Answers are grounded in this paper and cite their sources.
Promotional and SEO-oriented free-text content can undermine the neutrality and quality of OpenStreetMap (OSM). Because such text is multilingual and context-dependent, purely rule-based detection is challenging. This study evaluates large language models (LLMs) for identifying promotional content in OSM description=* tags and investigates their use in a community-based quality assurance workflow. A classification prompt was developed using a pilot dataset and evaluated on a balanced, manually annotated multilingual dataset of 1500 descriptions. Thirty-five open-weight model configurations were compared using precision, recall, F1-score, and runtime, and two conventional text-classification baselines were evaluated for classification performance. The best-performing model, llama3.3:70b, achieved 95.03% precision, 91.87% recall, and an F1-score of 93.42%, outperforming the evaluated conventional baselines. This model was then applied to 16,156 newly added or modified OSM descriptions in a real-world workflow. It identified 2738 descriptions as potentially promotional, resulting in 2599 public changeset comments. Within a 28-day follow-up period, 79.5% of analyzed commented cases were modified or removed, compared with 27.3% during a later comparison period without workflow-generated comments. Because the two conditions were temporally separated and non-randomized, this difference should be interpreted as an observational association.
No takes yet. Share an insight, caveat, or question.
Pascal Neis (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: