Do LLMs Grasp Emotion or Just Keywords? A Benchmark for Implicit Emotion Recognition

Abstract

Recent advances in Large Language Models (LLMs) have shown strong performance in natural language understanding tasks, including emotion recognition. Implicit Emotion Recognition (IER) focuses on identifying emotions in text without explicit emotional cues, posing a significant challenge in Natural Language Processing. This study presents a systematic evaluation of state-of-the-art LLMs, namely, ChatGPT-5, Phi-4, LLaMA-3, and Gemma-3, on IER under zero-shot and few-shot prompting strategies. The models were tested on two complementary datasets: ISEAR, containing naturally implicit emotions in coherent narratives, and TEST 2018, where explicit emotion words were masked to simulate context-only inference. Experimental results show that LLMs, particularly ChatGPT-5, can effectively capture implicit emotions, with few-shot prompting consistently improving performance. Performance differences between datasets highlight the role of context quality, with narrative texts facilitating inference and social media content introducing challenges.