inversedMixup: Data Augmentation via Inverting Mixed Embeddings
Researchers have developed inversedMixup, a novel data augmentation technique for natural language processing that combines the controllability of traditional Mixup with the interpretability of LLM-generated text. This method reconstructs mixed embeddings into human-readable sentences, offering insights into the manifold intrusion phenomenon in text Mixup. Experiments show inversedMixup is effective in both few-shot and fully supervised learning scenarios. AI
IMPACT Introduces a novel technique for improving NLP model performance through interpretable data augmentation.