Improving Multi-task Stance Detection with Multi-task Interaction Network

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2990 - 3000 December 7-11, 2022 ©2022 Association for Computational Linguistics Improving Multi-task Stance Detection with Multi-task Interaction Network Heyan Chai1, Siyu Tang1, Jinhao Cui1, Ye Ding2, Binxing Fang1,3 and Qing Liao1,3∗ 1 Harbin Institute of Technology, Shenzhen, China 2 Dongguan University of Technology, China 3 Peng Cheng Laboratory, Shenzhen, China {chaiheyan,tangsiyu999}@stu.hit.edu.cn,gcjh777@163.com dingye@dgut.edu.cn,fangbx@cae.cn,liaoqing@hit.edu.cn Abstract Stance detection aims to identify people’s standpoints expressed in the text towards a tar- get, which can provide powerful information for various downstream tasks. Recent stud- ies have proposed multi-task learning models that introduce sentiment information to boost stance detection. However, they neglect to ex- plore capturing the fine-grained task-specific interaction between stance detection and sen- timent tasks, thus degrading performance. To address this issue, this paper proposes a novel multi-task interaction network (MTIN) for im- proving the performance of stance detection and sentiment analysis tasks simultaneously. Specifically, we construct heterogeneous task- related graphs to automatically identify and adapt the roles that a word plays with respect to a specific task. Also, a multi-task interac- tion module is designed to capture the word- level interaction between tasks, so as to obtain richer task representations. Extensive exper- iments on two real-world datasets show that our proposed approach outperforms state-of- the-art methods in both stance detection and sentiment analysis tasks. 1 Introduction Stance detection is a task that identifies the stand- point or attitude (e.g. favor, against, or none, etc.) of user towards various targets (e.g. entity, event, opinion, people, etc.) (Küçük and Can, 2020; Liang et al., 2021). Stance detection is crucial for ma- chines to understand natural language and it has many practical application areas such as sarcasm detection (Agrawal et al., 2020; Tang et al., 2022), fake news detection (Liao et al., 2021a; Ma et al., 2022), relation extraction (Li et al., 2021b; Huang et al., 2021), etc. Some early studies attempt to extract the richer stance expressions by leveraging manually engineered features (Darwish et al., 2020; Kochkina et al., 2017; Zarrella and Marsh, 2016; ∗ Corresponding Author Stance Pregnant people have feelings, and the ability to make decisions about their health Tweet Favor Positive Sentiment Target Favor Negative Legalization of abortion Against Positive Favor Positive Legalization of abortion Wearing a face mask Wearing a face mask They have not the ability and shouldn’t make decisions that involve their health I have an immune system that works fine, masks harm our immune system I have next to no immune system right now so thanks to all wearing masks Example 1 Example 2 Figure 1: Examples of MTL stance detection. Mohammad et al., 2016; Sen et al., 2018) or em- ploying deep learning approaches (Wei et al., 2016; Augenstein et al., 2016; Zhou et al., 2019; Kawin- tiranon and Singh, 2021). However, limited anno- tated training tweets on social media platforms can not provide richer stance representations, leading to the degraded performance of models mentioned above. Motivated by recent advances in multi-task learn- …

Content truncated to 500 words