Stance classification model with knowledge-aware multi-feature attention network

Chao Meng, Binxing Fang, Hongli Zhang, Yuchen Yang, Gongzhu Yin, Kun Lu

Abstract: Stance classification aims to identify the stance conveyed in tweets toward a specific target. Recent works have been devoted to leveraging target word embedding to incorporate target information into the stance classification model. However, it is difficult to capture implicit target information solely through target word embedding. In addition, stance knowledge is often ignored in previous work. To address these issues, this paper proposes a novel stance classification model with knowledge-aware multi-feature attention network (SC-KMAN). Firstly, we introduce richer target information into SC-KMAN through the target information extractor T-BERT designed in this paper. Meanwhile, we introduce a sentiment feature extractor S-BERT by transfer learning. Then, we propose a knowledge-based multi-feature attention network (KMAN) to introduce stance knowledge into the stance classification model. KMAN utilizes the stance, sentiment, and target features of texts as inputs. Subsequently, it integrates stance knowledge to accurately infer the stance of texts. The experimental results on two Twitter datasets and one Chinese stance classification dataset demonstrate that SC-KMAN achieves relatively considerable performance.

External IDs:dblp:journals/nca/MengFZYYL25