Emotion-o1: Adaptive Long Reasoning for Emotion Understanding in LLMs

2025-05-28

AI Summary

Existing affective understanding methods—such as sentiment classification, sarcasm, and humor detection—rely on fixed-length chains-of-thought (CoT), limiting adaptability to varying task complexities. Method: We propose a task-adaptive long-chain reasoning framework featuring (i) the first emotion-task-driven dynamic control mechanism for reasoning depth; (ii) a four-objective composite reward function balancing accuracy, depth controllability, path diversity, and logical redundancy suppression; and (iii) an integrated training pipeline combining supervised fine-tuning and reinforcement learning atop DeepSeek-R1 to enable dynamic reasoning chain generation and structured reward modeling. Contribution/Results: Evaluated across sentiment, emotion, humor, and sarcasm tasks, our approach achieves F1 improvements of 3.56%–37.95% and accuracy gains of 2.76%–23.14%, significantly enhancing large language models’ capacity for deep, interpretable reasoning over complex affective phenomena.

Technology Category

Cognitive Modeling & Cognitive Systems: Affective ComputingNatural Language Processing: Sentence-level Semantics, Textual Inference, etc.Humans and AI: Emotional Intelligence

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsEconomics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Emotion understanding includes basic tasks (e.g., sentiment/emotion classification) and advanced tasks (e.g., sarcasm/humor detection). Current methods rely on fixed-length CoT reasoning, failing to adapt to the varying complexity of emotions. We propose a task-adaptive reasoning framework that employs DeepSeek-R1 to generate variable-length reasoning chains for different emotion tasks. By combining fine-tuning with reinforcement learning, we design a composite reward function that balances four objectives: prediction accuracy, adaptive reasoning depth control, structural diversity in reasoning paths, and suppression of repetitive logic. This approach achieves dynamic context-sensitive inference while enabling LLMs to autonomously develop deep reasoning capabilities. Experimental results demonstrate consistent improvements in both Acc and F1 scores across four tasks: emotion, sentiment, humor, and sarcasm. Notably, peak enhancements reached 3.56% F1 (2.76% Acc) for basic tasks and 37.95% F1 (23.14% Acc) for advanced tasks. Our work bridges rigid CoT reasoning and emotional complexity through adaptive-depth analysis.

Problem

• Adaptive reasoning for varying emotion task complexity
• Dynamic context-sensitive inference for emotion understanding
• Improving accuracy in basic and advanced emotion tasks

Innovation

• Adaptive reasoning framework for emotion tasks
• Composite reward with fine-tuning and RL
• Dynamic context-sensitive inference for LLMs