On Jun 30, 3:38 am, Peter Bienstman <[email protected]> wrote:
> On a more personal note, I've been using (the predecessor of) Mnemosyne since > 2003, learning roughly 5 cards a day. I'm now at 175 scheduled cards a day on > average, and 8500 cards in my database. There is still plenty of stuff I want > to learn, and if I keep up what I think is this steady, gently pace, I could > be at 350 reps daily in 5 more years... > > This very long term aspect is definitely something that needs thinking about, > either by making a more thorough analysis of the logs and tweaking the > algorithm, or by pruning the cards in my database. > > Peter I know you're waiting for the hard data, but allow me to share a few ideas. 1. One resists the idea of pruning cards, but after thinking about it I have found a good rationale (not proof). Some have opined that at very short intervals, beginning from the first glance at a fact, flashcards are not yet the ideal tool, that one should first learn the fact to some (arguable) degree. (I agree, without knowing how well that should be. I've tried stretching this out as long as eight days, memorizing material before flashcarding it for retention-only. An ideal is probably in there somewhere.) Now, similarly, maybe the single-fact, atomized-data style flashcard system becomes non-ideal again at long intervals too. For example, one of my earlier Chinese textbooks broke down into 900 flashcards total, which are still in mnemosyne. I can now read, aloud or not, this book fairly rapidly, and understand its audio. So, reading or listening I zoom over hundreds of "atoms", all nicely connected with context and grammar, etc.- real language. At some point, it might be a good idea to prune all 900 cards and make a "review scheduling style" card maybe like this: Front- "read Modern Chinese Reader aloud" Back- "Did you know (almost) everything?" (Where "almost everything" concedes that your brain is not a machine, after all; we all have a standard, and compromise on "perfection" for the sake of just carrying on living and learning.) I now intend to do this when I get around to it. (By the way, this would make it even more important that you're learning from something cohesive, like a book with lots of context, *so that* you could later prune all of the cards, knowing you can still hold them all securely in one hand.) I had thought that once cards were known perfectly well that each card would become sufficiently effortless. It doesn't; it is still many times harder than flying over that same fact in context while reading or listening. (You could prove that.) I had also thought that once they were promoted far enough, they would practically disappear. Well, your testimony above confirms my impression that they don't, quite well enough. This is what motivated me to think about this again. 2. Back to your discussion above about rep-buildup: Again without proof, after some years tweaking your program (trying to stay within the bounds of something that would fit into your project), the most convincing improvement I think is the idea of demotion by some factor instead of failing all the way back. I have a strong common-sense rationale for this too. At some point, we think we "know" a card, but a cautious compromise is to promote it by some factor instead of "all the way". When we miss a card, we could employ a symmetrical "cautious compromise" (cautiously avoiding a buildup at shorter intervals that would crowd other cards we *know* we don't know yet), and demote by a factor, instead of all the way. Price paid: we won't know until we see it again, at some shorter interval, whether we really still have some grip on it. Benefit gained: less clogging of the cards we *know* we don't yet retain well. (Common sense exercise: jot down the ones missed and review outside the program.) This is a common sense compromise, the degree of which is powerfully controlled by this one variable: very convenient and easy to understand. 3. Lastly, I'll mention this here because it is the *only* other important thing I finally decided on; I posted about it here some months ago. This is the idea of accelerated promote/demote, where your ef is still calculated as usual per button and saved, but an additional (one time, *this* time) factor accelerates the desired promotion/slowing/demotion. I contributed a tiny bit of code a couple of years ago. Now I have 1.99 running with virtualenv, etc., and hope to produce algorithm plugins. I won't mind if someone beats me to it though. I exchanged my very first all-Chinese emails the other day. Now I'm trying not to burst with premature satisfaction. Mnemosyne is at the heart of my language-learning world, every day, and I appreciate it very much. --~--~---------~--~----~------------~-------~--~----~ You received this message because you are subscribed to the Google Groups "mnemosyne-proj-users" group. To post to this group, send email to [email protected] To unsubscribe from this group, send email to [email protected] For more options, visit this group at http://groups.google.com/group/mnemosyne-proj-users?hl=en -~----------~----~----~----~------~----~------~--~---
