Knowledge Based Question And Answering Study Based On Deep Neural Networks

Compared with the traditional search engine,knowledge base question answering can directly return accurate answers.It has been widely used in intelligent customer service,voice assistant,recommendation system and other fields.It can partly replace manual labor and reduce labor cost,which has great research value.At present,there are three main methods of knowledge base question and answering,namely semantic analysis,information extraction and vector modeling.However,these methods involve rules and templates defined manually,which are not universal.With the development of deep neural network,the end-to-end neural network model is used to deal with knowledge base question and answer.On this basis,the research of this paper mainly includes two key technologies in knowledge base question and answer: entity link and relationship detection.The details are as follows:(1)In the stage of entity linking,aiming at the polysemy problem of the same name entity in natural language problems,the potential relationship between relational words and candidate entities in the knowledge base is explored,an entity linking method based on bidirectional long-term and short-term memory network conditional random field(bilstm-crf)model is designed.The method realizes entity disambiguation according to the similarity of problem relation words and candidate relation,firstly,the bilstm-crf model is used to label the sequence of natural language problems to get entities,which effectively overcomes the dependence on artificial features.Then,the relational words in natural language problems are extracted according to the part of speech according to certain rules.Finally,the string similarity and similarity between the problem relational words and candidate relations are calculated Semantic similarity,mapping problem relation words and candidate relations in knowledge base,sorting and screening the identified entities by using candidate relation information in knowledge base,so as to reduce noise data,realize entity disambiguation,and effectively improve the accuracy of entity link.(2)In the relationship detection stage,aiming at the semantic gap between the diverse natural language problems and the knowledge base triples,a multi granularity attention mechanism relationship detection method is proposed.Candidate relationships are represented from multiple perspectives to obtain word level and relationship level information,and the correlation between problems and knowledge base relationships is modeled from multiple granularity.To solve the problem that a single attention mechanism can not capture the correlation between two pieces of text,multiple attention mechanisms are introduced to aggregate vectors more effectively to retain the original information and achieve fine-grained alignment between characters in relation detection.Finally,according to the similarity between the candidate relationship and the question,the candidate answers are sorted and the final answer is obtained.This paper conducts experiments on simplequests data sets,and compares with the results of other knowledge base question and answer models.The experimental results show that the method proposed in this paper has significantly improved the accuracy and recall rate.