Exploiting Frame Information for Prepositional Phrase Semantic Role Labeling

Abstract

AbstractSemantic role expresses the underlying relations that an argument has with its governing predicate. Prepositional phrase semantic role labeling concentrates on such relations indicated by prepositional phrases. Previously, the problem has been formulated as a word sense disambiguation (WSD) problem and contextual words are used as important features. In the past years, there has been a growing interests in general semantic role labeling (SRL). Therefore, it would be interesting to compare the previous contextual features with argument related features specifically designed for semantic role labeling. In experiments, we showed that the argument related features are much better than the contextual features, improving classification accuracy from 84.96% to 90.25% on a 6 role task and 71.47% to 75.93% on a 33 role task. To further investigate dependency between frame elements, we also introduced new features based on semantic frame that consider the governing predicate, preposition, and content phrase at the same time. The use of frame based features further improves the accuracy to 91.25% and 83.48% on both tasks respectively. In the end, we found that by treating prepositional phrases carefully, the overall performance of a semantic role labeling system can be improved significantly.KeywordsContent WordSemantic RoleWord Sense DisambiguationShare TaskContextual WordThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.