ACL 2023 论文列表

The large scale of pre-trained language models poses a challenge for their deployment on various devices, with a growing emphasis on methods to compress these models, particularly knowledge distillation. However, current knowledge distillation methods rely on the model’s intermediate layer features and the golden labels (also called hard labels), which usually require aligned model architecture and enough labeled data respectively. Moreover, the parameters of vocabulary are usually neglected in existing methods. To address these problems, we propose a general language model distillation (GLMD) method that performs two-stage word prediction distillation and vocabulary compression, which is simple and surprisingly shows extremely strong performance. Specifically, GLMD supports more general application scenarios by eliminating the constraints of dimension and structure between models and the need for labeled datasets through the absence of intermediate layers and golden labels. Meanwhile, based on the long-tailed distribution of word frequencies in the data, GLMD designs a strategy of vocabulary compression through decreasing vocabulary size instead of dimensionality. Experimental results show that our method outperforms 25 state-of-the-art methods on the SuperGLUE benchmark, achieving an average score that surpasses the best method by 3%.

Human-annotated labels and explanations are critical for training explainable NLP models. However, unlike human-annotated labels whose quality is easier to calibrate (e.g., with a majority vote), human-crafted free-form explanations can be quite subjective. Before blindly using them as ground truth to train ML models, a vital question needs to be asked: How do we evaluate a human-annotated explanation’s quality? In this paper, we build on the view that the quality of a human-annotated explanation can be measured based on its helpfulness (or impairment) to the ML models’ performance for the desired NLP tasks for which the annotations were collected. In comparison to the commonly used Simulatability score, we define a new metric that can take into consideration the helpfulness of an explanation for model performance at both fine-tuning and inference. With the help of a unified dataset format, we evaluated the proposed metric on five datasets (e.g., e-SNLI) against two model architectures (T5 and BART), and the results show that our proposed metric can objectively evaluate the quality of human-annotated explanations, while Simulatability falls short.

Social biases and stereotypes are embedded in our culture in part through their presence in our stories, as evidenced by the rich history of humanities and social science literature analyzing such biases in children stories. Because these analyses are often conducted manually and at a small scale, such investigations can benefit from the use of more recent natural language processing (NLP) methods that examine social bias in models and data corpora. Our work joins this interdisciplinary effort and makes a unique contribution by taking into account the event narrative structures when analyzing the social bias of stories. We propose a computational pipeline that automatically extracts a story’s temporal narrative verb-based event chain for each of its characters as well as character attributes such as gender. We also present a verb-based event annotation scheme that can facilitate bias analysis by including categories such as those that align with traditional stereotypes. Through a case study analyzing gender bias in fairy tales, we demonstrate that our framework can reveal bias in not only the unigram verb-based events in which female and male characters participate but also in the temporal narrative order of such event participation.

Reflection is a crucial counselling skill where the therapist conveys to the client their interpretation of what the client said. Language models have recently been used to generate reflections automatically, but human evaluation is challenging, particularly due to the cost of hiring experts. Laypeople-based evaluation is less expensive and easier to scale, but its quality is unknown for reflections. Therefore, we explore whether laypeople can be an alternative to experts in evaluating a fundamental quality aspect: coherence and context-consistency. We do so by asking a group of laypeople and a group of experts to annotate both synthetic reflections and human reflections from actual therapists. We find that both laypeople and experts are reliable annotators and that they have moderate-to-strong inter-group correlation, which shows that laypeople can be trusted for such evaluations. We also discover that GPT-3 mostly produces coherent and consistent reflections, and we explore changes in evaluation results when the source of synthetic reflections changes to GPT-3 from the less powerful GPT-2.

Developing monolingual large Pre-trained Language Models (PLMs) is shown to be very successful in handling different tasks in Natural Language Processing (NLP). In this work, we present AraMUS, the largest Arabic PLM with 11B parameters trained on 529GB of high-quality Arabic textual data. AraMUS achieves state-of-the-art performances on a diverse set of Arabic classification and generative tasks. Moreover, AraMUS shows impressive few-shot learning abilities compared with the best existing Arabic PLMs.

On-device automatic speech recognition systems face several challenges compared to server-based systems. They have to meet stricter constraints in terms of speed, disk size and memory while maintaining the same accuracy. Often they have to serve several ap- plications with different distributions at once, such as communicating with a virtual assistant and speech-to-text. The simplest solution to serve multiple applications is to build application-specific (language) models, but this leads to an increase in memory. Therefore, we explore different data- and architecture-driven language modeling approaches to build a single application-agnostic model. We propose two novel feed-forward architectures that find an optimal trade off between different on-device constraints. In comparison to the application-specific solution, one of our novel approaches reduces the disk size by half, while maintaining speed and accuracy of the original model.

Multimodal aspect-based sentiment analysis (MABSA) aims to extract aspects from text-image pairs and recognize their sentiments. Existing methods make great efforts to align the whole image to corresponding aspects. However, different regions of the image may relate to different aspects in the same sentence, and coarsely establishing image-aspect alignment will introduce noise to aspect-based sentiment analysis (i.e., visual noise). Besides, the sentiment of a specific aspect can also be interfered by descriptions of other aspects (i.e., textual noise). Considering the aforementioned noises, this paper proposes an Aspect-oriented Method (AoM) to detect aspect-relevant semantic and sentiment information. Specifically, an aspect-aware attention module is designed to simultaneously select textual tokens and image blocks that are semantically related to the aspects. To accurately aggregate sentiment information, we explicitly introduce sentiment embedding into AoM, and use a graph convolutional network to model the vision-text and text-text interaction. Extensive experiments demonstrate the superiority of AoM to existing methods.

Spoken Question Answering (QA) is a key feature of voice assistants, usually backed by multiple QA systems. Users ask questions via spontaneous speech that can contain disfluencies, errors, and informal syntax or phrasing. This is a major challenge in QA, causing unanswered questions or irrelevant answers, leading to bad user experiences. We analyze failed QA requests to identify core challenges: lexical gaps, proposition types, complex syntactic structure, and high specificity. We propose a Semantic Question Reformulation (SURF) model offering three linguistically-grounded operations (repair, syntactic reshaping, generalization) to rewrite questions to facilitate answering. Offline evaluation on 1M unanswered questions from a leading voice assistant shows that SURF significantly improves answer rates: up to 24% of previously unanswered questions obtain relevant answers (75%). Live deployment shows positive impact for millions of customers with unanswered questions; explicit relevance feedback shows high user satisfaction.

In open-domain question answering, due to the ambiguity of questions, multiple plausible answers may exist. To provide feasible answers to an ambiguous question,one approach is to directly predict all valid answers, but this can struggle with balancing relevance and diversity. An alternative is to gather candidate answers and aggregate them, but this method can be computationally costly and may neglect dependencies among answers. In this paper, we present AmbigPrompt to address the imperfections of existing approaches to answering ambiguous questions. Specifically, we integrate an answering model with a prompting model in an iterative manner. The prompting model adaptively tracks the reading process and progressively triggers the answering model to compose distinct and relevant answers. Additionally, we develop a task-specific post-pretraining approach for both the answering model and the prompting model, which greatly improves the performance of our framework. Empirical studies on two commonly-used open benchmarks show that AmbigPrompt achieves state-of-the-art or competitive results while using less memory and having a lower inference latency than competing approaches. Additionally, AmbigPrompt also performs well in low-resource settings.

The usefulness of part-of-speech tags for parsing has been heavily questioned due to the success of word-contextualized parsers. Yet, most studies are limited to coarse-grained tags and high quality written content; while we know little about their influence when it comes to models in production that face lexical errors. We expand these setups and design an adversarial attack to verify if the use of morphological information by parsers: (i) contributes to error propagation or (ii) if on the other hand it can play a role to correct mistakes that word-only neural parsers make. The results on 14 diverse UD treebanks show that under such attacks, for transition- and graph-based models their use contributes to degrade the performance even faster, while for the (lower-performing) sequence labeling parsers they are helpful. We also show that if morphological tags were utopically robust against lexical perturbations, they would be able to correct parsing mistakes.

Implicit discourse relation classification is a challenging task due to the absence of discourse connectives. To overcome this issue, we design an end-to-end neural model to explicitly generate discourse connectives for the task, inspired by the annotation process of PDTB. Specifically, our model jointly learns to generate discourse connectives between arguments and predict discourse relations based on the arguments and the generated connectives. To prevent our relation classifier from being misled by poor connectives generated at the early stage of training while alleviating the discrepancy between training and inference, we adopt Scheduled Sampling to the joint learning. We evaluate our method on three benchmarks, PDTB 2.0, PDTB 3.0, and PCC. Results show that our joint model significantly outperforms various baselines on three datasets, demonstrating its superiority for the task.

A series of datasets and models have been proposed for summaries generated for well-formatted documents such as news articles. Dialogue summaries, however, have been under explored. In this paper, we present the first dataset with fine-grained factual error annotations named DIASUMFACT. We define fine-grained factual error detection as a sentence-level multi-label classification problem, and weevaluate two state-of-the-art (SOTA) models on our dataset. Both models yield sub-optimal results, with a macro-averaged F1 score of around 0.25 over 6 error classes. We further propose an unsupervised model ENDERANKER via candidate ranking using pretrained encoder-decoder models. Our model performs on par with the SOTA models while requiring fewer resources. These observations confirm the challenges in detecting factual errors from dialogue summaries, which call for further studies, for which our dataset and results offer a solid foundation.

In this work, we present a natural language processing (NLP) pipeline for the identification, extraction and linking of Research Infrastructure (RI) used in scientific publications. Links between scientific equipment and publications where the equipment was used can support multiple use cases, such as evaluating the impact of RI investment, and supporting Open Science and research reproducibility. These links can also be used to establish a profile of the RI portfolio of each institution and associate each equipment with scientific output. The system we are describing here is already in production, and has been used to address real business use cases, some of which we discuss in this paper. The computational pipeline at the heart of the system comprises both supervised and unsupervised modules to detect the usage of research equipment by processing the full text of the articles. Additionally, we have created a knowledge graph of RI, which is utilized to annotate the articles with metadata. Finally, examples of the business value of the insights made possible by this NLP pipeline are illustrated.

Although recent neural models for coreference resolution have led to substantial improvements on benchmark datasets, it remains a challenge to successfully transfer these models to new target domains containing many out-of-vocabulary spans and requiring differing annotation schemes. Typical approaches involve continued training on annotated target-domain data, but obtaining annotations is costly and time-consuming. In this work, we show that adapting mention detection is the key component to successful domain adaptation of coreference models, rather than antecedent linking. We also show annotating mentions alone is nearly twice as fast as annotating full coreference chains. Based on these insights, we propose a method for efficiently adapting coreference models, which includes a high-precision mention detection objective and requires only mention annotations in the target domain. Extensive evaluation across three English coreference datasets: CoNLL-2012 (news/conversation), i2b2/VA (medical notes), and child welfare notes, reveals that our approach facilitates annotation-efficient transfer and results in a 7-14% improvement in average F1 without increasing annotator time.

Existing research has shown that a multilingual pre-trained language model fine-tuned with one (source) language also performs well on downstream tasks for non-source languages, even though no fine-tuning is done on these languages. However, there is a clear gap between the performance of the source language and that of the non-source languages. This paper analyzes the fine-tuning process, discovers when the performance gap changes and identifies which network weights affect the overall performance most. Additionally, the paper seeks to answer to what extent the gap can be reduced by reducing forgetting. Based on the analysis results, a method named Fine-tuning slow and fast with four training policies is proposed to address these issues. Experimental results show the proposed method outperforms baselines by a clear margin.

Understanding Transformer-based models has attracted significant attention, as they lie at the heart of recent technological advances across machine learning. While most interpretability methods rely on running models over inputs, recent work has shown that a zero-pass approach, where parameters are interpreted directly without a forward/backward pass is feasible for some Transformer parameters, and for two-layer attention networks. In this work, we present a theoretical analysis where all parameters of a trained Transformer are interpreted by projecting them into the embedding space, that is, the space of vocabulary items they operate on. We derive a simple theoretical framework to support our arguments and provide ample evidence for its validity. First, an empirical analysis showing that parameters of both pretrained and fine-tuned models can be interpreted in embedding space. Second, we present two applications of our framework: (a) aligning the parameters of different models that share a vocabulary, and (b) constructing a classifier without training by “translating” the parameters of a fine-tuned classifier to parameters of a different model that was only pretrained. Overall, our findings open the door to interpretation methods that, at least in part, abstract away from model specifics and operate in the embedding space only.

Textual representations based on pre-trained language models are key, especially in few-shot learning scenarios. What makes a representation good for text classification? Is it due to the geometric properties of the space or because it is well aligned with the task? We hypothesize the second claim. To test it, we develop a task alignment score based on hierarchical clustering that measures alignment at different levels of granularity. Our experiments on text classification validate our hypothesis by showing that task alignment can explain the classification performance of a given representation.

Tabular data analysis is performed everyday across various domains. It requires an accurate understanding of field semantics to correctly operate on table fields and find common patterns in daily analysis. In this paper, we introduce the AnaMeta dataset, a collection of 467k tables with derived supervision labels for four types of commonly used field metadata: measure/dimension dichotomy, common field roles, semantic field type, and default aggregation function. We evaluate a wide range of models for inferring metadata as the benchmark. We also propose a multi-encoder framework, called KDF, which improves the metadata understanding capability of tabular models by incorporating distribution and knowledge information. Furthermore, we propose four interfaces for incorporating field metadata into downstream analysis tasks.

We present an efficient and reliable approach to Natural Language Querying (NLQ) on databases (DB) which is not based on text-to-SQL type semantic parsing. Our approach simplifies the NLQ on structured data problem to the following “bread and butter” NLP tasks: (a) Domain classification, for choosing which DB table to query, whether the question is out-of-scope (b) Multi-head slot/entity extraction (SE) to extract the field criteria and other attributes such as its role (filter, sort etc) from the raw text and (c) Slot value disambiguation (SVD) to resolve/normalize raw spans from SE to format suitable to query a DB. This is a general purpose, DB language agnostic approach and the output can be used to query any DB and return results to the user. Also each of these tasks is extremely well studied, mature, easier to collect data for and enables better error analysis by tracing problems to specific components when something goes wrong.