Research On Open Domain Question Answering Based On Knowledge Base

The automatic question answering system based on knowledge base refers to the information retrieval system which can accept the questions described by users in the form of natural language,and can find or infer the answers from the large-scale knowledge base.Automatic question answering is widely used in reading comprehension,chat robot and other fields.It will be the basic form of next generation search engine.The traditional question answering system relies on the size of the corpus of questions and answers collected.It is difficult to find the appropriate answers when facing the questions in the open domain,and the performance of the question answering system is seriously degraded.With the rapid development of knowledge base,a large number of knowledge items are stored.How to use the knowledge base to generate answers to questions has become a research hotspot in this field.The researches mainly focu on three aspects:(1)how to find the knowledge points of the answers from the knowledge base;(2)how to generate the answers in natural language form according to the knowledge points;(3)how to get the knowledge points of the answers through the multi hop reasoning mechanism.To solve the above problems,this paper studies the neural network model of automatic question answering generation based on knowledge base,and realizes the open domain automatic question answering system.The main contributions are summarized as follows:(1)We design and implement the semantic representation method and answer generation model of knowledge base based on TransE.Because the questions from users are in the form of text and the knowledge base is organized in the form of head-relationship-tail,there is a gap in semantic representation between the questions and the knowledge base.To solve this problem,we design and implement the semantic representation method of knowledge base based on the TransE model,and obtain the tail entity as the answer by using the head entity and attribute relationship of the triple of knowledge base Knowledge of the case.Specifically,we first initialize the state vector with the entity in the question,then predict the attribute relationship,and finally get the answer by adding the state vector and the predicted attribute relationship.In the named entity recognition of questions,we implement the named entity recognition model based on Bi LSTM and CRF.The evaluation results on the open data set World Cup2014 show that the precision rate reaches 85.1%,the recall rate reaches87.3%,and the F1 value reaches 86.1%.Compared with the baseline model,the performance has been improved to some extent.(2)We design a question answering model to generate natural language sentence answers.Natural language sentences are used in people's daily communication.How to transform the word level or phrase level answers from the knowledge base into natural sentences has become a challenge for automatic question answering.To solve this problem,we implement a neural network model combined with a copy mechanism.When generating answer sentences,three models are designed to predict the probability of generating words from question sentences,knowledge base and vocabulary respectively.The evaluation results on the open data set Gen QA show that BLEU value reaches 0.43,which is 0.01 higher than the traditional method.(3)We design an automatic question answering model for multi hop reasoning mechanism.Another challenge in automatic question answering is the generation of answers involving multi hop reasoning.Simple questions usually can get answers by retrieving a triple in the knowledge base,but it is difficult to get answers directly for a triple in complex question.It needs to deduce multiple triplets to get answers.To solve this problem,we designs an automatic question answering model which integrates multi hop reasoning mechanism.We integrate an iteration mechanism on the model of 1)above.By introducing the semantic representation vector of questions and the state vector of answer generation and updating the predicted attribute relations,we can get the answers of each round and the question representation of next round until we get the final answer.In addition,due to the semantic gap between knowledge base and question expression in the open domain,the coverage of answer acquisition is low.To solve this problem,this paper introduces attention mechanism between the semantic expression vector of question and the knowledge base,so that the expression of knowledge base can be concerned when the expression of question is updated.The experimental results on the open data set Path Question-2H/ 3H show that the precision is 0.989 and 0.972.The introduction of attention mechanism improves the accuracy of the automatic question answering model based on multi hop reasoning.This paper focuses on the task of automatic question answer generation in open domain,we design and implement the semantic representation method and answer generation model based on knowledge base by using TransE.We also implement the generation model of natural language sentence answer and the automatic question answer model with multi hop reasoning mechanism.The evaluation experiment resultson open dataset verify the validity of the models proposed in this paper.