Large Language Models for Quantitative Analysis of Statutory Stringency and Its Correlation with Sexual Assault Crime Rates
Abstract
Sexual assault (SA) crimes are among the most important issues for our current society to solve, with these crimes in particular having long-term consequences on victims and their communities. To solve this issue, one approach is to make laws and penalties for these crimes stricter. However, there have not been any studies on the relationship between stricter penalties and lower crime rates, especially about SA crimes, in a quantitative manner. The biggest barrier to conducting a quantitative study on the strictness of laws and statutes is the lack of a standardized measure of how strict a legal text is. This paper explores the use of Large Language Models (LLMs) to generate such a measure. The work utilizes LLMs to generate ratings of the relative strictness of SA crime penalties for each state in the United States of America and compares this to SA crime rates for each. Another question addressed in this work is how well LLMs can extract information and categories from statutes, and which LLMs perform better for these tasks. The paper concludes that among the top open-source LLMs available today tested for this work, Mistral AI performs the best for categorizing statutes according to relative strictness. The work suggests that there is a weak correlation between stricter statutes and lower crime rates for SA, indicating that directing efforts to other solutions, such as grassroots movements, might be more fruitful.
