SUB-100 FINALLY
Semantle #15 in 90 guesses. semantle.novalis.org
Conversation
basically same strategy as semantle #14, thought in terms of co-occurrence in similar documents. again ran into a mild issue with polysemous words but now that i know to look for it it's manageable
Quote Tweet
today i tried thinking a lot more about what words were likely to occur in similar positions in similar documents in word2vec’s training set and that worked a lot better than trying to use raw semantic distance only
Show this thread
1
4
the main mental move i'm doing is something like maintaining a "light grip" on what i think the word is "about." like too tight a grip and i don't move far away enough, too loose a grip and there are too many options where to go next
1
6
also slowly developing more of an intuition for what the similarity numbers mean. below 15 or so is incredibly noisy, 25ish is roughly when you're really getting somewhere it seems like?
Replying to
game state at the end. initially i got words that seemed like they could be descriptions of skin, then got to "feathery" on a whim and went for bird words
5
4
Replying to
Quote Tweet
Replying to @RoomeByTheSea
I got to “snout” which was 998 but kept getting mistake after mistake for over 40 more tries until I gave up and thought “whattt? they’re not that similar “ when i saw the desired word
1
3
it's not about how similar their meanings are, it's about what kinds of sentences they're likely to be found in. a sentence like "this animal has a large ____" could be filled in with many different body parts e.g.
1
2
Show replies
Replying to
this gamification of developing intuition for neural nets is... brilliant. I kind of want to build a generic version of this for any model, it actually seems like a great way to QA them
1
3
yeah def!
1
1
Show replies
Replying to
I also find 25-35 as the sweet spot that I'm directionally on the right track but then get stuck. Going to try your co-occurence mental model instead of going by "type" of word (eg, similar adverbs) or "topic" of word (which might overlap with co-occurence)