Overfitting in protein name recognition on biomedical literature and method of preventing it through use of transductive SVM

Machine learning methods have been used in research on protein name recognition. A classifier trained in a specific domain, however, could be overfit and so inflexible that it could be used only in that domain. We therefore developed a new corpus about breast cancer and investigated the flexibility of classifier trained on the GENIA (T. Ohta, 2002) corpus or the breast cancer corpus. To avoid overfitting we used the transductive support vector machine (SVM), and we evaluated the effect of transductive learning. We confirmed experimentally that the tranductive SVM prevented overfitting and yielded higher accuracies than the ordinary SVM did