Публикация

the unembedding layer's "default" output, given no further information, is "how likely is a word, in general." this is the zipf distribution. but the actual output says "i only care about how likely any word is relative to the other options."