BERT-Neural-Language-Interface-Explainability-Explorer

Author: Manai Mortadha

---

Description

Welcome to the Neural Language Interface (NLI) Explain project! This repository is dedicated to exploring and explaining the decision-making process of BERT models in the context of Natural Language Inference (NLI) tasks. We employ Feature Interaction methods to shed light on why BERT makes specific predictions in NLI.

Explainers Included

We offer several feature interaction explainers to understand BERT's decisions in NLI tasks:

1. IH - Integrated Hessians: Utilizes the integrated Hessians method for explanation.

2. Archipelago: Incorporates the Archipelago explainer for decision interpretation.

3. X-Archipelago (Ours): Offers a cross-sentence version specially designed for NLI examples with two sentences, building upon the Archipelago explainer.

4. MAFIA (Ours) - MAsk-based Feature Interaction Attribution: Introduces our own MAFIA explainer for enhanced feature interaction analysis.

Requirements

Ensure you have the following Python libraries installed:

nltk
pandas
numpy
torch
transformers

How to Use

Follow these steps to effectively use the repository and explore BERT's decisions in NLI tasks:

1. Download the e-SNLI Dataset:

   • Visit the e-SNLI dataset repository.
   • Download the dataset files and place them in the data/ folder of this repository.

2. Preprocess the Data:

   • Run the python prepare_data.py script to preprocess the dataset.

3. Generate Explanations:

   • For each explainer (lime, IH, Arch, Mask), navigate to the scripts/ folder.

   • Execute the respective script to generate explanations. These explanations will be stored as .json files in the explanations/ directory (created if not already).

   • For in-depth details on explanation generation, refer to the explainers/save_explanations.py file.

   • To explore the implementation of each explainer, navigate to the explainers/ directory.

4. Evaluation:

   • To evaluate the explanations, you can use the scripts/eval_explanation.sh script.
   • Modify the script according to your specific evaluation needs and preferences.

Feel free to explore, experiment, and enhance the explanations of BERT's decisions in NLI tasks using this repository.