Poetry Generation with PoeTryMe - Computational Creativity Research: Towards Creative Machines - page 251

= ˀ = . ˆ = ʴ =

the G&T strategy, with the following properties: n 750, 0 35, 15, 1,

ʽ = . ˁ = ˃ =

0 01, 2, 1.

que vai depor vai assentar
com arte santo e manobra
na estaca me amparar
eu fiquei obrante e obra

12.4.6 Contextualizing the Poem

We took advantage of PoeTryMe's contextualizer to generate comments that describe
the relation instances used for creating each verse. When the instances do not contain
a domain word, their connection to the domain is also provided. To have a more
friendly format, more easily understood by human readers, we manually developed
a grammar to express semantic relations as text. It is similar to those for the poem
lines, but simpler and with more objective sentences.
Figure 12.6 shows an 'English sonnet' (4-4-4-2), obtained with the G&T strategy,
without any given rhyme pattern and with n

= 750, ˀ = 0 . 35, ʴ = 1, ʽ = 0 . 01,

ˁ = 1. Also, for the generation of this poem, we added two words to our

previous domain, namely: criar (create) and gerar (generate). All the lines rhyme,
but it does not have the maximum possible score because the 4th and 6th lines have
one more syllable than the target 10. The same figure includes the contextualization
of the sonnet and its rough translation. Since two of the sonnet lines are repeated,
there are just 12 sentences in the contextualization.

2, ˃ =

12.5 Comparing Different Generation Strategies

It is not possible to show all the capabilities of PoeTryMe in just a few pages. While
comparing the more objective features of our generation strategies, this section gives
an idea of what we can do with it.
For each generation strategy, we produced 100 sonnets, with standard parameters.
In each generation, a domain was set, with 2 to 10 random words of the graph. For
comparison convenience, we used the same neighbourhood parameters, ʴ = 1 and

ʽ = 0 . 01, and we set the same bonuses and penalties ˁ = 3 and ˃ = 1. As a sonnet

has 14 lines, the best possible evaluation would be 21.

Figures 12.7 , 12.8 , 12.9 , and 12.10 have some numbers regarding the diversity
and the evaluation of the poems obtained respectively with the Ev, G&T, G&T+Ev
strategies and also for the rhyme-pattern variation of G&T (G&T-rp). About the
diversity, each figure reports on the distinct patterns used in the 100 poems, from
the total 4,107. This includes the number of distinct patterns, the average (Avg)
and standard deviation (StDev) of occurrences, and the number of times the most
frequently selected pattern was used. As for evaluation, they report the average and

−