Smart cities generate large volumes of heterogeneous data from Internet of Things (IoT) sensors, transportation systems, environmental monitoring platforms, public services, and other interconnected urban applications. Semantic annotation is the process of enriching such raw or heterogeneous data with machine-interpretable metadata and semantic descriptions, thereby enabling intelligent data integration, interoperability, discovery, and reuse. However, manual semantic annotation is time-consuming and difficult to scale, while fully automated approaches often suffer from semantic ambiguity, inconsistent annotations, limited adaptability to heterogeneous data sources, and the potential for hallucinated or unreliable outputs when Large Language Models (LLMs) are used.

This paper proposes an ontology-guided and uncertainty-aware LLM-based framework for semantic annotation of heterogeneous smart city data. The proposed architecture integrates data preprocessing, LLM-based semantic annotation, ontology-based concept guidance and validation, uncertainty estimation, and human-in-the-loop verification. Smart city data from multiple domains are first processed and transformed into a unified representation. An LLM then generates candidate semantic annotations, which are guided and constrained by relevant domain ontologies to improve semantic consistency and reduce invalid or hallucinated annotations. The framework further estimates the confidence and uncertainty associated with generated annotations, allowing uncertain cases to be selectively referred to human experts for validation. Validated annotations are subsequently used to construct or enrich a Smart City Knowledge Graph.

The proposed framework will be evaluated using heterogeneous datasets representing multiple smart city domains, such as transportation, environment, and IoT services. Experiments will compare multiple LLMs and annotation strategies, including standalone LLM-based annotation, ontology-guided annotation, and the proposed uncertainty-aware human-in-the-loop approach. Evaluation will consider annotation precision, recall, F1-score, semantic correctness, ontology consistency, completeness, human validation effort, and processing efficiency.

The expected findings are that ontology-guided LLM annotation will achieve higher semantic accuracy and consistency than unconstrained LLM-based approaches, while uncertainty-aware human intervention will improve the reliability of annotations by focusing expert effort on ambiguous or low-confidence cases. The study is also expected to demonstrate that combining LLMs, ontologies, uncertainty estimation, and selective human validation provides a scalable and reliable approach for semantic integration of heterogeneous smart city data. The resulting framework can support the development of interoperable Smart City Knowledge Graphs and improve data-driven decision-making across interconnected urban services.

Conference Title

An-Najah 1st Conference for Artificial Intelligence Applications

Conference Country

Palestine

Conference Date

Oct. 18, 2026 - Oct. 18, 2026

Conference Sponsor

An-Najah N. University

Additional Info

Conference Website