[3K1-OS-27a] OS-27

10:00 AM - 10:15 AM JST(1:00 AM - 1:15 AM UTC)

[3K1-OS-27a-05] A Robust Method for Structuring J-REIT Investment Property Information Across Diverse Tabular Formats

〇Mayuri Tanaka¹, Nobushige Doi¹ (1. Japan Exchange Group, Inc.)

Keywords:

LLM,J-REIT,Annual Securities Report,Table structure analysis,Real estate data

We propose a low-cost, robust method for structuring Japanese Real Estate Investment Trusts (J-REITs) investment property information in Annual Securities Reports using Large Language Models (LLMs). Because property tables vary widely across investment corporations, automated dataset creation is difficult. Whereas prior work relied on few-shot prompts tailored to each corporation, incurring substantial sample-preparation cost, our approach unifies these into a single common prompt. To address this issue, we propose a low-cost approach that standardizes prompts previously prepared on a per-corporation basis into a single common prompt, enabling the structuring of J-REIT investment property information at reduced cost. Based on a comprehensive survey of variables that may appear in J-REIT investment property information, we examine an output format that minimizes missing information while remaining easy for LLMs to transform into a structured form. Experiments indicate that the common prompt can be applied to arbitrary corporations, enabling a practical workflow for building a high-quality J-REIT property information dataset with reduced manual effort.