Poetry Generation with PoeTryMe - Computational Creativity Research: Towards Creative Machines - page 238

poems/lyrics by a seed author. Probabilistic language models are learned and then
used for generating text in the style of the seed author. In Barbieri et al.'s system, the
models include constrains for rhyme, metre, syntax and semantics.
In addition to the previous, alternative techniques have been proposed for this task
including constraint programming [ 33 ], or a blackboard multi-agent system [ 26 ].
The truth is that, as more systems are developed for the automatic generation of
poetry, the line between these categories is becoming thinner and there are, of course,
overlapping approaches. The combination of different approaches, some focused on
the form and others on the content, is related with the degree of articulation [ 12 ]of
a poetry generation system.

12.2.2 Properties of Generated Poetry

In his dissertation on the automatic generation of poetry, Manurung [ 24 ] affirms that
poetic text must hold all the three properties of meaningfulness (M), grammatical-
ity (G) and poeticness (P). Of course, this does not cover all kinds of poetry, as
it leaves out some alternative types, including concrete poetry, but these properties
were a starting point for much work on this area. Manurung adds that poetry must:
(i) convey a conceptual message, meaningful under some interpretation; (ii) obey
linguistic conventions prescribed by a given grammar and lexicon; and (iii) exhibit
poetic features. Having this in mind, he divides poetry generation systems according
to the properties held by the text they generate. To accommodate additional poetry
generation systems, we refined and expanded Manurung's categories with an addi-
tional group, thus resulting in the following categories:

• Word salad : random words are concatenated together, without following gram-

matical rules, therefore not holding any of the properties.

• Features-aware (P): generated text exhibits poetic features (e.g. metre, rhyme),

but there are no syntactical nor semantic constraints.

• Features+Syntax-aware (P+G): besides exhibiting poetic features, the choice of

words follows a (pre-defined) grammatically correct form.

• Poetry generation (P+G+M): generated text exhibits poetic features, obeys syn-

tactic rules and the choice of words follows some semantic constraints.

The random words strategy of Tra-la-Lyrics [ 16 , 17 ] falls into the features-aware
category. In this strategy, the selection of words is only constrained by the target
metre and, when in the end of a line, by their termination, which should rhyme with
previously used terminations. The choice of words does not follow any grammatical
rules nor semantic constraints, so the generated text does not convey a precise mean-
ing. The WASP system [ 10 ], POEVOLVE [ 22 ], and the generative grammar strategy
of Tra-la-Lyrics [ 16 , 17 ] are features+syntax-aware systems, because they generate
text that is grammatically correct, but do not have any kind of semantic constraints
nor concerns.