e and F d are respectively hierarchical encoder and decoder.	neutral
train_94105	The model (as well as its variants listed above) is implemented in PyTorch (Paszke et al., 2017), and the code is openly available 1 .	we use it as-is (ZSDG) as well its variation as follows.	neutral
train_94106	Given the target domain, we first train LAED model(s) on the MetaLWOz data -here we exclude from training every domain that might overlap with the target one.	for the NLU ZSDG setup, we annotated all available SMD data and randomly selected a subset of 1000 utterances from each source domain, and 200 utterances from the target domain.	neutral
train_94107	We first describe the task we are addressing in this paper, and the corresponding base model.	we perform an exhaustive ablation study of DiKTNet by comparing it to all its variations mentioned above: HRED, HRED+ELMo, and HRED+LAED.	neutral
train_94108	In addition, with the recently adopted technique of training dialogue systems end-to-end data-efficiency of such systems becomes the key question in their adoption in practical applications.	we have a set of dialogues in source domains and just a few seed dialogues in the target domain.	neutral
train_94109	All bold-face results are statistically significant to p < 0.01. that MGT results in more general representations of language, thereby facilitating better transfer.	across both base architectures, MGT outperforms ensembling.	neutral
train_94110	In such case, the personal information about "School" may be fake.	if n crt < n wrg , the worker's decision will be "Fraud" and the manager's decision will be "Fraud" too.	neutral
train_94111	Then, based on the KG, we present structured dialogue management (Section 3) to explore the optimal dialogue strategy with reinforcement learning.	we find that, compared with Full-S and HP-S, MP-S is unable to learn any useful dialogue policy during training.	neutral
train_94112	It means that using the hierarchical policy to unfold dialogues is necessary for our task.	the system actions in our task are selecting nodes in the KG to generate questions.	neutral
train_94113	Given the same or similar dialog context, there may exist many valid expressions for the responses with the same dialog act a, each corresponding to a certain configuration of z.	this is because that taking average metrics may cause that the safety responses get higher scores than meaningful and diverse responses if most of these valid responses are not related to the ground-truth.	neutral
train_94114	In this Corpus, the dialog act categories are {Inform, Question, Directive, Commissive}.	in our experiment, we also find that HAE-CT still faces serious safe response problems.	neutral
train_94115	To tackle the challenges, we propose to take advantage of the hierarchical nature of response generation.	cNN models have been shown to be efficient for NLP and have achieved excellent results in sentence modeling and classification.	neutral
train_94116	This method needs the preorganized topic/theme annotations for each conversation, which are prohibitively expensive to obtain.	note that if annotators rate different options, this triplet will be counted as "tie".	neutral
train_94117	For example, when "The Shining" is mentioned, some of the top biased words are "creepy", "gory" and "scary", which are consistent with the style of the horror movie, and "stephen", who is the original creator of the movie.	then it combines the information with its hidden representation to form its updated representation at the next layer.	neutral
train_94118	By further introducing the recommender system's knowledge of the items that have appeared in dialog, we guide the dialog system to generate responses that are more consistent with the user's interests.	kBRD (D) stands for only incorporating the dialog contents.	neutral
train_94119	These actions can be regarded as users' feedbacks that reflect users' interest.	an ideal recommender dialog system is an endto-end framework that can effectively integrate the two systems so that they can bring mutual benefits to one another.	neutral
train_94120	The word embedding is also 1000 dimension, trained from random initialization.	while moving towards z AE of an arXiv-style sentence "This would be an interesting viewpoint", the responses gradu-context Do you want to play a game?	neutral
train_94121	Table 6 presents the generated examples of our models and baselines, our model can extract the keywords from the context which is helpful to generate an informative response, but the HRED model often generates safe responses like "Metoo" or "Y es".	we first rewrite the contexts in the test set with CRN, and then retrieve 10 candidates with the rewritten context from the index 6 .	neutral
train_94122	Numbers in parentheses indicate the gains or losses after adding the persona conditions.	the gain achieved by the IMN ctx model is limited.	neutral
train_94123	For the relevance score, the retrievalindependent Seq2Seq-MMI establishes a strong baseline.	relevance Fluency ours 2.69 3.11 3.20 Seq2Seq-MMI 2.64 3.02 3.14 Table 4: Ablation study on the ranker.	neutral
train_94124	First, their skeleton extractor is pre-trained by the lexical overlap between the retrieved response and the golden response.	for a fair comparison, the lengths of the skeletons (the number of preserving words) generated by PMI and Keywords are kept as the same with the one generated by our skeleton extractor.	neutral
train_94125	DSTRead (Gao et al., 2019) formulate the dialogue state tracking task as a reading comprehension problem by asking slot specified questions to the BERT model and find the answer span in the dialogue history for each of the pre-defined combined slot.	then for each s in S, the third CMRD generates the value sequence V d,s based on the corresponding h s,d .	neutral
train_94126	When the size of paired data is small, the basic Seq2Seq model tends to generate more generic responses.	different from existing work, we assume that n is small (e.g., a few hundred thousands) and further assume that there is another set with T i a piece of plain text sharing the same characteristics with {Y i } n i=1 (e.g., both are questions) and N > n. Our goal is to learn a generation probability P (Y \|X) with both d P and d U .	neutral
train_94127	Indeed, existing big conversation data mix various intentions, styles, emotions, personas, and so on.	suppose that we have a dataset D is a pair of message-response, and n represents the number of pairs in D P .	neutral
train_94128	We conduct our experiments on a short-text conversation benchmark dataset (Shang et al., 2015) which contains about 4 million post-response pairs from the Sina Weibo 2 , a Chinese social platforms.	α denotes the attention weight.	neutral
train_94129	This is consistent with our motivation that open-domain short-text conversation covers a wide range of topics and areas, and the top frequent words are not enough to capture the content of most training pairs.	many enhanced encoder-decoder approaches have been proposed to improve the quality of generated responses.	neutral
train_94130	(1) The model tends to select more gain from a lower Kullback-Leibler (KL) divergence during training, which encourages the approximate posterior close to Gaussian prior, rendering the latent space of the former unused.	our work is supported by National Natural Science Foundation of China (61976154), Tianjin Natural Science Foundation (18JCY-BJC15500), National Natural Science Foundation of China (61771333), Tianjin Municipal Science and Technology Project (18ZXZNGX00330), and the Foundation of State Key Laboratory of Cognitive Intelligence, iFLYTEK(CIoS-20190001).	neutral
train_94131	(2) The speakers in a dyadic conversation have different linguistic characteristics, sentiments and personalities.	(3) Compared with the Gau-Generators in Generator KL, the vMF-Generators (i.e., SVN, SSVN Gau−E and SSVN) have the much higher KL values, indicating that the vMF is a better selection than Gaussian to solve the KL-vanishing problem.	neutral
train_94132	We compare our models with Information Retrieval (IR) based models and recommendation-only models.	additionally, the recommendations are not grounded in real observed movie preferences, which may make trained models less consistent with actual users.	neutral
train_94133	First, our response generation task only includes turns where the system's dialogue act is CONFORM SQL.	after annotation and reviewing, 3,007 of them were finished and kept in the final dataset.	neutral
train_94134	The difficulty of constructing dialogue-based NLIDB systems stems from these requirements.	the majority of previous work focus on converting a single, complex question into its corresponding SQL query.	neutral
train_94135	In task-oriented dialog systems, the user simulator's task is to complete a pre-defined goal by interacting with the system.	we used policy gradient method to train dialog systems (williams, 1992).	neutral
train_94136	When the dialog policy is bad, even if the NLG module can generate natural system responses, the users would still think the system is unnatural.	this is because SL-template's dialog manager is learned with supervised learning, less rigid than the agenda-based dialog policy, which further leads to a less rigid behaviour of the trained dialog system.	neutral
train_94137	The performance of the simulator directly impacts the RL policy.	another common automatic metrics used to assess the language model is BLEU, but since this is a user simulator study and we don't have ground truth, BLEU score is not available.	neutral
train_94138	One of the main drawbacks of HOCA is the generation of high-order tensor, the size of the high-order tensor will increases exponentially with the number of modalities as n i=1 t i , resulting in a lot of computation.	space (G) Training Time (s) HOCA-U 4.1 16418 HOCA-UBT 9.7 37040 L-HOCA-UBT 5.6 24615 The theoretical complexity of different attention mechanisms is illustrated in section 4.3.	neutral
train_94139	We simply compare the descriptions generated by HOCA-U and L-HOCA-UBT, respectively.	we implement the multimodal correlation between the different modalities in a more efficient way with lowrank approximation which has been widely used in the community of vision and language (Lei et al., 2015;Liu et al., 2018;.	neutral
train_94140	The goal of our work is to deal with unpaired image-caption data for image captioning.	baseline (10%) : a group of sheep standing next to each other.	neutral
train_94141	To compute the cross entropy loss, we use the pseudo-label assignment for the Unlabeled-COCO images.	these pseudo-labels are not noisefree, thus treating them equally with the paired ones is detrimental.	neutral
train_94142	After learning unimodal (H) and multimodal (Ĥ) representations of context, we use a Memory Fusion Network (MFN) (Zadeh et al., 2018a) to model the punchline ( , -∈/,$ , -∈/,0 , -∈/,1 2 .	a full understanding of humor requires analyzing the context of the punchline.	neutral
train_94143	Execution of the command then requires mapping the command into the physical visual space, after which the appropriate action can be taken.	talk2Car contains images in realistic settings accompanied by free language in contrast to curated datasets such as MS-COCO.	neutral
train_94144	We use only the articles from trustworthy media sources, according to our MBFC labels.	testing on Snopes + Reuters data.	neutral
train_94145	Table 1: Accuracy and Average Precision for individual feature types, calculated using 10-fold cross-validation using the Snopes dataset (S), and the Snopes+Reuters dataset (S+R).	there are cases when an image is completely legitimate, but it is published alongside some text that does not reflect its content accurately.	neutral
train_94146	Our new model generalizes the planning space of VPN to reason about intermediate goals and obstacles, and includes recurrent action generation for trajectories with multiple goals.	both metrics fail to reflect this.	neutral
train_94147	During fine-turning, we continuously generate additional examples using model failures.	without the implicit reasoning discriminator (-Implicit discriminator), the additional examples make learning too difficult, and do not help.	neutral
train_94148	E (I, B, R, T ) is a noncontextualized representation for each token and of its position in text, but also of the content and position of the bounding boxes.	vCR has much longer questions and answers compared to other popular visual Question Answering (vQA) datasets, such as vQA v1 (Antol et al., 2015), vQA v2 (Goyal et al., 2017) and GQA (Hudson and Manning, 2019), requiring more modeling capacity for language understanding.	neutral
train_94149	GQA (Hudson and Manning, 2019) uses real scenes from Visual Genome, but the language is artificially generated.	we found this to be surprisingly competitive on VCR compared to published baselines, perhaps due to our choice of powerful pretrained models.	neutral
train_94150	input sig- nals that can help to differentiate domains apart have has no impact on the task itself.	they are prone to fail when there is a severe change in the marginal distribution that is task irrelevant.	neutral
train_94151	The overall distribution of α values at convergence can be seen in Figure 3.	the decoder self and context attention only learn to distribute these parameters in a single mode.	neutral
train_94152	Encoder Architecture: Another way to encourage this robustness is through inductive bias in the encoder architecture.	in this work we present a probabilistic latent variable model capable of disentangling stylistic features of fonts from the underlying structure of each character.	neutral
train_94153	Since our model attempts to learn a smooth manifold over the latent style, we can also perform interpolation between the inferred font representations, something which is not directly possible using either of the baselines.	we propose a deep probabilistic approach that combines aspects of both these lines of past work.	neutral
train_94154	These are broken down into training, dev, and test splits of 7649, 1473, and 1560 fonts respectively.	we use a convolutional neural network which takes in a batch of characters from a single font, concatenated with their respective character type embedding.	neutral
train_94155	MacCartney and Manning (2009); Angeli and Manning (2014); and others incorporated knowledge of verb signatures within a natural logic framework (MacCart-ney, 2009;Sánchez Valencia, 1991) in order to perform natural language inference.	in this paper, we revisit the question of whether neural models of natural language inference-which are not explicitly endowed with knowledge of verbs' lexical semantic categorieslearn to make inferences about veridicality consistent with those made by humans.	neutral
train_94156	None of the traditional criteria for basic color terms hold up robustly.	the sequence gamma can have larger magnitude if, despite poor sifting, the order of the basic terms is correct.	neutral
train_94157	The results indicate that the BiLSTM-CRF framework is a better fit for encoding order information and long-range context dependency for such sequence labeling task.	in sentence S1, the negative focus is the propositional clause until the market close, yielding the interpretation that mutual fund trades take effect, but not until the market close.	neutral
train_94158	Although methods like (Elsner et al., 2007;Li and Jurafsky, 2017) attempt to combine these aspects of coherence, to our knowledge, no methods consider all the three aspects together in a single framework.	our bi-LSTM sentence encoder gives a representation h i ∈ R 2p for each sentence s i in the document.	neutral
train_94159	We observed that the naive pairwise ranking loss that uses a fixed margin unfairly penalizes the locally positive sentences during training.	the main differences between our implementation and the implementation referred in their paper are that we used a bi-LStM (as opposed to simple RNN) for sentence encoding and trained the network with the Adam optimizer (as opposed to AdaGrad).	neutral
train_94160	As shown in Table 6, the addition of the global model and LM loss to the local model improves performance on the standard discrimination task by 1.34%.	this section presents details of our experiment procedures and results.	neutral
train_94161	Subsequent studies extended the basic entity grid model.	we show that most of these models often fail on harder tasks with more realistic application scenarios.	neutral
train_94162	Here, except the random model, all of the baselines are based on neural networks, which are typically more competitive than traditional approaches (e.g., utilizing handcraft features).	the primary goal of sentence set ordering is to put an unordered set of sentences into a coherent paragraph in the correct order.	neutral
train_94163	This is related to the global dependencies of an entire paragraph.	we discuss the reason for our highest performance among previous models by classifying the factors affecting performance into four categories: sentence set modeling, topic latent vectors, permutation invariance, and attention at decoder.	neutral
train_94164	In this section, we present TGCM as a novel topic-guided coherence modeling for SSO.	in Table 2-3, the several values of some models are directly taken from (Cui et al., 2018), while we implemented the rest of models using their public code with the same experimental setup.	neutral
train_94165	Although TGCM-S do not utilize topical context features at encoder and decoder (such as t p and t s ), TGCM-S also employs the transformer-based attentive pointer decoder.	the pairwise model only learns the relative order from sentence-pair interactions.	neutral
train_94166	The topic latent vector t s i captures how topical a sentence s i is in itself (local context), whereas the topic latent vector t p representes the overall theme of a paragraph p (global context).	other parameters were initialized randomly based on He et al.	neutral
train_94167	Since our decoder directly receives the set of sentence vectors regardless of their input order via attention, our encoder is free from the constraint that all sentences must be compressed into a single fixedlength vector as a paragraph representation.	these models can capture global dependencies regardless of the order of input sentences.	neutral
train_94168	The attention layer repeatedly decides the next sentence from sentence-pair interactions between the current predicted sentence and each paragraph-independent candidate {s i } 5 1 .	with the multiple attention modules, TGCM draws global dependencies among sentences by attending over the topic-sensitive sentence vectors repeatedly.	neutral
train_94169	Its results were only slightly less good than those of stand alone GEN. To further investigate comparisons between different architectures for solving the attachment problem, we compared various local models extended with the MS-DAG decoding algorithm discussed in Section 5, giving the global results shown in the righthand columns of Tables 3 and 4.	we include LFs are expert-composed functions that make an attachment prediction for a given candidate: each LF returns a 1, a 0 or a -1 ("attached"/"do not know"/"not attached") for each candidate.	neutral
train_94170	To generate a large number of discourse structures via distant supervision from sentiment, we propose a four-step approach, shown in Figures 2 and 3.	although in principle new corpora could be created, the annotation process is expensive and time consuming.	neutral
train_94171	We study this question on the ACL and EMNLP paper collections and present an analysis on how well deep learning techniques can infer the authors of a paper.	a separate convolutional layer is dedicated to learning from part-of-speech sequences.	neutral
train_94172	For instance, in Figure 1, with identified key elements, one can easily categorize the incident in dimensions of "age of harasser" (adult), "single/multiple harasser(s)" (single), "type of harasser" (unspecified), "type of location" (park) , "time of day" (day time).	karlekar and Bansal (2018) were the first group to our knowledge that applied NLP to analyze large amount ( ∼10,000) of sexual harassment stories.	neutral
train_94173	This could explain why the regular BiLSTM model got lower performance than the CNN model.	the missing information is hence marked as "unspecified".	neutral
train_94174	In this study, we manually annotated those stories with labels in the dimensions of location, time, and harassers' characteristics, and marked the key elements related to these dimensions.	when training with key element extractions, it put almost all attention on the harasser "young man" (S2), which helped the model make correct prediction of "young harasser".	neutral
train_94175	The volume of social media data is huge, and the more we can extract from these data, the more powerful we can be as part of the efforts to build a safer and more inclusive communities.	the position of each word in this context sequence is crucial information for the CNN model to make the correct predictions (Nguyen and Grishman, 2015).	neutral
train_94176	The results show that our approach in general performs better in the Twitter dataset.	to the best of our knowledge, our work is the first that investigate VAEs as a tool for training data expansion, so as to enhance machine learning performance with limited amount of labeled data.	neutral
train_94177	We design a hierarchical seq2tree model that can better capture information of the AST of the equation.	these studies lost sight of important information of the math equation ASTs (e.g., parent and siblings of each node), despite of their promising results.	neutral
train_94178	One point to note is that our training procedure is based on finding shortest paths in a complete KB, and we use this same procedure on the incomplete KB setting.	the classifiers used for retrieval in the different iterations are identical.	neutral
train_94179	Our implementation shows comparable performance on the WikiMovies (MetaQA 1-hop) dataset, as shown in Table 3.	both models are limited by the number of facts and text that can fit into memory.	neutral
train_94180	As we discussed, different from the traditional RC task, the answer string will appear multiple times in different paragraphs in OpenQA.	besides, in the distantly supervised setup (Mintz et al., 2009), it is postulated that the paragraphs that contain the answer string are ground truths, while even some of the positive paragraphs are wrong-labeled as they do not concern the question, e.g.	neutral
train_94181	These paragraphs are inherently relevant to each other and provide enhanced evidence for the correct answer.	this indicates that SearchQA contains more positive paragraphs than other datasets.	neutral
train_94182	The scoring function of ComplEx is given by where e s , e o , w r ∈ C n are the embeddings of s, o, and r, respectively.	a matrix is block circulant if it can be written in the form 1b) .	neutral
train_94183	20162.96 Elsahar et al.	in this paper, we strive toward the above two issues via incorporating diversified contexts and answer-aware loss.	neutral
train_94184	Comparison with NAQANet The major difference between our model and NAQANet is that NAQANet does not have the reasoning module, i.e., M 0 is simply set as M P .	our NumNet model outperforms strong baselines with a large margin on the DRoP dataset.	neutral
train_94185	Hence, the learned model behaves more like a question language model with some loose context constraint, while it is unaware of the strong requirements that it should be closely grounded by the context and should be answered by the given answer.	these works either reduced the ground-truth data size or simplified the span-prediction QA task to answer sentence selection.	neutral
train_94186	In terms of syntactic correctness, KEAG and MHPGM both perform well thanks to their architectures of composing answer text and integrating knowledge.	an overview of the neural architecture of KEaG is depicted in Figure 1 3.2 Sequence-to-sequence model KEaG is built upon an extension of the sequenceto-sequence attentional model (Bahdanau et al., 2015;Nallapati et al., 2016;See et al., 2017).	neutral
train_94187	Each tuple (question, answer, and document) is fed into the BERT model as: " [CLS] As a remark, we want to highlight that both the DRD and ARD discriminators have been trained on different datasets than the final multiple choice QA model, which is trained and evaluated on the ARC Easy and Challenge datasets.	as shown in Figure 2, each supporting document is associated with a list of scores computed by the discriminators.	neutral
train_94188	This decision step is accomplished by a simple feed-forward network (Figure 3).	replacing GrU with Long-Short Term Memory (LSTM) cells (Hochreiter and Schmidhuber, 1997) gives similar performance but the training procedure is computationally more expensive.	neutral
train_94189	Humans capture that information by learning, by experiencing and from commonsense knowledge.	the weights in dictate to what extent the network should attend each document.	neutral
train_94190	In this work we use linguistic annotations as a basis for a Discourse-Aware Semantic Self-Attention encoder that we employ for reading comprehension on narrative texts.	), a h i is the output of head h i .	neutral
train_94191	2018for QA span prediction modeling.	to be specific, in the system without the paragraph-level neural retrieval module, we re-trained the sentence-level retrieval module with negative sentences directly sampled from the term-based retrieval set and then also re-trained the downstream QA or verification module.	neutral
train_94192	Progress on MRS has been made by improving individual IR or comprehension sub-modules with recent advancements on representative learning (Peters et al., 2018;Radford et al., 2018;Devlin et al., 2018).	this gives the final term-based retrieval set for HOtPOtQA.	neutral
train_94193	The task has gained popularity with its natural combination of information retrieval (IR) and machine comprehension (MC).	to validate the system design decisions in Sec 3 and reveal the importance of semantic retrieval towards downstream, we conducted a series of ablation and analysis experiments on all the modules.	neutral
train_94194	We add a "maybe" choice in PubMedQA to cover uncertain instances.	under reasoning-free setting where the annotator can see the conclusions, a single human achieves 90.4% accuracy and 84.2% macro-F1.	neutral
train_94195	To which organ system do the esophagus, liver, pancreas, small intestine, and colon belong?	although various neural Qa methods have achieved high performance on some of these datasets Trivedi et al., 2019;Tymoshenko et al., 2017;Seo et al., 2016;Wang and Jiang, 2016;De Cao et al., 2018;Back et al., 2018), we argue that more effort must be dedicated to explaining their inference process.	neutral
train_94196	Unsurprisingly, training and testing in the same domain (e.g., Fiction) leads to the best performance on sentence selection.	in MultiRC, AutoROCC outperforms our baseline of BERT + entire passage (row 10 vs 22) by 8.3% EM0, indicating that AutoROCC can filter out irrelevant content.	neutral
train_94197	Note that some of the works discussed here transfer knowledge from external datasets into the QA task they address (Chung et al., 2017;Pan et al., 2019;Min et al., 2017;Qiu et al., 2018;Chen et al., 2017).	the improvements are not large, e.g., the maximum improvement is 1.6% (when k = 4), which indicates that ROCC is robust to a certain extent to lexical variation.	neutral
train_94198	Open-domain question answering (QA) Inspired by the series of TREC QA competitions, 2 Chen et al.	centered around entity names, this model risks missing purely descriptive clues in the question.	neutral
train_94199	Because off-the-shelf IR systems generally optimize for shallow lexical similarity between query and candidate documents in favor of efficiency, a good proxy for this overlap is locating spans of text that have high lexical overlap with the intended supporting documents.	it is computationally inefficient, and has high variance especially for the second reasoning step and forward, because the context depends greatly on what queries have been chosen previously and their search results.	neutral

train_94100

To amortize both costs and annotators' effort at answering questions, each HIT requires the participants to annotate three stories after answering demographic questions.

my favorite set was all the art that was made out of stylish trash .

neutral

train_94101

S α : the flowers were in full bloom.

if annotators are asked to write their own text given a reference sentence, they may simply produce arbitrarily paraphrased output which does not exhibit a stylistic diversity.

neutral

train_94102

The TA shows that the annotators rarely indicate the same scenes, even if they are asked to annotate an event in the screenplay that is described by a specific synopsis sentence.

it still remains competitive against the baselines, indicating that tracking TPs in screenplays fully automatically is feasible.

neutral

train_94103

Traditional suicide risk assessment like Suicide Probability Scale (Bagge and Osman, 1998), Adult Suicide Ideation Questionnaire (Fu et al., 2007), Suicidal Affect-Behavior-Cognition Scale (Harris et al., 2015), etc.

the experimental result showed that by combining all of three emotion traits together, the proposed model could generate more discriminative suicidal prediction performance.

neutral

train_94104

Although highly beneficial, those models' use may not be sufficient for the case of dialogue as response generation for goal-oriented dialogue from extremely limited data requires specialized tools.

the base optimization objective is as follows: where x usr is user's query, x sys is the system's response, c is the dialogue context, and F e and F d are respectively hierarchical encoder and decoder.

neutral

train_94105

The model (as well as its variants listed above) is implemented in PyTorch (Paszke et al., 2017), and the code is openly available 1 .

we use it as-is (ZSDG) as well its variation as follows.

neutral

train_94106

Given the target domain, we first train LAED model(s) on the MetaLWOz data -here we exclude from training every domain that might overlap with the target one.

for the NLU ZSDG setup, we annotated all available SMD data and randomly selected a subset of 1000 utterances from each source domain, and 200 utterances from the target domain.

neutral

train_94107

We first describe the task we are addressing in this paper, and the corresponding base model.

we perform an exhaustive ablation study of DiKTNet by comparing it to all its variations mentioned above: HRED, HRED+ELMo, and HRED+LAED.

neutral

train_94108

In addition, with the recently adopted technique of training dialogue systems end-to-end data-efficiency of such systems becomes the key question in their adoption in practical applications.

we have a set of dialogues in source domains and just a few seed dialogues in the target domain.

neutral

train_94109

All bold-face results are statistically significant to p < 0.01. that MGT results in more general representations of language, thereby facilitating better transfer.

across both base architectures, MGT outperforms ensembling.

neutral

train_94110

In such case, the personal information about "School" may be fake.

if n crt < n wrg , the worker's decision will be "Fraud" and the manager's decision will be "Fraud" too.

neutral

train_94111

Then, based on the KG, we present structured dialogue management (Section 3) to explore the optimal dialogue strategy with reinforcement learning.

we find that, compared with Full-S and HP-S, MP-S is unable to learn any useful dialogue policy during training.

neutral

train_94112

It means that using the hierarchical policy to unfold dialogues is necessary for our task.

the system actions in our task are selecting nodes in the KG to generate questions.

neutral

train_94113

Given the same or similar dialog context, there may exist many valid expressions for the responses with the same dialog act a, each corresponding to a certain configuration of z.

this is because that taking average metrics may cause that the safety responses get higher scores than meaningful and diverse responses if most of these valid responses are not related to the ground-truth.

neutral

train_94114

In this Corpus, the dialog act categories are {Inform, Question, Directive, Commissive}.

in our experiment, we also find that HAE-CT still faces serious safe response problems.

neutral

train_94115

To tackle the challenges, we propose to take advantage of the hierarchical nature of response generation.

cNN models have been shown to be efficient for NLP and have achieved excellent results in sentence modeling and classification.

neutral

train_94116

This method needs the preorganized topic/theme annotations for each conversation, which are prohibitively expensive to obtain.

note that if annotators rate different options, this triplet will be counted as "tie".

neutral

train_94117

For example, when "The Shining" is mentioned, some of the top biased words are "creepy", "gory" and "scary", which are consistent with the style of the horror movie, and "stephen", who is the original creator of the movie.

then it combines the information with its hidden representation to form its updated representation at the next layer.

neutral

train_94118

By further introducing the recommender system's knowledge of the items that have appeared in dialog, we guide the dialog system to generate responses that are more consistent with the user's interests.

kBRD (D) stands for only incorporating the dialog contents.

neutral

train_94119

These actions can be regarded as users' feedbacks that reflect users' interest.

an ideal recommender dialog system is an endto-end framework that can effectively integrate the two systems so that they can bring mutual benefits to one another.

neutral

train_94120

The word embedding is also 1000 dimension, trained from random initialization.

while moving towards z AE of an arXiv-style sentence "This would be an interesting viewpoint", the responses gradu-context Do you want to play a game?

neutral

train_94121

Table 6 presents the generated examples of our models and baselines, our model can extract the keywords from the context which is helpful to generate an informative response, but the HRED model often generates safe responses like "Metoo" or "Y es".

we first rewrite the contexts in the test set with CRN, and then retrieve 10 candidates with the rewritten context from the index 6 .

neutral

train_94122

Numbers in parentheses indicate the gains or losses after adding the persona conditions.

the gain achieved by the IMN ctx model is limited.

neutral

train_94123

For the relevance score, the retrievalindependent Seq2Seq-MMI establishes a strong baseline.

relevance Fluency ours 2.69 3.11 3.20 Seq2Seq-MMI 2.64 3.02 3.14 Table 4: Ablation study on the ranker.

neutral

train_94124

First, their skeleton extractor is pre-trained by the lexical overlap between the retrieved response and the golden response.

for a fair comparison, the lengths of the skeletons (the number of preserving words) generated by PMI and Keywords are kept as the same with the one generated by our skeleton extractor.

neutral

train_94125

DSTRead (Gao et al., 2019) formulate the dialogue state tracking task as a reading comprehension problem by asking slot specified questions to the BERT model and find the answer span in the dialogue history for each of the pre-defined combined slot.

then for each s in S, the third CMRD generates the value sequence V d,s based on the corresponding h s,d .

neutral

train_94126

When the size of paired data is small, the basic Seq2Seq model tends to generate more generic responses.

different from existing work, we assume that n is small (e.g., a few hundred thousands) and further assume that there is another set with T i a piece of plain text sharing the same characteristics with {Y i } n i=1 (e.g., both are questions) and N > n. Our goal is to learn a generation probability P (Y |X) with both d P and d U .

neutral

train_94127

Indeed, existing big conversation data mix various intentions, styles, emotions, personas, and so on.

suppose that we have a dataset D is a pair of message-response, and n represents the number of pairs in D P .

neutral

train_94128

We conduct our experiments on a short-text conversation benchmark dataset (Shang et al., 2015) which contains about 4 million post-response pairs from the Sina Weibo 2 , a Chinese social platforms.

α denotes the attention weight.

neutral

train_94129

This is consistent with our motivation that open-domain short-text conversation covers a wide range of topics and areas, and the top frequent words are not enough to capture the content of most training pairs.

many enhanced encoder-decoder approaches have been proposed to improve the quality of generated responses.

neutral

train_94130

(1) The model tends to select more gain from a lower Kullback-Leibler (KL) divergence during training, which encourages the approximate posterior close to Gaussian prior, rendering the latent space of the former unused.

our work is supported by National Natural Science Foundation of China (61976154), Tianjin Natural Science Foundation (18JCY-BJC15500), National Natural Science Foundation of China (61771333), Tianjin Municipal Science and Technology Project (18ZXZNGX00330), and the Foundation of State Key Laboratory of Cognitive Intelligence, iFLYTEK(CIoS-20190001).

neutral

train_94131

(2) The speakers in a dyadic conversation have different linguistic characteristics, sentiments and personalities.

(3) Compared with the Gau-Generators in Generator KL, the vMF-Generators (i.e., SVN, SSVN Gau−E and SSVN) have the much higher KL values, indicating that the vMF is a better selection than Gaussian to solve the KL-vanishing problem.

neutral

train_94132

We compare our models with Information Retrieval (IR) based models and recommendation-only models.

additionally, the recommendations are not grounded in real observed movie preferences, which may make trained models less consistent with actual users.

neutral

train_94133

First, our response generation task only includes turns where the system's dialogue act is CONFORM SQL.

after annotation and reviewing, 3,007 of them were finished and kept in the final dataset.

neutral

train_94134

The difficulty of constructing dialogue-based NLIDB systems stems from these requirements.

the majority of previous work focus on converting a single, complex question into its corresponding SQL query.

neutral

train_94135

In task-oriented dialog systems, the user simulator's task is to complete a pre-defined goal by interacting with the system.

we used policy gradient method to train dialog systems (williams, 1992).

neutral

train_94136

When the dialog policy is bad, even if the NLG module can generate natural system responses, the users would still think the system is unnatural.

this is because SL-template's dialog manager is learned with supervised learning, less rigid than the agenda-based dialog policy, which further leads to a less rigid behaviour of the trained dialog system.

neutral

train_94137

The performance of the simulator directly impacts the RL policy.

another common automatic metrics used to assess the language model is BLEU, but since this is a user simulator study and we don't have ground truth, BLEU score is not available.

neutral

train_94138

One of the main drawbacks of HOCA is the generation of high-order tensor, the size of the high-order tensor will increases exponentially with the number of modalities as n i=1 t i , resulting in a lot of computation.

space (G) Training Time (s) HOCA-U 4.1 16418 HOCA-UBT 9.7 37040 L-HOCA-UBT 5.6 24615 The theoretical complexity of different attention mechanisms is illustrated in section 4.3.

neutral

train_94139

We simply compare the descriptions generated by HOCA-U and L-HOCA-UBT, respectively.

we implement the multimodal correlation between the different modalities in a more efficient way with lowrank approximation which has been widely used in the community of vision and language (Lei et al., 2015;Liu et al., 2018;.

neutral

train_94140

The goal of our work is to deal with unpaired image-caption data for image captioning.

baseline (10%) : a group of sheep standing next to each other.

neutral

train_94141

To compute the cross entropy loss, we use the pseudo-label assignment for the Unlabeled-COCO images.

these pseudo-labels are not noisefree, thus treating them equally with the paired ones is detrimental.

neutral

train_94142

After learning unimodal (H) and multimodal (Ĥ) representations of context, we use a Memory Fusion Network (MFN) (Zadeh et al., 2018a) to model the punchline ( , -∈/,$ , -∈/,0 , -∈/,1 2 .

a full understanding of humor requires analyzing the context of the punchline.

neutral

train_94143

Execution of the command then requires mapping the command into the physical visual space, after which the appropriate action can be taken.

talk2Car contains images in realistic settings accompanied by free language in contrast to curated datasets such as MS-COCO.

neutral

train_94144

We use only the articles from trustworthy media sources, according to our MBFC labels.

testing on Snopes + Reuters data.

neutral

train_94145

Table 1: Accuracy and Average Precision for individual feature types, calculated using 10-fold cross-validation using the Snopes dataset (S), and the Snopes+Reuters dataset (S+R).

there are cases when an image is completely legitimate, but it is published alongside some text that does not reflect its content accurately.

neutral

train_94146

Our new model generalizes the planning space of VPN to reason about intermediate goals and obstacles, and includes recurrent action generation for trajectories with multiple goals.

both metrics fail to reflect this.

neutral

train_94147

During fine-turning, we continuously generate additional examples using model failures.

without the implicit reasoning discriminator (-Implicit discriminator), the additional examples make learning too difficult, and do not help.

neutral

train_94148

E (I, B, R, T ) is a noncontextualized representation for each token and of its position in text, but also of the content and position of the bounding boxes.

vCR has much longer questions and answers compared to other popular visual Question Answering (vQA) datasets, such as vQA v1 (Antol et al., 2015), vQA v2 (Goyal et al., 2017) and GQA (Hudson and Manning, 2019), requiring more modeling capacity for language understanding.

neutral

train_94149

GQA (Hudson and Manning, 2019) uses real scenes from Visual Genome, but the language is artificially generated.

we found this to be surprisingly competitive on VCR compared to published baselines, perhaps due to our choice of powerful pretrained models.

neutral

train_94150

input sig- nals that can help to differentiate domains apart have has no impact on the task itself.

they are prone to fail when there is a severe change in the marginal distribution that is task irrelevant.

neutral

train_94151

The overall distribution of α values at convergence can be seen in Figure 3.

the decoder self and context attention only learn to distribute these parameters in a single mode.

neutral

train_94152

Encoder Architecture: Another way to encourage this robustness is through inductive bias in the encoder architecture.

in this work we present a probabilistic latent variable model capable of disentangling stylistic features of fonts from the underlying structure of each character.

neutral

train_94153

Since our model attempts to learn a smooth manifold over the latent style, we can also perform interpolation between the inferred font representations, something which is not directly possible using either of the baselines.

we propose a deep probabilistic approach that combines aspects of both these lines of past work.

neutral

train_94154

These are broken down into training, dev, and test splits of 7649, 1473, and 1560 fonts respectively.

we use a convolutional neural network which takes in a batch of characters from a single font, concatenated with their respective character type embedding.

neutral

train_94155

MacCartney and Manning (2009); Angeli and Manning (2014); and others incorporated knowledge of verb signatures within a natural logic framework (MacCart-ney, 2009;Sánchez Valencia, 1991) in order to perform natural language inference.

in this paper, we revisit the question of whether neural models of natural language inference-which are not explicitly endowed with knowledge of verbs' lexical semantic categorieslearn to make inferences about veridicality consistent with those made by humans.

neutral

train_94156

None of the traditional criteria for basic color terms hold up robustly.

the sequence gamma can have larger magnitude if, despite poor sifting, the order of the basic terms is correct.

neutral

train_94157

The results indicate that the BiLSTM-CRF framework is a better fit for encoding order information and long-range context dependency for such sequence labeling task.

in sentence S1, the negative focus is the propositional clause until the market close, yielding the interpretation that mutual fund trades take effect, but not until the market close.

neutral

train_94158

Although methods like (Elsner et al., 2007;Li and Jurafsky, 2017) attempt to combine these aspects of coherence, to our knowledge, no methods consider all the three aspects together in a single framework.

our bi-LSTM sentence encoder gives a representation h i ∈ R 2p for each sentence s i in the document.

neutral

train_94159

We observed that the naive pairwise ranking loss that uses a fixed margin unfairly penalizes the locally positive sentences during training.

the main differences between our implementation and the implementation referred in their paper are that we used a bi-LStM (as opposed to simple RNN) for sentence encoding and trained the network with the Adam optimizer (as opposed to AdaGrad).

neutral

train_94160

As shown in Table 6, the addition of the global model and LM loss to the local model improves performance on the standard discrimination task by 1.34%.

this section presents details of our experiment procedures and results.

neutral

train_94161

Subsequent studies extended the basic entity grid model.

we show that most of these models often fail on harder tasks with more realistic application scenarios.

neutral

train_94162

Here, except the random model, all of the baselines are based on neural networks, which are typically more competitive than traditional approaches (e.g., utilizing handcraft features).

the primary goal of sentence set ordering is to put an unordered set of sentences into a coherent paragraph in the correct order.

neutral

train_94163

This is related to the global dependencies of an entire paragraph.

we discuss the reason for our highest performance among previous models by classifying the factors affecting performance into four categories: sentence set modeling, topic latent vectors, permutation invariance, and attention at decoder.

neutral

train_94164

In this section, we present TGCM as a novel topic-guided coherence modeling for SSO.

in Table 2-3, the several values of some models are directly taken from (Cui et al., 2018), while we implemented the rest of models using their public code with the same experimental setup.

neutral

train_94165

Although TGCM-S do not utilize topical context features at encoder and decoder (such as t p and t s ), TGCM-S also employs the transformer-based attentive pointer decoder.

the pairwise model only learns the relative order from sentence-pair interactions.

neutral

train_94166

The topic latent vector t s i captures how topical a sentence s i is in itself (local context), whereas the topic latent vector t p representes the overall theme of a paragraph p (global context).

other parameters were initialized randomly based on He et al.

neutral

train_94167

Since our decoder directly receives the set of sentence vectors regardless of their input order via attention, our encoder is free from the constraint that all sentences must be compressed into a single fixedlength vector as a paragraph representation.

these models can capture global dependencies regardless of the order of input sentences.

neutral

train_94168

The attention layer repeatedly decides the next sentence from sentence-pair interactions between the current predicted sentence and each paragraph-independent candidate {s i } 5 1 .

with the multiple attention modules, TGCM draws global dependencies among sentences by attending over the topic-sensitive sentence vectors repeatedly.

neutral

train_94169

Its results were only slightly less good than those of stand alone GEN. To further investigate comparisons between different architectures for solving the attachment problem, we compared various local models extended with the MS-DAG decoding algorithm discussed in Section 5, giving the global results shown in the righthand columns of Tables 3 and 4.

we include LFs are expert-composed functions that make an attachment prediction for a given candidate: each LF returns a 1, a 0 or a -1 ("attached"/"do not know"/"not attached") for each candidate.

neutral

train_94170

To generate a large number of discourse structures via distant supervision from sentiment, we propose a four-step approach, shown in Figures 2 and 3.

although in principle new corpora could be created, the annotation process is expensive and time consuming.

neutral

train_94171

We study this question on the ACL and EMNLP paper collections and present an analysis on how well deep learning techniques can infer the authors of a paper.

a separate convolutional layer is dedicated to learning from part-of-speech sequences.

neutral

train_94172

For instance, in Figure 1, with identified key elements, one can easily categorize the incident in dimensions of "age of harasser" (adult), "single/multiple harasser(s)" (single), "type of harasser" (unspecified), "type of location" (park) , "time of day" (day time).

karlekar and Bansal (2018) were the first group to our knowledge that applied NLP to analyze large amount ( ∼10,000) of sexual harassment stories.

neutral

train_94173

This could explain why the regular BiLSTM model got lower performance than the CNN model.

the missing information is hence marked as "unspecified".

neutral

train_94174

In this study, we manually annotated those stories with labels in the dimensions of location, time, and harassers' characteristics, and marked the key elements related to these dimensions.

when training with key element extractions, it put almost all attention on the harasser "young man" (S2), which helped the model make correct prediction of "young harasser".

neutral

train_94175

The volume of social media data is huge, and the more we can extract from these data, the more powerful we can be as part of the efforts to build a safer and more inclusive communities.

the position of each word in this context sequence is crucial information for the CNN model to make the correct predictions (Nguyen and Grishman, 2015).

neutral

train_94176

The results show that our approach in general performs better in the Twitter dataset.

to the best of our knowledge, our work is the first that investigate VAEs as a tool for training data expansion, so as to enhance machine learning performance with limited amount of labeled data.

neutral

train_94177

We design a hierarchical seq2tree model that can better capture information of the AST of the equation.

these studies lost sight of important information of the math equation ASTs (e.g., parent and siblings of each node), despite of their promising results.

neutral

train_94178

One point to note is that our training procedure is based on finding shortest paths in a complete KB, and we use this same procedure on the incomplete KB setting.

the classifiers used for retrieval in the different iterations are identical.

neutral

train_94179

Our implementation shows comparable performance on the WikiMovies (MetaQA 1-hop) dataset, as shown in Table 3.

both models are limited by the number of facts and text that can fit into memory.

neutral

train_94180

As we discussed, different from the traditional RC task, the answer string will appear multiple times in different paragraphs in OpenQA.

besides, in the distantly supervised setup (Mintz et al., 2009), it is postulated that the paragraphs that contain the answer string are ground truths, while even some of the positive paragraphs are wrong-labeled as they do not concern the question, e.g.

neutral

train_94181

These paragraphs are inherently relevant to each other and provide enhanced evidence for the correct answer.

this indicates that SearchQA contains more positive paragraphs than other datasets.

neutral

train_94182

The scoring function of ComplEx is given by where e s , e o , w r ∈ C n are the embeddings of s, o, and r, respectively.

a matrix is block circulant if it can be written in the form 1b) .

neutral

train_94183

20162.96 Elsahar et al.

in this paper, we strive toward the above two issues via incorporating diversified contexts and answer-aware loss.

neutral

train_94184

Comparison with NAQANet The major difference between our model and NAQANet is that NAQANet does not have the reasoning module, i.e., M 0 is simply set as M P .

our NumNet model outperforms strong baselines with a large margin on the DRoP dataset.

neutral

train_94185

Hence, the learned model behaves more like a question language model with some loose context constraint, while it is unaware of the strong requirements that it should be closely grounded by the context and should be answered by the given answer.

these works either reduced the ground-truth data size or simplified the span-prediction QA task to answer sentence selection.

neutral

train_94186

In terms of syntactic correctness, KEAG and MHPGM both perform well thanks to their architectures of composing answer text and integrating knowledge.

an overview of the neural architecture of KEaG is depicted in Figure 1 3.2 Sequence-to-sequence model KEaG is built upon an extension of the sequenceto-sequence attentional model (Bahdanau et al., 2015;Nallapati et al., 2016;See et al., 2017).

neutral

train_94187

Each tuple (question, answer, and document) is fed into the BERT model as: " [CLS] As a remark, we want to highlight that both the DRD and ARD discriminators have been trained on different datasets than the final multiple choice QA model, which is trained and evaluated on the ARC Easy and Challenge datasets.

as shown in Figure 2, each supporting document is associated with a list of scores computed by the discriminators.

neutral

train_94188

This decision step is accomplished by a simple feed-forward network (Figure 3).

replacing GrU with Long-Short Term Memory (LSTM) cells (Hochreiter and Schmidhuber, 1997) gives similar performance but the training procedure is computationally more expensive.

neutral

train_94189

Humans capture that information by learning, by experiencing and from commonsense knowledge.

the weights in dictate to what extent the network should attend each document.

neutral

train_94190

In this work we use linguistic annotations as a basis for a Discourse-Aware Semantic Self-Attention encoder that we employ for reading comprehension on narrative texts.

), a h i is the output of head h i .

neutral

train_94191

2018for QA span prediction modeling.

to be specific, in the system without the paragraph-level neural retrieval module, we re-trained the sentence-level retrieval module with negative sentences directly sampled from the term-based retrieval set and then also re-trained the downstream QA or verification module.

neutral

train_94192

Progress on MRS has been made by improving individual IR or comprehension sub-modules with recent advancements on representative learning (Peters et al., 2018;Radford et al., 2018;Devlin et al., 2018).

this gives the final term-based retrieval set for HOtPOtQA.

neutral

train_94193

The task has gained popularity with its natural combination of information retrieval (IR) and machine comprehension (MC).

to validate the system design decisions in Sec 3 and reveal the importance of semantic retrieval towards downstream, we conducted a series of ablation and analysis experiments on all the modules.

neutral

train_94194

We add a "maybe" choice in PubMedQA to cover uncertain instances.

under reasoning-free setting where the annotator can see the conclusions, a single human achieves 90.4% accuracy and 84.2% macro-F1.

neutral

train_94195

To which organ system do the esophagus, liver, pancreas, small intestine, and colon belong?

although various neural Qa methods have achieved high performance on some of these datasets Trivedi et al., 2019;Tymoshenko et al., 2017;Seo et al., 2016;Wang and Jiang, 2016;De Cao et al., 2018;Back et al., 2018), we argue that more effort must be dedicated to explaining their inference process.

neutral

train_94196

Unsurprisingly, training and testing in the same domain (e.g., Fiction) leads to the best performance on sentence selection.

in MultiRC, AutoROCC outperforms our baseline of BERT + entire passage (row 10 vs 22) by 8.3% EM0, indicating that AutoROCC can filter out irrelevant content.

neutral

train_94197

Note that some of the works discussed here transfer knowledge from external datasets into the QA task they address (Chung et al., 2017;Pan et al., 2019;Min et al., 2017;Qiu et al., 2018;Chen et al., 2017).

the improvements are not large, e.g., the maximum improvement is 1.6% (when k = 4), which indicates that ROCC is robust to a certain extent to lexical variation.

neutral

train_94198

Open-domain question answering (QA) Inspired by the series of TREC QA competitions, 2 Chen et al.

centered around entity names, this model risks missing purely descriptive clues in the question.

neutral

train_94199

Because off-the-shelf IR systems generally optimize for shallow lexical similarity between query and candidate documents in favor of efficiency, a good proxy for this overlap is locating spans of text that have high lexical overlap with the intended supporting documents.

it is computationally inefficient, and has high variance especially for the second reasoning step and forward, because the context depends greatly on what queries have been chosen previously and their search results.

neutral