Sentence-aware Contrastive Learning for Open-Domain Passage Retrieval

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 81 - 91 May 22-27, 2022 c⃝2022 Association for Computational Linguistics JointCL: A Joint Contrastive Learning Framework for Zero-Shot Stance Detection Bin Liang1,2∗, Qinglin Zhu1∗, Xiang Li1∗, Min Yang3, Lin Gui4, Yulan He4,5, Ruifeng Xu1,6† 1School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen, China 2 Joint Lab of HITSZ and China Merchants Securities, Shenzhen, China 3 SIAT, Chinese Academy of Sciences, Shenzhen, China 4 Department of Computer Science, University of Warwick, U.K 5 The Alan Turing Institute, UK, 6 Peng Cheng Laboratory, Shenzhen, China {bin.liang, zhuqinglin, xiangli}@stu.hit.edu.cn min.yang@siat.ac.cn, {lin.gui, Yulan.He}@warwick.ac.uk xuruifeng@hit.edu.cn Abstract Zero-shot stance detection (ZSSD) aims to de- tect the stance for an unseen target during the inference stage. In this paper, we propose a joint contrastive learning (JointCL) framework, which consists of stance contrastive learning and target-aware prototypical graph contrastive learning. Specifically, a stance contrastive learning strategy is employed to better gener- alize stance features for unseen targets. Fur- ther, we build a prototypical graph for each instance to learn the target-based representa- tion, in which the prototypes are deployed as a bridge to share the graph structures between the known targets and the unseen ones. Then a novel target-aware prototypical graph con- trastive learning strategy is devised to general- ize the reasoning ability of target-based stance representations to the unseen targets. Extensive experiments on three benchmark datasets show that the proposed approach achieves state-of- the-art performance in the ZSSD task1. 1 Introduction Stance detection aims to automatically identify one’s opinionated standpoint/attitude (e.g. Pro, Con, or Neutral, etc.) expressed in text towards a specific proposition, topic, or target (Somasun- daran and Wiebe, 2010; Augenstein et al., 2016; Mohammad et al., 2016; Sobhani et al., 2017). For example, a text “Everyone is able to believe in whatever they want.” expresses a stance of “Pro” towards the target “Atheism”. Existing methods achieved promising perfor- mance in in-target stance detection when trained and tested on the datasets towards the same set of ∗ Equal contribution † Corresponding Author 1The source code of this work is released at https:// github.com/HITSZ-HLT/JointCL targets (Mohtarami et al., 2018; Graells-Garrido et al., 2020), and in cross-target stance detection that identifies the stance of a destination target us- ing models trained on a related source target in a one-to-one way (Xu et al., 2018; Zhang et al., 2020; Liang et al., 2021a). In practice, however, it is in- feasible to enumerate all possible (in-target) or re- lated (cross-target) targets beforehand for training stance detection models. Hence, zero-shot stance detection (ZSSD) (Allaway and McKeown, 2020), which aims to detect the stance for unseen targets during the inference stage is a promising scenario forward. To deal with ZSSD, intuitively, we can either reason the target-based stance features from the learned stance information based on the context (i.e., from the context-aware perspective), or iden- tify stance information that is potentially relevant with unseen targets from the learned target-related stance expressions …

Content truncated to 500 words