文献类型：会议论文

英文题名：Role-Guided Contrastive Learning for Event Argument Extraction

作者：Yao, Chunyu[1];Guo, Yi[1,2,3];Chen, Xue[1];Duan, Zhenzhen[1];Fu, Jiaojiao[1]

机构：[1]East China Univ Sci & Technol, Shanghai, Peoples R China;[2]Natl Engn Lab Big Data Distribut & Exchange Techn, Business Intelligence & Visualizat Res Ctr, Shanghai, Peoples R China;[3]Shanghai Engn Res Ctr Big Data & Internet Audienc, Shanghai, Peoples R China

会议论文集：46th European Conference on Information Retrieval (ECIR)

会议日期：MAR 24-28, 2024

会议地点：Glasgow, SCOTLAND

语种：英文

外文关键词：Event argument extraction; Contrastive learning; Information extraction

摘要：yEvent argument extraction is a subtask of information extraction. Recent efforts have predominantly focused on mitigating the issue of error propagation associated with pipeline methods for extracting event arguments, such as machine reading comprehension and generative approaches. However, these aforementioned methods necessitate the careful design of various templates, and the choice of templates can significantly impact the model's performance. Therefore, we propose a novel approach to extract event arguments using contrastive learning. Our approach aims to maximize the semantic similarity between role name semantics and actual argument semantics while minimizing the similarity between role name semantics and the semantics of other non-argument words, thereby enabling more precise extraction of argument boundaries. We investigate the impact of different templates on event argument extraction, and experimental results demonstrate that template adjustments have limited effects on our model. To attain more precise argument boundaries, we also introduce entity type boundary embeddings, which substantially enhance the effectiveness of event argument extraction.