基于句间信息的图注意力卷积网络的文档级关系抽取  

Document-level Relation Extraction of Graph Attention Convolutional Network Based on Inter-sentence Information

作　　者：段建勇[1,2] 杨潇 王昊 何丽[1,2] 李欣 DUAN Jianyong;YANG Xiao;WANG Hao;HE Li;and LI Xin(School of Information,North China University of Technology,Beijing 100144,China;CNONIX National Standard Application and Promotion Lab,Beijing 100144,China;Beijing Urban Governance Research Center,Beijing 100144,China)

机构地区：[1]北方工业大学信息学院,北京100144 [2]CNONIX国家标准应用与推广实验室,北京100144 [3]北京城市治理研究基地,北京100144

出　　处：《计算机科学》2023年第S01期181-186,共6页Computer Science

基　　金：国家自然科学基金(61972003);教育部人文社科基金项目(21YJA740052);北京哲社基地建设专项经费(110051360019XN142);北京市教育委员会科学研究计划项目(KM202210009002)。

摘　　要：为解决现有模型对文档的结构信息挖掘不足的问题,提出一种基于句间信息的图注意力卷积网络模型。该模型改进了一种文档级编码器,该编码器使用了一种新的注意力机制--句间注意力机制,使得句子的最终表示更加关注前一个句子和之前文档中的重要信息,更有利于挖掘文档的结构信息。实验结果表明,所提模型在DocRED数据集上的F 1评价指标达到56.3%,性能优于基线模型。在融入句间注意力机制时,由于模型需要对每一句话分别进行句间注意力操作,因此训练模型时需要消耗更多的内存和时间。基于句间信息的图注意力卷积网络模型可以有效地对文档中的相关信息进行聚合,并且增强对文档的结构信息的挖掘能力,从而使得模型在文档级关系抽取任务中效果得到提升。In order to solve the problem of insufficient structural information mining for documents in existing models,a graph attention convolution network model based on inter-sentence information is proposed.In this model,a document level encoder is improved,which uses a new attention mechanism,inter-sentence attention mechanism,so that the final representation of a sentence pays more attention to the important information in the previous sentence and the previous document,which is more conducive to mine the structural information of the document.Experiments show that the F 1 evaluation index of this model on DocRED data set reaches 56.3%,and its performance is better than that of the baseline model.When integrating the inter sentence attention mechanism,the model needs to perform inter-sentence attention operations for each sentence,so it needs to consume more memory and time when training the model.The graph attention convolution network model based on inter-sentence information can effectively aggregate the relevant information in the document,and enhance the mining ability of the document structure information,so that the model can improve the effect of the document level relationship extraction task.

关 键 词：文档级关系抽取 注意力机制 文档级编码器 图卷积网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]