Using Candidate Exploration and Ranking for Abbreviation Resolution in Clinical Document

In biomedical texts, abbreviations are frequently used due to their inclusion of many technical expressions of some length. Accordingly, appropriate recognition of abbreviations and their full form pairs is an essential task in automatic text processing of biomedical documents. However, unlike the biomedical literature, clinical notes have many abbreviations without their full forms available in the text or without standard definitions in dictionaries due to the nature of the documents. This causes difficulties in adapting traditional approaches for abbreviation disambiguation such as classification among fixed candidates or pattern-based definition extraction. Because of this reason, we consider the task as search problem and propose an approach with two steps: a) exploring possible full form candidates from various resources and b) choosing most acceptable one among retrieved candidates by ranking. To discover full form candidates and their features, we exploited external academic resources such as MEDLINE and UMLS as well as the clinical note corpus itself. To rank the candidates properly based on human criteria, we adopted Rank Boost, one of the learning-to-rank models developed from information retrieval and machine learning communities. Experimental results show the suggested two-step approach is promising for this kind of task.