Ontology-Based Feature Weighting for Biomedical Literature Classification

Ontology-based methods have been applied to biomedical literature classification tasks recently. By mapping lexically different but semantically similar words into features in the domain ontology that underlies the words, we can achieve at least two benefits: the dimensionality of the feature space can be reduced effectively, and the semantic information that underlies the lexical words can be incorporated into the classification process, leading to better classification accuracies. In this paper, we propose an ontology-based feature weighting strategy for the biomedical literature classification problem. We assign weights to the features into which the lexical words are mapped, according to the structure of the domain ontology, and further optimize the weights using cross-validation. Our experiments on MEDLINE-indexed journal abstracts demonstrate that our method can achieve a significant improvement on the classification accuracies, especially when the classification task is hard