id
stringlengths 7
12
| sentence1
stringlengths 6
1.27k
| sentence2
stringlengths 6
926
| label
stringclasses 4
values |
|---|---|---|---|
train_98500 | Despite those positive outcomes, several issues with multi-task learning for sequence labeling remain open. | actually, by training our model on NER, chunking and POS tagging, we report state-of-the-art (or highly competitive) results on each task, without using external knowledge (such as gazetteers that has been shown to be important for NER), or hand-picking tasks to combine. | neutral |
train_98501 | The authors acknowledge the Texas Advanced Computing Center (TACC) at The University of Texas at Austin for providing HPC resources used to conduct this research. | this establishes a new state-ofthe-art on the test set, outperforming concurrently published work (Xiong et al., 2019) and matching the performance of a BERt model (Devlin et al., 2018) on this task. | neutral |
train_98502 | Training data We train the relabeling function g on another synthetically-noised dataset D drop generated from the manually-labeled data D. To mimic the type distribution of the distantly-labeled examples, we take each example (s, m, t) and randomly drop each type with a fixed rate 0.7 independent of other types to produce a new type set t . | we show in the next section that this is not sufficient for denoising. | neutral |
train_98503 | We generate artificial subjectverb agreement errors from large amounts of data. | both neural models obtain higher F 0.5 scores than the rule-based baseline, on average and across the board, i.e., +10.6 for LSTM ESL and +15.7 for LSTM ESL+Art . | neutral |
train_98504 | We train these models using EM for 500 iterations or until convergence, and we select the model with the lowest perplexity from among 70 random restarts. | here we note a mixed result: whilst de, sv, and it do benefit from POS information, the other languages do not, obtaining great improvements from MUSE embed-dings instead. | neutral |
train_98505 | This challenge should be taken into consideration in model design. | we also thank USC Plus Lab and UCLA-NLP group for discussion and comments. | neutral |
train_98506 | Training on increasing amounts of target samples improves the model performances monotonically for each target language and the model leveraging the bilingual data consistently outperforms the monolingual baseline model. | in this section, we present our evaluation for zeroshot learning. | neutral |
train_98507 | For both SVD-aligned and ADV-aligned, we use the embeddings as provided by the original authors. | among the 5 considered languages, Turkish seemed to benefit the least from cross-lingual learning in all experiments. | neutral |
train_98508 | 10 serves as a regularization that encourages the output of R(G) to be distinguishable among classes. | the performances of these works are highly dependent on manually annotated training data while annotation process is time-consuming and expensive. | neutral |
train_98509 | Rotten Tomatoes and Idebate dataset (Wang and Ling, 2016) use online text as source, but they are relatively small in scale: 3.7K posts of Rotten-Tomatoes compared to 80K posts of TIFU-short as shown in Table 1. | second, we propose a novel abstractive summarization model named multilevel memory networks (MMN), equipped with multi-level memory to store the information of text from different levels of abstraction. | neutral |
train_98510 | Different from previous approaches, we propose to alleviate such bias issue by changing the source of summarization dataset. | high scores of the TIFU dataset in both metrics show that it is potentially an excellent benchmark for evaluation of abstractive summarization systems. | neutral |
train_98511 | We collect data from Reddit, which is a discussion forum platform with a large number of subreddits on diverse topics and interests. | instead of using only the last layer output of CNNs, we exploit the outputs of multiple layers of CNNs to construct S sets of memories. | neutral |
train_98512 | We combine the best variants from the three approaches into a single system by taking the majority vote from the models. | all content included is accurate, with no irrelevant details or repetitions. | neutral |
train_98513 | In order to asses the effectiveness of AL for neural text compression we extend the OpenNMT 7 implementations with our interactive framework following Algorithm 1. | + Coverage-AL: The urgency of the situation in Alaska , Defenders needs your immediate assistance . | neutral |
train_98514 | It is therefore indispensable to minimize the cost of data annotation. | neural sequence-to-sequence (Seq2Seq) models have shown remarkable success in many areas of natural language processing and specifically in natural language generation tasks, including text compression (Rush et al., 2015;Filippova et al., 2015;Yu et al., 2018;Kamigaito et al., 2018). | neutral |
train_98515 | Given a question, the system predicts an answer using an extractive summary as the source input. | these summaries should factually adhere to the content of the source text and present the reader with the key points therein. | neutral |
train_98516 | Abstractive methods can thus introduce new words to the summary that are not present in the source article. | the articles were evenly split across the four competing systems, and each HIt was completed by 5 turkers. | neutral |
train_98517 | As shown in Table 2, the linguistic output of SL & CL is closer to the language used by humans: Our agent is able to produce a much richer and less repetitive output than both BL and RL. | overall, the accuracy of the CL and RL models is close. | neutral |
train_98518 | Language is highly abstract: one dialog can correctly describe a lot of different scenes in real world, so why should we force a dialog to fit one single example among them? | the full training procedure is specified in Algorithm 1. | neutral |
train_98519 | (2016) and Arora et al. | we provide these results in Table 4 and observe that each component is indeed important for our model. | neutral |
train_98520 | Intuitively, we expect our model to have a lower sample complexity since training our model involves learning fewer parameters. | for multimodal personality traits recognition on POM (left side of Table 1), our baseline is able to additionally outperform more complicated memory-based recurrent models such as MfN on several metrics. | neutral |
train_98521 | The models on the leaderboard are evaluated on a private unseen test set which contains 18 new environments. | while these tasks are driven by different goals, they all require agents that can perceive their surroundings, understand the goal (either presented visually or in language instructions), and act in a virtual environment. | neutral |
train_98522 | Our speaker model is an enhanced version of the encoder-decoder model of Fried et al. | we introduce our environmental dropout method to mimic the "new" environment E , as described next in Sec. | neutral |
train_98523 | Inspired by reading strategies, with limited resources and a pretrained transformer, we propose three strategies to improve machine reading comprehension. | we use a sigmoid function instead of softmax at the final layer ( Figure 1) and regard the task as a binary (i.e., correct or incorrect) classification problem over each (document, question, answer option) instance. | neutral |
train_98524 | For experiments with two datasets, we use Algorithm 2; for experiments with three datasets we find the re-weighting mechanism in Section 4.2 to have a better performance (a detailed comparison will be presented in Section 5.4). | the similar gain indicates that our method is orthogonal to ELMo. | neutral |
train_98525 | Since the GCN layer retains important structural information and is sensitive to positional data from the syntax tree, we consider it as a position-based approach. | in this paper, we introduced the application of GCN and attention mechanism to identification of verbal MWEs and finally proposed and tested a hybrid approach integrating both models. | neutral |
train_98526 | Second, the attention-based variants further boosted performance in comparison with their counterparts without attention. | incorrect attention led to a large drop in segmentation accuracy. | neutral |
train_98527 | Su and Lee (2017) also introduce a pixel-based model that learns character features from font images. | the final embedding of the target word is indirectly affected by the visual information. | neutral |
train_98528 | The BPE algorithm constructs a subword list from raw data and lattice LSTM introduces subwords into character LSTM representation. | we examine their non-pretrained model performance for fair comparison. | neutral |
train_98529 | Our work is in line with their work in directly using word information for CWS. | as shown in Table 5, among the ten most improved words, seven words are domain-specific noun entities, including person names, disease names and chemical compound names. | neutral |
train_98530 | Although the in-word negative sampling method proposed above is expected to prevent our model from incorrectly splitting multi-character words, we still want our model to pay more attention to the segmentation of such words. | • On four datasets in different special domains, our model improves the word F-measure by more than 3.0%, compared with the state-ofthe-art baseline segmenter. | neutral |
train_98531 | In contrast, we model characterbased POS tagging. | to further investigate the robustness of our model, we conduct experiments with different levels of corrupted tokenization in English. | neutral |
train_98532 | While most pure character-level models cannot ensure consistent labels for each character of a token, our semi-CRF outputs correct segments in most cases (tokenization F 1 is 98.69%, see Table 4), and ensures a single label for all characters of a segment. | we calculate joint tokenization and UPOS (universal POS) F 1 scores. | neutral |
train_98533 | Using a dictionary with NN is also popular (Zhang et al., 2018b;. | we do not induce a uniform smoothing. | neutral |
train_98534 | Our unsupervised dynamic speaker model differs from previous work in that we build speaker embeddings as a weighted combination of latent modes with weights computed based on the utterance. | second, we use the learned dynamic speaker embed- dings in two representative tasks in dialogs: predicting user topic decisions in socialbot dialogs, and classifying dialog acts in human-human dialogs. | neutral |
train_98535 | Figure 1 shows a visual comparison between outputs generated by the two models. | to contextualize these results, we compare disfluency removal as a post-processing step after end-to-end speech translation with the original disfluent pardev test Model 1Ref 2Ref 1Ref 2Ref Postproc. | neutral |
train_98536 | While SACNN can focus on important segments and gain local features, DepRNN helps to handle long-distance dependency between two entities based on the SDP as well as provide subject and object roles of two entities for the directional relation. | in this paper, we present a new model combining both CNNs and RNNs, exploiting the information from both the raw sequence and the SDP. | neutral |
train_98537 | We also propose the SACNN which automatically focus on the essential segments and gains local features. | the relation classification task is treated as a multi-class classification problem. | neutral |
train_98538 | In the future work, we will consider to detect events and their sentence-level and documentlevel factuality with a joint framework, and we will also continue to expand the scale of our DLEF corpus. | (Time: November, 2017) (Document-level factuality of the event "reach" is CT-.) | neutral |
train_98539 | Table 2 indicates that sentence-level factuality usually agrees with document-level factuality in CT+ documents, making them straightforward to be identified. | the results is given in table 6, which shows that contexts can improve the performance more significantly on the Chinese corpus than the English corpus. | neutral |
train_98540 | To our best knowledge, this is the first document-level event factuality corpus. | no previous work annotated a document-level corpus. | neutral |
train_98541 | 4, where Mintz (Mintz et al., 2009), MultiR (Hoffmann et al., 2011) and MIMLRE (Surdeanu et al., 2012) are conventional feature-based methods, and (Lin et al., 2016) and are PCNN-based ones 4 . | table 5 compares the AUC values reported in these two papers and the results of our proposed models. | neutral |
train_98542 | During the training, the model learns a joint network including F, E and D to minimize the empirical loss Eq (1). | in order to have a reasonable comparisons, we report the average ranking score for each method. | neutral |
train_98543 | In order to transfer the rules into a new policy π r , the KL divergence between the posterior of π and π r should be minimized, this can be formally defined as minKL(P π (A t |S t , θ π )||P πr (A t |S t , θ π )) (5) Optimizing the constrained convex problem defined by Eq. | here we use a rule pattern as the Fig.1 shows (?). | neutral |
train_98544 | • We apply the PR REINFORCE to the instance selection task for DS dataset to alleviate the wrong label problem in DS. | unbiased methods, such as REINFORCE, could usually take much time to train. | neutral |
train_98545 | The Metropolis Hastings Walker (MHW) method (Li et al., 2014) scales well in the number of topics, and uses a collapsed inference algorithm, but it operates in the batch setting, so it is not scalable to large corpora. | the SCVB0 algorithm does not leverage sparsity, and hence requires O(K) operations per word token. | neutral |
train_98546 | (S3) W AE W +W AE L +Sim: It is similar to W AE W +Sim except we also include the average embeddings of attribute labels associated with the instance. | we want the words and their attribute labels to be close to each other in the embedding space and the embeddings of different labels to be far away from each other. | neutral |
train_98547 | Previous studies have mostly focused on estimating each annotator's overall reliability on the entire annotation task. | when constructing models that learn from noisy labels produced by multiple annotators, it is important to accurately estimate the reliability of annotators. | neutral |
train_98548 | The function H denotes the cross entropy between two distributions. | the review representation is obtained by averaging word embeddings. | neutral |
train_98549 | To solve the FDKB task using KBR, one feasible way is to exhaustively calculate the scores of all (r, t) combinations for the given head entity h. Afterwards, the highly-scored facts are returned as results. | semantic matching models such as REsCAL (Nickel et al., 2011), DistMult (Yang et al., 2014), Complex (Trouillon et al., 2016), HolE (Nickel et al., 2016) and ANALOGY (Liu et al., 2017) model the score of triples by the semantic similarity. | neutral |
train_98550 | RE has been widely studied in NLP community for many years. | we aim to address them and further extensions of our model in future works. | neutral |
train_98551 | Moreover, it is hard for machines to learn the attention weights from a long sequence of input text. | our RbSP model yields an F1-score of 86.3%, outperforms other comparative models, except Multi-Att-CNN model of with multi-level attention CNN. | neutral |
train_98552 | Our idea is based on fundamental notion that the syntactic structure of a sentence consists of binary asymmetrical relations between words (Nivre, 2005). | we propose a novel DNN framework which combines Long Short-Term Memory (LSTM) (Hochreiter and Schmidhuber, 1997) and Convolutional Neural Networks (CNN) (LeCun et al., 1989) with a multi-attention layer. | neutral |
train_98553 | Given questions in natural language (NL), the goal of KBQA is to automatically find answers from the underlying KB, which provides a more natural and intuitive way to access the vast underlying knowledge resources. | for instance, the same question can be expressed in various ways in NL while a KB usually has a canonical lexicon. | neutral |
train_98554 | Unlike a basic memory network (Weston et al., 2014), its addressing stage is based on the key memory while the reading stage uses the value memory, which gives greater flexibility to encode prior knowledge via functionality separation. | 4 shows the attention heatmap generated for a test question "who did location surrender to in number " (where "location" and " number " are entity types which replace the topic entity mention "France" and the constraint entity mention "ww2", respectively in the original question). | neutral |
train_98555 | Following previous work , we also try building entailment-like training data from SQuAD 2.0 (Rajpurkar et al., 2018). | molecular and quantum physics show that the electromagnetic force is the fundamental interaction responsible for contact forces. | neutral |
train_98556 | Preliminary experiments found that simply transferring the lower-level weights of extractive QA models was ineffective, so we instead consider three methods of con-structing entailment-like data from extractive QA data. | in this paper, we gather a large number of additional yes/no questions in order to construct a dedicated yes/no QA dataset. | neutral |
train_98557 | Classify: Finally we feed [v * ; p * ] into a fully connected layer, and then through a softmax layer to predict the output class. | part of the value of this dataset is that it contains questions that people genuinely want to answer. | neutral |
train_98558 | The learning rate is set to 0.001. | kVQU+AR: applies the approach introduced in this paper that additionally considers both the key and Value representations in the Query Updating (kVQU). | neutral |
train_98559 | Then with the hop size increasing, the performance drops. | [2016] achieves the state- Table 2: Running examples of addressed keys and corresponding relevance probabilities before and after introducing the STOP strategy. | neutral |
train_98560 | We perform this additional contextualization only when sentences form contiguous text. | given a premise sentence P i , the entailment function f e computes a single hypothesis-aware representation x i containing information in the premise that is relevant to entailing the answer hypothesis H qa . | neutral |
train_98561 | Models such as Decomposable Attention (Parikh et al., 2016) and ESIM (Chen et al., 2017), on the other hand, find alignments between the hypothesis and premise words through crossattention. | so we haveX This is similar to a standard attention mechanism, where attended representation is computed by summing the scaled representations. | neutral |
train_98562 | To understand why, consider the set of premises in Figure 1, which entail the hypothesis H c . | figure 4a illustrates this behavior. | neutral |
train_98563 | The goal of this module is to identify sentences in the paragraph that are important for the given hypothesis. | such scaled addition is not possible when the outputs from lower layers are not of the same shapes, as in the following case. | neutral |
train_98564 | Processed sentence after this step is sentence 2 in Figure 3. | since we do not have labeled data, we need to identify mentions in contexts, and assign gender labels to them. | neutral |
train_98565 | Gendered Language Gendered language is the use of words and phrases that discriminate 1 the gender of a subject. | we also look at classifier probability distribution for human decisions shown in box and whisker plot in Figure 5, where x-axis is the classifier probability of the mention being female. | neutral |
train_98566 | o t is the output and e t is the hidden state of the GRU. | so we first try to tackle it based on the sequence-to-sequence model, which is commonly used in machine translation. | neutral |
train_98567 | This means the Seq2Seq model can better explain the hate symbols when Twitter users intentionally misspell or abbreviate common slur terms. | so we first try to tackle it based on the sequence-to-sequence model, which is commonly used in machine translation. | neutral |
train_98568 | Distant supervision has been widely used in relation extraction tasks without hand-labeled datasets recently. | evaluations P@100 P@200 P@300 Automatic@NYT 76.2 73.1 67.4 Manual@NYT 96.0(+19.8) 95.5(+22.4) 91.0(+23.6) Automatic@A-NYT 93.0 89.5 88.0 Manual@A-NYT 96.0(+3.0) 92.5(+3.0) 90.7(+2.7) To further demonstrate the effectiveness of our training strategies, we compare Generative Adversarial Training (GAT) with other baselines on the partially labeled dataset A-NYT as shown in Figure 4(b). | neutral |
train_98569 | To further alleviate the effect of wrong labeling problem, soft-label training algorithm (Liu et al., 2017b), reinforcement learning methods (Feng et al., 2018;Qin et al., 2018b) and additional side information (Vashishth et al., 2018;Wang et al., 2018) have been used. | a former negative instance has a big chance to be credible negative if any of its entities is not mentioned in the description of the other one. | neutral |
train_98570 | The edges are weighted by the coreference and relation scores, which are trained according to the neural architecture explained in Section 3.1. | our Model We develop a general information extraction framework (DYGIE) to identify and classify entities, relations, and coreference in a multi-task setup. | neutral |
train_98571 | 3 This experiment envisions a pipeline where the noisy source is first automatically corrected and then translated. | combining this simple method with an automatic grammar correction system, we find that we can recover 1.5 BLEU. | neutral |
train_98572 | (2017) employ a common encoder to encode the sentences from both the in-domain and out-of-domain data and meanwhile add a discriminator to the encoder to make sure that only domain-invariant information is transferred to the decoder. | in addition to the common encoder, Zeng et al. | neutral |
train_98573 | In this paper, we present a method to make use of out-of-domain data to help in-domain translation. | our method can achieve a mild improvement on the out-of-domain compared to the baseline system. | neutral |
train_98574 | For Encoder Context integration, the HAN encoder (Miculicich et al., 2018) is the best for TED and News datasets, however, the results are statistically insignificant with respect to our best model. | the Hierarchical Attention module has four operations: 1. | neutral |
train_98575 | This amounts to 1,200 sentence pairs in the target side. | * " indicates that the correlation is significantly better than the next-best one. | neutral |
train_98576 | This makes it possible to evaluate the effectiveness of adversarial attacks or defenses either using goldstandard human evaluation, or approximations that can be calculated without human intervention. | for a word-based translation model M 6 , and given an input sentence w 1 , . | neutral |
train_98577 | The implementation of our method is available at https: //github.com/hassyGo/NLG-RL. | this section discusses our main contribution: how efficient our method is in accelerating reinforcement learning for sentence generation. | neutral |
train_98578 | For reference, we report the test set results in Table 4. | for the En-Ja (2M) and En-Ja (2M, SW) datasets, we used a single GPU of NVIDIA Tesla V100 4 to speedup our experiments. | neutral |
train_98579 | Better generation of rare words These BLEU scores suggest that our method for reinforcement learning has the potential to outperform the full softmax baseline. | we then review how reinforcement learning is used, and present a simple and efficient method to accelerate the training. | neutral |
train_98580 | This paper has presented how to accelerate reinforcement learning for sentence generation tasks by reducing large action spaces. | e(y t ) is the y t -th row vector in W p , and the technique has shown to be effective in machine translation (Hashimoto and Tsuruoka, 2017) and text summarization (Paulus et al., 2018). | neutral |
train_98581 | The work related to this paper falls into two sub topics, described as follows. | gal and ghahramani used dropout in DNNs as an approximate Bayesian inference in deep gaussian processes (gal and ghahramani, 2016) to mitigate the problem of representing uncertainty in deep learning without sacrificing the computational complexity. | neutral |
train_98582 | Dropout-based methods have also been extended to various tasks such as computer vision , autonomous vehicle safety (McAllister et al., 2017) and medical decision making (van der Westhuizen and Lasenby, 2017). | we used review data from the Sports and outdoors category, with 272,630 data samples and rating labels from 1 to 5. | neutral |
train_98583 | The Dropout operation will be randomly applied to the activations during the training and uncertainty measurement phrases, but will not be applied to the evaluation phrase. | the shortened intra-class distance and enlarged inter-class distance can reduce the prediction variance and increase the confidence for the accurate predictions. | neutral |
train_98584 | We experiment with the following state-of-the-art neural text classification methods: 1. | helpfulness might be conflated with other reasons such as humour, sentiment in certain domains. | neutral |
train_98585 | As a consequence the model has little freedom in discovering and concentrating on some natural label order. | • Sequence Generation Model (SGM) (Yang et al., 2018) which trains the RNN model similar to seq2seq-RNN but uses a new decoder structure that computes a weighted global embedding based on all labels as opposed to just the top one at each timestep. | neutral |
train_98586 | Between the above two extremes are Vinyals-RNN-max and set-RNN (we have omitted Vinyals-RNN-sample and Vinyals-RNN-maxdirect here as they are similar to Vinyals-RNNmax). | if for each document, RNN finds one good way of ordering relevant labels (such as hierarchically) and allocates most of the probability mass to the sequence in that order, the model still assigns low probabilities to the ground truth label sets and will be penalized heavily. | neutral |
train_98587 | For both lexicons, we keep 90% of the lexicon in train set and remaining 10% in validation set. | with this motivation, we focus on increasing accuracy on the most frequent words. | neutral |
train_98588 | Figure 2 and Table 5 categorize the errors of systems trained on 3 types of lexicons (Romanized version of our lexicon is discussed in section 6.3) by using Algorithm 2. | for the remaining 30K words (29105 words to be exact), we observe that at least one set provides different transcription. | neutral |
train_98589 | (2018) utilized additional side information from KBs for improved RE. | to the best of our knowledge, ours is the first principled framework to combine and jointly learn heterogeneous representations from both language and knowledge for the RE task. | neutral |
train_98590 | All above models are based on handcrafted feature. | p φ (y|x) is calculated by the relation classifier from the semi-supervised learning framework. | neutral |
train_98591 | In this paper, we proposed RCEND to fully exploit valid information of the noisy data in distant supervision relation classification. | sentence Ds Gold s1:Al Gore was waiting to board a commercial flight from Nashville to Miami... s2:There were also performers who were born in Louisiana , including Lucinda Williams... s3:Boggs was married, had three young children and lived in Brewster NA LivedIn it suffers from noisy labeling problem due to the irrelevance of aligned text and incompleteness of KB, which consists of false positives and false negatives. | neutral |
train_98592 | FastText optimizes the loss function in Eq.1, but uses the scoring function s F T defined in Eq.2. | when the amount of misspellings is higher, i.e., r ∈ {0.25, 0.375}, MOE improves the results over the baseline for all of the α values. | neutral |
train_98593 | In fact, it is the only component of the loss function which attempts to learn these relationships. | , w i+l } for some l set as a hyperparameter. | neutral |
train_98594 | Conversely, CC-DBP and NYT-FB contain dirtier sentences which mean a high probability of incurring wrong labeling. | for instance, T-REX has well-written textual mentions, because the sentences are extracted from Wikipedia. | neutral |
train_98595 | Formally, given is a new dense representation of w ij which encodes also the information of the whole sentence. | siamese neural networks (Bromley et al., 1993) are well suited to this task because they are specifically designed to compute the similarity between two instances. | neutral |
train_98596 | This kind of neural network has been used in both computer vision (Koch et al., 2015) and natural language processing (Mueller and Thyagarajan, 2016;Neculoiu et al., 2016) in order to map two similar instances close in a feature space. | the works in (Gladkova et al., 2016;Vylomova et al., 2016) explore the use of word vectors to model the semantic relations. | neutral |
train_98597 | In order to avoid this situation, and to prevent overfitting, we apply l 2 regularization to the noise model. | 3 We further investigate the performance of the proposed approach on instance-dependent label noise by flipping each class labels with different noise percentages as shown in Fig. | neutral |
train_98598 | The results of our analysis imply that early in training, representing part of speech is the natural way to get initial high performance. | 3 If the LM begins by maximizing mutual information with input, because the input is identical for the LM and tag models it may lead to these similar initial representations, followed by a decline in similarity as the compression narrows to properties specific to each task. | neutral |
train_98599 | Note that topic and UDP POS both apply to the same enwikipedia corpus, but PTB POS and SEM use two different unaligned sets from the GMB corpus. | we ask: is our first conceptual shift (to SVCCA) necessary? | neutral |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.