Catching metaphors

Metaphors are ubiquitous in language and developing methods to identify and deal with metaphors is an open problem in Natural Language Processing (NLP). In this paper we describe results from using a maximum entropy (ME) classifier to identify metaphors. Using the Wall Street Journal (WSJ) corpus, we annotated all the verbal targets associated with a set of frames which includes frames of spatial motion, manipulation, and health. One surprising finding was that over 90% of annotated targets from these frames are used metaphorically, underscoring the importance of processing figurative language. We then used this labeled data and each verbal target's PropBank annotation to train a maximum entropy classifier to make this literal vs. metaphoric distinction. Using the classifier, we reduce the final error in the test set by 5% over the verb-specific majority class baseline and 31% over the corpus-wide majority class baseline.