Path-Aware Knowledge Injection for Fine-Grained Emotion Recognition in Mental Health Counseling

Fine-grained, multi-label emotion recognition in textual dialogues, particularly in contexts like mental health counseling, presents significant challenges due to the complexity and concurrency of human emotions. Standard pre-trained language models, while powerful, often struggle to capture the underlying reasoning and emotional interdependencies that are not explicitly stated in the text. To address this, we propose PK-GAT, a novel framework for Path-aware Knowledge-injected Graph Attention Networks. Our approach moves beyond traditional knowledge fusion by actively reasoning over a commonsense knowledge graph (i..e, SenticNet). It identifies, encodes, and ranks multi-hop semantic paths that connect textual concepts to emotion labels, thereby injecting rich, inferential logic into the model. This path-aware mechanism is integrated with a Graph Attention Network (GAT) that dynamically models the relationships between emotion labels, allowing them to mutually refine one another. We conduct extensive experiments on the GoeMotions and SemEval-2018 datasets. Our results demonstrate that PK-GAT achieves new state-of-the-art performance, attaining a Macro-F1 score of 58.72% and a Micro-F1 of 69.85% on GoeMotions, outperforming strong baselines and recent models. Additionally, through detailed ablation and case studies, we validate the significant contributions of both the path-based reasoning and the dynamic label graph modeling, offering a robust and effective solution for nuanced emotion recognition.