A scoping review of transformer and large language model approaches for word sense disambiguation in multilingual and low resource settings from 2019 to 2026

Abstract Background In natural language processing, word sense disambiguation (WSD) is still a significant difficulty, especially in low-resource, multilingual environments with few annotated resources. Contextual semantic representation and WSD performance have been enhanced by recent developments in transformer architectures and large language models (LLMs). However, there is still little evidence synthesis in multilingual and genuine low-resource situations, and research is still dispersed across architectures, datasets, and evaluation systems. Methods In order to map evidence on transformer and LLM-based Word Sense Disambiguation techniques published between 2019 and January 2026, this scoping study used the PRISMA-ScR methodology. Google Scholar database was searched for relevant literatures. Studies that used transformer or LLM-based techniques for WSD and reported evaluation criteria like accuracy or F1-score were considered eligible. A structured framework comprising model architectures, datasets, evaluation techniques, adaption strategies, and important discoveries was used to chart the data descriptively. Results Nine studies were included in the analysis. Gloss-aware transformers, poly-encoder architectures, multilingual transfer learning, morphologically aware fine-tuning, parameter-efficient adapters, and LLM-based reasoning were among the main topics. Model scale alone does not seem to have as much of an impact on WSD performance as architectural design and contextual modeling. Evidence for really low-resource languages was still scarce, though. Conclusion Transformer and LLM-based methods show immense potential for improving WSD in multilingual contexts. However, there are still gaps in robustness assessment, benchmark standardization, and the wider inclusion of low-resource languages that are underrepresented.