FADA: Crafting Feature-Aware Data Augmentation Policies for Enhanced Text Classification

Anonymous

TL;DR: Leveraging AMR features and text quality metrics to craft targeted augmentation policies for individual text instances

Abstract: This paper introduces FADA, a novel data augmentation technique that creates feature-aware data augmentation policies. Unlike traditional dataset-level approaches, FADA utilizes a text's abstract meaning representation to extract high-level concepts, enabling targeted transformations for specific features. It evaluates transformation effectiveness through cheaply computed quality metrics like label alignment, fluency, and grammaticality. Our evaluations on four benchmark datasets show that our learned augmentation policies attain strong performance against baseline techniques and transfer surprisingly well to new domains.

Paper Type: short

Research Area: NLP Applications

Contribution Types: NLP engineering experiment, Approaches to low-resource settings

Languages Studied: English