antonis19 / nlp-robustness-evaluation

Robustness Evaluation for NLP

Download Data

Run sh data/download_data.sh to download the data needed for the experiments. This will download the preprocessed IMDB dataset, the word embeddings for the neural network and for the adversarial attack, and dictionaries of cached nearest neighbors.

Download StanfordCoreNLP

Run sh download_stanfordcorenlp.sh to download the Standford CoreNLP parser.

Train LSTM model

Run python train_lstm_model.py to train an LSTM model for sentiment classification on the IMDB dataset.

Perform verification using DeepGo

Run python verification.py to perform robustness analysis using DeepGo on the LSTM.

Obtain explanations for sentiment analysis.

Run the code in sbe_examples.ipynb to generate explanations for sentiment analysis, using the adaptation of Spectrum-Based Explanations to text classification.

Generate adversarial examples

Run the code in attack.ipynb to generate adversarial examples for sentiment analysis. To visualize the generated adversarial examples run the code in visualize_attack.ipynb. Some examples of adversarial examples are already shown in visualize_attack.ipynb.

Fix the classification

Run the code in fix_classification.ipynb to generate suggestions that change the classification of a text to the correct class. To visualize the suggestions run the code in visualize_fixing.ipynb.