【作者】 张磊；

【导师】 苏曙光；

【作者基本信息】 华中科技大学 ， 软件工程（专业学位）， 2024， 硕士

【摘要】 近年来,人们的健康问题日益突显,关于医学领域的问答系统备受关注;与此同时,传统问答系统依赖模板和关键词匹配,对于复杂和模糊问题表现不佳,难以满足多变的查询需求和非结构化数据的挑战。因此,急需新技术和策略来突破目前的局限。在数据预训练阶段,针对句子级处理的准确性不足的问题,提出一种结合BERT、LSTM和CRF的模型框架。该框架结合了三种技术的优点,充分利用BERT预训练模型的优势,通过BERT的强大语义表示能力来深入理解句子中的语义信息。在此基础上,使用BERT计算问句与属性之间的相似度,从而准确地进行属性选择。针对实体识别任务,提出通过结合数据增强策略和TF-IDF实体对齐技术,提升模型在该任务上的性能,并确保模型输出与知识图谱的高度一致性。针对实体匹配精度不足的问题,提出了采用相似度匹配优化的方法,通过计算输入语句的词组与实体特征词之间的字重叠率、词向量余弦相似度以及DP编辑距离,将这三个指标的平均值作为相似度评分,并选择相似度最高且超过设定阈值的实体特征词作为匹配结果。为了进一步提升意图识别的准确性,提出采用34B大模型进行意图识别的技术方法。通过对比了三种意图识别方案,即规则匹配、训练模型与提示工程,在准确性、多意图辨识能力、人力成本、推理效率及资源占用等多个维度上的差异,选定提示工程为首选策略。为了进一步精细化处理,结合模板匹配的方法,通过定义一系列与医学领域相关的模板,可以对识别出的意图进行进一步的校验和调整。在某市卫健委健康教育科普馆项目中,集成和应用了医疗智能问答功能,设计了一个基于RAG与大模型技术的医疗问答系统,利用Disease KG数据集与Neo4j构建知识图谱,结合BERT的命名实体识别和34b大模型的意图识别,通过精确的知识检索和问答生成,提升系统在医疗咨询中的性能,解决大模型在医疗领域应用的可靠性问题。更多还原

【Abstract】 In recent years,people’s health issues have become increasingly prominent,and question answering systems in the medical field have received much attention;At the same time,traditional question answering systems rely on templates and keyword matching,which perform poorly on complex and ambiguous questions,making it difficult to meet the changing query requirements and challenges of unstructured data.Therefore,there is an urgent need for new technologies and strategies to overcome current limitations.In the data pre training stage,a model framework combining BERT,LSTM,and CRF is proposed to address the issue of insufficient accuracy in sentence level processing.This framework combines the advantages of three technologies and fully utilizes the strengths of BERT pre trained models to deeply understand semantic information in sentences through BERT’s powerful semantic representation capabilities.On this basis,BERT is used to calculate the similarity between the question and the attribute,in order to accurately select the attribute.For entity recognition tasks,it is proposed to improve the performance of the model on this task by combining data augmentation strategies and TF-IDF entity alignment techniques,while ensuring high consistency between the model output and the knowledge graph.Aiming at the problem of insufficient accuracy in entity matching,a similarity matching optimization method is proposed.By calculating the word overlap rate,cosine similarity of word vectors,and DP editing distance between the input sentence phrase and entity feature words,the average of these three indicators is used as the similarity score,and the entity feature words with the highest similarity and exceeding the set threshold are selected as the matching results.In order to further improve the accuracy of intent recognition,a technical method of using 34B large model for intent recognition is proposed.By comparing three intent recognition schemes,namely rule matching,model training,and prompt engineering,and comparing their differences in accuracy,multi intent recognition ability,labor cost,inference efficiency,and resource utilization,prompt engineering was selected as the preferred strategy.Then,in order to further refine the processing,a template matching method was combined to define a series of templates related to the medical field,which can further verify and adjust the identified intentions.In the project of the Health Education Science Popularization Museum of a certain city’s Health Inspection Commission,the medical intelligent question answering function was integrated and applied.A medical question answering system based on RAG and large model technology was designed.A knowledge graph was constructed using the Diseasy KG dataset and Neo4j,combined with BERT’s named entity recognition and the intention recognition of the 34b large model.Through accurate knowledge retrieval and question answering generation,the performance of the system in medical consultation was improved,and the reliability problem of large model application in the medical field was solved.更多还原