|
When I was grad student in the late 1980's, I was involved in organizing the US Defense Department MUC evaluations of information extraction (IE) technology. (IE systems use natural language processing to extract pieces of information from text and fill out database records, or more elaborate structures.) Since I was designing tasks and effectiveness measures, I would constantly ask the DOD people what they were using IE for. Since I didn’t have a security clearance, they wouldn’t tell me. Finally I found one person who described an actual IE application. The information was extracted from news stories and used to populate relational database records, so that users could search for news stories using SQL! I was appalled: text retrieval is a much simpler problem than information extraction, and even in the 1980's there was perfectly good ranked retrieval technology available (shout out to Matt Koll). Even boolean search, which was plenty mature then, would have been better than what they were doing. (To be fair, by the way, there are other applications where you do need IE, and it's big business these days.) Anyway, I'd be curious to hear of other examples of solving the wrong problem in NLP. (This post is inspired by Richard Lipton's blog posts Problems, Not Algorithms and Big Problems with Big Iron. |
