COLING 2020 Papers

The release of BERT revolutionized the development of NLP. Various BERT-based reading comprehension models have been proposed, thus updating the performance ranking of reading comprehension tasks. However, the above BERT-based models inherently employ BERT’s combined input method, representing the input question and paragraph as a single packed sequence, without further modification for reading comprehension. This paper makes an in-depth analysis of this input method, proposes a problem of this approach. We call it attention deconcentration. Accordingly, this paper proposes ForceReader, a BERT-based interactive machine reading comprehension model. First, ForceReader proposes a novel solution called the Attention Separation Representation to respond to attention deconcentration. Moreover, starting from the logical nature of reading comprehension tasks, ForceReader adopts Multi-mode Reading and Interactive Reasoning strategy. For the calculation of attention, ForceReader employs Conditional Background Attention to solve the lack of the overall context semantic after the separation of attention. As an integral model, ForceReader shows a significant improvement in reading comprehension tasks compared to BERT. Moreover, this paper makes detailed visual analyses of the attention and propose strategies accordingly. This may be another argument to the explanations of the attention.

Recent work has shown that pre-trained Transformers obtain remarkable performance on many natural language processing tasks including automatic summarization. However, most work has focused on (relatively) data-rich single-document summarization settings. In this paper, we explore highly-abstractive multi-document summarization where the summary is explicitly conditioned on a user-given topic statement or question. We compare the summarization quality produced by three state-of-the-art transformer-based models: BART, T5, and PEGASUS. We report the performance on four challenging summarization datasets: three from the general domain and one from consumer health in both zero-shot and few-shot learning settings. While prior work has shown significant differences in performance for these models on standard summarization tasks, our results indicate that with as few as 10 labeled examples there is no statistically significant difference in summary quality, suggesting the need for more abstractive benchmark collections when determining state-of-the-art.

Recently, leveraging pre-trained Transformer based language models in down stream, task specific models has advanced state of the art results in natural language understanding tasks. However, only a little research has explored the suitability of this approach in low resource settings with less than 1,000 training data points. In this work, we explore fine-tuning methods of BERT - a pre-trained Transformer based language model - by utilizing pool-based active learning to speed up training while keeping the cost of labeling new data constant. Our experimental results on the GLUE data set show an advantage in model performance by maximizing the approximate knowledge gain of the model when querying from the pool of unlabeled data. Finally, we demonstrate and analyze the benefits of freezing layers of the language model during fine-tuning to reduce the number of trainable parameters, making it more suitable for low-resource settings.

Previous work on bridging anaphora recognition (Hou et al., 2013) casts the problem as a subtask of learning fine-grained information status (IS). However, these systems heavily depend on many hand-crafted linguistic features. In this paper, we propose a simple discourse context-aware BERT model for fine-grained IS classification. On the ISNotes corpus (Markert et al., 2012), our model achieves new state-of-the-art performances on fine-grained IS classification, obtaining a 4.8 absolute overall accuracy improvement compared to Hou et al. (2013). More importantly, we also show an improvement of 10.5 F1 points for bridging anaphora recognition without using any complex hand-crafted semantic features designed for capturing the bridging phenomenon. We further analyze the trained model and find that the most attended signals for each IS category correspond well to linguistic notions of information status.

Image text carries essential information to understand the scene and perform reasoning. Text-based visual question answering (text VQA) task focuses on visual questions that require reading text in images. Existing text VQA systems generate an answer by selecting from optical character recognition (OCR) texts or a fixed vocabulary. Positional information of text is underused and there is a lack of evidence for the generated answer. As such, this paper proposes a localization-aware answer prediction network (LaAP-Net) to address this challenge. Our LaAP-Net not only generates the answer to the question but also predicts a bounding box as evidence of the generated answer. Moreover, a context-enriched OCR representation (COR) for multimodal fusion is proposed to facilitate the localization task. Our proposed LaAP-Net outperforms existing approaches on three benchmark datasets for the text VQA task by a noticeable margin.

The recent dominance of machine learning-based natural language processing methods has fostered the culture of overemphasizing model accuracies rather than studying the reasons behind their errors. Interpretability, however, is a critical requirement for many downstream AI and NLP applications, e.g., in finance, healthcare, and autonomous driving. This study, instead of proposing any “new model”, investigates the error patterns of some widely acknowledged sentiment analysis methods in the finance domain. We discover that (1) those methods belonging to the same clusters are prone to similar error patterns, and (2) there are six types of linguistic features that are pervasive in the common errors. These findings provide important clues and practical considerations for improving sentiment analysis models for financial applications.

Unsupervised neural machine translation (NMT) utilizes only monolingual data for training. The quality of back-translated data plays an important role in the performance of NMT systems. In back-translation, all generated pseudo parallel sentence pairs are not of the same quality. Taking inspiration from domain adaptation where in-domain sentences are given more weight in training, in this paper we propose an approach to filter back-translated data as part of the training process of unsupervised NMT. Our approach gives more weight to good pseudo parallel sentence pairs in the back-translation phase. We calculate the weight of each pseudo parallel sentence pair using sentence-wise round-trip BLEU score which is normalized batch-wise. We compare our approach with the current state of the art approaches for unsupervised NMT.

In this paper, we introduce a state-of-the-art pseudo-labeling technique for few-shot intent detection. We devise a folding/unfolding hierarchical clustering algorithm which assigns weighted pseudo-labels to unlabeled user utterances. We show that our two-step method yields significant improvement over existing solutions. This performance is achieved on multiple intent detection datasets, even in more challenging situations where the number of classes is large or when the dataset is highly imbalanced. Moreover, we confirm this results on the more general text classification task. We also demonstrate that our approach nicely complements existing solutions, thereby providing an even stronger state-of-the-art ensemble method.

In this paper, we propose a new few-shot text classification method. Compared with supervised learning methods which require a large corpus of labeled documents, our method aims to make it possible to classify unlabeled text with few labeled data. To achieve this goal, we take advantage of advanced pre-trained language model to extract the semantic features of each document. Furthermore, we utilize an edge-labeling graph neural network to implicitly models the intra-cluster similarity and the inter-cluster dissimilarity of the documents. Finally, we take the results of the graph neural network as the input of a prototypical network to classify the unlabeled texts. We verify the effectiveness of our method on a sentiment analysis dataset and a relation classification dataset and achieve the state-of-the-art performance on both tasks.

Recently, spoken language understanding (SLU) has attracted extensive research interests, and various SLU datasets have been proposed to promote the development. However, most of the existing methods focus on a single individual dataset, the efforts to improve the robustness of models and obtain better performance by combining the merits of various datasets are not well studied. In this paper, we argue that if these SLU datasets are considered together, different knowledge from different datasets could be learned jointly, and there are high chances to promote the performance of each dataset. At the same time, we further attempt to prevent data leakage when unifying multiple datasets which, arguably, is more useful in an industry setting. To this end, we propose a federated learning framework, which could unify various types of datasets as well as tasks to learn and fuse various types of knowledge, i.e., text representations, from different datasets and tasks, without the sharing of downstream task data. The fused text representations merge useful features from different SLU datasets and tasks and are thus much more powerful than the original text representations alone in individual tasks. At last, in order to provide multi-granularity text representations for our framework, we propose a novel Multi-view Encoder (MV-Encoder) as the backbone of our federated learning framework. Experiments on two SLU benchmark datasets, including two tasks (intention detection and slot filling) and federated learning settings (horizontal federated learning, vertical federated learning and federated transfer learning), demonstrate the effectiveness and universality of our approach. Specifically, we are able to get 1.53% improvement on the intent detection metric accuracy. And we could also boost the performance of a strong baseline by up to 5.29% on the slot filling metric F1. Furthermore, by leveraging BERT as an additional encoder, we establish new state-of-the-art results on SNIPS and ATIS datasets, where we get 99.33% and 98.28% in terms of accuracy on intent detection task as well as 97.20% and 96.41% in terms of F1 score on slot filling task, respectively.

Learning on large text corpora, deep neural networks achieve promising results in the next word prediction task. However, deploying these huge models on devices has to deal with constraints of low latency and a small binary size. To address these challenges, we propose a fast word predictor performing efficiently on mobile devices. Compared with a standard neural network which has a similar word prediction rate, the proposed model obtains 60% reduction in memory size and 100X faster inference time on a middle-end mobile device. The method is developed as a feature for a chat application which serves more than 100 million users.

Linking of biomedical entity mentions to various terminologies of chemicals, diseases, genes, adverse drug reactions is a challenging task, often requiring non-syntactic interpretation. A large number of biomedical corpora and state-of-the-art models have been introduced in the past five years. However, there are no general guidelines regarding the evaluation of models on these corpora in single- and cross-terminology settings. In this work, we perform a comparative evaluation of various benchmarks and study the efficiency of state-of-the-art neural architectures based on Bidirectional Encoder Representations from Transformers (BERT) for linking of three entity types across three domains: research abstracts, drug labels, and user-generated texts on drug therapy in English. We have made the source code and results available at https://github.com/insilicomedicine/Fair-Evaluation-BERT.

Recent advancements in self-attention neural network architectures have raised the bar for open-ended text generation. Yet, while current methods are capable of producing a coherent text which is several hundred words long, attaining control over the content that is being generated—as well as evaluating it—are still open questions. We propose a controlled generation task which is based on expanding a sequence of facts, expressed in natural language, into a longer narrative. We introduce human-based evaluation metrics for this task, as well as a method for deriving a large training dataset. We evaluate three methods on this task, based on fine-tuning pre-trained models. We show that while auto-regressive, unidirectional Language Models such as GPT2 produce better fluency, they struggle to adhere to the requested facts. We propose a plan-and-cloze model (using fine-tuned XLNet) which produces competitive fluency while adhering to the requested content.

Most current extractive summarization models generate summaries by selecting salient sentences. However, one of the problems with sentence-level extractive summarization is that there exists a gap between the human-written gold summary and the oracle sentence labels. In this paper, we propose to extract fact-level semantic units for better extractive summarization. We also introduce a hierarchical structure, which incorporates the multi-level of granularities of the textual information into the model. In addition, we incorporate our model with BERT using a hierarchical graph mask. This allows us to combine BERT’s ability in natural language understanding and the structural information without increasing the scale of the model. Experiments on the CNN/DaliyMail dataset show that our model achieves state-of-the-art results.

A 2018 study led by the Media Insight Project showed that most journalists think that a clearmarking of what is news reporting and what is commentary or opinion (e.g., editorial, op-ed)is essential for gaining public trust. We present an approach to classify news articles into newsstories (i.e., reporting of factual information) and opinion pieces using models that aim to sup-plement the article content representation with argumentation features. Our hypothesis is thatthe nature of argumentative discourse is important in distinguishing between news stories andopinion articles. We show that argumentation features outperform linguistic features used previ-ously and improve on fine-tuned transformer-based models when tested on data from publishersunseen in training. Automatically flagging opinion pieces vs. news stories can aid applicationssuch as fact-checking or event extraction.

Recently, pre-trained language models such as BERT have shown state-of-the-art accuracies in text matching. When being applied to IR (or QA), the BERT-based matching models need to online calculate the representations and interactions for all query-candidate pairs. The high inference cost has prohibited the deployments of BERT-based matching models in many practical applications. To address this issue, we propose a novel BERT-based text matching model, in which the representations and the interactions are decoupled. Then, the representations of the candidates can be calculated and stored offline, and directly retrieved during the online matching phase. To conduct the interactions and generate final matching scores, a lightweight attention network is designed. Experiments based on several large scale text matching datasets show that the proposed model, called FASTMATCH, can achieve up to 100X speed-up to BERT and RoBERTa at the online matching phase, while keeping more up to 98.7% of the performance.

In this paper, we propose and experiment with techniques for extreme compression of neural natural language understanding (NLU) models, making them suitable for execution on resource-constrained devices. We propose a task-aware, end-to-end compression approach that performs word-embedding compression jointly with NLU task learning. We show our results on a large-scale, commercial NLU system trained on a varied set of intents with huge vocabulary sizes. Our approach outperforms a range of baselines and achieves a compression rate of 97.4% with less than 3.7% degradation in predictive performance. Our analysis indicates that the signal from the downstream task is important for effective compression with minimal degradation in performance.

Patient adherence is a critical factor in health outcomes. We present a framework to extract adherence information from electronic health records, including both sentence-level information indicating general adherence information (full, partial, none, etc.) and span-level information providing additional information such as adherence type (medication or nonmedication), reasons and outcomes. We annotate and make publicly available a new corpus of 3,000 de-identified sentences, and discuss the language physicians use to document adherence information. We also explore models based on state-of-the-art transformers to automate both tasks.

This paper proposes a general purpose relation extractor that uses Wikidata descriptions to represent the relation’s surface form. The results are tested on the FewRel 1.0 dataset, which provides an excellent framework for training and evaluating the proposed zero-shot learning system in English. This relation extractor architecture exploits the implicit knowledge of a language model through a question-answering approach.