Different types of annotations can be used. Components of nn system could be generated automatically depending on the input.

Proposition: To avoid having to restructure all code, simply append extra input to the end of the data file and create vectors, embeddings, layers for each extra input (tab-separated as before)

Extra input specified in scripts/config.py:

name, type, max size (NA when sequential)

list of inputs, w/ name, type (seq or nseq) and a maximum size)

extra_inputs = [("s_morph_de", "seq", 90)]