Poetry Generation with PoeTryMe - Computational Creativity Research: Towards Creative Machines - page 256

the idea that creativity lies “in the eyes of the beholder” [ 6 ]. PoeTryMe follows this
path in two ways: first, its modular architecture makes it adaptable to a wide range
of disparate situations, including the exploration of different generation strategies
(possibly considering different features), adaptation to an evolvingworld by learning,
exploitation of different lexical and semantic resources, and possible adaptation to
other languages; second, it explains the process behind each new poem, which is
particularly relevant for ruling out the idea of “randomness”.
According to the techniques used, our baseline implementation falls into simple
template-based generation, while the others, although using templates, are closer to
evolutionary and generate-and-test approaches. As for the generated poems, they are
grammatically correct and, when the generation strategy considers the metre and
the rhyme, they exhibit poetic features. Also, despite using templates, the richness
of the lexical-semantic resources and of the poetry collection exploited provide an
interesting degree of variation in the generated lines. Semantics is coherent in each
line and the poem uses words of a pre-defined semantic domain. Although minor
issues might arise regarding the connection of two contiguous lines, a meaning tends
to emerge from the full poem, possibly assisted by the provided contextualization.
Given their semantic concerns, we can say that all our strategies, but the baseline,
fall into Manurung's poetry generation category.
Although our approach targeted Portuguese, it should be emphasised that
PoeTryMe's architecture is flexible enough for generating poetry in other languages,
as long as we have the following language-specific components: (i) a semantic graph,
which could be obtained, for instance, from a wordnet-like resource; (ii) line tem-
plates, which could be learned from any corpus of poetry in the target language;
(iii) a syllable division tool; and (iv) a lexicon with morphological information. In
fact, PoeTryMe has recently been adapted to Spanish [ 18 ] and, in the future, other
languages might be targeted. We have also been working on the generation of poetry
with a predefined sentiment orientation (e.g. positive or negative), using available
polarity lexicons [ 15 ]. Other future directions might involve adding other kinds of
relations to the semantic network (e.g. word associations, as in [ 33 ]) and embracing
current trends in poetry generation, such as generating poetry based on a given piece
of text (e.g. news, as in [ 7 , 32 ]).
References
1. Agirrezabal, M., Arrieta, B., Astigarraga, A., Hulden, M.: Pos-tag based poetry generation with
wordnet. In: Proceedings of the 14th European Workshop on Natural Language Generation,
pp. 162-166. ACL Press, Sofia, Bulgaria, August 2013
2. Barbieri, G., Pachet, F., Roy, P., Esposti, M.D.: Markov constraints for generating lyrics with
style. In: Proceedings of 20th European Conference on Artificial Intelligence (ECAI), Frontiers
in Artificial Intelligence and Applications, vol. 242, pp. 115-120. IOS Press (2012)
3. Belz, A.: Automatic generation of weather forecast texts using comprehensive probabilistic
generation-space models. Nat. Lang. Eng. 14 (4), 431-455 (2008)