Generative Data Augmentation with Contrastive Learning for Zero-Shot Stance Detection

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6985 - 6995 December 7-11, 2022 ©2022 Association for Computational Linguistics Generative Data Augmentation with Contrastive Learning for Zero-Shot Stance Detection Yang Li Northeast Forestry University, Harbin 150004, China yli@nefu.edu.cn JiaWei Yuan Northeast Forestry University, Harbin 150004, China jwyuan@nefu.edu.cn Abstract Stance detection aims to identify whether the author of an opinionated text is in favor of, against, or neutral towards a given target. Re- markable success has been achieved when suf- ficient labeled training data is available. How- ever, it is labor-intensive to annotate sufficient data and train the model for every new target. Therefore, zero-shot stance detection, aiming at identifying stances of unseen targets with seen targets, has gradually attracted attention. Among them, one of the important challenges is to reduce the domain transfer between seen and unseen targets. To tackle this problem, we propose a generative data augmentation ap- proach to generate training samples containing targets and stances for testing data, and map the real samples and generated synthetic samples into the same embedding space with contrastive learning, then perform the final classification based on the augmented data. We evaluate our proposed model on two benchmark datasets. Experimental results show that our approach achieves state of-the-art performance on most topics in the task of zero-shot stance detection. 1 Introduction In recent years, social media websites have become an important platform for people to express their opinions on different targets (or topics in some lit- erature) ranging from politics, government policies, movies, sports and social issues, etc (ALDayel and Magdy, 2021). More often, users tend to take a stance, in Favor, Against or Neutral towards a par- ticular target (Mohammad et al., 2016). The task of stance detection aims to automatically identify the stance represented by the users in numerous texts. It is of great significance to applications like argu- ment mining, fake news detection, public opinion analysis and has caught the attention of researchers. Remarkable success has been achieved when sufficient labeled training data is available in the task of stance detection (Sun et al., 2018; Siddiqua Figure 1: An Illustration of Target Specific Stance De- tection and Zero-Shot Stance Detection. In ZSSD, the annotation data of the target to be tested will not appear in the training step of the model. et al., 2019a; Du et al., 2020). Most of the existing research work relies on a large number of labeled data for a specific target. However, in reality, new targets are constantly produced, and it is unpracti- cal to label sufficient data for each target. Based on the idea of transfer learning, some researchers use cross-target stance detection to train one classifier adapting from the current target to “new” target (Sobhani et al., 2017). However, this adaptation depends on the correlation between targets, and still needs annotation data of “new” targets. To solve the problem of missing annotating data of “new” targets in the real world scenario, the task of Zero-Shot Stance Detection …

Content truncated to 500 words