A Comparative Study of Lexicon-based and Multi-LLM Sentiment Analysis in Mental Health Discourse on Social Media

View Full Paper

Ask AI

Bookmark

Share

Key Points

This research aims to compare sentiment classification results between lexicon-based and multi-large language model analyses on mental health related texts in social media.
Collected 8,020 posts from Naver blogs and news between December 2025 and January 2026.
Utilized KNU Korean sentiment lexicon and three LLMs (EXAONE, Llama 3, GPT-4o) for sentiment classification.
Employed Cohen's Kappa coefficient for measuring agreement and benchmarked with NSMC's 1,000 entries.
All three LLMs showed higher accuracy in sentiment classification than the lexicon approach (Lexicon: 0.593, GPT-4o: 0.842, Llama 3: 0.806, EXAONE: 0.754).
The lexicon classified 74.5% of mental health data as negative, whereas LLM majority voting identified only 61.7% as negative.
60.6% of the discrepancies involved the lexicon classifying texts as negative, which LLMs reclassified as positive or neutral.

Answers are grounded in this paper and cite their sources.

Comparative study evaluates sentiment classification in mental health discourse on social media, suggesting LLMs may correct biases.

No abstract yet. The journal hasn't published one we can show.

No takes yet. Share an insight, caveat, or question.

JongHwi Song (2026) studied this question.

View Full Paper

Ask AI

Bookmark

Share