Attention Modulation for Zero-Shot Cross-Domain Dialogue State Tracking

Mathilde Veron (1) , Guillaume Bernard (2) , Olivier Galibert (2) , Sophie Rosset (1)

1 ILES - Information, Langue Ecrite et Signée
2 LNE - Laboratoire National de Métrologie et d’Essais

Dialog state tracking (DST) is a core step for task-oriented dialogue systems aiming to track the user's current goal during a dialogue. Recently a special focus has been put on applying existing DST models to new domains, in other words performing zero-shot crossdomain transfer. While recent state-of-theart models leverage large pre-trained language models, no work has been made on understanding and improving the results of first-developed zero-shot models like SUMBT. In this paper, we thus propose to improve SUMBT zero-shot results on MultiWOZ by using attention modulation during inference. This method improves SUMBT zero-shot results significantly on two domains and does not worsen the initial performance with the significant advantage of needing no additional training.

Domaines: Informatique et langage [cs.CL]

Mathilde Veron, Guillaume Bernard, Olivier Galibert, Sophie Rosset. Attention Modulation for Zero-Shot Cross-Domain Dialogue State Tracking. 3rd Workshop on Computational Approaches to Discourse (CODI) at COLING 2022, Oct 2022, Gyeongju, South Korea. ⟨hal-03839919⟩