How can we implement word sense disambiguation using word2vec representation?

Question

I know how word2vec works, but I am having trouble with finding out how to implement word sense disambiguation using word2vec. Can you help with the process?

Answer 1

As @sam-h mentions in his comment, this is an area of ongoing research.

There's no standard or automatic approach, so there's no one best-practice to recommend – you'll likely have to sift through the various papers, in the list `sam-h provided and from elsewhere, for ideas.

In many cases, approaches don't use standard word2vec – adding extra steps before or during training – because standard word2vec is oblivious to the fact that a single word-token might have multiple contrasting senses. As a result, the standard word2vec vectors for words with many senses can wind up with a single vector that "mushes together" the many distinct senses.

One interesting write-up that does manage to bootstrap a model of multiple-senses from existing, word-sense-oblivous word-vectors is described in the paper "Linear Algebraic Structure of Word Senses, with Applications to Polysemy", which also has a less-formal blogpost write-up.

Essentially, by assuming the rich space of all standard word-vectors actually draw from a smaller number of "discourses", and interpreting word-vectors as some combination of the alternate "atoms of discourse" (for their difference senses), they can tease-out the alternate senses of word-tokens that began with only a single vector.