ACL 2022 论文列表

Representations of events described in text are important for various tasks. In this work, we present SWCC: a Simultaneous Weakly supervised Contrastive learning and Clustering framework for event representation learning. SWCC learns event representations by making better use of co-occurrence information of events. Specifically, we introduce a weakly supervised contrastive learning method that allows us to consider multiple positives and multiple negatives, and a prototype-based clustering method that avoids semantically related events being pulled apart. For model training, SWCC learns representations by simultaneously performing weakly supervised contrastive learning and prototype-based clustering. Experimental results show that SWCC outperforms other baselines on Hard Similarity and Transitive Sentence Similarity tasks. In addition, a thorough analysis of the prototype-based clustering method demonstrates that the learned prototype vectors are able to implicitly capture various relations between events.

Weighted decoding methods composed of the pretrained language model (LM) and the controller have achieved promising results for controllable text generation. However, these models often suffer from a control strength/fluency trade-off problem as higher control strength is more likely to generate incoherent and repetitive text. In this paper, we illustrate this trade-off is arisen by the controller imposing the target attribute on the LM at improper positions. And we propose a novel framework based on existing weighted decoding methods called CAT-PAW, which introduces a lightweight regulator to adjust bias signals from the controller at different decoding positions. Experiments on positive sentiment control, topic control, and language detoxification show the effectiveness of our CAT-PAW upon 4 SOTA models.

Data-to-text generation focuses on generating fluent natural language responses from structured meaning representations (MRs). Such representations are compositional and it is costly to collect responses for all possible combinations of atomic meaning schemata, thereby necessitating few-shot generalization to novel MRs. In this work, we systematically study the compositional generalization of the state-of-the-art T5 models in few-shot data-to-text tasks. We show that T5 models fail to generalize to unseen MRs, and we propose a template-based input representation that considerably improves the model’s generalization capability. To further improve the model’s performance, we propose an approach based on self-training using fine-tuned BLEURT for pseudo-response selection. On the commonly-used SGD and Weather benchmarks, the proposed self-training approach improves tree accuracy by 46%+ and reduces the slot error rates by 73%+ over the strong T5 baselines in few-shot settings.

Chinese Grammatical Error Detection(CGED) aims at detecting grammatical errors in Chinese texts. One of the main challenges for CGED is the lack of annotated data. To alleviate this problem, previous studies proposed various methods to automatically generate more training samples, which can be roughly categorized into rule-based methods and model-based methods. The rule-based methods construct erroneous sentences by directly introducing noises into original sentences. However, the introduced noises are usually context-independent, which are quite different from those made by humans. The model-based methods utilize generative models to imitate human errors. The generative model may bring too many changes to the original sentences and generate semantically ambiguous sentences, so it is difficult to detect grammatical errors in these generated sentences. In addition, generated sentences may be error-free and thus become noisy data. To handle these problems, we propose CNEG, a novel Conditional Non-Autoregressive Error Generation model for generating Chinese grammatical errors. Specifically, in order to generate a context-dependent error, we first mask a span in a correct text, then predict an erroneous span conditioned on both the masked text and the correct span. Furthermore, we filter out error-free spans by measuring their perplexities in the original sentences. Experimental results show that our proposed method achieves better performance than all compared data augmentation methods on the CGED-2018 and CGED-2020 benchmarks.

Entity linking (EL) is the task of linking entity mentions in a document to referent entities in a knowledge base (KB). Many previous studies focus on Wikipedia-derived KBs. There is little work on EL over Wikidata, even though it is the most extensive crowdsourced KB. The scale of Wikidata can open up many new real-world applications, but its massive number of entities also makes EL challenging. To effectively narrow down the search space, we propose a novel candidate retrieval paradigm based on entity profiling. Wikidata entities and their textual fields are first indexed into a text search engine (e.g., Elasticsearch). During inference, given a mention and its context, we use a sequence-to-sequence (seq2seq) model to generate the profile of the target entity, which consists of its title and description. We use the profile to query the indexed search engine to retrieve candidate entities. Our approach complements the traditional approach of using a Wikipedia anchor-text dictionary, enabling us to further design a highly effective hybrid method for candidate retrieval. Combined with a simple cross-attention reranker, our complete EL framework achieves state-of-the-art results on three Wikidata-based datasets and strong performance on TACKBP-2010.

In document classification for, e.g., legal and biomedical text, we often deal with hundreds of classes, including very infrequent ones, as well as temporal concept drift caused by the influence of real world events, e.g., policy changes, conflicts, or pandemics. Class imbalance and drift can sometimes be mitigated by resampling the training data to simulate (or compensate for) a known target distribution, but what if the target distribution is determined by unknown future events? Instead of simply resampling uniformly to hedge our bets, we focus on the underlying optimization algorithms used to train such document classifiers and evaluate several group-robust optimization algorithms, initially proposed to mitigate group-level disparities. Reframing group-robust algorithms as adaptation algorithms under concept drift, we find that Invariant Risk Minimization and Spectral Decoupling outperform sampling-based approaches to class imbalance and concept drift, and lead to much better performance on minority classes. The effect is more pronounced the larger the label set.

Relation linking (RL) is a vital module in knowledge-based question answering (KBQA) systems. It aims to link the relations expressed in natural language (NL) to the corresponding ones in knowledge graph (KG). Existing methods mainly rely on the textual similarities between NL and KG to build relation links. Due to the ambiguity of NL and the incompleteness of KG, many relations in NL are implicitly expressed, and may not link to a single relation in KG, which challenges the current methods. In this paper, we propose an implicit RL method called ImRL, which links relation phrases in NL to relation paths in KG. To find proper relation paths, we propose a novel path ranking model that aligns not only textual information in the word embedding space but also structural information in the KG embedding space between relation phrases in NL and relation paths in KG. Besides, we leverage a gated mechanism with attention to inject prior knowledge from external paraphrase dictionaries to address the relation phrases with vague meaning. Our experiments on two benchmark and a newly-created datasets show that ImRL significantly outperforms several state-of-the-art methods, especially for implicit RL.

There has been a growing interest in developing machine learning (ML) models for code summarization tasks, e.g., comment generation and method naming. Despite substantial increase in the effectiveness of ML models, the evaluation methodologies, i.e., the way people split datasets into training, validation, and test sets, were not well studied. Specifically, no prior work on code summarization considered the timestamps of code and comments during evaluation. This may lead to evaluations that are inconsistent with the intended use cases. In this paper, we introduce the time-segmented evaluation methodology, which is novel to the code summarization research community, and compare it with the mixed-project and cross-project methodologies that have been commonly used. Each methodology can be mapped to some use cases, and the time-segmented methodology should be adopted in the evaluation of ML models for code summarization. To assess the impact of methodologies, we collect a dataset of (code, comment) pairs with timestamps to train and evaluate several recent ML models for code summarization. Our experiments show that different methodologies lead to conflicting evaluation results. We invite the community to expand the set of methodologies used in evaluations.

The ability to integrate context, including perceptual and temporal cues, plays a pivotal role in grounding the meaning of a linguistic utterance. In order to measure to what extent current vision-and-language models master this ability, we devise a new multimodal challenge, Image Retrieval from Contextual Descriptions (ImageCoDe). In particular, models are tasked with retrieving the correct image from a set of 10 minimally contrastive candidates based on a contextual description. As such, each description contains only the details that help distinguish between images. Because of this, descriptions tend to be complex in terms of syntax and discourse and require drawing pragmatic inferences. Images are sourced from both static pictures and video frames. We benchmark several state-of-the-art models, including both cross-encoders such as ViLBERT and bi-encoders such as CLIP, on ImageCoDe.Our results reveal that these models dramatically lag behind human performance: the best variant achieves an accuracy of 20.9 on video frames and 59.4 on static pictures, compared with 90.8 in humans. Furthermore, we experiment with new model variants that are better equipped to incorporate visual and temporal context into their representations, which achieve modest gains. Our hope is that ImageCoDE will foster progress in grounded language understanding by encouraging models to focus on fine-grained visual differences.

This paper studies the (often implicit) human values behind natural language arguments, such as to have freedom of thought or to be broadminded. Values are commonly accepted answers to why some option is desirable in the ethical sense and are thus essential both in real-world argumentation and theoretical argumentation frameworks. However, their large variety has been a major obstacle to modeling them in argument mining. To overcome this obstacle, we contribute an operationalization of human values, namely a multi-level taxonomy with 54 values that is in line with psychological research. Moreover, we provide a dataset of 5270 arguments from four geographical cultures, manually annotated for human values. First experiments with the automatic classification of human values are promising, with F1-scores up to 0.81 and 0.25 on average.

Identifying changes in individuals’ behaviour and mood, as observed via content shared on online platforms, is increasingly gaining importance. Most research to-date on this topic focuses on either: (a) identifying individuals at risk or with a certain mental health condition given a batch of posts or (b) providing equivalent labels at the post level. A disadvantage of such work is the lack of a strong temporal component and the inability to make longitudinal assessments following an individual’s trajectory and allowing timely interventions. Here we define a new task, that of identifying moments of change in individuals on the basis of their shared content online. The changes we consider are sudden shifts in mood (switches) or gradual mood progression (escalations). We have created detailed guidelines for capturing moments of change and a corpus of 500 manually annotated user timelines (18.7K posts). We have developed a variety of baseline models drawing inspiration from related tasks and show that the best performance is obtained through context aware sequential modelling. We also introduce new metrics for capturing rare events in temporal windows.

Recent works of opinion expression identification (OEI) rely heavily on the quality and scale of the manually-constructed training corpus, which could be extremely difficult to satisfy. Crowdsourcing is one practical solution for this problem, aiming to create a large-scale but quality-unguaranteed corpus. In this work, we investigate Chinese OEI with extremely-noisy crowdsourcing annotations, constructing a dataset at a very low cost. Following Zhang el al. (2021), we train the annotator-adapter model by regarding all annotations as gold-standard in terms of crowd annotators, and test the model by using a synthetic expert, which is a mixture of all annotators. As this annotator-mixture for testing is never modeled explicitly in the training phase, we propose to generate synthetic training samples by a pertinent mixup strategy to make the training and testing highly consistent. The simulation experiments on our constructed dataset show that crowdsourcing is highly promising for OEI, and our proposed annotator-mixup can further enhance the crowdsourcing modeling.

Natural language inference (NLI) has been widely used as a task to train and evaluate models for language understanding. However, the ability of NLI models to perform inferences requiring understanding of figurative language such as idioms and metaphors remains understudied. We introduce the IMPLI (Idiomatic and Metaphoric Paired Language Inference) dataset, an English dataset consisting of paired sentences spanning idioms and metaphors. We develop novel methods to generate 24k semiautomatic pairs as well as manually creating 1.8k gold pairs. We use IMPLI to evaluate NLI models based on RoBERTa fine-tuned on the widely used MNLI dataset. We then show that while they can reliably detect entailment relationship between figurative phrases with their literal counterparts, they perform poorly on similarly structured examples where pairs are designed to be non-entailing. This suggests the limits of current NLI models with regard to understanding figurative language and this dataset serves as a benchmark for future improvements in this direction.

Knowledge of difficulty level of questions helps a teacher in several ways, such as estimating students’ potential quickly by asking carefully selected questions and improving quality of examination by modifying trivial and hard questions. Can we extract such benefits of instance difficulty in Natural Language Processing? To this end, we conduct Instance-Level Difficulty Analysis of Evaluation data (ILDAE) in a large-scale setup of 23 datasets and demonstrate its five novel applications: 1) conducting efficient-yet-accurate evaluations with fewer instances saving computational cost and time, 2) improving quality of existing evaluation datasets by repairing erroneous and trivial instances, 3) selecting the best model based on application requirements, 4) analyzing dataset characteristics for guiding future data creation, 5) estimating Out-of-Domain performance reliably. Comprehensive experiments for these applications lead to several interesting results, such as evaluation using just 5% instances (selected via ILDAE) achieves as high as 0.93 Kendall correlation with evaluation using complete dataset and computing weighted accuracy using difficulty scores leads to 5.2% higher correlation with Out-of-Domain performance. We release the difficulty scores and hope our work will encourage research in this important yet understudied field of leveraging instance difficulty in evaluations.

Traditionally, a debate usually requires a manual preparation process, including reading plenty of articles, selecting the claims, identifying the stances of the claims, seeking the evidence for the claims, etc. As the AI debate attracts more attention these years, it is worth exploring the methods to automate the tedious process involved in the debating system. In this work, we introduce a comprehensive and large dataset named IAM, which can be applied to a series of argument mining tasks, including claim extraction, stance classification, evidence extraction, etc. Our dataset is collected from over 1k articles related to 123 topics. Near 70k sentences in the dataset are fully annotated based on their argument properties (e.g., claims, stances, evidence, etc.). We further propose two new integrated argument mining tasks associated with the debate preparation process: (1) claim extraction with stance classification (CESC) and (2) claim-evidence pair extraction (CEPE). We adopt a pipeline approach and an end-to-end method for each integrated task separately. Promising experimental results are reported to show the values and challenges of our proposed tasks, and motivate future research on argument mining.

To alleviate the data scarcity problem in training question answering systems, recent works propose additional intermediate pre-training for dense passage retrieval (DPR). However, there still remains a large discrepancy between the provided upstream signals and the downstream question-passage relevance, which leads to less improvement. To bridge this gap, we propose the HyperLink-induced Pre-training (HLP), a method to pre-train the dense retriever with the text relevance induced by hyperlink-based topology within Web documents. We demonstrate that the hyperlink-based structures of dual-link and co-mention can provide effective relevance signals for large-scale pre-training that better facilitate downstream passage retrieval. We investigate the effectiveness of our approach across a wide range of open-domain QA datasets under zero-shot, few-shot, multi-hop, and out-of-domain scenarios. The experiments show our HLP outperforms the BM25 by up to 7 points as well as other pre-training methods by more than 10 points in terms of top-20 retrieval accuracy under the zero-shot scenario. Furthermore, HLP significantly outperforms other pre-training methods under the other scenarios.

Knowledge-based visual question answering (QA) aims to answer a question which requires visually-grounded external knowledge beyond image content itself. Answering complex questions that require multi-hop reasoning under weak supervision is considered as a challenging problem since i) no supervision is given to the reasoning process and ii) high-order semantics of multi-hop knowledge facts need to be captured. In this paper, we introduce a concept of hypergraph to encode high-level semantics of a question and a knowledge base, and to learn high-order associations between them. The proposed model, Hypergraph Transformer, constructs a question hypergraph and a query-aware knowledge hypergraph, and infers an answer by encoding inter-associations between two hypergraphs and intra-associations in both hypergraph itself. Extensive experiments on two knowledge-based visual QA and two knowledge-based textual QA demonstrate the effectiveness of our method, especially for multi-hop reasoning problem. Our source code is available at https://github.com/yujungheo/kbvqa-public.

We consider the problem of generating natural language given a communicative goal and a world description. We ask the question: is it possible to combine complementary meaning representations to scale a goal-directed NLG system without losing expressiveness? In particular, we consider using two meaning representations, one based on logical semantics and the other based on distributional semantics. We build upon an existing goal-directed generation system, S-STRUCT, which models sentence generation as planning in a Markov decision process. We develop a hybrid approach, which uses distributional semantics to quickly and imprecisely add the main elements of the sentence and then uses first-order logic based semantics to more slowly add the precise details. We find that our hybrid method allows S-STRUCT’s generation to scale significantly better in early phases of generation and that the hybrid can often generate sentences with the same quality as S-STRUCT in substantially less time. However, we also observe and give insight into cases where the imprecision in distributional semantics leads to generation that is not as good as using pure logical semantics.

A pressing challenge in current dialogue systems is to successfully converse with users on topics with information distributed across different modalities. Previous work in multiturn dialogue systems has primarily focused on either text or table information. In more realistic scenarios, having a joint understanding of both is critical as knowledge is typically distributed over both unstructured and structured forms. We present a new dialogue dataset, HybriDialogue, which consists of crowdsourced natural conversations grounded on both Wikipedia text and tables. The conversations are created through the decomposition of complex multihop questions into simple, realistic multiturn dialogue interactions. We propose retrieval, system state tracking, and dialogue response generation tasks for our dataset and conduct baseline experiments for each. Our results show that there is still ample opportunity for improvement, demonstrating the importance of building stronger dialogue systems that can reason over the complex setting of informationseeking dialogue grounded on tables and text.