Generates syntactically diverse, label-preserving text with lexicalized PCFGs for few-shot classification, and pairs this with augmentation-aware train/validation splitting for stronger low-resource training.
Overview
Few-shot text classification needs more than copies of the same sentences. ALP uses lexicalized PCFGs to generate syntactically diverse, label-preserving text for data augmentation.
The method is paired with an augmentation-aware train/validation split so models benefit from the extra data without leaking inflated validation scores.
Why it matters
In low-resource settings, synthetic diversity can matter as much as volume. ALP focuses on structure-preserving augmentation that still changes surface form enough to help generalization.
Paper
The full paper is embedded below. You can also download the PDF or view it on the AAAI proceedings site.