Learning to identify antonyms

1. Introduction Antonymy is a common lexical relation that is intuitively clear (if not always easy to define) for humans, but challenging for machines. In Natural Language Processing, antonymy detection has applications in tasks of understanding language, such as paraphrase detection, question answering, and textual inference. For that reason, WordNet (Fellbaum, 2005) includes some antonymy annotation; but the relation is relatively rare, and a quick manual inspection reveals that there are many more antonym pairs (including very common ones) than those shown in WordNet. It also becomes clear that the relation is somewhat vague; masculine and neuter, for example, are listed as antonyms in WordNet, but many native speakers of English would not intuitively consider these words antonyms. The presence of these antonyms is WordNet makes an interesting resource for supervised learning, however, which opens the possibility of trying to automatically extend the annotation. Identifying pairs of antonyms in corpora is the task I propose in this paper. The approach is to train a classifier to distinguish between sets of sentences that contain pairs of antonyms from sets of sentences that do not. The intuition behind it is that antonyms are often used contrastively in the same sentence. The highest-performing classifier obtains 84% accuracy. Because the literature on this task is limited, it is hard to rigorously compare the performance of the classifier with existing published results; however, it does seem to in line with, and perhaps above, results reported from other research. An important aspect of this approach is that knowledge-rich feature engineering was deliberately avoided. The reason for this is that an approach to detecting antonymy is much more productive if it fits into a general framework for learning other lexical relations, such as synonymy, hypernymy, entailment etc.. Therefore, whereas linguistic knowledge about the antonymy relation can be useful for the task (for example, features indicating the presence of morphemes such as un-or dis-would clearly be informative), I instead opted for a distributional approach, where the features are