News for hackers. Stuff that matters.
Re:Deep Learning, NLP, and Representations by nl on Wednesday March 25, 2015 @04:24AM (#9261764)
The maths works when you are acting along the same vector(s), so things like king:male ~= queen:king work. In that example you are doing maths along the 'gender vector'.

In you water/boat/plane example I'm not sure what you are expecting. It seems unlikely that there is a 'planeiness' vector one can move along to find concepts like that.

Re:Deep Learning, NLP, and Representations by navanit on Wednesday March 25, 2015 @04:30AM (#9261778)
The vector for the water/boat/plane is "medium".
Re:Deep Learning, NLP, and Representations by nl on Wednesday March 25, 2015 @04:50AM (#9261814)
Yes, we understand that. Not sure word2vec does.

From memory the default word2vec implementation and data gives you ~500 dimensions. I think some of the examples given are failing because of that limitation.