Mystery Number Five - Computational Linguistics and Talking Robots - page 115

used: (i) looking several times at the same animal in different postures and (ii)
comparing several different instances of the same species (token). 27
The new zebra type may be constructed in two ways. If the type for horse
is already available to the artificial agent, the new type in question may be
constructed as a combination of the types for small horse and for black-and-
white stripes. If the type for horse is not available, in contrast, the type for
zebra may be constructed from more basic types, e.g., for head, body, legs,
tail, and black-and-white stripes - such that a later first encounter with a horse
will result in a type defined as big zebra without black-and-white stripes .
The alternative definitions of a zebra in terms of a horse or a horse in terms
of a zebra are equivalent for all practical purposes. Either type will allow the
agent to recognize any future zebras, resulting in instantiations (zebra tokens)
serving as the (procedural) core values of context proplets. In this way, an ar-
tificial agent without language may learn an open number of new concepts. 28
In vision, these concepts are built up as new combinations of elementary, uni-
versal types implemented as line, edge, ridge, color, etc. detectors.
Next consider an artificial agent with language learning a new word. After
the agent has acquired the concept type for a zebra, a human points at the ze-
bras and utters the French word zèbre . 29 Given the similarity between context
proplets and language proplets, the artificial agent is able to extend its French
vocabulary by (i) copying the context proplet to the language level and (ii)
inserting the French surface into the sur slot of the copied proplet.

6.6.9 L EARNING A NEW WORD

stage 1 stage 2

sur: zèbre

noun: zebra

fnc:
cat: sg
mdr:
prn:

language level

zèbre zèbre

sur: sur:

noun: zebra noun: zebra

fnc:
cat: sg
mdr:
prn: 465

fnc:
cat: sg
mdr:
prn: 465

zebra context level

27 In machine learning, this kind of learning is called learning by observation.
28 In other words, DBS does not attempt to bootstrap all cognition from a small set of universal, basic
language concepts, in contradistinction to Wierzbicka (1991).
29 This surface was chosen to differ from the core value, as in the examples in 6.6.3.