In this article, we tackle the issue of the limited quantity of manually\nsense annotated corpora for the task of word sense disambiguation, by\nexploiting the semantic relationships between senses such as synonymy,\nhypernymy and hyponymy, in order to compress the sense vocabulary of Princeton\nWordNet, and thus reduce the number of different sense tags that must be\nobserved to disambiguate all words of the lexical database. We propose two\ndifferent methods that greatly reduces the size of neural WSD models, with the\nbenefit of improving their coverage without additional training data, and\nwithout impacting their precision. In addition to our method, we present a WSD\nsystem which relies on pre-trained BERT word vectors in order to achieve\nresults that significantly outperform the state of the art on all WSD\nevaluation tasks.\n