Research On Cross Domain Dialogue State Tracking

Human-machine dialogue is an important research topic in artificial intelligence.The major difficulty in the research and application of human-machine dialogue is to achieve consistent and coherent open-domain multi-turn dialogue.As a core module of open-domain multi-turn dialogue system,cross-domain dialogue state tracking recognizes the information in different domains in the dialogue,and records corresponding states in the dialogue history.It is helpful to support subsequent dialogue strategy decision.It directly influence on the overall performance of the dialogue system.Therefore,cross-domain dialogue state tracking has high research value.The existing modeling methods of dialogue state tracking can be majorly camped into three categories: ontology retrieval method which relies on a predefined ontology to search the candidates;text segment extraction method which extracts a fragment from the original text as the predicted value;and text generation method which uses a text generator to generate the target value word by word.However,normally the existing works only adopted one of these methods,but fails to make full use of the synergy and complementarity among them.In addition,in the cross-domain dialogue state tracking,the data imbalance problem among the distributions of domain,slot and value will lead to fluctuation during training,which may make the model get trapped in poor local optimum.To address the problem that existing works fails to make full use of the synergy and complementarity among various modeling methods,this thesis proposes a Coarse-to-Fine Refining Framework(C2F)for cross-domain dialogue state tracking,which uses the ontology retrieval method as a probe,and text generation method as a refinement,so as to combines the high performance of ontology retrieval method on high-frequency seen value and the born advantage of text generation method on unseen value.Experimental results on Multi WOZ dataset show that,compared with the baseline,the proposed model improves2.38% on Joint Accuracy with variance decreased by 1.17%;10.43% on Slot Accuracy with variance decreased by 0.11%;and 0.95% on Joint F1 with variance decreased by0.445%,respectively.Targeting to the data imbalance problem in cross-domain dialogue state tracking,the Loss-over-pre Loading Optimizer(Lo L)is proposed.In each training step,the simulated training of several steps is applied to estimate the look-ahead loss.It is incorporated to update the the current-step parameters,so as to reduce the influence of fluctuation during training and realize self-balanced optimization of the model in a certain degree.Experimental results on Multi WOZ dataset show that this method can improve the baseline steadily by 0.38% on Joint Accuracy;by 0.02% on Slot Accuracy;and by 0.13% on Joint F1,respectively.