Contextual Embedding for Word Sense Disambiguation

Word Sense Disambiguation (WSD) requires users to identify the correct meaning of polysemous words through their surrounding text. The process of WSD becomes difficult for Marathi, which serves as a morphologically rich language with low resource availability, because there are few annotated corpora and its complex inflectional system, and lack of strong lexical resources. The paper shows a Marathi WSD system which uses linguistic processing to create transformer-based semantic models for its operation. The system first performs normalization and morphological analysis and lemma extraction before it creates contextual embeddings through the use of IndicBERT. The system fetches candidate gloss meanings from IndoWordNet, which it then transforms into semantic vectors to compare with sentence-level context vectors through cosine similarity for sense determination. The experimental results show that transfer learning effectively disambiguates both basic and advanced sentence constructions. This proves its usefulness for Indian language semantic analysis. The method demonstrates that NLP systems for low-resource domains achieve better results when they combine linguistic expertise with deep contextual embedding techniques.