Consistent Mnemonics, my experiment

EDIT: I just added this to the Wanikani Strategy Hub, but realize it doesn’t start with a decent explanation. WK uses some mnemonic devices to represent multiple different readings, which confuses some users. A common example is a cow, which might be for a reading of む, むう, かお, うし, and possibly others. This causes problems for some users as the different readings interfere with one another. Consistent Mnemonics is based upon memory technique like Memory Palaces, etc and requires any mnemonic symbol to mean exactly one thing, so that interference doesn’t occur.

So after reading threads like these:

I’ve been thinking about consistent mnemonics a bit and thought I’d try putting my own spin on the idea. I’ve done a bit of work on it and would love to get ideas/feedback. :grin:

I used the WK API to get every on’yomi and kun’yomi reading for every kanji in WK. There were something like 1200-1500 of them, of which ~900 were 2 or fewer syllables. Way too many (IMO) to uniquely mnemonic-ize.

So, I then split them into ‘syllables’. I put that in quotes because I split up things like long vs. short vowels and the small っ and a few other random bits that didn’t really fit/I was too lazy to code into my python script. (Maybe ‘beats’ would be a better description for these? Whatever.) For each of these, I kept 2 lists in WK order of kanji that used the particular element in its reading.

Anyway, this gave me a list of about 150 unique elements. About 130 of those are used alone as a reading and about 20 only appearing with some other beat.

I also pulled out all the long/multi vowel readings (ex,
あい, べい, etc). For want of a better place, I included combos that included the small っ (ばっ, がっ, etc) here as well. There were around 100 of those that I added to my list.

So that gave me approximately 250 separate elements. That’s a lot, but compared to 2000+ kanji, not too bad for building a library of mnemonics. I’ve gone over it and filled in the ones from the earlier levels I’ve already done, and plan on filling in more as I hit them. For each element, I am planning on creating one mnemonic for use as a primary/exclusive element and an (optional) second, related mnemonic (as discussed in those other threads) for when it’s the second syllable. Three+ syllable cases seem to be (almost?) entirely vocabulary masquerading as kun’yomi, so I’ll just deal with them like any other rare/non-standard reading as they come up.

So, I’m not planning on being restrictive on this. Two syllable readings (~750) that have obvious mnemonics for the pair or turn out to be really common will be added as they seem relevant.

Ok, that was a lot more verbiage than I expected, but for anyone who got here, what do you think? Good idea? Terrible idea? WTF are you thinking idea?

15 Likes

Will there be Hard Gay?

17 Likes

That’s similar to what I came up with, though your superior coding skills have allowed you to do a more thorough job.

Here is my excel sheet of mnemonics - I’ve not shared them before because they are so personalised, but since you asked :slightly_smiling_face: :

Rowena's mnemonics




There’s a combo of fictional characters, famous people and friends & family in there. You may very well be able to guess my age and cultural background if not exactly my country of origin from this stuff!

I have found that they work well enough for some of the later 3+ mora (that’s the actual word for the ‘beats’ you wrote of) words that are not already part of my vocab. Eg. ‘container’ 器 is うつわ, so I picture Uma う Thurman opening a shipping container and being buried in a tsunami つ of belt buckles わ!

8 Likes

That’s not really surprising, since I was inspired by your discussion in the first place. I don’t know that it’s ‘more thorough’ or just a different breakdown of the data. But yeah, being able to write some scripts to do it helped immensely in trying different breakdowns to find a good fit. You managed with intuition alone, so hats off.

I expect they will, for sure. Key to that is keeping elements ordered. Tips for that would be welcome, if you have any.

EDIT

Meant to mention, but forgot when writing the initial response. I didn’t use mora but probably should. I didn’t split ん from the prior kana and kept vowel combinations together, things like あお as well as plain long vowels. Making this switch drops the number of unique elements to 75. And surprisingly (at least to me), there are very few 2 element combos that jump to 3+ (<10).

1 Like

Nice! I’ve been thinking about trying something similar (right down to carving up the API using python…), with one slight difference - I was planning on encoding aspects like gemination, long vowel, and rendaku using size and colour on the relevant mental image. This might help you cut down on your element list if you do find 250 a bit too much to swallow comfortably once you start using it.

This is a good idea, and I’ve been doing it ad hoc already for vowel length. If I switch to mora like Rowena suggested, gemination and ん are just single mnemonics, but making them modifiers of whatever they follow could be a nice simplification.

This is based only on the kanji readings so far, so rendaku hasn’t really come into it. I have a mnemonic character (Puck from Midsummer’s Night’s Dream) I have been using for that purpose, but a modifier mnemonic might work as well or better.

1 Like

What’s the TL;DR on this?

2 Likes

I think if we’re all on the same page, some of us are vulnerable to tripping over mnemonics that are conceptually the same (e.g. “a cow”) but resolve into different readings in different contexts.

For an early example, compare the reading mnemonics for 向こう (where the cow represents ‘mu’) and 六日 (where the image of a cow is also used, but here represents the longer ‘mui’). Sometimes you can find the right reading by feel (or, obviously, just through repetition!) but “muka” is not an obviously wrong sounding pronunciation (and is in fact a perfectly acceptable Japanese word, 無化).

The idea is to settle on a different set of mental cues that are exactly 1:1 between cue and reading. This requires a lot more mental images, and probably means having to dig into symbols that are personally meaningful and memorable for each individual user. So I doubt there will ever be a globally usable general set of consistent mnemonics (the WK ones are already a pretty amazing attempt.) But the first step is to determine how many of these symbols are needed, and what rules you’ll use to slice up the WK information, which is what is being discussed here. Does this help?

7 Likes

Thank

1 Like

TL;DR? I want to do for kanji readings what radicals do for kanji symbols.

Slightly longer: WK’s mnemonics are a great first pass for learning kanji with a few problems. I’m an engineer, and I want to build my own personal mnemonic engine without those problems.

Full explanation: My goals are to come up with mnemonics that don’t clash. By clash, I mean issues like curiousjp mentions with ‘mu’ and ‘mui’, but also ones that don’t clash with 牛 later in the mnemonic for 失う as うし. Ie, no sheep → し. Mnemonics should be memorable and not confuse. And there also should not be too many components to learn either.

Competitive memory enthusiasts use certain techniques for chunking up large amounts of data to build 1-1 mnemonics with only a small number of components. So that’s what I (and others) are trying to do for the readings of the kanji and vocabulary. The issue is how best to do it.

6 Likes

Respect for your formal approach! I take an interest in methodologies and optimizations, so thanks for posting this and giving me some food for thought!

I’ll cut to the chase.

I second this. I hold that any system that comes up with mnemonics for all characters in a corpus the size of WaniKani’s is bound to have content that goes unused, even in entirely personalized systems. Predicting what is likely to go unused in a collection of mnemonics themselves is perhaps impossible because different people opt to not use different parts of the system.

For example, I find in my own studies completely emergent variables that affect my kanji learning efficiency. The more vocabulary I learn (even only as kana) outside of WaniKani, the more frequently I find myself not needing mnemonics for kanji used in those vocabulary when I learn the kanji on WaniKani after I’ve already committed the vocabulary to long-term memory using some other SRS. In this way, part of WaniKani’s effort is for naught, at least for me. Of course, others will use mnemonics I don’t, and I will use mnemonics others may skip. My point here, though, is that it’s nigh impossible to predict which parts will go unused.

WaniKani accounts for this by allowing custom user meanings and mnemonics. This approach is a valid, extremely low-cost solution to the problem of how to make mnemonics flexible. It simply sidesteps finding a clever way to make key parts of the mnemonic organic by instead allowing the user to make something entirely different. If you could somehow come up with a template for consistent mnemonics, I think you would be advancing the field. (In the meantime, I’m creating many of my own mnemonics from scratch as well.)

As for your quest for consistency, can it be done in a system that only you use? Absolutely. Can it be done in a system anyone could use with only limited modification allowed? I think this is an open question. The other big question here is, “how much time are you willing to invest?” Solving the case of the template system may take longer than you completing vanilla WaniKani.

Thanks again for posting this! I enjoyed thinking through this.

3 Likes

Thanks. I think in systems and patterns (well sometimes), so I’m always looking for them. Japanese orthography is just rife with them, so they scratch that itch.

So, I have thought about it, but came to the conclusion that I wasn’t the one to do it. At least, not yet. Such a system would ideally take into account the whole pattern, not just the ~15% I’ve seen so far. A good system would, for example, know where particular mnemonics should be used and ideally coach the student into picking ones that are memorable, non-conflicting, and work with that content. WK does this to some degree, with Mrs. Chou and Ken the samurai. It tells you up front that you’ll use this a lot, so make it a good one.

So I’d want the system to say “This is the first time you’ve seen this reading. It will later be used for concepts such as A, B, and C. Pick something that relates to those ideas.” I can’t do that quite yet since I don’t know that yet. I could write a script to pull the meanings, but ideally it would take the rarity of the readings into account as well, choosing the A, B, and C with the most bang for the buck. Some of the readings are used for 40+ kanji, but are rare readings, and there’s little point in crafting a mnemonic that won’t be frequently useful. I can’t do that quite yet, and yeah, I don’t think I want to go to the effort now, for just me.

But if I were to build such a thing, I can see such a system having a guided Madlibs approach, where it suggests “Pick a character/action/object/whatever” so that it can auto construct mnemonics for you : “Cap’n Crunch is packing a snorkel for his trip to the moon” for がつ/月. I could even forsee a system like they have at Kanji Koohii where users share their components and mnemonic fill-in-the-blank phrases so a new learner could scan the list and pick the one that was funniest/most memorable.

The system could also advise when to pick things in pairs/triples to facilitate learning dakuten/handakuten. Ie, Kirk/Spock == て/で, or Mickey/Donald/Goofy are は/ば/ぱ. I’m not sure how I’d want to handle gemination, syllabic ん, vowel length in the system, but probably modifiers of some sort like curiousjp suggested upthread.

TL;DR It’s pretty interesting to think about how you could build a mnemonic template generator that could be personalized. I’ve got a bunch of ideas I’m trying in this “experiment” and I’ll report back here if/when I make discoveries.

3 Likes

Hi Rowena, do you have the actual spreadsheet file for this amazing chart?
I would like to do something similar with different people for the mnemonics, but I am really bad at making things like this.

Here’s mine, if it helps:

Notes:

  • Column C has counts of WK kanji with the specific mora in 1st, 2nd, etc. position for any reading in WK, regardless of rarity.
  • F and G are lists of Kanji with the mora in first position for the WK reading (I think. It’s been awhile)
  • H is the glosses for all the Kanji in F & G. Originally I thought there might be some patterns that could be exploited here, but it doesn’t really work.

Let me know if you have any questions

2 Likes

Thanks, really interesting to look at.
I wonder if there was any reason to sort them alphabetically by Romaji?
What are the numbers in column I and P?
Edit: also, the totals for column C show all kanji from level 1-60, right?

I go back and forth on sorting, sometimes by romaji (A), sometimes by hiragana (B). I like having the rendaku-ed kana together, but I have trouble remembering the order.
So for ease of lookup, I’ll sort them romaji. For ease of comparison, I’ll sort them kana. So it depends on how lazy I feel/what I’m looking for. It just happened it was in romaji order at the moment when I built the dummy table. I also sorted by I and P (see below) while filling in the mnemonics, but only needed that at the beginning.

I is just the number of first position occurrences. P is the number of occurrences in positions 2+. I used that to sort the list by how often I’d see each mora at a specific location. That way, I could focus my efforts on the most common mora/position combinations and not waste time/good mnemonics on combos that would come up only rarely.

And yeah, C is for all kanji 1-60.

Enjoy!

1 Like

Sorry for the slow reply - are you happy with sporadic’s chart, or would you still like a copy of mine as well? I’m ok with excel, but might need some help in getting into a sharable format…

@Rowena
That’s no problem, I am good with Sporadic’s, thanks!

@sporadic
I hope that you answer some things that I may be thinking too much about.
I wonder about dealing with some of the longer and you say, rarer mnemonics, for example:

美しい
utsukushii - うつくしい
WK= “U too cushy!” (beautiful)

失う
ushinau - うしなう
WK= “ushi” “nun” (to lose)

Using your elements there would need to be 3 of them in the story in the right order for the mnemonic to work. So for these readings, where there is only one in WK, is there any difference in using the pre-generated mnemonic, or building it with your elements for unique readings?
I guess there is more consistency with your method, but there still needs to be a unique mnemonic to get the parts in order.

The “mnemonic 2” for う and い - these appear 605 and 309 times respectively, how do you deal with the high frequency of these elements appearing as secondary to their first syllable counterparts. Have the high volumes of these led to the “mnemonic 2” of う, い, く, etc to be main characters in your memory palace of WK kanji and vocabulary?

I know you have written about this in previous posts, but hopefully you can clarify it a bit for my particular situation.

1 Like

So this method works for 1-2 elements fairly well. With three or more, there appears to be potential for breakdowns, since order isn’t implied.

So far though, I haven’t had a problem with this. Maybe because I use the audio and repeat it, but I’ve found if I can get the first mora, the order of the rest flow naturally. Other orders typically sound wrong to me.

I did think about the problem a bit when I was first setting up, and had a couple of strategies I was going to try, should they be necessary. First was to just split into multiple characters, so everything was in pairs. This is similar to what competitive memorizers do when learning decks of cards. An alternative was to work the order into the mnemonic in some way, either by having a modifier (adverbs?) or just making the story context clarify it.
My final answer was just to punt. There aren’t that many cases, and most of them are vocabulary in their own right, so I’d take them as they come. No worse than vanilla WK, after all.

For vowels and ん, I tried to pick ideas that had lots of potential for variation. For example, my う is “undead”, which can mean anything from being a zombie/vampire/ghost to creating/raising those things. This has caused problems, because then I can’t use those ideas for other mora, like it would be nice if ぶ was a ghost, for example.

To deal with that problem, I’ve tried to use the “giant/tiny” idea for long/short vowels, but it’s not perfect, and it does break the pattern. Then I only use the mnemonic 2 for cases where it’s not a long vowel, which helps a bit.

I hate my く, mnemonic 2, btw. Initially it was “cooing”, then became “like a pigeon”, then “turning into a bird”, but it’s still giving me trouble. So yeah, not perfect.

Anyway, that’s what I’m doing right now. Maybe @Rowena has some other tips for dealing with the edge cases.

I can’t say the system works better than the one-off of WK proper, unfortunately. My correct/incorrect rates seem pretty typical. I have a really bad memory in general, so who knows?

I do like it though and have stuck with it, so that’s something. Maybe me just being stubborn and not caring for the WK mnemonics. But something. :wink:

Still, the lack of ambiguity is really nice. Not only with the clue words being used for multiple pronunciations, but also because the clue words often feature in other mnemonics. It’s really frustrating when you remember a story, but don’t know which part of the story is supposed to tell you. Was the car incidental, or is it “か”? Or “かる”? Or “かり”. Ugh. Just no.

Anyway, hope that helps. And if you have any ideas for disambiguating those edge cases, please share. We’re all trying to figure this method out, after all.

@sporadic

I am finding your spreadsheet useful, so thanks again for sharing. I have not fully populated it yet, and will continue to use it as a work in progress.

My wife speaks to our children in Japanese and I get to read some basic children’s books. Although they are starting to get a bit bored with the speed I read unless it is one that I basically memorized! My grammar is terrible, and I will never be able to interject in the conversation. Some words that WK teaches me however are bit of an “aha! moment”, and making a mnemonic would be going backwards. This level, 顔 (かお) and 頭 (あたま), are vocab that I have known for ages, and are already burnt in my mind! A bit like what @orphen says in his outside of WK studies.

So for me I think I will definitely augment my learning of WK with the things I have learnt from you, @Rowena and a YouTube overview of the Moon-walking with Einstein (I can’t pretend to ever have the time to read to book, but defiantly useful to listen to the author explain about it).

This method could have been more useful to me starting up at level 1, but I didn’t honestly think that I would be still doing this a year later (I’m slow?!). I think that I will keep adding to my take on your spread sheet, especially with the readings that are far to reach for me with a British accent. Even with trying a very bad American accent I can’t make HAWK likeはこ(I am using Hakone as it feels natural).

I do find with some master and most enlightened that when I see the kanji, the reading and/or meaning is usually automatic and the Kanji has lost any need for mnemonics. I think that my usage of your method will definitely help me get to this state faster by adding consistency to new readings and also sorting out once and for all the confusing ones: 月 (がつ) gatsu/ 月 (げつ) getsu.

I have made some combination elements for readings at level 10, trying to keep them based on original WK. An example is さ for Saw and small っfor Tui beer (it’s a New Zealand thing, but there is a bird too). 早速 (さっそく) mnemonic is: there are lots of these Central Otago guys at a wood Sawing competition, of course there is lots of Tui to go around, get me one At Once!

It’s not based on actual human characters, but just the act of creating it in a visual way has definitely improved my success rate in these early stages. I can’t say how it will work once it’s in the Master and Enlightened stages, but hopefully I will get there quicker.

I’m not sure what normal percentages are for correct readings. I found that after a couple of slow months where my review queue consistently had 100-200 reviews was 60-70%. When I am doing multiple daily WK sessions I am 75-90%. I think when there is pressure to go quickly, things start to break down!

Thanks again for sharing your inner workings, it’s great to learn from other people and keep each other motivated in this great community.

1 Like