EMNLP 2020 论文列表

Abstractive document summarization is usually modeled as a sequence-to-sequence (SEQ2SEQ) learning problem. Unfortunately, training large SEQ2SEQ based summarization models on limited supervised summarization data is challenging. This paper presents three sequence-to-sequence pre-training (in shorthand, STEP) objectives which allow us to pre-train a SEQ2SEQ based abstractive summarization model on unlabeled text. The main idea is that, given an input text artificially constructed from a document, a model is pre-trained to reinstate the original document. These objectives include sentence reordering, next sentence generation and masked document generation, which have close relations with the abstractive document summarization task. Experiments on two benchmark summarization datasets (i.e., CNN/DailyMail and New York Times) show that all three objectives can improve performance upon baselines. Compared to models pre-trained on large-scale data (larger than 160GB), our method, with only 19GB text for pre-training, achieves comparable results, which demonstrates its effectiveness.

We consider the problem of better modeling query-cluster interactions to facilitate query focused multi-document summarization. Due to the lack of training data, existing work relies heavily on retrieval-style methods for assembling query relevant summaries. We propose a coarse-to-fine modeling framework which employs progressively more accurate modules for estimating whether text segments are relevant, likely to contain an answer, and central. The modules can be independently developed and leverage training data if available. We present an instantiation of this framework with a trained evidence estimator which relies on distant supervision from question answering (where various resources exist) to identify segments which are likely to answer the query and should be included in the summary. Our framework is robust across domains and query types (i.e., long vs short) and outperforms strong comparison systems on benchmark datasets.

Sentence-level extractive text summarization is substantially a node classification task of network mining, adhering to the informative components and concise representations. There are lots of redundant phrases between extracted sentences, but it is difficult to model them exactly by the general supervised methods. Previous sentence encoders, especially BERT, specialize in modeling the relationship between source sentences. While, they have no ability to consider the overlaps of the target selected summary, and there are inherent dependencies among target labels of sentences. In this paper, we propose HAHSum (as shorthand for Hierarchical Attentive Heterogeneous Graph for Text Summarization), which well models different levels of information, including words and sentences, and spotlights redundancy dependencies between sentences. Our approach iteratively refines the sentence representations with redundancy-aware graph and delivers the label dependencies by message passing. Experiments on large scale benchmark corpus (CNN/DM, NYT, and NEWSROOM) demonstrate that HAHSum yields ground-breaking performance and outperforms previous extractive summarizers.

Evaluation of a document summarization system has been a critical factor to impact the success of the summarization task. Previous approaches, such as ROUGE, mainly consider the informativeness of the assessed summary and require human-generated references for each test summary. In this work, we propose to evaluate the summary qualities without reference summaries by unsupervised contrastive learning. Specifically, we design a new metric which covers both linguistic qualities and semantic informativeness based on BERT. To learn the metric, for each summary, we construct different types of negative samples with respect to different aspects of the summary qualities, and train our model with a ranking loss. Experiments on Newsroom and CNN/Daily Mail demonstrate that our new evaluation method outperforms other metrics even without reference summaries. Furthermore, we show that our method is general and transferable across datasets.

Modeling content importance is an essential yet challenging task for summarization. Previous work is mostly based on statistical methods that estimate word-level salience, which does not consider semantics and larger context when quantifying importance. It is thus hard for these methods to generalize to semantic units of longer text spans. In this work, we apply information theory on top of pre-trained language models and define the concept of importance from the perspective of information amount. It considers both the semantics and context when evaluating the importance of each semantic unit. With the help of pre-trained language models, it can easily generalize to different kinds of semantic units n-grams or sentences. Experiments on CNN/Daily Mail and New York Times datasets demonstrate that our method can better model the importance of content than prior work based on F1 and ROUGE scores.

Aspect-based sentiment analysis (ABSA) aims to predict the sentiment towards a specific aspect in the text. However, existing ABSA test sets cannot be used to probe whether a model can distinguish the sentiment of the target aspect from the non-target aspects. To solve this problem, we develop a simple but effective approach to enrich ABSA test sets. Specifically, we generate new examples to disentangle the confounding sentiments of the non-target aspects from the target aspect’s sentiment. Based on the SemEval 2014 dataset, we construct the Aspect Robustness Test Set (ARTS) as a comprehensive probe of the aspect robustness of ABSA models. Over 92% data of ARTS show high fluency and desired sentiment on all aspects by human evaluation. Using ARTS, we analyze the robustness of nine ABSA models, and observe, surprisingly, that their accuracy drops by up to 69.73%. We explore several ways to improve aspect robustness, and find that adversarial training can improve models’ performance on ARTS by up to 32.85%. Our code and new test set are available at https://github.com/zhijing-jin/ARTS_TestSet

As an important research issue in the natural language processing community, multi-label emotion detection has been drawing more and more attention in the last few years. However, almost all existing studies focus on one modality (e.g., textual modality). In this paper, we focus on multi-label emotion detection in a multi-modal scenario. In this scenario, we need to consider both the dependence among different labels (label dependence) and the dependence between each predicting label and different modalities (modality dependence). Particularly, we propose a multi-modal sequence-to-set approach to effectively model both kinds of dependence in multi-modal multi-label emotion detection. The detailed evaluation demonstrates the effectiveness of our approach.

Emotion-cause pair extraction (ECPE) is a new task that aims to extract the potential pairs of emotions and their corresponding causes in a document. The existing methods first perform emotion extraction and cause extraction independently, and then perform emotion-cause pairing and filtering. However, the above methods ignore the fact that the cause and the emotion it triggers are inseparable, and the extraction of the cause without specifying the emotion is pathological, which greatly limits the performance of the above methods in the first step. To tackle these shortcomings, we propose two joint frameworks for ECPE: 1) multi-label learning for the extraction of the cause clauses corresponding to the specified emotion clause (CMLL) and 2) multi-label learning for the extraction of the emotion clauses corresponding to the specified cause clause (EMLL). The window of multi-label learning is centered on the specified emotion clause or cause clause and slides as their positions move. Finally, CMLL and EMLL are integrated to obtain the final result. We evaluate our model on a benchmark emotion cause corpus, the results show that our approach achieves the best performance among all compared systems on the ECPE task.

The task of emotion-cause pair extraction deals with finding all emotions and the corresponding causes in unannotated emotion texts. Most recent studies are based on the likelihood of Cartesian product among all clause candidates, resulting in a high computational cost. Targeting this issue, we regard the task as a sequence labeling problem and propose a novel tagging scheme with coding the distance between linked components into the tags, so that emotions and the corresponding causes can be extracted simultaneously. Accordingly, an end-to-end model is presented to process the input texts from left to right, always with linear time complexity, leading to a speed up. Experimental results show that our proposed model achieves the best performance, outperforming the state-of-the-art method by 2.26% (p<0.001) in F1 measure.

Aspect based sentiment analysis, predicting sentiment polarity of given aspects, has drawn extensive attention. Previous attention-based models emphasize using aspect semantics to help extract opinion features for classification. However, these works are either not able to capture opinion spans as a whole, or not able to capture variable-length opinion spans. In this paper, we present a neat and effective structured attention model by aggregating multiple linear-chain CRFs. Such a design allows the model to extract aspect-specific opinion spans and then evaluate sentiment polarity by exploiting the extracted opinion features. The experimental results on four datasets demonstrate the effectiveness of the proposed model, and our analysis demonstrates that our model can capture aspect-specific opinion spans.

Aspect-category sentiment analysis (ACSA) aims to predict sentiment polarities of sentences with respect to given aspect categories. To detect the sentiment toward a particular aspect category in a sentence, most previous methods first generate an aspect category-specific sentence representation for the aspect category, then predict the sentiment polarity based on the representation. These methods ignore the fact that the sentiment of an aspect category mentioned in a sentence is an aggregation of the sentiments of the words indicating the aspect category in the sentence, which leads to suboptimal performance. In this paper, we propose a Multi-Instance Multi-Label Learning Network for Aspect-Category sentiment analysis (AC-MIMLLN), which treats sentences as bags, words as instances, and the words indicating an aspect category as the key instances of the aspect category. Given a sentence and the aspect categories mentioned in the sentence, AC-MIMLLN first predicts the sentiments of the instances, then finds the key instances for the aspect categories, finally obtains the sentiments of the sentence toward the aspect categories by aggregating the key instance sentiments. Experimental results on three public datasets demonstrate the effectiveness of AC-MIMLLN.

The state-of-the-art methods in aspect-level sentiment classification have leveraged the graph based models to incorporate the syntactic structure of a sentence. While being effective, these methods ignore the corpus level word co-occurrence information, which reflect the collocations in linguistics like “nothing special”. Moreover, they do not distinguish the different types of syntactic dependency, e.g., a nominal subject relation “food-was” is treated equally as an adjectival complement relation “was-okay” in “food was okay”. To tackle the above two limitations, we propose a novel architecture which convolutes over hierarchical syntactic and lexical graphs. Specifically, we employ a global lexical graph to encode the corpus level word co-occurrence information. Moreover, we build a concept hierarchy on both the syntactic and lexical graphs for differentiating various types of dependency relations or lexical word pairs. Finally, we design a bi-level interactive graph convolution network to fully exploit these two graphs. Extensive experiments on five bench- mark datasets show that our method outperforms the state-of-the-art baselines.

Contextualized word embeddings have been employed effectively across several tasks in Natural Language Processing, as they have proved to carry useful semantic information. However, it is still hard to link them to structured sources of knowledge. In this paper we present ARES (context-AwaRe Embeddings of Senses), a semi-supervised approach to producing sense embeddings for the lexical meanings within a lexical knowledge base that lie in a space that is comparable to that of contextualized word vectors. ARES representations enable a simple 1 Nearest-Neighbour algorithm to outperform state-of-the-art models, not only in the English Word Sense Disambiguation task, but also in the multilingual one, whilst training on sense-annotated data in English only. We further assess the quality of our embeddings in the Word-in-Context task, where, when used as an external source of knowledge, they consistently improve the performance of a neural model, leading it to compete with other more complex architectures. ARES embeddings for all WordNet concepts and the automatically-extracted contexts used for creating the sense representations are freely available at http://sensembert.org/ares.

We propose the novel Within-Between Relation model for recognizing lexical-semantic relations between words. Our model integrates relational and distributional signals, forming an effective sub-space representation for each relation. We show that the proposed model is competitive and outperforms other baselines, across various benchmarks.

State-of-the-art methods for Word Sense Disambiguation (WSD) combine two different features: the power of pre-trained language models and a propagation method to extend the coverage of such models. This propagation is needed as current sense-annotated corpora lack coverage of many instances in the underlying sense inventory (usually WordNet). At the same time, unambiguous words make for a large portion of all words in WordNet, while being poorly covered in existing sense-annotated corpora. In this paper, we propose a simple method to provide annotations for most unambiguous words in a large corpus. We introduce the UWA (Unambiguous Word Annotations) dataset and show how a state-of-the-art propagation-based model can use it to extend the coverage and quality of its word sense embeddings by a significant margin, improving on its original results on WSD.

Meta-embedding learning, which combines complementary information in different word embeddings, have shown superior performances across different Natural Language Processing tasks. However, domain-specific knowledge is still ignored by existing meta-embedding methods, which results in unstable performances across specific domains. Moreover, the importance of general and domain word embeddings is related to downstream tasks, how to regularize meta-embedding to adapt downstream tasks is an unsolved problem. In this paper, we propose a method to incorporate both domain-specific and task-oriented information into meta-embeddings. We conducted extensive experiments on four text classification datasets and the results show the effectiveness of our proposed method.

The challenge of both achieving task completion by querying the knowledge base and generating human-like responses for task-oriented dialogue systems is attracting increasing research attention. In this paper, we propose a “Two-Teacher One-Student” learning framework (TTOS) for task-oriented dialogue, with the goal of retrieving accurate KB entities and generating human-like responses simultaneously. TTOS amalgamates knowledge from two teacher networks that together provide comprehensive guidance to build a high-quality task-oriented dialogue system (student network). Each teacher network is trained via reinforcement learning with a goal-specific reward, which can be viewed as an expert towards the goal and transfers the professional characteristic to the student network. Instead of adopting the classic student-teacher learning of forcing the output of a student network to exactly mimic the soft targets produced by the teacher networks, we introduce two discriminators as in generative adversarial network (GAN) to transfer knowledge from two teachers to the student. The usage of discriminators relaxes the rigid coupling between the student and teachers. Extensive experiments on two benchmark datasets (i.e., CamRest and In-Car Assistant) demonstrate that TTOS significantly outperforms baseline methods.

Retrieving the proper knowledge relevant to conversational context is an important challenge in dialogue systems, to engage users with more informative response. Several recent works propose to formulate this knowledge selection problem as a path traversal over an external knowledge graph (KG), but show only a limited utilization of KG structure, leaving rooms of improvement in performance. To this effect, we present AttnIO, a new dialog-conditioned path traversal model that makes a full use of rich structural information in KG based on two directions of attention flows. Through the attention flows, AttnIO is not only capable of exploring a broad range of multi-hop knowledge paths, but also learns to flexibly adjust the varying range of plausible nodes and edges to attend depending on the dialog context. Empirical evaluations present a marked performance improvement of AttnIO compared to all baselines in OpenDialKG dataset. Also, we find that our model can be trained to generate an adequate knowledge path even when the paths are not available and only the destination nodes are given as label, making it more applicable to real-world dialogue systems.

We study multi-turn response generation for open-domain dialogues. The existing state-of-the-art addresses the problem with deep neural architectures. While these models improved response quality, their complexity also hinders the application of the models in real systems. In this work, we pursue a model that has a simple structure yet can effectively leverage conversation contexts for response generation. To this end, we propose four auxiliary tasks including word order recovery, utterance order recovery, masked word recovery, and masked utterance recovery, and optimize the objectives of these tasks together with maximizing the likelihood of generation. By this means, the auxiliary tasks that relate to context understanding can guide the learning of the generation model to achieve a better local optimum. Empirical studies with three benchmarks indicate that our model can significantly outperform state-of-the-art generation models in terms of response quality on both automatic evaluation and human judgment, and at the same time enjoys a much faster decoding process.