Deep Learning Based Criminal Case Judgement and penalty Prediction for Afaan Oromo

Criminal case judgment and penalty prediction has recently gained significant attention in the field of legal informatics. While numerous studies have explored deep learning techniques for high-resource languages like English, there remains a lack of research targeting low-resource languages such as Afaan Oromo. This study addresses this gap by proposing a deep learning-based system for predicting criminal case judgments and corresponding penalties in Afaan Oromo. A new dataset was prepared from scratch, collected from the Ilu Aba Bor Zone Supreme Court and selected First Instance Woreda Courts. The dataset includes 3,892 labeled cases for judgment prediction, categorized into two classes: Yakkama (Guilty) and Bilisa (Not Guilty), and 1,742 records for penalty prediction comprising 39 distinct penalty categories. To address the research questions, four deep learning architectures LSTM, BiLSTM, GRU, and BiGRU were implemented and evaluated. Among them, the BiLSTM model demonstrated the best performance for both tasks. For penalty classification, it achieved an accuracy of 98% and an F1-score of 98%. For judgment classification, the BiLSTM and BiGRU models attained the highest results, both with an F1- score of 86% and accuracy scores of 85.6% and 85.4%, respectively. This research contributes a newly developed and publicly unavailable dataset for a low-resource language, explores multiple deep learning models for dual classification tasks, and presents a comparative evaluation of their performance. The findings indicate that deep learning approaches, especially BiLSTM, are effective in handling legal text classification in Afaan Oromo and show potential for integration into real-time legal support systems, provided further testing and optimization are conducted.