Introduction: How to Build a Talking Robot - Computational Linguistics and Talking Robots - page 14

To be effective, an agent's autonomous control should be based on flexible
thought 33 rather than some rigid decision mechanism. In DBS, the analysis of
thought proceeds from the following assumption:
Assumption : We can study thought by studying natural language.
This is not new, except that DBS starts from the modality-dependent unana-
lyzed external surfaces of natural language to model the transmission of infor-
mation between the speaker and the hearer computationally.
The mechanism is based on pattern matching (Sect. 2.6) and works just as
well for agent-external nonlanguage phenomena (L&I'05). Treating nonlan-
guage and language contents alike in recognition and action suggests treating
them alike also in the think mode. This has the advantage that (i) they may be
processed the same in a “horizontal” time-linear fashion and (ii) their “verti-
cal” interaction for reference is facilitated (cf. FoCL'99, 5.4.1, 5.4.2).
The assumption of a close structural similarity between language and thought
has the following consequences on the DBS software design:
Consequences:
1. Language content and nonlanguage content are coded practically the same,
using flat feature structures (proplets) and addresses.
2. Language content and nonlanguage content are stored the same, using a
content-addressable database (Word Bank).
3. Language content and nonlanguage content are processed the same, using
a time-linear algorithm (LA-grammar).
4. The formal similarity between language content and nonlanguage (con-
text) content is used for a cognitive reconstruction of reference .
Is it possible for such a design to be psychologically real ? If humans and
machines can communicate freely with nothing but time-linear sequences of
modality-dependent unanalyzed external surfaces, then their cognitions may
be equivalent - at certain levels of abstraction. Each equivalence level must be
continuously verified by long-term upscaling for many different natural lan-
guages and applications, serving as a broad empirical foundation.
To give an idea of the functionalities demanded by a computational model of
a cognitive agent with language, the following Part I is organized around five
mysteries of natural language communication. Each mystery is answered with
a Mechanism of Communication, MoC-1-MoC-5.
33 The term thought is used here to refer to the cognitive processing of nonlanguage content, i.e., content
independent of language surfaces.