Word sense disambiguation in document space

We introduce a method for word sense disambiguation that uses an existing topical document map created with an unsupervised method (WEBSOM (Kohonen et al., 2000)) on a very large document collection. Results on the SENSEVAL-2 corpus indicate that the proposed method is statistically significantly better than the baselines and on a par with supervised methods. The method uses the document map as a representation of the semantic space of word contexts. The assumption is that similar meanings of a word have similar contexts, which are located in the same area on the self-organized document map. The results confirm this assumption. The benefit of the proposed method is that a single general purpose representation of the semantic space can be used for all words and their word senses.