ACL 2022 Papers

Input saliency methods have recently become a popular tool for explaining predictions of deep learning models in NLP. Nevertheless, there has been little work investigating methods for aggregating prediction-level explanations to the class level, nor has a framework for evaluating such class explanations been established. We explore explanations based on XLM-R and the Integrated Gradients input attribution method, and propose 1) the Stable Attribution Class Explanation method (SACX) to extract keyword lists of classes in text classification tasks, and 2) a framework for the systematic evaluation of the keyword lists. We find that explanations of individual predictions are prone to noise, but that stable explanations can be effectively identified through repeated training and explanation. We evaluate on web register data and show that the class explanations are linguistically meaningful and distinguishing of the classes.

The performance of multilingual pretrained models is highly dependent on the availability of monolingual or parallel text present in a target language. Thus, the majority of the world’s languages cannot benefit from recent progress in NLP as they have no or limited textual data. To expand possibilities of using NLP technology in these under-represented languages, we systematically study strategies that relax the reliance on conventional language resources through the use of bilingual lexicons, an alternative resource with much better language coverage. We analyze different strategies to synthesize textual or labeled data using lexicons, and how this data can be combined with monolingual or parallel text when available. For 19 under-represented languages across 3 tasks, our methods lead to consistent improvements of up to 5 and 15 points with and without extra monolingual text respectively. Overall, our study highlights how NLP methods can be adapted to thousands more languages that are under-served by current technology.

To understand a story with multiple events, it is important to capture the proper relations across these events. However, existing event relation extraction (ERE) framework regards it as a multi-class classification task and do not guarantee any coherence between different relation types, such as anti-symmetry. If a phone line “died” after “storm”, then it is obvious that the “storm” happened before the “died”. Current framework of event relation extraction do not guarantee this coherence and thus enforces it via constraint loss function (Wang et al., 2020). In this work, we propose to modify the underlying ERE model to guarantee coherence by representing each event as a box representation (BERE) without applying explicit constraints. From our experiments, BERE also shows stronger conjunctive constraint satisfaction while performing on par or better in F1 compared to previous models with constraint injection.

Open-ended text generation tasks, such as dialogue generation and story completion, require models to generate a coherent continuation given limited preceding context. The open-ended nature of these tasks brings new challenges to the neural auto-regressive text generators nowadays. Despite these neural models are good at producing human-like text, it is difficult for them to arrange causalities and relations between given facts and possible ensuing events. To bridge this gap, we propose a novel two-stage method which explicitly arranges the ensuing events in open-ended text generation. Our approach can be understood as a specially-trained coarse-to-fine algorithm, where an event transition planner provides a “coarse” plot skeleton and a text generator in the second stage refines the skeleton. Experiments on two open-ended text generation tasks demonstrate that our proposed method effectively improves the quality of the generated text, especially in coherence and diversity. We will release the codes to the community for further exploration.

Automated simplification models aim to make input texts more readable. Such methods have the potential to make complex information accessible to a wider audience, e.g., providing access to recent medical literature which might otherwise be impenetrable for a lay reader. However, such models risk introducing errors into automatically simplified texts, for instance by inserting statements unsupported by the corresponding original text, or by omitting key information. Providing more readable but inaccurate versions of texts may in many cases be worse than providing no such access at all. The problem of factual accuracy (and the lack thereof) has received heightened attention in the context of summarization models, but the factuality of automatically simplified texts has not been investigated. We introduce a taxonomy of errors that we use to analyze both references drawn from standard simplification datasets and state-of-the-art model outputs. We find that errors often appear in both that are not captured by existing evaluation metrics, motivating a need for research into ensuring the factual accuracy of automated simplification models.

Several natural language processing (NLP) tasks are defined as a classification problem in its most complex form: Multi-label Hierarchical Extreme classification, in which items may be associated with multiple classes from a set of thousands of possible classes organized in a hierarchy and with a highly unbalanced distribution both in terms of class frequency and the number of labels per item. We analyze the state of the art of evaluation metrics based on a set of formal properties and we define an information theoretic based metric inspired by the Information Contrast Model (ICM). Experiments on synthetic data and a case study on real data show the suitability of the ICM for such scenarios.

Several high-profile events, such as the mass testing of emotion recognition systems on vulnerable sub-populations and using question answering systems to make moral judgments, have highlighted how technology will often lead to more adverse outcomes for those that are already marginalized. At issue here are not just individual systems and datasets, but also the AI tasks themselves. In this position paper, I make a case for thinking about ethical considerations not just at the level of individual models and datasets, but also at the level of AI tasks. I will present a new form of such an effort, Ethics Sheets for AI Tasks, dedicated to fleshing out the assumptions and ethical considerations hidden in how a task is commonly framed and in the choices we make regarding the data, method, and evaluation. I will also present a template for ethics sheets with 50 ethical considerations, using the task of emotion recognition as a running example. Ethics sheets are a mechanism to engage with and document ethical considerations before building datasets and systems. Similar to survey articles, a small number of carefully created ethics sheets can serve numerous researchers and developers.

Shannon entropy is often a quantity of interest to linguists studying the communicative capacity of human language. However, entropymust typically be estimated from observed data because researchers do not have access to the underlying probability distribution. While entropy estimation is a well-studied problem in other fields, there is not yet a comprehensive exploration of the efficacy of entropy estimators for use with linguistic data. In this work, we fill this void, studying the empirical effectiveness of different entropy estimators for linguistic distributions. In a replication of two recent information-theoretic linguistic studies, we find evidence that the reported effect size is over-estimated due to over-reliance on poor entropy estimators. We end this paper with a concrete recommendation for the entropy estimators that should be used in future linguistic studies.

Natural Language Processing (NLP) models risk overfitting to specific terms in the training data, thereby reducing their performance, fairness, and generalizability. E.g., neural hate speech detection models are strongly influenced by identity terms like gay, or women, resulting in false positives, severe unintended bias, and lower performance. Most mitigation techniques use lists of identity terms or samples from the target domain during training. However, this approach requires a-priori knowledge and introduces further bias if important terms are neglected. Instead, we propose a knowledge-free Entropy-based Attention Regularization (EAR) to discourage overfitting to training-specific terms. An additional objective function penalizes tokens with low self-attention entropy. We fine-tune BERT via EAR: the resulting model matches or exceeds state-of-the-art performance for hate speech classification and bias metrics on three benchmark corpora in English and Italian.EAR also reveals overfitting terms, i.e., terms most likely to induce bias, to help identify their effect on the model, task, and predictions.

In this paper, we propose an entity-based neural local coherence model which is linguistically more sound than previously proposed neural coherence models. Recent neural coherence models encode the input document using large-scale pretrained language models. Hence their basis for computing local coherence are words and even sub-words. The analysis of their output shows that these models frequently compute coherence on the basis of connections between (sub-)words which, from a linguistic perspective, should not play a role. Still, these models achieve state-of-the-art performance in several end applications. In contrast to these models, we compute coherence on the basis of entities by constraining the input to noun phrases and proper names. This provides us with an explicit representation of the most important items in sentences leading to the notion of focus. This brings our model linguistically in line with pre-neural models of computing coherence. It also gives us better insight into the behaviour of the model thus leading to better explainability. Our approach is also in accord with a recent study (O’Connor and Andreas, 2021), which shows that most usable information is captured by nouns and verbs in transformer-based language models. We evaluate our model on three downstream tasks showing that it is not only linguistically more sound than previous models but also that it outperforms them in end applications.

Typed entailment graphs try to learn the entailment relations between predicates from text and model them as edges between predicate nodes. The construction of entailment graphs usually suffers from severe sparsity and unreliability of distributional similarity. We propose a two-stage method, Entailment Graph with Textual Entailment and Transitivity (EGT2). EGT2 learns the local entailment relations by recognizing the textual entailment between template sentences formed by typed CCG-parsed predicates. Based on the generated local graph, EGT2 then uses three novel soft transitivity constraints to consider the logical transitivity in entailment structures. Experiments on benchmark datasets show that EGT2 can well model the transitivity in entailment graph to alleviate the sparsity, and leads to signifcant improvement over current state-of-the-art methods.

Controllable summarization aims to provide summaries that take into account user-specified aspects and preferences to better assist them with their information need, as opposed to the standard summarization setup which build a single generic summary of a document. We introduce a human-annotated data set EntSUM for controllable summarization with a focus on named entities as the aspects to control. We conduct an extensive quantitative analysis to motivate the task of entity-centric summarization and show that existing methods for controllable summarization fail to generate entity-centric summaries. We propose extensions to state-of-the-art summarization approaches that achieve substantially better results on our data set. Our analysis and results show the challenging nature of this task and of the proposed data set.

In this paper, we investigate improvements to the GEC sequence tagging architecture with a focus on ensembling of recent cutting-edge Transformer-based encoders in Large configurations. We encourage ensembling models by majority votes on span-level edits because this approach is tolerant to the model architecture and vocabulary size. Our best ensemble achieves a new SOTA result with an F0.5 score of 76.05 on BEA-2019 (test), even without pre-training on synthetic datasets. In addition, we perform knowledge distillation with a trained ensemble to generate new synthetic training datasets, “Troy-Blogs” and “Troy-1BW”. Our best single sequence tagging model that is pretrained on the generated Troy- datasets in combination with the publicly available synthetic PIE dataset achieves a near-SOTA result with an F0.5 score of 73.21 on BEA-2019 (test). The code, datasets, and trained models are publicly available.

We study how to enhance text representation via textual commonsense. We point out that commonsense has the nature of domain discrepancy. Namely, commonsense has different data formats and is domain-independent from the downstream task. This nature brings challenges to introducing commonsense in general text understanding tasks. A typical method of introducing textual knowledge is continuing pre-training over the commonsense corpus. However, it will cause catastrophic forgetting to the downstream task due to the domain discrepancy. In addition, previous methods of directly using textual descriptions as extra input information cannot apply to large-scale commonsense. In this paper, we propose to use large-scale out-of-domain commonsense to enhance text representation. In order to effectively incorporate the commonsense, we proposed OK-Transformer (Out-of-domain Knowledge enhanced Transformer). OK-Transformer effectively integrates commonsense descriptions and enhances them to the target text representation. In addition, OK-Transformer can adapt to the Transformer-based language models (e.g. BERT, RoBERTa) for free, without pre-training on large-scale unsupervised corpora. We have verified the effectiveness of OK-Transformer in multiple applications such as commonsense reasoning, general text classification, and low-resource commonsense settings.

Cross-lingual natural language inference (XNLI) is a fundamental task in cross-lingual natural language understanding. Recently this task is commonly addressed by pre-trained cross-lingual language models. Existing methods usually enhance pre-trained language models with additional data, such as annotated parallel corpora. These additional data, however, are rare in practice, especially for low-resource languages. Inspired by recent promising results achieved by prompt-learning, this paper proposes a novel prompt-learning based framework for enhancing XNLI. It reformulates the XNLI problem to a masked language modeling problem by constructing cloze-style questions through cross-lingual templates. To enforce correspondence between different languages, the framework augments a new question for every question using a sampled template in another language and then introduces a consistency loss to make the answer probability distribution obtained from the new question as similar as possible with the corresponding distribution obtained from the original question. Experimental results on two benchmark datasets demonstrate that XNLI models enhanced by our proposed framework significantly outperform original ones under both the full-shot and few-shot cross-lingual transfer settings.

Chinese pre-trained language models usually exploit contextual character information to learn representations, while ignoring the linguistics knowledge, e.g., word and sentence information. Hence, we propose a task-free enhancement module termed as Heterogeneous Linguistics Graph (HLG) to enhance Chinese pre-trained language models by integrating linguistics knowledge. Specifically, we construct a hierarchical heterogeneous graph to model the characteristics linguistics structure of Chinese language, and conduct a graph-based method to summarize and concretize information on different granularities of Chinese linguistics hierarchies. Experimental results demonstrate our model has the ability to improve the performance of vanilla BERT, BERTwwm and ERNIE 1.0 on 6 natural language processing tasks with 10 benchmark datasets. Further, the detailed experimental analyses have proven that this kind of modelization achieves more improvements compared with previous strong baseline MWA. Meanwhile, our model introduces far fewer parameters (about half of MWA) and the training/inference speed is about 7x faster than MWA.

Aspect Sentiment Triplet Extraction (ASTE) is an emerging sentiment analysis task. Most of the existing studies focus on devising a new tagging scheme that enables the model to extract the sentiment triplets in an end-to-end fashion. However, these methods ignore the relations between words for ASTE task. In this paper, we propose an Enhanced Multi-Channel Graph Convolutional Network model (EMC-GCN) to fully utilize the relations between words. Specifically, we first define ten types of relations for ASTE task, and then adopt a biaffine attention module to embed these relations as an adjacent tensor between words in a sentence. After that, our EMC-GCN transforms the sentence into a multi-channel graph by treating words and the relation adjacent tensor as nodes and edges, respectively. Thus, relation-aware node representations can be learnt. Furthermore, we consider diverse linguistic features to enhance our EMC-GCN model. Finally, we design an effective refining strategy on EMC-GCN for word-pair representation refinement, which considers the implicit results of aspect and opinion extraction when determining whether word pairs match or not. Extensive experimental results on the benchmark datasets demonstrate that the effectiveness and robustness of our proposed model, which outperforms state-of-the-art methods significantly.

Code switching (CS) refers to the phenomenon of interchangeably using words and phrases from different languages. CS can pose significant accuracy challenges to NLP, due to the often monolingual nature of the underlying systems. In this work, we focus on CS in the context of English/Spanish conversations for the task of speech translation (ST), generating and evaluating both transcript and translation. To evaluate model performance on this task, we create a novel ST corpus derived from existing public data sets. We explore various ST architectures across two dimensions: cascaded (transcribe then translate) vs end-to-end (jointly transcribe and translate) and unidirectional (source -> target) vs bidirectional (source <-> target). We show that our ST architectures, and especially our bidirectional end-to-end architecture, perform well on CS speech, even when no CS training data is used.

In this paper, we bring a new way of digesting news content by introducing the task of segmenting a news article into multiple sections and generating the corresponding summary to each section. We make two contributions towards this new task. First, we create and make available a dataset, SegNews, consisting of 27k news articles with sections and aligned heading-style section summaries. Second, we propose a novel segmentation-based language generation model adapted from pre-trained language models that can jointly segment a document and produce the summary for each section. Experimental results on SegNews demonstrate that our model can outperform several state-of-the-art sequence-to-sequence generation models for this new task.