C-STANCE: A Large Dataset for Chinese Zero-Shot Stance Detection

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 13369–13385 July 9-14, 2023 ©2023 Association for Computational Linguistics C-STANCE: A Large Dataset for Chinese Zero-Shot Stance Detection Chenye Zhao♣ Yingjie Li♡ Cornelia Caragea♣ ♣ University of Illinois at Chicago ♡ Westlake University {czhao43,cornelia}@uic.edu liyingjie@westlake.edu.cn Abstract Zero-shot stance detection (ZSSD) aims to de- termine whether the author of a text is in fa- vor of, against, or neutral toward a target that is unseen during training. Despite the grow- ing attention on ZSSD, most recent advances in this task are limited to English and do not pay much attention to other languages such as Chinese. To support ZSSD research, in this pa- per, we present C-STANCE that, to our knowl- edge, is the first Chinese dataset for zero-shot stance detection. We introduce two challeng- ing subtasks for ZSSD: target-based ZSSD and domain-based ZSSD. Our dataset includes both noun-phrase targets and claim targets, covering a wide range of domains. We provide a detailed description and analysis of our dataset. To es- tablish results on C-STANCE, we report perfor- mance scores using state-of-the-art deep learn- ing models. We publicly release our dataset and code to facilitate future research.1 1 Introduction Stance detection aims to automatically predict whether the author of a text is in favor of, against, or neutral toward a specific target (Mohammad et al., 2016b; Küçük and Can, 2020; ALDayel and Magdy, 2021), e.g., epidemic prevention, gasoline price, or equal rights. The stance can provide use- ful information for important events such as policy- making and presidential elections. Early works focus on two types of stance de- tection tasks: in-target stance detection, where classifiers are trained and tested on data from the same set of targets (Hasan and Ng, 2014; Moham- mad et al., 2016b; Graells-Garrido et al., 2020) and cross-target stance detection, where classifiers are trained on source targets that are related to desti- nation targets (Augenstein et al., 2016; Wei and Mao, 2019), but destination targets are unseen dur- ing training. However, it is impractical to include 1https://github.com/chenyez/C-STANCE all possible or related targets in the training set. More recently, zero-shot stance detection (ZSSD) has been identified as a promising direction (All- away and McKeown, 2020) to evaluate classifiers on a large number of unseen (and unrelated) targets. ZSSD is more similar to the situations in practice and has received a lot of attention (Liu et al., 2021; Luo et al., 2022; Liang et al., 2022b). Despite the growing interest in stance detection, the task has several limitations. First, most recent advances in stance detection are limited to English (Mohammad et al., 2016b; Allaway and McKeown, 2020; Conforti et al., 2020b; Li et al., 2021a; Glandt et al., 2021), and pay little attention to other lan- guages such as Chinese (Xu et al., 2016) although large amounts of online data with expressions of stance are available in other languages. Second, the current ZSSD task (Allaway and McKeown, 2020) aims to detect the …

Content truncated to 500 words