IDyOT: A Computational Theory of Creativity as Everyday Reasoning from Learned Information - Computational Creativity Research: Towards Creative Machines - page 133

(a)

group t1 t2 t3

group s1 s2 s3 s4 s5 s6

word The horse raced past the barn

morpheme h s re s t p s t b n

phoneme h s r e s t p s t b n

time

v2 v3

(b)

u3 u5 u7

group v1

passed the barn

group u1 u2

v2

u4 u6 u8

word The horse race past the barn

morpheme h s re s p s t b n

phoneme h s r e s p s t b n

time

Fig. 7.2 Representation and recognition. a A sequence of symbols in IDyOT memory is composed
of subsequences and supersequences; here phonemes are given in International Phonetic Alphabet
for UK standard English; larger groups are marked as words or arbitrarily named groups; note
though that some of these groups correspond with fairly conventional syntactic labels, such as
“noun phrase”. In this simplified diagram, no alternatives are shown; however, each arrow in the
diagram is associated with two distributions (one in each direction) over the set of symbols that
may be at the appropriate end. The arrow heads are a rough guide to the flow of information
around the diagram as phonemes are perceived and words assembled. Dotted arrows are relatively
low-probability implications. It is important to understand that each higher level is inferred from
the level below, so the linguistically motivated labels on the group level and words themselves
are added to assist the reader, and would be arbitrary symbol names in the software. b Amore
complex parse, including one ambiguous possiblity. The shaded angles indicate where two values
from a distribution are included in the diagram. Note that the lower reading leads to expectation
of a continuing sentence (indicated by dotted rightwards arrows ). Necessarily for a 2-dimensional
diagram, the two parses are not synchronised on the time dimension

interacting features, associated together by means of sequences of multi-dimensional
symbols, to admit multi-dimensional prediction. This is the system used in IDyOM
[ 33 ] (see Sect. 7.3.1 ), and adapted for multidimensional language models byWiggins
[ 54 ]. A key feature of IDyOM is its ability to integrate information from different
features with weights determined by their information content [ 34 ], and the same
idea is used in IDyOT.
Given Conklin's notion of viewpoint [ 8 ] and the associated mathematics, it
becomes possible also to represent propositional meaningwithin the statistical frame-
work: one simply incorporates representations of the meaning (perhaps drawn from