Research On The Method Of Obtaining Answers In Knowledge Base Question And Answer Based On Representation Learning

Knowledge based question answering?KBQA?is to give right answer to natural language questions by using structured knowledge in knowledge base.In recent years,the emergence of large-scale knowledge base such as Freebase and DBpedia has provided abundant knowledge sources for KBQA research.At the same time,the progress of deep representation learning technology has also brought a turning point for KBQA research.Under the above background,this paper adopts an answer acquisition method based on representation learning to improve traditional information extraction and an answer acquisition method based on distributed knowledge base representation learning,which further improves the accuracy of answer acquisition in KBQA.The specific contents are as follows:The answer acquisition based on representation learning to improve traditional information extraction method mainly use the idea of traditional information extraction,combined with representation learning and dynamic memory network?DMN?model.Firstly,name entity recognition?NER?is used to get question headword,and link the headword to correspond entity in the knowledge base.Secondly,the knowledge base sub-graph centered on the entity is queried in the knowledge base as candidate answer set.So far,the steps of this method are basically same as traditional information extraction method.Then combined with the word embedding representation technology and the DMN model to obtain answers.The improved model overcomes the defects of traditional methods lying too much on semantic knowledge and feature engineering,and further improves the accuracy of answer acquisition.The answer acquisition method that integrates knowledge base representation learning is mainly aimed at the simple questions of single fact.By NER?entity link and relationship recognition steps to get correspond entity0)₄)and relation₅)of the question.Combining TransE?TransR and TransH knowledge representation learning model,the candidate answer set and its initial ranking are obtained by link prediction through the entity0)₄)and relation₅),then get final ranking by union the score of the semantic relevance of candidate answer and the question,return the first one as right answer.This method makes full use of the structural information of entities and relationships in the knowledge base,which significantly improved the accuracy of the answer acquisition for simple questions of single fact type.In this paper,the above two methods are tested on the NLPCC KBQA and SimpleQuestion public datasets respectively.The results are measured by multiple indicators and comparative experiments.The final accuracy scores are compared with other methods on the correspond datasets,which further proves the superiority of the two answer acquisition models proposed in this paper.