tasksource/scinli · Datasets at Hugging Face

id stringlengths 7 12	sentence1 stringlengths 6 1.27k	sentence2 stringlengths 6 926	label stringclasses 4 values
train_4500	manner, where they first conduct entity recognition and then predict relations between extracted entities.	the pipeline framework ignores the relevance of entity identification and relation prediction (Li and Ji, 2014).	contrasting
train_4501	The NovelTagging method finds triplets through tagging the words.	they assume that only one tag could be assigned to just one word.	contrasting
train_4502	When extracting relation from sentences that contains 1 triplets, NovelTagging model achieve the best performance.	when the number of triplets increases, the performance of NovelTagging model decreases significantly.	contrasting
train_4503	Event Type: Transport It is certain that spurious features often result from the semantically pseudo-related context, and during training, a neural network may mistakenly and unconsciously preserve the memory to produce the fakes.	it is difficult to determine which words are pseudo-related in a specific case, and when they will "jump out" to mislead the generation of latent features during testing.	contrasting
train_4504	• The RNN models contribute to achieving a higher recall.	the precision is lower.	contrasting
train_4505	Multi-angled cognition enables Hybrid to be more precise.	it is built using a single-channel architecture, concatenating the RNN and the CNN.	contrasting
train_4506	At each time step, it chooses what to memorize and forget, so patterns over arbitrary time intervals can be recognized.	the memory in LSTM is still short-term.	contrasting
train_4507	For example, sentences are naturally fit units for such models, since a sentence starts only after the preceding sentence is finished, and LSTM may be an adequate tool to process sentences.	when the sequences are not contiguous, as in temporal and other discourse-scale relations, LSTM models do not have the capability to look for input pieces across sequences.	contrasting
train_4508	By connecting the output layer, which has a softmax activation, we hand the final decisions made by the pairwise model to GCL.	the hidden layer provides higher layers with cruder but richer information.	contrasting
train_4509	By doing the rotation, some pairs in the final chunk of epoch 1 will show up in the first chunk in epoch 2 as well.	within each chunk, we do not randomize pairs, so narrative order is preserved at this level.	contrasting
train_4510	Tuning the model for each pair type separately, as well as resampling to deal with class imbalance would, perhaps, improve performance.	the point of these experiments was not to get the largest improvement, but to show that the GCL mechanism can replace heuristic-based timegraph conflict resolution, improving the performance of an otherwise very similar model.	contrasting
train_4511	For instance, the Pointer Networks (Vinyals et al., 2015) uses attention over input timesteps.	it has no power to rewrite information for later use, since they have no "memory" except for the RNN states.	contrasting
train_4512	All previous works converged to a shared assessment: both CNNs and RNNs provide relevant, but different kinds of information for text classification.	though several works have studied linguistic structures inherent in RNNs, to our knowledge, none of them have focused on CNNs.	contrasting
train_4513	As we can see, when the z-test is the highest, the TDS is also the highest and the TDS values are high also for the neighbor words (for example around the word castra).	this is not always the case: for example small words as que or et are also high in z-test but they do not impact the network at the same level.	contrasting
train_4514	We also experimented with Glove (Pennington et al., 2014), which has more vocabulary coverage than word2vec -Glove covers 89.77% of our vocabulary items, whereas word2vec covers 85.66%.	glove did not perform well giving F 1 score of 86% in the standard discrimination task.	contrasting
train_4515	To our knowledge, we are the first to use coherence models for this problem.	our goal in this paper is not to build a state-of-the-art system for thread reconstruction, rather to evaluate coherence models by showing its effectiveness in scoring candidate tree hypotheses.	contrasting
train_4516	However, our goal in this paper is not to build a state-of-the-art system for thread reconstruction, rather to evaluate coherence models by showing its effectiveness in scoring candidate tree hypotheses.	to previous methods, our approach therefore considers the whole thread structure at once, and computes coherence scores for all possible candidate trees of a conversation.	contrasting
train_4517	In English, there are few omissions of arguments, and thus PA is relatively easy, around 83% accuracy , while CR is relatively difficult, around 70% accuracy .	in Japanese and Chinese, where arguments are often omitted, PA is a dif-ficult task, and even state-of-the-art systems only achieve around 50% accuracy.	contrasting
train_4518	Although most of studies did not consider the notion entity, Sasano and Kurohashi (2011) consider an entity, and its salience score is calculated based on simple rules.	they used gold coreference links to form the entities, and reported the salience score did not improve the performance.	contrasting
train_4519	However, they used gold coreference links to form the entities, and reported the salience score did not improve the performance.	we perform CR automatically, and capture the entity salience by using RNNs.	contrasting
train_4520	The performance of ZAR also matters.	the performance of ZAR in our baseline model is extremely low, and thus there are few worsen examples and Table 2: Performance of case analysis and zero anaphora resolution for each case, and each argument position for zero anaphora resolution.	contrasting
train_4521	The above selectional system restricts the parser's search space sufficiently well that it is feasible to generate an initial MG treebank for many of the sentences in the PTB, particularly the shorter ones and those longer ones which do not require the full range of null heads to be allowed into the chart 16 .	for longer sentences requiring null heads such as extraposers, topicalizers or focalizers, parsing remains impractically slow.	contrasting
train_4522	Division by power, i.e., high (SH and LH groups) vs. low (SL and LL groups), does not result in a salient difference in slopes, as it can be seen that the slopes of high (solid) and low (dashed) power lines do not differ much from each other within the same prime utterance length group (indicated by color).	division by prime utterance length, i.e.	contrasting
train_4523	We are not denying the existence of accommodation caused by the social distance between interlocutors.	we want to stress the difference between the priming-induced alignment at lower linguistic levels and the intentional accommodation that is caused by higher-level perception of social power.	contrasting
train_4524	We manually inspected 30 reference answers which were annotated incorrectly and found that of those, about 95% were indeed incorrect.	62% are actually answerable from some paragraph, indicating that the real ceiling performance on this dataset is around 90% and that there is still room for improvement on this task.	contrasting
train_4525	Sentence scoring and sentence selection are two main steps in extractive document summarization systems.	previous works treat them as two separated subtasks.	contrasting
train_4526	On one hand, the RNN is used to remember the partial output summary by feeding the selected sentence into it.	it is used to provide a sentence extraction state that can be used to score sentences with their representations.	contrasting
train_4527	One is used to rank the sentences to select the first sentence, and the other one is used to model the redundancy during sentence selection.	their model of measuring the redundancy only considers the redundancy between the sentence that has the maximal score, which lacks the modeling of all the selection history.	contrasting
train_4528	We think the second step selection benefits from the first step in NEUSUM since it can remember the selection history, while the separated models lack this ability.	we can notice the trend that the precision drops fast after each selection.	contrasting
train_4529	Paths having less than z words or that do not contain a verb are filtered out (z is a tuning parameter).	unlike in Boudin and Morin (2013), we rerank the K best paths with the following novel weighting scheme (the lower, the better), and the path with the lowest score is used as the compression: The denominator takes into account the length of the path, and its fluency (F ), coverage (C), and diversity (D).	contrasting
train_4530	This set of sentences can already be considered to be a summary of the meeting.	it might exceed the maximum size allowed, and still contain some redundancy or off-topic sections unrelated to the general theme of the meeting (e.g., chit-chat).	contrasting
train_4531	Given that our extractor performs a nondifferentiable hard extraction, we apply standard policy gradient methods to bridge the backpropagation and form an end-to-end trainable (stochastic) computation graph.	simply starting from a randomly initialized network to train the whole model in an end-to-end fashion is infeasible.	contrasting
train_4532	(2017) added controlling parameters to adapt the summary to length, style, and entity preferences.	none of these used RL to bridge the non-differentiability of neural models.	contrasting
train_4533	Regression has primarily been attempted through support vector regression (Kim et al., 2006;Zhang and Varadarajan, 2006;Yang et al., 2015).	probabilistic matrix factorization (Tang et al., 2013), linear regression (Lu et al., 2010), and extended tensor factorization models (Moghaddam et al., 2012) have successfully been used to integrate sophisticated constraints into the learning process and have achieved improvements over regular regression models.	contrasting
train_4534	Back in 2012, in the case of "Nagoya", many native English speakers posted their pleasant travel experiences in Nagoya on Twitter.	chinese people overwhelmingly greeted the city with anger and condemnation on Weibo (a chinese version of Twitter), because the city mayor denied the truthfulness of the Nanjing Massacre.	contrasting
train_4535	Due to the dynamic nature of language and trending issues on Twitter, it is impracticable to construct a list of all possible phrases one can expect to appear in tweets.	because politicians are known for sticking to certain talking points, these phrases can be abstracted into higher-level phrases that are more stable and thus easier to identify and extract.	contrasting
train_4536	The difference in performance between the GOLD and MFD results shows that directly mapping the expected MFD unigrams to politicians' tweets is not informative enough for party affiliation prediction.	by using abstract representations of language, the PSL model is able to achieve results closer to that which can be attained when using the actual annotations as features.	contrasting
train_4537	The ablation results indicate that model uncertainty plays the most important role among the confidence metrics.	removing the metrics of data uncertainty affects performance less, because most examples in the datasets are in-domain.	contrasting
train_4538	Recent work has managed to learn crosslingual word embeddings without parallel data by mapping monolingual embeddings to a shared space through adversarial training.	their evaluation has focused on favorable conditions, using comparable corpora or closely-related languages, and we show that they often fail in more realistic scenarios.	contrasting
train_4539	(2017) showed that an iterative self-learning method is able to bootstrap a high quality mapping from very small seed dictionaries (as little as 25 pairs of words).	their analysis reveals that the self-learning method gets stuck in poor local optima when the initial solution is not good enough, thus failing for smaller training dictionaries.	contrasting
train_4540	(2017b) adopt the earth mover's distance for training, optimized through a Wasserstein generative adversarial network followed by an alternating optimization procedure.	all this previous work used comparable Wikipedia corpora in most experiments and, as shown in Section 5, face difficulties in more challenging settings.	contrasting
train_4541	Previous studies (Collobert and Weston, 2008;Dong et al., 2015;Luong et al., 2016;Liu et al., 2018;Yang et al., 2017) have proven that MTL is an effective approach to boost the performance of related tasks such as MT and parsing.	most of these previous efforts focused on tasks and languages which have sufficient labeled data but hit a performance ceiling on each task alone.	contrasting
train_4542	In (Yang et al., 2017), the authors simulated a low-resource setting for English and Spanish by downsampling the training data for the target task.	for most low-resource languages, the data sparsity problem also lies in related tasks and languages.	contrasting
train_4543	The bidirectional LSTM layer is essential to extract character, word, and contextual information from a sentence.	with a large number of parameters, it cannot be fully trained only using the low-resource task data.	contrasting
train_4544	As Figure 7 shows, the performance goes up when we raise the sample rate from 1% to 20%.	we do not observe significant improvement when we further increase the sample rate.	contrasting
train_4545	By using BWEs based on only Subtitles, we lose too many embeddings of similar English and Spanish tweets.	if we use only tweet-based BWEs we lose good quality semantic knowledge which can be learned from more standard text domains.	contrasting
train_4546	Cloze-style reading comprehension is a task setting where the question is formed by replacing a token in a sentence of the story with a placeholder (left part of Figure 1).	to many previous complex models Dhingra et al., 2017;Cui et al., 2017;Munkhdalai and Yu, 2016; that perform multi-turn reading of a story and a question before inferring the correct answer, we aim to tackle the cloze-style RC task in a way that resembles how humans solve it: using, in addition, background knowledge.	contrasting
train_4547	(2017), by contrast, explicitly reads the acquired additional knowledge sequentially after reading the document and question, but transfers the background knowledge implicitly, by refining the word embeddings of the words in the document and the question with the words from the supporting knowledge that share the same lemma.	to the implicit knowledge transfer of Weissenborn et al.	contrasting
train_4548	In summary, previous work on text adversaries change semantics, only generate local (instancespecific) adversaries (Zhao et al., 2018;Iyyer et al., 2018), or are tailored for white-box models (Ebrahimi et al., 2018) or specific tasks (Jia and Liang, 2017).	sEAs expose oversensitivity for specific predictions of black-box models for a variety of tasks, while sEARs are intuitive and actionable global rules that induce a high number of high-quality adversaries.	contrasting
train_4549	We reproduced the results reported in (Hu et al., 2017) using their tasks and data.	the same model trained on our political slant datasets (described in §3), obtained an almost random accuracy of 50.98% in style transfer.	contrasting
train_4550	Unlike DBpedia and Freebase, Wikidata usually contains a very concise description for many of its entities.	because Wikidata is based on user contributions, many new entries are created that still lack such descriptions.	contrasting
train_4551	In this regard, such a model thus resembles our approach.	there are important differences between this baseline and our model.	contrasting
train_4552	However, the major shortcoming of these sorts of methods, including those aiming at more fine-grained typing, is that they assume that the set of candidate types is given as input, and the main remaining challenge is to pick the correct one(s).	our work yields descriptions that often indicate the type of entity, but typically are more natural-sounding and descriptive (e.g.	contrasting
train_4553	(2016) take Wikipedia infobox data as input and train a custom form of neural language model that, conditioned on occurrences of words in the input table, generates biographical sentences as output.	their system is limited to a single kind of description (biographical sentences) that tend to share a common structure.	contrasting
train_4554	They investigated an extensive array of current state-of-the-art neural pointer methods but found that template-based models outperform all neural models on this task by a significant margin.	their method requires specific templates for each domain (for example, basketball games in their case).	contrasting
train_4555	They find this technique can generate text of the desired genre, but the movie plots are not interpretable (as the model outputs events, not raw text).	we are not aware of previous work that has used hierarchical generation from a textual premise to improve the coherence and structure of stories.	contrasting
train_4556	When analyzing 500 150-word generated stories from test-set prompts, the average longest common subsequence is 8.9.	the baseline Conv seq2seq model copies 10.2 words on average and the KNN baseline copies all 150 words from a story in the training set.	contrasting
train_4557	The second seq2seq model learns to focus on rare words, such as horned and robe.	the fusion model has limitations.	contrasting
train_4558	Using automatic metrics can ensure rapid prototyping and testing new models with fewer expensive human evaluation.	they have been criticized to be biased and correlate poorly with human judgments, especially in many generative tasks like response generation (Lowe et al., 2017;Liu et al., 2016), dialogue system (Bruni and Fernández, 2017) and machine translation (Callison-Burch et al., 2006).	contrasting
train_4559	In order to approximate the Reward Boltzmann distribution towards the "real" data distribution p * (W ), we design a min-max two-player game, where the Reward Boltzmann distribution p θ aims at maximizing the its similarity with empirical distribution p e while minimizing that with the "faked" data generated from policy model π β .	the policy distribution π β tries to maximize its similarity with the Boltzmann distribution p θ .	contrasting
train_4560	However, the references to the same image sequence are photostream different from each other, so the score is very low and not suitable for this task.	our AREL framework can lean a more robust reward function from human-annotated stories, which is able to provide better guidance to the policy and thus improves its performances over different metrics.	contrasting
train_4561	Glove+Picturebook improves over the Glove baseline for image search but falls short on image annotation.	using contextual gating results in improvements over the baseline on all metrics except R@1 for image annotation.	contrasting
train_4562	Previous works (She et al., 2014;Misra et al., 2015;Chai, 2016, 2017) explicitly model verb semantics as desired goal states and thus linking natural language commands with underlying planning systems for action planning and execution.	these studies were carried out either in a simulated world or in a carefully curated simple environment within the limitation of the robot's manipulation system.	contrasting
train_4563	Indeed, we can also observe that another CNN-based baseline, i.e., CNN-ASP implemented by us, also obtains good results on TWITTER.	the performance of those comparison methods is mostly unstable.	contrasting
train_4564	This also led to existing and potential research in improving attention modeling (discussed in Section 5).	we observed that simply focusing on tackling the target-context detection problem and learning better attention are not sufficient to solve the problem found in sentences (2), (3) and (4).	contrasting
train_4565	First, an aspect word often expresses no sentiment, for example, "screen".	if the aspect term v t is simply removed from Eq.	contrasting
train_4566	This technique is also intuitive in neural networks.	notice that by using the non-linear projection (or adding more sophisticated hidden layers) over them in this way, we sacrifice some interpretability.	contrasting
train_4567	For some simple examples like "the battery is good", the context word "good" simply indicates clear sentiment, which can be captured by their first-order term.	notice that the modeling of second-order terms offers additional help in both general and target-sensitive scenarios.	contrasting
train_4568	In fact, solely improving attention does not solve our problem (see Sections 1 and 3).	better attention can certainly help achieve an overall better performance for the ASC task, as it makes the targeted-context detection more accurate.	contrasting
train_4569	The cause of this unavoidable error is that W c i is not conditioned on the target.	w J d i , •d t tanh(w 2 c i ) can change the sentiment polarity with the aspect vector d t encoded.	contrasting
train_4570	Most of the works have focused on learning a shared low dimensional representation of features that can be generalized across different domains.	none of the approaches explicitly analyses significance and polarity of words across domains.	contrasting
train_4571	In other words, addition of the iterative process with the shared representation given by SCL overcomes the errors introduced by SCL.	sCP given by our approach were able to produce a less erroneous system in oneshot.	contrasting
train_4572	Table 5 shows that K→E outperforms B→E and D→E, and E→K outperforms B→K and D→K.	dVd (d) and electronics are two very different domains unlike electronics and Kitchen, or dVd and books.	contrasting
train_4573	We argue that content preservation is also an indispensable evaluation metric.	when applied to the sentiment-to-sentiment translation task, the previously mentioned models share the same problem.	contrasting
train_4574	is computed based on the percentage of overlapping n-grams between the generated text and the reference text.	the overlapping n-grams contain not only content words but also function words, bringing the noisy results.	contrasting
train_4575	proposed to use discourse markers to help rep-resent the meanings of the sentences.	they represent each sentence by a single vector and directly concatenate them to predict the answer, which is too simple and not ideal for the largescale datasets.	contrasting
train_4576	We observe that the values are highly correlated among the synonyms like "people" with "man", "three" with "3" in both situations.	words that might have contradictory meanings like "hoods" with "bareheaded", "quiet" with "busy" perform worse without the discourse markers augmentation, which conforms to the conclusion that the "contradiction" label examples benefit a lot which is observed in the Section 5.5.	contrasting
train_4577	They also apply their pre-trained sentence encoder to a series of natural language understanding tasks such as sentiment analysis, question-type, entailment, and relatedness.	all those datasets are provided by Conneau et al.	contrasting
train_4578	However, they manually group the discourse markers into several categories based on human knowledge and predict the category instead of the explicit discourse marker phrase.	the size of their dataset is much smaller than that in , and sometimes there has been disagreement among annotators about what exactly is the correct categorization of discourse relations (Hobbs, 1990).	contrasting
train_4579	While this allows some multi-step inferential process, these networks lack a more complex reasoning mechanism, needed for more elaborated tasks such as inferring relations among entities (relational reasoning).	relation Networks (rNs), proposed in Santoro et al.	contrasting
train_4580	A possibility is that each attention head is being adapted for different groups of related tasks.	we did not investigate this further.	contrasting
train_4581	The model shares some similarities with the Memory Network model.	unlike the MemNN model, it operates in the input sequentially (as in the NTM model).	contrasting
train_4582	That is because the episodic buffer has an integration function that our model does not cover.	that can be an interesting source of inspiration for next versions of the model that integrate both visual and textual information for question answering.	contrasting
train_4583	Then, we can proceed using the same architecture for the reasoning and attention module that the one used in the textual QA model.	for the visual QA task, we used an additive attention mechanism.	contrasting
train_4584	On short text, we also found that attention (or the gated pooling mechanism from GRNN) did not really help make any significant improvements over the vanilla LSTM model and a qualitative explanation to why this is so is deferred to the next section.	attention helps for long text (such as debates), resulting in Attention LSTMs becoming the strongest baseline on the Debates datasets.	contrasting
train_4585	This is in concert with our intuition about modeling contrast and incongruity.	both ATT-LSTM and ATT-RAW learn very different attention maps.	contrasting
train_4586	This time, the attention maps of MIARN are not as distinct as before.	they focus on sentiment-bearing words, composing the words 'ignored sucks' to form the majority of the intraattentive representation.	contrasting
train_4587	This behavior of gracefully falling back to NCE is more desirable than the alternative of stalled training if p − (y\|x) does not have a simple p nce mixture component.	we would still like to avoid such collapse, as the adversarial samples provide greater learning signals than NCE samples.	contrasting
train_4588	As can be seen from our results in Table 2, ACE on RW is not always better and for the 100d and 300d Glove embeddings is marginally worse.	on WordSim353 ACE does considerably better across the board to the point where 50d Glove embeddings outperform the 300d baseline Glove model.	contrasting
train_4589	Divergence between F-Measure and Cross-Entropy Loss.	detection tasks are mostly evaluated using F-measure computed on positive classes, which makes it unsuitable to optimize classifiers using cross-entropy loss.	contrasting
train_4590	We can easily see that for accuracy metric, correct predictions of positive and negative instances are equally regarded (i.e., T P and T N are symmetric), which is consistent with crossentropy loss function.	when measuring using F-measure, this condition is no longer holding.	contrasting
train_4591	2) Static scaling sets the importance of negative instances statically in the entire training procedure.	as shown in Section 3, the rel- ative importance between different classes is dynamically changed during the training procedure, which makes static scaling incapable of achieving stable performance in different phases of training.	contrasting
train_4592	More formally, let us illustrate the model by taking the sequence prediction task (Figure 1) as illustration.	given an utterance with labels y 1 , .., y n , our Multi-task Tri-training loss consists of three task-specific (m 1 , m 2 , m 3 ) tagging loss functions (where h is the uppermost Bi-LSTM encoding): to classic tri-training, we can train the multi-task model with its three model-specific outputs jointly and without bootstrap sampling on the labeled source domain data until convergence, as the orthogonality constraint enforces different representations between models m 1 and m 2 .	contrasting
train_4593	Note that we constrain the size of S t not to exceed \|S * \| (the 7th line in Algorithm 1) to avoid that too many fluency boost pairs overwhelm the effects of the original error-corrected pairs on model learning.	to back-boost learning whose core idea is originally from NMT, self-boost learning is original, which is specially devised for neural GEC.	contrasting
train_4594	Take our best system (the last row in Table 2) as an example, among 1,312 sentences in the CoNLL-2014 dataset, seq2seq inference with shallow fusion LM edits 566 sentences.	fluency boost inference additionally edits 23 sentences during the second round inference, improving F 0.5 from 52.59 to 52.72.	contrasting
train_4595	Most of advanced GEC systems are classifierbased (Chodorow et al., 2007;De Felice and Pulman, 2008;Han et al., 2010;Leacock et al., 2010;Tetreault et al., 2010a;Dale and Kilgarriff, 2011) 7 The state-of-the-art result on CoNLL-2014 dataset has been recently advanced by Chollampatt and Ng (2018) (F0.5=54.79) and Grundkiewicz and Junczys-Dowmunt (2018) (F0.5=56.25), which are contemporaneous to this paper.	to the basic seq2seq model in this paper, they used advanced approaches for modeling (e.g., convolutional seq2seq with pre-trained word embedding, using edit operation features, ensemble decoding and advanced model combinations).	contrasting
train_4596	The success of deep neural networks (DNNs) is heavily dependent on the availability of labeled data.	obtaining labeled data is a big challenge in many real-world problems.	contrasting
train_4597	We initialize the embedding matrix E in our network with word embeddings that are pretrained on a large crisis dataset (Subsection 2.5).	embedding matrix E can also be initialize randomly.	contrasting
train_4598	Typically graphs are constructed based on a relational knowledge source, e.g., citation links in (Lu and Getoor, 2003), or distance between instances (Zhu, 2005).	we do not have access to such a relational knowledge in our setting.	contrasting
train_4599	On one hand, this is actually determined by real distribution of ratings for a particular prompt, e.g., how many essays are with an extreme quality for a given prompt in the target data.	a fine-grained tuning of the RankSVM (e.g., tuning C + and C − for positive and negative exam-ples separately) may partially resolve the problem, which is left for the future work.	contrasting

train_4500

manner, where they first conduct entity recognition and then predict relations between extracted entities.

the pipeline framework ignores the relevance of entity identification and relation prediction (Li and Ji, 2014).

contrasting

train_4501

The NovelTagging method finds triplets through tagging the words.

they assume that only one tag could be assigned to just one word.

contrasting

train_4502

When extracting relation from sentences that contains 1 triplets, NovelTagging model achieve the best performance.

when the number of triplets increases, the performance of NovelTagging model decreases significantly.

contrasting

train_4503

Event Type: Transport It is certain that spurious features often result from the semantically pseudo-related context, and during training, a neural network may mistakenly and unconsciously preserve the memory to produce the fakes.

it is difficult to determine which words are pseudo-related in a specific case, and when they will "jump out" to mislead the generation of latent features during testing.

contrasting

train_4504

• The RNN models contribute to achieving a higher recall.

the precision is lower.

contrasting

train_4505

Multi-angled cognition enables Hybrid to be more precise.

it is built using a single-channel architecture, concatenating the RNN and the CNN.

contrasting

train_4506

At each time step, it chooses what to memorize and forget, so patterns over arbitrary time intervals can be recognized.

the memory in LSTM is still short-term.

contrasting

train_4507

For example, sentences are naturally fit units for such models, since a sentence starts only after the preceding sentence is finished, and LSTM may be an adequate tool to process sentences.

when the sequences are not contiguous, as in temporal and other discourse-scale relations, LSTM models do not have the capability to look for input pieces across sequences.

contrasting

train_4508

By connecting the output layer, which has a softmax activation, we hand the final decisions made by the pairwise model to GCL.

the hidden layer provides higher layers with cruder but richer information.

contrasting

train_4509

By doing the rotation, some pairs in the final chunk of epoch 1 will show up in the first chunk in epoch 2 as well.

within each chunk, we do not randomize pairs, so narrative order is preserved at this level.

contrasting

train_4510

Tuning the model for each pair type separately, as well as resampling to deal with class imbalance would, perhaps, improve performance.

the point of these experiments was not to get the largest improvement, but to show that the GCL mechanism can replace heuristic-based timegraph conflict resolution, improving the performance of an otherwise very similar model.

contrasting

train_4511

For instance, the Pointer Networks (Vinyals et al., 2015) uses attention over input timesteps.

it has no power to rewrite information for later use, since they have no "memory" except for the RNN states.

contrasting

train_4512

All previous works converged to a shared assessment: both CNNs and RNNs provide relevant, but different kinds of information for text classification.

though several works have studied linguistic structures inherent in RNNs, to our knowledge, none of them have focused on CNNs.

contrasting

train_4513

As we can see, when the z-test is the highest, the TDS is also the highest and the TDS values are high also for the neighbor words (for example around the word castra).

this is not always the case: for example small words as que or et are also high in z-test but they do not impact the network at the same level.

contrasting

train_4514

We also experimented with Glove (Pennington et al., 2014), which has more vocabulary coverage than word2vec -Glove covers 89.77% of our vocabulary items, whereas word2vec covers 85.66%.

glove did not perform well giving F 1 score of 86% in the standard discrimination task.

contrasting

train_4515

To our knowledge, we are the first to use coherence models for this problem.

our goal in this paper is not to build a state-of-the-art system for thread reconstruction, rather to evaluate coherence models by showing its effectiveness in scoring candidate tree hypotheses.

contrasting

train_4516

However, our goal in this paper is not to build a state-of-the-art system for thread reconstruction, rather to evaluate coherence models by showing its effectiveness in scoring candidate tree hypotheses.

to previous methods, our approach therefore considers the whole thread structure at once, and computes coherence scores for all possible candidate trees of a conversation.

contrasting

train_4517

In English, there are few omissions of arguments, and thus PA is relatively easy, around 83% accuracy , while CR is relatively difficult, around 70% accuracy .

in Japanese and Chinese, where arguments are often omitted, PA is a dif-ficult task, and even state-of-the-art systems only achieve around 50% accuracy.

contrasting

train_4518

Although most of studies did not consider the notion entity, Sasano and Kurohashi (2011) consider an entity, and its salience score is calculated based on simple rules.

they used gold coreference links to form the entities, and reported the salience score did not improve the performance.

contrasting

train_4519

However, they used gold coreference links to form the entities, and reported the salience score did not improve the performance.

we perform CR automatically, and capture the entity salience by using RNNs.

contrasting

train_4520

The performance of ZAR also matters.

the performance of ZAR in our baseline model is extremely low, and thus there are few worsen examples and Table 2: Performance of case analysis and zero anaphora resolution for each case, and each argument position for zero anaphora resolution.

contrasting

train_4521

The above selectional system restricts the parser's search space sufficiently well that it is feasible to generate an initial MG treebank for many of the sentences in the PTB, particularly the shorter ones and those longer ones which do not require the full range of null heads to be allowed into the chart 16 .

for longer sentences requiring null heads such as extraposers, topicalizers or focalizers, parsing remains impractically slow.

contrasting

train_4522

Division by power, i.e., high (SH and LH groups) vs. low (SL and LL groups), does not result in a salient difference in slopes, as it can be seen that the slopes of high (solid) and low (dashed) power lines do not differ much from each other within the same prime utterance length group (indicated by color).

division by prime utterance length, i.e.

contrasting

train_4523

We are not denying the existence of accommodation caused by the social distance between interlocutors.

we want to stress the difference between the priming-induced alignment at lower linguistic levels and the intentional accommodation that is caused by higher-level perception of social power.

contrasting

train_4524

We manually inspected 30 reference answers which were annotated incorrectly and found that of those, about 95% were indeed incorrect.

62% are actually answerable from some paragraph, indicating that the real ceiling performance on this dataset is around 90% and that there is still room for improvement on this task.

contrasting

train_4525

Sentence scoring and sentence selection are two main steps in extractive document summarization systems.

previous works treat them as two separated subtasks.

contrasting

train_4526

On one hand, the RNN is used to remember the partial output summary by feeding the selected sentence into it.

it is used to provide a sentence extraction state that can be used to score sentences with their representations.

contrasting

train_4527

One is used to rank the sentences to select the first sentence, and the other one is used to model the redundancy during sentence selection.

their model of measuring the redundancy only considers the redundancy between the sentence that has the maximal score, which lacks the modeling of all the selection history.

contrasting

train_4528

We think the second step selection benefits from the first step in NEUSUM since it can remember the selection history, while the separated models lack this ability.

we can notice the trend that the precision drops fast after each selection.

contrasting

train_4529

Paths having less than z words or that do not contain a verb are filtered out (z is a tuning parameter).

unlike in Boudin and Morin (2013), we rerank the K best paths with the following novel weighting scheme (the lower, the better), and the path with the lowest score is used as the compression: The denominator takes into account the length of the path, and its fluency (F ), coverage (C), and diversity (D).

contrasting

train_4530

This set of sentences can already be considered to be a summary of the meeting.

it might exceed the maximum size allowed, and still contain some redundancy or off-topic sections unrelated to the general theme of the meeting (e.g., chit-chat).

contrasting

train_4531

Given that our extractor performs a nondifferentiable hard extraction, we apply standard policy gradient methods to bridge the backpropagation and form an end-to-end trainable (stochastic) computation graph.

simply starting from a randomly initialized network to train the whole model in an end-to-end fashion is infeasible.

contrasting

train_4532

(2017) added controlling parameters to adapt the summary to length, style, and entity preferences.

none of these used RL to bridge the non-differentiability of neural models.

contrasting

train_4533

Regression has primarily been attempted through support vector regression (Kim et al., 2006;Zhang and Varadarajan, 2006;Yang et al., 2015).

probabilistic matrix factorization (Tang et al., 2013), linear regression (Lu et al., 2010), and extended tensor factorization models (Moghaddam et al., 2012) have successfully been used to integrate sophisticated constraints into the learning process and have achieved improvements over regular regression models.

contrasting

train_4534

Back in 2012, in the case of "Nagoya", many native English speakers posted their pleasant travel experiences in Nagoya on Twitter.

chinese people overwhelmingly greeted the city with anger and condemnation on Weibo (a chinese version of Twitter), because the city mayor denied the truthfulness of the Nanjing Massacre.

contrasting

train_4535

Due to the dynamic nature of language and trending issues on Twitter, it is impracticable to construct a list of all possible phrases one can expect to appear in tweets.

because politicians are known for sticking to certain talking points, these phrases can be abstracted into higher-level phrases that are more stable and thus easier to identify and extract.

contrasting

train_4536

The difference in performance between the GOLD and MFD results shows that directly mapping the expected MFD unigrams to politicians' tweets is not informative enough for party affiliation prediction.

by using abstract representations of language, the PSL model is able to achieve results closer to that which can be attained when using the actual annotations as features.

contrasting

train_4537

The ablation results indicate that model uncertainty plays the most important role among the confidence metrics.

removing the metrics of data uncertainty affects performance less, because most examples in the datasets are in-domain.

contrasting

train_4538

Recent work has managed to learn crosslingual word embeddings without parallel data by mapping monolingual embeddings to a shared space through adversarial training.

their evaluation has focused on favorable conditions, using comparable corpora or closely-related languages, and we show that they often fail in more realistic scenarios.

contrasting

train_4539

(2017) showed that an iterative self-learning method is able to bootstrap a high quality mapping from very small seed dictionaries (as little as 25 pairs of words).

their analysis reveals that the self-learning method gets stuck in poor local optima when the initial solution is not good enough, thus failing for smaller training dictionaries.

contrasting

train_4540

(2017b) adopt the earth mover's distance for training, optimized through a Wasserstein generative adversarial network followed by an alternating optimization procedure.

all this previous work used comparable Wikipedia corpora in most experiments and, as shown in Section 5, face difficulties in more challenging settings.

contrasting

train_4541

Previous studies (Collobert and Weston, 2008;Dong et al., 2015;Luong et al., 2016;Liu et al., 2018;Yang et al., 2017) have proven that MTL is an effective approach to boost the performance of related tasks such as MT and parsing.

most of these previous efforts focused on tasks and languages which have sufficient labeled data but hit a performance ceiling on each task alone.

contrasting

train_4542

In (Yang et al., 2017), the authors simulated a low-resource setting for English and Spanish by downsampling the training data for the target task.

for most low-resource languages, the data sparsity problem also lies in related tasks and languages.

contrasting

train_4543

The bidirectional LSTM layer is essential to extract character, word, and contextual information from a sentence.

with a large number of parameters, it cannot be fully trained only using the low-resource task data.

contrasting

train_4544

As Figure 7 shows, the performance goes up when we raise the sample rate from 1% to 20%.

we do not observe significant improvement when we further increase the sample rate.

contrasting

train_4545

By using BWEs based on only Subtitles, we lose too many embeddings of similar English and Spanish tweets.

if we use only tweet-based BWEs we lose good quality semantic knowledge which can be learned from more standard text domains.

contrasting

train_4546

Cloze-style reading comprehension is a task setting where the question is formed by replacing a token in a sentence of the story with a placeholder (left part of Figure 1).

to many previous complex models Dhingra et al., 2017;Cui et al., 2017;Munkhdalai and Yu, 2016; that perform multi-turn reading of a story and a question before inferring the correct answer, we aim to tackle the cloze-style RC task in a way that resembles how humans solve it: using, in addition, background knowledge.

contrasting

train_4547

(2017), by contrast, explicitly reads the acquired additional knowledge sequentially after reading the document and question, but transfers the background knowledge implicitly, by refining the word embeddings of the words in the document and the question with the words from the supporting knowledge that share the same lemma.

to the implicit knowledge transfer of Weissenborn et al.

contrasting

train_4548

In summary, previous work on text adversaries change semantics, only generate local (instancespecific) adversaries (Zhao et al., 2018;Iyyer et al., 2018), or are tailored for white-box models (Ebrahimi et al., 2018) or specific tasks (Jia and Liang, 2017).

sEAs expose oversensitivity for specific predictions of black-box models for a variety of tasks, while sEARs are intuitive and actionable global rules that induce a high number of high-quality adversaries.

contrasting

train_4549

We reproduced the results reported in (Hu et al., 2017) using their tasks and data.

the same model trained on our political slant datasets (described in §3), obtained an almost random accuracy of 50.98% in style transfer.

contrasting

train_4550

Unlike DBpedia and Freebase, Wikidata usually contains a very concise description for many of its entities.

because Wikidata is based on user contributions, many new entries are created that still lack such descriptions.

contrasting

train_4551

In this regard, such a model thus resembles our approach.

there are important differences between this baseline and our model.

contrasting

train_4552

However, the major shortcoming of these sorts of methods, including those aiming at more fine-grained typing, is that they assume that the set of candidate types is given as input, and the main remaining challenge is to pick the correct one(s).

our work yields descriptions that often indicate the type of entity, but typically are more natural-sounding and descriptive (e.g.

contrasting

train_4553

(2016) take Wikipedia infobox data as input and train a custom form of neural language model that, conditioned on occurrences of words in the input table, generates biographical sentences as output.

their system is limited to a single kind of description (biographical sentences) that tend to share a common structure.

contrasting

train_4554

They investigated an extensive array of current state-of-the-art neural pointer methods but found that template-based models outperform all neural models on this task by a significant margin.

their method requires specific templates for each domain (for example, basketball games in their case).

contrasting

train_4555

They find this technique can generate text of the desired genre, but the movie plots are not interpretable (as the model outputs events, not raw text).

we are not aware of previous work that has used hierarchical generation from a textual premise to improve the coherence and structure of stories.

contrasting

train_4556

When analyzing 500 150-word generated stories from test-set prompts, the average longest common subsequence is 8.9.

the baseline Conv seq2seq model copies 10.2 words on average and the KNN baseline copies all 150 words from a story in the training set.

contrasting

train_4557

The second seq2seq model learns to focus on rare words, such as horned and robe.

the fusion model has limitations.

contrasting

train_4558

Using automatic metrics can ensure rapid prototyping and testing new models with fewer expensive human evaluation.

they have been criticized to be biased and correlate poorly with human judgments, especially in many generative tasks like response generation (Lowe et al., 2017;Liu et al., 2016), dialogue system (Bruni and Fernández, 2017) and machine translation (Callison-Burch et al., 2006).

contrasting

train_4559

In order to approximate the Reward Boltzmann distribution towards the "real" data distribution p * (W ), we design a min-max two-player game, where the Reward Boltzmann distribution p θ aims at maximizing the its similarity with empirical distribution p e while minimizing that with the "faked" data generated from policy model π β .

the policy distribution π β tries to maximize its similarity with the Boltzmann distribution p θ .

contrasting

train_4560

However, the references to the same image sequence are photostream different from each other, so the score is very low and not suitable for this task.

our AREL framework can lean a more robust reward function from human-annotated stories, which is able to provide better guidance to the policy and thus improves its performances over different metrics.

contrasting

train_4561

Glove+Picturebook improves over the Glove baseline for image search but falls short on image annotation.

using contextual gating results in improvements over the baseline on all metrics except R@1 for image annotation.

contrasting

train_4562

Previous works (She et al., 2014;Misra et al., 2015;Chai, 2016, 2017) explicitly model verb semantics as desired goal states and thus linking natural language commands with underlying planning systems for action planning and execution.

these studies were carried out either in a simulated world or in a carefully curated simple environment within the limitation of the robot's manipulation system.

contrasting

train_4563

Indeed, we can also observe that another CNN-based baseline, i.e., CNN-ASP implemented by us, also obtains good results on TWITTER.

the performance of those comparison methods is mostly unstable.

contrasting

train_4564

This also led to existing and potential research in improving attention modeling (discussed in Section 5).

we observed that simply focusing on tackling the target-context detection problem and learning better attention are not sufficient to solve the problem found in sentences (2), (3) and (4).

contrasting

train_4565

First, an aspect word often expresses no sentiment, for example, "screen".

if the aspect term v t is simply removed from Eq.

contrasting

train_4566

This technique is also intuitive in neural networks.

notice that by using the non-linear projection (or adding more sophisticated hidden layers) over them in this way, we sacrifice some interpretability.

contrasting

train_4567

For some simple examples like "the battery is good", the context word "good" simply indicates clear sentiment, which can be captured by their first-order term.

notice that the modeling of second-order terms offers additional help in both general and target-sensitive scenarios.

contrasting

train_4568

In fact, solely improving attention does not solve our problem (see Sections 1 and 3).

better attention can certainly help achieve an overall better performance for the ASC task, as it makes the targeted-context detection more accurate.

contrasting

train_4569

The cause of this unavoidable error is that W c i is not conditioned on the target.

w J d i , •d t tanh(w 2 c i ) can change the sentiment polarity with the aspect vector d t encoded.

contrasting

train_4570

Most of the works have focused on learning a shared low dimensional representation of features that can be generalized across different domains.

none of the approaches explicitly analyses significance and polarity of words across domains.

contrasting

train_4571

In other words, addition of the iterative process with the shared representation given by SCL overcomes the errors introduced by SCL.

sCP given by our approach were able to produce a less erroneous system in oneshot.

contrasting

train_4572

Table 5 shows that K→E outperforms B→E and D→E, and E→K outperforms B→K and D→K.

dVd (d) and electronics are two very different domains unlike electronics and Kitchen, or dVd and books.

contrasting

train_4573

We argue that content preservation is also an indispensable evaluation metric.

when applied to the sentiment-to-sentiment translation task, the previously mentioned models share the same problem.

contrasting

train_4574

is computed based on the percentage of overlapping n-grams between the generated text and the reference text.

the overlapping n-grams contain not only content words but also function words, bringing the noisy results.

contrasting

train_4575

proposed to use discourse markers to help rep-resent the meanings of the sentences.

they represent each sentence by a single vector and directly concatenate them to predict the answer, which is too simple and not ideal for the largescale datasets.

contrasting

train_4576

We observe that the values are highly correlated among the synonyms like "people" with "man", "three" with "3" in both situations.

words that might have contradictory meanings like "hoods" with "bareheaded", "quiet" with "busy" perform worse without the discourse markers augmentation, which conforms to the conclusion that the "contradiction" label examples benefit a lot which is observed in the Section 5.5.

contrasting

train_4577

They also apply their pre-trained sentence encoder to a series of natural language understanding tasks such as sentiment analysis, question-type, entailment, and relatedness.

all those datasets are provided by Conneau et al.

contrasting

train_4578

However, they manually group the discourse markers into several categories based on human knowledge and predict the category instead of the explicit discourse marker phrase.

the size of their dataset is much smaller than that in , and sometimes there has been disagreement among annotators about what exactly is the correct categorization of discourse relations (Hobbs, 1990).

contrasting

train_4579

While this allows some multi-step inferential process, these networks lack a more complex reasoning mechanism, needed for more elaborated tasks such as inferring relations among entities (relational reasoning).

relation Networks (rNs), proposed in Santoro et al.

contrasting

train_4580

A possibility is that each attention head is being adapted for different groups of related tasks.

we did not investigate this further.

contrasting

train_4581

The model shares some similarities with the Memory Network model.

unlike the MemNN model, it operates in the input sequentially (as in the NTM model).

contrasting

train_4582

That is because the episodic buffer has an integration function that our model does not cover.

that can be an interesting source of inspiration for next versions of the model that integrate both visual and textual information for question answering.

contrasting

train_4583

Then, we can proceed using the same architecture for the reasoning and attention module that the one used in the textual QA model.

for the visual QA task, we used an additive attention mechanism.

contrasting

train_4584

On short text, we also found that attention (or the gated pooling mechanism from GRNN) did not really help make any significant improvements over the vanilla LSTM model and a qualitative explanation to why this is so is deferred to the next section.

attention helps for long text (such as debates), resulting in Attention LSTMs becoming the strongest baseline on the Debates datasets.

contrasting

train_4585

This is in concert with our intuition about modeling contrast and incongruity.

both ATT-LSTM and ATT-RAW learn very different attention maps.

contrasting

train_4586

This time, the attention maps of MIARN are not as distinct as before.

they focus on sentiment-bearing words, composing the words 'ignored sucks' to form the majority of the intraattentive representation.

contrasting

train_4587

This behavior of gracefully falling back to NCE is more desirable than the alternative of stalled training if p − (y|x) does not have a simple p nce mixture component.

we would still like to avoid such collapse, as the adversarial samples provide greater learning signals than NCE samples.

contrasting

train_4588

As can be seen from our results in Table 2, ACE on RW is not always better and for the 100d and 300d Glove embeddings is marginally worse.

on WordSim353 ACE does considerably better across the board to the point where 50d Glove embeddings outperform the 300d baseline Glove model.

contrasting

train_4589

Divergence between F-Measure and Cross-Entropy Loss.

detection tasks are mostly evaluated using F-measure computed on positive classes, which makes it unsuitable to optimize classifiers using cross-entropy loss.

contrasting

train_4590

We can easily see that for accuracy metric, correct predictions of positive and negative instances are equally regarded (i.e., T P and T N are symmetric), which is consistent with crossentropy loss function.

when measuring using F-measure, this condition is no longer holding.

contrasting

train_4591

2) Static scaling sets the importance of negative instances statically in the entire training procedure.

as shown in Section 3, the rel- ative importance between different classes is dynamically changed during the training procedure, which makes static scaling incapable of achieving stable performance in different phases of training.

contrasting

train_4592

More formally, let us illustrate the model by taking the sequence prediction task (Figure 1) as illustration.

given an utterance with labels y 1 , .., y n , our Multi-task Tri-training loss consists of three task-specific (m 1 , m 2 , m 3 ) tagging loss functions (where h is the uppermost Bi-LSTM encoding): to classic tri-training, we can train the multi-task model with its three model-specific outputs jointly and without bootstrap sampling on the labeled source domain data until convergence, as the orthogonality constraint enforces different representations between models m 1 and m 2 .

contrasting

train_4593

Note that we constrain the size of S t not to exceed |S * | (the 7th line in Algorithm 1) to avoid that too many fluency boost pairs overwhelm the effects of the original error-corrected pairs on model learning.

to back-boost learning whose core idea is originally from NMT, self-boost learning is original, which is specially devised for neural GEC.

contrasting

train_4594

Take our best system (the last row in Table 2) as an example, among 1,312 sentences in the CoNLL-2014 dataset, seq2seq inference with shallow fusion LM edits 566 sentences.

fluency boost inference additionally edits 23 sentences during the second round inference, improving F 0.5 from 52.59 to 52.72.

contrasting

train_4595

Most of advanced GEC systems are classifierbased (Chodorow et al., 2007;De Felice and Pulman, 2008;Han et al., 2010;Leacock et al., 2010;Tetreault et al., 2010a;Dale and Kilgarriff, 2011) 7 The state-of-the-art result on CoNLL-2014 dataset has been recently advanced by Chollampatt and Ng (2018) (F0.5=54.79) and Grundkiewicz and Junczys-Dowmunt (2018) (F0.5=56.25), which are contemporaneous to this paper.

to the basic seq2seq model in this paper, they used advanced approaches for modeling (e.g., convolutional seq2seq with pre-trained word embedding, using edit operation features, ensemble decoding and advanced model combinations).

contrasting

train_4596

The success of deep neural networks (DNNs) is heavily dependent on the availability of labeled data.

obtaining labeled data is a big challenge in many real-world problems.

contrasting

train_4597

We initialize the embedding matrix E in our network with word embeddings that are pretrained on a large crisis dataset (Subsection 2.5).

embedding matrix E can also be initialize randomly.

contrasting

train_4598

Typically graphs are constructed based on a relational knowledge source, e.g., citation links in (Lu and Getoor, 2003), or distance between instances (Zhu, 2005).

we do not have access to such a relational knowledge in our setting.

contrasting

train_4599

On one hand, this is actually determined by real distribution of ratings for a particular prompt, e.g., how many essays are with an extreme quality for a given prompt in the target data.

a fine-grained tuning of the RankSVM (e.g., tuning C + and C − for positive and negative exam-ples separately) may partially resolve the problem, which is left for the future work.

contrasting