EMNLP 2019 Papers

We propose a fine-grained entity typing model with a novel attention mechanism and a hybrid type classifier. We advance existing methods in two aspects: feature extraction and type prediction. To capture richer contextual information, we adopt contextualized word representations instead of fixed word embeddings used in previous work. In addition, we propose a two-step mention-aware attention mechanism to enable the model to focus on important words in mentions and contexts. We also present a hybrid classification method beyond binary relevance to exploit type inter-dependency with latent type representation. Instead of independently predicting each type, we predict a low-dimensional vector that encodes latent type features and reconstruct the type vector from this latent representation. Experiment results on multiple data sets show that our model significantly advances the state-of-the-art on fine-grained entity typing, obtaining up to 6.1% and 5.5% absolute gains in macro averaged F-score and micro averaged F-score respectively.

The task of Chinese text spam detection is very challenging due to both glyph and phonetic variations of Chinese characters. This paper proposes a novel framework to jointly model Chinese variational, semantic, and contextualized representations for Chinese text spam detection task. In particular, a Variation Family-enhanced Graph Embedding (VFGE) algorithm is designed based on a Chinese character variation graph. The VFGE can learn both the graph embeddings of the Chinese characters (local) and the latent variation families (global). Furthermore, an enhanced bidirectional language model, with a combination gate function and an aggregation learning function, is proposed to integrate the graph and text information while capturing the sequential information. Extensive experiments have been conducted on both SMS and review datasets, to show the proposed method outperforms a series of state-of-the-art models for Chinese spam detection.

Phonological processes are context-dependent sound changes in natural languages. We present an unsupervised approach to learning human-readable descriptions of phonological processes from collections of related utterances. Our approach builds upon a technique from the programming languages community called constraint-based program synthesis. We contribute a novel encoding of the learning problem into Boolean Satisfiability constraints, which enables both data efficiency and fast inference. We evaluate our system on textbook phonology problems and datasets from the literature, and show that it achieves high accuracy at interactive speeds.

We present set to ordered text, a natural language generation task applied to automatically generating discharge instructions from admission ICD (International Classification of Diseases) codes. This task differs from other natural language generation tasks in the following ways: (1) The input is a set of identifiable entities (ICD codes) where the relations between individual entity are not explicitly specified. (2) The output text is not a narrative description (e.g. news articles) composed from the input. Rather, inferences are made from the input (symptoms specified in ICD codes) to generate the output (instructions). (3) There is an optimal order in which each sentence (instruction) should appear in the output. Unlike most other tasks, neither the input (ICD codes) nor their corresponding symptoms appear in the output, so the ordering of the output instructions needs to be learned in an unsupervised fashion. Based on clinical intuition, we hypothesize that each instruction in the output is mapped to a subset of ICD codes specified in the input. We propose a neural architecture that jointly models (a) subset selection: choosing relevant subsets from a set of input entities; (b) content ordering: learning the order of instructions; and (c) text generation: representing the instructions corresponding to the selected subsets in natural language. In addition, we penalize redundancy during beam search to improve tractability for long text generation. Our model outperforms baseline models in BLEU scores and human evaluation. We plan to extend this work to other tasks such as recipe generation from ingredients.

Classical Chinese poetry is a jewel in the treasure house of Chinese culture. Previous poem generation models only allow users to employ keywords to interfere the meaning of generated poems, leaving the dominion of generation to the model. In this paper, we propose a novel task of generating classical Chinese poems from vernacular, which allows users to have more control over the semantic of generated poems. We adapt the approach of unsupervised machine translation (UMT) to our task. We use segmentation-based padding and reinforcement learning to address under-translation and over-translation respectively. According to experiments, our approach significantly improve the perplexity and BLEU compared with typical UMT models. Furthermore, we explored guidelines on how to write the input vernacular to generate better poems. Human evaluation showed our approach can generate high-quality poems which are comparable to amateur poems.

In this paper, we propose a novel framework for sarcasm generation; the system takes a literal negative opinion as input and translates it into a sarcastic version. Our framework does not require any paired data for training. Sarcasm emanates from context-incongruity which becomes apparent as the sentence unfolds. Our framework introduces incongruity into the literal input version through modules that: (a) filter factual content from the input opinion, (b) retrieve incongruous phrases related to the filtered facts and (c) synthesize sarcastic text from the incongruous filtered and incongruous phrases. The framework employs reinforced neural sequence to sequence learning and information retrieval and is trained only using unlabeled non-sarcastic and sarcastic opinions. Since no labeled dataset exists for such a task, for evaluation, we manually prepare a benchmark dataset containing literal opinions and their sarcastic paraphrases. Qualitative and quantitative performance analyses on the data reveal our system’s superiority over baselines built using known unsupervised statistical and neural machine translation and style transfer techniques.

Word embeddings have been widely used to study gender stereotypes in texts. One key problem regarding existing bias scores is to evaluate their validities: do they really reflect true bias levels? For a small set of words (e.g. occupations), we can rely on human annotations or external data. However, for most words, evaluating the correctness of them is still an open problem. In this work, we utilize word association test, which contains rich types of word connections annotated by human participants, to explore how gender stereotypes spread within our minds. Specifically, we use random walk on word association graph to derive bias scores for a large amount of words. Experiments show that these bias scores correlate well with bias in the real world. More importantly, comparing with word-embedding-based bias scores, it provides a different perspective on gender stereotypes in words.

Automatic post-editing (APE), which aims to correct errors in the output of machine translation systems in a post-processing step, is an important task in natural language processing. While recent work has achieved considerable performance gains by using neural networks, how to model the copying mechanism for APE remains a challenge. In this work, we propose a new method for modeling copying for APE. To better identify translation errors, our method learns the representations of source sentences and system outputs in an interactive way. These representations are used to explicitly indicate which words in the system outputs should be copied. Finally, CopyNet (Gu et.al., 2016) can be combined with our method to place the copied words in correct positions in post-edited translations. Experiments on the datasets of the WMT 2016-2017 APE shared tasks show that our approach outperforms all best published results.

Most machine translation systems generate text autoregressively from left to right. We, instead, use a masked language modeling objective to train a model to predict any subset of the target words, conditioned on both the input text and a partially masked target translation. This approach allows for efficient iterative decoding, where we first predict all of the target words non-autoregressively, and then repeatedly mask out and regenerate the subset of words that the model is least confident about. By applying this strategy for a constant number of iterations, our model improves state-of-the-art performance levels for non-autoregressive and parallel decoding translation models by over 4 BLEU on average. It is also able to reach within about 1 BLEU point of a typical left-to-right transformer model, while decoding significantly faster.

For machine translation, a vast majority of language pairs in the world are considered low-resource because they have little parallel data available. Besides the technical challenges of learning with limited supervision, it is difficult to evaluate methods trained on low-resource language pairs because of the lack of freely and publicly available benchmarks. In this work, we introduce the FLORES evaluation datasets for Nepali–English and Sinhala– English, based on sentences translated from Wikipedia. Compared to English, these are languages with very different morphology and syntax, for which little out-of-domain parallel data is available and for which relatively large amounts of monolingual data are freely available. We describe our process to collect and cross-check the quality of translations, and we report baseline performance using several learning settings: fully supervised, weakly supervised, semi-supervised, and fully unsupervised. Our experiments demonstrate that current state-of-the-art methods perform rather poorly on this benchmark, posing a challenge to the research community working on low-resource MT. Data and code to reproduce our experiments are available at https://github.com/facebookresearch/flores.

Recently, biomedical version of embeddings obtained from language models such as BioELMo have shown state-of-the-art results for the textual inference task in the medical domain. In this paper, we explore how to incorporate structured domain knowledge, available in the form of a knowledge graph (UMLS), for the Medical NLI task. Specifically, we experiment with fusing embeddings obtained from knowledge graph with the state-of-the-art approaches for NLI task (ESIM model). We also experiment with fusing the domain-specific sentiment information for the task. Experiments conducted on MedNLI dataset clearly show that this strategy improves the baseline BioELMo architecture for the Medical NLI task.

Natural language inference (NLI) datasets (e.g., MultiNLI) were collected by soliciting hypotheses for a given premise from annotators. Such data collection led to annotation artifacts: systems can identify the premise-hypothesis relationship without observing the premise (e.g., negation in hypothesis being indicative of contradiction). We address this problem by recasting the CommitmentBank for NLI, which contains items involving reasoning over the extent to which a speaker is committed to complements of clause-embedding verbs under entailment-canceling environments (conditional, negation, modal and question). Instead of being constructed to stand in certain relationships with the premise, hypotheses in the recast CommitmentBank are the complements of the clause-embedding verb in each premise, leading to no annotation artifacts in the hypothesis. A state-of-the-art BERT-based model performs well on the CommitmentBank with 85% F1. However analysis of model behavior shows that the BERT models still do not capture the full complexity of pragmatic reasoning, nor encode some of the linguistic generalizations, highlighting room for improvement.

We introduce WIQA, the first large-scale dataset of “What if...” questions over procedural text. WIQA contains a collection of paragraphs, each annotated with multiple influence graphs describing how one change affects another, and a large (40k) collection of “What if...?” multiple-choice questions derived from these. For example, given a paragraph about beach erosion, would stormy weather hasten or decelerate erosion? WIQA contains three kinds of questions: perturbations to steps mentioned in the paragraph; external (out-of-paragraph) perturbations requiring commonsense knowledge; and irrelevant (no effect) perturbations. We find that state-of-the-art models achieve 73.8% accuracy, well below the human performance of 96.3%. We analyze the challenges, in particular tracking chains of influences, and present the dataset as an open challenge to the community.

Pre-trained BERT contextualized representations have achieved state-of-the-art results on multiple downstream NLP tasks by fine-tuning with task-specific data. While there has been a lot of focus on task-specific fine-tuning, there has been limited work on improving the pre-trained representations. In this paper, we explore ways of improving the pre-trained contextual representations for the task of automatic short answer grading, a critical component of intelligent tutoring systems. We show that the pre-trained BERT model can be improved by augmenting data from the domain-specific resources like textbooks. We also present a new approach to use labeled short answering grading data for further enhancement of the language model. Empirical evaluation on multi-domain datasets shows that task-specific fine-tuning on the enhanced pre-trained language model achieves superior performance for short answer grading.

Contextualized word embeddings have boosted many NLP tasks compared with traditional static word embeddings. However, the word with a specific sense may have different contextualized embeddings due to its various contexts. To further investigate what contextualized word embeddings capture, this paper analyzes whether they can indicate the corresponding sense definitions and proposes a general framework that is capable of explaining word meanings given contextualized word embeddings for better interpretation. The experiments show that both ELMo and BERT embeddings can be well interpreted via a readable textual form, and the findings may benefit the research community for a better understanding of what the embeddings capture.

In this work, we propose an aggregation method to combine the Bidirectional Encoder Representations from Transformer (BERT) with a MatchLSTM layer for Sequence Matching. Given a sentence pair, we extract the output representations of it from BERT. Then we extend BERT with a MatchLSTM layer to get further interaction of the sentence pair for sequence matching tasks. Taking natural language inference as an example, we split BERT output into two parts, which is from premise sentence and hypothesis sentence. At each position of the hypothesis sentence, both the weighted representation of the premise sentence and the representation of the current token are fed into LSTM. We jointly train the aggregation layer and pre-trained layer for sequence matching. We conduct an experiment on two publicly available datasets, WikiQA and SNLI. Experiments show that our model achieves significantly improvement compared with state-of-the-art methods on both datasets.

How do adjectives project from a noun to its parts? If a motorcycle is red, are its wheels red? Is a nuclear submarine’s captain nuclear? These questions are easy for humans to judge using our commonsense understanding of the world, but are difficult for computers. To attack this challenge, we crowdsource a set of human judgments that answer the English-language question “Given a whole described by an adjective, does the adjective also describe a given part?” We build strong baselines for this task with a classification approach. Our findings indicate that, despite the recent successes of large language models on tasks aimed to assess commonsense knowledge, these models do not greatly outperform simple word-level models based on pre-trained word embeddings. This provides evidence that the amount of commonsense knowledge encoded in these language models does not extend far beyond that already baked into the word embeddings. Our dataset will serve as a useful testbed for future research in commonsense reasoning, especially as it relates to adjectives and objects

Most deep learning approaches for text-to-SQL generation are limited to the WikiSQL dataset, which only supports very simple queries over a single table. We focus on the Spider dataset, a complex and cross-domain text-to-SQL task, which includes complex queries over multiple tables. In this paper, we propose a SQL clause-wise decoding neural architecture with a self-attention based database schema encoder to address the Spider task. Each of the clause-specific decoders consists of a set of sub-modules, which is defined by the syntax of each clause. Additionally, our model works recursively to support nested queries. When evaluated on the Spider dataset, our approach achieves 4.6% and 9.8% accuracy gain in the test and dev sets, respectively. In addition, we show that our model is significantly more effective at predicting complex and nested queries than previous work.

Deep reinforcement learning (RL) has been a commonly-used strategy for the abstractive summarization task to address both the exposure bias and non-differentiable task issues. However, the conventional reward Rouge-L simply looks for exact n-grams matches between candidates and annotated references, which inevitably makes the generated sentences repetitive and incoherent. In this paper, instead of Rouge-L, we explore the practicability of utilizing the distributional semantics to measure the matching degrees. With distributional semantics, sentence-level evaluation can be obtained, and semantically-correct phrases can also be generated without being limited to the surface form of the reference sentences. Human judgments on Gigaword and CNN/Daily Mail datasets show that our proposed distributional semantics reward (DSR) has distinct superiority in capturing the lexical and compositional diversity of natural language.