EMNLP 2020 Papers

The neural attention mechanism plays an important role in many natural language processing applications. In particular, multi-head attention extends single-head attention by allowing a model to jointly attend information from different perspectives. However, without explicit constraining, multi-head attention may suffer from attention collapse, an issue that makes different heads extract similar attentive features, thus limiting the model’s representation power. In this paper, for the first time, we provide a novel understanding of multi-head attention from a Bayesian perspective. Based on the recently developed particle-optimization sampling techniques, we propose a non-parametric approach that explicitly improves the repulsiveness in multi-head attention and consequently strengthens model’s expressiveness. Remarkably, our Bayesian interpretation provides theoretical inspirations on the not-well-understood questions: why and how one uses multi-head attention. Extensive experiments on various attention models and applications demonstrate that the proposed repulsive attention can improve the learned feature diversity, leading to more informative representations with consistent performance improvement on multiple tasks.

One reason pretraining on self-supervised linguistic tasks is effective is that it teaches models features that are helpful for language understanding. However, we want pretrained models to learn not only to represent linguistic features, but also to use those features preferentially during fine-turning. With this goal in mind, we introduce a new English-language diagnostic set called MSGS (the Mixed Signals Generalization Set), which consists of 20 ambiguous binary classification tasks that we use to test whether a pretrained model prefers linguistic or surface generalizations during finetuning. We pretrain RoBERTa from scratch on quantities of data ranging from 1M to 1B words and compare their performance on MSGS to the publicly available RoBERTa_BASE. We find that models can learn to represent linguistic features with little pretraining data, but require far more data to learn to prefer linguistic generalizations over surface ones. Eventually, with about 30B words of pretraining data, RoBERTa_BASE does consistently demonstrate a linguistic bias with some regularity. We conclude that while self-supervised pretraining is an effective way to learn helpful inductive biases, there is likely room to improve the rate at which models learn which features matter.

Most modern NLP systems make use of pre-trained contextual representations that attain astonishingly high performance on a variety of tasks. Such high performance should not be possible unless some form of linguistic structure inheres in these representations, and a wealth of research has sprung up on probing for it. In this paper, we draw a distinction between intrinsic probing, which examines how linguistic information is structured within a representation, and the extrinsic probing popular in prior work, which only argues for the presence of such information by showing that it can be successfully extracted. To enable intrinsic probing, we propose a novel framework based on a decomposable multivariate Gaussian probe that allows us to determine whether the linguistic information in word embeddings is dispersed or focal. We then probe fastText and BERT for various morphosyntactic attributes across 36 languages. We find that most attributes are reliably encoded by only a few neurons, with fastText concentrating its linguistic structure more than BERT.

To measure how well pretrained representations encode some linguistic property, it is common to use accuracy of a probe, i.e. a classifier trained to predict the property from the representations. Despite widespread adoption of probes, differences in their accuracy fail to adequately reflect differences in representations. For example, they do not substantially favour pretrained representations over randomly initialized ones. Analogously, their accuracy can be similar when probing for genuine linguistic labels and probing for random synthetic tasks. To see reasonable differences in accuracy with respect to these random baselines, previous work had to constrain either the amount of probe training data or its model size. Instead, we propose an alternative to the standard probes, information-theoretic probing with minimum description length (MDL). With MDL probing, training a probe to predict labels is recast as teaching it to effectively transmit the data. Therefore, the measure of interest changes from probe accuracy to the description length of labels given representations. In addition to probe quality, the description length evaluates “the amount of effort” needed to achieve the quality. This amount of effort characterizes either (i) size of a probing model, or (ii) the amount of data needed to achieve the high quality. We consider two methods for estimating MDL which can be easily implemented on top of the standard probing pipelines: variational coding and online coding. We show that these methods agree in results and are more informative and stable than the standard probes.

Deep pre-trained contextualized encoders like BERT demonstrate remarkable performance on a range of downstream tasks. A recent line of research in probing investigates the linguistic knowledge implicitly learned by these models during pre-training. While most work in probing operates on the task level, linguistic tasks are rarely uniform and can be represented in a variety of formalisms. Any linguistics-based probing study thereby inevitably commits to the formalism used to annotate the underlying data. Can the choice of formalism affect probing results? To investigate, we conduct an in-depth cross-formalism layer probing study in role semantics. We find linguistically meaningful differences in the encoding of semantic role- and proto-role information by BERT depending on the formalism and demonstrate that layer probing can detect subtle differences between the implementations of the same linguistic formalism. Our results suggest that linguistic formalism is an important dimension in probing studies, along with the commonly used cross-task and cross-lingual experimental settings.

Deep learning models for linguistic tasks require large training datasets, which are expensive to create. As an alternative to the traditional approach of creating new instances by repeating the process of creating one instance, we propose doing so by first collecting a set of seed examples and then applying human-driven natural perturbations (as opposed to rule-based machine perturbations), which often change the gold label as well. Such perturbations have the advantage of being relatively easier (and hence cheaper) to create than writing out completely new examples. Further, they help address the issue that even models achieving human-level scores on NLP datasets are known to be considerably sensitive to small changes in input. To evaluate the idea, we consider a recent question-answering dataset (BOOLQ) and study our approach as a function of the perturbation cost ratio, the relative cost of perturbing an existing question vs. creating a new one from scratch. We find that when natural perturbations are moderately cheaper to create (cost ratio under 60%), it is more effective to use them for training BOOLQ models: such models exhibit 9% higher robustness and 4.5% stronger generalization, while retaining performance on the original BOOLQ dataset.

The aim of all Question Answering (QA) systems is to generalize to unseen questions. Current supervised methods are reliant on expensive data annotation. Moreover, such annotations can introduce unintended annotator bias, making systems focus more on the bias than the actual task. This work proposes Knowledge Triplet Learning (KTL), a self-supervised task over knowledge graphs. We propose heuristics to create synthetic graphs for commonsense and scientific knowledge. We propose using KTL to perform zero-shot question answering, and our experiments show considerable improvements over large pre-trained transformer language models.

Despite the rapid progress in multihop question-answering (QA), models still have trouble explaining why an answer is correct, with limited explanation training data available to learn from. To address this, we introduce three explanation datasets in which explanations formed from corpus facts are annotated. Our first dataset, eQASC contains over 98K explanation annotations for the multihop question answering dataset QASC, and is the first that annotates multiple candidate explanations for each answer. The second dataset eQASC-perturbed is constructed by crowd-sourcing perturbations (while preserving their validity) of a subset of explanations in QASC, to test consistency and generalization of explanation prediction models. The third dataset eOBQA is constructed by adding explanation annotations to the OBQA dataset to test generalization of models trained on eQASC. We show that this data can be used to significantly improve explanation quality (+14% absolute F1 over a strong retrieval baseline) using a BERT-based classifier, but still behind the upper bound, offering a new challenge for future research. We also explore a delexicalized chain representation in which repeated noun phrases are replaced by variables, thus turning them into generalized reasoning chains (for example: “X is a Y” AND “Y has Z” IMPLIES “X has Z”). We find that generalized chains maintain performance while also being more robust to certain perturbations.

Recent work by Clark et al. (2020) shows that transformers can act as “soft theorem provers” by answering questions over explicitly provided knowledge in natural language. In our work, we take a step closer to emulating formal theorem provers, by proposing PRover, an interpretable transformer-based model that jointly answers binary questions over rule-bases and generates the corresponding proofs. Our model learns to predict nodes and edges corresponding to proof graphs in an efficient constrained training paradigm. During inference, a valid proof, satisfying a set of global constraints is generated. We conduct experiments on synthetic, hand-authored, and human-paraphrased rule-bases to show promising results for QA and proof generation, with strong generalization performance. First, PRover generates proofs with an accuracy of 87%, while retaining or improving performance on the QA task, compared to RuleTakers (up to 6% improvement on zero-shot evaluation). Second, when trained on questions requiring lower depths of reasoning, it generalizes significantly better to higher depths (up to 15% improvement). Third, PRover obtains near perfect QA accuracy of 98% using only 40% of the training data. However, generating proofs for questions requiring higher depths of reasoning becomes challenging, and the accuracy drops to 65% for “depth 5”, indicating significant scope for future work.

We frame the task of machine translation evaluation as one of scoring machine translation output with a sequence-to-sequence paraphraser, conditioned on a human reference. We propose training the paraphraser as a multilingual NMT system, treating paraphrasing as a zero-shot translation task (e.g., Czech to Czech). This results in the paraphraser’s output mode being centered around a copy of the input sequence, which represents the best case scenario where the MT system output matches a human reference. Our method is simple and intuitive, and does not require human judgements for training. Our single model (trained in 39 languages) outperforms or statistically ties with all prior metrics on the WMT 2019 segment-level shared metrics task in all languages (excluding Gujarati where the model had no training data). We also explore using our model for the task of quality estimation as a metric—conditioning on the source instead of the reference—and find that it significantly outperforms every submission to the WMT 2019 shared task on quality estimation in every language pair.

Many valid translations exist for a given sentence, yet machine translation (MT) is trained with a single reference translation, exacerbating data sparsity in low-resource settings. We introduce Simulated Multiple Reference Training (SMRT), a novel MT training method that approximates the full space of possible translations by sampling a paraphrase of the reference sentence from a paraphraser and training the MT model to predict the paraphraser’s distribution over possible tokens. We demonstrate the effectiveness of SMRT in low-resource settings when translating to English, with improvements of 1.2 to 7.0 BLEU. We also find SMRT is complementary to back-translation.

The term translationese has been used to describe features of translated text, and in this paper, we provide detailed analysis of potential adverse effects of translationese on machine translation evaluation. Our analysis shows differences in conclusions drawn from evaluations that include translationese in test data compared to experiments that tested only with text originally composed in that language. For this reason we recommend that reverse-created test data be omitted from future machine translation test sets. In addition, we provide a re-evaluation of a past machine translation evaluation claiming human-parity of MT. One important issue not previously considered is statistical power of significance tests applied to comparison of human and machine translation. Since the very aim of past evaluations was investigation of ties between human and MT systems, power analysis is of particular importance, to avoid, for example, claims of human parity simply corresponding to Type II error resulting from the application of a low powered test. We provide detailed analysis of tests used in such evaluations to provide an indication of a suitable minimum sample size for future studies.

The quality of automatic metrics for machine translation has been increasingly called into question, especially for high-quality systems. This paper demonstrates that, while choice of metric is important, the nature of the references is also critical. We study different methods to collect references and compare their value in automated evaluation by reporting correlation with human evaluation for a variety of systems and metrics. Motivated by the finding that typical references exhibit poor diversity, concentrating around translationese language, we develop a paraphrasing task for linguists to perform on existing reference translations, which counteracts this bias. Our method yields higher correlation with human judgment not only for the submissions of WMT 2019 English to German, but also for Back-translation and APE augmented MT output, which have been shown to have low correlation with automatic metrics using standard references. We demonstrate that our methodology improves correlation with all modern evaluation metrics we look at, including embedding-based methods. To complete this picture, we reveal that multi-reference BLEU does not improve the correlation for high quality output, and present an alternative multi-reference formulation that is more effective.

Social media platforms have become an essential venue for online deliberation where users discuss arguments, debate, and form opinions. In this paper, we propose an unsupervised method to detect the stance of argumentative claims with respect to a topic. Most related work focuses on topic-specific supervised models that need to be trained for every emergent debate topic. To address this limitation, we propose a topic independent approach that focuses on a frequently encountered class of arguments, specifically, on arguments from consequences. We do this by extracting the effects that claims refer to, and proposing a means for inferring if the effect is a good or bad consequence. Our experiments provide promising results that are comparable to, and in particular regards even outperform BERT. Furthermore, we publish a novel dataset of arguments relating to consequences, annotated with Amazon Mechanical Turk.

When summarizing a collection of views, arguments or opinions on some topic, it is often desirable not only to extract the most salient points, but also to quantify their prevalence. Work on multi-document summarization has traditionally focused on creating textual summaries, which lack this quantitative aspect. Recent work has proposed to summarize arguments by mapping them to a small set of expert-generated key points, where the salience of each key point corresponds to the number of its matching arguments. The current work advances key point analysis in two important respects: first, we develop a method for automatic extraction of key points, which enables fully automatic analysis, and is shown to achieve performance comparable to a human expert. Second, we demonstrate that the applicability of key point analysis goes well beyond argumentation data. Using models trained on publicly available argumentation datasets, we achieve promising results in two additional domains: municipal surveys and user reviews. An additional contribution is an in-depth evaluation of argument-to-key point matching models, where we substantially outperform previous results.

Argumentation accommodates various rhetorical devices, such as questions, reported speech, and imperatives. These rhetorical tools usually assert argumentatively relevant propositions rather implicitly, so understanding their true meaning is key to understanding certain arguments properly. However, most argument mining systems and computational linguistics research have paid little attention to implicitly asserted propositions in argumentation. In this paper, we examine a wide range of computational methods for extracting propositions that are implicitly asserted in questions, reported speech, and imperatives in argumentation. By evaluating the models on a corpus of 2016 U.S. presidential debates and online commentary, we demonstrate the effectiveness and limitations of the computational models. Our study may inform future research on argument mining and the semantics of these rhetorical devices in argumentation.

Finding attackable sentences in an argument is the first step toward successful refutation in argumentation. We present a first large-scale analysis of sentence attackability in online arguments. We analyze driving reasons for attacks in argumentation and identify relevant characteristics of sentences. We demonstrate that a sentence’s attackability is associated with many of these characteristics regarding the sentence’s content, proposition types, and tone, and that an external knowledge source can provide useful information about attackability. Building on these findings, we demonstrate that machine learning models can automatically detect attackable sentences in arguments, significantly better than several baselines and comparably well to laypeople.