Classifying Biomedical Articles by Making Localized Decisions

We describe a system developed for the Categorization task of the Text Retrieval Conference (TREC) 2005 Genomics track, and experiments we conducted in the process of developing our system. Our research effort for this task explored the hypothesis that more accurate predictions could be achieved by considering only selected passages in the documents being processed. We investigated methods that involve (i) basing classifications on selected passages from test articles, and (ii) adjusting the classifier training process such that certain putatively relevant passages affect the learned model more than other passages. Whereas the first approach was effective at improving predictive accuracy in our experiments, the latter approach was not.