interpret.md

XAI Interpretability instructions

Main README

All commands should be used from root directory.

Retrieve stats for visualisation

Stats csv files are saved in stats-results/ folder, run this command before running the notebooks:

### Example on our best BasicLSTM trained model
python -m src.evaluation.test_save_stats --model BasicLSTM --saved_model_path saved-models/BasicLSTM_2021-12-08_01-04-25_trained_testAcc=0.7107.pth --loss_criterion bcelosswithlogits --only_test 0 --stats_label 1

Word importance visualisations with Captum:

We provide two notebooks to visualize which parts of the input sentence are used for an inference of a trained model.

In the current state, we use Integrated Gradients from Captum library to obtain the attribution scores for each word in a given sentence.

• For CNN/RNN-based models, please use this XAI LSTM notebook (Example on our best BasicLSTM trained model).

• For BERT-based models, please use this XAI Bert notebook (Example on our best DistillBert trained model).

Details on XAI Bert notebook

Here is the Confusion Matrix of our best trained model DistillBert_2021-12-08_16-39-08_trained_testAcc=0.7960.pth used in the XAI Bert notebook:

True Positive

False Positive

True Negative

False Negative