Poetry Generation with PoeTryMe - Computational Creativity Research: Towards Creative Machines - page 255

Properties

Strategy G&T-rp

Generations/line 800

mais testante que a pessoa pura
tao triviais os fez e taobatidos
dos gastos feitos grandes e comidos
jubiloso e ovante sou eu

Progressivemultiplier ( ) 0.35

Rhyme pos. multiplier ( ) 10

Diversity

Distinct patterns 590

Avg(uses/pattern) 1.99

StDev(uses/pattern) 2.0

de ti h´aumfesto pela largura
do louro deuscalcados e batidos
um dos batidos ministros vencidos
mais testante que a pessoa pura

Most usedpattern 20 times

Evaluation

Avg(gen. time) 45.13 seconds

Avg(score) -1.61

StDev(score) 2.83

de todo o navio de cargueira
tudo e naturals´oeu for¸ada
uma janota etaful terceira

Best poem

Score -15

Domain africano (african)

odiento (hateful)
cangalhas (packsaddles)
parlamentario (negotiator)
festivo (merry)
batido (whipped)
debilitante (debilitating)
debito (debit)
gatunar (to steal)

dascangalhasquetrouxeram puxada
duma certa pessoa de veleira
abarcou-me a festa churrascada

Fig. 12.10 Figures of 100 sonnets with rhyme pattern ABBA ABBA CDC DCD , obtained with the
G&T strategy, together with the best-scoring sonnet

12.6 Concluding Remarks

This chapter presented PoeTryMe, which can be seen as an exploratory creativity
system acting in the complex conceptual space of poetry. One of its distinctive char-
acteristics is definitely the wide range of possible parameterisations, which make it
highly customisable and thus a powerful creative support tool. PoeTryMe, originally
described in 2012 [ 13 ], may be used with different lexical resources and generate
poems with different configurations. Generation grammars may be edited manually,
in order to cover new linguistic constructions. Generated sentences can be selected
and organised according to different criteria, in what we called the generation strate-
gies. Since PoeTryMe's earlier instantiations, significant improvements were made,
including an improved G&T strategy, where rhymes are more common, and also the
new contextualizer module.
In this context, we described an incremental approach to poetry generation, in
Portuguese, using PoeTryMe as a starting point. We detail each step we took towards
poetry generation and the main generation strategies implemented—evolutionary
(Ev), generate-and-test (G&T), and their combination (G&T+Ev). In the end, the
most objective features of their results were compared, together with illustrative
examples of the best scoring poems.
PoeTryMe was conceived having inmind the idea that, for a computational system
to be accepted as creative, it must convey a perception of creativity into its public, i.e.,