Google works so well because if you can think of something to say, the odds are good (well, OK at least) that somebody has thought about saying the same thing in the same way before. Add a good relevancy engine and you're rocking.
Then people start complaining. 'I wrote film, and the page I was looking for talked about movies instead. What a stupid search engine', and so the dream of natural language search is born. In reality it is not the ability to frame the questions in natural language that is so sought after, but rather the listeners ability to extract a broad meaning from the question instead of a stupid text match.
Dave Winer had an idea about search through an open directory of directories, i.e. a general search of anything remotely resembling a catogorization of knowledge.
I sort of had that idea myself a while back (before reading Dave, honest!)

Now from another perspective comes Jon Udell's idea of XPath everywhere. By cleaning up our writing and using XPath we suddenly have a powerful search functionality on our hands.

Now combine the two: Suppose we had a directory of directories/ontologies. Suppose that this directory was XPath searchable. On this data, XPath does exactly what I claim we would really want natural language search to do: Search world models in a form sufficiently close to language to be useful.

The complexity of the search and of the ontology storage must be adressed of course (clearly generating any possible ontology consistent with a given set of documents (based on word occurence) is unlikely to be feasible).

Some more related Udell thinking on this matter.

And then of course the main reason to push forward