A Multi-Domain Dialogue State Tracking Method with Prompt and Graph Attention Mechanism

This paper introduces an innovative Multi-Domain Dialogue State Tracking Method with Prompt and Graph Attention Mechanism (DST-PGAM). The focus of this research is on optimizing the prediction of shared values across domain slots by integrating representations from both dialogue and prompts to capture contextual information and semantic relationships. Additionally, the derived representations from the graph attention network are utilized to enhance the GPT-2 model, enabling causal sequence prediction of slot values and explicit modeling of relationships between cross-domain slots and values. Through this approach, precise prediction of domain slot values is achieved, improving the efficiency of the dialogue state tracking module and enabling more accurate analysis of user demands for providing convenient and efficient services. This research provides a new approach to improve the performance and user experience of multi-domain dialog systems.