Deep Learning Judicial Decision Prediction for Afaan Oromo Multi-Defendant Cases

Abstract

The goal of Judicial Decision Prediction (JDP) is to predict court rulings based on the accusations in a criminal case. Criminal cases may feature a single defendant or numerous defendants; the latter type of case is referred to as a "multiple defendant case". Multi-defendant cases are difficult since many defendants are accused of a single offense. By automating the process using a real dataset, the use of a deep learning prediction model can reduce complexity and boost the quality of decision-making. Using a newly collected dataset of 1005 criminal cases from the Oromia Supreme Court (OSC), we constructed a deep learning-based model for predicting court outcomes. Our dataset comprises 3101 instances with ten variables, three of which are target variables. We utilized Microsoft's Optical Character Recognition (OCR) to extract text from the scanned text that was in image format. After that, we applied several data preparation techniques such as cleaning and consolidating before performing feature extraction. Models have been created for both the Judgment on Conviction (JOC) and the Penalty, which are the two main components of the final judicial decision in Ethiopia. JOC involves deciding whether a defendant is guilty or not, and a penalty is a sentence imposed on the defendant who is found guilty. Deep learning models including Bidirectional Long Short Term Memory (BiLSTM), Hybrid Convolutional Neural Network and Bidirectional Long Short Term Memory (CNN-BiLSTM), Bidirectional Gated Recurrent Unit (Bi-GRU), and Convolutional Neural Network (CNN) with Word2vec and FastText feature extraction are used for both JOC and penalty prediction. With performance parameters like accuracy, recall, f1- score, and precision as well as a confusion matrix, the performance of the models is assessed under stratified 10-fold cross-validation. The performance evaluation findings showed that the CNN model with FastText outperforms other models for judgment on conviction prediction. It achieved an f1-score of 99% and an accuracy of 98.74%. In terms of penalty prediction, hybrid CNN and BiLSTM with FastText outperform other models with 73.29% accuracy and 74% f1- score. While this study uses deep learning to predict judicial decisions in multi-defendant cases, it would be preferable to use similar techniques on a larger dataset that might include all possible punishments.