Automatic Resolution of Ambiguous Abbreviations in Biomedical Texts using Support Vector Machines and One Sense Per Discourse Hypothesis

We present an algorithm to disambiguate abbreviations in Medline abstracts using Support Vector Machines (SVM) and one sense per discourse hypothesis. In contrast to other work using SVM for natural language disambiguation which always depend on handcrafted training and testing data, the algorithm provided here automatically extracts the training and testing data through searching long form of abbreviation in the texts and using one sense per discourse hypothesis. In the phase of testing, we also use this hypothesis to unify the outputs of the classifier via majority voting. The results obtained in our experiments demonstrate that SVM is a promising technique for abbreviation disambiguation and using majority voting in the phase of testing can improve the accuracy from 82.35% to 84.31%.