I have Swype on my phone, which I like a lot: it lets me type by sliding my finger around the keyboard instead of tapping each letter separately. It might be able to do better, however, by considering word frequencies conditional on the last few words and their parts of speech. [1] I wonder if they're already working on this? Or if they tried it and it turned out to be a bad idea?

(Swype already biases me away from abbreviations and shortened forms that aren't in its dictionary, and this would make it favor proper writing even more by getting confused over elided articles and other minor words.)

[1] I'm pretty sure it's not already doing this, but to test it I decided to start a sentence with either "I like to", suggesting a verb, or "I like red", suggesting a noun. After typing the sentence beginning, I put my figure on 't' and gave it a complex wild 5-10 stroke wiggle. It tried its best to find something that matched, and these are the results:

I like to ____	I like red ____
refurbish	televisual
integer	travels
eunichs	telepathic
epistemic	Escondido
fizzled	telephonist
troubleshoot	rebalanced
radiochemistry	Yaroslavi
telepathic	Randolph
unorthodox	rainfalls
Yevtushenko	Tautvidas
telepathic	tranquil
transposition	rejig
unorthodox	hairbrushes
glycosidic	helminthic

Following "I like to" we got two words acceptable in the verb context and eleven acceptable in the noun context ('transposition' was bad in both). Following "I like red" we got twelve words acceptable in the noun context and one acceptable in the verb context ('Escondido' was bad in both). It doesn't look like they're paying attention to conditional probabilities.