About

The repository showcases the use of model agnostic prediction explainers (SHAP and LIME). They allow to get more insights into the model decision-making. Each token is being analysed - how it influences the final output. Custom terminal insight display has been designed, to be able to use outside jupyter notebook.

Usage

To try out the explainer one needs to provide a path to a model compatible with HerbertSentiment class. Then run explain_file.py or explain_single.py file. Provide --model_path argument and choose explainer with --explainer. When running explain_file.py one can also specify --dataset which will be used as a file for prediction (should be txt, one example per line).

Sentiment colors:

• red for negative
• yellow for neutral
• green for positive

Sample output

First each token is shown, along with its sentiment contribution and class it is contributing to:

Then summary information is displayed, to easily analyse aggregated output. Overall sentiment probability distrbution, sentence colored token-wise and then the sentence in dominating sentiment color:

XAI Blog post

I have written a blog post describing the methods in detail and providing examples with a deeper dive into the explanation of explanation.

Link to the article: https://voicelab.ai/explainable-artificial-intelligence-xai-in-sentiment-analysis