ACL 2021 Papers

In recent years, we have seen a colossal effort in pre-training multilingual text encoders using large-scale corpora in many languages to facilitate cross-lingual transfer learning. However, due to typological differences across languages, the cross-lingual transfer is challenging. Nevertheless, language syntax, e.g., syntactic dependencies, can bridge the typological gap. Previous works have shown that pre-trained multilingual encoders, such as mBERT (CITATION), capture language syntax, helping cross-lingual transfer. This work shows that explicitly providing language syntax and training mBERT using an auxiliary objective to encode the universal dependency tree structure helps cross-lingual transfer. We perform rigorous experiments on four NLP tasks, including text classification, question answering, named entity recognition, and task-oriented semantic parsing. The experiment results show that syntax-augmented mBERT improves cross-lingual transfer on popular benchmarks, such as PAWS-X and MLQA, by 1.4 and 1.6 points on average across all languages. In the generalized transfer setting, the performance boosted significantly, with 3.9 and 3.1 points on average in PAWS-X and MLQA.

We study the problem of leveraging the syntactic structure of text to enhance pre-trained models such as BERT and RoBERTa. Existing methods utilize syntax of text either in the pre-training stage or in the fine-tuning stage, so that they suffer from discrepancy between the two stages. Such a problem would lead to the necessity of having human-annotated syntactic information, which limits the application of existing methods to broader scenarios. To address this, we present a model that utilizes the syntax of text in both pre-training and fine-tuning stages. Our model is based on Transformer with a syntax-aware attention layer that considers the dependency tree of the text. We further introduce a new pre-training task of predicting the syntactic distance among tokens in the dependency tree. We evaluate the model on three downstream tasks, including relation classification, entity typing, and question answering. Results show that our model achieves state-of-the-art performance on six public benchmark datasets. We have two major findings. First, we demonstrate that infusing automatically produced syntax of text improves pre-trained models. Second, global syntactic distances among tokens bring larger performance gains compared to local head relations between contiguous tokens.

While the use of character models has been popular in NLP applications, it has not been explored much in the context of psycholinguistic modeling. This paper presents a character model that can be applied to a structural parser-based processing model to calculate word generation probabilities. Experimental results show that surprisal estimates from a structural processing model using this character model deliver substantially better fits to self-paced reading, eye-tracking, and fMRI data than those from large-scale language models trained on much more data. This may suggest that the proposed processing model provides a more humanlike account of sentence processing, which assumes a larger role of morphology, phonotactics, and orthographic complexity than was previously thought.

Open pit mines left many regions worldwide inhospitable or uninhabitable. Many sites are left behind in a hazardous or contaminated state, show remnants of waste, or have other restrictions imposed upon them, e.g., for the protection of human or nature. Such information has to be permanently managed in order to reuse those areas in the future. In this work we present and evaluate an automated workflow for supporting the post-mining management of former lignite open pit mines in the eastern part of Germany, where prior to any planned land reuse, aforementioned information has to be acquired to ensure the safety and validity of such an endeavor. Usually, this information is found in expert reports, either in the form of paper documents, or in the best case as digitized unstructured text—all of them in German language. However, due to the size and complexity of these documents, any inquiry is tedious and time-consuming, thereby slowing down or even obstructing the reuse of related areas. Since no training data is available, we employ active learning in order to perform multi-label sentence classification for two categories of restrictions and seven categories of topics. The final system integrates optical character recognition (OCR), active-learning-based text classification, and geographic information system visualization in order to effectively extract, query, and visualize this information for any area of interest. Active learning and text classification results are twofold: Whereas the restriction categories were reasonably accurate (>0.85 F1), the seven topic-oriented categories seemed to be complex even for human annotators and achieved mediocre evaluation scores (<0.70 F1).

We present an annotation approach to capturing emotional and cognitive empathy in student-written peer reviews on business models in German. We propose an annotation scheme that allows us to model emotional and cognitive empathy scores based on three types of review components. Also, we conducted an annotation study with three annotators based on 92 student essays to evaluate our annotation scheme. The obtained inter-rater agreement of α=0.79 for the components and the multi-π=0.41 for the empathy scores indicate that the proposed annotation scheme successfully guides annotators to a substantial to moderate agreement. Moreover, we trained predictive models to detect the annotated empathy structures and embedded them in an adaptive writing support system for students to receive individual empathy feedback independent of an instructor, time, and location. We evaluated our tool in a peer learning exercise with 58 students and found promising results for perceived empathy skill learning, perceived feedback accuracy, and intention to use. Finally, we present our freely available corpus of 500 empathy-annotated, student-written peer reviews on business models and our annotation guidelines to encourage future research on the design and development of empathy support systems.

How does the input segmentation of pretrained language models (PLMs) affect their interpretations of complex words? We present the first study investigating this question, taking BERT as the example PLM and focusing on its semantic representations of English derivatives. We show that PLMs can be interpreted as serial dual-route models, i.e., the meanings of complex words are either stored or else need to be computed from the subwords, which implies that maximally meaningful input tokens should allow for the best generalization on new words. This hypothesis is confirmed by a series of semantic probing tasks on which DelBERT (Derivation leveraging BERT), a model with derivational input segmentation, substantially outperforms BERT with WordPiece segmentation. Our results suggest that the generalization capabilities of PLMs could be further improved if a morphologically-informed vocabulary of input tokens were used.

The Lottery Ticket Hypothesis suggests that an over-parametrized network consists of ”lottery tickets”, and training a certain collection of them (i.e., a subnetwork) can match the performance of the full model. In this paper, we study such a collection of tickets, which is referred to as ”winning tickets”, in extremely over-parametrized models, e.g., pre-trained language models. We observe that at certain compression ratios, the generalization performance of the winning tickets can not only match but also exceed that of the full model. In particular, we observe a phase transition phenomenon: As the compression ratio increases, generalization performance of the winning tickets first improves then deteriorates after a certain threshold. We refer to the tickets on the threshold as ”super tickets”. We further show that the phase transition is task and model dependent — as the model size becomes larger and the training data set becomes smaller, the transition becomes more pronounced. Our experiments on the GLUE benchmark show that the super tickets improve single task fine-tuning by 0.9 points on BERT-base and 1.0 points on BERT-large, in terms of task-average score. We also demonstrate that adaptively sharing the super tickets across tasks benefits multi-task learning.

Active Learning (AL) has been successfully applied to Deep Learning in order to drastically reduce the amount of data required to achieve high performance. Previous works have shown that lightweight architectures for Named Entity Recognition (NER) can achieve optimal performance with only 25% of the original training data. However, these methods do not exploit the sequential nature of language and the heterogeneity of uncertainty within each instance, requiring the labelling of whole sentences. Additionally, this standard method requires that the annotator has access to the full sentence when labelling. In this work, we overcome these limitations by allowing the AL algorithm to query subsequences within sentences, and propagate their labels to other sentences. We achieve highly efficient results on OntoNotes 5.0, only requiring 13% of the original training data, and CoNLL 2003, requiring only 27%. This is an improvement of 39% and 37% compared to querying full sentences.

Every natural text is written in some style. Style is formed by a complex combination of different stylistic factors, including formality markers, emotions, metaphors, etc. One cannot form a complete understanding of a text without considering these factors. The factors combine and co-vary in complex ways to form styles. Studying the nature of the covarying combinations sheds light on stylistic language in general, sometimes called cross-style language understanding. This paper provides the benchmark corpus (XSLUE) that combines existing datasets and collects a new one for sentence-level cross-style language understanding and evaluation. The benchmark contains text in 15 different styles under the proposed four theoretical groupings: figurative, personal, affective, and interpersonal groups. For valid evaluation, we collect an additional diagnostic set by annotating all 15 styles on the same text. Using XSLUE, we propose three interesting cross-style applications in classification, correlation, and generation. First, our proposed cross-style classifier trained with multiple styles together helps improve overall classification performance against individually-trained style classifiers. Second, our study shows that some styles are highly dependent on each other in human-written text. Finally, we find that combinations of some contradictive styles likely generate stylistically less appropriate text. We believe our benchmark and case studies help explore interesting future directions for cross-style research. The preprocessed datasets and code are publicly available.

Misinformation has recently become a well-documented matter of public concern. Existing studies on this topic have hitherto adopted a coarse concept of misinformation, which incorporates a broad spectrum of story types ranging from political conspiracies to misinterpreted pranks. This paper aims to structurize these misinformation stories by leveraging fact-check articles. Our intuition is that key phrases in a fact-check article that identify the misinformation type(s) (e.g., doctored images, urban legends) also act as rationales that determine the verdict of the fact-check (e.g., false). We experiment on rationalized models with domain knowledge as weak supervision to extract these phrases as rationales, and then cluster semantically similar rationales to summarize prevalent misinformation types. Using archived fact-checks from Snopes.com, we identify ten types of misinformation stories. We discuss how these types have evolved over the last ten years and compare their prevalence between the 2016/2020 US presidential elections and the H1N1/COVID-19 pandemics.

Structured sentiment analysis attempts to extract full opinion tuples from a text, but over time this task has been subdivided into smaller and smaller sub-tasks, e.g., target extraction or targeted polarity classification. We argue that this division has become counterproductive and propose a new unified framework to remedy the situation. We cast the structured sentiment problem as dependency graph parsing, where the nodes are spans of sentiment holders, targets and expressions, and the arcs are the relations between them. We perform experiments on five datasets in four languages (English, Norwegian, Basque, and Catalan) and show that this approach leads to strong improvements over state-of-the-art baselines. Our analysis shows that refining the sentiment graphs with syntactic dependency information further improves results.

Large pre-trained language models achieve state-of-the-art results when fine-tuned on downstream NLP tasks. However, they almost exclusively focus on text-only representation, while neglecting cell-level layout information that is important for form image understanding. In this paper, we propose a new pre-training approach, StructuralLM, to jointly leverage cell and layout information from scanned documents. Specifically, we pre-train StructuralLM with two new designs to make the most of the interactions of cell and layout information: 1) each cell as a semantic unit; 2) classification of cell positions. The pre-trained StructuralLM achieves new state-of-the-art results in different types of downstream tasks, including form understanding (from 78.95 to 85.14), document visual question answering (from 72.59 to 83.94) and document image classification (from 94.43 to 96.08).

Pre-trained language models (PrLMs) have demonstrated superior performance due to their strong ability to learn universal language representations from self-supervised pre-training. However, even with the help of the powerful PrLMs, it is still challenging to effectively capture task-related knowledge from dialogue texts which are enriched by correlations among speaker-aware utterances. In this work, we present SPIDER, Structural Pre-traIned DialoguE Reader, to capture dialogue exclusive features. To simulate the dialogue-like features, we propose two training objectives in addition to the original LM objectives: 1) utterance order restoration, which predicts the order of the permuted utterances in dialogue context; 2) sentence backbone regularization, which regularizes the model to improve the factual correctness of summarized subject-verb-object triplets. Experimental results on widely used dialogue benchmarks verify the effectiveness of the newly introduced self-supervised tasks.

Knowledge distillation is a critical technique to transfer knowledge between models, typically from a large model (the teacher) to a more fine-grained one (the student). The objective function of knowledge distillation is typically the cross-entropy between the teacher and the student’s output distributions. However, for structured prediction problems, the output space is exponential in size; therefore, the cross-entropy objective becomes intractable to compute and optimize directly. In this paper, we derive a factorized form of the knowledge distillation objective for structured prediction, which is tractable for many typical choices of the teacher and student models. In particular, we show the tractability and empirical effectiveness of structural knowledge distillation between sequence labeling and dependency parsing models under four different scenarios: 1) the teacher and student share the same factorization form of the output structure scoring function; 2) the student factorization produces more fine-grained substructures than the teacher factorization; 3) the teacher factorization produces more fine-grained substructures than the student factorization; 4) the factorization forms from the teacher and the student are incompatible.

Transformer-based language models pre-trained on large amounts of text data have proven remarkably successful in learning generic transferable linguistic representations. Here we study whether structural guidance leads to more human-like systematic linguistic generalization in Transformer language models without resorting to pre-training on very large amounts of data. We explore two general ideas. The “Generative Parsing” idea jointly models the incremental parse and word sequence as part of the same sequence modeling task. The “Structural Scaffold” idea guides the language model’s representation via additional structure loss that separately predicts the incremental constituency parse. We train the proposed models along with a vanilla Transformer language model baseline on a 14 million-token and a 46 million-token subset of the BLLIP dataset, and evaluate models’ syntactic generalization performances on SG Test Suites and sized BLiMP. Experiment results across two benchmarks suggest converging evidence that generative structural supervisions can induce more robust and humanlike linguistic generalization in Transformer language models without the need for data intensive pre-training.

There are two major classes of natural language grammars — the dependency grammar that models one-to-one correspondences between words and the constituency grammar that models the assembly of one or several corresponded words. While previous unsupervised parsing methods mostly focus on only inducing one class of grammars, we introduce a novel model, StructFormer, that can induce dependency and constituency structure at the same time. To achieve this, we propose a new parsing framework that can jointly generate a constituency tree and dependency graph. Then we integrate the induced dependency relations into the transformer, in a differentiable manner, through a novel dependency-constrained self-attention mechanism. Experimental results show that our model can achieve strong results on unsupervised constituency parsing, unsupervised dependency parsing, and masked language modeling at the same time.

Auditing NLP systems for computational harms like surfacing stereotypes is an elusive goal. Several recent efforts have focused on benchmark datasets consisting of pairs of contrastive sentences, which are often accompanied by metrics that aggregate an NLP system’s behavior on these pairs into measurements of harms. We examine four such benchmarks constructed for two NLP tasks: language modeling and coreference resolution. We apply a measurement modeling lens—originating from the social sciences—to inventory a range of pitfalls that threaten these benchmarks’ validity as measurement models for stereotyping. We find that these benchmarks frequently lack clear articulations of what is being measured, and we highlight a range of ambiguities and unstated assumptions that affect how these benchmarks conceptualize and operationalize stereotyping.

A stereotype is an over-generalized belief about a particular group of people, e.g., Asians are good at math or African Americans are athletic. Such beliefs (biases) are known to hurt target groups. Since pretrained language models are trained on large real-world data, they are known to capture stereotypical biases. It is important to quantify to what extent these biases are present in them. Although this is a rapidly growing area of research, existing literature lacks in two important aspects: 1) they mainly evaluate bias of pretrained language models on a small set of artificial sentences, even though these models are trained on natural data 2) current evaluations focus on measuring bias without considering the language modeling ability of a model, which could lead to misleading trust on a model even if it is a poor language model. We address both these problems. We present StereoSet, a large-scale natural English dataset to measure stereotypical biases in four domains: gender, profession, race, and religion. We contrast both stereotypical bias and language modeling ability of popular models like BERT, GPT-2, RoBERTa, and XLnet. We show that these models exhibit strong stereotypical biases. Our data and code are available at https://stereoset.mit.edu.

Relational triple extraction is critical to understanding massive text corpora and constructing large-scale knowledge graph, which has attracted increasing research interest. However, existing studies still face some challenging issues, including information loss, error propagation and ignoring the interaction between entity and relation. To intuitively explore the above issues and address them, in this paper, we provide a revealing insight into relational triple extraction from a stereoscopic perspective, which rationalizes the occurrence of these issues and exposes the shortcomings of existing methods. Further, a novel model is proposed for relational triple extraction, which maps relational triples to a three-dimension (3-D) space and leverages three decoders to extract them, aimed at simultaneously handling the above issues. A series of experiments are conducted on five public datasets, demonstrating that the proposed model outperforms the recent advanced baselines.