GENERATION OF NATURAL LANGUAGE EXPLANATIONS FOR AMCC DERIVED COUNTERFACTUALS

Abstract

This thesis explores 2 different approaches to generate simple natural language sentences for counterfactual explanations to better help user understand a machine learning model’s decision. The objective is to determine a domain independent natural language generation approach to generate effective and simple sentences to explain counterfactual explanations. Counterfactual explanations generated from the achievable minimally-contrastive counterfactual explanations (AMCC) algorithm is used to conduct experiments for the two approaches, a template-based approach, enhanced with WordNet synonyms and lexical substitution using a BERT model, and a template-guided approach using a fine-tuned T5 model. Evaluations are performed using an automatic metric, BLEURT, to identify the effectiveness of each approach. To test the effectiveness of domain independence of the proposed approaches, experiments were conducted on a dataset from a new domain to see the effectiveness of generated natural language explanations.