ACL 2022 论文列表

Writing is, by nature, a strategic, adaptive, and, more importantly, an iterative process. A crucial part of writing is editing and revising the text. Previous works on text revision have focused on defining edit intention taxonomies within a single domain or developing computational models with a single level of edit granularity, such as sentence-level edits, which differ from human’s revision cycles. This work describes IteraTeR: the first large-scale, multi-domain, edit-intention annotated corpus of iteratively revised text. In particular, IteraTeR is collected based on a new framework to comprehensively model the iterative text revisions that generalizes to a variety of domains, edit intentions, revision depths, and granularities. When we incorporate our annotated edit intentions, both generative and action-based text revision models significantly improve automatic evaluations. Through our work, we better understand the text revision process, making vital connections between edit intentions and writing quality, enabling the creation of diverse corpora to support computational modeling of iterative text revisions.

Knowledge base (KB) embeddings have been shown to contain gender biases. In this paper, we study two questions regarding these biases: how to quantify them, and how to trace their origins in KB? Specifically, first, we develop two novel bias measures respectively for a group of person entities and an individual person entity. Evidence of their validity is observed by comparison with real-world census data. Second, we use the influence function to inspect the contribution of each triple in KB to the overall group bias. To exemplify the potential applications of our study, we also present two strategies (by adding and removing KB triples) to mitigate gender biases in KB embeddings.

We present a new dataset containing 10K human-annotated games of Go and show how these natural language annotations can be used as a tool for model interpretability. Given a board state and its associated comment, our approach uses linear probing to predict mentions of domain-specific terms (e.g., ko, atari) from the intermediate state representations of game-playing agents like AlphaGo Zero. We find these game concepts are nontrivially encoded in two distinct policy networks, one trained via imitation learning and another trained via reinforcement learning. Furthermore, mentions of domain-specific terms are most easily predicted from the later layers of both models, suggesting that these policy networks encode high-level abstractions similar to those used in the natural language annotations.

Gender bias is largely recognized as a problematic phenomenon affecting language technologies, with recent studies underscoring that it might surface differently across languages. However, most of current evaluation practices adopt a word-level focus on a narrow set of occupational nouns under synthetic conditions. Such protocols overlook key features of grammatical gender languages, which are characterized by morphosyntactic chains of gender agreement, marked on a variety of lexical items and parts-of-speech (POS). To overcome this limitation, we enrich the natural, gender-sensitive MuST-SHE corpus (Bentivogli et al., 2020) with two new linguistic annotation layers (POS and agreement chains), and explore to what extent different lexical categories and agreement phenomena are impacted by gender skews. Focusing on speech translation, we conduct a multifaceted evaluation on three language directions (English-French/Italian/Spanish), with models trained on varying amounts of data and different word segmentation techniques. By shedding light on model behaviours, gender bias, and its detection at several levels of granularity, our findings emphasize the value of dedicated analyses beyond aggregated overall results.

Uncertainty estimation (UE) of model predictions is a crucial step for a variety of tasks such as active learning, misclassification detection, adversarial attack detection, out-of-distribution detection, etc. Most of the works on modeling the uncertainty of deep neural networks evaluate these methods on image classification tasks. Little attention has been paid to UE in natural language processing. To fill this gap, we perform a vast empirical investigation of state-of-the-art UE methods for Transformer models on misclassification detection in named entity recognition and text classification tasks and propose two computationally efficient modifications, one of which approaches or even outperforms computationally intensive methods.

In many natural language processing (NLP) tasks the same input (e.g. source sentence) can have multiple possible outputs (e.g. translations). To analyze how this ambiguity (also known as intrinsic uncertainty) shapes the distribution learned by neural sequence models we measure sentence-level uncertainty by computing the degree of overlap between references in multi-reference test sets from two different NLP tasks: machine translation (MT) and grammatical error correction (GEC). At both the sentence- and the task-level, intrinsic uncertainty has major implications for various aspects of search such as the inductive biases in beam search and the complexity of exact search. In particular, we show that well-known pathologies such as a high number of beam search errors, the inadequacy of the mode, and the drop in system performance with large beam sizes apply to tasks with high level of ambiguity such as MT but not to less uncertain tasks such as GEC. Furthermore, we propose a novel exact n-best search algorithm for neural sequence models, and show that intrinsic uncertainty affects model uncertainty as the model tends to overly spread out the probability mass for uncertain tasks and sentences.

Vision-Language Pre-training (VLP) has achieved impressive performance on various cross-modal downstream tasks. However, most existing methods can only learn from aligned image-caption data and rely heavily on expensive regional features, which greatly limits their scalability and performance. In this paper, we propose an end-to-end unified-modal pre-training framework, namely UNIMO-2, for joint learning on both aligned image-caption data and unaligned image-only and text-only corpus. We build a unified Transformer model to jointly learn visual representations, textual representations and semantic alignment between images and texts. In particular, we propose to conduct grounded learning on both images and texts via a sharing grounded space, which helps bridge unaligned images and texts, and align the visual and textual semantic spaces on different types of corpora. The experiments show that our grounded learning method can improve textual and visual semantic alignment for improving performance on various cross-modal tasks. Moreover, benefiting from effective joint modeling of different types of corpora, our model also achieves impressive performance on single-modal visual and textual tasks. Our code and models are public at the UNIMO project page https://unimo-ptm.github.io/.

High-quality phrase representations are essential to finding topics and related terms in documents (a.k.a. topic mining). Existing phrase representation learning methods either simply combine unigram representations in a context-free manner or rely on extensive annotations to learn context-aware knowledge. In this paper, we propose UCTopic, a novel unsupervised contrastive learning framework for context-aware phrase representations and topic mining. UCTopic is pretrained in a large scale to distinguish if the contexts of two phrase mentions have the same semantics. The key to the pretraining is positive pair construction from our phrase-oriented assumptions. However, we find traditional in-batch negatives cause performance decay when finetuning on a dataset with small topic numbers. Hence, we propose cluster-assisted contrastive learning (CCL) which largely reduces noisy negatives by selecting negatives from clusters and further improves phrase representations for topics accordingly. UCTopic outperforms the state-of-the-art phrase representation model by 38.2% NMI in average on four entity clustering tasks. Comprehensive evaluation on topic mining shows that UCTopic can extract coherent and diverse topical phrases.

Grammatical Error Correction (GEC) aims to automatically detect and correct grammatical errors. In this aspect, dominant models are trained by one-iteration learning while performing multiple iterations of corrections during inference. Previous studies mainly focus on the data augmentation approach to combat the exposure bias, which suffers from two drawbacks. First, they simply mix additionally-constructed training instances and original ones to train models, which fails to help models be explicitly aware of the procedure of gradual corrections. Second, they ignore the interdependence between different types of corrections. In this paper, we propose a Type-Driven Multi-Turn Corrections approach for GEC. Using this approach, from each training instance, we additionally construct multiple training instances, each of which involves the correction of a specific type of errors. Then, we use these additionally-constructed training instances and the original one to train the model in turn. Experimental results and in-depth analysis show that our approach significantly benefits the model training. Particularly, our enhanced model achieves state-of-the-art single-model performance on English GEC benchmarks. We release our code at Github.

The retriever-reader pipeline has shown promising performance in open-domain QA but suffers from a very slow inference speed. Recently proposed question retrieval models tackle this problem by indexing question-answer pairs and searching for similar questions. These models have shown a significant increase in inference speed, but at the cost of lower QA performance compared to the retriever-reader models. This paper proposes a two-step question retrieval model, SQuID (Sequential Question-Indexed Dense retrieval) and distant supervision for training. SQuID uses two bi-encoders for question retrieval. The first-step retriever selects top-k similar questions, and the second-step retriever finds the most similar question from the top-k questions. We evaluate the performance and the computational efficiency of SQuID. The results show that SQuID significantly increases the performance of existing question retrieval models with a negligible loss on inference speed.

Recall and ranking are two critical steps in personalized news recommendation. Most existing news recommender systems conduct personalized news recall and ranking separately with different models. However, maintaining multiple models leads to high computational cost and poses great challenges to meeting the online latency requirement of news recommender systems. In order to handle this problem, in this paper we propose UniRec, a unified method for recall and ranking in news recommendation. In our method, we first infer user embedding for ranking from the historical news click behaviors of a user using a user encoder model. Then we derive the user embedding for recall from the obtained user embedding for ranking by using it as the attention query to select a set of basis user embeddings which encode different general user interests and synthesize them into a user embedding for recall. The extensive experiments on benchmark dataset demonstrate that our method can improve both efficiency and effectiveness for recall and ranking in news recommendation.

Modern Irish is a minority language lacking sufficient computational resources for the task of accurate automatic syntactic parsing of user-generated content such as tweets. Although language technology for the Irish language has been developing in recent years, these tools tend to perform poorly on user-generated content. As with other languages, the linguistic style observed in Irish tweets differs, in terms of orthography, lexicon, and syntax, from that of standard texts more commonly used for the development of language models and parsers. We release the first Universal Dependencies treebank of Irish tweets, facilitating natural language processing of user-generated content in Irish. In this paper, we explore the differences between Irish tweets and standard Irish text, and the challenges associated with dependency parsing of Irish tweets. We describe our bootstrapping method of treebank development and report on preliminary parsing experiments.

Models pre-trained with a language modeling objective possess ample world knowledge and language skills, but are known to struggle in tasks that require reasoning. In this work, we propose to leverage semi-structured tables, and automatically generate at scale question-paragraph pairs, where answering the question requires reasoning over multiple facts in the paragraph. We add a pre-training step over this synthetic data, which includes examples that require 16 different reasoning skills such as number comparison, conjunction, and fact composition. To improve data efficiency, we sample examples from reasoning skills where the model currently errs. We evaluate our approach on three reasoning-focused reading comprehension datasets, and show that our model, PReasM, substantially outperforms T5, a popular pre-trained encoder-decoder model. Moreover, sampling examples based on model errors leads to faster training and higher performance.

We propose a benchmark to measure whether a language model is truthful in generating answers to questions. The benchmark comprises 817 questions that span 38 categories, including health, law, finance and politics. We crafted questions that some humans would answer falsely due to a false belief or misconception. To perform well, models must avoid generating false answers learned from imitating human texts. We tested GPT-3, GPT-Neo/J, GPT-2 and a T5-based model. The best model was truthful on 58% of questions, while human performance was 94%. Models generated many false answers that mimic popular misconceptions and have the potential to deceive humans. The largest models were generally the least truthful. This contrasts with other NLP tasks, where performance improves with model size. However, this result is expected if false answers are learned from the training distribution. We suggest that scaling up models alone is less promising for improving truthfulness than fine-tuning using training objectives other than imitation of text from the web.

Triangular machine translation is a special case of low-resource machine translation where the language pair of interest has limited parallel data, but both languages have abundant parallel data with a pivot language. Naturally, the key to triangular machine translation is the successful exploitation of such auxiliary data. In this work, we propose a transfer-learning-based approach that utilizes all types of auxiliary data. As we train auxiliary source-pivot and pivot-target translation models, we initialize some parameters of the pivot side with a pre-trained language model and freeze them to encourage both translation models to work in the same pivot language space, so that they can be smoothly transferred to the source-target translation model. Experiments show that our approach can outperform previous ones.

Recent Quality Estimation (QE) models based on multilingual pre-trained representations have achieved very competitive results in predicting the overall quality of translated sentences. However, detecting specifically which translated words are incorrect is a more challenging task, especially when dealing with limited amounts of training data. We hypothesize that, not unlike humans, successful QE models rely on translation errors to predict overall sentence quality. By exploring a set of feature attribution methods that assign relevance scores to the inputs to explain model predictions, we study the behaviour of state-of-the-art sentence-level QE models and show that explanations (i.e. rationales) extracted from these models can indeed be used to detect translation errors. We therefore (i) introduce a novel semi-supervised method for word-level QE; and (ii) propose to use the QE task as a new benchmark for evaluating the plausibility of feature attribution, i.e. how interpretable model explanations are to humans.

Translate-train is a general training approach to multilingual tasks. The key idea is to use the translator of the target language to generate training data to mitigate the gap between the source and target languages. However, its performance is often hampered by the artifacts in the translated texts (translationese). We discover that such artifacts have common patterns in different languages and can be modeled by deep learning, and subsequently propose an approach to conduct translate-train using Translationese Embracing the effect of Artifacts (TEA). TEA learns to mitigate such effect on the training data of a source language (whose original and translationese are both available), and applies the learned module to facilitate the inference on the target language. Extensive experiments on the multilingual QA dataset TyDiQA demonstrate that TEA outperforms strong baselines.

Transformer architecture has become the de-facto model for many machine learning tasks from natural language processing and computer vision. As such, improving its computational efficiency becomes paramount. One of the major computational inefficiency of Transformer based models is that they spend the identical amount of computation throughout all layers. Prior works have proposed to augment the Transformer model with the capability of skimming tokens to improve its computational efficiency. However, they suffer from not having effectual and end-to-end optimization of the discrete skimming predictor. To address the above limitations, we propose the Transkimmer architecture, which learns to identify hidden state tokens that are not required by each layer. The skimmed tokens are then forwarded directly to the final output, thus reducing the computation of the successive layers. The key idea in Transkimmer is to add a parameterized predictor before each layer that learns to make the skimming decision. We also propose to adopt reparameterization trick and add skim loss for the end-to-end training of Transkimmer. Transkimmer achieves 10.97x average speedup on GLUE benchmark compared with vanilla BERT-base baseline with less than 1% accuracy degradation.

Representation of linguistic phenomena in computational language models is typically assessed against the predictions of existing linguistic theories of these phenomena. Using the notion of polarity as a case study, we show that this is not always the most adequate set-up. We probe polarity via so-called ‘negative polarity items’ (in particular, English ‘any’) in two pre-trained Transformer-based models (BERT and GPT-2). We show that – at least for polarity – metrics derived from language models are more consistent with data from psycholinguistic experiments than linguistic theory predictions. Establishing this allows us to more adequately evaluate the performance of language models and also to use language models to discover new insights into natural language grammar beyond existing linguistic theories. This work contributes to establishing closer ties between psycholinguistic experiments and experiments with language models.