Self-Training with Purpose Preserving Augmentation Improves Few-shot Generative Dialogue State Tracking

In dialogue state tracking (DST), labeling the dataset involves considerable human labor. We propose a new self-training framework for fewshot generative DST that utilize unlabeled data. Our self-training method iteratively improves the model by pseudo labeling and employs Purpose Preserving augmentation (PPaug) to prevent overfitting. We increase the few-shot (10%) performance by approximately 4% on Figure 1: Dialogue example of DST dataset and its belief MultiWOZ 2.1 (Eric et al., 2019) and enhances state. The underlined part of the dialogue is the the slot-recall 8.34% for unseen values compared value of the belief state and has specific information to baseline.