面向领域知识图谱的问答方法研究

Research on question answering method oriented to domain knowledge graph

针对现有多数领域知识图谱问答方法存在实体间关系的语义信息利用不足,多跳问题处理能力弱的问题,难以满足用户对领域知识深度查询的需求,该文引入预训练模型,构建了一种面向领域知识图谱的问答方法。首先采用意图分类模型判断用户问题查询类型,约束知识查询类别。其次通过实体识别模型识别出问题中的实体提及,结合实体链接词典定位主题实体,进而召回相应查询路径。最后通过语义匹配模型计算问题与查询路径的语义相似程度,实现查询路径排序,选择最优查询路径得到答案。通过在地震灾害防治领域知识图谱上验证,该文构建的模型均优于同类对比模型,总体准确率达到86%,能够有效应对多跳问题,满足领域知识图谱问答的实际需求。

Aiming at the problem that most of the existing domain knowledge graph question answering methods have insufficient use of semantic information of relationships between entities,weak ability to handle multi hop problems,and are difficult to meet users'needs for in-depth query of domain knowledge,the pre training model was introduced and a question answering method oriented to domain knowledge graph was constructed in this paper.First,the intention classification model was used to judge the query type of user questions and constrain the knowledge graph category.Second,the entity reference in the question was identified through the entity recognition model.Then,the subject entity was located with the entity reference dictionary,and the corresponding query path was recalled.Finally,the semantic similarity between the question and the query path was calculated through the semantic matching model to achieve the query path sequencing,and then the optimal query path was selected to get the answer.Through the verification on the knowledge graph in the field of earthquake disaster prevention and control,the models built in this paper were superior to similar comparison models,and the overall accuracy of the system reached 86%,which could effectively deal with multi hop problems and meet the actual needs of the field knowledge graph question and answer.