Prompt-based Dialogue State Tracking Method Jointly Modeled with Natural Language Understanding

Anonymous

Abstract: Cross-domain dialogue state tracking has become a hot topic in recent years, it profoundly influences the generalizability of task-oriented dialogue systems. In this paper, we propose a prompt-based dialogue state tracking method jointly modeled with natural language understanding (PLDT) to address the problem of multi-domain adaptation in the state tracking task and optimize the existing models. We introduce the joint modeling method to reduce the cumulative errors between DST and NLU in pipeline dialogue system. Based on this, in analyzing current dialogue state tracking methods, we combine T5 with Ptr-Net in a proper way to solve both the redundancy and inaccuracy shortcomings in generative methods and the out-of-vocabulary (OOV) problem in pointer network methods, respectively. We also design a continuous prompt learning approach that uses a few discrete samples (labeled by a keyword extraction algorithm in an automatic way) to train the model in an unsupervised way and generate a suitable prompt. Our model outperforms other existing approaches on MultiWOZ2.0 and CrossWOZ in both slot and joint accuracy and has better performance in zero-shot tasks than other cross-domain models.

Paper Type: long

Research Area: Dialogue and Interactive Systems

Languages Studied: English, Chinese