TACL 2021 论文列表

Idiomatic expressions are an integral part of natural language and constantly being added to a language. Owing to their non-compositionality and their ability to take on a figurative or literal meaning depending on the sentential context, they have been a classical challenge for NLP systems. To address this challenge, we study the task of detecting whether a sentence has an idiomatic expression and localizing it when it occurs in a figurative sense. Prior research for this task has studied specific classes of idiomatic expressions offering limited views of their generalizability to new idioms. We propose a multi-stage neural architecture with attention flow as a solution. The network effectively fuses contextual and lexical information at different levels using word and sub-word representations. Empirical evaluations on three of the largest benchmark datasets with idiomatic expressions of varied syntactic patterns and degrees of non-compositionality show that our proposed model achieves new state-of-the-art results. A salient feature of the model is its ability to identify idioms unseen during training with gains from 1.4% to 30.8% over competitive baselines on the largest dataset.

We adopt an evolutionary view on language change in which cognitive factors (in addition to social ones) affect the fitness of words and their success in the linguistic ecosystem. Specifically, we propose a variety of psycholinguistic factors—semantic, distributional, and phonological—that we hypothesize are predictive of lexical decline, in which words greatly decrease in frequency over time. Using historical data across three languages (English, French, and German), we find that most of our proposed factors show a significant difference in the expected direction between each curated set of declining words and their matched stable words. Moreover, logistic regression analyses show that semantic and distributional factors are significant in predicting declining words. Further diachronic analysis reveals that declining words tend to decrease in the diversity of their lexical contexts over time, gradually narrowing their ‘ecological niches’.

Debugging a machine learning model is hard since the bug usually involves the training data and the learning process. This becomes even harder for an opaque deep learning model if we have no clue about how the model actually works. In this survey, we review papers that exploit explanations to enable humans to give feedback and debug NLP models. We call this problem explanation-based human debugging (EBHD). In particular, we categorize and discuss existing work along three dimensions of EBHD (the bug context, the workflow, and the experimental setting), compile findings on how EBHD components affect the feedback providers, and highlight open problems that could be future research directions.

Interpretable rationales for model predictions are crucial in practical applications. We develop neural models that possess an interpretable inference process for dependency parsing. Our models adopt instance-based inference, where dependency edges are extracted and labeled by comparing them to edges in a training set. The training edges are explicitly used for the predictions; thus, it is easy to grasp the contribution of each edge to the predictions. Our experiments show that our instance-based models achieve competitive accuracy with standard neural models and have the reasonable plausibility of instance-based explanations.

We introduce a simple but flexible mechanism to learn an intermediate plan to ground the generation of abstractive summaries. Specifically, we prepend (or prompt) target summaries with entity chains—ordered sequences of entities mentioned in the summary. Transformer-based sequence-to-sequence models are then trained to generate the entity chain and then continue generating the summary conditioned on the entity chain and the input. We experimented with both pretraining and finetuning with this content planning objective. When evaluated on CNN/DailyMail, XSum, SAMSum, and BillSum, we demonstrate empirically that the grounded generation with the planning objective improves entity specificity and planning in summaries for all datasets, and achieves state-of-the-art performance on XSum and SAMSum in terms of rouge. Moreover, we demonstrate empirically that planning with entity chains provides a mechanism to control hallucinations in abstractive summaries. By prompting the decoder with a modified content plan that drops hallucinated entities, we outperform state-of-the-art approaches for faithfulness when evaluated automatically and by humans.

Human evaluation of modern high-quality machine translation systems is a difficult problem, and there is increasing evidence that inadequate evaluation procedures can lead to erroneous conclusions. While there has been considerable research on human evaluation, the field still lacks a commonly accepted standard procedure. As a step toward this goal, we propose an evaluation methodology grounded in explicit error analysis, based on the Multidimensional Quality Metrics (MQM) framework. We carry out the largest MQM research study to date, scoring the outputs of top systems from the WMT 2020 shared task in two language pairs using annotations provided by professional translators with access to full document context. We analyze the resulting data extensively, finding among other results a substantially different ranking of evaluated systems from the one established by the WMT crowd workers, exhibiting a clear preference for human over machine output. Surprisingly, we also find that automatic metrics based on pre-trained embeddings can outperform human crowd workers. We make our corpus publicly available for further research.

Multi-head attention, a collection of several attention mechanisms that independently attend to different parts of the input, is the key ingredient in the Transformer. Recent work has shown, however, that a large proportion of the heads in a Transformer’s multi-head attention mechanism can be safely pruned away without significantly harming the performance of the model; such pruning leads to models that are noticeably smaller and faster in practice. Our work introduces a new head pruning technique that we term differentiable subset pruning. ntuitively, our method learns per- head importance variables and then enforces a user-specified hard constraint on the number of unpruned heads. he importance variables are learned via stochastic gradient descent. e conduct experiments on natural language inference and machine translation; we show that differentiable subset pruning performs comparably or better than previous works while offering precise control of the sparsity level.1

Several metrics have been proposed for assessing the similarity of (abstract) meaning representations (AMRs), but little is known about how they relate to human similarity ratings. Moreover, the current metrics have complementary strengths and weaknesses: Some emphasize speed, while others make the alignment of graph structures explicit, at the price of a costly alignment step. In this work we propose new Weisfeiler-Leman AMR similarity metrics that unify the strengths of previous metrics, while mitigating their weaknesses. Specifically, our new metrics are able to match contextualized substructures and induce n:m alignments between their nodes. Furthermore, we introduce a Benchmark for AMR Metrics based on Overt Objectives (Bamboo), the first benchmark to support empirical assessment of graph-based MR similarity metrics. Bamboo maximizes the interpretability of results by defining multiple overt objectives that range from sentence similarity objectives to stress tests that probe a metric’s robustness against meaning-altering and meaning- preserving graph transformations. We show the benefits of Bamboo by profiling previous metrics and our own metrics. Results indicate that our novel metrics may serve as a strong baseline for future work.

This paper contains prompts and model outputs that are offensive in nature. When trained on large, unfiltered crawls from the Internet, language models pick up and reproduce all kinds of undesirable biases that can be found in the data: They often generate racist, sexist, violent, or otherwise toxic language. As large models require millions of training examples to achieve good performance, it is difficult to completely prevent them from being exposed to such content. In this paper, we first demonstrate a surprising finding: Pretrained language models recognize, to a considerable degree, their undesirable biases and the toxicity of the content they produce. We refer to this capability as self-diagnosis. Based on this finding, we then propose a decoding algorithm that, given only a textual description of the undesired behavior, reduces the probability of a language model producing problematic text. We refer to this approach as self-debiasing. Self-debiasing does not rely on manually curated word lists, nor does it require any training data or changes to the model’s parameters. While we by no means eliminate the issue of language models generating biased text, we believe our approach to be an important step in this direction.1

During production of this paper, an error was introduced to the formula on the bottom of the right column of page 1020. In the last two terms of the formula, the n and m subscripts were swapped. The correct formula is:Lc=∑n=1k∑m=n+1kDKL(Qnri∥Qmri)+DKL(Qmri∥Qnri)The paper has been updated.

Progress in cross-lingual modeling depends on challenging, realistic, and diverse evaluation sets. We introduce Multilingual Knowledge Questions and Answers (MKQA), an open- domain question answering evaluation set comprising 10k question-answer pairs aligned across 26 typologically diverse languages (260k question-answer pairs in total). Answers are based on heavily curated, language- independent data representation, making results comparable across languages and independent of language-specific passages. With 26 languages, this dataset supplies the widest range of languages to-date for evaluating question answering. We benchmark a variety of state- of-the-art methods and baselines for generative and extractive question answering, trained on Natural Questions, in zero shot and translation settings. Results indicate this dataset is challenging even in English, but especially in low-resource languages.1

We describe a parser of English effectuated by biologically plausible neurons and synapses, and implemented through the Assembly Calculus, a recently proposed computational framework for cognitive function. We demonstrate that this device is capable of correctly parsing reasonably nontrivial sentences.1 While our experiments entail rather simple sentences in English, our results suggest that the parser can be extended beyond what we have implemented, to several directions encompassing much of language. For example, we present a simple Russian version of the parser, and discuss how to handle recursion, embedding, and polysemy.

Recent improvements in the predictive quality of natural language processing systems are often dependent on a substantial increase in the number of model parameters. This has led to various attempts of compressing such models, but existing methods have not considered the differences in the predictive power of various model components or in the generalizability of the compressed models. To understand the connection between model compression and out-of-distribution generalization, we define the task of compressing language representation models such that they perform best in a domain adaptation setting. We choose to address this problem from a causal perspective, attempting to estimate the average treatment effect (ATE) of a model component, such as a single layer, on the model’s predictions. Our proposed ATE-guided Model Compression scheme (AMoC), generates many model candidates, differing by the model components that were removed. Then, we select the best candidate through a stepwise regression model that utilizes the ATE to predict the expected performance on the target domain. AMoC outperforms strong baselines on dozens of domain pairs across three text classification and sequence tagging tasks.1

We introduce Generative Spoken Language Modeling, the task of learning the acoustic and linguistic characteristics of a language from raw audio (no text, no labels), and a set of metrics to automatically evaluate the learned representations at acoustic and linguistic levels for both encoding and generation. We set up baseline systems consisting of a discrete speech encoder (returning pseudo-text units), a generative language model (trained on pseudo- text), and a speech decoder (generating a waveform from pseudo-text) all trained without supervision and validate the proposed metrics with human evaluation. Across 3 speech encoders (CPC, wav2vec 2.0, HuBERT), we find that the number of discrete units (50, 100, or 200) matters in a task-dependent and encoder- dependent way, and that some combinations approach text-based systems.1

We study learning named entity recognizers in the presence of missing entity annotations. We approach this setting as tagging with latent variables and propose a novel loss, the Expected Entity Ratio, to learn models in the presence of systematically missing tags. We show that our approach is both theoretically sound and empirically useful. Experimentally, we find that it meets or exceeds performance of strong and state-of-the-art baselines across a variety of languages, annotation scenarios, and amounts of labeled data. In particular, we find that it significantly outperforms the previous state-of-the-art methods from Mayhew et al. (2019) and Li et al. (2021) by +12.7 and +2.3 F1 score in a challenging setting with only 1,000 biased annotations, averaged across 7 datasets. We also show that, when combined with our approach, a novel sparse annotation scheme outperforms exhaustive annotation for modest annotation budgets.1

We study continual learning for natural language instruction generation, by observing human users’ instruction execution. We focus on a collaborative scenario, where the system both acts and delegates tasks to human users using natural language. We compare user execution of generated instructions to the original system intent as an indication to the system’s success communicating its intent. We show how to use this signal to improve the system’s ability to generate instructions via contextual bandit learning. In interaction with real users, our system demonstrates dramatic improvements in its ability to generate language over time.

Much of the existing linguistic data in many languages of the world is locked away in non- digitized books and documents. Optical character recognition (OCR) can be used to produce digitized text, and previous work has demonstrated the utility of neural post-correction methods that improve the results of general- purpose OCR systems on recognition of less- well-resourced languages. However, these methods rely on manually curated post- correction data, which are relatively scarce compared to the non-annotated raw images that need to be digitized. In this paper, we present a semi-supervised learning method that makes it possible to utilize these raw images to improve performance, specifically through the use of self-training, a technique where a model is iteratively trained on its own outputs. In addition, to enforce consistency in the recognized vocabulary, we introduce a lexically aware decoding method that augments the neural post-correction model with a count-based language model constructed from the recognized texts, implemented using weighted finite-state automata (WFSA) for efficient and effective decoding. Results on four endangered languages demonstrate the utility of the proposed method, with relative error reductions of 15%–29%, where we find the combination of self-training and lexically aware decoding essential for achieving consistent improvements.1

To develop commonsense-grounded NLP applications, a comprehensive and accurate commonsense knowledge graph (CKG) is needed. It is time-consuming to manually construct CKGs and many research efforts have been devoted to the automatic construction of CKGs. Previous approaches focus on generating concepts that have direct and obvious relationships with existing concepts and lack an capability to generate unobvious concepts. In this work, we aim to bridge this gap. We propose a general graph-to-paths pretraining framework that leverages high-order structures in CKGs to capture high-order relationships between concepts. We instantiate this general framework to four special cases: long path, path-to-path, router, and graph-node-path. Experiments on two datasets demonstrate the effectiveness of our methods. The code will be released via the public GitHub repository.

Measuring bias is key for better understanding and addressing unfairness in NLP/ML models. This is often done via fairness metrics, which quantify the differences in a model’s behaviour across a range of demographic groups. In this work, we shed more light on the differences and similarities between the fairness metrics used in NLP. First, we unify a broad range of existing metrics under three generalized fairness metrics, revealing the connections between them. Next, we carry out an extensive empirical comparison of existing metrics and demonstrate that the observed differences in bias measurement can be systematically explained via differences in parameter choices for our generalized metrics.