NAACL 2021 Papers

We propose EASE, a simple diagnostic tool for Visual Question Answering (VQA) which quantifies the difficulty of an image, question sample. EASE is based on the pattern of answers provided by multiple annotators to a given question. In particular, it considers two aspects of the answers: (i) their Entropy; (ii) their Semantic content. First, we prove the validity of our diagnostic to identify samples that are easy/hard for state-of-art VQA models. Second, we show that EASE can be successfully used to select the most-informative samples for training/fine-tuning. Crucially, only information that is readily available in any VQA dataset is used to compute its scores.

Word Sense Disambiguation (WSD) is a historical NLP task aimed at linking words in contexts to discrete sense inventories and it is usually cast as a multi-label classification task. Recently, several neural approaches have employed sense definitions to better represent word meanings. Yet, these approaches do not observe the input sentence and the sense definition candidates all at once, thus potentially reducing the model performance and generalization power. We cope with this issue by reframing WSD as a span extraction problem — which we called Extractive Sense Comprehension (ESC) — and propose ESCHER, a transformer-based neural architecture for this new formulation. By means of an extensive array of experiments, we show that ESC unleashes the full potential of our model, leading it to outdo all of its competitors and to set a new state of the art on the English WSD task. In the few-shot scenario, ESCHER proves to exploit training data efficiently, attaining the same performance as its closest competitor while relying on almost three times fewer annotations. Furthermore, ESCHER can nimbly combine data annotated with senses from different lexical resources, achieving performances that were previously out of everyone’s reach. The model along with data is available at https://github.com/SapienzaNLP/esc.

Coarse-grained linguistic information, such as named entities or phrases, facilitates adequately representation learning in pre-training. Previous works mainly focus on extending the objective of BERT’s Masked Language Modeling (MLM) from masking individual tokens to contiguous sequences of n tokens. We argue that such contiguously masking method neglects to model the intra-dependencies and inter-relation of coarse-grained linguistic information. As an alternative, we propose ERNIE-Gram, an explicitly n-gram masking method to enhance the integration of coarse-grained information into pre-training. In ERNIE-Gram, n-grams are masked and predicted directly using explicit n-gram identities rather than contiguous sequences of n tokens. Furthermore, ERNIE-Gram employs a generator model to sample plausible n-gram identities as optional n-gram masks and predict them in both coarse-grained and fine-grained manners to enable comprehensive n-gram prediction and relation modeling. We pre-train ERNIE-Gram on English and Chinese text corpora and fine-tune on 19 downstream tasks. Experimental results show that ERNIE-Gram outperforms previous pre-training models like XLNet and RoBERTa by a large margin, and achieves comparable results with state-of-the-art methods. The source codes and pre-trained models have been released at https://github.com/PaddlePaddle/ERNIE.

Most of privacy protection studies for textual data focus on removing explicit sensitive identifiers. However, personal writing style, as a strong indicator of the authorship, is often neglected. Recent studies, such as SynTF, have shown promising results on privacy-preserving text mining. However, their anonymization algorithm can only output numeric term vectors which are difficult for the recipients to interpret. We propose a novel text generation model with a two-set exponential mechanism for authorship anonymization. By augmenting the semantic information through a REINFORCE training reward function, the model can generate differentially private text that has a close semantic and similar grammatical structure to the original text while removing personal traits of the writing style. It does not assume any conditioned labels or paralleled text data for training. We evaluate the performance of the proposed model on the real-life peer reviews dataset and the Yelp review dataset. The result suggests that our model outperforms the state-of-the-art on semantic preservation, authorship obfuscation, and stylometric transformation.

Framing involves the positive or negative presentation of an argument or issue depending on the audience and goal of the speaker. Differences in lexical framing, the focus of our work, can have large effects on peoples’ opinions and beliefs. To make progress towards reframing arguments for positive effects, we create a dataset and method for this task. We use a lexical resource for “connotations” to create a parallel corpus and propose a method for argument reframing that combines controllable text generation (positive connotation) with a post-decoding entailment component (same denotation). Our results show that our method is effective compared to strong baselines along the dimensions of fluency, meaning, and trustworthiness/reduction of fear.

Representation learning is widely used in NLP for a vast range of tasks. However, representations derived from text corpora often reflect social biases. This phenomenon is pervasive and consistent across different neural models, causing serious concern. Previous methods mostly rely on a pre-specified, user-provided direction or suffer from unstable training. In this paper, we propose an adversarial disentangled debiasing model to dynamically decouple social bias attributes from the intermediate representations trained on the main task. We aim to denoise bias information while training on the downstream task, rather than completely remove social bias and pursue static unbiased representations. Experiments show the effectiveness of our method, both on the effect of debiasing and the main task performance.

We introduce Dynabench, an open-source platform for dynamic dataset creation and model benchmarking. Dynabench runs in a web browser and supports human-and-model-in-the-loop dataset creation: annotators seek to create examples that a target model will misclassify, but that another person will not. In this paper, we argue that Dynabench addresses a critical need in our community: contemporary models quickly achieve outstanding performance on benchmark tasks but nonetheless fail on simple challenge examples and falter in real-world scenarios. With Dynabench, dataset creation, model development, and model assessment can directly inform each other, leading to more robust and informative benchmarks. We report on four initial NLP tasks, illustrating these concepts and highlighting the promise of the platform, and address potential objections to dynamic benchmarking as a new standard for the field.

Recent neural text-to-SQL models can effectively translate natural language questions to corresponding SQL queries on unseen databases. Working mostly on the Spider dataset, researchers have proposed increasingly sophisticated solutions to the problem. Contrary to this trend, in this paper we focus on simplifications. We begin by building DuoRAT, a re-implementation of the state-of-the-art RAT-SQL model that unlike RAT-SQL is using only relation-aware or vanilla transformers as the building blocks. We perform several ablation experiments using DuoRAT as the baseline model. Our experiments confirm the usefulness of some techniques and point out the redundancy of others, including structural SQL features and features that link the question with the schema.

Robustness and counterfactual bias are usually evaluated on a test dataset. However, are these evaluations robust? If the test dataset is perturbed slightly, will the evaluation results keep the same? In this paper, we propose a “double perturbation” framework to uncover model weaknesses beyond the test dataset. The framework first perturbs the test dataset to construct abundant natural sentences similar to the test data, and then diagnoses the prediction change regarding a single-word substitution. We apply this framework to study two perturbation-based approaches that are used to analyze models’ robustness and counterfactual bias in English. (1) For robustness, we focus on synonym substitutions and identify vulnerable examples where prediction can be altered. Our proposed attack attains high success rates (96.0%-99.8%) in finding vulnerable examples on both original and robustly trained CNNs and Transformers. (2) For counterfactual bias, we focus on substituting demographic tokens (e.g., gender, race) and measure the shift of the expected prediction among constructed sentences. Our method is able to reveal the hidden model biases not directly shown in the test dataset. Our code is available at https://github.com/chong-z/nlp-second-order-attack.

Domain divergence plays a significant role in estimating the performance of a model in new domains. While there is a significant literature on divergence measures, researchers find it hard to choose an appropriate divergence for a given NLP application. We address this shortcoming by both surveying the literature and through an empirical study. We develop a taxonomy of divergence measures consisting of three classes — Information-theoretic, Geometric, and Higher-order measures and identify the relationships between them. Further, to understand the common use-cases of these measures, we recognise three novel applications – 1) Data Selection, 2) Learning Representation, and 3) Decisions in the Wild – and use it to organise our literature. From this, we identify that Information-theoretic measures are prevalent for 1) and 3), and Higher-order measures are more common for 2). To further help researchers choose appropriate measures to predict drop in performance – an important aspect of Decisions in the Wild, we perform correlation analysis spanning 130 domain adaptation scenarios, 3 varied NLP tasks and 12 divergence measures identified from our survey. To calculate these divergences, we consider the current contextual word representations (CWR) and contrast with the older distributed representations. We find that traditional measures over word distributions still serve as strong baselines, while higher-order measures with CWR are effective.

Finetuning deep pre-trained language models has shown state-of-the-art performances on a wide range of Natural Language Processing (NLP) applications. Nevertheless, their generalization performance drops under domain shift. In the case of Arabic language, diglossia makes building and annotating corpora for each dialect and/or domain a more challenging task. Unsupervised Domain Adaptation tackles this issue by transferring the learned knowledge from labeled source domain data to unlabeled target domain data. In this paper, we propose a new unsupervised domain adaptation method for Arabic cross-domain and cross-dialect sentiment analysis from Contextualized Word Embedding. Several experiments are performed adopting the coarse-grained and the fine-grained taxonomies of Arabic dialects. The obtained results show that our method yields very promising results and outperforms several domain adaptation methods for most of the evaluated datasets. On average, our method increases the performance by an improvement rate of 20.8% over the zero-shot transfer learning from BERT.

Aspect-based Sentiment Analysis (ABSA), aiming at predicting the polarities for aspects, is a fine-grained task in the field of sentiment analysis. Previous work showed syntactic information, e.g. dependency trees, can effectively improve the ABSA performance. Recently, pre-trained models (PTMs) also have shown their effectiveness on ABSA. Therefore, the question naturally arises whether PTMs contain sufficient syntactic information for ABSA so that we can obtain a good ABSA model only based on PTMs. In this paper, we firstly compare the induced trees from PTMs and the dependency parsing trees on several popular models for the ABSA task, showing that the induced tree from fine-tuned RoBERTa (FT-RoBERTa) outperforms the parser-provided tree. The further analysis experiments reveal that the FT-RoBERTa Induced Tree is more sentiment-word-oriented and could benefit the ABSA task. The experiments also show that the pure RoBERTa-based model can outperform or approximate to the previous SOTA performances on six datasets across four languages since it implicitly incorporates the task-oriented syntactic information.

Machine reading comprehension is a challenging task especially for querying documents with deep and interconnected contexts. Transformer-based methods have shown advanced performances on this task; however, most of them still treat documents as a flat sequence of tokens. This work proposes a new Transformer-based method that reads a document as tree slices. It contains two modules for identifying more relevant text passage and the best answer span respectively, which are not only jointly trained but also jointly consulted at inference time. Our evaluation results show that our proposed method outperforms several competitive baseline approaches on two datasets from varied domains.

Large Transformers pretrained over clinical notes from Electronic Health Records (EHR) have afforded substantial gains in performance on predictive clinical tasks. The cost of training such models (and the necessity of data access to do so) coupled with their utility motivates parameter sharing, i.e., the release of pretrained models such as ClinicalBERT. While most efforts have used deidentified EHR, many researchers have access to large sets of sensitive, non-deidentified EHR with which they might train a BERT model (or similar). Would it be safe to release the weights of such a model if they did? In this work, we design a battery of approaches intended to recover Personal Health Information (PHI) from a trained BERT. Specifically, we attempt to recover patient names and conditions with which they are associated. We find that simple probing methods are not able to meaningfully extract sensitive information from BERT trained over the MIMIC-III corpus of EHR. However, more sophisticated “attacks” may succeed in doing so: To facilitate such research, we make our experimental setup and baseline probing models available at https://github.com/elehman16/exposing_patient_data_release.

Event extraction has long been treated as a sentence-level task in the IE community. We argue that this setting does not match human informative seeking behavior and leads to incomplete and uninformative extraction results. We propose a document-level neural event argument extraction model by formulating the task as conditional generation following event templates. We also compile a new document-level event extraction benchmark dataset WikiEvents which includes complete event and coreference annotation. On the task of argument extraction, we achieve an absolute gain of 7.6% F1 and 5.7% F1 over the next best model on the RAMS and WikiEvents dataset respectively. On the more challenging task of informative argument extraction, which requires implicit coreference reasoning, we achieve a 9.3% F1 gain over the best baseline. To demonstrate the portability of our model, we also create the first end-to-end zero-shot event extraction framework and achieve 97% of fully supervised model’s trigger extraction performance and 82% of the argument extraction performance given only access to 10 out of the 33 types on ACE.

Analysing whether neural language models encode linguistic information has become popular in NLP. One method of doing so, which is frequently cited to support the claim that models like BERT encode syntax, is called probing; probes are small supervised models trained to extract linguistic information from another model’s output. If a probe is able to predict a particular structure, it is argued that the model whose output it is trained on must have implicitly learnt to encode it. However, drawing a generalisation about a model’s linguistic knowledge about a specific phenomena based on what a probe is able to learn may be problematic: in this work, we show that semantic cues in training data means that syntactic probes do not properly isolate syntax. We generate a new corpus of semantically nonsensical but syntactically well-formed Jabberwocky sentences, which we use to evaluate two probes trained on normal data. We train the probes on several popular language models (BERT, GPT-2, and RoBERTa), and find that in all settings they perform worse when evaluated on these data, for one probe by an average of 15.4 UUAS points absolute. Although in most cases they still outperform the baselines, their lead is reduced substantially, e.g. by 53% in the case of BERT for one probe. This begs the question: what empirical scores constitute knowing syntax?

Sequence-to-sequence models have delivered impressive results in word formation tasks such as morphological inflection, often learning to model subtle morphophonological details with limited training data. Despite the performance, the opacity of neural models makes it difficult to determine whether complex generalizations are learned, or whether a kind of separate rote memorization of each morphophonological process takes place. To investigate whether complex alternations are simply memorized or whether there is some level of generalization across related sound changes in a sequence-to-sequence model, we perform several experiments on Finnish consonant gradation—a complex set of sound changes triggered in some words by certain suffixes. We find that our models often—though not always—encode 17 different consonant gradation processes in a handful of dimensions in the RNN. We also show that by scaling the activations in these dimensions we can control whether consonant gradation occurs and the direction of the gradation.

While the predictive performance of modern statistical dependency parsers relies heavily on the availability of expensive expert-annotated treebank data, not all annotations contribute equally to the training of the parsers. In this paper, we attempt to reduce the number of labeled examples needed to train a strong dependency parser using batch active learning (AL). In particular, we investigate whether enforcing diversity in the sampled batches, using determinantal point processes (DPPs), can improve over their diversity-agnostic counterparts. Simulation experiments on an English newswire corpus show that selecting diverse batches with DPPs is superior to strong selection strategies that do not enforce batch diversity, especially during the initial stages of the learning process. Additionally, our diversity-aware strategy is robust under a corpus duplication setting, where diversity-agnostic sampling strategies exhibit significant degradation.

We propose a practical instant question answering (QA) system on product pages of e-commerce services, where for each user query, relevant community question answer (CQA) pairs are retrieved. User queries and CQA pairs differ significantly in language characteristics making relevance learning difficult. Our proposed transformer-based model learns a robust relevance function by jointly learning unified syntactic and semantic representations without the need for human labeled data. This is achieved by distantly supervising our model by distilling from predictions of a syntactic matching system on user queries and simultaneously training with CQA pairs. Training with CQA pairs helps our model learning semantic QA relevance and distant supervision enables learning of syntactic features as well as the nuances of user querying language. Additionally, our model encodes queries and candidate responses independently allowing offline candidate embedding generation thereby minimizing the need for real-time transformer model execution. Consequently, our framework is able to scale to large e-commerce QA traffic. Extensive evaluation on user queries shows that our framework significantly outperforms both syntactic and semantic baselines in offline as well as large scale online A/B setups of a popular e-commerce service.