Get word from vector

[sheavalentine]

It would be useful to have a method which returns the closest word when the model is given a vector.

[piskvorky]

What model?

[sheavalentine]

Sorry... word2vec. Eg

vectors =  Word2Vec.load_word2vec_format(file)
vectors.word_from_vec(vectors['cat']) == 'cat'

[gojomo]

You can already more-or-less do this by passing a single positive example to most_similar(), like so:

vectors.most_similar(positive=[vectors['cat']],topn=1) 

I suppose as a convenience, it could detect a single ndarray (like it detects a single string) and just assume this is what the caller wants...

[sheavalentine]

! Oh... that's not entirely obvious. The expectation was that positive needed to be a list of strings.

[gojomo]

Yes, the doc-comment should be more clear... and any future/revised tutorials should include a few examples of the multiple styles of arguments it takes.

[sheavalentine]

Thank you - also thank you for making this tool.

[piskvorky]

Good point.

We'll probably split this API into two distinct methods (similar_by_word, similar_by_vector or something) if we want to promote it. Overloading arguments ultimately brings too much trouble and confusion, this was originally supposed to be just an internal quick hack.

@sheavalentine can you do it, and open a pull request?

I assume the by_word version will be just a thin wrapper around the vector version, which is more general.

[sheavalentine]

I can do it, but it'll take me a few days to get the bandwidth. :)

[piskvorky]

No problem.

We plan to make a new release this weekend, so this change will be in the release after that, then.

[tmylk]

@sheavalentine Is this still a needed feature? Do you think we could include your Pull Request into the January release?

[tmylk]

Closing as abandoned