基于XLNET和GAT的句法信息增强事件抽取模型

Syntax-Enhanced Event Extraction Model Based on XLNET and GAT

【目的】解决序列建模对触发词之间的长距离依赖关系和触发词与论元实体关系捕获不足的问题,提升事件抽取任务上的效果。【方法】提出一种基于预训练模型XLNET和图注意力网络GAT的句法信息增强事件抽取模型SEM-XG,通过预训练语言模型进行文本表示,引入依存句法树中依赖弧增强信息流,将单词看作图中的节点,使用图注意力网络进行图信息建模,得到融入句法信息的单词表示,从而联合抽取句子中的事件触发词和论元角色。在CNC数据集和ACE2005数据集上,开展实证研究。【结果】在CNC数据集上,SEM-XG在触发词分类任务上的F1值为94.4%,在论元分类任务上的F1值为94.0%。在ACE2005数据集上,SEM-XG在触发词分类任务上的F1值为76.7%,在论元分类任务上的F1值为66.3%。实验结果表明,本文模型能够有效提升事件抽取的效果。【局限】尚未探究联合事件抽取模型迁移到搜索引擎、智能问答等任务上的效果。【结论】通过句法信息增强以及图注意力网络建模,能够显著提升联合事件抽取的效果。本文对于触发词分类和论元分类,提升事件抽取在科技文献分析、信息检索等领域的应用效果具有重要参考意义。

[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 F1 value of 94.4%on the trigger word classification task and an F1 value of 94.0%on the argument classification task.On the ACE2005 dataset,the SEM-XG achieved an F1 value of 76.7%on the trigger word classification task and an F1 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.