[Paper Review] EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks

Jason Wei, Kai Zou|arXiv (Cornell University)|Jan 31, 2019

Topic ModelingComputer Science33 references197 citations

TL;DR

The paper introduces EDA, a simple set of four text augmentation operations (synonym replacement, random insertion, random swap, random deletion) that improve text classification performance, especially on small datasets, without requiring external models.

ABSTRACT

We present EDA: easy data augmentation techniques for boosting performance on text classification tasks. EDA consists of four simple but powerful operations: synonym replacement, random insertion, random swap, and random deletion. On five text classification tasks, we show that EDA improves performance for both convolutional and recurrent neural networks. EDA demonstrates particularly strong results for smaller datasets; on average, across five datasets, training with EDA while using only 50% of the available training set achieved the same accuracy as normal training with all available data. We also performed extensive ablation studies and suggest parameters for practical use.

Motivation & Objective

• Motivate the need for simple, universal data augmentation in NLP.
• Propose four lightweight augmentation operations that preserve label integrity.
• Evaluate EDA across multiple datasets and model architectures to assess robustness and utility.
• Identify practical parameter settings and guidelines for using EDA in real tasks.

Proposed method

• Define four augmentation operations: synonym replacement, random insertion, random swap, and random deletion.
• Change amount n of edits proportional to sentence length with parameter alpha; set p for deletion equal to alpha.
• Generate n_aug augmented sentences per original sentence.
• Evaluate on five benchmark datasets using CNN and RNN architectures with multiple seeds.
• Perform ablation studies to assess contribution of each operation.
• Provide practical usage recommendations based on results.

Experimental results

Research questions

• RQ1 Do simple text editing augmentations improve text classification performance across multiple datasets and models?
• RQ2 Which of the four augmentation operations contribute most to performance gains and under what conditions?
• RQ3 How do augmentation level and number of augmented samples affect performance, especially on small vs large datasets?
• RQ4 Do augmented sentences largely preserve original labels as evidenced by latent-space analyses?

Key findings

• EDA improves accuracy across five datasets for both CNN and RNN models.
• Average improvement is 0.8% on full datasets and 3.0% on the smallest subset (N_train=500).
• All four operations contribute to gains; using alpha around 0.1 provides a sweet spot.
• More augmented samples (n_aug) help more on small datasets but offer diminishing returns on larger datasets.
• Augmented sentences largely preserve their labels, as shown by latent-space clustering in a t-SNE visualization.
• EDA is most beneficial when training data are limited, enabling comparable performance with half the data.