Alignment-HMM-based Extraction of Abbreviations from Biomedical Text

We present an algorithm for extracting abbreviation definitions from biomedical text. Our approach is based on an alignment HMM, matching abbreviations and their definitions. We report 98% precision and 93% recall on a standard data set, and 95% precision and 91% recall on an additional test set. Our results show an improvement over previously reported methods and our model has several advantages. Our model: (1) is simpler and faster than a comparable alignment-based abbreviation extractor; (2) is naturally generalizable to specific types of abbreviations, e.g., abbreviations of chemical formulas; (3) is trained on a set of unlabeled examples; and (4) associates a probability with each predicted definition. Using the abbreviation alignment model we were able to extract over 1.4 million abbreviations from a corpus of 200K full-text PubMed papers, including 455,844 unique definitions.