Listen, Watch, and Learn to Feel: Retrieval-Augmented Emotion Reasoning for Compound Emotion Generation

ACL ARR 2025 February Submission2446 Authors

Abstract: The ability to comprehend human emotion using multimodal large language models (MLLMs) is essential for advancing human-AI interaction and multimodal sentiment analysis. While psychology theory-based human annotations have contributed to multimodal emotion tasks, the subjective nature of emotional perception often leads to inconsistent annotations, limiting the robustness of current models. Addressing these challenges requires more fine-grained methods and evaluation frameworks. In this paper, we propose the Retrieval-Augmented Emotion Reasoning (RAER) framework, a plug-and-play module that enhances MLLMs' ability to tackle compound and context-rich emotion tasks. To systematically evaluate model performance, we introduce the Stimulus-Armed Bandit (SAB) framework, designed to benchmark emotional reasoning capabilities. Additionally, we construct the Compound Emotion QA dataset, an AI-generated multimodal dataset aimed at strengthening emotion understanding in MLLMs. Experimental results demonstrate the effectiveness of RAER across both traditional benchmarks and SAB evaluations, highlighting its potential to enhance emotional intelligence in multimodal AI systems.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: Emotion Detection and Analysis; Human-AI Interaction; Retrieval-Augmented Generation; Multimodal QA; Evaluation Methodologies

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Data resources, Data analysis

Languages Studied: English; Chinese

Submission Number: 2446