结合Roberta和Bi-FLASH-SRU的中文事件因果关系抽取  

Combining Roberta and Bi-FLASH-SRU for Chinese Event Causality Extraction

作　　者：陈泉林 贾珺[1] 樊硕 CHEN Quan-Lin;JIA Jun;FAN Shuo(Institute of War,Academy of Military Sciences,Beijing 100091,China)

机构地区：[1]军事科学院战争研究院,北京100091

出　　处：《计算机系统应用》2024年第6期259-267,共9页Computer Systems & Applications

基　　金：国家自然科学基金(6203366);全军军事类研究生资助课题(JY2022C251)。

摘　　要：针对现有基于填表的事件关系抽取方法填表数目过多、表格特征获取不充分的问题,本文提出了结合Roberta和Bi-FLASH-SRU的中文事件因果关系抽取方法 TF-ChineseERE.该方法通过制定填表策略,利用文本中已标记关系,将其转化为带有标签的表格;借助Roberta预训练模型和本文提出的双向内置闪电注意力简单循环单元(bidirectional built-in flash attention simple recurrent unit,Bi-FLASH-SRU)获取主客体事件特征,再利用表格特征循环学习模块挖掘全局特征,最后进行表格解码获得事件因果关系三元组.实验采用金融领域两个公开数据集进行验证,结果表明,本文提出的方法 F_(1)值分别达到59.2%和62.5%,且Bi-FLASH-SRU模型训练速度更快,填表数目更少,证明了该方法的有效性.Aiming at the problems of too many forms and insufficient acquisition of form features in the existing formbased event relationship extraction methods,this study proposes TF-ChineseERE,a Chinese event causality extraction method that combines Roberta and Bi-FLASH-SRU.The method transforms the text into labeled forms by formulating a form-filling strategy that takes advantage of the labeled relationships in the text.The study proposes the Roberta pretraining model and the bidirectional built-in flash attention simple recurrent unit(Bi-FLASH-SRU) to obtain the subjectobject event features.It then uses the table feature recurrent learning module to mine the global features and finally performs table decoding to obtain event causality triples.The experiments are validated with two public datasets in the financial domain.The results show that the F_(1) values of the proposed method reach 59.2% and 62.5%,respectively,with a faster training speed of the Bi-FLASH-SRU model and less number of filled forms,which proves the effectiveness of the method.

关 键 词：事件因果关系抽取 Roberta Bi-FLASH-SRU 闪电注意力 填表式抽取 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]