A graph-to-text counterfactual prompting framework called GATOC-ERC enables frozen large language models to perform interpretable emotion recognition in conversation while achieving competitive task accuracy. By bridging structured conversational graphs with generative language models, the approach generates verifiable reasoning traces alongside final emotion classifications.
Emotion recognition in conversation requires computational systems to determine utterance-level emotional states by analyzing both local lexical wording and broader dialogue-level context. Established graph-based methods capture intricate speaker interactions and conversational dependencies across dialogue turns. However, the structural graph representations generated by these methods are typically consumed by specialized, task-specific classifiers. Because these representations remain embedded inside closed classification architectures, they cannot be readily exposed to large language models as transparent, inspectable reasoning evidence.
Direct prompting of large language models presents an alternative strategy for conversation analysis, yet it introduces significant drawbacks of its own. Although standard prompts allow language models to generate natural language explanations for their predictions, these explanations frequently lack explicit structural grounding in conversational topologies. Moreover, conventional prompting techniques rarely evaluate or contrast competing emotion hypotheses in a systematic manner, leaving the underlying decision-making process unverified across alternative conversational interpretations.
To overcome these complementary limitations, the research team designed GATOC-ERC as a modular graph-to-text prompting pipeline. The framework begins by constructing an utterance-level dialogue graph specifically structured to encode local conversational dependencies and dialogue interactions.
Following graph construction, a dedicated processing pipeline consisting of a Producer and a GraphTranslator converts the underlying graph data into natural language. This pipeline systematically extracts node attributes, neighboring connections, and relational information directly from the conversational graph, transforming them into coherent textual graph-context summaries. Because the resulting summaries consist entirely of descriptive text, they can be seamlessly consumed by a frozen large language model using standard text prompts, bypassing the need for architectural modifications or additional fine-tuning.
Conditioned on this graph-derived textual evidence, the frozen language model executes a counterfactual emotion verification procedure. Rather than generating a single emotion prediction directly, the model produces candidate-specific rationales for multiple possible emotion labels. The system then contrasts these alternative rationales against each other, evaluating their contextual plausibility to identify and select the specific emotion label that demonstrates the greatest overall consistency with the surrounding dialogue.
The authors validated the GATOC-ERC framework through empirical experiments on two standard benchmark datasets, IEMOCAP and MELD, evaluating text-only emotion recognition performance. Across both benchmarks, GATOC-ERC demonstrated competitive classification accuracy while successfully producing interpretable, inspectable reasoning traces that document how conversational context informed each prediction.
To further evaluate the system, the researchers conducted detailed ablation studies, faithfulness assessments, cross-dataset transfer experiments, prompt sensitivity tests, and computational cost analyses. These comprehensive evaluations clarified the distinct functional contributions of the graph-to-text translation interface and the counterfactual verification process, while providing objective insight into the operational boundaries and limitations of the combined architecture.
Subject of Research: Interpretable emotion recognition in conversation using graph-to-text counterfactual prompting
Article Title: Graph-to-text counterfactual verification for interpretable emotion recognition in conversation
Article References: Liu, J., Jin, S., Deng, Z., Zhang, X., & Xu, Z. (2026). Graph-to-text counterfactual verification for interpretable emotion recognition in conversation. Frontiers in Artificial Intelligence, 9, Article 1926067. https://doi.org/10.3389/frai.2026.1926067
Image Credits: AI Generated
DOI: 10.3389/frai.2026.1926067
Keywords: Emotion Recognition in Conversation, GATOC-ERC, Counterfactual Prompting, Graph-to-Text, Large Language Models, Frozen LLM, Dialogue Graph, IEMOCAP, MELD, Counterfactual Verification, GraphTranslator, Utterance-Level Emotion, Tree-of-Counterfactual, Interpretable Reasoning Traces, Conversational Dependencies, Speaker Interactions, Prompt Sensitivity, Cross-Dataset Transfer
News Source: Denise Maddox. (November 18, 2026). GATOC-ERC Framework Combines Conversational Graphs and Language Models for Emotion Recognition. Scienmag.



