How to Write a Spelling Corrector (2007) • Hacker News Board

Re: How to Write a Spelling Corrector (2007)

#21

Post by brian_cloutier » Fri, Sep 17, 2021, 3:09 PM UTC

mtreis86 wrote: post #17

One of my most common spelling mistakes is physical mistypes on the keyboard, yet no spell checker seems to account for keyboard layout and locality of keys, or for something like my hand being one position off on the board but typing all the keys relatively correct only positionally shifted.

spell checkers which account for keyboard layout absolutely exist. [1] is one example, but in general it's hard to believe that any trained model would fail to notice that certain mistakes are more likely than others, and it's hard to believe that any spell checker google releases these days would not be based on a trained model.

[1] https://ai.googleblog.com/2017/05/the-machine-intelligence-b...

Re: How to Write a Spelling Corrector (2007)

#22

Post by m463 » Fri, Sep 17, 2021, 3:22 PM UTC

mtreis86 wrote: post #17

One of my most common spelling mistakes is physical mistypes on the keyboard, yet no spell checker seems to account for keyboard layout and locality of keys, or for something like my hand being one position off on the board but typing all the keys relatively correct only positionally shifted.

I think the problem is that there are multiple layers of correction - both the physical and the spelling/semantic.

On ios I just had to turn off autocorrect, because I had reminders I jotted down quickly which missed the physical correction and were irretrievably corrected semantically into garbage.

now I find a note with "spekk" and I can figure out I meant "spell" (need a better example)

Re: How to Write a Spelling Corrector (2007)

#23

Post by tyingq » Fri, Sep 17, 2021, 3:31 PM UTC

mtreis86 wrote: post #17

One of my most common spelling mistakes is physical mistypes on the keyboard, yet no spell checker seems to account for keyboard layout and locality of keys, or for something like my hand being one position off on the board but typing all the keys relatively correct only positionally shifted.

I would guess that the iPhone autocorrect does this implicitly since it's ML trained.

Re: How to Write a Spelling Corrector (2007)

#24

Post by z3t4 » Fri, Sep 17, 2021, 3:32 PM UTC

mtreis86 wrote: post #17

One of my most common spelling mistakes is physical mistypes on the keyboard, yet no spell checker seems to account for keyboard layout and locality of keys, or for something like my hand being one position off on the board but typing all the keys relatively correct only positionally shifted.

The problem is that if you take an english word and change one letter it will likely become another valid word.

Re: How to Write a Spelling Corrector (2007)

#25

Post by wootest » Fri, Sep 17, 2021, 8:52 PM UTC

mtreis86 wrote: post #17

One of my most common spelling mistakes is physical mistypes on the keyboard, yet no spell checker seems to account for keyboard layout and locality of keys, or for something like my hand being one position off on the board but typing all the keys relatively correct only positionally shifted.

Many designs does this implicitly. At least one, the original iPhone keyboard autocomplete/suggestion algorithm, did it explicitly.

Ken Kocienda's book Creative Selection has a very good chapter on the algorithm being built piece by piece, but finding out words being created by surrounding keys was part of collecting all the candidates.

They even used this in marketing. One of the pre-original-iPhone-launch videos was focused just on the on-screen keyboard (probably because almost everyone thought it was a really kooky idea at the time), and used the example of explicitly pressing "O-U-Z-Z-A" but still getting "pizza" as the autocomplete because it was the closest recommendation.

One of the iOS versions a few years ago became incredibly fond of including the space bar and considering alternatives with slightly off key presses near the space bar split into two or more words. When you're using a language with a lot of compound words like Swedish, this yielded some almost postmodern corrections with one or more words often completely different (but close on the keyboard, of course). I don't know if this was a tweak to the manual algorithm going off the rails or an AI version that wasn't quite tuned well yet.