How to make Autocorrect stronger?

[gd4c]

If I type multiplr, the suggestion multiple is not in the autocorrect color, and doesn't correct on space/enter. How do I make autocorrect more eager with such mistypes?

Autocorrect confidence: x1.5
Adaptive confidence: off
Learn from my fixes: tried both
Mistype tolerance: 200%
Fix several missed keys: Eager

How can it be improved?

I am trying things like lowering mistype % to let the keyboard by less uncertain, and increasing the needed confidence a bit and making it "Careful". But even with settings like mistype tolerance at 1, vpud suggests your first. and coid suggests could Both I aim at void

I guess what I am trying to get the keyboard to respect is:

• I will almost never type void with 3 or 5 letters. Worst that could happen is that I will "raise" my thumb in my mind from the desired o and back down for the i but will actually just remain 2x time stuck to the screen because the thumb won't lift fast enough. Exceptions of maybe +/- one letter in long words with 10+ letters. This concerns typing not gliding.
• If I am too hectic, I can misplace 2 or 3 letters in a four letter word. Mostly on an adjacent letter but occasionally, if I type too fast, I may even just land on the correct quarter of the keyboard. void is Lower left > Upper right > Upper right > Upper/lower left. I am not trying to make things hard for you. This is just the best I can do on a phone when I am excited!
• Significant respect to actual dictionary words. What on earth is "mil riple" to be the bold one, when multiple is right there?

[dgesteves]

You're already at the most eager setting. x1.5 is the bottom of the Autocorrect confidence slider (range), and the engine clamps the final value to that floor after adaptive confidence and typing rhythm are applied (clamp), so there is nothing lower. Raising it makes autocorrect stricter, not looser.

What the number means: with Learned correction gate on (the default), the gate estimates the chance that the top candidate is the word you meant, and the slider sets the bar at c/(c+1). x1.5 needs 3 in 5, the default x4 needs 4 in 5 (Confidence.kt). With the gate off, the top candidate's score has to beat the runner-up's by that factor.

Bold isn't the autocorrect pick. The strip bolds its top-ranked chip, and the autocorrect colour goes only on the word space will actually commit (KeyboardScreen.kt), so in your milriple screenshot nothing cleared the bar. mil riple is a two-word split from Add missing spaces; turning that off removes split chips from the strip.

What I'd change for how you type:

• Fix several missed keys: Careful (the default). Both modes apply same-length fixes with several wide taps, like milriple → multiple. Eager also applies fixes that add or drop a letter, which is how coid can become could; Careful only offers those on the strip (filter). That matches "I will almost never type void with 3 or 5 letters".
• Mistype tolerance below 200%. At 200% any key within 2.5 key widths of a tap counts as a slip, against 1.25 at 100% (touch model), so more common words become cheap readings of what you typed and can outrank the word you aimed at.
• Learned correction gate off is worth an A/B: x1.5 then means a plain 1.5-to-1 lead over the runner-up, and a word that both the word list and your own typing history pick is applied regardless of the slider.

Change one at a time and leave the slider at x1.5. I read this against the v0.6.4 source but couldn't run the engine, so I can't say which change flips multiplr for your taps. "Assume the length is right, allow several far misses" isn't a mode today; an issue with your vpud and coid examples is the way to ask for it.

[gd4c]

Thanks a lot @dgesteves That was really helpful (although I won't claim I understood it perfectly).

Raising it makes autocorrect stricter, not looser.

I understand and that was my first choice. I just thought maybe raising the confidence threshold helps throw out the lower alternatives instead of making them all acceptable thus no autocorrect.

Careful and Eager

One letter for any word or does this allowance for missing letters increase with word length?

Mistype tolerance

Yeah. I lowered that back closer to the default. Thanks

an issue with your vpud and coid examples is the way to ask for it.

Reported here yesterday.

---

The help popup for the confidence slider says that corrections are applied if two sources agree. I hadn't had any other source beyond a few dozen personal dictionary words. So, I extracted the word lists from the language packs and imported them as custom dictionaries as a "second" source. And it worked! For most tried English words. Autocorrect.

For Arabic, I tried many things including the same as above and 3rd-party word lists but not even once, did it autocorrect @wasi-master

[dgesteves]

At v0.6.4 (latest tag) it's one letter at any length, in both modes. Autocorrect only applies a word at most one real edit away (a letter added, dropped or swapped), plus any number of slips onto nearby keys (rule). The search allows two edits from 5 letters up (budget), but those only reach the strip. Careful won't mix that one edit with slips; Eager will.

Imports aren't a second source: they join the word lists, and the other source is your own words, learned ones plus Android's personal dictionary (sources). With Learned correction gate on, agreement is just one input; "always applies" only holds with it off (here). So English most likely improved because the extra words gave it better candidates.

Arabic stops earlier: a word any list holds is never corrected (check), apart from narrow cases like a spelling ranked below 50,000 whose fix is 1,000x commoner (escape). Per the code's own note, in the 210,000-word Arabic list a one-letter slip is usually a real word (note), so those stay as typed. Extra Arabic lists tend to make it worse: each adds known spellings. Check Autocorrect this language is on for Arabic and drop the extra lists; beyond that, the known-word check itself would need to change. I read this in source but couldn't run it.

If the English part is sorted now, marking the answer helps others who land here.

[gd4c]

Imports

It was the exact same file. So no extra words, just 2 of everyone. That was the only variable in that step so it has got to be it somehow.

Arabic list a one-letter slip is usually a real word

I tried with non-existent one-letter-off words (think the Arabic equivalent of mulriple). Still no autocorrect to the Arabic multiple

If the English part is sorted now, marking the answer helps others who land here.

Of course. But I'd still love for wasi-master to take a look.