Research On End-to-End Task-oriented Dialogue System Based On Deep Learning

Nowadays,task-oriented dialogue system has become the focus of research at universities and many large Internet companies as well as some startups.At present,the research on task-oriented dialogue system is mainly based on pipeline method or end-to-end method.The pipeline method requires a large number of domain-specific handcrafting,as well as credit assignment problem and interdependencies of modules in the process.However,these problems can be avoided by using the end-to-end method.This thesis proposes an end-to-end approach to implement a task-oriented dialogue system based on supervised deep learning,which is used to help users find restaurants.Because of the difficulty in collecting task-specific dialogue data and the high noise of data,this system adopts delexicalisation and weight sharing strategy to reduce the data required by the training model and improve the generalization capability of the system.This system consists of five modules:natural language understanding module,dialogue state tracking module,database operation module,dialogue strategy module,and natural language generation module.The main research contents of this work include:(1)This system uses delexicalisation to preprocess the sentences input by users.According to the slot-value pairs defined in the ontology,the slots in the sentences and their corresponding values are delexicalized to improve the generalization capability of the dialogue state tracker.(2)In the natural language understanding module,this system uses bidirectional LSTM to model the NLU module.The model takes the delexicalised sequence as input and outputs the distributed representation of user's intention.(3)This system uses CNNs to extract the features of user input sequence.and applies RNNs to model the dialogue state tracker.The dialog state of the current turn is tracked from the current user input and the historical dialog,which is represented by the probability distribution of the values of each slot.The values with the highest probability of each information slot are combined to query the database.(4)The dialog policy module integrates the current user intention,the current dialog state and the database query results to learn the system action of the current turn.The system action that the dialog policy module outputs is used to constrain the natural language generation module to generate the system response.The natural language generation module applies LSTM to model the decoder and uses attention mechanism to improve model performance.