Information Extraction

• In 1994, when I was working on my PhD, we participated in the Message Understanding Competitions (MUCs).
• The idea is that your system would take a newspaper article, and generate a database template about its important information.
• They were domain specific tasks, so we worked on Central American terrorist events, rocket launches, and corporate executive job movements.
• The templates were pretty accurate (60 percent), but not as good as people (85 percent inter-annotater agreement).
• They ran these competitions 8 times, and in the end they stopped because they essentially solved the problem.
• This technology still works, though you have to build up the domain specific knowledge.