I agree that there's no reason that these properties are axis-aligned. Isn't the normal approach to look at whether word2vec('king') - word2vec('man') ?= word

I agree that there's no reason that these properties are axis-aligned. Isn't the normal approach to look at whether word2vec('king') - word2vec('man') ?= word2vec('queen') - word2vec('woman') There's an entertaining investigation of this applied to Game of Thrones at https://towardsdatascience.com/game-of-thrones-word-embeddings-does-r-l-j-part-1-8ca70a8f1fad https://towardsdatascience.com/game-of-thrones-word-embeddin...!