Context-based Word Definition Classification for Arabic: Leveraging Pre-trained Language Models

Abstract

This research addresses the challenge of context-based word definition classification, a critical task in Natural Language Processing (NLP), particularly for morphologically rich languages like Arabic. Using a dataset of 3,670 annotated examples, we propose a methodology combining pre-trained BERT embeddings with cosine similarity to evaluate the contextual alignment of ambiguous word senses. Our approach achieved 80% accuracy, outperforming traditional methods like Word2Vec and Lesk by a significant margin. These findings highlight the potential of pre-trained transformers for tasks such as machine translation, information retrieval, and conversational AI in low-resource languages. Limitations include computational demands and limited gloss coverage, which will be addressed in future work by expanding resources and refining classification techniques for dialectal and multi-lingual scenarios.