Tapio Pahikkala, Sampo Pyysalo, Filip Ginter, Jorma Boberg, Jouni Järvinen, and Tapio Salakoski, University of Turku
In this paper, we introduce a new kernel function designed for the problem of word sense disambiguation. The presented kernel function employs two different types of positional information related to words present in the contexts of the words to be disambiguated. For each pair of words in two contexts, the proposed kernel takes into account both their distances from the ambiguous words and also the difference of their mutual positions. We apply the kernel to contextsensitive spelling correction with SVMs and show that it significantly outperforms other considered kernels.