When I studied NLP one of the first thing that was drilled into my head by that book (<a href="http://www-nlp.stanford.edu/fsnlp/promo/" rel="nofollow">http://www-nlp.stanford.edu/fsnlp/promo/</a>) was that these are all <i>statistical</i> techniques and context based understanding of a corpus is still a while away. The problem really is to condense all of that knowledge about the real world into a form that is computable. As that really famous example goes, how do you tell a computer that water is 'wet'?<p>However, this is still an extremely important development and it shows that sooner or later by leveraging the things computers are good at we will be able to solve such problems. I really don't think that it is something impossible to do. It just requires a new approach that none of us have thought about. After all not too long ago nuclear fission was said to be impossible, until a chain reaction was conceptualized.<p>Perhaps AI will go this way too, the difficult almost intractable problems will turn out to be easy under some new paradigm. Perhaps not. We'll never really know until we try.<p>So kudos Regina Barzilay and her team for pushing the limits.