Skip to content

Latest commit

 

History

History
28 lines (20 loc) · 2.61 KB

language-learning-treated-as-breadth-first-search.md

File metadata and controls

28 lines (20 loc) · 2.61 KB
Raw
title date tags
Language learning treated as breadth-first search
2024-04-22
secondary-language-acquisition
automation
law-of-diminishing-returns
anti-advice
simple-suboptimal-solutions

I am, emphatically, not the language learning type. I've done enough of it over my life to know this. It's certainly one of the better hobbies out there: You can do it for free (in principle), you can sink as many hours as you care to into it, and if you get good enough at it you get to reap some unique cultural and economic1 benefits. But there's a reason I picked up Python when 14 year old me decided he wanted to get ahead in life instead of (say) German.

That being said: Of the languages I have studied, I liked Latin the best, hands down. Why? Because of the emphasis on reading and understanding individual words and sentences. Reading lets you take your time, scour your memory for the meanings of words you haven't seen in a while, jump around the sentence to puzzle out the grammatical complexity.

I place a heavy emphasis on vocabulary instead of grammar. So from the inside learning a language like this feels an awful lot like doing a breadth-first search:

  1. Read the sentence.
  2. IF you recognize all of the vocabulary terms in the sentence, great! Proceed to figuring out the grammar, and parsing what it means (if it didn't already happen unconsciously for you).
  3. If you didn't -- toss the word into GPT-4 and generate 10 example sentences.
  4. Read these new "synthetic sentences".

I also use an "early cutoff" here, because I don't want to make 90% of my reading intake machine-generated. I never throw words from these syntheic sentences back into the LLM. I wait until I run across them again in a human-spoken context, and if I still fail to recognize them there, then they get 10 of their own. If I knew fewer words, I might not do this early cutoff, trading off fidelity in how people actually use the languages for more depth in the search.

I freely admit immersion is the more efficient method for many kinds of goals. I neither have those goals, nor have I ever found the efficiency gain of immersion to be worth its subjective unpleasantness. Class this under "simple, suboptimal solutions" that I can do ten times as much of under my own limited willpower for this drudgery.

Footnotes

  1. The current lingua franca of the business/scientific world gets an exception from my "I don't think this passes the cost/benefit sniff test". There really are such massive returns to being good at English in the modern day, French 200 years ago, Latin 800 years ago, etc. ↩