> We need a lot of data to do this well.
Yup. And once you've got a sufficiently large tagged error corpus, spelling correction will be as simple as a lookup: almost all misspellings (in terms of frequency) will be present in the corpus, and you can drop the (rather simplistic) algorithmic part.
