Page 83 - Biomimetics : Biologically Inspired Technologies

Bar-Cohen : Biomimetics: Biologically Inspired Technologies DK3163_c003 Final Proof page 69 21.9.2005 11:40pm

Mechanization of Cognition 69

The solution is to invoke two new confabulation architecture elements: a language hierarchy
and consensus building. These are sketched next.

3.3.2 Language Hierarchies

There are many reasons why the architecture of Figure 3.1 does not solve the phrase completion and
sentence continuation problems. First of all, this architecture disallows the learning and application
of standard language constructions such as multi-word conceptual units (e.g., New York Stock
Exchange, which we will refer to as phrases), variable element constructions (VECs, e.g., ___
went to the ___), and pendent clauses (e.g., the success of her daughter was, except for
ordinary daily distractions, foremost on her mind). Standard constructions are important
elements of all human languages (although, they differ and can take on different forms, in different
languages) and a comprehensive architecture must include provisions for learning and representing
them.
For the problems of phrase completion and sentence continuation, the architecture of Figure 3.2
is much more capable than that of Figure 3.1. For example, consider again the problem of finding
the next word for the assumed fact phrase The canoe trip was going smoothly when all of a
sudden. Now, the first thing that happens is that this phrase is parsed, meaning that the words are
re-represented at the phrase level. This happens almost instantly by using the knowledge bases
which proceed from the word level to the phrase level. The parsing process, which is described
next, proceeds in a rapid ‘‘rippling wave’’ of thought processes running from the beginning of the
assumed fact word string to the end.
To start, the first word, The, of the string goes up first. These links (which in accordance with the
knowledge base design described in the caption of Figure 3.2) only go to the first phrase lexicon.
A C1F on this first phrase lexicon yields an expectation consisting of those symbols which
represent phrases that begin with the word The. The second word lexicon then sends links upward
to the first and second phrase regions from the symbol for canoe. C1Fs on phrase regions, one and

Figure 3.2 Single-sentence hierarchical confabulation architecture for proper English phrase or sentence com-
pletion. The lower row of lexicons is used to represent words, as in the architecture of Figure 3.1. Again, knowledge
bases link each of the first 19 of these 20 word-level lexicons to all of the lexicons to their right. Positioned exactly
above the word-level lexicon row is a row of 20 phrase-level lexicons. These phrase lexicons represent word
groups (and other standard language constructions, although these will not be discussed much in this introductory
chapter). Each phrase lexicon has at least 126,000 symbols (63,000 single words and punctuations and the 63,000
most common multiple word groups). Knowledge bases connect each phrase lexicon to each of the phrase lexicons
which follow it. Knowledge bases also connect each phrase lexicon to all of the word lexicons except those that lie
to its left. Finally, knowledge bases connect each word region with all of the phrase regions except those that lie to
its right. This architecture has a total of 800 knowledge bases. On average, each knowledge base contains roughly
a million individual items of knowledge. The capability of this architecture is a practical demonstration of the main
premise of the author’s theory of vertebrate cognition; namely, that lots of simple knowledge, along with a single,
simple, information processing operation can implement all of cognition.