Lately I've been becoming more and more of a fan of using multiple spaced repetition decks, even within the same area of study. The disadvantage of doing this is that you have to switch between files, so one might wonder, why not just put all the cards in one deck? Thing is, when you start dealing with a very large deck, if you have you take a break from daily reviews (and you will), the deck can get out of hand. Right now, my main Japanese deck is "stuck" around 300 failed items. It's slowly going down, but it's gonna take months before it reaches 0 again. I could just grit my teeth and go through all those cards in one giant sitting, which would take the better part of a day, but then the next month I'd be dealing with tons of things coming due every day. I don't want to devote my life to SRSing. Flashcards are a tool to serve you, but too many people become tools who serve their decks. So, I only do a set number of reviews per day, as explained here.
When your deck gets "stuck", you can't add more cards without mucking it up even further. When the carddeck is one of your main study methods, that means you can't learn new stuff for awhile (in my case, a long while).
For a long time I struggled with this barrier. I didn't want to just start a 2nd deck exactly the same as the first, because it'd just get to where I have the same problem, only double. Finally I came upon the solution. I call it the Satellite Model of Spaced Repetion.
In the Satellite Model, you maintain one large "central" pile, and augment it with smaller helper piles. Since the helper piles are small, they never accumulate too many reviews due. Even if you neglect them for a couple days, it's not a big problem. And when the flagship is overburdened and clogged up with failed/new cards, you can add additional cards to the smaller sets.
Maintaining a small SRSfile presents some interesting challenges on its own, which is giving me further insight into spacing theory as well as language learning in general. I'm starting to rethink certain habits which I took before as ideology. The main thing is, by definition, you want to keep small piles small. This constraint forces you to moderate your sentence mining. I'll discuss this some more below. Anyway, by handling the new cardsets in different ways, I'm doing new things, thereby increasing my language-learning intelligence (see my more general article, Become More Intelligent By Doing New Things). I'm opening my mind to the possibility that I was wrong before in some of my habits. (Well, "wrong" isn't the right word, there really is no "wrong" way to SLA as long as you've got regular exposure! "Suboptimal" might be better.) I'll discuss some of this further on.
MY SATELLITE DECKS
Right now, for Japanese study, I have one main ginormous flagship pile of flashcards, and three tiny orbiters. The "motherload" currently has 9003 facts, which are mostly sentences, along with about 1700 Heisig kanji, and a much smaller amount of raw vocabulary. You can read about its Anki Statistics here. I've been working on this for years now. When I went to Japan for thirty days, I totally neglected this cardset and when I got back it had blown up to thousands of reviews due. I didn't get it back down to managableness until finally getting my hands dirty and doing some huge review marathons while killing time last December while my GF was in Tokyo. (The reviews-per-day is still recovering from that.)
Then, I made a new file, "japtemp". At first, this was mainly for vocabulary words for the JP101-103 independent study I was taking to cheatingly get up to the 9 units I'm required to take as a TA. (I'm overqualified for the class, and classes suck anyway, but there was still some unfamiliar vocab.) This file doesn't contain any sentences. Later, the purpose shifted toward words I picked up when talking in Japanese to my girlfriend. Again, just vocab drill, no s-mining. Currently it has 182 entries, all vocab.
Encouraged by good results with japtemp, I made another satellite, "kanjiaid". Here, as the name suggests, the purpose is to help familiarize me with kanji which I'm particularly struggling with. Heisig's philosophy says you should learn all the kanji before you start learning to read them or actually speak at all. But I've come to really question the wisdom in that. I'm finding that when I spend only a limited amount of time per day on reviews, certain Chinese characters are a serious pain for one reason or another (usually it has little to do with how complicated they are, some of the hardest characters happen to be some of the simpler ones, just because they don't have as nice a structure as the more convoluted ones). In this helper deck, I include both vocab items and sentence items containing the kanji I struggle with. But to avoid running into the same congestion I'm trying to avoid, I'm taking a very minimalist approach. For most kanji, I pick just one word and one sentence involving that word. This is a world of difference from my old sentence mining philosophy of going for as many sentences and clauses as I could get my hands on. Kanjiaid works nicely, it's really helping learn those tricky Chinese characters, because it's much easier to remember them when they're familiar to your eyes.
Finally, just last night, I lauched my newest pack, which I call "latekanji". This is to finally finish off the last couple hundred Heisig chars, which I've been neglecting ever since December just because I haven't had a deck sufficiently cleared up to add more material since then. With this deck, it's just Heisicards, with an English keyword (and possibly an example nihongo word) on the question side and the kanji on the answer side. This time around, I'm also going to be writing the memory-enhancing visualization stories on the answer side, using invisible text.
QUESTIONING OLD PRACTICES
Dealing with these lean, mean new helper decks has made me question some things I used to take as gospel, largely because of exposure to a kind of SLA "ideology".
Raw Vocabulary Cards
I've been a champion of rooting out the raw vocab cards from the Spaced Repetition System, replacing them with lots of sentences full of context. I still believe that's the optimal system for one particular word, but now I'm really in more of a time crunch than I was back then, and I'm seeing that in order to make real headway, it's necessary to sacrifice some quality for some quantity. By using rote vcards in my helper decks, I'm able to expand my acquired lexicon at a much faster rate, and at the same time familiarize myself with tons of delicious Chinese chars.
Besides, we mustn't underestimate the power of the subconscious mind to figure out the bigger pattern, given only very limited data. We are master interpolators and with only just one or two example sentences for a word, your mind can make long strides toward deducing how that word is properly used. There's no need to go overboard and import every clause you find in every dictionary. Moderation, moderation!
English->Target Language
In the japtemp file I made, since part of the purpose was to cram for a specific class, and since that class sometimes involved English->Japanese direct translation, I went ahead and made some entries with English on the question side and the "Devil's Tongue" on the answer side.
Among the hard core language self-teaching community, this is pure heresy. The problem is that there's never a single fixed translation for a given English phrase. Often there isn't even a direct correlation even on the word-to-word level.
What I realized, though, is that it doesn't really matter. The Golden Rule of Language Learning says that all that matters is exposure. It doesn't even really matter whether the translations are correct. I could make a card with question side "I want to eat" and answer side 「猫がかわいい」 ("cats are cute"). In the short term it might screw me up quite a bit, but in the long run the English cue will just be blurred out and all I'll even "see" will be the target answer-- which is exposure, any way you look at it.
The only thing to watch out for is making answering impossible by making two entries with the same question side, like two "body" questions with two synonyms for body in Japanese as their answers. The way to get around this is to paste the ENTIRE English gloss from your dictionary. At least with the WWWJDIC as a dictionary, this always solves the problem. Using the body example, the English gloss for one synonym (体) is
ARE SPACED REPETITION ALGORITHMS BIASED TOWARD SMALLER DECKS?
This is purely, 100% speculation, but think about it for a second. The guys who develop spaced repetition software probably don't test them very thoroughly on massive decks that have been slowly built up and reviewed for years. When you start getting in the high four digits, you're really departing from where the algorithms have been carefully tested. I'm sure whatever theory underlies everything, carries over fine to arbitrary numbers. But often theory and practice clash.
On the other hand, the spacing computations have been very rigorously tested for small, lean little card decks. So you're guaranteed to get really great system performance for them.
FURTHER READING
For more on rote memorization, check out my article, Rote Memorization In Mathematics. Incidentally, I wonder whether anyone has successfully used SRS technology for any great length of time to study math?
Here's an interesting and refreshing look at the Japanese language. The Adverb Model of Japanese. I think this is probably the only language in the world where this crazy model can be pulled off so easily.
For a break from all this seriousness, treat yourself to a chuckle from my article, You Might Be An Autodidact If...
When your deck gets "stuck", you can't add more cards without mucking it up even further. When the carddeck is one of your main study methods, that means you can't learn new stuff for awhile (in my case, a long while).
For a long time I struggled with this barrier. I didn't want to just start a 2nd deck exactly the same as the first, because it'd just get to where I have the same problem, only double. Finally I came upon the solution. I call it the Satellite Model of Spaced Repetion.
In the Satellite Model, you maintain one large "central" pile, and augment it with smaller helper piles. Since the helper piles are small, they never accumulate too many reviews due. Even if you neglect them for a couple days, it's not a big problem. And when the flagship is overburdened and clogged up with failed/new cards, you can add additional cards to the smaller sets.
Maintaining a small SRSfile presents some interesting challenges on its own, which is giving me further insight into spacing theory as well as language learning in general. I'm starting to rethink certain habits which I took before as ideology. The main thing is, by definition, you want to keep small piles small. This constraint forces you to moderate your sentence mining. I'll discuss this some more below. Anyway, by handling the new cardsets in different ways, I'm doing new things, thereby increasing my language-learning intelligence (see my more general article, Become More Intelligent By Doing New Things). I'm opening my mind to the possibility that I was wrong before in some of my habits. (Well, "wrong" isn't the right word, there really is no "wrong" way to SLA as long as you've got regular exposure! "Suboptimal" might be better.) I'll discuss some of this further on.
MY SATELLITE DECKS
Right now, for Japanese study, I have one main ginormous flagship pile of flashcards, and three tiny orbiters. The "motherload" currently has 9003 facts, which are mostly sentences, along with about 1700 Heisig kanji, and a much smaller amount of raw vocabulary. You can read about its Anki Statistics here. I've been working on this for years now. When I went to Japan for thirty days, I totally neglected this cardset and when I got back it had blown up to thousands of reviews due. I didn't get it back down to managableness until finally getting my hands dirty and doing some huge review marathons while killing time last December while my GF was in Tokyo. (The reviews-per-day is still recovering from that.)
Then, I made a new file, "japtemp". At first, this was mainly for vocabulary words for the JP101-103 independent study I was taking to cheatingly get up to the 9 units I'm required to take as a TA. (I'm overqualified for the class, and classes suck anyway, but there was still some unfamiliar vocab.) This file doesn't contain any sentences. Later, the purpose shifted toward words I picked up when talking in Japanese to my girlfriend. Again, just vocab drill, no s-mining. Currently it has 182 entries, all vocab.
Encouraged by good results with japtemp, I made another satellite, "kanjiaid". Here, as the name suggests, the purpose is to help familiarize me with kanji which I'm particularly struggling with. Heisig's philosophy says you should learn all the kanji before you start learning to read them or actually speak at all. But I've come to really question the wisdom in that. I'm finding that when I spend only a limited amount of time per day on reviews, certain Chinese characters are a serious pain for one reason or another (usually it has little to do with how complicated they are, some of the hardest characters happen to be some of the simpler ones, just because they don't have as nice a structure as the more convoluted ones). In this helper deck, I include both vocab items and sentence items containing the kanji I struggle with. But to avoid running into the same congestion I'm trying to avoid, I'm taking a very minimalist approach. For most kanji, I pick just one word and one sentence involving that word. This is a world of difference from my old sentence mining philosophy of going for as many sentences and clauses as I could get my hands on. Kanjiaid works nicely, it's really helping learn those tricky Chinese characters, because it's much easier to remember them when they're familiar to your eyes.
Finally, just last night, I lauched my newest pack, which I call "latekanji". This is to finally finish off the last couple hundred Heisig chars, which I've been neglecting ever since December just because I haven't had a deck sufficiently cleared up to add more material since then. With this deck, it's just Heisicards, with an English keyword (and possibly an example nihongo word) on the question side and the kanji on the answer side. This time around, I'm also going to be writing the memory-enhancing visualization stories on the answer side, using invisible text.
QUESTIONING OLD PRACTICES
Dealing with these lean, mean new helper decks has made me question some things I used to take as gospel, largely because of exposure to a kind of SLA "ideology".
Raw Vocabulary Cards
I've been a champion of rooting out the raw vocab cards from the Spaced Repetition System, replacing them with lots of sentences full of context. I still believe that's the optimal system for one particular word, but now I'm really in more of a time crunch than I was back then, and I'm seeing that in order to make real headway, it's necessary to sacrifice some quality for some quantity. By using rote vcards in my helper decks, I'm able to expand my acquired lexicon at a much faster rate, and at the same time familiarize myself with tons of delicious Chinese chars.
Besides, we mustn't underestimate the power of the subconscious mind to figure out the bigger pattern, given only very limited data. We are master interpolators and with only just one or two example sentences for a word, your mind can make long strides toward deducing how that word is properly used. There's no need to go overboard and import every clause you find in every dictionary. Moderation, moderation!
English->Target Language
In the japtemp file I made, since part of the purpose was to cram for a specific class, and since that class sometimes involved English->Japanese direct translation, I went ahead and made some entries with English on the question side and the "Devil's Tongue" on the answer side.
Among the hard core language self-teaching community, this is pure heresy. The problem is that there's never a single fixed translation for a given English phrase. Often there isn't even a direct correlation even on the word-to-word level.
What I realized, though, is that it doesn't really matter. The Golden Rule of Language Learning says that all that matters is exposure. It doesn't even really matter whether the translations are correct. I could make a card with question side "I want to eat" and answer side 「猫がかわいい」 ("cats are cute"). In the short term it might screw me up quite a bit, but in the long run the English cue will just be blurred out and all I'll even "see" will be the target answer-- which is exposure, any way you look at it.
The only thing to watch out for is making answering impossible by making two entries with the same question side, like two "body" questions with two synonyms for body in Japanese as their answers. The way to get around this is to paste the ENTIRE English gloss from your dictionary. At least with the WWWJDIC as a dictionary, this always solves the problem. Using the body example, the English gloss for one synonym (体) is
(1) body; (2) healthand the gloss for another synonym (身) is
(1) body; (2) oneself; (3) one's place; one's position; (4) main part; meat (as opposed to bone, skin, etc.); wood (as opposed to bark); blade (as opposed to its handle); container(as opposed to its lid).A lot of the second gloss is obscure stuff we don't really care about, but it sufficiently distinguishes the two cards, and in the long run, all we'll "see" when that review comes up is the overall shape of the question.
ARE SPACED REPETITION ALGORITHMS BIASED TOWARD SMALLER DECKS?
This is purely, 100% speculation, but think about it for a second. The guys who develop spaced repetition software probably don't test them very thoroughly on massive decks that have been slowly built up and reviewed for years. When you start getting in the high four digits, you're really departing from where the algorithms have been carefully tested. I'm sure whatever theory underlies everything, carries over fine to arbitrary numbers. But often theory and practice clash.
On the other hand, the spacing computations have been very rigorously tested for small, lean little card decks. So you're guaranteed to get really great system performance for them.
FURTHER READING
For more on rote memorization, check out my article, Rote Memorization In Mathematics. Incidentally, I wonder whether anyone has successfully used SRS technology for any great length of time to study math?
Here's an interesting and refreshing look at the Japanese language. The Adverb Model of Japanese. I think this is probably the only language in the world where this crazy model can be pulled off so easily.
For a break from all this seriousness, treat yourself to a chuckle from my article, You Might Be An Autodidact If...
Posts
1 comments:
An insightful post, I've wondered about dividing my core deck into several smaller ones for a while, so it was nice to see some input on this.
However my main question that I want to ask is, would it not be easier to just use different categories as opposed to making a brand new deck? Anki allows you to switch categories on and off at will after after all.
Post a Comment