Automated Classification of Moroccan Procedural Court Orders Using CAMeL-BERT

The digitization of judicial systems creates new opportunities for automating document management tasks that have traditionally relied on manual processing. This study addresses the automated classification of Arabic procedural court orders issued by Moroccan courts of first instance into seven official procedural categories, a low-resource legal document type characterized by highly standardized procedural language and limited publicly available resources. The primary contribution is a leakage-controlled evaluation demonstrating that frozen CAMeL-BERT contextual representations, combined with lightweight conventional machine learning classifiers, achieve highly accurate classification without computationally expensive transformer fine-tuning. The proposed pipeline integrates automatic structured extraction from heterogeneous Word documents with contextual embeddings and a comparative evaluation of five classifiers under a strict train–validation–test protocol. Experimental results show that Logistic Regression achieves 99.74% test accuracy (κ = 0.9966), while even the simplest baseline, Gaussian Naive Bayes, reaches 96.66%, highlighting the strong discriminative characteristics of Moroccan procedural court orders. An ablation study demonstrates that replacing TF-IDF representations with CAMeL-BERT embeddings reduces classification errors by 71.4%, while the complete classification pipeline further improves overall performance and reduces the remaining errors. These findings show that the highly regular linguistic structure of Moroccan procedural court orders enables computationally efficient and deployment-oriented classification using frozen contextual embeddings, providing a practical solution for Arabic legal document management and supporting future e-Justice applications in Morocco.