Domain-Aligned Data Augmentation for Low-Resource and Imbalanced Text Classification

Data Augmentation approaches often use Language Models, pretrained on large quantities of unlabeled generic data, to conditionally generate examples. However, the generated data can be of subpar quality and struggle to maintain the same characteristics as the original dataset. To this end, we propose a Data Augmentation method for low-resource and imbalanced datasets, by aligning Language Models to in-domain data prior to generating synthetic examples. In particular, we propose the alignment of existing generic models in task-specific unlabeled data, in order to create better synthetic examples and boost performance in Text Classification tasks. We evaluate our approach on three diverse and well-known Language Models, four datasets, and two settings (i.e. imbalance and low-resource) in which Data Augmentation is usually deployed, and study the correlation between the amount of data required for alignment, model size, and its effects in downstream in-domain and out-of-domain tasks. Our results showcase that in-domain alignment helps create better examples and increase the performance in Text Classification. Furthermore, we find a positive connection between the number of training parameters in Language Models, the volume of fine-tuning data, and their effects in downstream tasks.This preprint has not undergone peer review or any post-submission improvements or corrections. The Version of Record of this contribution is published in Advances in Information Retrieval. ECIR 2023. Lecture Notes in Computer Science, vol 13981, and is available online at https://doi.org/10.1007/978-3-031-28238-6_1