Emotion-Aware Graph Neural Meta-Learning for Few-Shot Manipulation Detection in Dialogues

Detecting subtle forms of manipulation in multi-turn conversations is essential for mitigating misinformation and improving the quality of online dialogue.This work introduces a novel end-to-end manipulation detection pipeline that integrates semantic, emotional, and structural reasoning.Utterance-level semantic embeddings from RoBERTa are combined with emotional cues from EmoBERTa and keyword-emotion rules to provide weak supervision for manipulation strategies such as guilt, coercion, flattery, and distraction.Each dialogue is modeled as a feature-enriched graph in which utterances function as nodes and conversational transitions define edges, enabling more expressive discourse representation.A Graph Convolutional Network optimized with Model-Agnostic Meta-Learning learns contextual manipulation cues and demonstrates strong generalization under limited annotated data.To further improve interpretability, the system incorporates rule-based keyword matching and SHAP-based feature attribution, offering clear and transparent explanations for model predictions.Experimental results show that the proposed approach achieves a Macro F1 score of 89.4% and an accuracy of 95.3%, highlighting its effectiveness, robustness, and interpretability for manipulation detection in dialogues.