Information-Extraction-Guided LLM Generation of WoT Thing Descriptions for Services

Urban service ecosystems expose heterogeneous APIs described through diverse and often unstructured documentation, hindering automated discovery, comparison, and composition. Although the W3C Web of Things Thing Description (TD) standard enables machine-interpretable service representations, TD generation remains largely manual and error-prone. Large Language Models (LLMs) enable automatic generation but suffer from hallucinations, omissions, and high token costs when processing raw documentation.This paper proposes a pipeline that combines Information Extraction and LLM-based TD generation. A GLiNER-based step extracts domain-specific functional and non-functional properties to produce compact, structured inputs that guide TD generation. Proof-of-concept experiments on heterogeneous real-world services show improved semantic completeness, reduced hallucinations, and lower token usage compared to baseline prompting. The pipeline supports semantic indexing within a multi-agent service orchestration framework.