Extending the Norwegian Wordnet using word embeddings.

Abstract

It is highly expensive to manually extend and maintain taxonomies such as wordnet; they tend to suffer from inefficient coverage. This thesis explores the possibility of automatically extending the Norwegian Wordnet, adapting a method based on distributional similarity and hierarchical structures. The actual extension of the wordnet is performed using word embeddings, which are estimated on PoS tagged lemmas from a Norwegian news corpus. These word embeddings are then the basis for a distributional similarity used in the adapted scoring function to score candidate hypernyms for a target word. A bi-product of is also a modification of the wordnet, where scripts to generate these modifications will be made available.