eda_nlp

Why we included this project

Hand-labeling a text classification dataset is expensive, and when you only have a few hundred examples, every extra training sentence counts. This reference implementation of EDA, the data augmentation method from EMNLP 2019, stretches what you have by generating paraphrase-like variants of your existing sentences using four simple editing operations: swapping words for synonyms, inserting random synonyms, swapping word positions, and deleting a few words at random. No external language model is trained, only WordNet synonyms and NLTK are needed, so you can plug it into a current pipeline in minutes and rerun your classifier to see the difference. The paper behind it reported the biggest gains on training sets under 500 examples, which is exactly where scarce hand-labeled data makes augmentation worthwhile. If you want a lightweight, reproducible way to expand a corpus before trying heavier synthetic-data tools, this is a solid place to start.