Are Human Explanations Always Helpful? Towards Objective Evaluation of Human Natural Language Explanations

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 14698–14713 July 9-14, 2023 ©2023 Association for Computational Linguistics Are Human Explanations Always Helpful? Towards Objective Evaluation of Human Natural Language Explanations Bingsheng Yao Rensselaer Polytechnic Institute Prithviraj Sen ∗ Amazon Lucian Popa IBM Research James Hendler Rensselaer Polytechnic Institute Dakuo Wang † Northeastern University Abstract Human-annotated labels and explanations are critical for training explainable NLP mod- els. However, unlike human-annotated labels whose quality is easier to calibrate (e.g., with a majority vote), human-crafted free-form expla- nations can be quite subjective. Before blindly using them as ground truth to train ML mod- els, a vital question needs to be asked: How do we evaluate a human-annotated expla- nation’s quality? In this paper, we build on the view that the quality of a human-annotated explanation can be measured based on its help- fulness (or impairment) to the ML models’ per- formance for the desired NLP tasks for which the annotations were collected. In comparison to the commonly used Simulatability score, we define a new metric that can take into con- sideration of the helpfulness of an explanation for model performance at both fine-tuning and inference. With the help of a unified dataset format, we evaluated the proposed metric on five datasets (e.g., e-SNLI) against two model architectures (T5 and BART), and the results show that our proposed metric can objectively evaluate the quality of human-annotated expla- nations, while Simulatability falls short. 1 Introduction Despite the recent advances of large-scale language models (LLM) (Devlin et al., 2019; Qin et al., 2023; Lewis et al., 2019; Raffel et al., 2020), which ex- hibit close-to-human performance on many natural language processing (NLP) tasks (e.g., Question Answering (Rajpurkar et al., 2016; Koˇcisk`y et al., 2018; Mou et al., 2020, 2021; Xu et al., 2022), Natural Language Inference (Bowman et al., 2015; Williams et al., 2017; Wang et al., 2018), and Text ∗† Work done while Prithviraj was at IBM Research. †d.wang@northeastern.edu Corresponding Author. Generation (Duan et al., 2017; Yao et al., 2022; Zhao et al., 2022)), humans are eager to know how State-of-the-Art (SOTA) models arrive at a pre- diction. Researchers working on natural language explanations1 turned to human annotators for help by recruiting crowd-workers or experts to annotate both the labels and corresponding natural language explanations (Camburu et al., 2018; Rajani et al., 2019; Aggarwal et al., 2021; Wang et al., 2019b); Researchers can thus leverage human-annotated ex- planations to boost models’ prediction performance or train models to generate human-understandable natural language explanations. However, the quality issue of human-annotated explanations has yet to be explored. Researchers of- ten leverage popular Natural Language Generation (NLG) metrics such as BLEU (Papineni et al., 2002) and ROUGE (Lin, 2004) to evaluate the similarity between model-generated and human-annotated ex- planations, with a strong assumption that human- annotated ones are the gold standard. Neverthe- less, unlike providing labels for classification or multiple-choice QA tasks (Chen et al., 2021), dif- ferent people may …

Content truncated to 500 words