Abstract
Detecting nuanced emotions such as empathy and depression in adolescent social media is difficult due to sarcasm, slang, and context-dependent cues. Large language models (LLMs) show promise but often misclassify or hallucinate when applied to youth discourse. This study evaluates Retrieval-Augmented Generation (RAG) for enhancing LLM-based emotion detection in YouTube comments authored by teenagers. From over 15, comments on youth mental health videos, we curated 8 comments—4 challenging and 4 typical—for systematic analysis. Three state-of-the-art LLMs (ChatGPT, Gemini, Llama2) were tested under simple, transcript-only, and multi-contextual RAG prompting. To our knowledge, this is the first systematic, cross-model evaluation of multi-contextual RAG for adolescent YouTube comments, showing statistically significant improvements across both typical and challenging cases (McNemar’s, Cochran’s Q). Findings highlight RAG’s potential to reduce hallucinations, improve robustness, and support ethically responsible adolescent mental health monitoring on social media. This pilot study establishes proof-of-concept and directions for scalable research.
Recommended Citation
Vahdati, Farahnaz; Atif, Amara; and Saberi, Morteza, "A Novel Context-Aware RAG Approach for Detecting Empathy
and Depression in Youth YouTube Comments" (2025). ACIS 2025 Proceedings. 64.
https://aisel.aisnet.org/acis2025/64