DEEP LEARNING-BASED AMHARIC TEXT CLASSIFICATION WITH SEMANTIC UNDERSTANDING

Abstract

This study explores the application of deep learning for Amharic text classification, addressing the challenges associated with its morphological complexity and limited computational resources. A Bidirectional Encoder Representations from Transformers -based architecture is employed using a two-stage process. First, a BERT model is pretrained on a large-scale Amharic corpus to capture semantic and morphological intricacies. This is followed by fine-tuning on labeled datasets for specific classification tasks, enabling both general language understanding and task-specific adaptability. Experimental evaluation demonstrates that the proposed approach improves classification performance, achieving an accuracy of 90.34% despite the scarcity of labeled data. The model effectively addresses linguistic challenges such as morphological ambiguity and contextdependent word meanings. This research contributes to Amharic natural language processing (NLP) by providing a deep learning framework with applications in sentiment analysis, spam detection, and personalized recommendation systems.