Researchers Release First Large Crime-Domain Dataset for Extracting Information from Law Enforcement Reports

Extracting actionable intelligence from law enforcement documents—press releases, crime reports, and attack summaries—remains a challenging task for agencies trying to organize and analyze large volumes of unstructured text. A new research effort introduces CrimeNER, a specialized dataset and benchmark to help train machine learning models for this purpose.

The Problem#

Most off-the-shelf named-entity recognition (NER) systems are trained on general news text and perform poorly when applied to crime-related documents, which use distinct vocabulary and entity types not well represented in mainstream datasets. This gap has left agencies without adequate tools to automatically extract structured information like suspect names, weapon types, locations, and organizations from their internal reports.

Introducing CrimeNER#

Researchers have assembled CrimeNER-db, a corpus of more than 1,500 manually annotated documents drawn from public records: terrorist attack reports and press releases from the US Department of Justice. The dataset defines 4 broad crime entity categories and breaks those down into 21 specialized entity types that capture the granularity needed for law enforcement analysis.

The team tested the database by fine-tuning existing NER models and evaluating their performance in both fully supervised and zero/few-shot learning scenarios—techniques that measure how well models generalize to unseen crime documents without extensive labeled examples.

Next Steps#

The dataset has been made publicly available on GitHub to support further research in crime-domain NLP. The work establishes a foundation for building AI tools tailored to the specific information extraction challenges that law enforcement agencies face in practice.