Towards Unified Emotion Understanding: A One-Step End-to-End Model for Emotion–Cause Pair Extraction

Emotion analysis in textual data has evolved from simple emotion detection to sophisticated Emotion-Cause Pair Extraction (ECPE), which jointly identifies emotions and their corresponding causes without requiring pre-annotated emotion labels. Traditional ECPE approaches employ multi-stage pipelines that separately extract emotions and causes before pairing, leading to significant error propagation that degrades overall performance. This paper proposes a novel one-step, end-to-end deep learning framework that directly extracts emotion-cause pairs from unannotated documents, eliminating intermediate error accumulation while capturing their inherent interdependency through unified contextual modeling. Our model leverages advanced transformer architectures with multi-granular attention mechanisms to simultaneously encode document-level semantics, clause interactions, and position-aware dependencies. Extensive experiments on the benchmark ECPE-News dataset demonstrate superior F1-scores compared to state-of-the-art two-step and existing end-to-end methods, achieving substantial improvements in both precision and recall. The proposed approach exhibits enhanced generalization across diverse domains including social media and conversational texts, addressing key limitations of prior work. By providing a robust, efficient solution for real-world emotion analysis applications such as mental health monitoring, customer sentiment analysis, and dialogue systems, this work advances practical emotion understanding capabilities significantly.