I recently learned about the following phenomenon: Google BERT word embeddings of well-known state-of-the-art models seem to ignore the measure of semantical contrast betwee
