Designing Semantic Kernels as Implicit Superconcept Expansions

Recently, there has been an increased interest in the exploitation of background knowledge in the context of text mining tasks, especially text classification. At the same time, kernel-based learning algorithms like Support Vector Machines have become a dominant paradigm in the text mining community. Amongst other reasons, this is also due to their capability to achieve more accurate learning results by replacing standard linear kernel (bag-of-words) with customized kernel functions which incorporate additional apriori knowledge. In this paper we propose a new approach to the design of ‘semantic smoothing kernels’ by means of an implicit superconcept expansion using well-known measures of term similarity. The experimental evaluation on two different datasets indicates that our approach consistently improves performance in situations where (i) training data is scarce or (ii) the bag-ofwords representation is too sparse to build stable models when using the linear kernel.