CoPrompt: A Contrastive-prompt Tuning Method for Multiparty Dialogue Character Relationship Extraction

Information extraction is a fundamental task in natural language processing, and relation extraction is one of its primary subfields. However, most previous research focuses on extracting relations from standard texts, such as news articles and Wikipedia entries, while overlooking the challenges present in dialogues - sparse details, cross-sentence links, and complex character relationships. To address these challenges, we propose a contrastive-prompt tuning method (CoPrompt) to better capture the relationships between characters in dialogues. Our method improves the performance of the relationship extraction model by constructing positive and negative sample pairs to obtain better embedding representations of relationship features through prompt learning. To evaluate our method, we separately built manual and continuous templates and conducted experiments on the DialogRE and CRECIL datasets. Our method consistently outperformed other methods, notably achieving state-of-the-art results on Chinese datasets(DialogRE_cn and CRECIL), underscoring its robust efficiency in Chinese language relation extraction tasks. We also demonstrate the effectiveness of our method in low-resource scenarios. Our code is available at https://github.com/LIyu810/CoPrompt_main.