基于混合神经网络的实体与关系联合抽取模型  

Joint entity and relationship extraction method based on mixed network

作　　者：蒋伟强 李凤英[1] 董荣胜[1] JIANG Weiqiang;LI Fengying;DONG Rongsheng(Guangxi Key Laboratory of Trusted Software,Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学广西可信软件重点实验室,广西桂林541004

出　　处：《桂林电子科技大学学报》2022年第1期60-65,共6页Journal of Guilin University of Electronic Technology

基　　金：国家自然科学基金(62062029,61762024);广西自然科学基金(2017GXNSFDA198050)。

摘　　要：针对实体关系联合抽取任务中的重叠实体关系识别问题,提出了一种混合神经网络模型(MNN-RE)。MNN-RE使用膨胀卷积神经网络结合门控线性单元作为输入共享编码层,增强了实体与关系之间关联度。采用自注意力机制结合指针网络的标注策略提取主实体。基于贪心策略,对每个预定义的关系都进行关系存在的预测并抽取目标实体,以此得到完整三元组。实验结果表明,该模型在中文的实体与关系联合抽取任务的表现比经典的联合抽取模型Noveltagging及Multi-head模型要好。A mixed neural network model(MNN-RE)is proposed for the recognition of overlapping entity relationship in the task of joint entity relationship extraction.MNN-RE uses dilated convolutional neural network combined with gated linear units as input to share coding layer,which enhances the correlation between entities and relationships.The self-attention mechanism and labeling strategy of pointer network is used to extract the primary entity.Based on greedy strategy,the existence of each predefined relationship is predicted and the target entity is extracted to get a complete triple.Experimental results show that the performance of this method in Chinese entity and relationship joint extraction task is better than the classical joint extraction model Noveltagging and Multi-head model.

关 键 词：实体关系抽取 联合抽取 门控线性单元 注意力机制 残差连接 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]