Research On Human-robot Interactive Intelligent Question-answering Method With Knowledge Graph

With the development of robots and artificial intelligence technologies,intelligent Question-Answering(QA)Systems have gradually become one of the key technologies in human robot natural interaction,providing intelligent services for the users by the QA interaction.In smart medical,the Knowledge Graph(KG)based intelligent QA systems can generate concise and accurate answers with structured medical knowledge,and automatically reply to questions,helping users to acquire knowledge more conveniently.The traditional QA system technologies can not deal with the complex question with multiple intentions effectively.This may lead to incomprehensive or incorrect intention understanding and unsatisfied answering.Currently,the Chinese medical field lacks suitable KG for medical consultation and diagnosis services,and it is difficult to generate high quality answers.This brings challenges and difficulties to the development of the QA system based on KG.To solve above problems,this thesis investigates into the basic theory and techniques in the field of deep learning and natural language processing technologies and develops a KG-QA for medical guidance robots.Particularly,this thesis proposes a medical entity recognition model and constructs a medical KG.We propose a method of understanding the intention of complex questions.We also have developed an intelligent medical consultation QA system for a medical guidance robot.All the research contents are as follows.1)In order to tackle the problem of the shortcomings of the traditional named entity recognition model in the representation of contextual semantic information in the construction of medical KG,this thesis proposes a medical KG construction method based on deep learning and st ructure prior.We develop web data analysis software to obtain medical data on medical websites,and propose a named entity recognition method fused with Language-Modeling(LM)to extract medical knowledge from it and store knowledge in a local database in a structured form.Then we build a medical KG containing about 140,000 entities of 6 typical categories.And We propose to fuse the LM with Bi LSTM-CRF and IDCNN-CRF models.The Word2 Vec word vectors and pre-trained LM are used to extract the semantic feature information of the text for comparison experiments.Experimental results show that fusion of the LM and the medical entity recognition model can significantly improve the performance of medical entity recognition.The specific keywords and relationship extraction rules are futher used to complete the medical relationship extraction task.Finally,the graph database is used to store medical knowledge data to complete the task of building a medical KG.2)In order to solve the problem of decomp osing the input multi-intention question,an intelligent understanding method of complex question based on semantic analysis and deep learning is proposed.The method is composed of two parts:standardization of syntexal components of the input sentences a nd medical intention recognition.For input sentence standarization,the medical entities extraction and dependency parsing are performed on the input question to obtain the core dependency syntax tree and the correlation between the entities.Then,the syntax standardization is proposed to decompose the input multi-intention question into several simple questions about attribute or relation.The part of medical intention recognition,we use the pre-trained LM to extract features from the standardized questions,and then we input them into the Text CNN to classify the question intentions.The experimental results show that the proposed method can effectively understand the complete intention of the question after decomposing the complex question with multiple intentions.It can not only reduce the impact of natural language colloquial expression on the understanding of the intentions of the questions,but also improve the ability to parse the complete intentions of the questions.3)Based on the medical knowledge graph,medical entity recognition model and complex question intention recognition method proposed in this thesis,a new intelligent question answering system framework is designed,including six modules:human-robot interaction module,question grammatical analysis,question standardization,intelligent knowledge retrieval and natural language answer generation.The design and implementation of an intelligent medical consultation QA system for medical guidance robots is completed.The intelligent QA system supports different modal interactive methods such as voice or text to input user natural questions.We obtain medical entities and relational predicates from the semantic analysis results of the question.The retrieval sentence generated with the corresponding medical entities and the relational predicates is used to extract knowledge from the KG for the answer generation.