Interpretable Information Retrieval

This is an interpretable information retrieval research project that implements and evaluates RAG (Retrieval-Augmented Generation) systems using three interpretability methods:

1. Gradient-based methods - Using gradients to identify important tokens/sentences
2. Perturbation-based methods - Measuring impact by removing or modifying text elements
3. Surrogate-based methods - Using SHAP values to explain retrieval scores

The project evaluates these methods using the MS MARCO and SCIREX datasets.

Environment Setup

IMPORTANT: PyTorch must be installed from source first (it will not work properly with conda).

After PyTorch is installed, create the conda environment:

conda env create -f environment.yml

This creates an environment named interpretability_project with Python 3.10.18.

Activate the environment:

conda activate interpretability_project

Methods

1. Gradient-Based Method

Coming soon

2. Perturbation-Based Method

The permutation_method folder conaints the implementation and evaluation of permutation-based importance scoring. For methodology and results, see:

• pertubation_method/README.md Overview of the formulation and motivation of permutation-based importance scoring
• pertubation_method/permutation_scores_test.py Implementation of the pertubation based method and analysis in multiple different settings on two different datasets

3. SHAP-Based Surrogate Method

The shap_method/ folder contains the complete implementation and evaluation of the SHAP-based surrogate method. For detailed methodology, results, and analysis, see:

• shap_method/README.md - Implementation documentation
• shap_method/SHAP_METHOD_REPORT.md - Evaluation report for SHAP method
• shap_method/notebooks/ - Jupyter notebooks with EDA and SHAP evaluation outputs

Note: Other methods will be added soon.

Dataset

• MS MARCO v2.1 (microsoft/ms_marco) for retrieval evaluation
• SCIREX for entity-level evaluation (used in SHAP method)