Geometric Analysis of Synthetic Data Augmentation for NLP

This research evaluates how the cosine distance of Llama 3.1-generated synthetic data from real RoBERTa embeddings affects discourse-pragmatic function classification. The study finds that the geometric relationship to real training data is as critical as data quantity.