A Research On Open Domain Question Answering System Over Knowledge Base

Question Answering is a comprehensive application of natural language processing tasks.Compared to traditional search engines,question answering is designed to allow users to ask questions in natural language and return to the user with a concise answer or candidate answer set.For the era of large data,search engines gradually fail to meet the needs of users through simple keyword matching retrieval techniques.The research of question answering systems has important significance and prospects.On the one hand,it benefits from the outstanding achievements of natural language processing technology represented by deep learning on multiple tasks.On the other hand,it is due to the emergence of large-scale structured knowledge bases in open domain such as Freebase and YAGO,which promote the knowledge base question answering system(KBQA)development.Because of the diversity of language and the huge search space of the knowledge base,the question answering system needs deep semantic understanding of questions.At present,even with simple knowledge base questions and answering in open domain,the current system cannot handle it well.The purpose of this work is to solve the semantic matching problem in this kind of simple knowledge question answering,so as to improve the system performance.The knowledge base question answering system usually needs to perform semantic parsing on the question sentence first,and then performs knowledge base query and reasoning on the parsed semantic representation.According to the different forms of intermediate semantics,question and answering methods can be divided into two lines based on traditional symbol representation and vector-based representation.The method based on symbolic representation will bring the problem of semantic gap,and it depends on high-quality labeling data,so it is difficult to extend to large-scale open domain question and answering.Therefore,the focus of this paper is to better model and match the question and the knowledge in knowledge base under the semantic representation of distributed vectors,and the main work is as follows:1.Fully investigating the basic methods and ideas of the knowledge base question answering system,and building a neural network question and answering baseline system in open domain.The evolution of knowledge base question answering systems from traditional symbol-based semantic analysis methods to representation-based learning methods based on distributed vectors has been analyzed over the past decade.Based on Freebase,a complete neural network question-answering system has been built to serve as a baseline for subsequent research in this paper;2.From the perspective of learning better relation representations and better matching strategies,this paper has explored some ways to improve the performance of relation detection subtasks.In terms of relation representation,this paper first proposes the modeling of the importance of relation structures and the learning of different levels of relation representations.In the matching strategy,an improved parallel match model and an attention-enhanced interactive match model are used.Experiments show that the method proposed in this paper is close to the current best level in the public data set;3.In order to support the follow-up research on the question answering system,this paper builds a Web-based simple question answering system.It mainly provides question answering,mention detection,entity prediction,relation detection and other functions.The system can well identify the entity and relation intent of a simple question,and give the candidate answer set and related evidence at the same time for the user's question.The data and code is available at http://github.com/geofftong/NJU_KBQA.