We have something like that at VISL, but with zero statistical or machine learning or AI aspects.
We instead write a few thousand rules by hand (largest language has 10000 rules) that look at the context - where context is the entire sentence, and possibly previous or next sentences - to figure out what meaning of a word is being used and what it attaches to.
Input: "They're looking at writing an AI which can in some sense understand what is being said."
Output: http://dl.dropbox.com/u/62647212/visl-eng.txt , http://dl.dropbox.com/u/62647212/visl-eng.png
This kind of system takes longer to develop and refine, but it also doesn't have any of the statistical problems. 95-99% "understanding" of text? Sure, we can do that. Statistics top out long before, and then have to add in rules to get the last 5-10%. And where statistics require giga- or terabytes of text, rule based systems only require a single example of a valid grammatical construct or word usage.