The Reasoning Visualization Model for Rational Crime Analysis and Fact Finding - A Study on Error Detection of Reasoning in Fact Finding

 

[ko]  Legal document automation : leveraging large language models in generating crime facts within criminal complaints

 

Park, Jeewon and Park, RoSeop.

Master's thesis, Legal document automation 

​

In response to recent legal reforms mandating the acceptance of all criminal complaints and intensifying investigative workloads, this study explores the automation of legal document drafting using Large Language Models (LLMs). Specifically, the research focuses on the generation of “crime facts” within criminal complaints involving loan and merchandise fraud.

To achieve this, expert-driven guidelines were developed to extract legally relevant information from criminal court judgments. Utilizing GPT-4 with Chain-of-Thought (CoT) and few-shot prompting techniques, a training dataset was constructed. This dataset was then used to fine-tune LLMs for generating criminal facts within complaints. Evaluation of the generated content showed high fidelity to legal standards, with BLEU scores of 76% and cosine similarity of 92% for loan fraud, and BLEU scores of 86% and cosine similarity of 94% for merchandise fraud.

The findings demonstrate the potential of LLMs to generate accurate and structured legal text aligned with existing judgments. This study offers a promising direction for enhancing the efficiency and accuracy of criminal complaint drafting in law enforcement, particularly under increasing investigative pressure.