--- Mark Waser <[EMAIL PROTECTED]> wrote: > > we want computers to > > understand natural language because we think: if you know the syntax, > > the semantics follow easily > > Huh? "We" don't think anything of the sort. Syntax is relatively easy. > Semantics are AGI.
Not really. Semantics is an easier problem. Information retrieval and text classification systems work pretty well by ignoring word order. (I know there are exceptions where word order is important, like "Bob hit Alice"). On the other hand, parsing natural language is an unsolved problem. Word order is extremely important in artificial languages such as C. Almost any rearrangement of the words in a program will change its meaning. When you write a compiler, you develop it in this order: lexical, syntax, semantics. This approach does not work for natural language. Syntax depends on semantics. The parse of "I ate pizza with X" depends on the meaning of X. Children learn language in the order: lexical, semantics, syntax. Artificial languages were designed to be processed efficiently on fast, sequential computers in a clean environment. Thus we have strictly unambiguous parsing, a stack-based short term memory (context free grammar), long chains of sequential dependencies, and a lack of error recovery capabilities. On the other hand, natural language evolved to be processed on massively parallel computers with slow, unreliable components in a noisy environment with a small short term memory (about 100 bits), and a slow long term learning rate (a few bits per second). Thus, we have noisy, ambiguous messages whose resolution requires integrating thousands of learned lexical, semantic, syntactic, and pragmatic constraints simultaneously, and a language structure that allows for gradual but efficient learning and updating of these constraints in the course of normal communication. This is not to say that a neural architecture like the brain is the best way to process natural language. Rather, natural language is the most efficient language for communication between the neural computers we already have. -- Matt Mahoney, [EMAIL PROTECTED] ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936
