The Amazing Power of Word Vectors (2016) | Hacker News Reader

The Amazing Power of Word Vectors (2016)

blog.acolyer.org

127 pointsSamuelKillin9 comments

More interesting information from a DOD research lab that resulted in vectors https://www.kaggle.com/c/word2vec-nlp-tutorial/discussion/12...

Department of Energy, not to be confused with Department of Defense. Granted there is a lot of nuclear weapons work at DOE, but LBNL does mostly big science work up the hill from Berkeley

Recently HN featured a podcast interview with "one of the creators of Word2Vec and fastText", documenting some of his professional history.

https://news.ycombinator.com/item?id=13630678

This is an awesome explanation of those papers! Does anyone have any cool examples of word2vec being used in a project? I'd be interested in seeing what people could make with it.

Document type classification. We wanted to predict which of these k classes a new text document was.

We trained 100-dim word vectors on all the text content we currently have, plus some 30,000 wiki articles related to the business. New content comes in, convert words to vecs, average them, and use that resulting vec as the input to a basic classifier.

For how simple that is, the method is unreasonably good. Widely applicable too.

For anyone looking for a simple javascript explorable explanation of this you can quickly download and run in a browser, I just found the following GitHub Project.

Demo:

http://turbomaze.github.io/word2vecjson/

Code:

https://github.com/turbomaze/word2vecjson

The code looks pretty straightforward so I look forward to exploring this playground of a new and fascinating concept.

Update: I'm having a lot of fun exploring the relationships between "nerd," "geek," "dork," etc. in this demo. : )

Word vectors are great. We've also written about them at length.[0] But any one interested in word vectors should also be looking at newer ways of applying neural nets to text. Specifically, convolutional nets with pooling for time are producing great results for clustering and classification.

[0] https://deeplearning4j.org//word2vec

You nailed it with the "German + airlines" example. Up until that point it was tough to read for a newbie like me. Great blog post

Reply on news.ycombinator.com