SAFE: sensitive annotation finding and extraction from multi-type Chinese maps via hybrid intelligence and knowledge graph

Sensitive annotations typically contain key geographic elements or sensitive information vital for geographic information security. Considering the challenges of processing multi-type Chinese maps (e.g. topographic, administrative, and web maps), such as the scarcity of training samples for sensitive annotations due to their confidential nature and access restrictions, and the complexity of Chinese glyphs, this study proposes a hybrid intelligence approach named sensitive annotation finding and extraction (SAFE). By distilling expert knowledge into a knowledge graph, a human – machine collaborative sensitive sample synthesizer is developed, creating the first expert knowledge-integrated sensitive annotation dataset to address the scarcity issue. An improved Chinese sensitive annotation interpretation model is introduced, addressing the unique properties of Chinese sensitive annotations. A knowledge graph-driven extraction method then processes annotation results and determines saliency. Experiments validate the effectiveness of SAFE: in detection tasks, SAFE achieves an Hmean of 96.44%, approximately ten percentage points higher than the baseline model; in recognition tasks, SAFE attains an accuracy of 96.73%, which is 15.59% higher than the original algorithm. Finally, the knowledge graph-driven saliency determination ensures interpretability. SAFE not only provides crucial data and technical support for research in geographic information security but also advances the intelligent development of map interpretation.