Mystery Number Five - Computational Linguistics and Talking Robots - page 110

Adaptation and learning differ also in that they apply to different ranges of
time and different media of storage (gene memory vs. brain memory).
In artificial agents, in contrast, improvement of the hardware is the work of
engineers, while development of an automatically adjusting cognition is the
work of software designers. Because of this division of labor between hard-
ware and software development, the automatic adjustment of artificial agents
corresponds more to learning than to adaptation. That artificial agents lack the
natural inheritance from parent to child may be compensated for by copying
the cognition software from one hardware model to the next.
The similarity between natural and artificial cognitive agents regarding their
theoretical (declarative) structure and the dissimilarity in their practical (pro-
cedural) implementation, i.e., hardware vs. wetware, may also be illustrated
with language acquisition. Take for example the learning of word forms.
For humans, this is a slow process, taking several years in childhood (first
language acquisition) and in adulthood (second language acquisition). An ar-
tificial agent, in contrast, may simply be uploaded with an online dictionary
and associated software for automatic word form recognition - not just for one
language, but for as many languages as available or desired.
The word form analyses provided to the artificial agent in this way specify
(i) the morphosyntactic properties (e.g., part of speech) formalized as proplet
shells (cf. NLC'06, Sect. 4.1), and (ii) the core values, represented by place-
holders. Thereby, the proplet shells and the core values are orthogonal to each
other in the sense that (i) a given proplet shell may take different core values
and (ii) a given core value may be embedded into different proplet shells.
The following example shows a proplet shell taking different core values:

6.6.1 O NE PROPLET SHELL TAKING DIFFERENT CORE VALUES

proplet shell context proplets

⎡
⎣

⎤
⎦

⎡
⎣

⎤
⎦

⎡
⎣

⎤
⎦

⎡
⎣

⎤
⎦

⎡
⎣

⎤
⎦

sur:
noun: α
cat: pn
sem: count pl
fnc:
mdr:
prn:

sur:
noun: dog
cat: pn
sem: count pl
fnc:
mdr:
prn:

sur:
noun: book
cat: pn
sem: count pl
fnc:
mdr:
prn:

sur:
noun: child
cat: pn
sem: count pl
fnc:
mdr:
prn:

sur:
noun: apple
cat: pn
sem: count pl
fnc:
mdr:
prn:

⇒

The context proplets derived from the proplet shell differ in only one value,
namely that of the core attribute noun , which should facilitate learning.
Context proplets may be turned into language proplets by inserting the ap-
propriate sur values, as in the following example for English: