Using graph-kernels to represent semantic information in text classification

Abstract. Most text classification systems use bag-of-words represen-tation of documents to find the classification target function. Linguistic structures such as morphology, syntax and semantic are completely ne-glected in the learning process. This paper proposes a new document representation that, while includ-ing its context independent sentence meaning, is able to be used by a structured kernel function, namely the direct product kernel. The proposal is evaluated using a dataset of articles from a Portuguese daily newspaper and classifiers are built using the SVM algorithm. The results show that this structured representation, while only partially de-scribing document’s significance has the same discriminative power over classes as the traditional bag-of-words approach.