Realization And Optimization Of Domain Dialogue System Based On Rasa

Dialogue systems are systems that communicate with users through natural language.It is usually embedded in the application system to provide customer-service conversations,content search and other functions,bringing users a more intelligent experience.Enterprise Service Management Cloud Platform is a platform for government management agencies to collect enterprise information in order to provide services for enterprises.In order to conveniently understand the enterprise information within the jurisdiction,it is necessary to provide an intelligent retrieval tool.Therefore,based on the Rasa multi-turn dialogue framework,the paper designes and realizes the Intelligent Retrieval Assistant for the application field of enterprise information,and realizes the enterprise information query and statistics function based on voice question and answer.At the same time,the algorithm of Rasa framework is optimized and improved to improve the dialogue service quality.The main work is as follows:Intelligent Retrieval Assistant consists of Query APP,Multi-turn Dialogue Service and Data Query Interface.Based on the common dialogue data set in enterprise service field,the paper uses Rasa to implement Multi-turn Dialogue Service through data construction,data enhancement,action definition and component configuration,and realizes the functions of intention classification,entity recognition,multi-round dialogue management and action feedback.Data Query Interface completes the interaction between Multi-turn Dialogue Service and enterprise information database.Query APP provides users with dialogue and interaction applications.Intelligent Retrieval Assistant has been deployed online and can provide services stably.Aiming at the problem of low dialogue quality of functional components in Rasa framework,the paper analyzes the reasons and puts forward the corresponding algorithm optimization and function improvement.Aiming at the low entity recognition accuracy of the natural language understanding module,BERT+CRF model is applied to the assistant instead of CRF model to improve the entity recognition accuracy.Aiming at the problem of recognition of pronouns,the paper proposes a rule-based method to identify simple pronouns.Aiming at the problem of company abbreviation match after entity recognition,the paper proposes a method based on the Pinyin similarity of the shortest editing distance and cosine similarity to calculate the similarity of candidate words to match words,finally it improves the recognition accuracy of company names.To solve the problem of low prediction accuracy of action prediction in the dialogue management module,experiments on LSTM,GRU,GRU+Attention and MLP models are carried out in this paper.The MLP model with the best effect is selected to replace the original LSTM model to improve the accuracy of action prediction.The paper designs and realizes an intelligent dialogue system for the Enterprise Service Management Cloud Platform to provide more convenient enterprise information query service for the platform manager.