Scholarlink

We present a pipeline that automatically assigns concise, human-readable labels to under-represented mental-health topics discovered by a seed-guided nonnegative matrix factorization model. Given each topic’s word distribution, we rank documents via Jensen–Shannon divergence, extract anchored n-gram candidates, score them on informativeness, phraseness, and seed overlap, and finally ask a large language model (LLM) to choose and justify a label. The approach amplifies minority voices while remaining fully automatic and language-agnostic. We demonstrate its efficacy on Finnish-language online discussion of YouTuber vlogs containing mental health minority themes. In two human evaluations of label quality, our model attains high expert scores and outperforms a baseline approach.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Voices Between Lines: Interpretable Labeling of Mental Health Minority Topics with Seed Guidance and LLMs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Seyedeh Fatemeh Ebrahimi,
  • Jaakko Peltonen

摘要

We present a pipeline that automatically assigns concise, human-readable labels to under-represented mental-health topics discovered by a seed-guided nonnegative matrix factorization model. Given each topic’s word distribution, we rank documents via Jensen–Shannon divergence, extract anchored n-gram candidates, score them on informativeness, phraseness, and seed overlap, and finally ask a large language model (LLM) to choose and justify a label. The approach amplifies minority voices while remaining fully automatic and language-agnostic. We demonstrate its efficacy on Finnish-language online discussion of YouTuber vlogs containing mental health minority themes. In two human evaluations of label quality, our model attains high expert scores and outperforms a baseline approach.