Dialogue Management Algorithm Based On LSTM With History Attention And Its Application

Dialogue System is a computer system that can communicate with human beings using natural language.It provides a convenient human-computer interface.Unlike chat bot,Task-oriented Dialog Systems are often used in several single scenarios,it will complete a specific task in each scenario.The difficulty of developing task-oriented dialogue system lies in the design of dialogue management.With the help of Rasa Dialogue System development framework,this paper proposes a dialogue management method based on LSTM with History Attention,and uses this method to develop a dialogue robot for illegal vehicle query and driver's license score query.Dialogue management includes two parts: dialog status tracking and dialog strategy.The Rasa Core module keeps the historical information of dialog and the user's current intention,it records the dialog state of each round of dialog,and achieves the tracking of dialog state.Therefore,this paper focuses on the design of dialog strategy.Rasa builtin dialog policy algorithm REDP has excellent test results on hotel consultation dialog data.In this paper,we use Word2 vec model to pre-train the dialog state,map user intentions and system actions to the same semantic space,and get their vectors.At the same time,we design an LSTM dialog policy model based on history attention by using REDP algorithm.After many comparative tests,the final model in the test data performance is not bad,basically reached the same test level with the REDP algorithm.This paper also develops a dialogue robot in the field of transportation by using the rasa dialogue robot development framework and the above dialogue policy model.The development process focuses on the definition of the dialogue field and the collection of the discourse data.Domain definitions include user intentions,named entities and system actions,while corpus is manually constructed on the basis of existing data of Dalian Traffic Police Wechat.On rasa development platform,after adding necessary functional components and dialogue policy model,editing configuration files and training each module separately,the system is designed and completed.The system basically realizes two functions: the query of illegal vehicles and the query of driver's license score.The system has been debugged and optimized many times,and the final result is satisfactory.