Knowledge-Enhanced Prompt-Tuning for Stance Detection

Investigating public attitudes on social media is important in opinion mining systems. Stance detection aims to analyze the attitude of an opinionated text (e.g., favor, neutral, or against) toward a given target. Existing methods mainly address this problem from perspective fine-tuning. Recently, prompt-tuning has achieved success natural language processing tasks. However, conducting for stance real-world remains challenge several reasons: (1) The form usually short and informal, which makes it difficult design label words verbalizer. (2) tweet may not explicitly give attitude. Instead, users use various hashtags background knowledge express stance-aware perspectives. In article, we first propose prompt-tuning-based framework that performs cloze question manner. Specifically, knowledge-enhanced (KEprompt) method designed, consists automatic verbalizer (AutoV) injection (BKI). AutoV, introduce semantic graph build better mapping predicted word pretrained model labels. BKI, topic learning hashtag representation ConceptGraph as supplement At last, present challenging dataset detection, where all categories are expressed implicit Extensive experiments large demonstrate superiority KEprompt over state-of-the-art methods.