Incorporate opinion-towards for stance detection

Stance detection can help gain different perspectives into important events, e.g., whether people are in favor of or against certain claim. Most previous work use sentiment information to assist in stance detection. However, they do not consider the critical opinion-towards information, i.e. whether the opinions are aimed at target or other objects. In this work, we incorporate opinion-towards information into a multi-task learning model to facilitate our proposed model for better understanding the sentiment information, which effectively improves the performance of stance detection. In particular, we have constructed a novel label relation matrix which constrains two auxiliary tasks in multi-task learning: (1) sentiment classification, and (2) opinion-towards classification. Our extensive experimental results on three publicly available benchmark datasets demonstrate the effectiveness of the proposed model. In addition, we show the importance of opinion-towards information for stance detection through ablation study and visualization analysis. (C)2022 Elsevier B.V. All rights reserved.

Automated summary

Stance detection is important for gaining different perspectives on important events, but previous methods often overlook the impact of opinion-towards information. This paper proposes a multi-task learning model that incorporates opinion-towards information to enhance sentiment understanding and improve stance detection performance. Experimental results demonstrate the effectiveness of the model, and ablation study and visualization analysis highlight the importance of opinion-towards information in stance detection.