基于图注意力和表指针网络的中文事件抽取方法  

Chinese Event Extraction Method Based on Graph Attention and Table Pointer Network

作　　者：刘炜[1,2] 马亚威 彭艳 李卫民[1] LIU Wei;MA Yawei;PENG Yan;LI Weimin(School of Computer Engineering and Science,Shanghai University,Shanghai 200444;Shanghai Artificial Intelligence Laboratory,Shanghai 200232;Institute of Artificial Intelligence,Shanghai University,Shanghai 200444)

机构地区：[1]上海大学计算机工程与科学学院,上海200444 [2]上海人工智能实验室,上海200232 [3]上海大学人工智能研究院,上海200444

出　　处：《模式识别与人工智能》2023年第5期459-470,共12页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金重大项目(No.61991410);浦江国家实验室项目(No.P22KN00391)资助。

摘　　要：现有的中文事件抽取方法存在触发词和论元依赖建模不足的问题,削弱事件内的信息交互,导致论元抽取性能低下,特别是论元角色存在重叠的情况下.对此,文中提出基于图注意力和表指针网络的中文事件抽取方法(Chinese Event Extraction Method Based on Graph Attention and Table Pointer Network,ATCEE).首先,融合预训练字符向量和词性标注向量作为特征输入,并利用双向长短期记忆网络,得到事件文本的强化语义特征.再将字符级建模的依存句法图引入图注意力网络,捕获文本中各组成成分的长距离依赖关系.然后,使用表填充的方法进行特征融合,进一步增强触发词和其对应的所有论元之间的依赖性.最后,将学习得到的表特征输入全连接层和表指针网络层,进行触发词和论元的联合抽取,使用表指针网络对论元边界进行解码,更好地识别长论元实体.实验表明:ATCEE在ACE2005和DuEE1.0这两个中文基准数据集上都有明显的性能提升,并且字符级依存特征和表填充策略在一定程度上可以解决论元角色重叠问题.ATCEE源代码地址如下:https://github.com/event6/ATCEE.The existing Chinese event extraction methods suffer from inadequate modeling of dependencies between an event trigger word and all its corresponding arguments,which results in weakened information interaction within an event and poor performance in argument extraction,especially when there is argument role overlap.To address this issue,a Chinese event extraction method based on graph attention and table pointer network(ATCEE)is proposed in this paper.Firstly,pre-trained character vectors and part-of-speech tagging vectors are fused as feature inputs.Then,the enhanced feature of the event text is obtained by a bidirectional long short-term memory network.Next,a character-level dependency syntax graph is constructed and introduced into multi-layer graph attention network to capture long-range dependencies among constituents of the event text.Subsequently,dependencies between an event trigger word and all its corresponding arguments are further enhanced via a table filling strategy.Finally,the learned table feature is input into a fully connected layer and table pointer network layer for joint extraction of trigger words and arguments.Consequently,long argument entities can be identified better by decoding argument boundaries with a table pointer network.Experimental results indicate that ATCEE method significantly outperforms previous event extraction methods on Chinese benchmark datasets,ACE2005 and DuEE1.0.In addition,the overlap problem of the event argument role is solved by introducing character-level dependency feature and table filling strategy to some extent.The source code of ATCEE can be found at the following website:https://github.com/event6/ATCEE.

关 键 词：中文事件抽取 论元角色重叠 图注意力网络 表填充 表指针网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]