nn-interpretability

This repository contains the code to a university project on interpretability measures in text classification. It investigates and compares interpretability measures for Support Vector Machine and transformer-based (DistilBERT) models for text classification on the 20-newsgroups corpus.

Usage

Dataset

• data set preprocessing (excluding headers, footers, quotes and truncating) is done in dataset/dataset.py

SVM

• tune model and vectorizer parameters: svm-parameter-tuning.py
• train/test model and generate coefficient outputs: svm.py

DistilBERT

You can either fine-tune the model yourself, or download our finetuned model and compute attributions.

• fine-tune the model yourself: models/model.py
• download the fine-tuned model from Google Drive and save it to the models/ directory; you may need to adjust the file path in captum-explain.py to run the attribution computations with the captum pipeline in explainer.py
• analyse the attributions in outputs/distilbert/

Analyses

Analyses are conducted in 3 different Notebooks:

• analysis-distilbert.ipynb contains the analyses of the DistilBERT attributions
• analysis-predictions.ipynb compares the scores for the test set instances (SVM coefficients in outputs/coefs_test.csv and DistilBERT attributions in outputs/distilbert_attributions.csv), creates some visualizations of specific instances in outputs/viz
• analysis-vocabs.ipynb conpares the scores over the general vocabulary (SVM coefficients in outputs/vocab_coef_svm.csv and DistilBERT attributions in outputs/vocab_attr_dist4_gold.csv and outputs/vocab_attr_dist4_pred.csv

Authors

Lydia Körber and Lisanne Rüh