lesson that looks more like structured prediction

[jeisner]

I'd like to add a lesson somewhere where the distribution is over 8 sequences of length 3.
Each sequence is (triangle, square, pentagon).
But each shape in the sequence could be either solid or hollow.

Features are "solid triangle," "solid square," "solid pentagon",
"triangle matches square", "square matches pentagon".

This is analogous to CRF POS tagging. Maybe we can even bring out the connection via discussion or questions in the instructions.

[jeisner]

To handle this, we'd have to allow the "shape" and "fill" features in the lesson file to be equal-length lists.
The area for showing the shapes would have to be divided into subareas, and we render a shape of the appropriate size each subarea separately.
https://github.com/fmof/loglin/wiki/Available-Graphics

The elements of the list should be laid out in a horizontal row, to simulate text.
I suppose if we wanted columns as well, we could do a list of lists ...