Classification of narcotics indictment texts: a comparison of Naive Bayes and Logistic Regression

Abstract

Narcotics crimes in Indonesia generate a massive accumulation of legal judgment documents, particularly at the Medan District Court, which handles the highest case load. Analyzing primary indictment documents manually is time-consuming and prone to subjectivity, necessitating a computational approach. This study aims to classify primary indictment texts of narcotics cases into light sentence (label 0) and heavy sentence (label 1) classes using naïve bayes and logistic regression methods. The methodological novelty of this research lies in the determination of an objective binary threshold calculated from the median sentence duration of 500 sample documents. Text processing involves text preprocessing and feature extraction based on Term Frequency-Inverse Document Frequency (TF-IDF) weighting, followed by evaluation using a stratified 10-fold cross-validation scheme. Experimental results show a significant performance gap, with logistic regression dominating by recording a consistent classification accuracy (CA), f1-score, and recall of 91.6%, along with an AUC of 0.978 (excellent classification category). Conversely, naïve bayes only achieves an accuracy and recall of 73.0% and an AUC of 0.784, producing a high misclassification rate of 135 documents due to the strict assumption of word independence and limited training data. The superiority of logistic regression is driven by ridge regularization, which effectively reduces the overlapping information in legal data matrices densely packed with bound phrases. Thus, the TF-IDF-based logistic regression model is highly recommended as a computational tool to assist in the binary classification of narcotics court judgment documents.