Coveting Your Neighbor's Wife: Using Lexical Neighborhoods in Substitution-based Word Sense Disambiguation

Report, 2022

We explore a simple approach to word sense disambiguation for the case where a graph-structured lexicon of word sense identifiers is available, but no definitions or annotated training examples. The key idea is to consider the neighborhood in a lexical graph to generate a set of potential substitutes of the target word, which can then be compared to a set of substitutes suggested by a language model for a given context. We applied the proposed method to the SALDO lexicon for Swedish and used a BERT model to propose contextual substitutes. The system was evaluated on sense-annotated corpora, and despite its simplicity we see a strong improvement over previously proposed models for unsupervised SALDO-based word sense disambiguation.