Arabic Gloss WSD Using BERT

Word Sense Disambiguation (WSD) aims to predict the correct sense of a word given its context. This problem is extreme importance in Arabic, as written words can be highly ambiguous; 43% diacritized have multiple interpretations and percentage increases 72% for non-diacritized words. Nevertheless, most Arabic text does not diacritical marks. Gloss-based WSD methods measure semantic similarity or overlap between context target that needs disambiguated dictionary definition (gloss word). gloss suffers from lack context-gloss datasets. In this paper, we present an gloss-based technique. We utilize celebrated Bidirectional Encoder Representation Transformers (BERT) build two models efficiently perform WSD. These trained with few training samples since they BERT were pretrained on large corpus. Our experimental results show our outperform recent WSDs when test them against same data used evaluate model. Additionally, model achieves F1-score 89% compared best-reported 85% knowledge-based Another contribution paper introducing benchmark may help overcome standardized