Balanced Data Augmentation for Dialogue State Tracking
Xiangyi Kong · 2023
Data augmentation methods for dialogue state tracking (DST) have made progress on producing more data based on internal or external datasets. However, they fall short in the balancing the generated dataset. We propose a balanced data augmentation method for DST to improve the quality of generated dataset. Our approach has two steps: (1) balance method in selecting slots and values: modify the possibility of selecting each slot and value when adding or replacing slots; (2) balance method for under-generation and over-generation: generate or delete cases based on their overall frequency of slots. We apply our method to MultiWOZ and the dataset generated by our method is more balance than the original and COCO augmented data. Then we evaluate a strong DST model trained on our generated training set. The performance of this DST model improves 5.93% (from 55.04% to 60.97%) joint goal accuracy than original model, and also surpasses the performance of the model trained on the data generated by COCO. This demonstrates the advantages and potential of our approach to be incorporated into data augmentation for DST.