Syllables and other String Kernel Extensions

During the last years, the use of string kernels that compare documents has been shown to achieve good results on text classification problems. In this paper we introduce the application of the string kernel in conjunction with syllables. Using syllables shortens the representation of documents compared to a character based representation and as a result reduces computation time. Moreover syllables provide a more natural representation of text; rather than the traditional coarse representation given by the bag-of-words, or the too fine one resulting from considering individual letters only. We give some experimental results which show that syllables can be effectively used in text-categorisation problems. In this paper we also propose two extensions to the string kernel. The first introduces a lambda-weighting scheme, where different symbols can be given differing decay weightings. This may be useful in text and other applications where the insertion of certain symbols may be known to be less significant. We also introduce the concept of ‘soft matching’, where symbols can match (possibly weighted by relevance) even if they are not identical. Again, this provides a method of incorporating prior knowledge where certain symbols can be regarded as a partial or exact match and contribute to the overal similarity measure for two data items. We also give an overview of other kernels for comparing strings that have evolved over the past few years. In the appendix we give a detailed description of a method for computing the string kernel efficiently. [email protected] This text is mainly an extension of [STST02] written by Craig Saunders and John Shawe-Taylor together with the author. Sections describing other string kernels and the efficient computation of the string subseqence kernel have been added, few passages altered to fit into the new context, few errors of the underlying paper corrected, and some hints to recently published literature added.