Antonymy and Contrast Relations英文电子图书.pdf

Antonymy and Contrast Relations Jennifer Spenader Gert Stulp j.spenader@ai.rug.nl g.stulp@student.rug.nl Artificial Intelligence, University of Groningen Abstract In this paper we report on experiments to determine if antonymy is a good predictor of contrast, using 124 texts from the British National Corpus and the antonymy relations for adjectives recognized by WordNet. Further, we considered whether antonyms are key arguments in the inferences that license contrast. We looked at the frequency of both indirect and direct antonym pairs in contrastive sentences marked with but and sentences with antonym pairs without but. Antonyms and but co-occurred in only 1% of the 218,017 sentences studied. However, in 81% of but marked sentences with true antonyms pairs, the feature the antonyms described was a source of the contrast. In the non- but marked sentences, antonymy alone was a poor predictor, licensing a contrast in only 15% of the cases. We also found that direct antonyms are better predictors of contrast than indirect antonyms, and certain antonym pairs, like same-different, are consistently good predictors. These results could be used to find unmarked contrast relations with antonyms alone. 1 Introduction contrast is considered a major category in information organization, ev- idenced by its inclusion in almost all major taxonomies of rhetorical rela- tions (e.g. Hobbs 1990, Martin 1992, Mann and Thompson 1988, Asher and Lascarides 2003). Since Lakoff (1971), two types of contrast are generally recognized, denial of expectation (1) and semantic opposition (2): (1) It’s raining but I’m taking an umbrella (2) John is tall but Bill is short. (Lakoff 1971: 133) A common analysis for (1) is that the first conjunct implies something that an implication of the second conjunct co