SoftEDA: Rethinking Rule-Based Data Augmentation with Soft Labels

TL;DR

The paper tackles the brittleness of rule-based text data augmentation by mitigating label noise through soft labeling. It introduces softEDA, which applies label smoothing to augmented samples while keeping EDA's perturbation operations, using the soft label $\hat{\bm{y}} = (1-\alpha)\bm{y} + \frac{\alpha}{N_{Class}}$. Across seven text classification tasks with CNN, LSTM, and BERT, softEDA consistently improves accuracy over EDA and AEDA, with strong performance on CoLA. The work offers a simple, reproducible technique to boost rule-based augmentation and suggests extension to other augmentation strategies.

Abstract

Rule-based text data augmentation is widely used for NLP tasks due to its simplicity. However, this method can potentially damage the original meaning of the text, ultimately hurting the performance of the model. To overcome this limitation, we propose a straightforward technique for applying soft labels to augmented data. We conducted experiments across seven different classification tasks and empirically demonstrated the effectiveness of our proposed approach. We have publicly opened our source code for reproducibility.