CrimeNER Demo: Named-Entity Recognition in the Crim... | AI Research

CrimeNER Demo is an AI-powered platform designed to help law enforcement agencies and researchers automatically extract and classify crime-related information from large volumes of documents. By utilizing Named-Entity Recognition (NER), the system identifies key details within text—such as the individuals involved, the types of crimes committed, and relevant logistical data—and organizes them into a structured format. This tool aims to bridge the gap in existing NER literature, which has historically focused on general news or biomedical texts rather than the specific needs of criminal investigations.

How the System Works

The platform operates through a four-stage pipeline: document preprocessing, optional fine-tuning, entity extraction, and postprocessing. Users can upload documents and choose to use pretrained models or provide their own annotated data to fine-tune the system for specific languages or unique case requirements. Once processed, the system identifies entities at two levels of granularity: "coarse" (broad categories like Crime, Actor, Agent & Agency, and Logistic) and "fine-grained" (specific sub-types like Homicide, Terrorist Organization, Law Enforcement, or Date). The final output provides users with the original documents featuring highlighted entities and, if requested, a structured JSON file containing the extracted data.

Granular Entity Classification

To ensure the extracted information is highly detailed, the platform uses a hierarchical classification system. The four coarse categories are further broken down into 21 fine-grained types. For example, the "Crime" category includes specific labels such as Terrorism, Fraud, Drug-related, and Sexual crimes. The "Actor" category distinguishes between criminal individuals and organizations, while the "Logistic" category captures essential details like locations, dates, and weapons. This two-level approach allows investigators to contextualize raw data effectively, moving from general identification to specific case details.

Performance and Flexibility

The platform is built upon the CrimeNER database, which contains over 1,500 annotated documents from real-world scenarios, including terrorist reports and U.S. Department of Justice documents. The developers tested several well-known language models, such as XLM-RoBERTa, DeBERTa, and BERT, to evaluate their effectiveness in identifying these entities. The results show that while different models produce varying levels of accuracy, the system is robust enough to handle diverse document types.

Limitations and Future Directions

While the platform provides a practical tool for forensic document analysis, the authors note some current limitations. There is occasional misalignment between the extracted entity positions and the visual text in PDF documents. Additionally, the current system relies on text-based analysis. Future development plans include integrating multimodal architectures to process text images more effectively and incorporating tools to detect AI-generated content, document manipulation, and other potential risks associated with digital document repositories.