How to inject training examples in gensim Word2Vec?

Question

I'm using gensim to create a Word2Vec model. I'm wondering if there is a way to feed the gensim class Word2Vec with my examples [(target, context1), (target, context2), ...] instead of feeding it with sentences.

Thanks in advance

Answer 1

The Gensim Word2Vec class expects a re-iterable sequence where each item is a list of string word tokens. It then does the construction of the inner 'micro-examples' (context-word -> target-word in skip-gram, or context-window -> target-window in CBOW) itself.

There's not an alternate interface, or easy extension-hook, for changing the micro-examples. (Though, as the source code is available, it's possible even when not easy to change it arbitrarily.)

If you only ever need single-word contexts to single-word targets, and are OK with (as in standard word2vec) every A B pair to imply both an A -> B prediction and a B -> A prediction, you may be able to approximate your desired effect by the proper preprocessing of your corpus, completely outside Word2Vec code.

Specifically, only ever provide 2-word texts, of exactly the word pairs you want trained, as if they were full texts.