A basic requirement for natural language processing is some representation of lexical semantics -- either through the word's relations to others, through dictionary definitions, or as distributional representations in terms of its collocates from a large corpus. Combining these representations is not straightforward, as they have different requirements -- in particular, relational models of meaning rely on differentiating word senses, while distributional representations do not (cannot) make such distinctions.

By using Wikipedia, and exploiting its various structured/semi-structured sources of information, we built WikiNet, a multi-lingual network of concepts. WikiNet combines all the above mentioned representations for an inventory of concepts obtained from Wikipedia. We show how WikiNet was built, and how its combination of representations allows us to tackle the task of metonymy resolution.