Presentation Information
[1Yin-B-20]Multistep Reasoning Approach with Large Language Models for Interpreting Deadline Provisions in Real Estate Loan Contracts
〇Megumi Hori1, Yukihiro Azuma1 (1. IBM Japan Systems Engineering Co., Ltd.)
Keywords:
LLM,Extraction
Loan agreements for a real estate securitization scheme contain not only information on the real estate properties but also various deadlines, such as loan execution, installment repayment, and reporting dates.Some deadlines recur, for example quarterly, and both the deadlines and their reference dates are often scattered across multiple pages, leading to time-consuming manual checks. This paper presents a multi-step workflow that uses the structured output capabilities of large language models (LLMs) to extract and interpret these deadline-related descriptions, along with a web system that allows users to view the supporting evidence sentences. The approach enables extraction of deadlines in calendar-date form even when defined across pages, such as "within five business days from the reference date", and provides explainability by displaying the evidence. Quantitative evaluation experiments and validation by real estate division staff confirm that the proposed method is effective in improving the efficiency of deadline management.