Question Answering Approaches For Simple Questions

Question answering is a classic problem in the field of natural language processing.With the rapid development of the knowledge base,more and more open knowledge bases have been established.As an excellent way to store and organize knowledge,the knowledge base can not only store a large amount of knowledge,but also preserve the connection between knowledge.However,the structured storage model of the knowledge base(for example,RDF)makes it difficult for users to directly access the knowledge.The question answering system is an effective way to bridge the gap between the user and the structured storage knowledge base.Question answering over the knowledge base can take advantages of the knowledge base and provide knowledge services quickly.Natural language is the ideal way of human-computer interaction.Through question and answer,people can easily acquire the knowledge of structured storage in the knowledge base.However,the semantic gap between natural language and the huge amount of knowledge structured storage knowledge base has brought great challenges to questions answering.It is an important research topic that how to let the computer understand the natural language and quickly retrieve the corresponding answer.This thesis mainly studies the question answering over knowledge base for simple questions.The simple question means that the question can be answered by a fact(subject,predicate,object)in the knowledge base.The main research work of this thesis is as follows:(1)The question answering based on a pipeline model is studied.The task of question answering is decomposed into two sub-tasks: entity detection and relation detection.The entity recognition model and relation matching model are cascaded to complete the question answering task.In addition,by establishing a matching approach of the question and the entity type,the context information is used to improve the question answering effect.(2)The end-to-end question answering approach is studied where the two sub-tasks of entity detection and relation detection are combined into one framework.The question,entity and relation are encoded by the recurrent neural network.The semantic vectors of the three are mapped to the same semantic space,and then the cosine similarity of the entity,relation and question semantic vector are calculated,so as to sort and select the candidate answers.In addition,a negative sampling method is used to train the end-to-end model.(3)A multi-task end-to-end model is constructed with fusing context information.On a public question answering dataset,our approach improves the accuracy of the existing end-to-end model from 71.2% to 71.8% and reduces the training time from 48 h to 4h.Moreover,a detailed result analysis is done,which finds that self-attention mechanism,the entity candidates set size and the proportion of negative samples have affected the quality of question answering.