A Study on Machine Learning in Criminal Sentencing Assessment: Centering on the Sentencing Standards of Article 57 of the Taiwan Criminal Code

Sentencing is among the most contested tasks in criminal adjudication, as judges must balance retribution, rehabilitation, deterrence, and proportionality within broad judicial discretion. In Taiwan, Article 57 of the Criminal Code mandates consideration of multiple statutory factors, yet substantial inter-case variation persists. This study proposes a framework integrating large language models (LLMs) with traditional machine learning (tree-ensemble regressors) to predict imprisonment duration. Using a question-driven labeling pipeline derived from Article 57, LLMs extract sentencing factors from unstructured judgments and transform them into transparent, auditable feature vectors for regression models. The results show that stronger LLM architectures improve downstream prediction: under the same 72-question and Gradient Boosting Regressor with Optuna setting, Qwen3-30B-A3B reduced MAE by 42.8% and RMSE by 41.6% relative to DeepSeek-R1-Distill-Qwen-32B. Additionally, recursive feature elimination reduced the questionnaire from 72 to 28 items (61.11% reduction) with only a 1.00% increase in MAE and a 1.57% increase in RMSE. The proposed system serves as a judicial decision-support tool providing reference values and consistency checks while preserving due process, transparency, and judicial discretion.