Corpus-Based Generation of Content and Form in Poetry
Jukka M. Toivanen, Hannu Toivonen, Alessandro Valitutti and Oskar Gross
Department Of Computer Science and
Helsinki Institute for Information Technology, HIIT
University of Helsinki, Finland
Abstract
We employ a corpus-based approach to generate content
and form in poetry. The main idea is to use two different
corpora, on one hand, to provide semantic content for
new poems, and on the other hand, to generate a specific
grammatical and poetic structure. The approach uses
text mining methods, morphological analysis, and morphological
synthesis to produce poetry in Finnish. We
present some promising results obtained via the combination
of these methods and preliminary evaluation
results of poetry generated by the system.
Introduction
Computational poetry is a challenging research area of computer
science, at the cross section of computational linguistics
and artificial intelligence. Since poetry is one of the
most expressive ways to use verbal language, computational
generation of texts recognizable as good poems is difficult.
Unlike other types of texts, both content and form contribute
to the expressivity and the aesthetical value of a poem. The
extent to which the two aspects are interrelated in poetry is
a matter of debate (Kell 1965).
In this paper we address the issues of generating content
and form using corpus-based approaches. We present a poetry
generator in which the processing of content and form is
performed through access to two separate corpora, with minimal
manual specification of linguistic or semantic knowledge.
In order to automatically obtain world knowledge necessary
for building the content, we use text mining on a background
corpus. We construct a word association network
based on word co-occurrences in the corpus and then use
this network to control the topic and semantic coherence of
poetry when we generate it.
Many issues with the form, especially the grammar, we
solve by using a grammar corpus. Instead of using an explicit,
generative specification of the grammar, we take random
instances of actual use of language from the grammar
corpus and copy their grammatical structure to the generated
poetry. We do this by substituting most words in the example
text by ones that are related to the given topic in the word
association network.
Our current focus is on testing these corpus-based principles
and their capability to produce novel poetry of good
quality on a given topic. At this stage of research, we have
not yet considered rhyme, rhythm or other phonetic features
of the form. These will be added in the future, as will more
elaborate mechanisms of controlling the content.
As a result of the corpus-based design, the input to the
current poetry generator consists of the background and the
grammar corpora, and the topic of the poem. In the intended
use case, the topic is directly controlled by the user, but
we allow the grammar corpus to influence the content, too.
Control over form is indirectly over the choice of the two
corpora. The only directly language-dependent component
in the system is an off-the-shelf module for morphological
analysis and synthesis.
The current version of our poetry generation system
works in Finnish. Its rich morphology adds another characteristic
to the current implementation. However, we believe
that the flexible corpora-based design will be useful in
transferring the ideas to other languages, as well as in developing
applications that can adapt to new styles and contents.
A possible application could be a news service in the web,
with a poem of the day automatically generated from recent
news and possibly triggering, in the mind of the reader, new
views to the events of the world.
After briefly reviewing related work in the next section,
we will describe the corpus-based approach in more detail.
Then, we will give some examples of generated poetry, with
rough English translations. We have carried out an empirical
evaluation of the generated poetry with twenty subjects, with
encouraging results. We will describe this evaluation and its
results, and will then conclude by discussing the proposed
approach and the planned future work.
Related Work
The high complexity of creative language usage poses substantial
challenges for poetry generation. Nevertheless, several
interesting research systems have been developed for
the task (Manurung, Ritchie, and Thompson 2000; Gervas´
2001; Manurung 2003; Diaz-Agudo, Gervas, and Gonz ´ alez- ´
Calero 2002; Wong and Chun 2008; Netzer et al. 2009).
These systems vary a lot in their approaches, and many different
computational and statistical methods are often combined
in order to handle the linguistic complexity and creativity
aspects. State of the art in lexical substitution but not
in poetical context is presented, for instance, by Guerini et
International Conference on Computational Creativity 2012 175
al. (2011). We next review some representative poetry generation
systems.
ASPERA (Gervas 2001) employs a case-based reasoning ´
approach. It generates poetry out of a given input text via
a composition of poetic fragments that are retrieved from a
case-base of existing poems. In the system case-base each
poetry fragment is annotated with a prose string that expresses
the meaning of the fragment in question. This prose
string is then used as the retrieval key for each fragment.
Finally, the system combines the fragments by using additional
metrical rules. In contrast, our “case-base” is a plain
text corpus without annotations. Additionally, our method
can benefit from the interaction of two distinct corpora for
content and form.
The work of Manurung et al. (2000) draws on rich
linguistic knowledge (semantics, grammar) to generate a
metrically constrained poetry out of a given topic via a
grammar-driven formulation. This approach requires strong
formalisms for syntax, semantics, and phonetics, and there
is a strong unity between the content and form. Thus, this
system is quite different from our approach. The GRIOT
system on its part (Harrell 2005) is able to produce narrative
poetry about a given theme. It models the theory of conceptual
blending (Fauconnier and Turner 2002) from which an
algorithm based on algebraic semantics was implemented.
In particular, the approach employs “semantics based interaction”.
This system allows the user to affect the computational
narrative and produce new meanings.
The above mentioned systems have rather complex structures
involving many different interacting components. Simpler
approaches have also been used to generate poetry. In
particular, Markov chains (n-grams) have been widely used
as the basis of poetry generation systems as they provide a
clear and simple way to model some syntactic and semantic
characteristics of language (Langkilde and Knight 1998).
However, the characteristics are local in nature, and therefore
standard use of Markov chains tends to result in poor
sentence and poem structures. Furthemore, form and content
are learned from a single corpus and cannot be easily
separated.
Methods
We next present our approach to poetry generation. In the
basic scenario, a topic is given by the user, and the proposed
method then aims to give as output a novel and non-trivial
poem in grammatically good form, and with coherent content
related to the given topic.
A design principle of the method is that explicit specifications
are kept at minimum, and existing corpora are used to
reduce human effort in modeling the grammar and semantics.
Further, we try to keep language-dependency of the
methods small.
The poetry generator is based on the following principles.
1. Content: The topics and semantic coherence of generated
poetry are controlled by using a simple word association
network. The network is automatically constructed
from a so-called background corpus, a large body of text
used as a source of common-sense knowledge. More specifically,
the semantic relatedness of word pairs is extracted
from their co-occurrence frequency in the corpus. In the experiments
of this paper, the background corpus is Finnish
Wikipedia.
2. Form (grammatical): The grammar, including the syntax
and morphology of the generated poetry, is obtained in
an instance-based manner from a given grammar corpus. Instead
of explicitly representing a generative grammar of the
output language, we copy a concrete instance from an existing
sentence or poem but replace the contents. In our experiments,
the corpus consists mainly of old Finnish poetry.
3. Form (phonetic): Rhythm, rhyme, and other phonetic
features can, in principle, be controlled when substituting
words in the original text by new ones. This part has not
been implemented yet but will be considered in future work.
The current poetry generation procedure can now be outlined
as follows:
• A topic is given (or randomly chosen) for the new poem.
The topic is specified by a single word.
• Other words associated with the topic are extracted from
the background graph (see below).
• A piece of text of the desired length is selected randomly
from the grammar corpus.
• Words in the text are analyzed morphologically for their
part of speech, singular/plural, case, verb tense, clitics etc.
• Words in the text are substituted independently, one by
one, by words associated with the topic. The substitutes are
transformed to similar morphological forms with the original
words. The original word is left intact, however, if there
are no words associated with the topic that can be transformed
to the correct morphological form.
• After all words have been considered, the novelty of the
poem is measured by the percentage of replaced words. If
the poem is sufficiently novel it is output. Otherwise the
process can be re-tried with a different piece of text.
For the experiments of this paper, we require that at least
one half of the words were replaced. This seems sufficient
to make readers perceive the new topic as the semantic core
of the poem.
We next describe in some more detail the background
graph construction process as well as the morphological
tools used.
Background Graph
A background graph is a network of common-sense associations
between words. These associations are extracted
from a corpus of documents, motivated by the observation
that (frequent) co-occurrence of words tends to imply some
semantic relatedness between them (Miller 1995).
The background graph is constructed from the given background
corpus using the log-likelihood ratio test (LLR). The
log-likelihood ratio, as applied here for measuring associations
between words, is based on a multinomial model of
word co-occurrences (see, e.g., Dunning (1993) for more information).
The multinomial model for a given pair {x, y} of words
has four parameters p11, p12, p21, p22 corresponding to the
probability of their co-occurrence as in the contingency table
below.
International Conference on Computational Creativity 2012 176
x ¬x ⌃
y p11 p12 p(y; C)
¬y p21 p22 1 ! p(y; C)
⌃ p(x; C) 1 ! p(x; C) 1
Here, p(x; C) and p(y; C) are the marginal probabilities of
word x or word y occurring in a sentence in corpus C, respectively.
The test is based on the likelihoods of two such multinomial
models, a null model and an alternative model. For
both models, the parameters are obtained from relative frequencies
in corpus C. The difference is that the null model
assumes independence of words x and y (i.e., by assigning
p11 = p(x; C)p(y; C) etc.), whereas the alternative model
is the maximum likelihood model which assigns all four
parameters from their observed frequencies (i.e., in general
p11 6= p(x; C)p(y; C)).
The log-likelihood ratio test is then defined as
LLR(x, y) = !2
X
2
i=1
X
2
j=1
kij log(pnull
ij /pij ), (1)
where kij is the respective number of occurrences. It can be
seen as a measure of how much the observed joint distribution
of words x and y differs from their distribution under
the null hypothesis of independence, i.e., how strong the association
between them is. More complex models, such as
LSA, pLSA or LDA could be used just as well.
Finally, edges in the background graph are constructed
to connect any two words x, y that are associated with
LLR(x, y) greater than an empirically chosen threshold. To
find words that are likely semantically related to the given
topic, first-level neighbours (i.e., words association with the
topic word) are extracted from the background graph. If this
set is not large enough (ten words or more in the experiments
of this paper), we add randomly selected second-level
neighrbours (i.e., words associated to any of the first-level
neighbors).
In the future, we plan to use edge weights to control the
selection of substitutes, and possibly to perform more complex
graph algorithms on the background graph to identify
and choose content words.
Morphological Analysis and Synthesis
Morphological analysis is essential and non-trivial for morphologically
rich languages such as Finnish or Hungarian.
In these languages, much of the language’s syntactic and
semantic information is carried by morphemes joined to
the root words. For instance, the Finnish word “juoksentelisinkohan”
(I wonder if I would run around) is formed out
of the root word “juosta” (run). Hence, morphological analysis
provides valuable information of the syntax and to some
degree of the semantics. In our current system, morphological
analysis and synthesis are carried out using Omorfi1,
a morphological analyzer and generator of Finnish language
based on finite state automata methodology (Linden, Silfver- ´
berg, and Pirinen 2009).
1
URL: http://gna.org/projects/omorfi
With the help of Omorfi we can thus generate substitutes
that have similar morphological forms with the original
words. For instance, assume that the topic of the poetry
is “ageing” and we want so substitute “juoksentelisinkohan”
by a word based on “muistaa” (remember). Omorfi can
now generate “muistelisinkohan” (I wonder if I would think
back) as a morphologically matching word.
Examples
We next give some example poems generated by the current
system with the original example texts used to provide
structure for these poems. We also give their rough English
translations, even though we suspect that poetical aesthetics
somewhat change in translation. The substituted words are
indicated by italics.
The first example poem is generated around the topic
“(children’s) play”. We first give the Finnish poem with the
template used to construct it (on the right) and then the English
translation of both the generated and original poems.
Kuinka han leikki ¨ silloin kuinka han leikki kerran ¨
uskaliaassa, uskaliaassa // suuressa vihreass ¨ a // ¨
kuiskeessa puistossa
vaaleiden puiden alla. ihanien puiden alla.
Han oli ¨ kuullut huvikseen, Han oli katsellut huvikseen, ¨
kuinka hanen ¨ kuiskeensa kuinka hanen hymyns ¨ a¨
kanteli helkkeina tuuloseen ¨ . putosi kukkina maahan,
Original by Uuno Kailas: Satu meista kaikista, 1925 ¨
How she played then how she played once
in a daring, daring whispering in a big green park
under the pale trees. under the lovely trees.
She had heard for fun She had watched for fun
how her whispering how her smile
drifted as jingle to the wind. falled down as flowers,
The next poem is generated with “hand” as the topic. The
template used is shown below the generated poem and thereafter
the translations, respectively.
Vaaleassa kourassa
sopusuhtaisessa kourassa ovat nuput niin kalpeita
kuvassasi lepa¨a¨ lapsikulta jumala.
Alakuloisessa metsass ¨ a¨
Ham¨ ar¨ ass ¨ a mets ¨ ass ¨ a ovat kukat niin kalpeita ¨
Varjossa lepa¨a sairas jumala ¨
Original by Edith Sodergran: Mets ¨ an h ¨ am¨ ar¨ a, 1929 ¨
In a pale fist
in a well-balanced fist, the buds are so pale
in your image lies a dear child god.
In a gloomy forest
In a dim forest flowers are so pale
In the shadow lies a sick god
The final example poem has “snow” as its topic.
Elot sai karkelojen teita, Aallot kulki tuulten teit ¨ a,¨
lumi ajan kotia, aurinko ajan latua,
hiljaa soi kodit autiot, hiljaa hiihti paiv ¨ at pitk ¨ at, ¨
hiljaa sai armaat karkelot - hiljaa hiipi pitkat y ¨ ot - ¨
laiho sai lumien riemut. paiv ¨ a kutoi kuiden ty ¨ ot, ¨
Original by Eino Leino: Alkusointu, 1896
International Conference on Computational Creativity 2012 177
Lives got the frolic ways,
snow the home of time,
softly chimed abandoned homes,
softly got frolics beloved -
ripening crop got the snows’ joys.
Waves fared the wind’s ways,
sun the track of time,
slowly skied for long days,
slowly crept for long nights -
day wove the deeds of moons
Subjectively judging, the generated poems show quite a
wide range of grammatical structures, and they are grammatically
well formed. The cohesion of the contents can also be
regarded as fairly high. However, the quality of generated
poetry varies a lot. Results from an objective evaluation are
presented in the next section.
Evaluation
Evaluation of creative language use is difficult. Previous
suggestions for judging the quality of automatically generated
poetry include passing the Turing test or acceptance for
publishing in some established venue. Because the intended
audience of poetry consists of people, the most pragmatic
way of evaluating computer poetry is by an empirical validation
by human subjects. In many computer poetry studies
both human written and computationally produced poetry
have been evaluated for qualities like aesthetic appreciation
and grammaticality.
In this study we evaluated poetry using a panel of twenty
randomly selected subjects (typically university students).
Each subject independently evaluated a set of 22 poems, of
which one half were human-written poems from the grammar
corpus and the other half computer-generated ones with
at least half of the words replaced. The poems were presented
in a random order, and the subjects were not explicitly
informed that some of the poems were computer-generated.
Each subject evaluated each text (poem) separately. The
first question to answer was if the subject considered the
piece of text to be a poem or not, with a binary yes/no
answer. Then each text was evaluated qualitatively along
six dimensions: (1) How typical is the text as a poem?
(2) How understandable is it? (3) How good is the language?
(4) Does the text evoke mental images? (5) Does
the text evoke emotions? (6) How much does the subject
like the text? These dimensions were evaluated on the scale
from one (very poor) to five (very good). (The interesting
question of how the amount of substituted words affects the
subjective experience of topic, novelty and quality is left for
future research.)
Evaluation results averaged over the subjects and poems
are shown in Figures 1 and 2. Human-written poems were
considered to be poems in 90.4% of the time and computergenerated
poems 81.5% of the time (Figure 1). Intervals
containing 66.7% of the poems show that there was more
variation in the human-written poetry than in the computer
generated poetry. Overall, these are promising results, even
though statistically the difference between human-written
and computer generated poetry is significant (p-value with
Wilcoxon rank-sum test is 0.02).
Figure 1: Relative amounts of texts (computer-generated
and human-written poetry) subjectively considered to be poems,
averaged over all subjects. The whiskers indicate an
interval of 66.7% of poems around the median. Points indicate
the best and worst poems in the both groups.
Figure 2: Subjective evaluation of computer-generated and
human-written poetry along six dimensions: (1) typicality
as a poem, (2) understandability, (3) quality of language,
(4) mental images, (5) emotions, and (6) liking (see text for
details). Results are averaged over all subjects and poems.
The whiskers indicate one standard deviation above and below
the mean.
The evaluated qualities have a similar pattern (Figure
2): The average difference between human-written and
computer-generated poetry is not large, and in many cases
there is a lot of overlap in the ranges of scores, indicating
that a good amount of (best) computer-generated poems
were as good as (worst) human-written ones. Statistically,
however, the differences are highly significant (all p-values
below 0.001). The biggest drop in quality was in understandability
(dimension 2). However, somewhat controversially,
the language remained relatively good (dimension 3). An
interesting observation is that some of the generated poems
were rated to be quite untypical but their language quality
and pleasantness were judged to be relatively high.
International Conference on Computational Creativity 2012 178
Discussion
We have proposed a flexible poetry generation system which
is potentially able to produce poetry out of a wide variety
of different topics and in different styles. The flexibility
is achieved by automating the processes of acquiring
and applying world knowledge and grammatical knowledge.
We use two separate corpora: background corpus for mining
lexical associations, and grammar corpus for providing
grammatical and structural patterns for the basis of new poetry.
We have implemented the system for Finnish, a morphologically
rich language. We carried out a preliminary
evaluation on the produced poetry, with promising results.
It may be questioned whether the current approach exhibits
creative behaviour, and whether the system is able to
produce poetry that is interesting and novel with respect to
the text that is used as the basis of new poetry. First, the
generated poems are usually very different from the original
texts (our subjective view, to be evaluated objectively in the
future). Second, some of the generated texts were rated to
be quite untypical, even though recognized as poems. The
pleasantness and language quality of these poems were still
judged to be relatively high. According to these observations
we think that at least some of the system’s output can
be considered to be creative. Thus, the system could be argued
to automatically piggyback on linguistic conventions
and previously written poetry to produce novel and reasonably
high quality poems.
Our aim is to develop methods that can be applied to
other languages with minimal effort. In our current system,
morphological analysis and synthesis are clearly the most
strongly language-specific components. They are fairly well
isolated and could, in principle, be replaced by similar components
for some other language. However, it may prove
to be problematic to apply the presented approach to more
isolating languages (i.e., with a low morpheme-per-word ratio),
such as English. In agglutinative languages (with higher
morpheme-per-word ratio), such as Finnish, a wide variety
of grammatical relations are realized by the use of affixation
and the word order is usually quite free. We currently consider
implementing the system for other languages, in order
to identify and test principles that could carry over to some
other languages.
So far, we have not considered controlling rhythm, rhyme,
alliteration or other phonetic aspects. We plan to use constraint
programming methods in the lexical substitution step
for this purpose. At the same time, we doubt this will be always
sufficient in practice since the space of suitable substitutes
can be severely constrained by grammar and semantics.
Another interesting technical idea is to use n-gram language
models for computational assessment of the coherence of
produced poetry.
We consider the approach described in this paper to be
a plausible building block of more skillful poetry generation
systems. The next steps we plan to take, in addition to
considering phonetic aspects, includes trying to control the
emotions that the poetry exhibits or evokes. We are also interested
in producing computer applications of adaptive or
instant poetry.
Acknowledgements: This work has been supported by the
Algorithmic Data Analysis (Algodan) Centre of Excellence
of the Academy of Finland.
<references_biblio/>
References
Diaz-Agudo, B.; Gervas, P.; and Gonz ´ alez-Calero, P. A. ´
2002. Poetry generation in COLIBRI. In ECCBR 2002,
Advances in Case Based Reasoning, 73–102.
Dunning, T. 1993. Accurate methods for the statistics of surprise
and coincidence. Computational linguistics 19(1):61–
74.
Fauconnier, G., and Turner, M. 2002. The Way We Think:
Conceptual Blending and the Mind’s Hidden Complexities.
New York: Basic Books.
Gervas, P. 2001. An expert system for the composition of ´
formal spanish poetry. Journal of Knowledge-Based Systems
14(3–4):181–188.
Guerini, M.; Strapparava, C.; and Stock, O. 2011. Slanting
existing text with valentino. In Pu, P.; Pazzani, M. J.; Andre,´
E.; and Riecken, D., eds., Proceedings of the 2011 International
Conference on Intelligent User Interfaces, 439–440.
ACM.
Harrell, D. F. 2005. Shades of computational evocation and
meaning: The GRIOT system and improvisational poetry
generation. In In Proceedings, Sixth Digital Arts and Culture
Conference, 133–143.
Kell, R. 1965. Content and form in poetry. British Journal
of Aesthetics 5(4):382–385.
Langkilde, I., and Knight, K. 1998. The practical value of
n-grams in generation. In Proceedings of the International
Natural Language Generation Workshop, 248–255.
Linden, K.; Silfverberg, M.; and Pirinen, T. 2009. HFST ´
tools for morphology - an efficient open-source package for
construction of morphological analyzers. In Proceedings
of the Workshop on Systems and Frameworks for Computational
Morphology, 28–47.
Manurung, H. M.; Ritchie, G.; and Thompson, H. 2000.
Towards a computational model of poetry generation. In
Proceedings of AISB Symposium on Creative and Cultural
Aspects and Applications of AI and Cognitive Science, 79–
86.
Manurung, H. 2003. An evolutionary algorithm approach
to poetry generation. Ph.D. Dissertation, University of Edinburgh,
Edinburgh, United Kingdom.
Miller, G. A. 1995. Wordnet: a lexical database for english.
Communications of the ACM 38(11):39–41.
Netzer, Y.; Gabay, D.; Goldberg, Y.; and Elhadad, M. 2009.
Gaiku : Generating haiku with word associations norms.
In Proceedings of NAACL Workshop on Computational Approaches
to Linguistic Creativity, 32–39.
Wong, M. T., and Chun, A. H. W. 2008. Automatic haiku
generation using VSM. In Proceedings of ACACOS, 318–
323.
International Conference on Computational Creativity 2012 179