Text Extraction

• I competed in the Message Understanding Competitions (10 years ago).
• Systems competed to automatically extract database entries from text.
• The best systems got about 60% precision and 60% recall.
• The text was open and often ungrammatical.
• It was in a restricted domain (central american terrorist events, or rocket launches).
• When MUC-1 started, most systems couldn't even get one result.
• Over the 10 years or so they occured the systems went from toy to marketable.
• The competitions stopped, because there were no more advances.
• The best systems used cascades of finite state automata.
• The key was to build these automata quickly and this was done by people (novices even) marking up documents.