|国家预印本平台
首页|EMER-Ranker: Learning to Rank Emotion Descriptions in the Absence of Ground Truth

EMER-Ranker: Learning to Rank Emotion Descriptions in the Absence of Ground Truth

EMER-Ranker: Learning to Rank Emotion Descriptions in the Absence of Ground Truth

来源:Arxiv_logoArxiv
英文摘要

With the recent success of large language models, Explainable Multimodal Emotion Recognition (EMER), also known as Descriptive MER (DMER), has attracted growing attention from researchers. Unlike traditional discriminative methods that rely on predefined emotion taxonomies, EMER aims to describe a person's emotional state using free-form natural language, thereby enabling fine-grained and interpretable emotion representations. However, this free-form prediction paradigm introduces significant challenges in evaluation. Existing approaches either depend on ground-truth descriptions, which require extensive manual annotations and often fail to capture the full complexity of human emotions, or simplify the evaluation task by shifting focus from assessing descriptions to evaluating emotion labels. However, this simplification overlooks critical aspects such as emotional temporal dynamics, intensity, and uncertainty. To address these limitations, we propose EMER-Ranker, a novel evaluation strategy that reformulates the traditional ``prediction-ground truth'' comparison into the ``prediction-prediction'' comparison, eliminating the need for ground-truth descriptions. We then apply the Bradley-Terry algorithm to convert pairwise comparison outcomes into model-level rankings. Additionally, we explore the potential for automatic preference prediction and introduce EMER-Preference, the first preference dataset specifically designed for human emotions. Our work advances the field of EMER and lays the foundation for more intelligent human-computer interaction systems.

Zheng Lian、Licai Sun、Haoyu Chen、Zebang Cheng、Fan Zhang、Ziyu Jia、Ziyang Ma、Fei Ma、Xiaojiang Peng、Jianhua Tao

计算技术、计算机技术

Zheng Lian,Licai Sun,Haoyu Chen,Zebang Cheng,Fan Zhang,Ziyu Jia,Ziyang Ma,Fei Ma,Xiaojiang Peng,Jianhua Tao.EMER-Ranker: Learning to Rank Emotion Descriptions in the Absence of Ground Truth[EB/OL].(2025-07-14)[2025-07-16].https://arxiv.org/abs/2507.04278.点此复制

评论