EMNLP 2021 Papers

Traditional hand-crafted linguistically-informed features have often been used for distinguishing between translated and original non-translated texts. By contrast, to date, neural architectures without manual feature engineering have been less explored for this task. In this work, we (i) compare the traditional feature-engineering-based approach to the feature-learning-based one and (ii) analyse the neural architectures in order to investigate how well the hand-crafted features explain the variance in the neural models’ predictions. We use pre-trained neural word embeddings, as well as several end-to-end neural architectures in both monolingual and multilingual settings and compare them to feature-engineering-based SVM classifiers. We show that (i) neural architectures outperform other approaches by more than 20 accuracy points, with the BERT-based model performing the best in both the monolingual and multilingual settings; (ii) while many individual hand-crafted translationese features correlate with neural model predictions, feature importance analysis shows that the most important features for neural and classical architectures differ; and (iii) our multilingual experiments provide empirical evidence for translationese universals across languages.

As an important task in opinion mining, comparative opinion mining aims to identify comparative sentences from product reviews, extract the comparative elements, and obtain the corresponding comparative opinion tuples. However, most previous studies simply regarded comparative tuple extraction as comparative element extraction, but ignored the fact that many comparative sentences may contain multiple comparisons. The comparative opinion tuples defined in these studies also failed to explicitly provide comparative preferences. To address these limitations, in this work we first introduce a new Comparative Opinion Quintuple Extraction (COQE) task, to identify comparative sentences from product reviews and extract all comparative opinion quintuples (Subject, Object, Comparative Aspect, Comparative Opinion, Comparative Preference). Secondly, based on the existing comparative opinion mining corpora, we make supplementary annotations and construct three datasets for the COQE task. Finally, we benchmark the COQE task by proposing a new BERT-based multi-stage approach as well as three baseline systems extended from previous methods. %The new approach significantly outperforms three baseline systems on three datasets and represents a strong benchmark for COQE. Experimental results show that the new approach significantly outperforms three baseline systems on three datasets for the COQE task.

In this paper, we introduce the task of political coalition signal prediction from text, that is, the task of recognizing from the news coverage leading up to an election the (un)willingness of political parties to form a government coalition. We decompose our problem into two related, but distinct tasks: (i) predicting whether a reported statement from a politician or a journalist refers to a potential coalition and (ii) predicting the polarity of the signal – namely, whether the speaker is in favour of or against the coalition. For this, we explore the benefits of multi-task learning and investigate which setup and task formulation is best suited for each sub-task. We evaluate our approach, based on hand-coded newspaper articles, covering elections in three countries (Ireland, Germany, Austria) and two languages (English, German). Our results show that the multi-task learning approach can further improve results over a strong monolingual transfer learning baseline.

Curriculum learning, a machine training strategy that feeds training instances to the model from easy to hard, has been proven to facilitate the dialogue generation task. Meanwhile, knowledge distillation, a knowledge transformation methodology among teachers and students networks can yield significant performance boost for student models. Hence, in this paper, we introduce a combination of curriculum learning and knowledge distillation for efficient dialogue generation models, where curriculum learning can help knowledge distillation from data and model aspects. To start with, from the data aspect, we cluster the training cases according to their complexity, which is calculated by various types of features such as sentence length and coherence between dialog pairs. Furthermore, we employ an adversarial training strategy to identify the complexity of cases from model level. The intuition is that, if a discriminator can tell the generated response is from the teacher or the student, then the case is difficult that the student model has not adapted to yet. Finally, we use self-paced learning, which is an extension to curriculum learning to assign weights for distillation. In conclusion, we arrange a hierarchical curriculum based on the above two aspects for the student model under the guidance from the teacher model. Experimental results demonstrate that our methods achieve improvements compared with competitive baselines.

Recent works have found evidence of gender bias in models of machine translation and coreference resolution using mostly synthetic diagnostic datasets. While these quantify bias in a controlled experiment, they often do so on a small scale and consist mostly of artificial, out-of-distribution sentences. In this work, we find grammatical patterns indicating stereotypical and non-stereotypical gender-role assignments (e.g., female nurses versus male dancers) in corpora from three domains, resulting in a first large-scale gender bias dataset of 108K diverse real-world English sentences. We manually verify the quality of our corpus and use it to evaluate gender bias in various coreference resolution and machine translation models. We find that all tested models tend to over-rely on gender stereotypes when presented with natural inputs, which may be especially harmful when deployed in commercial systems. Finally, we show that our dataset lends itself to finetuning a coreference resolution model, finding it mitigates bias on a held out set. Our dataset and models are publicly available at github.com/SLAB-NLP/BUG. We hope they will spur future research into gender bias evaluation mitigation techniques in realistic settings.

Unsupervised text style transfer aims to alter the underlying style of the text to a desired value while keeping its style-independent semantics, without the support of parallel training corpora. Existing methods struggle to achieve both high style conversion rate and low content loss, exhibiting the over-transfer and under-transfer problems. We attribute these problems to the conflicting driving forces of the style conversion goal and content preservation goal. In this paper, we propose a collaborative learning framework for unsupervised text style transfer using a pair of bidirectional decoders, one decoding from left to right while the other decoding from right to left. In our collaborative learning mechanism, each decoder is regularized by knowledge from its peer which has a different knowledge acquisition process. The difference is guaranteed by their opposite decoding directions and a distinguishability constraint. As a result, mutual knowledge distillation drives both decoders to a better optimum and alleviates the over-transfer and under-transfer problems. Experimental results on two benchmark datasets show that our framework achieves strong empirical results on both style compatibility and content preservation.

Pre-trained models for Natural Languages (NL) like BERT and GPT have been recently shown to transfer well to Programming Languages (PL) and largely benefit a broad set of code-related tasks. Despite their success, most current methods either rely on an encoder-only (or decoder-only) pre-training that is suboptimal for generation (resp. understanding) tasks or process the code snippet in the same way as NL, neglecting the special characteristics of PL such as token types. We present CodeT5, a unified pre-trained encoder-decoder Transformer model that better leverages the code semantics conveyed from the developer-assigned identifiers. Our model employs a unified framework to seamlessly support both code understanding and generation tasks and allows for multi-task learning. Besides, we propose a novel identifier-aware pre-training task that enables the model to distinguish which code tokens are identifiers and to recover them when they are masked. Furthermore, we propose to exploit the user-written code comments with a bimodal dual generation task for better NL-PL alignment. Comprehensive experiments show that CodeT5 significantly outperforms prior methods on understanding tasks such as code defect detection and clone detection, and generation tasks across various directions including PL-NL, NL-PL, and PL-PL. Further analysis reveals that our model can better capture semantic information from code. Our code and pre-trained models are released at https://github.com/salesforce/CodeT5.

We propose CodeQA, a free-form question answering dataset for the purpose of source code comprehension: given a code snippet and a question, a textual answer is required to be generated. CodeQA contains a Java dataset with 119,778 question-answer pairs and a Python dataset with 70,085 question-answer pairs. To obtain natural and faithful questions and answers, we implement syntactic rules and semantic analysis to transform code comments into question-answer pairs. We present the construction process and conduct systematic analysis of our dataset. Experiment results achieved by several neural baselines on our dataset are shown and discussed. While research on question-answering and machine reading comprehension develops rapidly, few prior work has drawn attention to code question answering. This new dataset can serve as a useful research benchmark for source code comprehension.

Spoken dialogue systems need to be able to handle both multiple languages and multilinguality inside a conversation (e.g in case of code-switching). In this work, we introduce new pretraining losses tailored to learn generic multilingual spoken dialogue representations. The goal of these losses is to expose the model to code-switched language. In order to scale up training, we automatically build a pretraining corpus composed of multilingual conversations in five different languages (French, Italian, English, German and Spanish) from OpenSubtitles, a huge multilingual corpus composed of 24.3G tokens. We test the generic representations on MIAM, a new benchmark composed of five dialogue act corpora on the same aforementioned languages as well as on two novel multilingual tasks (i.e multilingual mask utterance retrieval and multilingual inconsistency identification). Our experiments show that our new losses achieve a better performance in both monolingual and multilingual settings.

Existing relation extraction (RE) methods typically focus on extracting relational facts between entity pairs within single sentences or documents. However, a large quantity of relational facts in knowledge bases can only be inferred across documents in practice. In this work, we present the problem of cross-document RE, making an initial step towards knowledge acquisition in the wild. To facilitate the research, we construct the first human-annotated cross-document RE dataset CodRED. Compared to existing RE datasets, CodRED presents two key challenges: Given two entities, (1) it requires finding the relevant documents that can provide clues for identifying their relations; (2) it requires reasoning over multiple documents to extract the relational facts. We conduct comprehensive experiments to show that CodRED is challenging to existing RE methods including strong BERT-based models.

Existing text classification methods mainly focus on a fixed label set, whereas many real-world applications require extending to new fine-grained classes as the number of samples per label increases. To accommodate such requirements, we introduce a new problem called coarse-to-fine grained classification, which aims to perform fine-grained classification on coarsely annotated data. Instead of asking for new fine-grained human annotations, we opt to leverage label surface names as the only human guidance and weave in rich pre-trained generative language models into the iterative weak supervision strategy. Specifically, we first propose a label-conditioned fine-tuning formulation to attune these generators for our task. Furthermore, we devise a regularization objective based on the coarse-fine label constraints derived from our problem setting, giving us even further improvements over the prior formulation. Our framework uses the fine-tuned generative models to sample pseudo-training data for training the classifier, and bootstraps on real unlabeled data for model refinement. Extensive experiments and case studies on two real-world datasets demonstrate superior performance over SOTA zero-shot classification baselines.

Large-Scale Multi-Label Text Classification (LMTC) includes tasks with hierarchical label spaces, such as automatic assignment of ICD-9 codes to discharge summaries. Performance of models in prior art is evaluated with standard precision, recall, and F1 measures without regard for the rich hierarchical structure. In this work we argue for hierarchical evaluation of the predictions of neural LMTC models. With the example of the ICD-9 ontology we describe a structural issue in the representation of the structured label space in prior art, and propose an alternative representation based on the depth of the ontology. We propose a set of metrics for hierarchical evaluation using the depth-based representation. We compare the evaluation scores from the proposed metrics with previously used metrics on prior art LMTC models for ICD-9 coding in MIMIC-III. We also propose further avenues of research involving the proposed ontological representation.

Knowledge-grounded dialogue generation has achieved promising performance with the engagement of external knowledge sources. Typical approaches towards this task usually perform relatively independent two sub-tasks, i.e., knowledge selection and knowledge-aware response generation. In this paper, in order to improve the diversity of both knowledge selection and knowledge-aware response generation, we propose a collaborative latent variable (CoLV) model to integrate these two aspects simultaneously in separate yet collaborative latent spaces, so as to capture the inherent correlation between knowledge selection and response generation. During generation, our proposed model firstly draws knowledge candidate from the latent space conditioned on the dialogue context, and then samples a response from another collaborative latent space conditioned on both the context and the selected knowledge. Experimental results on two widely-used knowledge-grounded dialogue datasets show that our model outperforms previous methods on both knowledge selection and response generation.

Shared tasks have a long history and have become the mainstream of NLP research. Most of the shared tasks require participants to submit only system outputs and descriptions. It is uncommon for the shared task to request submission of the system itself because of the license issues and implementation differences. Therefore, many systems are abandoned without being used in real applications or contributing to better systems. In this research, we propose a scheme to utilize all those systems which participated in the shared tasks. We use all participated system outputs as task teachers in this scheme and develop a new model as a student aiming to learn the characteristics of each system. We call this scheme “Co-Teaching.” This scheme creates a unified system that performs better than the task’s single best system. It only requires the system outputs, and slightly extra effort is needed for the participants and organizers. We apply this scheme to the “SHINRA2019-JP” shared task, which has nine participants with various output accuracies, confirming that the unified system outperforms the best system. Moreover, the code used in our experiments has been released.

Substantial amounts of work are required to clean large collections of digitized books for NLP analysis, both because of the presence of errors in the scanned text and the presence of duplicate volumes in the corpora. In this paper, we consider the issue of deduplication in the presence of optical character recognition (OCR) errors. We present methods to handle these errors, evaluated on a collection of 19,347 texts from the Project Gutenberg dataset and 96,635 texts from the HathiTrust Library. We demonstrate that improvements in language models now enable the detection and correction of OCR errors without consideration of the scanning image itself. The inconsistencies found by aligning pairs of scans of the same underlying work provides training data to build models for detecting and correcting errors. We identify the canonical version for each of 17,136 repeatedly-scanned books from 58,808 scans. Finally, we investigate methods to detect and correct errors in single-copy texts. We show that on average, our method corrects over six times as many errors as it introduces. We also provide interesting analysis on the relation between scanning quality and other factors such as location and publication year.

Contracts are a common type of legal document that frequent in several day-to-day business workflows. However, there has been very limited NLP research in processing such documents, and even lesser in generating them. These contracts are made up of clauses, and the unique nature of these clauses calls for specific methods to understand and generate such documents. In this paper, we introduce the task of clause recommendation, as a first step to aid and accelerate the authoring of contract documents. We propose a two-staged pipeline to first predict if a specific clause type is relevant to be added in a contract, and then recommend the top clauses for the given type based on the contract context. We pre-train BERT on an existing library of clauses with two additional tasks and use it for our prediction and recommendation. We experiment with classification methods and similarity-based heuristics for clause relevance prediction, and generation-based methods for clause recommendation, and evaluate the results from various methods on several clause types. We provide analyses on the results, and further outline the limitations and future directions of this line of research.

Understanding the origins of militarized conflict is a complex, yet important undertaking. Existing research seeks to build this understanding by considering bi-lateral relationships between entity pairs (dyadic causes) and multi-lateral relationships among multiple entities (systemic causes). The aim of this work is to compare these two causes in terms of how they correlate with conflict between two entities. We do this by devising a set of textual and graph-based features which represent each of the causes. The features are extracted from Wikipedia and modeled as a large graph. Nodes in this graph represent entities connected by labeled edges representing ally or enemy-relationships. This allows casting the problem as an edge classification task, which we term dyad classification. We propose and evaluate classifiers to determine if a particular pair of entities are allies or enemies. Our results suggest that our systemic features might be slightly better correlates of conflict. Further, we find that Wikipedia articles of allies are semantically more similar than enemies.

Sentence-level Quality estimation (QE) of machine translation is traditionally formulated as a regression task, and the performance of QE models is typically measured by Pearson correlation with human labels. Recent QE models have achieved previously-unseen levels of correlation with human judgments, but they rely on large multilingual contextualized language models that are computationally expensive and make them infeasible for real-world applications. In this work, we evaluate several model compression techniques for QE and find that, despite their popularity in other NLP tasks, they lead to poor performance in this regression setting. We observe that a full model parameterization is required to achieve SoTA results in a regression task. However, we argue that the level of expressiveness of a model in a continuous range is unnecessary given the downstream applications of QE, and show that reframing QE as a classification problem and evaluating QE models using classification metrics would better reflect their actual performance in real-world applications.

We consider the hierarchical representation of documents as graphs and use geometric deep learning to classify them into different categories. While graph neural networks can efficiently handle the variable structure of hierarchical documents using the permutation invariant message passing operations, we show that we can gain extra performance improvements using our proposed selective graph pooling operation that arises from the fact that some parts of the hierarchy are invariable across different documents. We applied our model to classify clinical trial (CT) protocols into completed and terminated categories. We use bag-of-words based, as well as pre-trained transformer-based embeddings to featurize the graph nodes, achieving f1-scoresaround 0.85 on a publicly available large scale CT registry of around 360K protocols. We further demonstrate how the selective pooling can add insights into the CT termination status prediction. We make the source code and dataset splits accessible.