Research On Deep Learning Based Question Answering Over Knowledge Base

With the ever-growing amount of data on the network,the knowledge bases(KBs)also become larger and larger.However,such a large volume of data and complex structures make it extremely hard for users to access the information efficiently.Therefore,it is particularly urgent and important to use natural language processing technology to automatically answer the natural language questions raised by users with the triple information(knowledge)stored in the KBs.At present,the development of knowledge base-based question answering(KBQA)system has attracted extensive attention in recent years.According to the number of triples required to answer questions,KBQA can be divided into two categories: single-relation KBQA and multi-relation KBQA.In recent years,with the development of deep learning,a lot of works have made some progress by using the powerful representation learning ability of deep learning.However,for one hand,the single-relation KBQA still faces the following challenges:(1)Although the neural network has excellent feature fitting abilities,it lacks interpretability.And the user cannot estimate the credibility of the model predictions,which will bring security risks in practical applications;(2)In the current end-to-end frameworks,(question-subject)and(question-predicate)are often matched separately,ignoring the interaction between each other.For another,relation detection,as the core task of multi-relation KBQA,the existing approaches for KBQA often regard it as a single-label classification task,ignoring the fact that some complex questions correspond to multiple relation paths.To tackle the above shortcomings,the main work of this article includes the following:1)This paper proposes a novel end-to-end KBQA model based on Bayesian Neural Network.It can select entity and predicate simultaneously considering the relevance between entities and predicates existed in KBs.Furthermore,the model can estimate two types of uncertainties,model uncertainty and data uncertainty,of the predictions.Compared with the existing end-to-end KBQA methods,the proposed model achieves the best results on the Simple Question dataset.In addition,this paper additionally exploits the uncertainty measure to carry out misclassification detection experiment and error cause detection experiment.The experimental results prove that the effectiveness of the proposed uncertainty.2)This paper formalizes the relation detection task for KBQA as a multi-label multi-hop relation generation problem for the first time,and proposes a hierarchical sequence relation generation model(HSRGM)based on the sequence-to-sequence framework.By proposing a new decoder relationship generation paradigm,the relation can be predicted at different levels,and the correlation between the relations is considered.The experimental results on the Freebase QA dataset prove the effectiveness of the proposed model.There are five chapters in this thesis.The first chapter defines the task of KBQA,explains the research background and significance,describes the current status of related work,the research motivation and content.Chapter 2 describes the content of KBs and related technology,Chapter 3 introduces a novel end-to-end KBQA model based on Bayesian Neural Network.The model can estimates the uncertainty of the model and data while predicting the answer,providing the user with a basis for judgment.Chapter 4 proposes a multi-label multi-hop relation detection method and related experiments.Chapter 5 concludes the thesis and introduces the future work for KBQA.