基于SimCSE的疾病知识图谱问答系统

Disease Knowledge Base Question Answering Based on SimCSE

系统以疾病为中心的知识图谱为基础,构建问答系统,以帮助用户能够在线实时得到医疗问题的解答。将传统意图识别的分类任务转化为语义相似度计算任务,使用基于对比学习的SimCSE (Simple Contrastive Learning of Sentence Embeddings)模型进行微调,通过对比损失,得到更有区分度的语义向量表示,进而更准确识别用户意图。系统将SimCSE应用到疾病知识图谱问答系统领域中,并通过实验结果分析SimCSE与传统分类模型的性能,验证了SimCSE更适合完成意图识别任务,最终该系统回答准确率达到96.1%。

Based on the knowledge base centered on diseases,the system builds a question answer system to help users get answers to medical questions online in real time.The traditional classification task of intention recognition is transformed into a semantic similarity computing task,and the Simple Contrastive Learning of Sentence Embeddings(SimCSE) model based on contrast learning is used to fine-tune this task.By comparing the loss,a more differentiated semantic vector representation can be obtained,thus identifying the user's intention more accurately.The system applied SimCSE to the field of disease knowledge base question answering,and analyzed the performance of SimCSE and traditional classification model through experimental results,and verified that SimCSE is more suitable to complete the task of intention recognition,and finally the answer accuracy of the system reached 96.1%.