多层次权重优化的远程监督关系抽取模型  

Distant Supervision Relation Extraction via Multi-level Weight Optimization

作　　者：刘正 刘永坚[1] 解庆 李琳[1] LIU Zheng;LIU Yongjian;XIE Qing;LI Lin(School of Computer Science and Artificial Intelligence,Wuhan University of Technology,Wuhan,Hubei 430070,China;Chongqing Research Institute of Wuhan University of Technology,Chongqing 401135,China)

机构地区：[1]武汉理工大学计算机与人工智能学院,湖北武汉430070 [2]武汉理工大学重庆研究院,重庆401135

出　　处：《中文信息学报》2024年第7期40-50,62,共12页Journal of Chinese Information Processing

基　　金：重庆市自然科学基金(cstc2021jcyj-msxmX1013);湖北省重点研发计划项目(2021BAA030)。

摘　　要：针对目前基于远程监督的关系抽取方法存在句袋权重分配不合理和关系抽取模型对句子关键特征提取不充分的问题,该文提出了一种多层次权重优化的远程监督关系抽取模型。在句袋层面,通过构建编解码网络获取句子的表征向量并对句袋进行重构,使得句袋划分更加均衡;在句子层面,采用依赖路径注意力机制,增加模型对关键词的权重,提高关键特征的提取能力。模型在公共数据集NYT上的平均准确率达到79%,与有竞争力的主流方法相比有大约3%的提升,表明模型能够通过不同层次的权重优化降低噪声数据对模型的影响,从而提高关系抽取任务的准确率。To deal with the unproper sentence bag weight distribution and insufficient extraction of key features of sentences in current relation extraction models,this paper proposes a multi-level weight optimization method for distant supervision relation extraction.At the sentence bag level,an encoding and decoding network is applied to obtain the representation vector of the sentence and then the sentence bags are reconstructed,which makes the division of sentence bags more balanced.At the sentence level,the shortest dependency path attention mechanism is adopted to increase the weight of keywords and the ability to extract key features.Experiments on the public data set NYT show the method reaches 79%accuracy in average,which is about 3%improvement compared with the competitive mainstream methods.

关 键 词：远程监督 关系抽取 注意力机制 语义相似度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]