Ideas

[Disiok]

- Cluster character by rhyming, and remove generation and limit scope of first character to subset of rhyming ones
- Polishing: separate meaning modeling and pronunciation, additional pass to fix ping ze and rhyming
- Repeated training: downwards and upwards
- Restrict to nouns
- Fix embedding for end character

Add position information to word2vec, for example 3rd char on every line