Rationale-Inspired Natural Language Explanations with Commonsense

Extractive rationales (i.e., subsets of input features) and natural language explanations (NLEs) are two predominant types of explanations for machine learning models. While NLEs can be more comprehensive than extractive rationales, machine-generated NLEs have been shown to fall short in terms of commonsense knowledge. In this paper, we show that commonsense knowledge can act as a bridge between extractive rationales and NLEs, rendering both types of explanations better. We introduce a self-rationalizing framework, called RExC, that (1) extracts rationales as most responsible features for the predictions, (2) expands the extractive rationales using commonsense resources, and (3) selects the best-suited commonsense knowledge to generate NLEs and give the final prediction. Our framework surpasses by a large margin the previous state-of-the-art in generating NLEs across five tasks in both natural language and vision-language understanding. Self-rationalization with commonsense also strongly improves the quality of the extractive rationale and task performances over the previous best performing models that also produce explanations.