Research On Dialogue State Tracking Based On Open Vocabulary

A dialogue system is a machine-based system that simulates human interaction in natural language.It is typically categorized into three types: chatbot systems,question-answering systems,and task-oriented systems.Task-oriented systems are designed for specific goals and can accurately identify the user’s intentions and actions.As the central module in task-oriented dialogue systems,Dialog State Tracking(DST)is essential for achieving high-quality taskoriented dialogue systems.Therefore,research on DST is of great significance for building effective task-oriented dialogue systems.There are two types of task-oriented dialogue systems: end-to-end and pipeline.Due to its flexibility and strong interpretability,the pipeline model is widely used in many task-oriented dialogue systems.In this model,DST is often based on Natural Language Understanding(NLU),but this approach can cause errors in subsequent DST tasks.Therefore,many models combine NLU and DST using Multi-Task Learning(MTL),achieving excellent results.However,there are still some problems with the DST module,including the lack of utilization of domain-slotrelated information,the uneven distribution of dataset difficulty,and the problem of unseen slots.This paper addresses the above three problems by conducting a series of research after investigating the current state of DST-related research.The main contents are as follows:To address the problem of the lack of utilization of domain-slot-related information,a DST model based on Slot Information-Aware Extractor(SIAE)is proposed.The model uses SIAE to model the relevant information between domain-slot pairs,enabling the model to extract domain-slot-related information.To ensure the extraction ability of SIAE,a slot-aware supervision training task is designed for the SIAE module,which is introduced in the form of MTL,further improving the performance of SIAE.The experimental results on public datasets show that the DST model based on SIAE can effectively extract relevant information between slots and achieve better performance than existing models.To address the problem of the uneven distribution of data difficulty,the Learning Optimal Sample Weights(LOW)strategy is used to optimize the training process of the DST model.This paper designs a training method suitable for MTL to adjust the weight of each sample dynamically in the training of the slot classifier to improve the performance of the model.The experimental results show that the performance of the DST model is improved with the use of the LOW strategy,and the DST model based on SIAE achieves the best performance among all models.To address the problem of poor performance when facing unseen slot values,this paper explores the use of data augmentation methods to improve the robustness and generalization of the DST model in the context of the DST task.The effects of value replacement,generating controllable counterfactual dialogue algorithm,and adversarial training on the performance of the model are compared and analyzed.The experimental results show that adversarial training is not suitable for the DST task,while the generating controllable counterfactual dialogue algorithm can significantly improve the generalization ability of the model,to some extent,solve the problem of poor generalization ability when facing unseen slot values.