are not close if they've got no semantic similarity. The embeddings are device-generated dependant on the term or token neighbours in texts, not by a human meta-rationalization of what a word suggests. So „king“ can be each near to „throne“ and „checkmate“ based upon both of these contexts in Englis