Target Really Matters: Target-aware Contrastive Learning and Consistency Regularization for Few-shot Stance Detection

Proceedings of the 29th International Conference on Computational Linguistics, pages 6944–6954 October 12–17, 2022. 6944 Target Really Matters: Target-aware Contrastive Learning and Consistency Regularization for Few-shot Stance Detection Rui Liu1,2, Zheng Lin1,2∗, Huishan Ji1,2, Jiangnan Li1,2, Peng Fu1 and Weiping Wang1 1Institute of Information Engineering, Chinese Academy of Sciences 2School of Cyber Security, University of Chinese Academy of Sciences {liurui1995,linzheng,jihuishan,lijiangnan}@iie.ac.cn {fupeng,wangweiping}@iie.ac.cn Abstract Stance detection aims to identify the attitude from an opinion towards a certain target. De- spite the significant progress on this task, it is extremely time-consuming and budget- unfriendly to collect sufficient high-quality la- beled data for every new target under fully- supervised learning, whereas unlabeled data can be collected easier. Therefore, this paper is devoted to few-shot stance detection and investigating how to achieve satisfactory re- sults in semi-supervised settings. As a target- oriented task, the core idea of semi-supervised few-shot stance detection is to make better use of target-relevant information from labeled and unlabeled data. Therefore, we develop a novel target-aware semi-supervised framework. Specifically, we propose a target-aware con- trastive learning objective to learn more distin- guishable representations for different targets. Such an objective can be easily applied with or without unlabeled data. Furthermore, to thor- oughly exploit the unlabeled data and facili- tate the model to learn target-relevant stance features in the opinion content, we explore a simple but effective target-aware consistency regularization combined with a self-training strategy. Experimental results demonstrate that our approach can achieve state-of-the-art per- formance on multiple benchmark datasets in the few-shot setting. 1 Introduction Stance detection is intended to identify the attitude of opinions towards certain targets, where labels can be favor, against, and neutral. For example, the opinion “True equality allows all to be born.” is against the target “Legalization of Abortion”. The settings of stance detection can be generally di- vided into in-target and cross-target ones. Specif- ically, in-target stance detection aims to train an exclusive classifier for prediction on the same set ∗Zheng Lin is the corresponding author. of targets (Mohammad et al., 2016b; Augenstein et al., 2016; Li et al., 2021). Obviously, in-target stance detection is the most ideal setting when there is sufficient labeled data. However, in practice, a severe challenge is the scarcity of annotations for new targets. Cross-target stance detection (Zhang et al., 2020a; Xu et al., 2018; Wei and Mao, 2019) is to train on the labeled data of one source target and test on the destination target, which is based on an assumption that there is a strong correla- tion between the two targets. No doubt the harsh demand from the assumption above limits the ex- tension of cross-target stance detection, in which it still requires a large amount of annotated data for the source target. In this paper, we focus on the in-target few-shot stance detection with unlabeled data and limited labeled data, which is to alleviate the demand for human supervision. Different from common classification tasks (e.g., sentiment classification), the identification of stance is heavily dependent on the specific …

Content truncated to 500 words