Extracting Semantic Relations Using Dependency Paths
()unpublished.Abstract
Lexical semantic relations are useful in a variety of natural language applications, but to collect, update and maintain them by hand is tedious and costly. We experiment with the use of dependency paths, rather than regular expressions, as the formalism for representing patterns that may be indicative of a semantic relation between noun pairs. Our corpus is Wikipedia article abstracts, and we tag our training set using WordNet. We collect frequent dependency paths and using these as features, train two classification algorithms to predict the existence of a semantic relation among noun pairs in the test corpus, which predictions we evaluate using human judgment as the gold standard. Our experiments with hypernymy were successful; those with meronymy less so.