While trawling some old Textkit threads I found a gem that got me thinking:
I thought we might open the floor for discussion on this and the good people of Textkit could contribute their own ideas. Specifically, the question I’m interested in is as follows:
Say a ‘dream engine’ were developed by modern AI methods (or at some point in the future), which can produce as much Greek (or Latin etc.) as we want, and all of it entirely ‘correct’ and coherent (so readable, in theory, on a large-scale). That is to say, the engine can speak and write Greek ‘perfectly.’
What further features would you want this engine to have, to be most useful to learners of Greek? Let’s place the restriction that we haven’t developed the technology for the engine to surreptitiously understand from, say, a student’s responses, where intrinsically the student is weak and what they need: instead, if, for instance, you don’t know a certain word, you’d have to tell the engine (e.g. select the word) - but the subsequent text could then be immediately regenerated to avoid that word except in the most comprehensible contexts (where it would be used repeatedly). Or, you might start by letting learners choose frequency levels and/or grammatical constructions, and have the text created restricted to that. As designers it’s our job to include the functionalities that would be most useful or desirable to students. I’m looking for brainstorming on what that might look like, from the starting-point of an engine capable of producing correct Greek. Your ideas could have a completely different angle to mine.
I’m most eager to hear your thoughts! Particularly those of you who most strongly believe in comprehensible input as the way forward, but really anyone who thinks classical languages could be improved by practice.
Hi, if we had that ideal engine, there would be two things that I would want:
Some way of knowing that it is producing reliable content. I’m thinking of the short film Sunspring — where AI was used to produce a film script, which is hilarious gobbledygook, and then actors played it out. If anyone hasn’t seen it, google it - it’s fantastic.
We can tell that the content is gobbledygook because it’s in our language. You would need some way of testing that in the engine — can it give you reasons for its constructions (imitated models from actual texts, grammar references, etc.)?
So the first question is figure out how we can know that we have moved beyond the Sunspring phase, and aren’t just listening to gobbledygook which technically complies with the patterns.
(Of course, in reality we are nowhere near the Sunspring phase today — I’m assuming, as you asked, that we are already at the ideal engine stage.)
An ability to use different subsets of the corpus of texts would be great. Today I will practice patterns drawn from Plato, tomorrow from Demosthenes.