Modern data infrastructures rely on data virtualization services (DVS) to create many virtual versions of source datasets for diverse processing tasks. A single multivariate or multimodal dataset may serve classification, detection, segmentation, forecasting and other machine learning tasks. Data selection, participation, integration and preprocessing across tasks generate a large number of virtual datasets[Khan et al., SSC 2026]. Users also need to discover existing virtual datasets quickly to avoid duplicated effort. Therefore, automatically generating keywords for each virtual dataset is crucial for an effective DVS search engine.\n\nWe present a solution that leverages large language models (LLMs) to generate keywords. Five strategies are proposed to select different collections of textual descriptions for LLM processing. These strategies are applied to virtual datasets created in five case studies with distinct source data. Experiments involve three LLM generators and two embedding techniques. The results show that the performance of each strategy depends heavily on the quality of the text descriptions. Guided by these findings, we deploy a multi‑strategy keyword generation service (KGS) as part of an extended DVS, demonstrating that automated keyword generation is indispensable for high‑volume, dataset‑rich environments.\n\nReview: The study convincingly argues for the necessity of automated keyword generation in large‑scale DVS and offers a practical, multi‑strategy framework that can be adopted in future data management systems.