A SURVEY OF TRANSFORMER AND LARGE LANGUAGE MODEL STRATEGIES FOR CLINICAL ABBREVIATION DISAMBIGUATION: A SPECTRUM OF PARADIGMS

The clinical narrative is a treasure trove of patient information, yet its value is often obscured by a high density of ambiguous abbreviations. For years, resolving this ambiguity was a classic supervised learning problem, held back by data scarcity and the need to create custom models for each case. This survey charts the significant advances of the field since the advent of Transformers and Large Language Models (LLMs). We argue that the central challenge has shifted from one of data acquisition to one of model generalization, knowledge tapping, interpretability, and practical deployment. Recent advancements are categorized into a conceptual spectrum, critically analyzing the trade-offs, and assumptions of each approach. These categories include (1) foundational fine-tuning paradigms, (2) sophisticated knowledge-augmented models, and (3) leading edge data-centric models capable of operating in low-resource environments. An analysis of these paradigms will also consider how they address emerging issues including interpretability and deployment. A framework will then be proposed to guide practitioners in identifying the most suitable methodology for their specific use-case. Finally, the future directions of research will be discussed including generative models, prompt engineering, evolving evaluation paradigms and the necessity of developing explainable systems. Ultimately, a roadmap for researchers aiming to build robust and trustworthy Natural Language Processing (NLP) systems for disambiguation task in the clinical domain will be outlined.