Word2Vec Skip-Gram ak CBOW

The original Word2Vec papers present CBOW and continuous Skip-Gram architectures for learning distributed word representations. CBOW uses context words to predict a target; Skip-Gram uses a target to predict surrounding context words. The later paper describes negative sampling, subsampling of frequent words, and phrase detection. The learned vectors are useful representations evaluated on specified similarity and relationship tasks; they are not context-sensitive and the papers do not establish particular commercial deployments. The learned input weight matrix supplies the word vectors used after training. The later paper reports subsampling frequent words as a way to speed training and obtain more regular representations. Mikolov and colleagues at Google introduced the architecture papers in 2013. Their vector-analogy example includes king − man + woman ≈ queen; this illustrates a reported geometric relationship, not a guarantee for every vocabulary.