Monday, September 3, 2007
The resurgence of rule-based systems
Given our experience in the 1980's with rule-based systems, I am having trouble understanding their resurgence since 2000. From what I can tell, the various products are using the old, weak methods from the past. They are easy to understand and can give explanations for what they decide. They continue to be brittle and inadequate. I keep hearing that some product is being used, despite those facts. "Strong methods" are statistical-based or neural-network based. They can produce impressive results, but are difficult to understand and the results can't be explained easily. My efforts right now are limited to applying weak methods to machine translation. My usual application is translating Dutch to English. Web translators, such as Babelfish, are hampered by lack of vocabulary. Leading edge machine translation is using statistical methods. They have the drawback of depending on having corresponding texts the the source and the destination languages.