Investigating the Benefits of Free-Form Rationales

TL;DR

This paper explores how free-form rationales can enhance model interpretability and performance in commonsense question answering, revealing that high-quality rationales significantly boost model accuracy when used as supervision during training.

Contribution

It demonstrates that incorporating a small percentage of high-quality rationales during training substantially improves model performance and highlights the importance of rationale quality for interpretability.

Findings

Incorporating 5% rationales boosts model performance by over 47% for CoS-E.

High-quality human rationales improve model interpretability and accuracy.

Generated rationales are less helpful than crowdsourced ones for both models and humans.

Abstract

Free-form rationales aim to aid model interpretability by supplying the background knowledge that can help understand model decisions. Crowdsourced rationales are provided for commonsense QA instances in popular datasets such as CoS-E and ECQA, but their utility remains under-investigated. We present human studies which show that ECQA rationales indeed provide additional background information to understand a decision, while over 88% of CoS-E rationales do not. Inspired by this finding, we ask: can the additional context provided by free-form rationales benefit models, similar to human users? We investigate the utility of rationales as an additional source of supervision, by varying the quantity and quality of rationales during training. After controlling for instances where rationales leak the correct answer while not providing additional background knowledge, we find that incorporating only 5% of rationales during training can boost model performance by 47.22% for CoS-E and 57.14% for ECQA during inference. Moreover, we also show that rationale quality matters: compared to crowdsourced rationales, T5-generated rationales provide not only weaker supervision to models, but are also not helpful for humans in aiding model interpretability.