A research team has released a labeled dataset designed to train AI models to extract structured information from crime-related text.
The dataset, called CrimeNER-db, contains more than 1,500 annotated documents drawn from public reports on terrorist attacks and press notes from the US Department of Justice. The researchers defined 4 broad entity categories and 21 fine-grained types — think locations, weapons, perpetrators, and victims — to cover the range of facts buried in these documents. They tested the dataset against both fully supervised models and zero- and few-shot approaches to gauge how well trained models generalize. The dataset is publicly available on GitHub.
Named-entity recognition — the task of pulling structured facts like names, places, and dates from raw text — is well-studied in general domains but thin on labeled data for law enforcement contexts. A dataset purpose-built for crime documents could give agencies, journalists, and researchers a foundation for building tools that surface relevant facts faster than manual review allows.
The DOJ publishes thousands of press releases a year; getting a model to reliably parse them is useful, though whether real agencies adopt academic NER tools in operational workflows is a different question entirely.