20050508-1741

I find today that schools are increasingly using computer programs to grade writing assignments.[1] While this would be awesome if they could get it to work reliably, I fear that like the one student quoted states, all that students will really learn is how to beat the system. Natural language parsers are a notoriously hard problem to solve, and key word searches are incredibly difficult to get working well. While much progress has been made, look at the evolution of search engines, some of this has been done by avoiding the problem entirely, google's page rank for instance, specifically avoids part of the natural language problem by categorizing its keyword search by the page rank, essentially trusting the collective judgment of real humans. Grading an essay is rather more complex than the incredibly simplistic form of data mining involved in a search engine. The computer-grader needs to not only detect the presence of certain key words, but needs to determine if a correct relationship has been demonstrated between them, in any of the various ways a student could word it. Grammar checking is fairly straight forward, but even the best spelling dictionaries we have available to date do not contain a full set of English, including the jargon specific to the various fields that the high school and college student will need. However, while I maintain that grammar is "fairly straight forward," it is never the less a significantly difficult problem that it is not unheard of for a human writer to disagree with them various grammatics and be right. Thus overall, while I hold hope that this sort of grading might at some point be feasible, and, to the extent that it is achieved, does provide a useful service in removing some of the repetition and drudgery of teaching, I maintain that it is likely a mistake to deploy it in any sort of unverified (by which I mean a real live human verifies it came up with a reasonable result) manner.

[1] http://channels.netscape.com/ns/news/story.jsp?id=2005050713030001479371&dt=20050507130300&w=APO&coview=