Natural Language Understanding in Soar
The Soar
architecture was not designed to
support natural language understanding (NLU). However, NL-Soar,
as described by
Lehman, et al (1991) and Lewis (1993),
is a system that comprehends language utterances (in this case written
utterances) that was built on top of the architecture, utilizing its
features. In particular:
- Problem Spaces
exist for comprehension, language, constraint checking and semantics.
Comprehension is the top space.
- Productions
represent different types of knowledge -- morphology, syntax, semantics, etc.
The parallel nature of the
elaboration phase
allows the application of different knowledge sources simultaneously.
- The attribute-value
representation in working memory consists of two
network models. The
utterance model reflects the
(structural) dependencies in the utterance and refers to the
situation model which is a representation of the utterance's meaning.
- Impasses
arise in
the comprehensional problem space when comprehension is not
immediate. This causes the different problem spaces to be accessed
in order to elaborate and change the utterance and situational models.
- Chunking allows
recognitional productions to be built
in the comprehension space, making comprehension automatic.
Additionally, chunking's
generalization ability is utilized so
that chunks are not built for specific word combinations but only for those
words (and structural identifiers) upon which problem solving actually
depended. As recognitional
chunks are built, there is a corresponding speedup in comprehension;
this is an example of Soar's ability to move from
problematic to routine behavior.
Other capabilities for NLU implemented in NL-Soar include expectation
(the application of top-down knowledge to comprehension) and
recognitional repair (the
resolution of local ambiguity such as garden path effects). NL-Soar
may be considered not only as a NLU capability for cognitive
architectures but as a theory of human comprehension as well. As such,
it compares favorably to the predictions and capabilities
of other cognitive models of
language processing. Additionally, because the model is placed on top
of a general cognitive architecture (as opposed to some stand-alone
mechanism), the architecture provides insights into possible actual mechanisms
of language comprehension, strengthening Soar's claim as a
unified theory of cognition.
To go to the Soar capabilities list, press UP
To return, press HOME.
To go to the next document, press NEXT.