There are two kinds of automatic semantic distance measures: lexical-resource-based measures of concept-distance and corpus-based distributional measures of word-distance. Each approach has its own unique advantages and limitations. We present a hybrid measure of semantic distance computed from distributional profiles of concepts that retains the advantages of both and avoids their limitations. We use thesaurus categories as coarse-grained concepts.

Antonyms convey a sense of both distance and closeness simultaneously, and only a handful of automatic measures exist. We use our thesaurus-based concept-distance measure in combination with simple affix patterns to compute word-pair antonymy. Lastly, we show how these contrasting word pairs can be classified as positive and negative evaluative words, thereby creating a high-coverage semantic orientation lexicon.

This is joint work with Graeme Hirst from University of Toronto. Parts of this work were carried out in collaboration with: Bonnie Dorr, Philip Resnik, Cody Dunne (University of Maryland); and Iryna Gurevych, Torsten Zesch (Technische Universität Darmstadt).

The talk will be followed by a round table discussion at 11:15 am, in the same room, where you can exchange ideas with the visitor about your own projects.