Are Shortest Rationales the Best Explanations For Human Understanding?

Anonymous

Abstract: Existing self-explaining models typically favor extracting the shortest rationales possible (“shortest yet coherent subset of input to predict the same label”), with the assumption that short rationales are more intuitive to humans, even though short rationales lead to lower accuracy. However, there is a lack of human studies on validating the effect of rationale length on human understanding. Is the shortest rationale indeed the most understandable for humans? To answer this question, we design a self-explaining model that can take control on rationale length. Our model incorporates contextual information and supports flexibly extracting rationales at any target length. Through quantitative evaluation on model performance, we further verify that our method LIMITEDINK outperforms existing self-explaining baselines on both end-task prediction and human-annotated rationale agreement. We use it to generate rationales at 5 length levels, and conduct user studies to understand how much rationale would be sufficient for humans to confidently make predictions. We show that while most prior work extracts 10%-30% of the text to be the rationale, human accuracy tends to stabilize after seeing 40% of the full text. Our result suggests the need for more careful design of the best human rationales.