This is the most critical thing missing for my alternations experiment. I want to match pictures to words differently across participants, but once a picture-word association has been chosen for a participant, I need the related word to get the related picture so that singulars and plurals share semantics as expected.
I'm not sure the best way to do this. A general way would be to add something that knows how to set per-participant ground truth and then train on that. But that is probably too general; constraints and training styles probably vary too much.
Another way would be an object like SampleFrom, something like MapFrom. Whereas SampleFrom takes a string from a list, MapFrom would take a string from a dictionary. The key used to access the string would be a page id, so it would mean "get me the corresponding value to the value page 1 used." The experimenter could write in the other half of the correspondence, by putting in correlated text in these two pages for instance.
So I could have two banks of pictures. The singular bank would be a list. The plural bank would be a dictionary from singulars to plurals. When a string was sampled from the singular bank, the page id it went to would become the key to the plural in the plural bank. When MapFrom is called, it accesses that plural.
It's not terribly pretty but it's a lot prettier than all the other ways I can think of. I'll have to think about more complex versions we might want and ways to make it more flexible. At least this would mean you could map from different pages in the same page, and different aspects of a page.