Answers are grounded in this paper and cite their sources.
Emotion Recognition in Conversation (ERC) requires models to infer utterance-level emotions from both local wording and dialogue-level context. Graph-based ERC methods capture speaker interactions and contextual dependencies, but their graph representations are usually consumed by task-specific classifiers. They are therefore difficult to expose to large language models (LLMs) as transparent reasoning evidence. In contrast, direct LLM prompting can generate explanations, but it often lacks explicit structural grounding and rarely compares alternative emotion hypotheses in a systematic way. To address these limitations, we propose GATOC-ERC, a graph-to-text counterfactual prompting framework for text-only ERC. The framework first builds an utterance-level dialogue graph to encode local conversational dependencies. A Producer and GraphTranslator pipeline then converts graph-derived node, neighbor, and relation information into textual graph-context summaries. These summaries can be used by a frozen LLM through standard text prompts. Given this graph-conditioned evidence, the model performs counterfactual emotion verification. It generates candidate-specific rationales for possible emotions and contrasts them to select the most contextually consistent label. Experiments on IEMOCAP and MELD show that GATOC-ERC achieves competitive text-only ERC performance while producing interpretable reasoning traces. Ablation, faithfulness, cross-dataset transfer, prompt sensitivity, and cost analyses further clarify the contribution and limitations of the graph-to-text interface and the counterfactual verification process.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: