Unsupervised Resolution of Acronyms and Abbreviations in Nursing Notes Using Document-Level Context Models

Automatic simplification of clinical notes continues to be an important challenge for NLP systems. A frequent obstacle to developing more robust NLP systems for the clinical domain is the lack of annotated training data. This study investigates unsupervised techniques for one key aspect of medical text simplification, viz. the expansion and disambiguation of acronyms and abbreviations. Our approach combines statistical machine translation with document-context neural language models for the disambiguation of multi-sense terms. In addition we investigate the use of mismatched training data and self-training. These techniques are evaluated on nursing progress notes and obtain a disambiguation accuracy of 71.6% without any manual annotation effort.