Poetry Generation with PoeTryMe - Computational Creativity Research: Towards Creative Machines - page 239

ASPERA [ 11 ], COLIBRI [ 8 ] and McGonnagall [ 24 , 25 ] fall in the last category,
as they satisfy the three properties. In the first two systems, words are combined
according to the syntax of the language and should make sense according to a prose
message provided by the user. When occurring at the end of lines, words may be addi-
tionally constrained by the strophic form (e.g. rhymes). McGonnagall's goal state is
precisely a text that satisfies all the three properties. However, after several exper-
iments, Manurung et al. [ 25 ] admit that it is difficult to produce both semantically
coherent text in a strict agreement to a predefined metre.
In the last years, a series of poetry generators have been developed. Although not
producing poems towards a precise conceptual message, most of them constrain the
selection of words according to their meaning. These, which would also fit in the last
category, include systems that produce syntactically correct text in different forms
of poetry, such as haikus [ 27 , 34 ] or song lyrics [ 2 , 28 , 29 ], in different languages,
including Tamil [ 28 , 29 ], Chinese [ 35 ], Finnish [ 33 ], or Basque [ 1 ]. To define a
starting point for semantics, most systems start generation from a given theme or set
of seed words [ 27 , 34 , 35 ]. This constrains the space of possible generations in a
way that the poem should use these exact words, or others semantically related. The
choice of relevant words may be achieved either by exploring models of semantic
similarity, extracted from corpora [ 29 , 33 - 35 ], with the help of lexical-semantic
knowledge bases [ 1 , 27 , 28 ], or both [ 7 ]. As for systems that learn a language
model [ 2 ], it is expected that model already holds the three properties.
In addition to poetic features such as rhyme and metre, the Full-FACE system
by Colton et al. [ 7 ] generates poems according to the mood for a certain day, given
by newspaper articles. Poems are produced with line templates, collected from the
articles and from short phrases in the Web. Among other constraints, the selection
of words considers the sentiment they transmit, as given by a polarity lexicon. Fur-
thermore, this system provides comments, supporting the choices made (e.g. mood,
used sentences, aesthetic measures), which contextualize the poem and are useful
for more objective evaluations of the results.
12.3 PoeTryMe
PoeTryMe is not a poetry generation system, but more like a poetry generation
platform, on the top of which different approaches for poetry generation can be
implemented. It relies on a modular architecture (see Fig. 12.1 ), which enables the
independent development of each module. This architecture provides a high level
of customisation, depending on the needs of the system and ideas of the user. It is
possible to define the semantic relation instances to be used, the sentence templates
of the generation grammar, the generation strategy and the configuration of the poem.
In this section, the modules, their inputs and interactions are presented.