Similar things have been around for a long time.
For instance, you can train something like LDA on the contexts that words appear in. Also I've seen a system that uses a neural net to generate a 16,000-bit vector and very interesting things happen when you use AND and OR.