I could imagine using a neural net to make a first guess at classifying documents, but actually using one for parsing <i>structured</i> data is crazy.
Neural nets are notoriously hard to debug since after you've trained them all you have to work with is a set of coefficients. I can't imagine the silly things they probably had to do to get it to reliably handle corner cases.<p>Using a net to generate gibberish is also pretty silly, especially when there's a much better tool designed for just that purpose: <a href="http://dev.null.org/dadaengine" rel="nofollow">http://dev.null.org/dadaengine</a><p>An online version is here: <a href="http://herbert.the-little-red-haired-girl.org/en/dada/index.html" rel="nofollow">http://herbert.the-little-red-haired-girl.org/en/dada/index....</a>