New Methods for Tackling SparsityProblems in WordNet

Abstract

WordNet is one of the premier natural language processing tools used to represent words in a semantically meaningful way. Since it was founded in the 1980s, successive research has tried to expand WordNet and the number of semantic relationships it includes. A number of research efforts have hypothesized that cross-part-of-speech links, in particular, should be a focus for further research. But, such extensions are constrained by the costs of human review required for the addition of new links. This research focuses on adding new ``action links'' that will connect noun synsets to their primary actions. We also show how to leverage new advancements in word embedding techniques to come up with link suggestions that can then be reviewed by human supervisors. Our simple techniques have reasonable performance in proposing correct links and the final links are of high quality. In fact, the greatest issue we find was the performance of the human reviewers.