Research On Dialogue State Tracking Methods Based On Interactions Between State Elements

Dialogue state tracking(DST)is a core component of task-oriented dialogue system,which estimates the belief distribution of user goal in each dialogue turn.It is found that potential correlations are common between basic elements of the dialogue state,namely domains,slots or values.And the correlations may contain useful information for DST task.Current DST methods tend to predict the belief state of each dialogue element independently,but fail to model the potential correlations between the state elements,which may limit their prediction accuracy.Hence,the performance of DST methods and the dialogue systems can be improved if incorperating the interaction between state elements to take advantage of the latent features.In this research,we focus on the multi-domain scenario and propose to model multiple semantic interactions between the state elements by utilizing the semantic information contained by the elements and the dialogue context.The proposed interaction mechanisms are incorperated to DST baselines,and the prediction accuracy of these methods are improved.The main contributions of this thesis are as follows:Firstly,we reproduce representative DST methods to maximize the improvement our proposed mechanisms bring to the DST task.In both fix-vocabulary and open-vocabulary catagory,we implement neural belief tracker and non-autoregressive dialogue state tracker,which provide basic platforms for the incorperation of the proposed interaction mechanisms.Secondly,to address the limitation in performance caused by estimating belief state of each state element independently,a multiple interaction mechanism for joint modeling of state output.We manage to model the interactions between feature vectors of different state elements before output the belief state ultilizing the message passing mechanism of the recurrent relational network.The improvement of performance the proposed mechanism brings to the baseline model is verified by the experiment.Thirdly,a fine-grained interaction mechanism for the generation of structure information is proposed.We apply multi-head attention and graph attention network respectively to implement the fine-grained interaction between state elements and the dialogue utterances at token level,and the high-level interation between the elements after incorporating information from the dialogue context.In the experiment,we compare our proposed method with the baseline models.The result shows the positive effect our proposed method brings to the baseline model.