基于预训练模型的军事领域问答关系检测方法  

Method of question answering relation detection in military field based on pre-training model

作　　者：陈琳 段立[1] 张显峰[1] 罗兵[1] CHEN Lin;DUAN Li;ZHANG Xianfeng;LUO Bing(College of Electronic Engineering,Naval Univ.of Engineering,Wuhan 430033,China)

机构地区：[1]海军工程大学电子工程学院,武汉430033

出　　处：《海军工程大学学报》2024年第5期27-33,共7页Journal of Naval University of Engineering

基　　金：国家自然科学基金资助项目(62371079)。

摘　　要：为解决军事领域问答服务中用户问询语句的关系链接错误问题,提高知识库问答的准确性,提出了一种基于预训练语言模型的关系检测方法。首先,摒弃了用户问句中的实体名称信息,加入约束性本体信息,结合预训练语言模型嵌入注意力机制进行了关系检测模型研究;然后,将该关系检测方法结合军事语料应用于军事知识库问答任务中,进行了实验验证。结果表明:约束性本体信息的加入扩展了本体层级信息量与本体知识拓扑结构,对关系检测结果进行了约束,测试关系链接精准率提升了6.2%左右;预训练模型的嵌入为军事数据注入了更多背景知识,相比于未嵌入前,训练精准率提升了10%左右,说明结合知识库的信息特征,整体增强了关系检测在军事领域问答服务中的实际应用效果。In order to solve the problem of relationship link error of user query statements in military domain question answering service and improve the accuracy of knowledge base question answering,a relationship detection method based on pre-trained language model was proposed,in which the entity name information in user questions was abandoned and binding ontology information was added.On this basis,the relationship detection model combined with the pre-trained language model embedded attention mechanism was studied.The above relationship detection method was applied to the military knowledge base question answering task in combination with the military corpus.The experimental results show that,on one hand,the constrained ontology information is added to expand the ontology level information and the ontology knowledge topology,and the relationship detection results are constrained,resulting in an increase of about 6.2%of the accuracy of the test relationship link;on the other hand,more background knowledge is injected into military data through the pre-trained language model,which increases the training accuracy by about 10%compared with the unembedded pre-trained language model.It indicates that combined with the information characteristics of the know-ledge base,the practical application effect of relationship detection in military domain question answering service is enhanced.

关 键 词：关系检测 注意力机制 约束性本体 预训练语言模型 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]