Task-Oriented Dialogue Augmentation while Changing Flow for Dialogue State Tracking and Response Generation

Shubham Garg

Abstract: Due to the lack of high-quality datasets for training task-oriented dialogue systems, data augmentation has been widely used to increase the size and diversity of the training dataset. However, augmentation of such dialogues has traditionally followed simple methods in paraphrasing utterances or combining existing dialogues. Consequently, existing methods produce dialogues with flows similar to the source data, which is suboptimal for data augmentation. In this work, we introduce DiaFlow: an augmentation framework to construct dialogues while changing the dialogue flow. Exploiting the complete text generation capabilities of large language models, we facilitate the generation of dialogues with a wide range of flows and further obtain dialogue act annotations. Our experimental results show significant improvements on dialogue state tracking and response generation tasks.