word2vec-two-architectures-cbow-skipgram

IN premise — summaries/2026/08/24/wiki-Word_embedding-chunk-1.md

Created 2026-08-24T17:11:28+00:00

Word2vec (Mikolov et al., 2013, Google) provides two training architectures: CBOW (predict target word from surrounding context) and Skip-gram (predict context words from target word), with Skip-gram generally preferred for lower-frequency words

Summary

Word2Vec can learn word meanings in two directions: either guessing a word from its neighbors (CBOW) or guessing the neighbors from a given word (Skip-gram), and the second approach tends to capture rare words better. This matters because the choice of architecture shapes how well a system understands uncommon vocabulary, which affects downstream tasks like translation and search.