Research On Dialogue State Tracking Algorithm Based On Open Ontolog

The dialogue state tracking(DST)module is an important component of the pipeline system for tracking user’s intention information.Traditional ontology-based DST methods,which rely on pre-defined slot value candidates,have significant limitations in practical applications.Ontology-free dialogue state tracking methods,which do not require manual definition of dialogue ontologies,have superior extensibility and can handle slot values that have not yet been explicitly defined.They have become a hot research area in recent years.In this article,three highly related DST methods based on the challenges and difficulties faced by ontology-free DST were designed.The specific work is as follows:(1)Aiming at the problem that the existing ontology-free DST models are difficult to obtain the slot representation related to the slot,This paper proposes a Dialogue State Tracking model based on stacked Self-attention layers and Copy mechanism(SC-DST).The model sends all the slot representations together with the context representation into the stacked self-attention layer to obtain better slot and context-related slot representations,and introduces the copy mechanism to decthe slot that needs to be updated,getting rid of the dependence on the slot value candidate list.Experimental results show that SC-DST has better performance on both noisy(Multi WOZ 2.1)and clean(Multi WOZ 2.4)datasets.(2)Aiming at the problem of error transmission caused by using historical dialogue state based on the existing open ontology model,This paper proposes a Dialogue state tracking model based on dialogue level state and historical dialogue state deletion deletion,DLDS-DST).The model uses the dialogue level state as the prediction target,and randomly removes the dialogue state of the previous round with a certain probability during model training,forcing the model to learn to correct errors in the incompletely trusted dialogue state information.Experimental results on noisy(Multi WOZ 2.1)and clean(Multi WOZ 2.4)datasets have shown that this method can effectively alleviate errors in transmission,achieving good performance with a joint accuracy(Multi WOZ 2.4)of 70.95%.Aiming at the problem that existing ontology-free dialogue state tracking models perform suboptimal on noisy training data,In this paper,we propose Label noise Robust dialogue state tracking with dialogue state deletion(ASSIST-SD).Based on the ASSIST learning framework,the historical dialogue state deletion mechanism(SD)was introduced into the STAR model,which alleviated the overfitting phenomenon of the STAR model in the case of a small number of training samples,and effectively improved the few-shot learning ability and generalization ability of the STAR model.Experiments show that.The quality of pseudo-labels generated by using the STAR-SD model as an auxiliary model is better,which effectively improves the performance of the model trained using the robust framework.