Research Of Knowledge Base Question Answering With Query Graph Generation

With the rapid growth of information,people have higher and higher requirements for the convenience of obtaining information.Unlike traditional search engines that require users to summarize keywords and find information from results,knowledge base question answering system can return accurate and concise answers based on the user's natural language input,which can save people's time and attention.Knowledge base question answering includes two tasks:answer-supervised question answering and query-supervised question answering.The former does not rely on query annotations and can accelerate the implementation of applications.Among them,the method of generating query graphs combined with some rules can handle some complex questions under the premise of weak supervision.The latter trains neural network with the supervising of query annotations to transform question to query,the method of automatically generating query graphs can achieve acceptable results and maintain good portability without relying on manual rules.Therefore,the research of knowledge base question answering based on query graph generation is of great significance.The existing answer-supervised query graph generation methods ignore the structural information of query graph in the sorting stage.The query-supervised query graph generation method faces the problem that the syntactic information such as the structure of the question is ignored by the encoder.This paper has launched a series of work on the above issues,the main contents include:Aiming at the problem that the structural information in the answer-supervised query graph generation method is ignored by the ranking neural network,this paper constructs a query graph generation method based on feature arrangement.The method introduces the Tree-LSTM to encode the structure feature of query graph and performs the query graph ranking together with the similarity feature extracted by the pre-trained model BERT and other manual features.The query graph generation method based on the feature arrangement is 0.8%and 0.1%higher than the best comparison methods on average F₁ value of tow commonly used English question answering datasets,of which the structural encoding contributes 0.3%and 0.2%.The experiment results verify the effectiveness of the structure encoding and the entire model.In view of the problem that the query-supervised query graph generation method is facing the problem that the syntactic information is ignored by the encoder,this paper designs an automatic query graph generation model based on the encoder-decoder framework.The model uses Bi-LSTM to encode word embedding and part-of-speech tagging information.Use graph neural network to encode dependency tree information and dependency label information.Experiments on two public English datasets show that the query graph generation model that incorporates syntactic information has better performance.The average F₁ value of the two datasets is 3.2%and 3.1%higher than the best comparison models respectively,and all the syntactic information contributes 1.8%and 2.9%.Finally,this paper applies the proposed query graph generation model fusing syntactic information together with general entity recognition and entity linking methods to the music domain knowledge base question answering system.By showing the function of the question answering system,the practicability of the query graph generation method in the music domain knowledge base question answering task is verified.