When Large Language Models are Reliable for Judging Empathic Communication

TL;DR

It is shown that expert agreement offers a more informative benchmark for contextualizing LLM performance than standard classification metrics and can support transparency and oversight in emotionally sensitive applications including their use as conversational companions.

Abstract

Large language models (LLMs) excel at generating empathic responses in text-based conversations. But, how reliably do they judge the nuances of empathic communication? We investigate this question by comparing how experts, crowdworkers, and LLMs annotate empathic communication across four evaluative frameworks drawn from psychology, natural language processing, and communications applied to 200 real-world conversations where one speaker shares a personal problem and the other offers support. Drawing on 3,150 expert annotations, 2,844 crowd annotations, and 3,150 LLM annotations, we assess inter-rater reliability between these three annotator groups. We find that expert agreement is high but varies across the frameworks' sub-components depending on their clarity, complexity, and subjectivity. We show that expert agreement offers a more informative benchmark for contextualizing LLM performance than standard classification metrics. Across all four frameworks, LLMs consistently approach this expert level benchmark and exceed the reliability of crowdworkers. These results demonstrate how LLMs, when validated on specific tasks with appropriate benchmarks, can support transparency and oversight in emotionally sensitive applications including their use as conversational companions.

Figures (17)

• Figure 1: Reliability Across Annotator Pairs and Sub-Components.Top: Inter-rater reliability (quadratically weighted $\kappa_w$) across annotator pairs for each empathic communication sub-components. In the evaluation of experts with each other, the circle represents the median $\kappa_w$ and the error bar represents the $\kappa_w$ range between the three pairs. Experts comparisons with the LLM (Gemini 2.5 Pro) and crowd compare median expert annotations with the LLM and crowd, respectively. Bottom: Empathic communication frameworks with verbatim annotation questions.
• Figure 2: Reliability across Annotator Pairs and Frameworks. Inter-rater reliability (quadratically weighted $\kappa_w$)) for empathic communication across annotator pairs for each sub-component grouped by framework. In the evaluation of experts with each other, circle represents the the median $\kappa_w$ between the three expert pairs. Experts comparisons with the LLM (Gemini 2.5 Pro) and crowd compare median expert annotations with the LLM and crowd, respectively. The beeswarm plot ensures all data points are visible by applying a horizontal jitter to avoid overlap. Dotted lines represent the 25th, 50th, and 75th percentiles of Weighted Cohen's Kappa.
• Figure 3: Comparing Contextualized Inter-Rater Reliability with Multi-Class and Binary F1 Scores. Experts vs Expert1 (purple), Experts vs LLM ((blue)), and Experts vs Crowd (red) present agreement metrics between the median expert annotation and expert 1, Gemini 2.5 Pro, and the crowd. We present results for four datasets (A) EPITOME, (B) Perceived Empathy, (C) Empathetic Dialogues, and (D) Lend an Ear. Random classifier baselines are indicated by dotted (multi-class) and dashed (binary) lines.
• Figure 4: Expert Inter-rater Reliability across Frameworks. Pairwise expert reliability across four empathic communication frameworks
• Figure 5: Expert Inter-rater Reliability across Sub-components. Inter-rater agreement as Krippendorff's alpha among the three expert annotators across frameworks' sub-components. Error bars are 95% confidence intervals obtained by bootstrapping conversations, and the background color indicates the degree of agreement regier2013dsm.