Poetry Generation with PoeTryMe - Computational Creativity Research: Towards Creative Machines - page 254

que a uma naceja ounarceja
eu tenho uma narcejanaceja
de segunda pessoa davideira
dos do¸ais ou das aguassem pereira

Properties

Strategy G&T

Generations/line 800

Progressivemultiplier ( ˀ

) 0.35

Diversity

Distinct patterns 593

a pereira em todas asdo¸ais
as videiraspasmadasdefolhais
dos meusdivinos do¸ais de videira
das do¸ais quetrouxeram macieira

Avg(uses/pattern) 1.95

2.00

StDev(uses/pattern)

Most usedpattern 14 times

Evaluation

Avg(gen. time) 3.89 seconds

eu tenho uma videira parreira
nos fortes do¸ais na grande macieira
as videiraspasmadas de do¸ais

Avg(score) -6.25

3.32

StDev(score)

Best poem

Score -17

Domain do¸al (variety of grape/pear)

naceja (South American snipe)
revalorizar (to revalue)
dissoluvel (dissoluble)

dos do¸ais ou das aguassemvideira
eu tenho uma videira parreira
a pereira em todas asdo¸ais

Fig. 12.8 Figures of 100 sonnets with the G&T strategy, and the best-scoring sonnet

Properties

soumauversejador masseiversar
circunspeccionareponderar
versarem frente ir a estudar
tratarem frente ir a combinar

Strategy G&T+Ev

Generations/line 3

Reproduction probability 0.4

Crossingprobability 0.5

Remain probability 0.3

Stopafter stagnation 8 generations

circunspeccionareponderar
vai abordar jeito p'raversar
quero versarevolver rosa branca
ponderar a dor, pensar a labuta

Diversity
Distinct patterns 517

Avg(uses/pattern) 2.25

StDev(uses/pattern) 2.34

Most usedpattern 14 times

soumauversejador masseiversar
minha mae manda-me volver avolta
constar a dor, consistir a labuta

Evaluation
Avg(gen. time) 36.25 seconds

Avg(score) -5.23

StDev(score) 3.31

Best poem

vaipraticar jeito p'raversar
sempre a consistir e a constar
usarem frente ir a praticar

Score -18

Domain abezerrado (like a calf)

versar (to rhyme)

Fig. 12.9 Figures of 100 sonnets with the G&T+Ev strategy and the best-scoring sonnet

We should also stress that the domains of these specific poems were defined by
random words. Therefore, although the generated sentences should be semantically
coherent and related to a domainword, when compared to amanually defined domain,
it will be harder to have semantic coherence across lines.