Sense Embeddings in Knowledge-Based Word Sense Disambiguation

In this paper, we develop a new way of creating sense vectors for any dictionary, by using an existing word embeddings model, and summing the vectors of the terms inside a sense’s definition, weighted in function of their part of speech and their frequency. These vectors are then used for finding the closest senses to any other sense, thus creating a semantic network of related concepts, automatically generated. This network is hence evaluated against the existing semantic network found in WordNet, by comparing its contribution to a knowledge-based method for Word Sense Disambiguation. This method can be applied to any other language which lacks such semantic network, as the creation of word vectors is totally unsupervised, and the creation of sense vectors only needs a traditional dictionary. The results show that our generated semantic network improves greatly the WSD system, almost as much as the manually created one.