so, obviously there's been loads of talking about this, but all we really need are some statistics (there's a ton of people on RTK using anki) and a mathematician (i think there's quite a few on here?)

ok, i'm not a mathematician, but, vosimura wrote a plugin for anki before that showed results on mature cards. i'm not sure of the maths of it, but the same kind of thing could be used to get the results we need. i think what we need are:
1. no. of cards added per day
2. percentage of very good / good / difficult / failed mature cards.

Lots of data from everyone. Compare the data for cards added to results +21 days later (or more, if possible), making some kind of scatter graph... the results should show if there is a correlation between mature results and cards added, and roughly what the optimum amount of cards to add per day is.

Maybe someone good at this stuff could say if this sounds about right / possible? Obviously, it's not perfect, cos it can't measure results outside anki, in the real world, and depth of knowledge, etc. but, it might be a start... what do you think? our own scientific study

To me stats are meaningless beyond a basic personal guide. No point sharing. I grade my cards and have data on them and construct my decks in my own peculiar ways. ^_^

I might even say I don't use stats, period. As long as the reviews are spaced apart when I want them to. Not sure I've ever looked at my stats. The 'maturity-based' ecology idea was kind of new for me and only had meaning insofar as: the fact that a card became 'mature' meant that for a given length of time, I had stopped my arbitrary changes from day to day in grading, and had nothing else to learn in that context.

IceCream Wrote:hmm, i didn't really mean like that. An individuals stats are fairly meaningless i think, i don't look at mine too often either. But as a whole, a lot of people's stats can create patterns. Maybe we could think of a way to annoymously give them, too...
it also is more meaningless when you're not using study resources, because the sentences are different, and you do grade them different ways.
But, when i was using study resources, and i'm sure people doing RTK itself probably grade cards a lot more standardly. Especially when they reach maturity, since mature cards are graded a lot more standardly for everyone. Either you know it or you've forgotten. sort of...

maybe even, fabrice would have access to these kind of stats for people using this site?

I guess. Stills seems like too many variables to me to become anything more than a self-fulfilling, double-edged metric. I understand the connection to the previous thread (RTK in a week?), I think, as I mentioned I think there's probably a 'sweet spot' between trying to do RTK as fast as one can, or pacing oneself and doing complementary aspects of the language while doing RTK, and devising strategies to balance reinforcement/interference. (If I could go back, I'd try to combine complementary assembly-line style w/ momentum for cross-strengthening redundancy.) (For example.) But I really do think finding that spot can't be measured.

Oops, and anyway, sorry to be such a downer. I'm sure others might find the data useful/interesting, as arbitrary and self-fulfilling and prone to misuse as I think it might be! ;p

PS - I don't think the resources matter, there's a ton of variables either way.

IceCream, is your goal to maximize the amount learned per time studying?

If that's the case, it's pretty clear you shouldn't use an SRS at all.

http://supermemo.com/help/faq/memory.htm...um%20speed

I'm not a mathematician or statistician or anything, so I'd have no idea how you would even go about analysing the stats. But I think you'd find that studying more/adding more cards per day is a function with diminishing marginal returns. Each extra hour gives you increasingly less benefit; but it still does give you benefit. In other words if you were just looking at the benefits side of the equation, there probably wouldn't be an amount of cards that, if you added more, you would be worse off afterwards. Which means that the optimal amount is not based on how long you study, but more on the costs of studying more. And that would make it hard to really pinpoint anything useful, because the opportunity cost of studying would vary wildly from person to person.

EDIT: And that was just my extremely verbose way of saying "Do what works for you"

IceCream Wrote:not exactly, the goal was to find out whether and how much increasing number of cards added per day relate to memory retention.

As long as you keep up with reviews, long term retention is independent of of how fast you initially learn the material.

Quote:The answers given on the mnemosyne website (thanks for the link) go some way, but seem like it is the forgetting rate that is fixed, and other factors tested. Therefore, none of them can provide an answer to the question of how how many cards added relate to retention rate.

Actually, the forgetting rate isn't fixed. If each item has a stability S (measured in units time) and a retrievability of R (probability it will be remembered) then the forgetting rate (change in R per unit time) for that item is proportional to R/S. The overall forgetting rate is the sum of the forgetting rates of each item.

The assumption that the total forgetting rate is fixed comes from the assumption that you study a fixed number of cards per day at a fixed forgetting index. For example, if you study 250 items per day at a FI of 15%, you forget an average of 37.5 items per day.

Quote:Apart from figure 3 on http://supermemo.com/articles/theory.htm...um%20speed, but, i have only gcse maths, and i can't figure out why the acquisition rate is varying along with the forgetting index, it seems like not a fair test. If anyone can explain in simple terms, id appreciate it... (should this graph be 3d or something??)

Current versions of SuperMemo allow you to change the target forgetting index, and the algorithm changes its scheduling to meet that goal.

That graph was made by taking statistics from a SuperMemo user (I think Dr Wozniak himself, but I'm not sure) and simulating what happens if the forgetting index is varied. At very low forgetting indices, SuperMemo schedules the reviews close together. These close-scheduled reviews take a lot of time and the spacing effect causes items to take a long time to mature. At very high forgetting indices, reviews are scheduled a long time apart and it only takes a few reviews for items to mature. However, you spend so much time reviewing the forgotten cards that you don't have time to learn new material. The best trade-off in terms of speed is a FI between 20 and 30 percent, the mostly-flat peak of the curve.

Interestingly, this is similar to the natural forgetting index. If you study something that you understand well (in my case, something like car repair, English vocabulary, math, or physics) frequetly, but without any review, just relying on the natural redundancy of the material to remember what's important, you end up with a forgetting index of about 20 to 30 percent. The brain is optimized for learning speed.

Quote:Mnemosyne gives some other answers, but they don't really seem very well backed up to me:
http://supermemo.com/help/faq/memory.htm#21052-461
http://supermemo.com/help/faq/memory.htm#19357-4432
it does say about interference though, which presumably would increase along with more cards added...

Interference occurs when two items collide. For example,
ELUDE -> 逸
EVADE -> 避
ESCAPE -> 逃
can collide because the keywords are near-synonyms (and it doesn't help that the characters all have the same radical). My solution was too incorporate some English word-play in my mnemonics to keep them separate. Anyway, my experience is that interference is worst between young cards. Mature cards are more sharply defined and less likely to interfere. As a result, interference is usually more of a limit on how fast you can learn than how much you can learn.

Quote:um, i dunno why im even bothered, im pretty happy with my learning rate tbh, i guess just theoretically, it's interesting.

But, there are some interesting things on there. Like how your cards should be active recall in some way. And how it really doesn't help in remembering things without a trigger. Which suggests that adding context really would be the best way to help triggering in active things like speech and writing...

I don't think SRS is the right tool for training speech and writing. There isn't One Right Answer to speech and writing situations, so you can't possibly make items that say "this simple question leads to this simple answer." The only way you can know that what you're saying or writing is correct is if
a) you understand Japanese really well, and
b) native Japanese understand you.

To put it bluntly: input first, then make Japanese friends.

However, you've hit on something important in noticing that when using SRS to train input, you should do active recall of some kind. This is one of the reasons I don't particularly like the sentences method. (The other one is complexity. I'm supposed to remember how many vocab and how many readings all at once? You've got to be kidding me!) It is very difficult to grade "understanding."

The ideal low-level items are things like: kanji from L1 keywords, readings of single words (for both 音 and 訓) or even single characters in a word (for 音), and (may Khatz have mercy on me) L1 glosses of L2 words--ideally with pictures as well.

The ideal high-level items are general knowledge (including vocabulary) expressed in L2, both question and answer.

Sentences strike me as intermediate items that 1) give examples of context and grammar and 2) bridge the gab between basic items and advanced ones.

Anyway, I've spent enough time on this today. I'm going to continue experimenting with mnemonics for basic passive vocabulary. I think I've found a good method (it even does 訓読み!), but I need to get to a thousand words or so before I'm really confident that it works.