Publicação
GraphEME: graph neural networks with multi-expert fusion for emotion-cause pair extraction
| Resumo: | Emotion-Cause Pair Extraction (ECPE) aims to identify clause-level pairs expressing emotions and their underlying causes. Existing methods rely on monolithic architectures that model causal relations through homogeneous cues, such as syntactic or sequential proximity, making them fragile to long-range or semantically com plex associations. We argue that emotion-cause reasoning is inherently heterogeneous, emerging from the confluence of diverse linguistic signals including syntactic, semantic role, coreference, temporal, and commonsense. To capture this complexity, we propose GraphEME, a multi-expert graph neural network that decomposes inference into specialized linguistic perspectives, each encoded as a subgraph and processed independently. A learnable gating mechanism dynamically fuses expert outputs, enhancing both robustness and interpretability. Experiments on Chinese ECPE benchmarks show that GraphEME achieves state-of-the-art performance, with ablation and case studies validating the necessity and effectiveness of multi-expert fusion. |
|---|---|
| Autores principais: | Gu, Xue |
| Outros Autores: | Meng, Ziyao; Tavares, Adriano; Gomes, Tiago Manuel Ribeiro; Xu, Hao |
| Assunto: | Emotion-cause pair extraction Multi-expert Graph neural network |
| Ano: | 2026 |
| País: | Portugal |
| Tipo de documento: | comunicação em conferência |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Emotion-Cause Pair Extraction (ECPE) aims to identify clause-level pairs expressing emotions and their underlying causes. Existing methods rely on monolithic architectures that model causal relations through homogeneous cues, such as syntactic or sequential proximity, making them fragile to long-range or semantically com plex associations. We argue that emotion-cause reasoning is inherently heterogeneous, emerging from the confluence of diverse linguistic signals including syntactic, semantic role, coreference, temporal, and commonsense. To capture this complexity, we propose GraphEME, a multi-expert graph neural network that decomposes inference into specialized linguistic perspectives, each encoded as a subgraph and processed independently. A learnable gating mechanism dynamically fuses expert outputs, enhancing both robustness and interpretability. Experiments on Chinese ECPE benchmarks show that GraphEME achieves state-of-the-art performance, with ablation and case studies validating the necessity and effectiveness of multi-expert fusion. |
|---|