Research On Key Reasoning Techniques For Factoid Question Answering System

Benefiting from the rapid development of computer and the explosive growth of data,natural language processing technology based on deep neural networks has been widely used in information retrieval,machine question answering and other fields,and has achieved good results.The current intelligent question answering technology has surpassed the performance of traditional question answering systems on multiple question answering data sets,but existing intelligent question answering systems still need to use the matching degree of related articles and questions to extract answers.The application of knowledge and common sense is relatively insufficient.Based on this,this thesis conducts research on key technologies of reasoning for fact-oriented question answering systems.Factoid questions that can be answered with simple facts expressed.Commonsense reasoning is needed to answer factoid questions.Enhancing the capacity of commonsense reasoning is a difficult task.This thesis designs and implements a factoid questioning system based on a pre-trained language model.The commonsense reasoning capacity of the system is improved by enhancing the commonsense semantic understanding of the pre-training model as well as by increasing and improving knowledge sources.Finally,the performance on the Commonsense QA is improved.The main contents of this thesis are as follows:(1)A method to enhance commonsense semantics using open corpus is proposed.First,this thesis analyzes the English definition corpus in the dictionary,and determines the existence of co-occurrence relationships among question words and answer words in the corpus.Then a method to acquire and filter the open corpus is proposed.This method combines the recommendation engine and Elastic Search fuzzy search to supplement missing word definitions.And this method also filters out some corpora that are useless for commonsense semantic enhancement.Finally,based on the pre-training model ALBERT,a commonsense reasoning factoid question answering system fused with corpus is designed and implemented.And the effectiveness of the enhancement on commonsense semantics is proved through experiments.(2)A method to complete the knowledge graph using pre-training model is proposed.The incompleteness of the knowledge graph weakens its advantage as a knowledge source for question answering systems.In order to improve the knowledge source,this thesis proposes a link prediction model KGALBERT,which uses the semantic information in the entity description to enhance the generalization.Experiments on multiple data sets prove that the model can effectively reduce the incompleteness of the knowledge graph.(3)A method to enhance the commonsense reasoning capacity of the factoid question answering system using knowledge graphs is proposed.First,this thesis proposes a method of knowledge selection oriented to Concept Net.Accoring to the confidence,this method provides the most relevant Concept Net commonsense triples for the commonsense question answering data set.Meanwhile,a Concept Net entity prediction method based on KGALBERT is proposed,which supplements the commonsense knowledge of ConceptNet.Finally,based on the pre-training model,a factoid question answering model fused with knowledge graphs is designed and implemented.Experimental results demonstrate that compared with other models,this method can effectively exploit knowledge graphs for common sense reasoning.