Ode to a Keatsian Turn: Creating Meaningful and Poetic Instances of Rhetorical Forms - Computational Creativity Research: Towards Creative Machines - page 197

10.4 Once More, with Feeling!

As shown in [ 24 ], it is a simple matter to filter a set of stereotypes by affect, to reliably
identify the metaphors that impart a mostly positive or negative “spin”. But poems
are emotion-stirring texts that exploit much more than a crude two-tone polarity. A
system like Stereotrope must also model the emotions that a metaphorical conceit
will stir in a reader. Yet before Stereotrope can appreciate the emotions stirred by
the properties of a poetic conceit, it must model how properties reinforce and imply
each other.
A stereotype is a simplified but coherent representation of a complex real-
world phenomenon. So we cannot simply model stereotypes as mere sets of dis-
crete properties—we must also model how these properties cohere with each
other. For example, the property lush suggests the properties green and fertile ,
while green suggests new and fresh .Let coher e denote the set of proper-

ties that suggest and reinforce p -ness in a stereotype-based description. Thus e.g.
coher e

( p )

( lush ) ={ green , fertile , humid ,... } and coher e ( hot ) ={ humid , spicy ,

sultry . The set of properties that coherently reinforce another property

is easily acquired through corpus analysis—we need only look for similes where
multiple properties are ascribed to a single topic, as in e.g. “ as hot and humid as a
jungle .” To this end, Stereotrope trawls the Web for instances of the pattern “ as X and
Yas ”, and assumes for each X and Y pair that Y

, arid ,... }

.
Many properties have an emotional resonance, though some evoke more obvi-
ous feelings than others. The linguistic mapping from properties to feelings is also
more transparent for some property / feeling pairs than others. Consider the prop-
erty appalling , which is stereotypical of tyrants : the common linguistic usage “ feel
appalled by ” suggests that an entity with this property is quite likely to make us
“ feel appalled. ” Corpus analysis allows a system to learn a mapping from proper-
ties to feelings for these obvious cases, by mining instances of the n-gram pattern
“ feel P+ed by ” where P can be mapped to the property of a stereotype via a sim-
ple morphology rule. Let

∈ coher e ( X ) and X ∈ coher e ( Y )

denote the set of feelings that is learnt in
this way for the property p . Thus, feeling

feeling ( p )

( disgusting ) ={ feel _ disgusted _ by }

while feeling ( humid ) ={} . Naturally, because this approach can only find obvious

mappings, feeling ( p ) ={} for most p .

However, coher e ( p ) can be used to interpolate a range of feelings for almost any

property p .Let evoke denote the set of feelings that are likely to be stirred by a

property p . We can now interpolate evoke

( p )

( p ) as follows:

( 6 ) evoke ( p ) = feeling ( p ) ∪ feeling ( c )

c ∈ cohere ( p )

So a property p is likely to evoke a feeling f in an audience if p suggests another
property c that is known to evoke f . We can predict the range of emotional responses
to a stereotype S in the same way:
(

7 ) evoke ( S ) = evoke ( p )

p ∈ typical ( S )