The last few years, the ability of software to generate convincing natural-language prose has become more and more impressive. I first recognized the technology in my life when there was an earthquake last year and the LA Times' website instantly posted a bare bones computer-generated article, far faster than any human could type.<p>At the time, I was not particularly impressed. It seemed like the software had been working from a prewritten template into which it plugged in details like the location, timing, and intensity of the quake.<p>The program in this article is more sophisticated, because it looks for patterns in the data in order to determine which facts should be included in the article. But it sounds like most of the pattern-matching involves simple, prewritten, domain-specific rules. Some person had to make a list of as many basketball concepts as possible and then write functions for identifying them in the box-score. The result is cool, but the program doesn't demonstrate that much more intelligence than the Derrida generator.<p>What would be <i>really</i> cool is if someone figured out how to automate the creation of the robo-journalist's conceptual ontology. I don't think that would be an AI-complete problem, as long as the corpus was limited enough.<p>For sports, articles about baseball games would probably be the easiest place to start, because there are so many games, each of which involve a predictable, linear, number-oriented narrative.