Reducing Large Semantic Graphs to Improve Semantic Relatedness

Abstract

In the previous research the authors developed a family of semantic measures that are adaptable to any semantic graph, being automatically tuned with a set of parameters. The research presented in this paper extends this approach by also tuning the graph. This graph reduction procedure starts with a disconnected graph and incrementally adds edge types, until the quality of the semantic measure cannot be further improved. The validation performed used the three most recent versions of WordNet and, in most cases, this approach improves the quality of the semantic measure.