A Knowledge Recall Framework for Biomedical Abbreviation Disambiguation Using Pre-Trained Transformer Models

Background: The inherent polysemy of abbreviations in biomedical text presents a persistent challenge in natural language processing systems for Word Sense Disambiguation. Current approaches often struggle to fully leverage the vast, latent knowledge embedded within pre-trained transformer models. Method: This paper introduces a novel deep learning framework designed to resolve this ambiguity by directly querying the internal knowledge of such models. The core technical contribution is theRecalled Concept Embedding (RCE)method. RCE algorithmically retrieves a high-fidelity vector representation for each potential meaning of an abbreviation by directly probing the model's latent semantic space. This RCE is then fused with contextual embeddings and structured ontological features from the Unified Medical Language System (UMLS) to construct a comprehensive semantic profile. By formulating Word Sense Disambiguation as a semantic type classification problem, the proposed framework leverages a conceptually novel interaction that bridges the gap between passive fine-tuning and generative prompting. Results: The framework demonstrates state-of-the-art performance on two public benchmarks: the MSH WSD dataset and a subset of the MeDAL dataset, achieving weighted F1-scores of 99.4% and 98.5%, respectively. Conclusion: These results underscore that directly recalling latent knowledge is a more effective and robust technique than conventional fine-tuning or complex feature engineering. This study proposes a new paradigm in actively utilizing the deep, implicit knowledge within pre-trained models, offering a path towards more efficient, robust, and generalizable biomedical WSD systems.