Search
Browse By Day
Browse By Time
Browse By Person
Browse By Area
Browse By Session Type
Search Tips
ASC Home
Sign In
Introduction. This study uses NLP and machine learning to examine whether welfare-support vocabulary is structurally embedded in Japanese criminal sentencing, across 2,002 first-instance theft judgments.
Design/Methods. Judgments from Westlaw Japan (2010-2023) were vectorized via TF-IDF (5,000 features). Two interpretable models were trained: L1-regularized logistic regression with exact SHAP values, and fine-tuned Japanese BERT with three-fold Integrated Gradients (IG) attribution. A pre-specified 59-term welfare lexicon, organized into six semantic subcategories, was evaluated against importance rankings.
Results. Linear AUC = 0.913; BERT AUC = 0.951 (+3.7 points, concentrated in welfare-dense judgments, Mann-Whitney p = 0.007). Of 59 welfare terms, 14 fell in SHAP top 10% (2.4x chance). Three-fold IG corroborates at the BERT representation level: welfare tokens receive 2.23x (range 2.05-2.50x) the mean IG magnitude of non-welfare tokens (p < 10^-20 per fold), dominating the top-500 ranks at 5.21x chance (range 5.14-5.33x). The 24 suspension / 35 custody split is category-structured: abstract social-rehabilitation frames drive suspension, while concrete family-guarantor frames lean toward custody.
Discussion. As the first NLP/ML study of Japanese criminal sentencing, this shows judges use welfare concepts as structured discursive resources—not peripheral vocabulary. The subcategory-resolved findings ground the PSI reform debate, specifying welfare-language domains a formal system should systematize.