MT-DST: Multi-Task Strategy for Few-shot Dialog State Tracking

The remarkable effectiveness of large pre-trained models in Task-Oriented Dialogue (TOD) systems, particularly in the area of Dialog State Tracking (DST), has been well documented, even in scenarios with limited labeled data. In the past, researchers mainly focused on two approaches: generating all belief states simultaneously or generating values sequentially by prompting for each individual slot. The first technique heavily relies on the model’s comprehension ability, which often leads to inferior results, while the latter requires going through every possible slot, thus compromising processing speed. To address these challenges, we propose a new multi-task strategy for few-shot Dialog State Tracking (MT-DST) to achieve favorable outcomes while maintaining a fast pace. This method utilizes the former approach as the main task, with the secondary task serving as an auxiliary job for incorporating additional data. Our model has demonstrated state-of-the-art performance in a few-shot DST tasks through experiments on the widely used MultiWOZ 2.1 benchmark.