A Study On Semantic Parsing For Querying Knowledge Bases

Semanic parsing for querying knowledge bases refers to translating natural language sentence into the form that knowledge bases can execute,and helps computer obtain target information from knowledge bases.In recent years,with the continuous development of the Internet and the explosive growth of human knowledge,the knowledge base has been designed for knowledge storing.Querying knowledge bases with natural language can effectively and conveniently obtain information from them.Different natural language querying tasks have been designed for different types of knowledge bases,such as question answering task for knowledge graph,querying language generation task for tabular knowledge base,reading comprehension for document knowledge base and so on.Semantic parsing for querying knowledge bases is the primary core technique to deal with these tasks.Thanks to the development of machine learning and neural network,semantic parsing on natural language has also shifted from traditional rule-based methods to statistical methods,while deep learning methods have predominated among them in recent years.Compared with the traditional rule-based methods,the advantages of deep learning methods are embodied in the facts that no manual designed features are required,model learning is simple,and models can be conducted in an end-to-end way.This article mainly focuses on the question answering task for the knowledge graph and the querying language generation task for the tabular knowledge base.At this stage,the deep learning method has made positive progress in these two tasks,but there are still many shortcomings.For knowledge base-based question answering task,existing works mainly focus on how to design models to represent natural language questions in vectors.The model performance is often restricted by the database size.At the same time,existing methods ignore the structural information of the knowledge base,which can help the model better understand input natural language sentences.For querying language generation task,most encoders in existing methods simply use pre-trained language models for semantic representation.On the other hand,The historical information in multi-turn scenario is utilized in coarse ways.All of these shortcomings caused insufficient parsing accuracy on natural language sentences.This article works on the knowledge base-based question answering task and the querying language generation task for deal with these shortcomings.First,for the knowledge base-based question answering task,this paper studies the method of training data expansion based on question generation.By constructing a natural language question generation model with knowledge base triples as input,the training set in the question answering task is expanded.Then the topic entity linking module and the relation detection module in the knowledge base-based question answering system are trained on the expanded training set,and the system accuracy is improved.Second,for the knowledge base-based question answering task,this paper studies a multi-level attentive pooling method for question representation.By considering the relational organization structure in the knowledge base,a multi-level attentive pooling method is proposed to represent the relations in the knowledge base in different levels,and then represent questions with the influence of these relation representations.The accuracy of the relation detection in the knowledge base-based question answering system is finally improved.Third,for querying language generation task,this paper studies the semantic parsing method with information integration for the multi-turn scenario.This paper proposes a novel information integration method to realize the full use of historical information with multiple integration modules.And the methed improves the querying language generation accuracy in the multi-turn scenario.Fourth,for querying language generation task,this paper studies the interaction state tracking method for the multi-turn scenario.This paper introduces the idea of state tracking into the multi-turn querying language generation task for the first time.The natural language sentences can be better parsed by tracking and understanding the intent of them in interactions,which leads a better model performance in the multi-turn scenario.