基于ENCODER_ATT机制的远程监督关系抽取  

Remote Supervision Relationship Extraction Based on Encoder and Attention Mechanism

作　　者：王健[1] 郑七凡 李超 石晶 WANG Jian;ZHENG Qifan;LI Chao;SHI Jing(College of Information and Computer Engineering,Northeast Forestry University,Harbin Heilongjiang 150040,China)

机构地区：[1]东北林业大学信息与计算机工程学院

出　　处：《广西师范大学学报（自然科学版）》2019年第4期53-60,共8页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金(31700643);中央高校基本科研业务费专项资金(DL11AB01)

摘　　要：在信息抽取中,关系抽取是一项准确识别自然语言中实体间关系的关键技术。针对关系抽取模型中容易丢失关键语义特征问题及远程监督的基本假设容易引入噪声数据的问题,本文提出一种基于远程监督的ENCODER_ATT关系抽取模型。基于循环神经网络构造的ENCODER模型在以词级别进行特征记忆提取,并在句子层面进行语义特征信息整合,保证不遗失关键语义特征的同时去除冗余特征。然后在句子层面引入了注意力机制来降低噪声数据对实验结果的影响。在真实的数据集上进行实验,并绘制准确率-召回率曲线,实验结果表明ENCODER_ATT模型对比同类型的关系抽取方法有明显的提升。In information extraction,relation extraction is a key technology to accurately identify the relationships between entities in natural language.Aiming at the problem that the key semantics in the relation extraction model are easy to lose and the basic assumptions of remote supervision are easy to introduce noise data,an ENCODER_ATT relationship extraction model based on remote supervision is proposed.Firstly,the ENCODER model based on the construction of the cyclic neural network extracts the feature memory at the word level and integrates the semantic feature information at the sentence level to ensure that the key features are removed without removing the redundant features.Secondly,attention mechanism is introduced at the sentence level to reduce the influence of noise data on the test results.Based on the actual experimental data,the experiment was carried out and the accuracy-recall rate curve was drawn to prove that the ENCODER_ATT model has a better improvement over the relationship extraction method of the same type.

关 键 词：关系抽取 远程监督 ENCODER 注意力机制 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]