Some crude hacking on memory-training software.

deck-player/: SuperMemo-2 algorithm on a deck in Anki import format

deck-builder/: Tries to build a good deck automatically out of an
aligned parallel corpus. Germ of a more specialized language-learning
tool which maybe I'll tackle someday.


Notes:

What's great about Anki:

* Useful learning in 10 minutes/day with no special effort. Beats TV!

Shortcomings of Anki:

* I tend not to learn a card well enough in its session, because it's
  taken out of the rotation as soon as I recall it at all. So then by
  the time it next comes up, I've forgotten it, instead of *almost*
  forgotten it. Wozniak's scheme seems better.
* 'Flat' sessions (e.g. consider what happens if you try learning 200
  new cards/session).
* Contributed decks are often ill-structured ('ashtray' first instead
  of the most frequent & useful words).
* Flashcards aren't much like the real-world tasks you're training
  for. I don't know if this is a shortcoming, in net -- maybe the
  greater focus of practice makes up for the greater abstraction --
  but other more contextual schemes seem worth a try.
* It has no way of knowing when I tend to confuse two words. In such
  cases, I could use focused practice on distinguishing them.
* Next showings 'in N hours' probably not useful, compared to daily
  sessions.
* Lumpy schedules -- later repetitions should get spread out over a
  few days.
* Not obviously well-engineered. (E.g. a huge directory of backups of
  DB files. Thousands of lines of Python, not counting the UI. It's
  hard to find the core logic in all that code.)
* Someone has to build the decks.
* It'd be helpful if it could notice when you're going about learning
  in a suboptimal way, and advise you. For a start, it could
  extrapolate your performance to your expected items/year per
  minute/day, and compare to a benchmark. Likewise for forgetting
  rate.


http://www.mail-archive.com/kragen-tol@canonical.org/msg00243.html
So can we test this theory? E.g. how much Spanish vocab could I learn
in 4 hours by adapting sm2.py and my parallel-corpus deck? Then what's
needed for retention?


http://news.ycombinator.com/item?id=1603562
http://www.fourhourworkweek.com/blog/2007/11/07/how-to-learn-but-not-master-any-language-in-1-hour-plus-a-favor/

~/prama/python/darius/guesslanguage/parallel-corpora/leipzig/unpacked/se100k/words.txt

http://llt.msu.edu/vol5num3/pdf/stjohn.pdf

>  I'd like to see a program or website for learning a reading knowledge
>  of French (or whatever), assuming you've previously learned a
>  different foreign language like Spanish and don't need
>  elaborately-structured lessons -- where you just allocate it, say, 15
>  minutes a day, and it grows your vocabulary, etc., adaptively according
>  to the time you give it, starting with the roughly most important
>  stuff so you can get some benefit no matter how far you take it. Have
>  you seen anything like that via your class? You could make a start at
>  this sort of thing with a corpus of French text to get word
>  frequencies, plus a French-English dictionary and an adaptive
>  flashcard program. Of course that leaves out grammar and idioms and
>  conversation.