SSG-ECPE: Semantics-Structured Generation with Alignment for Emotion-Cause Pair Extraction

Xue Gu, Ziyao Meng, Tiago Gomes, Adriano Tavares, Hao Xu

Keywords: Emotion Cause Pair Extraction, Joint Multi-task Learning, Generative Perspective

Abstract: Emotion Cause Pair Extraction (ECPE) aims to jointly identify emotion clauses and their corresponding cause clauses, forming emotion-cause pairs (ECPs). Existing approaches either rely on complex discriminative architectures to model pair boundaries or adopt generic text-to-text frameworks that flatten ECPE into plain sequence generation. Both paradigms overlook rich semantic dependencies, such as clause roles, emotion types, and clue words, and struggle in multi-pair scenarios with nested or overlapping structures. In this paper, we propose a task-adaptive generative multi-task learning framework that rethinks ECPE as a structured text-to-text generation task. We design semantics-structured output formats that explicitly encode clause roles, emotion types, and trigger words as semantic markers, allowing the model to capture inter-label dependencies and co-occurrence patterns during generation. For emotion clause extraction (EE), outputs are formatted as \textit{(clause, emotion type, trigger words)} triplets; for ECPE, emotion–cause pairs are directly generated, enabling implicit modeling of emotional reasoning. A shared encoder with task-specific decoders supports both clause- and pair-level generation within a unified pipeline. To enhance reliability, we further introduce a Clause Prediction Alignment (CPA) strategy that grounds generated clauses to input spans, mitigating hallucinations and ensuring faithfulness. Extensive experiments demonstrate that CPA is indispensable: without it, performance collapses, whereas with it, our framework achieves consistent state-of-the-art results, including a +21.3 F1 improvement on the English benchmark.

Primary Area: generative models

Submission Number: 11665