An Unsupervised Clinical Acronym Disambiguation Method Based on Pretrained Language Model

Clinical concept normalization plays a vital role in extracting information from clinical documents, specifically clinical notes. The presence of abbreviations within these texts has a substantial impact on concept normalization performance. To address this issue, our objective is to propose an unsupervised learning approach for automatic disambiguation of clinical abbreviations. Our proposed pipeline consists of three main modules: a) Prompt-based contextualized token prediction, b) embedding-based semantic similarity calculation , and c) candidate ranking and selection. Our method achieves accuracies of 73.6% and 74.3% on two distinct clinical datasets, respectively. An ablation study demonstrates the beneficial contributions of all modules within our pipeline for acronym disambiguation. Our study highlights the effectiveness of the prompt-based unsupervised method in the clinical acronym disambiguation task, showcasing its potential application within existing clinical NLP pipelines for entity concept normalization.