A Knowledge-Based Framework for Urban Event Detection via Temporal Knowledge Graphs and Large Language Models

Urban agencies need event analytics that are accurate and explainable. We present Knowledge-based Event Detection (KED) method, which models events as temporal knowledge graphs to preserve participants, actions, places, and times. The specificity of KED is that it integrates LLM-assisted extraction of entities and relations, temporal knowledge graph construction with geocoded toponyms and proximity edges. In addition it incorporates semantic/contextual links allowing parametric community detection and adjustable event granularity. On Event2012 benchmarks, KED is competitive on long-tailed corpora and outperforms strong text baselines in balanced, lower-data regimes (a setting common for city-specific monitoring) while delivering auditable, graph-structured outputs that facilitate downstream situational awareness and decision support. By aligning geo-temporal structure with semantic context, KED advances explainable, transferable event detection for smart-city applications such as incident monitoring, mobility disruptions, and environmental alerts.