Research On Question Answering Over Knowledge Base Based On Automatic Generated Template

Question answering over knowledge base,as an important task in the field of natural language processing,is an information retrieval method to search or infer answers in the knowledge base for natural language questions.The study of question answering over knowledge base is of great value and significance.It is a classical method to use manual template for question answering over knowledge base.However,when it comes to complex natural language and increasing number of questions,traditional manual template method exposes the problems of high difficulty in construction and lack of quantity.This paper proposes an automatic generated template method for question answering over knowledge base.The templates are generated from training question-answers pair utilizing relation dictionary and are used to answer test questions.For test questions that do not have template matching or return an empty set of answers,this thesis proposes a method which is based on sentence similarity.The main research contents of this thesis are as follows:?1?Proposes an automatic template generation method.A relation dictionary is constructed by distant supervision,in order to build the mapping between natural language and the knowledge base.Extract question template constructed by the part-of-speech and dependent syntax from the training question-answering pairs,and extract query template from the query graph.The combination of question template and query template constitutes the template of this paper.?2?Answer question with the automatically generated templates.For test question,match the questions with the template bank.And generate candidate queries based on successfully matched templates.This paper use Learning-To-Rank method to rank candidate queries,and select the query result of the best candidate query as the answer for the question.?3?For test questions that no template matching or all the query results of candidate queries are empty,this paper proposes a method using similarity calculation to supplement answering.This paper construct a deep learning sentence similarity computation model with convolutional neural network and LSTM.Compute the similarity between the training question set and the test question.Construct query of the test question with the assistance of the training question which has the highest similarity and execute the query in the knowledge base.The effectiveness of the proposed method is verified experimentally on common datasets.Experimental results show that the method this paper proposed has improved accuracy,recall rate,precision rate and F₁ value on WebQuestionsSP and SimpleQuestions.