Word sequence kernels

• DOI: 10.5555/944919.944963
• Publication Date: 2003-03-01
• Authors: Nicola Cancedda; Éric Gaussier; Cyril Goutte; J.-M. Renders
• Keywords: Benchmark (surveying); Sequence (biology); Kernel (algebra); Similarity (geometry)

We address the problem of categorising documents using kernel-based methods such as Support Vector Machines. Since work Joachims (1998), there is ample experimental evidence that SVM standard word frequencies features yield state-of-the-art performance on a number benchmark problems. Recently, Lodhi et al. (2002) proposed use string kernels, novel way computing document similarity based matching non-consecutive subsequences characters. In this article, we propose technique with sequences words rather than This approach has several advantages, in particular it more efficient computationally and ties closely linguistic pre-processing techniques. present some extensions to sequence kernels dealing symbol-dependent match-dependent decay factors, empirical evaluations these Reuters-21578 datasets.