Using Translations to Distinguish Between Homonymy and Polysemy

Abstract

Distinguishing between homonymy and polysemy can facilitate Word Sense Disambiguation (WSD), as WSD systems use the standard sense inventories that are excessively fine-grained and include many polysemous senses. We classify words as either homonymous or polysemous by building graphs of word senses as nodes and the semantic relatedness between them as edges. To find these edges, we build upon a previous hypothesis on the role of shared translations in semantic relatedness, and after developing new theorems, we propose methods of obtaining semantic relatedness information from translations of multilingual sense inventories. Additionally, we create a new homonym resource, as well as updating an existing one. These resources include lists of homonyms and their short definitions, manually mapped to their corresponding WordNet senses. We evaluate our methods on a balanced dataset of homonymous and polysemous words, and our methods achieve state-of-the-art results in the task of homonym detection.