Task-Oriented Dialogue State Tracking And Application

Dialogue state tracking is a crucial component of the task-oriented multi-round dialogue.It aims to evaluate the user goals and requests in each round according to the history context of dialogue,and provide decision-making basis for the dialogue decision.The traditional dialogue state tracking methods usually regard the result of spoken understanding as the input,however,there is a problem of error propagation.In addition,with the number of rounds between users and t he system increases,the traditional methods are difficult to achieve high accuracy in identifying the new or rare slot values.Targeting these issues,we introduce a neural network-based state tracking method and study the effectiveness of the dialogue st ate tracking module in the knowledge-driven dialogue system.In this paper,we propose a model to solve the problem of identifying new slot values and rare slot values in multi-round dialogue system,which has capacity to capture the local slot features.The proposed model takes the local slot information as features,then learns the representation of the current input of different slots,the previous round of system actions and the candidate slot values by attention mechanism and LSTM network.The learned representation is used to generate the next dialogue state.We evaluate the proposed model on two public datasets.The experimental results show that this model achieves better performance compared to the baselines on both datasets.The proposed model is simple but effective,can learn the effective features of different slots information,which is helpful for identifying rare slot values and new slot values.Aiming at the problem of modeling all candidate slot values in the LSDST model,we propose a model which based on the hierarchical neural network.The proposed model try to model the sequence relations between the dialogue states through a LSTM-based hierarchical neural network.The bottom module utilizes different methods to encode the global information in different dialogue rounds.The top module takes the global information of each round as input features,and perform LSTM to capture the sequence features of the history multi-round information to achieve the purpose of dialogue state tracking.We also evaluate the New State model on the two public datasets,the experimental results show that New State achieves better performance in comparison with last model in identifying the normal slot values,which is more practical.In addition,in order to explore the effectiveness of the dialogue state tracking model in the dialogue system.We further design a framework to integrate the dialogue state tracking into the knowledge-driven dialogue system,to evaluate whether the dialogue state tracking module can improve the quality of dialogue generation.The proposed framework adopts the results of state tracking to assign weights to the entries in the knowledge base by attention mechanism,and combines the traditional Seq2 Seq model to generate the final response.The experimental results show that the proposed framework which combined with the above dialogue state tracking module can generate smoother and more accurate system response.