Very Large Neural Networks for Word Sense Disambiguation

The use of neural networks for word sense disambiguation has been suggested, but previous approaches did not provide any practical means to build the proposed networks. Therefore, it has not been clear that the proposed models scale up to realistic dimensions. We demonstrate how textual sources such as machine readable dictionaries can be used to automatically create Very Large Neural Networks which embody a simple local model and contain several thousands of nodes. Our results show that such networks are able to disambiguate words at a rate three times higher than chance, and thus provide real-size validation for connectionnist approaches to word sense disambiguation.