EDA-NLP

Easy data augmentation techniques for boosting performance on text classification tasks.

We propose the following data augmentation techniques:

1. Synonym replacement (SR): Randomly choose n non-stop words from the sentence, and replace those words a randomly selected synonyms.
2. Random insertion (RI): Retrieve n words that are synonyms of any non-stop word in the sentence. Randomly insert those words into the sentence.
3. Random swap (RS): Randomly choose two words in the sentence and swap their positions. Do this n times.
4. Random deletion (RD): Randomly choose n words from the sentence and remove them.

We run a grid search over n = min(1, floor(alpha/l)), where alpha = {0.05, 0.1, 0.2, 0.3, 0.5}

Also run a grid search for s augmented sentences per technique per sentence for s = {1, 2, 3, 5, 10}

Usage

Data format

First place the training file in the format label\tsentence in datasets/dataset/train_orig.txt and the testing file in the same format in datasets/dataset/test.txt.

Word embeddings

Download GloVe word embeddings and place in a folder named word2vec.

The config file

Take a look at the hyperparameters in config.py before you begin. What increments do you want? Are the file names correct?

Augment the data and load the word2vec dictionary

python code/1_data_process.py

Train the model and evaluate it

python code/2_train_eval.py

EDA-7?:

5. Sliding window: slide a window of size w with stride s over the text input
6. Jittering: add c to n random word embeddings or take e^c for n random word embeddings
7. Random noising: add Gaussian noise to n random word embeddings