A Semantic Parsing-based Approach For Knowledge Base Question Answering

Knowledge base question answering is an important task in the field of natural language processing.It aims to retrieve the answers to the natural language questions from the knowledge base.How to cross the gap between natural language and knowledge base query language is a challenge.Utilizing semantic parsing to translate natural language questions into formal queries is a direction that has been widely studied.However,for increasingly complex natural language questions,the existing semantic parsing methods have exposed some shortcomings such as lower accuracy for relation detection and excessive noise produced in formal query building.This thesis aims to leverage semantic parsing to transform natural language questions into formal queries,so as to realize automatic question answering.The overall process of question answering consists of three stages: entity linking,relation detection,and formal query building.In this thesis,assuming that the entity linking has been completed,and focus on relation detection and formal query building.The main contributions of this thesis are as follows:(1)A multi-head attention based model for relation detection.First,use this model to calculate matching scores of the knowledge base relations and the questions.Then,top-k relations with highest scores are selected as candidate components for formal query building.Comparing to baselines,the proposed model in this thesis is able to handle "long sequence".(2)A formal query building approach based on structure prediction.First,predict the query structure based on the input question.Then,use the results of entity linking and relation detection to generate candidate queries according to the predicted structure.Score candidate queries and output the top-scored query.Finally,execute this query in the knowledge base to retrieve the answer.Different from the traditional structure enumeration-based strategy,this thesis first proposes to leverage predicted query structure to constrain candidate generation.In addition,in order to ensure the flexibility,this method does not use any predefined structure template,but employs a framework that can automatically generate the query structure.(3)The evaluation of the proposed method on three common knowledge base datasets.Experimental results show that the proposed method achieves the state-of-the-art results on complex questions,and staying competitive on simple questions.The significance of this work is to provide an effective and interpretable solution to the knowledge base question answering task.This solution shows good performance when dealing with complex problems involving multiple relationships and constraints.At the same time,the pipeline process of semantic parsing and the output of logical form can intuitively reflect the understanding of the semantics,which benefits for error analysis.Therefore,this thesis is of great significance to the increasingly complex question answering scenarios.