Research On Key Technologies Of Automatic Question Answering Based On Fusion Knowledge Representation

In recent years,with the development of artificial intelligence and deep learning technology,automatic question answering systems have been widely used in many fields.At present,there are mainly two technical routes for the implementation of the automatic question answering system.One is to use search methods,including traditional search models and search for candidate answer documents using a similarity calculation model based on deep learning,and second is to automatically generate answers.Using a large amount of question and answer corpus to train and learn to get the correlation model between questions and answers,automatically generate answer sequences.In the process of constructing an automatic question answering system,no matter whether it is a search or an automatically generated method,it is inevitably faced with the problem of lack of knowledge.For the understanding of question sentences,if there is no context or related knowledge,it is difficult to understand the meaning of the question,and this poses a challenge to the retrieval or generation of answers.In addition,due to the diversity of knowledge existing forms,the use of knowledge information in the deep semantic representation model becomes more important.Therefore,this paper discusses the application of three different forms of knowledge in question answering systems,include the keywords information,the extensive questions and answers on the Internet,and the currently widely studied knowledge graphs.The research content is as follows:1.In order to solve the limitation of lack of knowledge for question sentences in deep learning model,this paper proposes a deep learning model with knowledge memory unit(KM-Bi LSTM),adding keyword knowledge features to knowledge memory unit for question and sentence.The measurement of the similarity between the answers is used for the sorting of the candidate answers.The fusion of the representation of the question and the knowledge information makes the semantic representation of the question more accurate.This article examines the effect of the model on answer selection tasks through experiments on the datasets selected from the Trec QA and Wiki QA answers.2.For the problem of insufficient representation of the question in the community question answering retrieval,the paper proposed a deep semantic model(ECQ-Bi LSTM)that integrates the external candidate question and answer to express the knowledge information to model the semantic expression of the question.It dynamically adds the knowledge information contained in the relevant question and answer pair to the question representation according to the degree of relevance.This article discusses the method of obtaining external knowledge information and the effective fusion method of question representation and external knowledge information through the Trec Live QA contest and experiments on Trec QA and Wiki QA.3.We propose a KV-Seq2 seq model that integrates the Knowledge graph triplet knowledge information to solve the problem of lack of control over target answer generation in the Factual Q & A model.This method designs a triple-knowledge storage structure with a key-value structure.Based on the traditional Seq2 seq model,it fuses the intermediate encoding state with the knowledge graph knowledge information,enriches the representation of questions and makes the answer sequence,to generation more accurate and diverse answers.This paper makes extensive and in-depth study of the application of knowledge representation in question answering system.A large number of experimental results show that the question answering system with integrated knowledge representation can improve the effect of answer selection and answer generation,and can answer users problems more accurately and intelligently.