kimcando / BoostcampAITech3-PaperReading-Embedding

Boostcamp AI Tech 3rd / Basic Paper reading w.r.t Embedding
13 stars 3 forks source link

[week3] Negative sampling 과 subsampling의 차이 #16

Closed kimcando closed 2 years ago

kimcando commented 2 years ago

NCE(Noise Contrastive Estimation) posits that a good model should be able to differentiate data from noise by means of logistic regression. We define Negative sampling (NEG) by the objective which is used to replace every log P (wO |wI ) term in the Skip-gram objective. Thus the task is to distinguish the target word wO from draws from the noise distribution Pn(w) using logistic regression, where there are k negative samples for each data sample.

sub- sampling of frequent words during training results in a significant speedup the vector representations of frequent words do not change significantly after training on several million examples.

  • 헷갈리는 포인트
  • negative sampling은 negative 라벨인 단어를 넣어줌으로써 모델 성능을 높이겠다는 것이고,
  • sub sampling 은 많이 등장하는 단어일 수록 정보성이 적다. 즉, 쓸데없는 단어로 취급해서 덜 학습에 사용하겠다는 것

Summary