change(web): adjust autocorrect aggressiveness / thresholding

[jahorton]

I've been doing some tests. Initial experience is great! Some thoughts:

[...]
2. Inconsistent: [Testing] was not offered automatically when I typed [Tesring]. (See video)
[...]
4. [Givd] offers "Give" but doesn't autocorrect?
[...]
7. [frie] does not auto-offer [friend]

Originally posted by @mcdurdin in #16223 (comment)

In each of these cases, the lack of autocorrect is caused due to low "aggressiveness" settings.

For item 2 above - tesring - the following may all be presented at once:

• tearing (s -> a, also a single-char correction)
• texting (s -> x, r -> t)
• bearing (t -> b, s -> a)
• teasing (s -> a, r -> s)
• tasting (e -> a, r -> t)
• gearing (t -> g, s -> a)

If the sum probability of all non-testing suggestions exceeds the probability of the testing suggestion, it will not be auto-selected. That's part of the conservative autoselection process currently in place.

For item 4, givd:

• given
• gives
• giving
• find (g -> f, v -> n)
• five (g -> f, d -> e)
• giveaway

For item 7, frie:

• friends (exceeds friend!)
• friendly
• friendship
• friend's
• fried
• friendships
• fries

It's hard to know which is correct when we're not part-of-speech aware, let alone gramatically aware, etc. We may be able to aim for the "default", singular wordform if we were part-of-speech aware (#3058), I suppose, assuming we could note that the bulk of these options came from the same word stem - friend.

[mcdurdin]

I wonder if we should offer primarily corrections for autocorrrect. So then givd would have give (d -> e) as the only 1-letter-change option and I would think that would be a better default than leaving givd unchanged.

Of course the difference between completion and correction isn't a solid boundary.

What to do when a word isn't in the dictionary

When what has been typed is not in the dictionary, corrections should be auto-offered, if there is a nearby choice that is a real word. ("Nearby": this threshold will be more important for now because we don't have learning or user dictionary.) If there are multiple choices, pick the highest.

That seems like a better outcome than leaving a broken word, which is never a good outcome.

If we autocorrect when a word isn't in the dictionary, there are four possible outcomes:

• autocorrect chooses correctly (yay!)
• autocorrect guesses wrong -- the user has to choose an alternative word from the banner
• or the user manually corrects the word
• the word isn't in the dictionary (boo!), user has to choose the typed word from the banner

I think the threshold could be higher for autocompletion, but open to discussion on this.

[jahorton]

I think the threshold could be higher for autocompletion, but open to discussion on this.

So, "item 7" above - frei - would not autocomplete to either friends or friend due to requiring autocompletion.

I do like how it would help solve the first two, though. Shouldn't be too tricky to work something out for a corrections-only check that could help those two cases.

[mcdurdin]

So, "item 7" above - frei - would not autocomplete to either friends or friend due to requiring autocompletion.

Yeah, that's a shame but until we have learning / user dict I think we'll just have to hold off on all the goodness!

Shouldn't be too tricky to work something out for a corrections-only check that could help those two cases.

Cool!