Review Scheduling Mechanics: A Discussion


#1

Overview

For a description of WaniKani’s SRS Mechanics, see: How long between reviews

After using this site for a few months, I’m quite happy with my progress, but I have had some lingering doubts about certain correlation-related aspects of the review scheduling system.
The mechanics lead to the following behaviors (not technical definitions):

  • Grouping: Lessons are usually completed in batches and tend to be reviewed together. As long as they are reviewed together, then these items will remain in the same review group. Review groups provide context to items in the same group as well as in relation to other groups.
  • Unlock Ordering: As we know, radicals unlock kanji, which unlock vocabulary. If you consistently review your items, you will generally see a “radical -> kanji -> vocabulary” order to your reviews with gaps of 3 to 4 days between them. The review mechanics cause review items to remain fixed to their unlock ordering.

Are certain aspects detrimental? Beneficial? What options could be provided to improve the user experience?

Grouping

Items within a group provide context for other items in the same group. After seeing some items in a review session, it becomes easier to predict the answers for the rest of the items in the same review. If one can consciously or subconsciously identify a group, one can also rule out members of other review groups.

For example, when learning kanji, you might recall that there were 2 kanji named “death” or that 3 of them had "しょう” as onyomi readings.

While some radicals or kanji look like each other, it isn’t too hard to differentiate if you can vaguely remember the levels they came from. For radicals, it might be very hard to mix up “bed” and “snake”, since the snake radical definitely came after bed.

Good: Grouping makes it a little bit easier to recall items during the apprentice levels. This boosts early accuracy.
Bad: WaniKani’s intent is to have you type in your answer (recall, not recognize), but grouping makes reviews a lot easier than intended. Grouping can be used as a crutch, which can lead to unexpected lower accuracy at later SRS stages (Master, Enlighten).

Mitigations: Reviews come out of order, and often, they get mixed in with a ton of other items. If a user can burn items based on Grouping alone, that user will probably be okay. In other words, the bad might not matter in the long run.

Unlock Ordering

Reviews follow a predictable schedule, so they will tend to follow the “unlock ordering” (radicals -> kanji -> vocabulary) with about a 3 to 4 day gap. Users on the forum have said it’s bad practice to specifically review items before you are about to get them in a WaniKani review, so this rigid order might pose a problem.

Good: It could be argued that the reverse ordering (vocabulary -> kanji -> radical) is worse; the mnemonics make it much easier to recall the dependent items. At the very least, “unlock ordering” guarantees that we don’t prime ourselves in an obviously bad way.
Bad: Even if “unlock ordering” is better than the reverse order, wouldn’t it be better to have no correlation at all? This order still primes memories a little more than necessary.

Suggestion: Randomized Review Delays

For the most part, I think that the early cycles don’t need changing and that users should have a choice if they want to do anything about these problems.

Randomized Review Delays: Allow users to set a preference where later-stage SRS review items will receive a randomized review delay. Rather than using the exact SRS delay, reviewed items may incur a random increase or decrease in days. (The expected value of the randomized delay can equal the existing delay.)

Randomness avoids the “grouping” and “unlock ordering” phenomena; both of them are derived from how strongly correlated reviews are. Since the hour and day differences matter much less after Guru, it should be relatively safe to do this.

Optional: Allow randomizing over hours, not just days.

  • Good: Better load-balancing and review mixing. A lot of users prefer a more even review distribution.
  • Bad: Pretty much every hour will have a review. Those who don’t review at fixed times may find themselves checking WaniKani constantly. Ultimate Timelines will be trashed.

Conclusion

While I do trust in the SRS system, the way it is setup leads to too much correlation between items. Overall, I think it would be beneficial (but not necessary) to provide users with more control over review timing. One suggestion is to randomize the review delay for later-stage SRS items.

It probably should also be stated that WaniKani shouldn’t be your only learning source. If WaniKani is supplemented with natural language practice, the issues I’ve highlighted might not amount to anything significant at all.

If you’ve read this much, I hope you’ve found my perspective interesting. Please discuss and share your thoughts on this topic!


[Feature Request] Unlock any item
#2

Very well thought out, I definitely agree that the grouping that happens tends to make recall easier. It is hard to say whether is is overall detrimental or if at a certain point it will cease to matter because each item is so well learned.


#3

I personally like having a very rigid review time structure, so it’s easy for me to predict my reviews without having to rely on a review timeline script… so I won’t really address that part…but I did want to mention that you can control when your reviews are going to be …to some extent.

You probably already realize this, but assuming you do all your reviews on time, your reviews will be permanently 12 hours after your lesson time. For example, if you do the lesson at 6 am, then do the 10 am and 6 pm review, you can always do the reviews at 6 pm in the future. (i know, it’s actually an hour sooner, but you get the gist). By spreading out the lessons at different hours you’ll end up getting reviews hourly instead.

Sorry if that was obvious…it’s a good idea (even though I wouldn’t use it if it were an option, exactly for the con’s that you listed…)


#4

Yeah, I was aware and agree. Unfortunately, maintaining that even spread necessitates checking WaniKani every hour. Any slip ups, and the once-staggered reviews will bunch up together again. :smiley: It’s certainly not what I’d want to do either.

I do think many users take advantage of the (Lesson, 4 hour review, and 12 hour review) to have 3 daily reviews. Earlier, I suggested days as the smallest delay increment, but perhaps 12 hours wouldn’t be so bad.

Anyway, I think it would be nice if WaniKani’s backend could do the balancing for us. I was trying to think of a way to userscript this, but I can’t think of a way that would be compatible with the other scripts people usually use.

Throwaway userscript idea:

  • Assign random, unofficial time delays (>= official SRS delay) to reviews as they are completed.
  • When a review session is started, hide reviews whose additional wait time has not been completed. [Given the existing scripts out there, it’s possible to reorder reviews and change the review screen to pretend certain items don’t exist.]
  • When using the script, the delayed reviews will remain unreviewed and unreviewable until their unofficial review time passes.

This would successfully ungroup reviews, but it comes at a price. In reality, the delayed reviews would still be in the true review queue, which may be confusing in conjunction with other scripts like Ultimate Timeline.


#5

That guy who posted the SRS mechanics post was a genius