Design And Implementation Of Task-Oriented Dialogue System Based On Deep Learning

With the rise of Artificial Intelligence,various personal assistants and customer service systems are developing rapidly.As the core of the services mentioned above,dialogue systems have attracted more and more attention.According to the applications,dialogue systems can be categorized into two groups:task-oriented systems and non-task-oriented ones.Non-task-oriented systems converse with human on open domains to provide reasonable responses and entertainment.Task-oriented ones aim to assist users to complete certain tasks,which is most widely used and is key to automation,thus this thesis focuses on the design and implementation of task-oriented dialogue systems.Up to now,most deployed task-oriented dialogue systems are based on manual features or handcrafted rules.However,the rules designed for a typical domain in a rule-based dialogue system can hardly suit other domains.As for the systems based on statistical models,extracting manual features is not only expensive but also time-consuming.To solve the problems above,in this thesis,we design and implement a task-oriented dialogue system based on deep learning.The system implemented in this thesis includes seven modules:Web server,text preprocessing,natural language understanding(NLU),dialogue state tracking(DST),dialogue policy(DP),natural language generation(NLG)and background knowledge management.In particular,we propose a natural language understanding model based on hierar-chical attention named HA-NLU and a dialogue state summary model named E-DSS.HA-NLU parses semantic feature at different levels including the character level,the word level and the sentence level.For a particular level,this thesis uses a semantic parsing network based on LSTM and CNN in HA-NLU to extract features from the text.The structure based on hierarchical attention in HA-NLU offers a tunnel transferring the semantic feature from lower levels to higher levels dynamically.E-DSS calculates the entropy of each probability distribution maintained in the dialogue state to reduce the complexity of deep learning models implemented in the following modules,and thus reducing the complexity of the whole system.Experiments show that comparing with models focusing on only semantic feature at word level,HA-NLU can raise the accuracy of system response,E-DSS can effectively accelerate the system without negatively impacting the accuracy of system response.Finally,we perform functional and performance testing on our system and the result shows the system meets the requirements.