Research And Application Of Question Answering Matching Model Based On Deep Learning

Question answering uses computer natural language processing technology to parse the natural language questions asked by users and then identify the intention of users' questions.Then it answers users according to the knowledge learned by the system.With the explosive growth of data and the improvement of computing capability,the research of question answering technology has gradually shifted from traditional method based on feature engineering to method based on deep learning.In the medical field,there are a large number of complex medical data,which provides a good foundation for the application of question answering technology in the medical field.According to the source of knowledge,question answering can be divided into two parts: question answering based on reading comprehension and question answering based on knowledge graph.This thesis focuses on question answering based on knowledge graph.Knowledge graph provides structured knowledge for a question answering system in the form of entity-relation-entity.In this thesis,we apply information retrieval to realize knowledge graph question answering.After a user has asked a natural language question,the system parses the question and then identify the entity mention in the question,makes entity linking to get the candidate entity and candidate relationship.The system matches the candidate entity and the candidate relationship with the question semantically and selects the triple corresponding to the candidate entity and candidate relationship with the highest matching score as answer.This thesis mainly studies and improves the related technologies of entity mention recognition and semantic matching,and applies them to COVID-19 knowledge graph question answering system.The main work is as follows:(1)In the entity mention recognition model,the embedding representation layer and encoding layer are improved respectively.For the embedding representation layer,a multi-granularity representation method is proposed,which combines the representation methods of character level and word level.The character level vector is generated by BERT model.This method conveys rich semantic information in the process of transforming natural language questions into representations that computers can understand.BERT is also introduced as an encoding layer.In the experiment,the two improved methods are compared.It is found that the former improved method is better than that latter.In addition,ablation experiments are carried out to verify the effectiveness of the multi-granularity representation method.(2)In the process of semantic matching,according to the characteristics of the knowledge graph question answering task,this thesis proposes a multi-stage candidate answer re-ranking model.First,rough sorting model filters a large number of negative samples,and then the final answer is obtained by fine sorting.In the rough sorting stage,RCNN-Att semantic matching model introducing attention mechanism is proposed.In the fine sorting stage,SA-BERT,a more complex semantic matching model based on adversarial training,is proposed.The effectiveness of the multi-stage design is proved by the ablation experiment.And the effects of the attention mechanism and adversarial training part of the model are also verified by the ablation experiment.(3)In order to verify the practical value of the algorithm proposed in this thesis,this thesis uses the proposed model to design and implement a COVID-19 knowledge graph question answering system,which is according to the procedure of the knowledge graph question answering based on information retrieval.And the model is pretrained on medical domain data to complete the model migration,which verifies the value of the model in COVID-19 knowledge graph question answering application.