Research On Semantic Matching Method Of Deep Neural Network For Knowledge Base Question Answering

With the rapid development of the Internet,the amount of available information has grown exponentially,the Internet has become an important way for people to obtain information.Currently,search engines can help users search for information conveniently to a certain extent,but they may also return a large number of results that are not related to user needs.The automatic question answering system based on knowledge base is aimed at the questions of natural language form,and the corresponding answer is obtained by using natural language processing technology to query,reason,and match in the structured knowledge base.The system not only provides users with friendly human-machine interaction that supports natural language,but also helps users to obtain valid information accurately and efficiently.Therefore,the question answering over knowledge base is an important research direction of natural language processing and information retrieval.Traditional methods of symbol-based knowledge base question answering are often difficult to solve the problem of text semantic gap.The vector space modeling method attempts to solve the problem in the form of vectors,but the method only involves the shallow semantic representation.In addition,the grammatical structure and expression form of Chinese are more complicated than English,so the research of Chinese knowledge base question answering is facing more challenges.This thesis mainly studies text semantic learning and matching based on deep neural network,then applies the method to the knowledge base question answering.The process of question answering is regarded as the semantic representation matching of questions and knowledge bases.The main content consists of the following three aspects:(1)For the semantic learning and matching of text,this thesis proposes a text semantic matching method based on Bi-LSTM with attention mechanism.The method firstly uses the LSTM and attention mechanism to encode the text,then aggregates the text semantic matrix into semantic vectors from different angles through three pooling mechanisms.Finally,the semantic vectors of the text pair are sharpened and passed to the multi-layer perceptron to obtain the result of semantic matching.(2)For the entity linking task,this thesis designs a binary classification method based on knowledge base.In order to simplify the task,the candidate entities of the corresponding question are firstly found by the knowledge base,then the feature of the candidate entity is extracted through feature engineering,and the candidate entities are classified by the Gradient Boosting Decision Tree algorithm.Finally,the candidate entity who is the highest score is selected as the best entity according to the prediction score of the model.(3)Based on the deep text semantic matching method and entity linking technique,this thesis applies them to the knowledge base question answering.Firstly,the entity of the question is recognized by the module of entity linking,and the relevant candidate answers are found in the knowledge base,and then the deep neural network is used to learn the semantic representation of the question and the candidate answer information.Finally,the best answer is selected according to their semantic matching results.This method can better solve the semantic gap problem in the traditional knowledge base question answering method,thus improving the accuracy of question answering.In addition,this thesis also uses the model fusion method to further improve the accuracy of the knowledge base question answering.