Are Hard Examples also Harder to Explain? A Study with Human and Model-Generated Explanations

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 2121 - 2131 December 7-11, 2022 ©2022 Association for Computational Linguistics Are Hard Examples also Harder to Explain? A Study with Human and Model-Generated Explanations Swarnadeep Saha1 Peter Hase1 Nazneen Rajani2 Mohit Bansal1 1UNC Chapel Hill 2Hugging Face {swarna, peter, mbansal}@cs.unc.edu nazneen@huggingface.co Abstract Recent work on explainable NLP has shown that few-shot prompting can enable large pre- trained language models (LLMs) to generate grammatical and factual natural language ex- planations for data labels. In this work, we study the connection between explainability and sample hardness by investigating the fol- lowing research question – “Are LLMs and hu- mans equally good at explaining data labels for both easy and hard samples?” We an- swer this question by first collecting human- written explanations in the form of generaliz- able commonsense rules on the task of Wino- grad Schema Challenge (Winogrande dataset). We compare these explanations with those gen- erated by GPT-3 while varying the hardness of the test samples as well as the in-context samples. We observe that (1) GPT-3 expla- nations are as grammatical as human expla- nations regardless of the hardness of the test samples, (2) for easy examples, GPT-3 gener- ates highly supportive explanations but human explanations are more generalizable, and (3) for hard examples, human explanations are sig- nificantly better than GPT-3 explanations both in terms of label-supportiveness and generaliz- ability judgements. We also find that hardness of the in-context examples impacts the quality of GPT-3 explanations. Finally, we show that the supportiveness and generalizability aspects of human explanations are also impacted by sample hardness, although by a much smaller margin than models.1 1 Introduction Prior work on explainable NLP (Wiegreffe and Marasovic, 2021) has explored different forms of explanations ranging from extractive ratio- nales (Zaidan et al., 2007; DeYoung et al., 2020), semi-structured, and structured explana- tions (Jansen et al., 2019; Mostafazadeh et al., 1Supporting code and data are available at https:// github.com/swarnaHub/ExplanationHardness. 2020; Saha et al., 2021) to free-text explana- tions (Camburu et al., 2018). Due to the flexibility of free-text explanations, they have emerged as a popular form of explanations with multiple bench- marks developed around them, as well as models that generate such explanations using seq2seq lan- guage models (Ehsan et al., 2018; Camburu et al., 2018; Rajani et al., 2019; Narang et al., 2020). Few-shot prompting (Radford et al., 2019; Schick and Schütze, 2021) with Large Language Mod- els (LLMs) like GPT-3 (Brown et al., 2020) has been shown to produce highly fluent and factual natural language explanations that are often pre- ferred over crowdsourced explanations in existing datasets (Wiegreffe et al., 2022). However, past work has not yet explored a critical dimension of datapoint-level explanations, which is how hard the data point is to classify correctly. Given re- cent work on measuring hardness of individual data points (Swayamdipta et al., 2020), we study how sample hardness influences both LLMs’ and humans’ ability to explain data labels. In sum- mary, …

Content truncated to 500 words