融合GAT与QA提取范式的事件抽取方法

Method of extraction of GAT and QA extracting paradigm

针对传统抽取方法中事件标签语义与事件关系信息利用率低的问题,提出一种融合GAT(图注意力网络)与QA(问题-回答)提取范式的事件抽取方法。将事件类型与论元角色作为查询语句,根据不同事件类型之间的关联构建事件关系图,通过图注意力网络优化事件类型表示,使用注意力机制获取文本与标签的丰富语义,获取事件触发词与事件论元角色。该方法在DuEE数据集上的实验结果表明,触发词识别和论元角色识别的F1值比传统的BERT_QA_Arg模型分别提升4.49%和10.02%,验证了其有效性。

To address the low utilization of semantic information of event labels and the event relation in traditional extraction methods,an event extraction method that combined graph attention network(GAT)with question-answer(QA)extraction para-digm was proposed.Event types and argument roles were treated as query statements and an event relation graph based on the associations between different event types was constructed.The event type representation was optimized using graph attention networks,and an attention mechanism was employed to capture rich semantic information between the text and labels,thereby extracting event trigger words and event argument roles.Experimental results on the DuEE dataset demonstrate the effectiveness of the proposed method,and F1 values for trigger word recognition and argument role recognition are improved by 4.49%and 10.02%,respectively,compared to that of the traditional BERT_QA_Arg model.