This is driving me insane.

Additionally 'hehe' gets autocorrected to 'hebe'.

it hurts, @luke, any suggestions?

After discussing with Tim, the simplest way is to add terms to the wordlist case-by-case since there hasn't been a mechanism to keep the list up to date yet. I can take this.

Comment on attachment 8642311 [details] [gaia] luke-chang:1094833_add_lol_hehe_to_wordlist > mozilla-b2g:master Hi Jan, I'm not sure if the procedure of modifying wordlist/dict I made is correct or not. Could you please take a look? Many thanks.

Comment on attachment 8642311 [details] [gaia] luke-chang:1094833_add_lol_hehe_to_wordlist > mozilla-b2g:master Hi, Can you add them with f=1 instead of f=0? f=0 is reserved for profane words, so they're handled differently. Tim, do you think we can start adding words to the dicts directly? We should have a way of adding new ones regardless of what AOSP dictionaries think, but if we do a upstream sync changes might be lost. How should we handle that?

It was already being updated and merged manually I think. See bug 902120. https://github.com/mozilla-b2g/gaia/commit/3b763d688d8b2993acc2e4ab566750b406715735 https://github.com/mozilla-b2g/gaia/commits/master/keyboard/dictionaries/en_us_wordlist.xml https://github.com/mozilla-b2g/gaia/commits/master/dictionaries/en_us_wordlist.xml We could continue to do that if rebasing our changes isn't an issue from time to time. If it is we could figure out something more automated.

There's a problem that it doesn't work if I set them to f=1. However, it does work if I try f=80. I guess the algorithm ignores the last few candidates with lower frequency if there's already some candidates with much higher value. In this case, "Lil" is set to f=88 so "lol" with f=1 seems to be ignored. To make it work, I can adjust the frequency manually to fit this requirement but I don't think it's a right way. Tim, do you have any suggestion?

Hi Jan, You're more familiar with autocorrect function. Could you give us some advice? Thanks.

Is the engine auto correct correct words just because it's frequency is not high enough? What's the point to keep these words in the blob if the engine do this?

f=0 gets bumped to f=100 or something internally, so that explains it. Anyway, for such short words f=1 is indeed not good. There's a good reason we take frequency info in account, because if there's a short word with f=1 and another short word with close key distance (i=>o) with f=88 then chances are just way bigger you mean the f=88 word. Example of the real dict: if you type edy, would you expect 'edy' or 'edu'? Edy is valid word (f=1), edu is also valid word (f=89). I guess edu would be good suggestion based on this. But now that I think of it f=0 is what we want to use. We shouldn't suggest the words in autocorrect, but when user types them we want to keep them, and in that sense it's exactly the same as profane words. So I'll revoke my r- here.

Master: https://github.com/mozilla-b2g/gaia/commit/1c9a169f4d2ad956a7f1533d2432dbe3cc238b6d

This seems to have broken GU10 on gaia-master, which will also break on b2g-inbound most likely. Compounded by the fact that the pull request did not have a valid try run attached (likely some infra issue), this is getting backed out. Backout: https://github.com/mozilla-b2g/gaia/commit/947a6dd6703118214d3d69cd27fad9b609121fe4 Gu10 error: https://treeherder.mozilla.org/logviewer.html#?job_id=404624&repo=gaia-master

I found the cause and fixed in the patch above. Waiting for the test.

test passed: https://treeherder.mozilla.org/#/jobs?repo=gaia&revision=6426618e1418b3534f3af460c96f9441f3730fdc land on master: https://github.com/mozilla-b2g/gaia/commit/a695f46955eec8b8f29651efa53a6d0b04caf5fe