Research On Relation Detection For Question Answering Over Knowledge Bases

In recent years,the amount of available information has grown exponentially.How to efficiently and accurately obtain the information that users need has become an urgent problem.Consequently,automatic question answering systems have developed rapidly.Behind these systems,there are often huge and complex Knowledge Graphs(KG),also known as Knowledge Bases(KB),which describe concepts and relations in the world using a more intuitive way.Question Answering over Knowledge Bases(KBQA)analyzes natural language questions and finds answers in the knowledge graph.During this process,how to identify the question's corresponding relation path in KB,known as relation detection,has become a challenge and attracted much attention.Most existing relation detection methods only consider semantic information of question and relation which pays attention to relations seen in the training data,and ignores lexical information,thus unseen relations can't be detected.And the relation path is limited to two hops which needs to be further expanded.Therefore,this paper focuses on relation detection,the main work and contributions are as follows:1.This paper first proposes a multi-strategy scorer,which combines word-level scorer and sentence-level scorer to score relations.Word-level scorer focuses on lexical interaction between question and relation,but can't remember seen relations.To make up for this deficiency,sentence-level scorer constructs a QA set using the question template and corresponding relation in training data,thus converting the relation detection task to the common task of sentence matching.2.However,training the two models separately is more difficult and less effective than training a single model.Therefore,this paper proposes a kind of method whichintegrates semantic and lexical matching,and implements two models which are based on word vectors and pre-trained BERT encoding respectively.The model based on word vectors encodes semantic representation of questions and relations with RNN to complete semantic match,and constructs the similarity matrix of question words and relation words to complete lexical match.The model based on BERT combines rich information of pre-trained models with attention mechanism to complete semantic and lexical matching.Also,the two kinds of methods above can handle multi-relation path.3.In order to fully verify the effectiveness of proposed relation detection methods,this paper implements two systems based on current popular KBQA frameworks: Information Retrieval based Question Answering(IRQA for short)and Semantic PARsing for Question Answering(SPARQA for short).Finally,this paper demonstrates the effectiveness of relation detection methods through experiments,which not only achieves state-of-the-art in relation detection tasks,but also helps our KBQA systems get good results on different datasets.