EMNLP 2020 论文列表

Unsupervised speech representation learning has shown remarkable success at finding representations that correlate with phonetic structures and improve downstream speech recognition performance. However, most research has been focused on evaluating the representations in terms of their ability to improve the performance of speech recognition systems on read English (e.g. Wall Street Journal and LibriSpeech). This evaluation methodology overlooks two important desiderata that speech representations should have: robustness to domain shifts and transferability to other languages. In this paper we learn representations from up to 8000 hours of diverse and noisy speech data and evaluate the representations by looking at their robustness to domain shifts and their ability to improve recognition performance in many languages. We find that our representations confer significant robustness advantages to the resulting recognition systems: we see significant improvements in out-of-domain transfer relative to baseline feature sets and the features likewise provide improvements in 25 phonetically diverse languages.

Recent advances in Knowledge Graph Embedding (KGE) allow for representing entities and relations in continuous vector spaces. Some traditional KGE models leveraging additional type information can improve the representation of entities which however totally rely on the explicit types or neglect the diverse type representations specific to various relations. Besides, none of the existing methods is capable of inferring all the relation patterns of symmetry, inversion and composition as well as the complex properties of 1-N, N-1 and N-N relations, simultaneously. To explore the type information for any KG, we develop a novel KGE framework with Automated Entity TypE Representation (AutoETER), which learns the latent type embedding of each entity by regarding each relation as a translation operation between the types of two entities with a relation-aware projection mechanism. Particularly, our designed automated type representation learning mechanism is a pluggable module which can be easily incorporated with any KGE model. Besides, our approach could model and infer all the relation patterns and complex relations. Experiments on four datasets demonstrate the superior performance of our model compared to state-of-the-art baselines on link prediction tasks, and the visualization of type clustering provides clearly the explanation of type embeddings and verifies the effectiveness of our model.

Natural language data exhibit tree-like hierarchical structures such as the hypernym-hyponym hierarchy in WordNet. FastText, as the state-of-the-art text classifier based on shallow neural network in Euclidean space, may not represent such hierarchies precisely with limited representation capacity. Considering that hyperbolic space is naturally suitable for modelling tree-like hierarchical data, we propose a new model named HyperText for efficient text classification by endowing FastText with hyperbolic geometry. Empirically, we show that HyperText outperforms FastText on a range of text classification tasks with much reduced parameters.

Current reading comprehension methods generalise well to in-distribution test sets, yet perform poorly on adversarially selected data. Prior work on adversarial inputs typically studies model oversensitivity: semantically invariant text perturbations that cause a model’s prediction to change. Here we focus on the complementary problem: excessive prediction undersensitivity, where input text is meaningfully changed but the model’s prediction does not, even though it should. We formulate an adversarial attack which searches among semantic variations of the question for which a model erroneously predicts the same answer, and with even higher probability. We demonstrate that models trained on both SQuAD2.0 and NewsQA are vulnerable to this attack, and then investigate data augmentation and adversarial training as defences. Both substantially decrease adversarial vulnerability, which generalises to held-out data and held-out attack spaces. Addressing undersensitivity furthermore improves model robustness on the previously introduced ADDSENT and ADDONESENT datasets, and models generalise better when facing train / evaluation distribution mismatch: they are less prone to overly rely on shallow predictive cues present only in the training set, and outperform a conventional model by as much as 10.9% F1.

Semantic role labeling is primarily used to identify predicates, arguments, and their semantic relationships. Due to the limitations of modeling methods and the conditions of pre-identified predicates, previous work has focused on the relationships between predicates and arguments and the correlations between arguments at most, while the correlations between predicates have been neglected for a long time. High-order features and structure learning were very common in modeling such correlations before the neural network era. In this paper, we introduce a high-order graph structure for the neural semantic role labeling model, which enables the model to explicitly consider not only the isolated predicate-argument pairs but also the interaction between the predicate-argument pairs. Experimental results on 7 languages of the CoNLL-2009 benchmark show that the high-order structural learning techniques are beneficial to the strong performing SRL models and further boost our baseline to achieve new state-of-the-art results.

This paper studies sentiment classification in the lifelong learning setting that incrementally learns a sequence of sentiment classification tasks. It proposes a new lifelong learning model (called L2PG) that can retain and selectively transfer the knowledge learned in the past to help learn the new task. A key innovation of this proposed model is a novel parameter-gate (p-gate) mechanism that regulates the flow or transfer of the previously learned knowledge to the new task. Specifically, it can selectively use the network parameters (which represent the retained knowledge gained from the previous tasks) to assist the learning of the new task t. Knowledge distillation is also employed in the process to preserve the past knowledge by approximating the network output at the state when task t-1 was learned. Experimental results show that L2PG outperforms strong baselines, including even multiple task learning.

Logic grid puzzle (LGP) is a type of word problem where the task is to solve a problem in logic. Constraints for the problem are given in the form of textual clues. Once these clues are transformed into formal logic, a deductive reasoning process provides the solution. Solving logic grid puzzles in a fully automatic manner has been a challenge since a precise understanding of clues is necessary to develop the corresponding formal logic representation. To meet this challenge, we propose a solution that uses a DistilBERT-based classifier to classify a clue into one of the predefined predicate types for logic grid puzzles. Another novelty of the proposed solution is the recognition of comparison structures in clues. By collecting comparative adjectives from existing dictionaries and utilizing a semantic framework to catch comparative quantifiers, the semantics of clues concerning comparison structures are better understood, ensuring conversion to correct logic representation. Our approach solves logic grid puzzles in a fully automated manner with 100% accuracy on the given puzzle datasets and outperforms state-of-the-art solutions by a large margin.

In this paper, we propose Multi^2OIE, which performs open information extraction (open IE) by combining BERT with multi-head attention. Our model is a sequence-labeling system with an efficient and effective argument extraction method. We use a query, key, and value setting inspired by the Multimodal Transformer to replace the previously used bidirectional long short-term memory architecture with multi-head attention. Multi^2OIE outperforms existing sequence-labeling systems with high computational efficiency on two benchmark evaluation datasets, Re-OIE2016 and CaRB. Additionally, we apply the proposed method to multilingual open IE using multilingual BERT. Experimental results on new benchmark datasets introduced for two languages (Spanish and Portuguese) demonstrate that our model outperforms other multilingual systems without training data for the target languages.

Deep neural networks have made great success on video captioning in supervised learning setting. However, annotating videos with descriptions is very expensive and time-consuming. If the video captioning algorithm can benefit from a large number of unlabeled videos, the cost of annotation can be reduced. In the proposed study, we make the first attempt to train the video captioning model on labeled data and unlabeled data jointly, in a semi-supervised learning manner. For labeled data, we train them with the traditional cross-entropy loss. For unlabeled data, we leverage a self-critical policy gradient method with the difference between the scores obtained by Monte-Carlo sampling and greedy decoding as the reward function, while the scores are the negative K-L divergence between output distributions of original video data and augmented video data. The final loss is the weighted sum of losses obtained by labeled data and unlabeled data. Experiments conducted on VATEX, MSR-VTT and MSVD dataset demonstrate that the introduction of unlabeled data can improve the performance of the video captioning model. The proposed semi-supervised learning algorithm also outperforms several state-of-the-art semi-supervised learning approaches.

Evaluating the trustworthiness of a model’s prediction is essential for differentiating between ‘right for the right reasons’ and ‘right for the wrong reasons’. Identifying textual spans that determine the target label, known as faithful rationales, usually relies on pipeline approaches or reinforcement learning. However, such methods either require supervision and thus costly annotation of the rationales or employ non-differentiable models. We propose a differentiable training–framework to create models which output faithful rationales on a sentence level, by solely applying supervision on the target task. To achieve this, our model solves the task based on each rationale individually and learns to assign high scores to those which solved the task best. Our evaluation on three different datasets shows competitive results compared to a standard BERT blackbox while exceeding a pipeline counterpart’s performance in two cases. We further exploit the transparent decision–making process of these models to prefer selecting the correct rationales by applying direct supervision, thereby boosting the performance on the rationale–level.

While recent advances in language modeling has resulted in powerful generation models, their generation style remains implicitly dependent on the training data and can not emulate a specific target style. Leveraging the generative capabilities of a transformer-based language models, we present an approach to induce certain target-author attributes by incorporating continuous multi-dimensional lexical preferences of an author into generative language models. We introduce rewarding strategies in a reinforcement learning framework that encourages the use of words across multiple categorical dimensions, to varying extents. Our experiments demonstrate that the proposed approach can generate text that distinctively aligns with a given target author’s lexical style. We conduct quantitative and qualitative comparisons with competitive and relevant baselines to illustrate the benefits of the proposed approach.

As an essential component of task-oriented dialogue systems, Dialogue State Tracking (DST) takes charge of estimating user intentions and requests in dialogue contexts and extracting substantial goals (states) from user utterances to help the downstream modules to determine the next actions of dialogue systems. For practical usages, a major challenge to constructing a robust DST model is to process a conversation with multi-domain states. However, most existing approaches trained DST on a single domain independently, ignoring the information across domains. To tackle the multi-domain DST task, we first construct a dialogue state graph to transfer structured features among related domain-slot pairs across domains. Then, we encode the graph information of dialogue states by graph convolutional networks and utilize a hard copy mechanism to directly copy historical states from the previous conversation. Experimental results show that our model improves the performances of the multi-domain DST baseline (TRADE) with the absolute joint accuracy of 2.0% and 1.0% on the MultiWOZ 2.0 and 2.1 dialogue datasets, respectively.

Neural network (NN) based data2text models achieve state-of-the-art (SOTA) performance in most metrics, but they sometimes drop or modify the information in the input, and it is hard to control the generation contents. Moreover, it requires paired training data that are usually expensive to collect. Template-based methods have good fidelity and controllability but require heavy human involvement. We propose a novel template-based data2text system powered by a text stitch model. It ensures fidelity and controllability by using templates to produce the main contents. In addition, it reduces human involvement in template design by using a text stitch model to automatically stitch adjacent template units, which is a step that usually requires careful template design and limits template reusability. The text stitch model can be trained in self-supervised fashion, which only requires free texts. The experiments on a benchmark dataset show that our system outperforms SOTA NN-based systems in fidelity and surpasses template-based systems in diversity and human involvement.

Accurate detection of emotions in user- generated text was shown to have several applications for e-commerce, public well-being, and disaster management. Currently, the state-of-the-art performance for emotion detection in text is obtained using complex, deep learning models trained on domain-specific, labeled data. In this paper, we propose ESTeR , an unsupervised model for identifying emotions using a novel similarity function based on random walks on graphs. Our model combines large-scale word co-occurrence information with word-associations from lexicons avoiding not only the dependence on labeled datasets, but also an explicit mapping of words to latent spaces used in emotion-enriched word embeddings. Our similarity function can also be computed efficiently. We study a range of datasets including recent tweets related to COVID-19 to illustrate the superior performance of our model and report insights on public emotions during the on-going pandemic.

We present PhoBERT with two versions, PhoBERT-base and PhoBERT-large, the first public large-scale monolingual language models pre-trained for Vietnamese. Experimental results show that PhoBERT consistently outperforms the recent best pre-trained multilingual model XLM-R (Conneau et al., 2020) and improves the state-of-the-art in multiple Vietnamese-specific NLP tasks including Part-of-speech tagging, Dependency parsing, Named-entity recognition and Natural language inference. We release PhoBERT to facilitate future research and downstream applications for Vietnamese NLP. Our PhoBERT models are available at https://github.com/VinAIResearch/PhoBERT

Existing question answering datasets focus on dealing with homogeneous information, based either only on text or KB/Table information alone. However, as human knowledge is distributed over heterogeneous forms, using homogeneous information alone might lead to severe coverage problems. To fill in the gap, we present HybridQA, a new large-scale question-answering dataset that requires reasoning on heterogeneous information. Each question is aligned with a Wikipedia table and multiple free-form corpora linked with the entities in the table. The questions are designed to aggregate both tabular information and text information, i.e., lack of either form would render the question unanswerable. We test with three different models: 1) a table-only model. 2) text-only model. 3) a hybrid model that combines heterogeneous information to find the answer. The experimental results show that the EM scores obtained by two baselines are below 20%, while the hybrid model can achieve an EM over 40%. This gap suggests the necessity to aggregate heterogeneous information in HybridQA. However, the hybrid model’s score is still far behind human performance. Hence, HybridQA can serve as a challenging benchmark to study question answering with heterogeneous information.

Recent advances in commonsense reasoning depend on large-scale human-annotated training sets to achieve peak performance. However, manual curation of training sets is expensive and has been shown to introduce annotation artifacts that neural models can readily exploit and overfit to. We propose a novel generative data augmentation technique, G-DAUGˆC, that aims to achieve more accurate and robust learning in a low-resource setting. Our approach generates synthetic examples using pretrained language models and selects the most informative and diverse set of examples for data augmentation. On experiments with multiple commonsense reasoning benchmarks, G-DAUGˆC consistently outperforms existing data augmentation methods based on back-translation, establishing a new state-of-the-art on WinoGrande, CODAH, and CommonsenseQA, as well as enhances out-of-distribution generalization, proving to be robust against adversaries or perturbations. Our analysis demonstrates that G-DAUGˆC produces a diverse set of fluent training examples, and that its selection and training approaches are important for performance.

Modeling the parser state is key to good performance in transition-based parsing. Recurrent Neural Networks considerably improved the performance of transition-based systems by modelling the global state, e.g. stack-LSTM parsers, or local state modeling of contextualized features, e.g. Bi-LSTM parsers. Given the success of Transformer architectures in recent parsing systems, this work explores modifications of the sequence-to-sequence Transformer architecture to model either global or local parser states in transition-based parsing. We show that modifications of the cross attention mechanism of the Transformer considerably strengthen performance both on dependency and Abstract Meaning Representation (AMR) parsing tasks, particularly for smaller models or limited training data.

Motion recognition is one of the basic cognitive capabilities of many life forms, yet identifying motion of physical entities in natural language have not been explored extensively and empirically. We present the Literal-Motion-in-Text (LiMiT) dataset, a large human-annotated collection of English text sentences describing physical occurrence of motion, with annotated physical entities in motion. We describe the annotation process for the dataset, analyze its scale and diversity, and report results of several baseline models. We also present future research directions and applications of the LiMiT dataset and share it publicly as a new resource for the research community.