3 ms·
As a ML layperson I had tried the first part a few months ago, using word2vec to find closest neighbors of word pairs and filter close clues of opponent words a
by jerkstate 5y ago
As a ML layperson I had tried the first part a few months ago, using word2vec to find closest neighbors of word pairs and filter close clues of opponent words and it was really effective at finding single word clues (nearly 100%) but only about 20% at finding clues for pairs. It looks like their research finds clues for pairs with 57% effectiveness, but to my layperson reading they did not attempt to find clues for 3 or 4 words, which humans can regularly do in this game.
- andyxor 5y agotry transformer models, like BERT, for pairs or sentences, I saw similarly that BERT is not as good as word2vec for single word similarity but shines for sentences (where context is important)