FigSIM: A Dataset for Fine-grained Suicide Severity and Figurative Language in Suicide Memes

The dataset's real innovation isn't just labeling memes; it's the annotation schema itself. By marking figurative language separately from severity, researchers created a tool that lets classifiers learn which linguistic patterns correlate with genuine risk versus performative dark humor. This distinction is absent from most content moderation datasets, which treat memes as binary problems.

This connects directly to the emergency department triage work from June 1st, which showed that hybrid ML pipelines combining traditional classifiers with LLM screening can detect self-harm signals that rule-based systems miss. FigSIM operates in the same detection space but upstream, on social media rather than clinical notes. Both projects share the same insight: harm detection requires moving beyond keyword matching to understanding context and linguistic nuance. The triage paper proved the clinical case; FigSIM builds the infrastructure for social platforms to do the same work at scale.

If major platforms (Meta, TikTok, X) integrate FigSIM-trained classifiers into their moderation pipelines within 12 months and publish false positive/negative rates on held-out meme datasets, that confirms the dataset has real operational value. If adoption stalls and the dataset remains academic, it suggests the gap between research annotation and production moderation is wider than the paper acknowledges.