Emotion-Cause Pair Extraction in Conversations (ECPEC) aims to identify the set of causal relations between emotion utterances and their triggering causes within a dialogue. Most existing approaches formulate ECPEC as independent pairwise classification, overlooking the distinct semantics of emotion diffusion and cause explanation, and failing to capture globally consistent many-to-many conversational causality. To address these limitations, we revisit ECPEC from a semantic perspective and seek to disentangle emotion-oriented semantics from cause-oriented semantics, mapping them into two complementary representation spaces to better capture their distinct conversational roles. Building on this semantic decoupling, we naturally formulate ECPEC as a global alignment problem between the emotion-side and cause-side representations, and employ optimal transport to enable many-to-many and globally consistent emotion-cause matching. Based on this perspective, we propose a unified framework SCALE that instantiates the above semantic decoupling and alignment principle within a shared conversational structure. Extensive experiments on several benchmark datasets demonstrate that SCALE consistently achieves state-of-the-art performance.

Recent advances in Large Language Models have transformed Emotion Recognition in Conversation, yet existing instruction-tuned approaches still produce high-confidence misjudgments on utterances whose surface text is weakly informative but whose emotion depends on the surrounding situation. We propose SRERC, a selective reflective correction framework for LLM-based ERC. SRERC uses a verdict teacher to assess contextual support for the first-pass prediction and triggers reflective decoding only when reflection has an actionable role. For selected cases, an SRC generator constructs a Scene-aware Reflection Context, which provides target-specific situational grounding and verification probes for an SRC-conditioned second pass. This gated design enables targeted correction while preserving non-gated predictions and limiting unnecessary over-correction. Experiments on IEMOCAP, MELD, and EmoryNLP show that SRERC achieves the highest average weighted-F1 across the three benchmarks, with diagnostic gains on context-sensitive cases where surface cues are insufficient or misleading.