MEDLINE abstracts classification based on noun phrases extraction

Abstract

Many algorithms have come up in the last years to tackle automated text categorization. They have been exhaustively studied, leading to several variants and combinations not only in the particular procedures but also in the treatment of the input data. A widely used approach is representing documents as Bag-Of-Words (BOW) and weighting tokens with the TFIDF schema. Many researchers have thrown into precision and recall improvements and classification time reduction enriching BOW with stemming, n-grams, feature selection, noun phrases, metadata, weight normalization, etc. We contribute to this field with a novel combination of these techniques. For evaluation purposes, we provide comparisons to previous works with SVM against the simple BOW. The well known OHSUMED corpus is exploited and different sets of categories are selected, as previously done in the literature. The conclusion is that the proposed method can be successfully applied to existing binary classifiers such as SVM outperforming the mixture of BOW and TFIDF approaches. © 2008 Springer-Verlag.