EMNLP 2021 Papers

In this paper, we present a new verification style reading comprehension dataset named VGaokao from Chinese Language tests of Gaokao. Different from existing efforts, the new dataset is originally designed for native speakers’ evaluation, thus requiring more advanced language understanding skills. To address the challenges in VGaokao, we propose a novel Extract-Integrate-Compete approach, which iteratively selects complementary evidence with a novel query updating mechanism and adaptively distills supportive evidence, followed by a pairwise competition to push models to learn the subtle difference among similar text pieces. Experiments show that our methods outperform various baselines on VGaokao with retrieved complementary evidence, while having the merits of efficiency and explainability. Our dataset and code are released for further research.

Prior studies on text-to-text generation typically assume that the model could figure out what to attend to in the input and what to include in the output via seq2seq learning, with only the parallel training data and no additional guidance. However, it remains unclear whether current models can preserve important concepts in the source input, as seq2seq learning does not have explicit focus on the concepts and commonly used evaluation metrics also treat them equally important as other tokens. In this paper, we present a systematic analysis that studies whether current seq2seq models, especially pre-trained language models, are good enough for preserving important input concepts and to what extent explicitly guiding generation with the concepts as lexical constraints is beneficial. We answer the above questions by conducting extensive analytical experiments on four representative text-to-text generation tasks. Based on the observations, we then propose a simple yet effective framework to automatically extract, denoise, and enforce important input concepts as lexical constraints. This new method performs comparably or better than its unconstrained counterpart on automatic metrics, demonstrates higher coverage for concept preservation, and receives better ratings in the human evaluation. Our code is available at https://github.com/morningmoni/EDE.

Deriving and modifying graphs from natural language text has become a versatile basis technology for information extraction with applications in many subfields, such as semantic parsing or knowledge graph construction. A recent work used this technique for modifying scene graphs (He et al. 2020), by first encoding the original graph and then generating the modified one based on this encoding. In this work, we show that we can considerably increase performance on this problem by phrasing it as graph extension instead of graph generation. We propose the first model for the resulting graph extension problem based on autoregressive sequence labelling. On three scene graph modification data sets, this formulation leads to improvements in accuracy over the state-of-the-art between 13 and 24 percentage points. Furthermore, we introduce a novel data set from the biomedical domain which has much larger linguistic variability and more complex graphs than the scene graph modification data sets. For this data set, the state-of-the art fails to generalize, while our model can produce meaningful predictions.

Exposure bias has been regarded as a central problem for auto-regressive language models (LM). It claims that teacher forcing would cause the test-time generation to be incrementally distorted due to the training-generation discrepancy. Although a lot of algorithms have been proposed to avoid teacher forcing and therefore alleviate exposure bias, there is little work showing how serious the exposure bias problem actually is. In this work, we focus on the task of open-ended language generation, propose metrics to quantify the impact of exposure bias in the aspects of quality, diversity, and consistency. Our key intuition is that if we feed ground-truth data prefixes (instead of prefixes generated by the model itself) into the model and ask it to continue the generation, the performance should become much better because the training-generation discrepancy in the prefix is removed. Both automatic and human evaluations are conducted in our experiments. On the contrary to the popular belief in exposure bias, we find that the the distortion induced by the prefix discrepancy is limited, and does not seem to be incremental during the generation. Moreover, our analysis reveals an interesting self-recovery ability of the LM, which we hypothesize to be countering the harmful effects from exposure bias.

Several studies have been carried out on revealing linguistic features captured by BERT. This is usually achieved by training a diagnostic classifier on the representations obtained from different layers of BERT. The subsequent classification accuracy is then interpreted as the ability of the model in encoding the corresponding linguistic property. Despite providing insights, these studies have left out the potential role of token representations. In this paper, we provide a more in-depth analysis on the representation space of BERT in search for distinct and meaningful subspaces that can explain the reasons behind these probing results. Based on a set of probing tasks and with the help of attribution methods we show that BERT tends to encode meaningful knowledge in specific token representations (which are often ignored in standard classification setups), allowing the model to detect syntactic and semantic abnormalities, and to distinctively separate grammatical number and tense subspaces.

Adverse Events (AE) are harmful events resulting from the use of medical products. Although social media may be crucial for early AE detection, the sheer scale of this data makes it logistically intractable to analyze using human agents, with NLP representing the only low-cost and scalable alternative. In this paper, we frame AE Detection and Extraction as a sequence-to-sequence problem using the T5 model architecture and achieve strong performance improvements over the baselines on several English benchmarks (F1 = 0.71, 12.7% relative improvement for AE Detection; Strict F1 = 0.713, 12.4% relative improvement for AE Extraction). Motivated by the strong commonalities between AE tasks, the class imbalance in AE benchmarks, and the linguistic and structural variety typical of social media texts, we propose a new strategy for multi-task training that accounts, at the same time, for task and dataset characteristics. Our approach increases model robustness, leading to further performance gains. Finally, our framework shows some language transfer capabilities, obtaining higher performance than Multilingual BERT in zero-shot learning on French data.

Recently, there has been significant progress in studying neural networks for translating text descriptions into SQL queries under the zero-shot cross-domain setting. Despite achieving good performance on some public benchmarks, we observe that existing text-to-SQL models do not generalize when facing domain knowledge that does not frequently appear in the training data, which may render the worse prediction performance for unseen domains. In this work, we investigate the robustness of text-to-SQL models when the questions require rarely observed domain knowledge. In particular, we define five types of domain knowledge and introduce Spider-DK (DK is the abbreviation of domain knowledge), a human-curated dataset based on the Spider benchmark for text-to-SQL translation. NL questions in Spider-DK are selected from Spider, and we modify some samples by adding domain knowledge that reflects real-world question paraphrases. We demonstrate that the prediction accuracy dramatically drops on samples that require such domain knowledge, even if the domain knowledge appears in the training set, and the model provides the correct predictions for related training samples.

Few-shot relation extraction (FSRE) focuses on recognizing novel relations by learning with merely a handful of annotated instances. Meta-learning has been widely adopted for such a task, which trains on randomly generated few-shot tasks to learn generic data representations. Despite impressive results achieved, existing models still perform suboptimally when handling hard FSRE tasks, where the relations are fine-grained and similar to each other. We argue this is largely because existing models do not distinguish hard tasks from easy ones in the learning process. In this paper, we introduce a novel approach based on contrastive learning that learns better representations by exploiting relation label information. We further design a method that allows the model to adaptively learn how to focus on hard tasks. Experiments on two standard datasets demonstrate the effectiveness of our method.

Commonsense reasoning benchmarks have been largely solved by fine-tuning language models. The downside is that fine-tuning may cause models to overfit to task-specific data and thereby forget their knowledge gained during pre-training. Recent works only propose lightweight model updates as models may already possess useful knowledge from past experience, but a challenge remains in understanding what parts and to what extent models should be refined for a given task. In this paper, we investigate what models learn from commonsense reasoning datasets. We measure the impact of three different adaptation methods on the generalization and accuracy of models. Our experiments with two models show that fine-tuning performs best, by learning both the content and the structure of the task, but suffers from overfitting and limited generalization to novel answers. We observe that alternative adaptation methods like prefix-tuning have comparable accuracy, but generalize better to unseen answers and are more robust to adversarial splits.

Document-level event extraction is critical to various natural language processing tasks for providing structured information. Existing approaches by sequential modeling neglect the complex logic structures for long texts. In this paper, we leverage the entity interactions and sentence interactions within long documents and transform each document into an undirected unweighted graph by exploiting the relationship between sentences. We introduce the Sentence Community to represent each event as a subgraph. Furthermore, our framework SCDEE maintains the ability to extract multiple events by sentence community detection using graph attention networks and alleviate the role overlapping issue by predicting arguments in terms of roles. Experiments demonstrate that our framework achieves competitive results over state-of-the-art methods on the large-scale document-level event extraction dataset.

In this paper, we explore Non-AutoRegressive (NAR) decoding for unsupervised text style transfer. We first propose a base NAR model by directly adapting the common training scheme from its AutoRegressive (AR) counterpart. Despite the faster inference speed over the AR model, this NAR model sacrifices its transfer performance due to the lack of conditional dependence between output tokens. To this end, we investigate three techniques, i.e., knowledge distillation, contrastive learning, and iterative decoding, for performance enhancement. Experimental results on two benchmark datasets suggest that, although the base NAR model is generally inferior to AR decoding, their performance gap can be clearly narrowed when empowering NAR decoding with knowledge distillation, contrastive learning, and iterative decoding.

This paper explores the effect of using multitask learning for abstractive summarization in the context of small training corpora. In particular, we incorporate four different tasks (extractive summarization, language modeling, concept detection, and paraphrase detection) both individually and in combination, with the goal of enhancing the target task of abstractive summarization via multitask learning. We show that for many task combinations, a model trained in a multitask setting outperforms a model trained only for abstractive summarization, with no additional summarization data introduced. Additionally, we do a comprehensive search and find that certain tasks (e.g. paraphrase detection) consistently benefit abstractive summarization, not only when combined with other tasks but also when using different architectures and training corpora.

The task of generating explanatory notes for language learners is known as feedback comment generation. Although various generation techniques are available, little is known about which methods are appropriate for this task. Nagata (2019) demonstrates the effectiveness of neural-retrieval-based methods in generating feedback comments for preposition use. Retrieval-based methods have limitations in that they can only output feedback comments existing in a given training data. Furthermore, feedback comments can be made on other grammatical and writing items than preposition use, which is still unaddressed. To shed light on these points, we investigate a wider range of methods for generating many feedback comments in this study. Our close analysis of the type of task leads us to investigate three different architectures for comment generation: (i) a neural-retrieval-based method as a baseline, (ii) a pointer-generator-based generation method as a neural seq2seq method, (iii) a retrieve-and-edit method, a hybrid of (i) and (ii). Intuitively, the pointer-generator should outperform neural-retrieval, and retrieve-and-edit should perform best. However, in our experiments, this expectation is completely overturned. We closely analyze the results to reveal the major causes of these counter-intuitive results and report on our findings from the experiments.

Fact verification based on structured data is challenging as it requires models to understand both natural language and symbolic operations performed over tables. Although pre-trained language models have demonstrated a strong capability in verifying simple statements, they struggle with complex statements that involve multiple operations. In this paper, we improve fact verification by decomposing complex statements into simpler subproblems. Leveraging the programs synthesized by a weakly supervised semantic parser, we propose a program-guided approach to constructing a pseudo dataset for decomposition model training. The subproblems, together with their predicted answers, serve as the intermediate evidence to enhance our fact verification model. Experiments show that our proposed approach achieves the new state-of-the-art performance, an 82.7% accuracy, on the TabFact benchmark.

Relative position embedding (RPE) is a successful method to explicitly and efficaciously encode position information into Transformer models. In this paper, we investigate the potential problems in Shaw-RPE and XL-RPE, which are the most representative and prevalent RPEs, and propose two novel RPEs called Low-level Fine-grained High-level Coarse-grained (LFHC) RPE and Gaussian Cumulative Distribution Function (GCDF) RPE. LFHC-RPE is an improvement of Shaw-RPE, which enhances the perception ability at medium and long relative positions. GCDF-RPE utilizes the excellent properties of the Gaussian function to amend the prior encoding mechanism in XL-RPE. Experimental results on nine authoritative datasets demonstrate the effectiveness of our methods empirically. Furthermore, GCDF-RPE achieves the best overall performance among five different RPEs.

Semantic sentence embeddings are usually supervisedly built minimizing distances between pairs of embeddings of sentences labelled as semantically similar by annotators. Since big labelled datasets are rare, in particular for non-English languages, and expensive, recent studies focus on unsupervised approaches that require not-paired input sentences. We instead propose a language-independent approach to build large datasets of pairs of informal texts weakly similar, without manual human effort, exploiting Twitter’s intrinsic powerful signals of relatedness: replies and quotes of tweets. We use the collected pairs to train a Transformer model with triplet-like structures, and we test the generated embeddings on Twitter NLP similarity tasks (PIT and TURL) and STSb. We also introduce four new sentence ranking evaluation benchmarks of informal texts, carefully extracted from the initial collections of tweets, proving not only that our best model learns classical Semantic Textual Similarity, but also excels on tasks where pairs of sentences are not exact paraphrases. Ablation studies reveal how increasing the corpus size influences positively the results, even at 2M samples, suggesting that bigger collections of Tweets still do not contain redundant information about semantic similarities. Code available at https://github.com/marco-digio/Twitter4SSE

We propose a novel Chain Guided Retriever-reader (CGR) framework to model the reasoning chain for multi-hop Science Question Answering. Our framework is capable of performing explainable reasoning without the need of any corpus-specific annotations, such as the ground-truth reasoning chain, or human-annotated entity mentions. Specifically, we first generate reasoning chains from a semantic graph constructed by Abstract Meaning Representation of retrieved evidence facts. A Chain-aware loss, concerning both local and global chain information, is also designed to enable the generated chains to serve as distant supervision signals for training the retriever, where reinforcement learning is also adopted to maximize the utility of the reasoning chains. Our framework allows the retriever to capture step-by-step clues of the entire reasoning process, which is not only shown to be effective on two challenging multi-hop Science QA tasks, namely OpenBookQA and ARC-Challenge, but also favors explainability.

Back-translation (BT) has become one of the de facto components in unsupervised neural machine translation (UNMT), and it explicitly makes UNMT have translation ability. However, all the pseudo bi-texts generated by BT are treated equally as clean data during optimization without considering the quality diversity, leading to slow convergence and limited translation performance. To address this problem, we propose a curriculum learning method to gradually utilize pseudo bi-texts based on their quality from multiple granularities. Specifically, we first apply crosslingual word embedding to calculate the potential translation difficulty (quality) for the monolingual sentences. Then, the sentences are fed into UNMT from easy to hard batch by batch. Furthermore, considering the quality of sentences/tokens in a particular batch are also diverse, we further adopt the model itself to calculate the fine-grained quality scores, which are served as learning factors to balance the contributions of different parts when computing loss and encourage the UNMT model to focus on pseudo data with higher quality. Experimental results on WMT 14 En-Fr, WMT 14 En-De, WMT 16 En-Ro, and LDC En-Zh translation tasks demonstrate that the proposed method achieves consistent improvements with faster convergence speed.

Our goal, in the context of open-domain textual question-answering (QA), is to explain answers by showing the line of reasoning from what is known to the answer, rather than simply showing a fragment of textual evidence (a “rationale”). If this could be done, new opportunities for understanding and debugging the system’s reasoning become possible. Our approach is to generate explanations in the form of entailment trees, namely a tree of multipremise entailment steps from facts that are known, through intermediate conclusions, to the hypothesis of interest (namely the question + answer). To train a model with this skill, we created ENTAILMENTBANK, the first dataset to contain multistep entailment trees. Given a hypothesis (question + answer), we define three increasingly difficult explanation tasks: generate a valid entailment tree given (a) all relevant sentences (b) all relevant and some irrelevant sentences, or (c) a corpus. We show that a strong language model can partially solve these tasks, in particular when the relevant sentences are included in the input (e.g., 35% of trees for (a) are perfect), and with indications of generalization to other domains. This work is significant as it provides a new type of dataset (multistep entailments) and baselines, offering a new avenue for the community to generate richer, more systematic explanations.