基于循环神经网络的作战文书实体关系抽取  

A Combat Documents Entity Relation Extraction Approach Based on RNN

作　　者：王学锋 杨若鹏 贾明亮 WANG Xuefeng;YANG Ruopeng;JIA Mingliang(College of Information Communications,National University of Defense Technology,Wuhan 430010,China;College of Army Logistics,Chongqing 401331,China)

机构地区：[1]国防科技大学信息通信学院,武汉430010 [2]陆军勤务学院勤务指挥系,重庆401331

出　　处：《智能安全》2022年第1期29-35,共7页

摘　　要：基于指挥信息系统的作战文书智能处理是未来指挥智能化的基础,采用自然语言处理的方法从非结构化作战文书中抽取出结构化的作战数据对于辅助指挥员决策有着重要意义.其中作战文书实体之间的语义关系是战场态势理解的基础,对于获取对抗双方中的作战编成、部署位置、目标状态、指挥关系具有重要价值.针对作战文书实体关系抽取中传统方法的人工构建特征不充分、军事领域中文分词不准确、输入与输出之间的相关性考虑不足等问题,本文提出了基于深度学习的关系抽取方法.结合双向长短时记忆(Bi-LSTM)神经网络对较长句子上下文的记忆能力、字向量对汉字语义的表示能力和注意力机制对输入与输出相关性的学习能力,构建了Character+Bi-LSTM+Attention实体关系抽取模型.为验证方法的有效性,在学员训练文书语料集上进行了实验,实验结果表明,该方法抽取效果较传统方法有进一步提高.The intelligent processing of combat documents based on command information system is the basis of future command intelligence,the use of natural language processing methods to extract structured combat data from unstructured combat documents is of great significance for assisting commanders.The semantic relationship between military named entities is the basis for understanding the battlefield situation,and it is important for acquiring the troops,positions,targets,and command of both sides.In view of the problems of insufficient artificial construction features,inaccurate Chinese word segmentation in the military field,and insufficient correlation between input and output in the current military named entity relationship extraction,the author proposes a relation extraction method based on deep learning.Combining Bi-directional Long Short-Term Memory(Bi-LSTM)neural network’s ability to remember long sentence context,the ability of character embedding to express Chinese characters and the ability of attention mechanism to learn the correlation between input and output,the Character+Bi-LSTM+Attention entity relationship extraction model was constructed.In order to verify the validity of the method,experiments were carried out on the military training corpus,and the experimental results show that the extraction effect of the method is further improved than the traditional method.

关 键 词：作战文书 实体关系抽取 深度学习 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]