1164421 - `im` should autocorrect to `I'm` instead of `in`

Status: VERIFIED FIXED

(Firefox OS Graveyard :: Gaia::Keyboard, defect)

Description

[Kevin Grandon :kgrandon]

One of the biggest painpoints right now that's shared between dogfooders is that our autocorrect *feels* off in for a wide variety of words. The biggest offender is probably `im` correcting to `in` instead of `I'm`.

Comment 1

[Kevin Grandon :kgrandon]

Recommending blocking here because currently typing on the keyboard is an extremely frustrating experience, and this is one of the most egregious offenders.

Comment 2

[Chris Lord [:cwiiis]]

Not to pile on, but if we are going to make this change, can we also stop autocorrect from changing 'jam' to 'ham' all the time? 'jam' is a word. This is really hampering my ability to write jamming-related messages.

Comment 3

[Jan Jongboom [:janjongboom] (inactive)]

Think this is a recent regression. Last time I dogfooded I doubt this was the case.

Comment 4

[Jan Jongboom [:janjongboom] (inactive)]

On Firefox OS 2.1 en-US keyboard Im becomes I'm.
On Firefox OS 2.2 en-US keyboard Im becomes In

Comment 5

[Kevin Grandon :kgrandon]

Thanks Jan. In requesting a regression window here if possible between 2.1 and 2.2.

Comment 6

[Yeojin Chung [:YeojinC]]

b2g-inbound Regression Window:

Last Working Environmental Variables:
Device: Flame 2.2
BuildID: 20141212003454
Gaia: 2e6a35dfa098aeb274835e812b18b5f6104622b0
Gecko: c73a1414a9b1
Version: 37.0a1 (2.2) 
Firmware Version: v18D-1
User Agent: Mozilla/5.0 (Mobile; rv:37.0) Gecko/37.0 Firefox/37.0

First Broken Environmental Variables:
Device: Flame 2.2
BuildID: 20141212014254
Gaia: 739dd45496e9d0f1806dc67faf0244c18c5dd3ac
Gecko: 6fe07f7f9c81
Version: 37.0a1 (2.2) 
Firmware Version: v18D-1
User Agent: Mozilla/5.0 (Mobile; rv:37.0) Gecko/37.0 Firefox/37.0

Last Working Gaia First Broken Gecko: Issue does NOT reproduce 
Gaia: 2e6a35dfa098aeb274835e812b18b5f6104622b0
Gecko: 6fe07f7f9c81

First Broken Gaia Last Working Gecko: Issue DOES reproduce
Gaia: 739dd45496e9d0f1806dc67faf0244c18c5dd3ac
Gecko: c73a1414a9b1

https://github.com/mozilla-b2g/gaia/compare/2e6a35dfa098aeb274835e812b18b5f6104622b0...739dd45496e9d0f1806dc67faf0244c18c5dd3ac


Caused by bug 971688

Comment 7

[KTucker [:KTucker][Inactive 3/4/2016]]

Jan, can you take a look at this please? This might have been caused by the work done for bug 971688.

Comment 8

[howie [:howie]]

Triage: regression, blocking.

Comment 9

[howie [:howie]]

Hi Jan, do you mind take a look? Thank you.

Comment 10

[howie [:howie]]

Hello David, do you have any suggestion here? Fixing bug 971688 is causing this bug, and I'd say the case here is more important/frequently used than the case in bug 971688.

I'd propose the back out bug 971688 if fine tuning the algorithm is not something we can easily achieve soon.

Comment 11

[David Flanagan [:djf]]

I don't think we should backout bug 971688. Jan argues convincingly in that bug that it can only make suggestions better by considering more possibilities.

I'm guessing that before 971688 landed, the autocorrect code was so busy looking at words that begin with "im" that it never even noticed "in" as a possibility. That was a bug, and 971688 fixed it so that "in" could be considered. (I should flash a 2.1 build and see what the suggestions for "im" are on that release; my guess is that they do not include "in").  So now that the autocorrection code correctly finds "I'm" and "in" as the two best possible matches for "im" the issue is that our weighting is off and we're picking the wrong one.

I'll take this and investigate whether a simple weighting tweak can fix it.

Comment 12

[David Flanagan [:djf]]

I've confirmed the hypothesis above. In 2.1, typing "im" suggests "I'm", "IMF" and "IMG". It doesn't even consider "in" as an option.

Comment 13

[David Flanagan [:djf]]

When we type "im", the prediction engine returns these results: 

[["in",21.743896484375],["I'm",16.929],["km",9.737735704393822],["um",9.118408203125]]

If I type "Im", it returns:

[["In",21.52645751953125],["I'm",17.099999999999998],["Km",9.737735704393822],["Um",9.118408203125]] 

I'm surprised that this would be different by such a small amount.

I'm surprised that we autocorrect to "in", given that "I'm" is also a highly-weighted candidate. Even if we don't autocorrect to I'm, we should at least not autocorrect to "in". So that is another minimal approach we could take here is tuning the threshold for deciding when to autocorrect instead of just suggesting.

Comment 14

[David Flanagan [:djf]]

Here's the relevant code in latin.js for how we decide to autocorrect:

  // Don't offer to autocorrect unless we're reasonably certain that the
  // user wants this correction. The first suggested word must be at least
  // this much more highly weighted than the second suggested word.
  var AUTO_CORRECT_THRESHOLD = 1.30;

21.53 / 17.1 is 1.26, which is less than this threshold, so this should not have been an autocorrect in the first place.  Something is going wrong here.

Comment 15

[David Flanagan [:djf]]

Nevermind. I misremembered. If the user's input is not a valid word, then we always autocorrect to a valid word. It is only when we're overriding one valid word with an autocorrection that we apply that threshold value.

Comment 16

[David Flanagan [:djf]]

Looking at the english wordlist now. Part of the problem here is that "in" is the 5th most common word in english, with a weight of 210. "I'm" is down below the 5000th word, and has a weight of 116. So it would take a pretty big adjustment to our weight to overcome that difference.

Here's a thought, though. English apostrophes are special, and users should never have to type them. So in an important sense, when the user types "im" we should treat that as having typed a valid word even though it does not exactly match any of the suggestions. If we did that, then the threshold thing above would come into play and I'm would be the autocorrection.

I'll see if I can do that in latin.js, which has the benefit of not needing to mess with predictions.js

Comment 17

[David Flanagan [:djf]]

'id' currently autocorrects to 'is'. By the logic of this bug, it should autocorrect to "I'd" instead. According to the wordlist "I'd" is much more common than "I'm" so autocorrecting "id" to "is" is even worse than autocorrecting "im" to "in".

Comment 18

[Chris Lord [:cwiiis]]

(In reply to David Flanagan [:djf] from comment #17)
> 'id' currently autocorrects to 'is'. By the logic of this bug, it should
> autocorrect to "I'd" instead. According to the wordlist "I'd" is much more
> common than "I'm" so autocorrecting "id" to "is" is even worse than
> autocorrecting "im" to "in".

for the record, I do actually find this really annoying too - especially if I capitalise the 'i' and it still auto-corrects to is (and same with Im auto-correcting to In).

Comment 19

[GitHub Autolander Bot]

Attachment: [gaia] davidflanagan:1164421 > mozilla-b2g:master

Comment 20

[David Flanagan [:djf]]

Comment on attachment 8609016 [details] [review]
[gaia] davidflanagan:1164421 > mozilla-b2g:master

This patch makes "im" autocorrect to "I'm" and "id" autocorrect to "I'd".

Rudy: can you review this, please, or find a reviewer? (Note that it is a 2.2+ bug)

Jan: if you have time to take a look, your feedback on this (or your review if you want to steal it from Rudy) would be great.

Comment 21

[Rudy Lu [:rudyl] (inactive after 2015/7/3)]

Comment on attachment 8609016 [details] [review]
[gaia] davidflanagan:1164421 > mozilla-b2g:master

I could rubber stamp on this patch since you understand latin.js much more than I do. :)

Thanks for your help to fix this blocker.

Comment 22

[David Flanagan [:djf]]

Comment on attachment 8609016 [details] [review]
[gaia] davidflanagan:1164421 > mozilla-b2g:master

[Approval Request Comment]
[Bug caused by] (feature/regressing bug #): 971688

[User impact] if declined: really frustrating autocorrect behavior where correct input "im" and "id" is turned into things that the user did not type.

[Testing completed]: yes, but the change needs some more dogfooding to make sure it has not introduced other unwelcome autocorrect behavior changes.

[Risk to taking this patch] (and alternatives if risky): the code itself is simple and not riskty: it simply re-arranges word suggestions from the suggestion engine so that for the user input `im` the word `I'm` comes before a more-common word `in` because if we ignore case and punctuation, the user typed the letters in `I'm`. This seems like it will always make autocorrection better. But we don't have a test suite for all possible autocorrections, so there is a chance that this change could break some other common autocorrection. I still think we should uplift it, however. I can't think of any simpler or safer way to fix this bug.

[String changes made]: none

Comment 23

[David Flanagan [:djf]]

The patch will not cherry-pick cleanly to v2.2. I will prepare a PR for that branch

Comment 24

[GitHub Autolander Bot]

http://docs.taskcluster.net/tools/task-graph-inspector/#jQsEm7pJRUW7pmgaD6TLOg

The pull request failed to pass integration tests. It could not be landed, please try again.

Comment 25

[David Flanagan [:djf]]

The tests are passing when I push the commit, so I assume these are infrastructure failures. Trying again.

Comment 26

[GitHub Autolander Bot]

http://docs.taskcluster.net/tools/task-graph-inspector/#z7atNthxSXqxURMk2sIyHA

The pull request failed to pass integration tests. It could not be landed, please try again.

Comment 27

[Kevin Grandon :kgrandon]

I think this is some autolander bug + issues since the tree closure yesterday, might have been resolved after a rebase, though you shouldn't need to do that. Sorry about that, let's manually land for now:  

https://github.com/mozilla-b2g/gaia/commit/99711e1fb0561d1413befed10d2eb87ce79e6c4f

Comment 28

[Kevin Grandon :kgrandon]

Please request uplift when you get a chance. Thanks!

Comment 29

[GitHub Autolander Bot]

Attachment: [gaia] davidflanagan:bug1164421-v2.2 > mozilla-b2g:v2.2

Comment 30

[David Flanagan [:djf]]

Comment on attachment 8609447 [details] [review]
[gaia] davidflanagan:bug1164421-v2.2 > mozilla-b2g:v2.2

This is the version of the patch that applies to 2.2, so moving the uplift request to this one.

See comment #22 for the approval uplift commenet

Comment 31

[Josh Cheng [:josh]]

Comment on attachment 8609447 [details] [review]
[gaia] davidflanagan:bug1164421-v2.2 > mozilla-b2g:v2.2

Thanks David!

Comment 32

[Josh Cheng [:josh]]

Requesting QA verifyme and exploratory testing for autocorrections on 2.2 and master.

Comment 33

[Ryan VanderMeulen [:RyanVM]]

v2.2: https://github.com/mozilla-b2g/gaia/commit/144673a413586f98b5e2c27b781c1a539611f754

Comment 34

[Shally Li（leave from Mozilla）]

This bug has been verified as pass on latest Nightly build of Flame v2.2&3.0 and Nexus 5 v2.2&3.0 by the STR in Comment 0.

See attachment: verified_v2.2&3.0.png
Reproduce rate: 0/6

STR:
1.Open Messages app.
2.Input "im" /"id" in message content area.
**it shows "I'm" /"I'd" correctly with highly-weighted candidate on keyboard.

------------------------------------------------------------------------
Device: Flame v2.2 build(Pass)
Build ID               20150526162504
Gaia Revision          15fa6e486b5c82956a8e8f8a99c39d297e8f0523
Gaia Date              2015-05-26 20:07:31
Gecko Revision         https://hg.mozilla.org/releases/mozilla-b2g37_v2_2/rev/663bf4703588
Gecko Version          37.0
Device Name            flame
Firmware(Release)      4.4.2
Firmware(Incremental)  eng.cltbld.20150526.201446
Firmware Date          Tue May 26 20:14:55 EDT 2015
Bootloader             L1TC000118D0

Device: Flame v3.0 build(Pass)
Build ID               20150526160204
Gaia Revision          8ca93673869a64e09ed6153c5402896822dfb253
Gaia Date              2015-05-26 19:31:37
Gecko Revision         https://hg.mozilla.org/mozilla-central/rev/1e4e369822ac
Gecko Version          41.0a1
Device Name            flame
Firmware(Release)      4.4.2
Firmware(Incremental)  eng.cltbld.20150526.195035
Firmware Date          Tue May 26 19:50:45 EDT 2015
Bootloader             L1TC000118D0

Device: Nexus 5 v2.2 build(Pass)
Build ID               20150526162504
Gaia Revision          15fa6e486b5c82956a8e8f8a99c39d297e8f0523
Gaia Date              2015-05-26 20:07:31
Gecko Revision         https://hg.mozilla.org/releases/mozilla-b2g37_v2_2/rev/663bf4703588
Gecko Version          37.0
Device Name            hammerhead
Firmware(Release)      5.1
Firmware(Incremental)  eng.cltbld.20150526.201116
Firmware Date          Tue May 26 20:11:30 EDT 2015
Bootloader             HHZ12f

Device: Nexus 5 v3.0 build(Pass)
Build ID               20150526160204
Gaia Revision          8ca93673869a64e09ed6153c5402896822dfb253
Gaia Date              2015-05-26 19:31:37
Gecko Revision         https://hg.mozilla.org/mozilla-central/rev/1e4e369822ac
Gecko Version          41.0a1
Device Name            hammerhead
Firmware(Release)      5.1
Firmware(Incremental)  eng.cltbld.20150526.195039
Firmware Date          Tue May 26 19:50:56 EDT 2015
Bootloader             HHZ12f

Comment 35

[Shally Li（leave from Mozilla）]

Attachment: verified_v3.0.png

Comment 36

[Shally Li（leave from Mozilla）]

Attachment: verified_v2.2&3.0.png

Comment 37

[Jan Jongboom [:janjongboom] (inactive)]

Comment on attachment 8609016 [details] [review]
[gaia] davidflanagan:1164421 > mozilla-b2g:master

Late f+