What does a Text Classifier Learn about Morality? An Explainable Method for Cross-Domain Comparison of Moral Rhetoric - code

Code for the paper "What does a Text Classifier Learn about Morality? An Explainable Method for Cross-Domain Comparison of Moral Rhetoric", published at ACL '23. This code implements Tomea, an Explainable AI method for investigating the difference in how language models represent morality across domains. Given a pair of datasets and models trained on the datasets, Tomea generates 10 m-distances and one d-distance to measure the difference between the datasets, based on the SHAP method. We make pairwise comparisons of seven models trained on the MFTC datasets (available at this DOI: 10.4121/646b20e3-e24f-452d-938a-bcb6ce30913c).