Comment by shdon | Hacker News Reader

Comment by shdon | Hacker News Reader

Spelling correction rather than spell checking and suggestions really bugs me. As soon as you have a somewhat international audience (and given that this is the internet - you probably do) or have any non-English content, automatic correction can be an actively user-hostile "feature".

Especially on search engines, sometimes I am searching for words in another language, mixed-language results, or even doing an intentional search for misspelt words. In my native Dutch his first example would already cause problems: "spelling" means the same as in English, but "speling" means margin/give (the noun) and I wouldn't want one to be auto-corrected to the other. The fuzzy search for names on Facebook makes the search pretty much worthless especially in case you have an unusual variant of a common name in your search - you can be sure the actual person you're looking for is listed after half a billion bogus results.

The whole trend of second-guessing the user and saying "you tried to do A, but we assume you meant B, so we're gonna do B instead" is one of my biggest pet peeves.

Not only there. Automatic word replacement is my absolute least favorite iOS mobile feature. I cannot tell you how many times I painstakingly, painfully thumb typed exactly what I meant just to have it changed out from under me. And I have gone back to see nonsense messages where I know I didn't type what was sent. Very frustrating.

I think this could be corrected if the spell corrector factored in the time taken to write the word out. If the user takes longer than average to write the word they are probably deliberately spelling it out, compared to just mashing the screen to get the words through.

Agreed. I am a touch typist on a regular computer. I literally hate writing on the phone. I find it an utterly odious process. So much for user friendliness from my point of view.

Throw RNN at it, like this toy example http://karpathy.github.io/2015/05/21/rnn-effectiveness/ and make it figure out which spelling correction is the most probable.

You could even train it on past messages making it more tailored towards particular user.

I distinctly remember the time when all of a sudden many of my friends started sending me sms messages with really odd mistakes in them. I didn't jump on the smartphone wagon for a number of years so I just couldn't understand how these errors were being made at the time.

Agreed !! Actually, it's a real problem that's not limited to spelling. Look at Siri for instance, granted I haven't used it in a few years, back in the days it was impossible to use for most of my contacts. Indeed, I have both English and French contacts and it's a nightmare to use with Siri.

My Siri is in English because I figured it was more reliable for everyday us then in French. But my mom's contact is as 'maman', can't use Siri to call her. Or I have to try and pronounce 'maman' with an english accent and hope Siri know what I'm talking about. And that's an easy one, try a name like 'Jean Brentois'.... But I assume setting up Siri in French will lead to the same problem, but with my non French contacts now. I have not tried however.

So obviously I get the challenge, but it makes it very hard to use (I can't use Siri like tools in parts because of this). Especially since so many of my contacts are international.

EDIT: a quick fix for contact pronounciation might be to ask the user how he pronounces a particular contact when adding it.

Similarly, why not allow for a special syntax allowing for multiple languages in a sentence ? Say my first language is English, my second is French, and my third is Spanish. Why not something like that: "Let's go to the $$casa $demain $avec $toi, what do you think ?"

I almost added autocorrection to a translation pipeline once. Curated a subset of unambiguous safe cases to reduce the manual workload. I parked it away because I didn't see a way to reliably convey this was just lowering the number of manual corrections needed.

A bad part about user-end autocorrection is that the end user is looking only at the case they have in front, not the overall average — if they have a case where it worked very well, they'll start overly trusting it and failing at producing good results. If they have a case where it wasn't useful, they'll be annoyed by it and feel they are doing more than they should. In all cases someone is going to be unhappy.

If I can't expect all ends to understand the value I'd mostly just look for pipelines where a-c could be an internal process and there was no need for the data provider and end user to be aware that it's happening. But then, ascertaining this is a problem on its own.

Reply on news.ycombinator.com