WILD-CRIME: a web-scale, semantically enriched knowledge graph for wildlife crime intelligence and analytics

Abstract

Wildlife crime is recognized as a key challenge under the United Nations Sustainable Development Goals, emphasizing its relevance in the modern world and the need for data-driven intelligence systems which can help with analysis and intervention. Wildlife crime impacts biodiversity, ecological balance, and conservation efforts worldwide, making it a critical problem. Unlike conventional crime domains, wildlife crime data is largely dispersed across unstructured sources, limiting its effective use by researchers, policymakers, and law enforcement agencies. The absence of standardized, semantically structured datasets hampers systematic analysis and evidence-driven interventions. In this paper we propose WILD-Crime (semantically enriched wildlife crime dataset) and an automated knowledge-centric framework that collates information from a variety of sources, filters for wildlife crime, performs semantic enrichment, consolidates and support analysis. The framework also includes a lightweight, novel expert-based model that classifies articles to wildlife-crime class with an improvement in F1-score by 4.2% against the best-performing baseline, RoBERTa. Relying on a single model proves insufficient for fully disambiguating nuanced cases where articles may be crime-related but not wildlife-specific, or vice versa. We validate the value of the dataset with downstream intelligence tasks, including the identification of wildlife crime hot-spots, species target patterns and the analysis of frequently used transit routes, complemented with API driven query and analysis interfaces designed for law enforcement agencies.