Research On Question Retrieval In Restricted Domains Question Answering System Based On Deep Text Semantic Matching

With the continuous development of today’s science and technology,search engines greatly facilitate people’s access to knowledge and information.However,due to the lack of understanding of natural language,search engines cannot identify the real intention of users.The emergence of question answering system has solved this problem well.With the development of question answering system,more and more scholars at home and abroad have carried out research in the domain of question answering system.The question retrieval of question answering system is one of the important research topics.In this paper,through the research confined domain question answering system question retrieval problems related research status at home and abroad,combined with the characteristics of restricted domain question answering system question,through the integrated use of a variety of text similarity calculation method and depth of text semantic matching method,character question surface feature extracting and deep semantic characteristics,combined with unsupervised algorithm and monitoring algorithm,this paper proposes a text semantic matching based on the depth of the restricted domain question answering system questions retrieval model.The question retrieval model based on deep text semantic matching is proposed in this paper,which mainly includes two sub-models,the similar question recall model based on multiple methods integration and the deep semantic text matching model based on Bert-TextCNN.For the questions to be retrieved entered by users,the model first uses a variety of string-based text similarity calculation methods,constructs a recall model of similar questions based on the integration of multiple methods,and selects TOPK candidate questions.Then,the selected TOPK candidate questions are matched with the questions to be retrieved one by one,and the deep semantic text matching model based on Bert-TextCNN is used for further matching.Finally,the remaining questions are sorted according to their similarity,and the retrieved similar questions are returned.In order to verify the model,this paper uses the data from the actual application scenarios of Ant Financial Brain provided by the Ant Financial Competition to conduct an experimental study.The results show that compared with the single string-based text similarity calculation method,the proposed similar question recall model based on the integration of multiple methods can effectively improve the recall rate of results.The traditional text matching method based on twin network architecture,namely,the vector representation of questions and then the vector similarity calculation for text matching,is not suitable for the question retrieval task of restricted domain question answering system.The text matching model based on Bert-TextCNN proposed in this paper has better performance.The experimental results finally verify the effectiveness of the proposed model.