haiku-for example, the final sample of four consecutive short sentences.
At first it seemed very funny and had a certain charm, but soon it became
rather stale. After reading a few pages of output one could sense the limits
of the space in which the program was operating; and after that, seeing
random points inside that space--even though each one was "new"-was
nothing new. This is, it seems to me, a general principle: you get bored with
something not when you have exhausted its repertoire of behavior, but
when you have mapped out the limits of the space that contains its be-
havior. The behavior space of a person is just about complex enough that it
can continually surprise other people; but that wasn't true of my program.
I realized that my goal of producing truly humorous output would require
that far more subtlety be programmed in. But what, in this case, was meant
by "subtlety"? It was clear that absurd juxtapositions of words were just too
unsubtle; I needed a way to ensure that words would be used in accordance
with the realities of the world. This was where thoughts about representa-
tion of knowledge began to enter the picture.
From RTN's to ATN's
The idea I adopted was to classify each word-noun, verb, prepositIOn,
etc.-in several different "semantic dimensions". Thus, each word was a
member of classes of various sorts; then there were also superclasses-
classes of classes (reminiscent of the remark by Vlam). In principle, such
aggregation could continue to any number of levels, but I stopped at two.
At any given moment, the choice of words was now semantically restricted,
because it was required that there should be agreement between the various
parts of the phrase being constructed. The idea was, for instance, that
certain kinds of acts could be performed only by animate objects; that only
certain kinds of abstractions could influence events, and so on. The deci-
sions about what categories were reasonable, and whether each category
was better thought of as a class or a superclass, were quite complicated. All
words were branded in several different dimensions. Common preposi-
tions-"of", "in", etc.-had several distinct entries, corresponding to their
distinct usages. Now, the output began to be much more
comprehensible-and for that reason it was funny in a new way.
A little Turing Test
Below, I have reproduced nine selections, carefully culled from many
pages of output from later versions of my program. Along with them are
three (seriously intended) human-written sentences. Which?
(I) Blurting may be considered as the reciprocal substitution of
semiotic material (dubbing) for a semiotic dialogical product
in a dynamic reflexion.Artificial Intelligence: Retrospects 621