CLSep 2, 2022

Random Text Perturbations Work, but not Always

arXiv:2209.00797v2297 citationsh-index: 4
Originality Synthesis-oriented
AI Analysis

This work addresses the problem of data augmentation effectiveness for NLP practitioners, showing that random text perturbations are not universally beneficial and depend on task-specific conditions, making it an incremental contribution.

The study evaluated random text perturbations for data augmentation in binary text matching across Chinese and English, finding that their effect on test performance of neural classifiers varies from negative to positive based on the amount of original training data, regardless of how perturbations are applied.

We present three large-scale experiments on binary text matching classification task both in Chinese and English to evaluate the effectiveness and generalizability of random text perturbations as a data augmentation approach for NLP. It is found that the augmentation can bring both negative and positive effects to the test set performance of three neural classification models, depending on whether the models train on enough original training examples. This remains true no matter whether five random text editing operations, used to augment text, are applied together or separately. Our study demonstrates with strong implication that the effectiveness of random text perturbations is task specific and not generally positive.

Code Implementations1 repo
Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes