Incorporating Polarity in Lexical Resources

The felicity of contrast and parallel relations as well as structures such as ellipsis are in part dependent on the presence of lexical elements contrasting or sharing. Traditional dictionaries and lexical resources like WordNet code information about antonyms, yet this is often not the type of polarity found in real contrast or parallel examples. It is not yet clear how negative or positive polarity can be objectively defined and how it can be identified for individual examples. Further, many adjectives are interpreted differently depending on the context in which they are used. This ability of the context to select the sense of the adjective is parallel to meaning coercion in other adjective-noun combinations and should also be treatable in a computational linguistic framework. This suggests that using a dynamic approach like the Generative Lexicon (GL) to allow polarity coercion would be more appropriate, but GL currently doesn’t even code antonym relationships, let alone polarity. Further, the use of contrastive or parallel markers like but or and play a role in the final polarity interpretation. This ability seems to be unique among function words and isn’t currently handled in GL. Since the sense shifts are completely local, it is similar to other coercion processes. We conclude that this type of information, and coercion processes would perhaps be best handled within a framework like GL, but that the determination of the default polarities themselves might be extractable from already existing lexical resources like WordNet.