Simple Markov chain based corrections and suggestions?

[nh2]

Hi,

for a long time I'm searching for a good open-source alternative to SwiftKey. FUTO looks quite good but there are still many inputs where SwiftKey does much better suggestions out-of-the-box.

• Today I typed I'm not sure bit; SwiftKey reliably corrects that to I'm not sure but and FUTO doesn't.

"I'm not sure bit" is just a very unlikely sentence that doesn't make sense in English. I suspect any Markov chain model should easily correct that.

Is it documented somehow in high-level terms how exactly FUTO's engine works? I see in the settings that there's a slider to weigh dictionary-based and neural network based suggestions, but couldn't find any details.

What does "dictionary" mean here? Does that use a Markov chain?

If not, I'd like to propose this to the project (perhaps as another model to combine in from). Markov chains

• are fast
• relatively cheap to store
• can do filtering and smoothing (predicting the next word or correcting multiple past ones)
• can be easily combined to support multiple languages simultaneously

[nh2]

Some other examples I'm collecting that SwiftKey does well and FUTO doesn't (all typing, no swiping):

• I'm not sure bit -> I'm not sure but
  • Autocorrect Threshold T = 0.7 does not help here
• I'm not sure id -> I'm not sure if
• The dog doesn't like the cst -> The dog doesn't like the cat
  • On Autocorrect Threshold T = 0.7 FUTO corrects it to cats instead of cat; not sure why, cat seems like a much better suggestion.
• II would like to eat some pizza with my friends and drunk some xoke
  • On T = 0.7 FUTO corrects xoke to comedy here; code is shown as a suggestion but not the default one. drunk some comedy makes no sense.
  • SwiftKey gets it right.
• Tyoinf -> Typing
  • FUTO T = 0.7 corrects this to Thing
  • SwiftKey gets it right.
• Ibwisg -> I wish
  • FUTO corrects this to Obstinent -- makes no sense
  • SwiftKey gets it right
• I can nitbswe -> I can not see
  • FUTO corrects this to I can bother
  • SwiftKey gets it right
• I don't like what is being done here because it is inoeofrssinal -> I don't like what is being done here because it is unprofessional
  • FUTO corrects inoeofrssinal to `impressive
  • But when disposing the keybaord and bringing it back, unprofessional is the first suggest but not applied when pressing space.
    • So not sure how FUTO makes suggestions; it seems nondeterministic or irreproducible
• If duxks -> If ducks
  • FUTO corrects this to If sucks -- that's not English
  • SwiftKey gets it right