2 citations · 2 across the 9 of their papers we have counts for
1 paper · 1 filter
Lisa Bylinina
A language model normally begins training with random word embeddings: whatever 'banana' means must be learned from training corpora. I implement St. Augustine's picture of word le…