融合知识图谱和语义匹配的医疗问答系统

Medical question answering system integrating knowledge graph and semantic matching

问答系统是自然语言处理领域中的一项重要任务,常应用于医疗服务。传统的问答系统通过知识图谱的实体关系匹配返回相应的尾实体作为答案,然而,倘若实体或关系无法识别,又或者在知识图谱中并不存在相应的实体关系,问答将无法继续进行。为了解决这一问题,建立一种融合知识图谱和语义匹配模型的中文医疗问答混合系统。当所提问题无法在知识图谱中进行实体关系匹配时,该模型能继续从问答对数据集中找到最相似的问题,并返回相应结果作为答案。在语义匹配模型方面,结合中文医疗相似问题对,在Sentence-BERT模型上进行微调训练,并引入双曲空间中的距离度量函数对句子对进行相似度度量。结果表明:在整体性能方面,所提模型相较于BERT这类大语言模型精度能提升7.16%;在度量能力方面,双曲度量相较于通用欧氏空间度量,如余弦度量,最高能有2.28%的精度提升和1.58%的F_1值提升。

Question answering system is an important task in the field of natural language processing,which is often used in medical service.The traditional question answering system can return the corresponding tail entity as the answer by means of the entity and relationship matching of the knowledge graph.However,if the entity or relationship is not recognized or there is no corresponding entity relationship in the knowledge graph,the question answering can not be continued.In order to solve this problem,a hybrid framework of Chinese medical question answering is proposed,which combines knowledge graph and semantic matching model.When the questions raised cannot be matched by the entity relationship in knowledge graph,the model can continue to find the most similar questions from the question answering on datasets and return corresponding results as answers.In terms of semantic matching models,combining Chinese medical similarity problems,fine-tuning training is conducted on the Sentence BERT model,and distance measurement functions in hyperbolic space are introduced to measure the sentence similarity.The results show that in terms of overall performance,the proposed model can improve accuracy by 7.16%compared to large language models like BERT.In terms of measurement ability,in combintion with the general Euclidean space metrics such as cosine metrics,hyperbolic metrics can achieve a maximum accuracy improvement of 2.28%and an F1 value improvement of 1.58%.