ChatGPT Rates Natural Language Explanation Quality Like Humans: But on Which Scales?

TL;DR

The paper investigates whether ChatGPT can rate natural language explanations (NLEs) as humans do across ordinal scales, addressing the cost and subjectivity of human evaluation. It compares ChatGPT's informativeness and clarity judgments to 900 human annotations on 300 NLE instances drawn from three datasets (e-SNLI, LIAR-PLUS, Latent Hatred) across binary, ternary, and 7-point scales, and also examines pairwise NLE comparisons and dynamic prompting. Results show strong alignment for coarse (binary/ternary) ratings but diminished accuracy for fine-grained 7-point judgments; pairwise comparisons are easier and dynamic prompting can improve agreement in comparison tasks, though not consistently for single-score ratings. The findings offer practical guidance for leveraging ChatGPT to supplement or reduce human annotation costs in responsible AI evaluation, while highlighting limitations and the need to validate across more models and tasks.

Abstract

As AI becomes more integral in our lives, the need for transparency and responsibility grows. While natural language explanations (NLEs) are vital for clarifying the reasoning behind AI decisions, evaluating them through human judgments is complex and resource-intensive due to subjectivity and the need for fine-grained ratings. This study explores the alignment between ChatGPT and human assessments across multiple scales (i.e., binary, ternary, and 7-Likert scale). We sample 300 data instances from three NLE datasets and collect 900 human annotations for both informativeness and clarity scores as the text quality measurement. We further conduct paired comparison experiments under different ranges of subjectivity scores, where the baseline comes from 8,346 human annotations. Our results show that ChatGPT aligns better with humans in more coarse-grained scales. Also, paired comparisons and dynamic prompting (i.e., providing semantically similar examples in the prompt) improve the alignment. This research advances our understanding of large language models' capabilities to assess the text explanation quality in different configurations for responsible AI development.

Figures (7)

• Figure 1: Correlation between ChatGPT and human evaluations for the three datasets and the two metrics, Informativeness and Clarity. Pearson's and Spearman's correlation coefficients are in the figure.
• Figure 2: Visualization of ChatGPT pair comparison f1-scores for various $\Delta_{i}$ and $\Delta_{c}$, from 0 to 6 rounded by integer bins, compared with additional human annotations specifically for the pair comparison task.
• Figure 3: Visualization of ChatGPT pair comparison f1-scores for various $\Delta_{i}$ and $\Delta_{c}$, from 0 to 6 rounded by integer bins, showing the difference between vanilla prompting and our proposed dynamic prompting. The measurement is only based on NLE evaluation metrics of Informativeness and Clarity.
• Figure 4: The screenshot of our design user interface used to collect the human evaluation scores of informativeness and clarity.
• Figure 5: The screenshot of the detailed instruction information in MTurk.