Lightweight Dialog State Tracking Methods Based on RoBERTa for Resource Constrained Dialog Systems

Dialog state tracking is a crucial component of task-oriented dialog systems. Pre-trained language models can effectively perform dialog state tracking for task-oriented dialog systems through model fine-tuning. However, it is difficult to apply most methods based on pre-trained language models to resource-constrained systems. To address this challenge, we first propose a model based on RoBERTa, which can perform dialog state tracking tasks more effectively for task-oriented dialog systems. Additionally, we propose two methods to enable our proposed model to be applied to resource-constrained dialog systems. Experimental results on two public datasets show that our proposed model based on RoBERTa can improve the joint goal accuracy of dialog state tracking, and our proposed methods can effectively implement dialog state tracking with less storage space and computing resources.