This paper proposes a new method for effectively summarizing tabular data in Korean administrative documents, the Tabular-TX pipeline, which is a theme-description structure-based tabular summarization method. To address the shortcomings of existing methods, which produce summary results that are difficult for humans to understand, Tabular-TX promotes deep tabular understanding of LLM through a multi-stage inference process and induces clear sentence generation using a reporter persona prompting strategy. In particular, it significantly improves readability by structuring the summary results into theme parts (adverbial phrases) and description parts (predicate phrases). It improves efficiency by utilizing in-context learning without the need for large-scale fine-tuning or label data, and experimental results show that it is a powerful and efficient solution for generating human-centered tabular summaries, especially in low-resource environments, by effectively processing complex tabular structures and metadata.
