CogSci Summaries home |
Peterson, J., K. Mahesh & A. Goel (1994) Situating natural
language understanding within experience-based
design. Int. J. Human-Computer Studies. 41, 881-913
Author of the summary: Jim Davies, 1999, email@example.com
Cite this paper for:
In the past there have been two ways of making a natural-language
understanding system (NLU) use inference and problem solving:
- The KA language system
- Structure is defined in terms of component and substance.
This research is new in that it is like the second method above except
that the linguistic and non-linguistic modules communicate. Both
modules come to partial conclusions, then negotiate the final
- give the system knowledge and the ability to reason about it
This has not been successful for the following reasons:
- The linguistic and non-linguistic known things do not
work well together
- It is unclear what knowledge to put in or how to acquire it.
- Take some other cognitive task (e.g. planning) and make a
specialized front end to the system. In these systems there is
typically no feedback because the NLU is just a front-end. Thus
it has trouble dealing with many parts of language.
The system is called KA. Here is how it works:
This aids ambiguity resolution in two ways: (p883)
- form tentative design spec
- retrieve an analogue
- use analogue to resolve ambiguities and correct errors
The designs in memory are represented in terms of Structure, Behavior,
and Function (SBF). Structure is defined in terms of component and
substance. Substances have locations with respect to components, and
they also have behavioral properties with corresponding parameters.
- the ontology of device designs insures some consistency
- interpretations most compatible with past experience are
produced because it uses past experience to reason.
Function is defined in terms of the behavioral state the device takes
as input and the state it has at output.
The internal causal behaviors of a device are viewed as sequences of
state transitions between behavioral states. They express the causal,
structural, and functional context in which the transformation of
state variables occur.
KA uses an early-committment/robust error recovery method to resolve
KA goes through each word of the sentence, and guesses at a
parse. There is a tentative interpretation made based on a the parts
of speech and a set of retrieved analogues. Content words in the text
activate corresponding parts of memory. (p893).
Consistency checking is done in accordance with the SBF ontology. The
semantic network identifies inconsistent inferences and resolves them
with the most common interpretation. Less common ones are retained for
later use if needed (p894).
A current interpretation's differences with the design spec are
resolved by making small modifications. (p899)
The third investigation showed that KA can usefully interpret NL
feedback about a design. (p903)
Summary author's notes:
Back to the Cognitive Science Summaries homepage
Cognitive Science Summaries Webmaster:
Last modified: Thu Apr 1 10:27:20 EST 1999