In the last post we built the data preprocessing required for word2vec training. In this post we will build the network and perform the training on the text8 dataset (source), a Wikipedia dump of ~17 million tokens.

Note that we are implementing the skip-gram version of word2vec since it has superior performance
Continue reading

One of the trends I've been seeing is the use of embeddings for similarity/recommendation in non-NLP domains (this talk and many others). While I've used/thought about word2vec for a while, I wanted to implement this to really get a sense of how it works and ways it can be extended to other domains. Intuitively, word2vec makes sense, and there are a lot of packages that will let you compute it without thinking much about the implementation (e.g. gensim
Continue reading

I've used keras a bit for deep learning recently. In general it is an excellent tool, but I know there are limitations in using a higher-level tool. In trying to implement an initial model in tensorflow, I found that you have to think about fundamentals of data flow and building computation graphs -- as opposed to machine learning a la sklearn
Continue reading