Investigating the Benefits of Free-Form Rationales
Jiao Sun, Swabha Swayamdipta, Jonathan May, Xuezhe Ma · 2022
Free-form rationales aim to aid model interpretability by supplying background knowledge that can help understand model decisions.Popular commonsense QA datasets such as CoS-E and ECQA provide crowdsourced freeform rationales for instances, but their utility remains under-investigated.We present studies which show that 88% of ECQA rationales indeed provide humans additional background information to understand a decision, while 93% 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 their effect on human users?We investigate the usefulness of rationales as an additional training signal, 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 a weaker training signal, but are also not helpful for humans in aiding model interpretability.* Work done prior to JM joining Amazon. 1 We use the terms "rationale" and "explanation" interchangeably.Please see Wiegreffe and Marasovic (2021) and Jacovi and Goldberg (2021) for more details on terminology.