Emotion-Cause Pair Extraction in Conversations via Semantic Decoupling and Graph Alignment

Abstract: 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 an independent pairwise classification task, 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. Our codes are released at https://github.com/CoCoSphere/SCALE.

• The paper introduces SCALE, which separates emotion and cause representations with graph attention and uses fused optimal transport to model globally consistent many-to-many causal links.
• SCALE achieves the best F1 on RECCON-DD (58.83), RECCON-IE (34.69), and ECF (57.70), including a 20% relative improvement over the strongest baseline on cross-domain RECCON-IE.
• The method handles multi-cause conversations more effectively and uses 8.2M parameters and 2.15 GFLOPs, but remains sensitive to local-versus-global evidence weighting and struggles with long-distance dependencies.

Motivation and problem formulation

Emotion–Cause Pair Extraction in Conversations (ECPEC) requires identifying the set of causal relations between emotion utterances and their triggering causes within a dialogue. The paper argues that existing approaches, whether following a select-then-pair pipeline (e.g., MECPE-2steps) or an end-to-end embed-then-pair scheme (e.g., PRG-MoE, Joint-Xatt, CENTER), share two structural deficiencies. First, they encode emotion-related and cause-related information in a unified representation space, implicitly assuming that emotion diffusion and cause explanation follow homogeneous relational patterns, even though emotional states propagate through contextual and speaker-dependent dynamics while causes rest on asymmetric explanatory dependencies. Second, they treat ECPEC as independent one-to-one binary pair classification, which cannot enforce globally consistent many-to-many causal structures where several causes jointly trigger an emotion or one cause influences multiple emotional outcomes.

The proposed framework, SCALE (Semantic Causal ALignment for ECPEC), addresses both issues by decoupling emotion-oriented and cause-oriented semantics into complementary representation spaces induced from a shared conversation graph, and by formulating pair extraction as a global alignment problem solved with optimal transport (OT).

Architecture

SCALE encodes each conversation as a graph whose nodes are RoBERTa-derived utterance embeddings concatenated with learnable speaker embeddings. Edges are of three types: global contextual edges created when cosine similarity exceeds a threshold $\tau_s$, local temporal edges within a sliding window $W$ with exponentially decaying weights, and intra-speaker edges capturing speaker-specific consistency. The adjacency matrix is treated as learnable and jointly optimized.

Two graph attention encoders operate over this shared topology: an emotion-aware encoder and a cause-aware encoder. Beyond producing node representations $H^{(E)}$ and $H^{(C)}$, each encoder's attention coefficients define task-specific refined adjacency matrices $A^{(E)}$ and $A^{(C)}$, which capture how relational importance differs between the two semantic views.

Pair inference is then cast as learning a transport plan $T \in \mathbb{R}^{N\times N}$ between the two representation sets. The alignment objective combines an attribute-level cost based on cosine distance between emotion and cause representations with a structure-consistency term penalizing disagreement between the relational patterns of $A^{(E)}$ and $A^{(C)}$ weighted by the plan itself — a fused Gromov-Wasserstein-style objective. Because the quadratic term is non-linear, the authors linearize it iteratively and solve entropic OT subproblems via Sinkhorn normalization, followed by a row-wise softmax with temperature to emphasize dominant alignments.

The final pair score blends global correspondence with local discriminative evidence: a lightweight MLP classifier over concatenated representations is combined with the OT-derived score through a mixing coefficient $\beta$. Utterance-level emotion extraction (EE) and cause extraction (CE) heads provide auxiliary supervision, and a KL-divergence regularizer encourages agreement between the local pairwise predictions and the OT alignment distribution. The full objective is trained jointly with AdamW on a single RTX 4090.

Main results

Experiments cover RECCON-DD, the cross-domain RECCON-IE subset, and the multimodal ECF benchmark, against eight baselines including generative LLM-based methods (GMEC). SCALE achieves the highest recall and F1 on all three datasets:

Method	RECCON-DD F1	RECCON-IE F1	ECF F1
PRG-MoE	57.26	28.90	50.86
GMEC	53.07	28.25	54.73
MECPE-2steps	48.34	10.24	52.71
SCALE	58.83	34.69	57.70

The most striking result is on cross-domain generalization: on RECCON-IE, SCALE improves F1 from 28.90 to 34.69, a relative gain of +20.0%, suggesting that the decoupled semantic spaces and global alignment transfer better than pairwise classifiers tuned to the training distribution. On ECF the relative improvement over the strongest baseline is +9.5%. The authors note explicitly that SCALE does not attain the highest precision; the soft OT alignment favors broader matching and thus trades precision for recall — a design consequence rather than an incidental artifact.

On dedicated multi-cause test subsets (dialogues where one emotion has two or more annotated causes), all models degrade substantially, but SCALE retains the best F1 everywhere: 38.33 on RECCON-DD, 25.33 on RECCON-IE (versus 21.62 for the next best), and 35.55 on ECF. This supports the central claim that many-to-many soft correspondences are better suited to dispersed causal evidence than independent binary decisions.

Ablations confirm both design principles: removing separated representation learning (SRL) costs roughly 2 F1 points per dataset, removing global alignment (GA) costs 3–4 points, and removing both costs up to 5 points on RECCON-DD. Auxiliary EE/CE supervision contributes more modestly but consistently. Hyperparameter analysis shows the framework is stable across window sizes and attribute/structure weightings, but critically sensitive to $W$0: forcing pure reliance on the OT alignment ($W$1) collapses F1 to 16.84 on ECF, indicating that the global alignment signal alone is insufficient without local classifier evidence.

Two further analyses strengthen the case. Against few-shot prompted LLMs (DeepSeek-V3.2, GPT-5.1 Instant, Gemini-3-pro-preview), SCALE outperforms the best LLM on both tested datasets (58.83 vs. 56.08 on RECCON-DD; 57.70 vs. 55.42 on ECF), supporting the claim that explicit structural modeling remains advantageous for this task. On efficiency, SCALE uses only 8.2M parameters, ~2.15 GFLOPs, and 7 GB peak memory, versus 110M parameters/~220 GFLOPs for PRG-MoE and 450M for GMEC — an order-of-magnitude cost reduction alongside superior accuracy. t-SNE visualizations show that baseline representations intertwine emotion and cause features into a single dense cluster, whereas SCALE yields well-separated subspaces with fine-grained local clusters, providing qualitative evidence for the semantic decoupling hypothesis.

A simple multimodal extension on ECF, concatenating audio and video features, raises F1 from 57.70 to 58.63, indicating the framework accommodates additional modalities without architectural change.

Limitations and open questions

The authors are candid about several constraints. SCALE is developed for textual conversations; although the multimodal appendix shows gains from naive feature concatenation, principled multimodal fusion within the alignment framework is unexplored. Qualitative analysis reveals persistent failures on long-distance dependencies such as $W$2 and $W$3 in the case study, and error analysis shows monotonically decreasing recall as the number of causes per emotion grows — exhaustive cause retrieval in complex multi-cause settings remains unsolved. The sensitivity to $W$4 also leaves open how to balance global alignment against local evidence without dataset-specific tuning. Finally, the authors caution that models trained on annotated corpora may inherit annotation bias or incomplete causal assumptions, positioning SCALE as an assistive tool rather than a definitive account of conversational causality.

Conclusion

SCALE reframes ECPEC as a global many-to-many alignment problem between semantically decoupled emotion-side and cause-side graph representations, solved with entropy-regularized fused Gromov-Wasserstein optimal transport. The approach delivers state-of-the-art F1 on three benchmarks with particularly large gains in cross-domain transfer (+20% relative on RECCON-IE) and multi-cause scenarios, while using an order of magnitude fewer parameters and FLOPs than competing systems. The evidence indicates that separating emotion diffusion from cause explanation and enforcing globally consistent correspondences are both necessary for robust conversational causal modeling, though long-range dependencies and exhaustive multi-cause retrieval remain open challenges.