Research On Dialogue State Tracking In Multi Domain

Dialogue state tracking(DST),a core module in task-oriented dialogue systems,has received extensive attention from researchers.Real task-oriented dialogues often involve multiple domains at the same time,therefore,the dialogue state tracking in multi-domain has become the focus of research.The multi-turn mapping in multi-domain dialogues makes the prediction of dialogue state need to refer to the dialogue history,so the effective modeling of context is significantly important for multi-domain dialogue state tracking.Traditional methods concatenate the utterances of context into a sequence as input and encode it with a sequential model,which may lead to the earlier key information loss and the interference from redundant turns as the dialogue gets longer.Some recent work tries to use the previous dialogue state together with the utterances as the input to the model.However,it is faced with the problem of error accumulation of predicted dialogue states and effective fusion of above two types of dialogue context input.From the above two angles,this thesis aims to explore the effective methods to model the context.The main contributions are summarized as follows:(1)From the perspective of regarding the concatenated utterances as input,we propose a novel method to model the contribution-aware context hierarchically.At first,a hierarchical encoder is introduced to model each dialogue turn.In this way,the information loss will be alleviated by reducing the length of sequence sent to each encoder.Then we propose a contribution perceptron which is used to get the context representation including the contribution information of different turns on target slot,so as to help prevent the decoder from the interference of lengthy turns.The experimental results show that this model has better ability to model the multi-domain dialogues with longer context.(2)From the perspective of fusing two types of dialogue context input(previous dialogue state and context utterances),we propose a method for multi-type context information fusion.Firstly,we introduce a new module,called multi-type context information fusion network,to achieve the alignment fusion of the relevant information corresponding to the target slot in the two types of dialogue context input containing multiple slot value information.Besides,a slot value state identification module is introduced as an auxiliary task jointly trained with DST so as to solve the error accumulation from the previous dialogue state during information fusion by improving the ability of the model to distinguish the wrong slot values.The experimental results show the proposed model achieves 55.51% joint accuracy in MultiWOZ 2.0,which illustrates the effectiveness of this model.To sum up,this thesis focuses on the multi-domain dialogue state tracking,and explores the methods to model the dialogue context from different perspectives.It has a great reference value to the relevant researches in multi-domain dialogue state tracking,and has certain value for theoretical studies and practical applications.