torvaney

Absolutely. This will be an even bigger issue for people texting in a second or third language (which I understand is quite common among countries and communities where feature phone use is high).

As I understand it, one of the big advantages of T9 had over other, more sophisticated forms of predictive text is that each word in the dictionary can be encoded in close to 1 byte. Given the constraints faced at the time, T9 feels to me like it is close to a local optimimum.

Yes, it's interesting that the T9 seems to have made a slight improvement on my "naive" alphabetical layout. You might even be able to brute-force a solution where the alphabetical constraint is imposed.

Hello! OP, here. I agree that the task is training for this specific subset of English writing, which isn't ideal.

For this task, I was primarily interested in whether the task would work at all. My assumption is that given we can optimise for these texts, we could optimise for more representative datasets, too. Perhaps you think this is a weak assumption?

Do you think testing on a sample of totally different texts from different authors would be more convincing?