From Prediction to Justification: Aligning Sentiment Reasoning with Human Rationale via Reinforcement Learning
Abstract
Although Aspect-based Sentiment Analysis (ABSA) systems attain high sentiment polarity detection accuracy, they act as "black boxes" without the explicit reasoning of human affective cognition, where humans form causal explanations for sentiment judgments. To address this, we propose ABSA-R1, a large language model framework mimicking the "reason-before-predict" cognitive process. Using reinforcement learning (RL), it generates natural language justifications to underpin sentiment predictions. We design a Cognition-Aligned Reward Model to ensure consistency between reasoning paths and emotional labels, and a performance-driven rejection sampling strategy (inspired by metacognitive monitoring) targeting hard cases with uncertain/inconsistent internal reasoning. Experiments on four benchmarks show this explicit reasoning boosts model interpretability and outperforms non-reasoning baselines in sentiment classification and triplet extraction.