I don't think this should have been accepted as an answer


#1

It came up with the “You made a small typo” text box, but the answer isn’t anywhere near the correct answer, so I don’t know what happened…

https://screenshots.firefox.com/csN65M6eKjpNQ67a/www.wanikani.com

(敗れる・やぶれる)
I answered to be released,
The correct answer was to be defeated.


#2

I was just coming to post something similar. I think the crabigator was a bit too lenient on me!


#3

It might not be ‘anywhere near the correct answer’ for you, because you understand the meaning of the words, but to a computer those answers seem pretty similar. It’s only a few letters that are different and WK thinks it’s close enough.


#4

Yes it shouldn’t have. But the spellchecker can’t differentiate between a typo and a wrong answer (short of an AI), and blacklisting every wrong answer for every vocab and kanji would be like infinite work.


#5

You got 11 out of the 14 characters right… I’m sure that’s on the edge of the range of acceptance, but it’s not that far off.


#6

Yea, that’s the biggest issue with this, is that for short words it can be too severe but for long words can be too lenient.


#7

They use the Levenshtein string distance algorithm. I’ve found that Jaro-Winkler distance works a bit better: more forgiving of typos in short words, and less lax on long words.

I’d love to see the result of an algorithm that takes keyboard layout into account, like more forgiveness for neighboring keys, and less forgiveness for distant ones.


#8

Hmm. I suppose WaniKani could benefit from checking that your answer isn’t closer to another word in the database than it is to the correct answer. It could still be kept client-side by pre-computing similar words server-side (ones that are close enough to produce a false positive when typed correctly), to then also check the Levenshtein distance to when checking your answer.


#9

^^^
I know of an algorithm that computes a signature for likely intended words, but it’s probably too big (due to data) to use on the client side. It would be cool, though.

Actually, what would be really useful is if WK started collecting wrong answers so they could be used as a dataset for evaluating answer-checking algorithms. Of course, it could also be used for identifying commonly confused vocab & kanji.


#10

Yeah unfortunately, as others have said, the accepted answers get more lenient as answer length increases. Biggest problem I have ran into is mixing up transitive/intransitive for longer words and having it accepted.

As for how to deal with this, I personally use a reorder script on my reviews. This way, if I undeservingly get the meaning correct, I can just answer the reading wrong on purpose.


#11

…yeah, it happens…