Syntax-Enhanced Event Extraction Model Based on XLNET and GAT
[Objective]This study addresses the issues of long-distance dependency between trigger words in sequence modeling and the insufficient capture of the relationship between trigger words and argument entities.It enhances the effectiveness of event extraction tasks.[Method]We proposed a Syntax-enhanced Event-extraction Model based on XLNET and GAT(SEM-XG).Using a pre-trained language model,we represented text and enhanced information flow by incorporating dependency arcs from dependency parse trees.Words were treated as nodes in a graph,and a graph attention network was used to model graph information.This yielded word representations that integrate syntactic information,facilitating the joint extraction of event triggers and argument roles in sentences.We conducted empirical studies on the CNC and ACE2005 datasets.[Results]The results demonstrate that,on the CNC dataset,the SEM-XG achieved an Fl value of 94.4％on the trigger word classification task and an Fl value of 94.0％on the argument classification task.On the ACE2005 dataset,the SEM-XG achieved an Fl value of 76.7％on the trigger word classification task and an Fl value of 66.3％on the argument classification task.Therefore,the proposed method is effective for event extraction tasks.[Limitations]We did not explore the effectiveness of the joint event extraction model in tasks such as search engines and intelligent question-answering systems.[Conclusions]Combining syntax enhancement and graph attention network modeling can significantly improve the performance of joint event extraction.This study has important implications for scientific literature analysis and information retrieval.
NETL
NSTL
万方数据
