How the model produces this

1. One encoder reads the letters of the word, once.
2. A boundary head marks where to cut, which gives split. A syllable head marks where syllables end, which gives syllables.
3. Classifier heads name the grammar (part of speech, case, number, tense, mood, aspect, voice, gender, polarity, root type) and how sure the model is. A multi-label head lists extra tags, in features.other.
4. A small decoder writes the root lemma and the ipa, letter by letter.
5. After the model, a bundled word list sets source and checks the root's type and part of speech. Cuts the model was unsure about become the alternatives.