Research On Intention Recognition In Medical Question Answering Based On Word Embedding And Domain Knowledge

With the continuous development of intelligent question-and-answer robot in the field of artificial intelligence,the market has gradually released voice dialogue products such as Tmall Wizard,Xiaomi intelligent speaker,and the underlying system logic and algorithm of such products have been accumulating and updating.Such products are currently not sensitive enough to medical terms and problems.Currently,the commonly used methods of intention recognition and classification include rule-based template or feature,etc.,which require a lot of manpower and have limited coverage and classification ability.Aiming at the above problems and combining with the characteristics of dialogue text in medical field,this paper proposes an intention recognition model(ETM-T)based on subject word embedding.In this model,the user's intention to seek medical treatment is regarded as a classification problem.Firstly,the ETM model is used to represent the topic semantics related to Chinese doctor-patient question and answer data mining by vectors.Then,the topic vectors are integrated with TF-IDF word weights,and the resulting sentence vectors are used as input of Transformer encoder model.Transformer encoder is used to extract semantic features of text and complete the classification of users' question and answer data.The experimental results show that this model has a good intention classification effect.However,the ETM-T model cannot extract semantic information related to medical knowledge such as complications related to the patient's condition and food avoidance by relying only on the Q&A text.In order to solve such problems,this paper also proposed the intention recognition model integrating domain knowledge(ETM-H-T).Based on the ETM model to extract the semantic representation of the topic,this model used the crawler technology to obtain the medical domain knowledge data and build the knowledge base,and obtained the entity word embedding representation of the knowledge through the Hole model.The lexical information,topic semantic information and knowledge entity information of the question and answer text are used as the input of Transformer encoder,and the semantic information of the three are integrated through aggregation operation to obtain the final semantic representation and complete the task of intention classification.The representation of knowledge entities enhances the semantic boundary information of question and answer texts,which enables ETM-H-T model to dig out multiple relationships of key medical knowledge.Through the experimental analysis,the intention recognition model proposed in this paper has a better performance.