Do the Wanikani team hate machine learning?

This is NOT a rant. I recently committed to Wanikani after joining last November. I tried it, moved on but couldn’t find anything better or anything that even came close. So I am committing to WK. However…

I do find it somewhat infuriating when a kanji I know very well but just mess up on spelling, I get sent straight back to the beginning for that kanji or vocab. eg saying one people instead of one person, or writing the reading instead of the meaning (and hitting enter before checking), etc.

Hasn’t the WK team heard of machine learning? I work in the field and there are many algorithms that can determine if you made a genuine mistake or you really haven’t learnt a kanji or piece of vocab.

Sorry its petty, but its happened to me on several occaisions now and I should have progressed to the next level if it wasn’t for a couple of minor mistakes stopping me from reaching 90% even though I have fully learned the items.



Yes. They hate machine learning and ice cream and everything that is good. It truly is a terrifying sight.

Maybe, you know, answering correctly is part of learning.


Install the ignore script, that’ll solve most of your problems


They do use an algorithm that will mark things correct even if you make a typo, assuming it’s close enough to the correct spelling (I’ve completely butchered the spelling of words before from typing too quickly but still gotten it marked correct, I believe the more letters in the correct answer, the more leniency from the algorithm)

I recommend installing the Override Script, you can ignore incorrect responses and be given the opportunity to put the correct answer/spelling next time around without affecting the SRS for that item

Also, you can include synonyms for some words in your reviews if you want, there are lots of ways to make your experience here on WK work for you! :crabigator:


As @MissMisc knwos, I make quite a few typos while typing. After installign that script, its pretty much made it a non issue when I am doing reviews. Would recommend


I work on similar stuff. Machine learning is good when you answer lots of times so that the system can learn about you…but when progressing on an item is just saying it right for around 10 times in its lifetime, which are all the SRS levels… I don’t think the AI can be really sure about it and anticipate these errors.

Typos are really annoying in WK. I installed the ignore script, by my favourite is: Double Check. I prefer this one much better than Ignore, which ignores the item and it will come back soon… I prefer to just retype what I did wrong when I know it was genuinely a typo.


I don’t know, it feels kind of like killing a fly with a cannonball to me.


Or maybe machine learning to determine whether a user intended to put the toilet seat down. :grin:


Oh and by the way, if you’re making “typos” that are really more about you not knowing how to actually write it in the properly way, that is not a typo ^^ So make sure not to override these type of answers (with the script above) or it will hurt your learning :smiley:


Lacking enough training data, as mentioned above, would be the biggest problem with it.

But even if you did have a big dataset, I wouldn’t welcome such a feature. Machine learning tries to approximate human intuition for what’s a right or wrong answer, but the precise line between accept or reject is often unpredictable. For example because I’ve seen WK accept “januar” for “january” before, I know it will also accept “februar” for “february” (it’s my pet shortcut to leave off final "y"s, the letter requiring the most finger motion across the keyboard to type). But with machine learning, who could say? It would depend entirely on an opaque training process.

In the end it’s much easier to take an extra 0.5 second per flashcard to make sure you typed something right. I’m intelligent enough to figure out the typo leniency rules on my own and take advantage of them without hurting the learning process; I don’t want a machine butting in on my turf.

context: I’m a computer science phd student who doesn’t personally work in machine learning but gets to see loads of talks on the subject.


Isn’t it kind of a stretch to assume that because they don’t use ML, they “hate” it? WK has been around for quite a while (in Internet years), long before the current ML/AI boom. Tofugu has like 4 developers.

Also, it’s worth mentioning, if they used ML to classify right/wrong answers, then

  1. Either they would have to send the answer to the server to be classified (which is REALLY slow when you have as many users as WK does), or
  2. The classifier has to be run in the browser, which would mean sending the pre-trained data (I forget the exact term) to the browser. Unless they use an algorithm that doesn’t need it… In which case you’re now running a classifier, often on a mobile device, in JavaScript.
1 Like

Hey, I also like to blame everything else but myself when I make mistakes as well.


I know a ton of people like that rage

Hate is maybe the wrong word I agree. As I said, its not a rant and I plan to see WK through to the end as we are moving to Japan soon or at least buying a second home there. Its just a niggle I have.

Sometimes after being asked several times before to provide the onyomi reading I just type onyomi when I see a new kanji or vocab only to find it asked me for the meaning. My bad for sure and yes taking an extra second will help, as it is doing given the number of times I fallen into that trap, its just surely there could be some warning in that circumstance.

As for training data, assuming most mistakes are common, the entire WK inputs made each day would quickly provide such a dataset if they were captured.

WaniKani isn’t a team full of developers. Stuff takes time and there’s other priorities. For the type of problem you’re having, a solution already exists: the override script. There’s no need to prioritize something that’s already solved.


I mean, that’s the whole reason it exists in the first place, right?

^^^ Sarcasm or not, this is true to an extent. :slight_smile:

“So, what do you do for a living?”
“I make your toilet watch you while you pee.”


I’d be more interested in machine learning for analyzing the kinds of mistakes users make in order to improve the mnemonics.

Machine learning might be overkill for the problem, but still helpful if WK decides to go public.

This topic was automatically closed 365 days after the last reply. New replies are no longer allowed.