| Question answering（QA）system answers user’s questions accurately and concisely in the form of natural language.Because of its efficient form of information service,it has attracted the attention of many researchers.And it has gradually developed from information retrieval based on Feature Engineering to intelligent question answering based on machine understanding.The latter aims to understand the semantics of the question and infer the answer based on the context information.Current QA methods rely on pre-training language models to model text semantic information.However,context and questions in QA tasks differ greatly in both semantics and structure.Pre-training models based on Self-attention are not good at capturing this information.Additionally,Mainstream methods help the model understand the meaning of the question more accurately by retrieving additional background knowledge.However,the noise knowledge introduced by retrieval will interfere with model reasoning and reduce the effect of question answering system.Based on this,this master’s thesis carries out the research and implementation of question answering system based on semantic understanding.And conduct in-depth research on text representation,semantic understanding,knowledge reasoning,etc.In detail,the research work of this paper is as follows:（1）A semantic enhancement method based on bidirectional multihead attention is proposed.Pre-training language models rely on Self-attention to capture the global relationships of text sequences,it rarely models the semantic and structural difference information between context and question in question answering tasks,which causes the model to have a poor semantic representation of the question and context.At the same time,the model is affected by noise statements in the context,making it difficult to focus on critical information.To solve these problems,we build a semantic enhancement network based on bidirectional multi-head attention for the pre-training language model.A question-aware context alignment module and a context-aware question alignment module are implemented respectively to achieve better text representation.Experiments show that this method can enhance the semantic understanding ability of the model,and even for the strongest language model Albert,there is still 1.7 % optimization.（2）A knowledge inference method based on entity interpretation enhancement and context perception is proposed.The current reasoning method separates the reasoning process from the text information,which limits the structured reasoning ability of the model,At the same time,the inference process is disturbed by noise knowledge,which leads to the unsatisfactory effect of QA models.We propose a reasoning method based on context awareness and a weight calculation method based on entity interpretation.The influence of irrelevant knowledge on reasoning is weakened by calculating the semantic similarity between entity interpretation and context;In the process of reasoning,context information is added to connect semantic space and symbol space to enhance the reasoning ability of the model.The experimental results on two different datasets show that this reasoning method is better than the previous knowledge reasoning model.（3）A question answering system based on semantic understanding is constructed.The QA system consists of front-end module,question processing module,knowledge retrieval module and question-answering module.The core question-answering module is implemented based on the method of（1）（2）.In order to improve retrieval efficiency,a Wikipedia text retrieval system based on Elasticsearch is also constructed to provide the system with text knowledge related to problems.Finally,a functional requirements test of the system are carried out.The test results show that the system basically meets the use requirements and can provide users with Knowledge QA services. |
