Domain Generality
- The NL problems can be over come or hacked.
- The real problem is domain generality.
- Existing AI systems work well in particular domains.
- For example, the last Message Understading Competition (1997) was
on missle launch events.
- Lots of systems were switched to this domain with a month of
development time, and generated pretty solid database
templates. (NE rec > 95%)
- Existing technology can combine machine learning techniques
with hand coding to implement systems in specific, relatively
restricted domains.
- However, people know about a lot of domains. For example, pretty
much everyone knows about the 3D world, food, and interpersonal
relations. Additionally everyone has a range of domains that
they know about; e.g american politics, Arsenal, and
N. London traffic.
- The enCYClopedia approach (Lenat) doesn't seem to work.
- Calling out to the Internet might help, but the system still
needs to process that information.
