Learning towards Selective Data Augmentation for Dialogue Generation

As it is cumbersome and expensive to acquire a huge amount of data for training neural dialog models, augmentation proposed effectively utilize existing samples. However, current techniques on the generation task mostly augment all cases in dataset without considering intrinsic attributes between different cases. We argue that not are beneficial task, suitable should obey following two attributes: (1) low-quality (the model cannot generate high-quality response case), (2) representative case represent property whole dataset). Herein, we explore this idea by proposing Selective Data Augmentation framework (SDA) task. SDA employs dual adversarial network select lowest quality most points one stage. Extensive experiments conducted publicly available datasets, i.e., DailyDialog OpenSubtitles, show our can improve performance with respect various metrics