Mosaic Augmentation for Text: Cropping and Collaging as Cross-Domain Techniques

Abstract

We present new visually inspired cropping and collaging data augmentations for text. We test how these augmentations impact data-scarce scenarios over multiple NLP tasks: name entity recognition, extractive question answering and abstractive summarization, across 9 prominent datasets. Ablation studies show different prevailing reasons for the augmentations' effectiveness for the different tasks, but all benefit from our approach. We achieve significant improvements over baselines, particularly for limited data use cases.

Tasks

Reproductions