EA-MTIL: A Study of Stance Detection with Integrated Multitask Learning
Guirong Chen, Lian Duan, Aiwang Chen, Yangcheng Mu · 2024
Previous research on stance detection usually focuses on incorporating sentiment analysis to help stance detection, whereas texts with ironic maneuvers may reverse the entire semantics. In this paper, we propose the EA-MTIL model for Arabic with the additional introduction of sarcasm analysis for detecting stances (favor, against or neutral) on three selected topics (COVID-19 Vaccine, Digital Transformation, and Female Empowerment). The MTIL-based model can federate other information in the textual data, such as its inclusion of sentiment tendencies and ironic maneuvers, to improve the performance of the stance detection system. The model uses an integrated BERT model to complete the coding work, learns the general features of the three tasks by combining the External Attention (EA) on feature extraction, and efficiently fuses this part of the feature map. Experimental results show that the framework improves the performance of the primary task (i.e., stance detection) by utilizing secondary tasks (sarcasm and sentiment analysis) instead of unimodal tasks and single-task variables.