Abstract
The task of interpreting tabular data semantically is central to domains ranging from biomedical research to finance, where structured tables must be linked to rich domain knowledge. Ontology-driven methods, such as the iterative approaches could provide a structured, interpretable framework for entity recognition and semantic enrichment. However, these methods often struggle with handling ambiguous terms, evolving taxonomies, and rapid domain shifts. Simultaneously, recent advances in Large Language Models (LLMs) have demonstrated remarkable flexibility and context-aware reasoning capabilities, making them attractive complementary tools. In this perspective paper, we review the ontology-driven semantic analysis landscape, highlight its limitations, and discuss how LLM-based agents can address these challenges. We then outline a conceptual framework for integrating LLM reasoning with ontology-driven pipelines, enabling dynamic ontology extension, robust entity disambiguation, and scalable cross-domain adaptation. By bridging symbolic knowledge structures with the adaptive intelligence of LLMs, we propose a pathway to more flexible, accurate, and future-proof semantic interpretation of tabular data.