GALAXY: A Generative Pre-trained Model for Task-Oriented Dialog with Semi-Supervised Learning and Explicit Policy Injection

TL;DR

GALAXY is a pre-trained dialog model that explicitly learns dialog policy using semi-supervised learning, improving task-oriented dialog performance and few-shot capabilities on benchmark datasets.

Contribution

It introduces a novel semi-supervised pre-training approach with explicit policy learning and regularization, achieving state-of-the-art results in task-oriented dialog systems.

Findings

Significantly improves performance on benchmark datasets

Achieves new state-of-the-art scores

Demonstrates strong few-shot learning ability

Abstract

Pre-trained models have proved to be powerful in enhancing task-oriented dialog systems. However, current pre-training methods mainly focus on enhancing dialog understanding and generation tasks while neglecting the exploitation of dialog policy. In this paper, we propose GALAXY, a novel pre-trained dialog model that explicitly learns dialog policy from limited labeled dialogs and large-scale unlabeled dialog corpora via semi-supervised learning. Specifically, we introduce a dialog act prediction task for policy optimization during pre-training and employ a consistency regularization term to refine the learned representation with the help of unlabeled dialogs. We also implement a gating mechanism to weigh suitable unlabeled dialog samples. Empirical results show that GALAXY substantially improves the performance of task-oriented dialog systems, and achieves new state-of-the-art results on benchmark datasets: In-Car, MultiWOZ2.0 and MultiWOZ2.1, improving their end-to-end combined scores by 2.5, 5.3 and 5.5 points, respectively. We also show that GALAXY has a stronger few-shot ability than existing models under various low-resource settings.