Request Correction

Knowledge-Infused Hierarchy-Aware Emotion Recognition in Code-mixed Mental Health Counseling Conversations

Effective counseling is often best achieved in a client’s preferred language, allowing better emotional resonance. Despite this, most existing research in emotion recognition in counseling focuses predominantly on English, overlooking the rich emotional and linguistic complexities of other widely spoken languages. Hinglish, a code-mixed blend of Hindi and English, is one such underexplored linguistic medium that millions use to express their emotions authentically. To address this gap, our research lays a foundational step in developing a mental-health conversation dataset in code-mixed Hinglish language, aka. IndieMH. We manually translate counseling conversations from publicly available sources into Hinglish. Moreover, we employ the dataset for emotion classification task for counseling patients. We prepare an exhaustive annotation guideline to annotate IndieMH with 13 emotional states under 3 board emotion categories. Our rigorous sanity check ensures that the quality of IndieMH adheres to research standards. Furthermore, we propose a novel knowledge-cum-hierarchy aware method named Healer for counseling emotion classification in the Hinglish language. To evaluate the model’s performance, we benchmark Healer against 11 potential baseline methods and report standard classification metrics, including accuracy, weighted-F1, and weighted-precision.