<p>Overall system architecture of the study.</p>

The exponential growth of legal documents in Ethiopia has created an urgent need for efficient and accurate automated classification systems tailored to the country’s unique linguistic and legal contexts. This study presents an enhanced deep learning approach for multi-class classification of Ethiopian legal texts by leveraging deep neural architectures integrated with attention mechanisms. In this study, we proposed Hybrid deep learning algorithms. CNN, CNN + BiGRU and CNN + BiLSTM with and without an attention-based neural architecture that dynamically focuses on the most important textual features. The proposed hybrid architecture integrates hybrid models with an attention mechanism, allowing the model to capture contextual dependencies which is crucial in legal language understanding. Extensive experiments on a curated dataset of Ethiopian legal texts across multiple classes demonstrate significant improvements on multiple hybrid models like, CNN, CNN + BiGRU and CNN + BiLSTM integrated with Attention mechanism. Model performance is evaluated using an evaluation metrics of precision, recall, F1-score, and accuracy, with evaluation strategies like, 10-fold cross-validation and 80:20 train-test-split which showed notable gains in classification effectiveness. The experimental results show that CNN + BiLSTM with Self-Attention scores 99.53, 99.25, 99.37 and 99.38 for precision, recall, F1-score, and accuracy respectively with 80:20 train-test-split and 99, 98.99, 98.99, and 98.98 for precision, recall, F1-score, and accuracy respectively with 10-fold cross-validation.