StanceDigger: Achieving Efficient Text Stance Detection with Collaborative Attention and Consistent Learning

Text stance detection is the task of identifying the attitude of a text editor, e.g., supported, opposed, or neutral, which is widely used to grasp fine-grained public opinion. However, the emotions expressed by users on specific topics are highly correlated with the stance preferred by the users, making it difficult for models trained on certain topics to generalize to those from other domains. In this paper, we propose StanceDigger, a topic-insensitive stance detection method embedded with collaborative attention and consistent learning. Our model achieves insensitivity to topics with the help of the attention mechanism and is capable of mining various types of topics through consistent learning. Experimental results on several benchmark datasets demonstrate that StanceDigger outperforms state-of-the-art baseline models on the task of text stance detection, not only in analyzing the correlation between target topics and short texts but also in overall detection accuracy.