WhatIf: Leveraging Word Vectors for Small-Scale Data Augmentation

AbstractWe introduce WhatIf, a lightly supervised data augmentation technique that leverages word vectors to enhance training data for small-scale language models. Inspired by reading prediction strategies used in education, WhatIf creates new samples by substituting semantically similar words in the training data. We evaluate WhatIf on multiple datasets, demonstrating small but consistent improvements in downstream evaluation compared to baseline models. Finally, we compare WhatIf to other small-scale data augmentation techniques and find that it provides comparable quantitative results at a potential tradeoff to qualitative evaluation.