Parameter-efficient Word Sense Disambiguation Through Low- Rank and Decomposed Adaptation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Word Sense Disambiguation (WSD) remains a fundamental challenge in natural language processing (NLP), necessitating models that can resolve ambiguity across both general and specialized domains. While fine-tuning large pre-trained language models has yielded good performance gains, these approaches are often parameter-inefficient, requiring the update of millions of model parameters for task-specific adaptation. This research proposes a unified, parameter-efficient framework for WSD that addresses these limitations through innovative architectural modules and domain-specific optimization strategies. For general- domain WSD, we introduce an approach utilizing Low-Rank Adaptation (LoRA) modules coupled with a novel Part-of-Speech (POS) retrieval pipeline. This pipeline aligns target word contexts with WordNet lexical categories to construct optimized sentence-gloss pairs, enabling the model to outperform state-of-the-art benchmarks on Senseval and SemEval datasets using only 0.5% of the original model parameters. Extending this framework to the biomedical domain, where abbreviation ambiguity and context-dependent terminology present unique hurdles, we propose a multi-task learning architecture that jointly optimizes classification and ranking objectives. To enhance discriminative learning in this high-density semantic environment, we implement a category-aware balanced sampling strategy and structured negative sampling, which incorporates both hard and semantically similar negatives. By employing Weight- Decomposed Low-Rank Adaptation (DoRA), the model achieves superior adaptation while updating a mere 0.29% of total parameters. Extensive empirical evaluations on the MeDAL and MSH datasets demonstrate state-of-the-art performance, achieving a 92.9% macro F1 score and 98.1% accuracy, respectively. Systematic ablation studies further validate that the integration of multi-task structured objectives, balanced data sampling, and parameter-efficient adaptation yields a robust, scalable framework. By harmonizing these elements, this research establishes a new benchmark for WSD, proving that high- precision semantic disambiguation can be achieved with minimal computational overhead across diverse linguistic landscapes. This dual-domain approach provides a validated pathway for future advancements in specialized medical text applications and general- purpose natural language understanding.