Conference PaperPDF Available

Let's FACE it. Finnish Poetry Generation with Aesthetics and Framing


Abstract and Figures

We present a creative poem generator for the morphologically rich Finnish language. Our method falls into the master-apprentice paradigm, where a computationally creative genetic algorithm teaches a BRNN model to generate poetry. We model several parts of poetic aesthetics in the fitness function of the genetic algorithm, such as sonic features, semantic coherence, imagery and metaphor. Furthermore , we justify the creativity of our method based on the FACE theory on computational creativity and take additional care in evaluating our system by automatic metrics for concepts together with human evaluation for aesthetics , framing and expressions.
Content may be subject to copyright.
Proceedings of The 12th International Conference on Natural Language Generation, pages 290–300,
Tokyo, Japan, 28 Oct - 1 Nov, 2019. c
2019 Association for Computational Linguistics
Let’s FACE it. Finnish Poetry Generation with Aesthetics and Framing
Mika H¨
Department of Digital Humanities
University of Helsinki
Khalid Alnajjar
Department of Computer Science (HIIT)
University of Helsinki
We present a creative poem generator for
the morphologically rich Finnish language.
Our method falls into the master-apprentice
paradigm, where a computationally creative
genetic algorithm teaches a BRNN model to
generate poetry. We model several parts of po-
etic aesthetics in the fitness function of the ge-
netic algorithm, such as sonic features, seman-
tic coherence, imagery and metaphor. Further-
more, we justify the creativity of our method
based on the FACE theory on computational
creativity and take additional care in evaluat-
ing our system by automatic metrics for con-
cepts together with human evaluation for aes-
thetics, framing and expressions.
1 Introduction
This paper explores the topic of computational
creativity in the case of poem generation in
Finnish. Our work does not only aim to generate,
but rather create poems automatically. We take
the FACE model (Colton et al.,2011) for com-
putational creativity as our definition of creativity.
Through this model, we motivate and evaluate cre-
ativity exhibited by our system.
Methodologically, our work embraces the
master-apprentice method (Alnajjar and
ainen,2018) used in the past for com-
putationally creative tasks. This means using a
creative genetic algorithm as a master to teach an
apprentice which is sequence-to-sequence neural
network model. This way the overall system
can approximate creative autonomy (Jennings,
2010) if the apprentice was to be exposed to
data originating from another source than the
master. For further discussion on the topic of
autonomy, see the original work establishing the
master-apprentice method.
We pay special attention to evaluation of our
system, and we motivate it through the FACE
model. A creative system should be evaluated in
terms of what has actually been modelled rather
than on an ad-hoc and unjustified fashion. Addi-
tionally, our contribution lies on the fact that the
aesthetics of our system are motivated by exist-
ing non-computational literature in poetry analy-
sis. Furthermore, our system is capable of adjust-
ing its aesthetics based on existing poetry.
Our work sheds some more light into the na-
ture of a master-apprentice system. Especially by
seeking to answer the question of multiple mas-
ters raised in the original work on the topic (Alna-
jjar and H¨
ainen,2018), which the authors left
2 Related Work
While poetry generation has been tackled a num-
ber of times before by multiple authors (Gerv´
2001;Toivanen et al.,2012;Misztal and In-
durkhya,2014;Oliveira et al.,2017), and an ex-
cellent overview is provided by Oliveira (2017)
on the recent state of the research, we dedicate
this section in describing the most recent work
conducted in the field after the aforementioned
overview paper.
TwitSong (Lamb and Brown,2019) mines a cor-
pus for verses to be used in poetry based on how
well they rhyme together. They score the verses
in poems by four metrics (meter, emotion, topical-
ity and imagery) and use a genetic algorithm to
edit the worst scoring verse in the poem. How-
ever, they only assess poems on a verse level and
their algorithm lacks poem level metrics (i.e. each
verse is considered individually and not as a part
of a whole). They base their evaluation on com-
paring generated poetry of different groups based
on how the genetic algorithm was used. They use
very broad questions such as which poem is more
creative or which poem has better imagery. This
is potentially problematic as broad questions open
more room for subjective interpretation.
Last year, a myriad of work on generation of
Chinese poetry with machine learning methods
was conducted. Research ranging from mutual
reinforcement learning (Yi et al.,2018) and con-
ditional variational autoencoders (Li et al.,2018)
to sequence-to-sequence Bi-LSTMs (Yang et al.,
2018) was presented. However, none of these
methods has been motivated from the point of
view of computational creativity, but rather serve
for a purely generative purpose.
The work conducted by Colton et al. (2012), al-
though not recent, deserves special attention, as
they had used the same FACE model as a basis in
their poem generation. They take a template based
approach to generating poems from current news
articles. Unfortunately they do not provide an
evaluation of the generated poetry, which makes
meaningful comparison difficult.
The work presented by us in this paper has to
deal with the rich morphosyntax of Finnish, which
is an NLG problem far from solved. H¨
(2018) presents a solution for this problem in their
Finnish poem generator. However, their generator
relies on predefined rule-based structures, whereas
our aim is to have a system with more structural
versatility, and yet the capability of coping with
the morphosyntax.
3 Creativity
In order to separate our system from generative
non-creative systems, we have to provide some
justification as to why our system would exhibit
creativity in the first place. For this reason, we fol-
low the SPECS approach (Jordanous,2012) that
has been designed to evaluate creativity in a rea-
soned fashion. The approach requires creativity to
be defined first on an abstract level, and then, fol-
lowing the abstract definition, creativity should be
defined in the context of the creative task that is
to be solved. After establishing these definitions,
creativity of the system should then be evaluated
based on the definitions.
3.1 Creativity in General
For an abstract level definition of creativity, we use
FACE (Colton et al.,2011). The theory divides
creative act into two categories, one is the ground
level generative act of producing an artefact and
the other is on the level of the process. Both of
these categories are represented in the four aspects
of creativity: framing, aesthetics, concept and ex-
Framing consists of outputting a framing for a
creative artefact, and the process that generates
this output. The framing should be an explanation
in natural language, for instance, putting the cre-
ated artefact into a historical and cultural context
or describing the processes of creating the output
artefact. In other words, framing can be used as an
additional persuasive or explanatory message that
is delivered to the human perceiving the artefact
produced by a computational system.
Aesthetics consist of a function measuring the
aesthetic quality of the output and/or the program
producing it. On the process level, FACE takes
into account how the aesthetic measures came to
be in the system. The system should be able to as-
sess its own work and rate its creations. This aes-
thetic measure can also be used to computationally
assess artefacts produced by other systems or hu-
Concept is used to refer to the program that gen-
erates creative artefacts on the ground level. And
on the process level it refers to how such a program
was generated. Finally, the ground level expres-
sion is the creative output, or artefact, generated
by the system, whereas the process level of ex-
pression describes the method for generating out-
put for a given input.
3.2 Creativity for Our Poem Generator
As framing can exists in many different forms ac-
cording to the original FACE model, we follow a
more narrowed down notion of framing, which is
the intention of the computer in creating artefacts
(Charnley et al.,2012). In other words, the com-
puter should be able to output a justification ex-
plaining what certain aspects of the poem mean.
The importance of framing has recently been high-
lighted in the literature (Cook et al.,2019).
Framing does not have to be a creative act on
its own. In our case, the process of coming up
with a framing is a template based approach that
conveys the intent of the creative program in pro-
ducing the output poem. This intent, on the other
hand is captured by the aesthetic function of the
creative system. Therefore, the framing produced
should explain the poem in terms of the aesthetic
Poetry as a genre showcases a wide diversity in
terms of aesthetics; ranging from epic poetry fol-
lowing a strict meter to modern free form poetry.
Even to a degree that the poetic genre has become
fragmented ever since the 20th century (Juntunen,
2012). This diversity is not just limited to the level
of structure, but is also reflected in meaning - some
forms of poetry are meant to be read and inter-
preted literally, where as others rely on indirect
communication such as symbolism and metaphors
(see Kantokorpi et al.,1990). In our work, we are
not aiming to model the poetic genre as a whole,
but rather define a set of aesthetic functions that
capture different aspects in poetry ranging from
the structural to the meaning.
In terms of structure, our system should be able
to assess rhyming in its various forms (alliteration,
assonance, consonance and full rhymes) and the
meter of the poetry as defined by poetic foot and
syllable count.
For meaning, our system should be able appre-
ciate the presence of metaphors, the semantic co-
herence of the words forming the poem and in es-
pecial the presence of words forming different se-
mantic fields and the semantic difference of these
fields as an indicator of tension built by the choice
of words in a poem (cf. Lotman,1974).
Certain poems paint a mental image in the mind
of the reader; this qualia1provoking aspect of po-
etry is called imagery. As it is extremely difficult
for a computer to assess such rich mental sensory
phenomena provoked by poetry in humans, we
have to reduce the aesthetics related to imagery to
a more computationally manageable level, namely
that of sentiment. Sentiments expressed in a poem
can be indicators of the potential mood evoked by
the sensory imagery in the poem. Another indica-
tor of imagery is the use of concrete expressions
(see Burroway,2007).
Although the list of aesthetic measures is pre-
defined, from the point of view of the process, our
system should be able learn to adjust its aesthetic
measures based on existing poetry. Furthermore,
we aim towards a system that can learn aesthetics
of its own on its own level of abstraction, hence
the use of apprentice.
In our case, the system consists of two concepts.
One of them is a genetic algorithm (master) that
has been defined by us, the programmers. The role
of the master is to produce expressions through a
search informed by the aesthetic functions. These
1For more on the problem of qualia, see (Chalmers,1995)
expression are used to train the second concept,
which is a sequence-to-sequence BRNN model
(apprentice). This way, the overall system is given
the capability of producing new concepts of its
The expressions output by the system are com-
putationally created Finnish poems. Ultimately,
we evaluate the expressions produced by the ap-
prentice with real humans and by the master’s aes-
thetic measures.
3.3 Data
We use the 6,189 Finnish poems that are avail-
able on Wikisources2as our poem corpus. We
use the Finnish dependency parser (Haverinen
et al.,2014) to parse the poems for morphologi-
cal features, syntactic relations, part of speech and
lemma for each word. The parsing is done on a
verse-level. We split each poem into stanzas as
divided in Wikisources. From now on we refer
to a stanza of an existing poem simply as a poem.
The reason for this is to have shorter poems to deal
with in the generation step. This is especially im-
portant for the human evaluation as shorter poems
can be evaluated more accurately, as longer po-
ems have more room for unintentional character-
istics that can be interpreted too positively by hu-
man judges, such as a perceivably deeper meaning
that is due to the mere fact of having more context
to read more into. After splitting the poems into
stanzas, we have a total of 34,988 poems.
We use the word embeddings3that have been
trained on the Finnish Internet Parsebank (Kan-
erva et al.,2014). We prefer this model for two
reasons: first it has been trained on a 1.5 billion to-
ken corpus that is big on the Finnish scale and sec-
ond it has been trained on lemmas, which is an im-
portant factor for a highly agglutinating language
such as Finnish. In order to generate grammatical
Finnish, the words need to be inflected. This step
is easier if the replacement words are already in a
lemmatized from.
4 Generating Poetry
The master-apprentice approach outlined in Alna-
jjar and H¨
ainen (2018) consists of a creative
master, which is a genetic algorithm, and an ap-
prentice, which is a sequence-to-sequence model.
In this part of the paper, we describe how the aes-
thetics are implemented in the master and how it is
used to generate poems for the apprentice to learn
In this paper, we experiment with two differ-
ent masters, which will learn the weights for their
aesthetic functions form poems of different eras.
We use these masters to train one apprentice for
each of them. In addition, we train one apprentice,
which will learn from both of the masters.
4.1 Master
The master is a genetic algorithm following the
implementation presented in Alnajjar et al. (2018).
In practice, the algorithm takes in a random poem
from the poem corpus and uses it to produce an ini-
tial population of 100 individuals. These individu-
als produce an offspring of another 100 individuals
that go through mutation and crossover, and at the
end of each generation the individuals are scored
according to the aesthetic functions defined later
in this section. The 100 fittest individuals are se-
lected with NSGA-II algorithm (Deb et al.,2002)
to survive to the next generation. This process is
done for 50 generations.
All individuals in the initial population are
based on a randomly selected poem and a ran-
domly picked theme word. The theme is expanded
into the 30 most semantically similar words to
the theme word using word2vec (Mikolov et al.,
2013). Each poem in the initial population is as-
signed a random theme out of the 30 semantically
similar words to the theme. Additionally, we mod-
ify each poem in the initial population once by
using the mutation function. This is applied to
have more variety of poems in the initial popula-
tion given that all of them are based on the same
original poem from the corpus.
In mutation, a random content word is picked
in the poem and it is replaced by a word related
to the input theme (assigned to the poem) or by a
word that is similar to the original one, while en-
suring that the new replacement matches the origi-
nal in terms of its part-of-speech. To obtain words
that are related to the input theme, we build a
semantic relatedness model following Xiao et al.
(2016) using the flat 5-gram data provided by Kan-
erva et al. (2014) as the corpus. Regarding the
semantic similarity to the original word, we uti-
lize the word2vec word embeddings model. The
space of candidate replacements consists of the top
1,000 and 300 (empirically chosen) semantically
related and similar words, respectively. Out of
these candidates, only words that match the part-
of-speech of the original word, based on Uralic-
ainen,2019), are considered in the
random selection.
In terms of the crossover, we employ a single-
point crossover on a verse-level where one point in
both individuals is selected at random and verses
to the right of that point are swapped.
As mutations and crossovers are bound to break
the morphosyntax of Finnish, the new words are
always inflected to match the original morphol-
ogy with UralicNLP and Omorfi (Pirinen et al.,
2017). This will account for morphological agree-
ment, but not for case government. In case gov-
ernment, the case of the complements of the verb
depends on the verb itself. For this reason, we
inflect words with an object relation with Syntax
Maker (H¨
ainen and Rueter,2018) to produce
a grammatical surface form even if the predicate
verb is changed.
4.1.1 Aesthetics
To assess the sonic structure of poetry the follow-
ing rule-based aesthetic functions are defined on
an inter-verse level: full rhyme, assonance and
consonance. These count the number of rhyming
words between verses of the poem. Alliteration
is a metric calculated within a verse, as this type
of rhyming occurs typically inside of a verse in
Finnish poetry. As Finnish spelling is almost one
to one mapping with phonology, we can do this on
a character level without the need to approximate
the pronunciation.
Meter is captured by two aesthetic functions:
the number of syllables and the distribution of long
and short syllables within a verse. These two func-
tions are again solved by simple rules. The master
rates higher the meter it has learned from its train-
ing corpus.
A previous attempt to capture imagery in the lit-
erature is by comparing the number of abstract and
non-abstract words with the hypothesis that non-
abstract words provoke more mental imagery (Kao
and Jurafsky,2012). However, this notion can be
used only as a proxy to the quantity of imagery
in poetry, but it tells nothing about the nature of
the provoked imagery. For this reason, we have
also decided to use sentiment as an indicator of the
mood of the mental image painted by the poem.
For abstractness of words we use an existing
dataset for English that maps 40,000 common En-
glish words to an average concreteness score as
annotated by humans on a 5-point Likert scale
(Brysbaert et al.,2014). We translate this data
in Finnish with a Wiktionary based online dictio-
nary4in such a way that we consider the three top-
most translations that are verbs, nouns or adjec-
tives for each English word. To deal with poly-
semy, if multiple English words translate into one
Finnish word, we take the average of the concrete-
ness values of the English words for the Finnish
word. If the concreteness value is greater or equal
to 3, the word is considered concrete. The aes-
thetic function gives a ratio of concrete words over
concrete and abstract words in the poem.
For sentiment, due to the lack of resources for
Finnish, we use a recent state of the art method
(Feng and Wan,2019) that can learn sentiment
prediction for English with annotated data and
use the model for other languages by bilingual
word embeddings. We train the model with senti-
ment annotated data for English from the OpeNER
project (Agerri et al.,2013). We use their method
to map the pretrained Finnish and English fasttext
models from Grave et al. (2018) into a common
space. This aesthetic measure will score senti-
ments on verse level and output their variance on
the poem level.
Dividing words into semantic fields can be used
as an auxiliary tool in poem analysis in literature
studies as it can reveal tensions inside of a poem
(c.f Lotman,1974). By following this notion, we
cluster the open class part of speech words based
on their cosine similarity within a poem. For this
clustering, we use affinity propagation (Frey and
Dueck,2007), which takes a similarity matrix as
input and clusters the words based on the matrix.
The number of clusters is not fixed and affinity
propagation is free to divide the words in as many
clusters as necessary.
The clustering aesthetic function looks at the
number of clusters in a poem and the average se-
mantic distances of the clusters. The distance be-
tween two clusters is calculated by counting a cen-
troid for each cluster based on the word vectors of
a cluster and then calculating the cosine distance
of the centroids of the clusters. The values output
by the aesthetic function will set standards to how
semantically cohesive the words have to be with
each other, and how distant can their meanings be.
Although words in different clusters might be
distant semantically, they can be related prag-
matically. Therefore, we want to reveal possible
metaphorical interpretations of a given word in the
poem. We represent each semantic cluster found
in a poem by a single word. In doing so, we com-
pute the centroid vector of words in each cluster
and use the nearest word in the model’s vocabulary
to the centroid as the topic of the cluster. There-
after, we iterate over all the possible combinations
of having a certain topic as a tenor and another
as a vehicle and measure the metaphoricity of the
poem with respect to them. We measure that us-
ing the two metaphoriticy measurements defined
by Alnajjar et al. (2018), one for measuring how a
word in the poem relates to both concepts and the
other for measuring how related a word is to the
vehicle but not to the tenor5. The metaphoricity
value is then represented by the mean of the two
measurements in case both had a positive value,
otherwise zero is returned. Using the metaphoric-
ity value assigned to each tenor-vehicle combina-
tion, we define two metaphoricity aesthetics 1) the
maximum metaphoricity value and 2) the number
of metaphorical clusters (i.e. combinations where
the metaphoricity value is above zero).
As having many objectives is difficult in prac-
tice to handle for the NSGA-II algorithm (see
Tanigaki et al.,2014), we group the aesthetic func-
tions into four fitness functions. Sonic (rhyme, al-
literation, consonance, assonance, foot and sylla-
ble count), semantic (number of clusters and aver-
age and maximum distance between the clusters),
imagerial (concrete word ratio and variance of
sentiment) and metaphorical (the maximum score
for metaphoricity and the number of metaphorical
words) functions represent the four fitness func-
tions used by the genetic algorithm. These fitness
functions sum up the individual aesthetic functions
when they are used to score a poem.
4.2 Learning the Aesthetics
We divide our corpus into centuries: the 19th and
20th century. We have two masters learn their
aesthetics from either century making them spe-
cialized in that century in particular. We first
learn weights for the individual aesthetic functions
within the higher-level fitness function they belong
to. We do this by training four random forest clas-
sifiers (Breiman,2001), one for each of the four
5See (Richards,1936) for more on tenor and vehicle
higher level fitness functions. The classifiers get
the features produced by the aesthetic functions
belonging to the fitness function in question. The
classifiers are trained with the entire corpus to pre-
dict true for the desired century and false for other
The trained classifiers are only used for their
weights for each individual feature. These weights
are used in the genetic algorithm to multiply the
output of each aesthetic function adjusting their
importance for the century.
As the weights tell only little about the possible
values the aesthetic functions can or should have
within one century, we calculate a range of ac-
cepted values for each aesthetic function within a
century. The 25th percentile of the values is set as
the minimum boundary of an accepted value and
the 75th percentile as the maximum boundary. If
the value output by the aesthetic function is out-
side of this range, the output value is set to 0.
4.2.1 Master’s liking
For the evaluation purposes of the apprentices, it is
important to set standards to what is good poetry
according to the master. The master likes a poem
generated by the apprentice if the poem gets a pos-
itive value in each one of the four fitness functions.
If any of the values is 0, the master is considered
not liking the poems.
4.3 Apprentice
Apprentice is a sequence-to-sequence model that
learns to produce creatively altered verses out of
verses in existing poetry. To achieve this, we use a
BRNN model with a copy attention mechanism by
using OpenNMT (Klein et al.,2018). We use the
default settings which are two layers for encoding
and decoding and general global attention (Luong
et al.,2015).
One apprentice is trained from the output of
each master, and an additional one from the out-
put of both of the masters. We train the appren-
tices for 90000 steps to produce poems one verse
at a time, from the original poem to the master
generated ones. The master for the 19th century
produced 11903 poems and the 20th century one
11900 poems out of randomly picked initial po-
ems from the entire corpus. These constitute the
training data for the apprentices. The random seed
used in training is the same for all apprentices to
make intercomparison possible.
5 Results and Evaluation
Evaluation is one of the most important and diffi-
cult parts of computational creativity, however it is
oftentimes overlooked and conducted in an ad-hoc
manner with little to do with the actual problem
being modelled (Lamb et al.,2018). In practice
this means that a great deal of work is evaluated
based on questions and metrics that have not been
justified. This practice together with the issue ex-
pressed by Veale (2016) that people are ready to
read more into the output if it has a suitable lin-
guistic structure regardless of the actual underly-
ing creative intent of the system, are things that
should not go unnoticed when evaluating a com-
putationally creative system.
Mutta hy¨
aykset, jotka kest¨
a sain,
muistot, jotka rakkauden est¨
esiin ilmentyy vihaa kasvattain.
But the attacks I was to endure,
the memories that prevent love,
emerge amplifying the ire
Above is an example poem output by the master
in Finnish followed by its translation in English.
The example is of a typical length of a poem pro-
duced by the system as the human authored poems
were split into stanzas.
5.1 Concepts
The master as a concept is fixed and can only ad-
just its appreciation, but the apprentice is an en-
tirely new concept that is created from the out-
put of the master. In this section we evaluate the
apprentices by evaluating their output by masters’
liking. This is done only in an automatic fashion
by having all 3 of the apprentices create output for
100 randomly picked poems from the poem cor-
master 1800 master 1900
apprentice 1800 28% 33%
apprentice 1900 36% 39%
apprentice both 47% 51%
Table 1: The percentage of the output of the two mas-
ters liked
Table 1shows how many of the poems produced
by the different apprentices the masters liked. It
is clear from the results that the apprentices did
not do too well in terms of learning the century
specific aesthetics. Nevertheless, having both of
the centuries in the training boosted the results in
terms of the two masters liking the poems. This
is probably due to the fact of having more training
data available.
5.2 Framing and Aesthetics
In order to make it less likely that people read
more into the poems than what is there, we eval-
uate the poems with people based on the framing
produced by the system. The main purpose of this
evaluation is not to evaluate how good the output
poems are, but how often the aesthetic functions
agree with human judgment. The framing consists
of templates that the system fills based on its aes-
thetic functions. People are asked whether they
agree or disagree with the statements expressed in
the framing. In addition, people have the possibil-
ity of stating that they don’t know whether to agree
or disagree.
For the evaluation, we sampled 30 poems at ran-
dom from the poetry generated by the two mas-
ters. We printed each poem 5 times, and we di-
vided each set of 30 unique poems into 3 piles of
10 poems with their framing. Each pile was shuf-
fled so that no pile contained exactly the same po-
ems and no pile had the same order for the poems.
The shuffling was done to decrease any potential
bias introduced by the order of presentation of the
Initially, we recruited 15 people, each one to go
through one pile of 10 poems. However, 5 people
found the task too time consuming and stopped af-
ter evaluating a few poems. The unevaluated po-
ems from these piles were assigned to completely
new reviewers. In the end, each unique poem was
evaluated 5 times by different people and no indi-
vidual evaluator evaluated more than 10 poems.
A framing was generated for each poem. The
framing followed always the same structure. The
first statements relating to rhyming were presented
as questions whereas the rest of them were state-
ments. The statements were formed in the follow-
ing way (translated from Finnish):
1. Do the words written in italics have rhymes
(e.g. heikko peikko)?
2. Do the words written in italics have asso-
nance (e.g. talosano)?
3. Do the words written in italics have conso-
nance (e.g. sakkosokka)?
4. Does the poem have alliteration within a
verse (e.g. vanha vesi)?
5. Verse number X and Y have the same meter
6. The poem has X semantic fields: [semantic
cluster 1]... and [semantic cluster N]
7. The semantic fields [semantic cluster X] and
[semantic cluster Y] are the closest to each
8. The semantic fields [semantic cluster A] and
[semantic cluster B] are the furthest away
from each other
9. The following words in the poem [concrete
words] are concrete concepts
10. The verse number X is positive
11. The verse number Y is negative
12. The following words in the poem [metaphor-
ical words] can be understood metaphorically
13. The word X has a metaphorical connection to
word Y
For the questions on rhyming, the system high-
lights in italics all the words that have one of the
rhyming types. For the meter statement and neg-
ative and positive verse statements, random num-
bers are picked within the range of the length of
the poem. For these questions, people agreeing
does not produce the highest score, but rather if
people’s prediction is in line with the prediction
of the aesthetic function. Also, if the poem didn’t
have any metaphorical words, random words were
picked for the last two questions. Again, if peo-
ple disagreed when random words were presented
and agreed when actual metaphorical words were
presented, the accuracy of the system based on the
evaluation would go higher.
Figure 1: Evaluation results for aesthetics and framing
The accuracy reported in Figure 1shows how
often the prediction (agree/disagree) of the aes-
thetic functions matches that of the majority of
the people out of all the times a majority deci-
sion could be reached per poem. The tie shown in
the figure shows the percentage of time the state-
ment received an equal number of agreeing and
disagreeing opinions from people per poem. The
data show by I don’t know represents the number
of times people stated they did not know over all
the answers for the statement. Note that this is not
calculated per poem but per statement.
The statements related to semantics were the
most difficult ones for people to evaluate with
around 80% of the time people saying they did not
know whether to agree or not. Another difficult
statement to judge was the last metaphorical state-
ment including an interpretation for two words be-
ing metaphorically connected. This question also
included the highest number of ties in people’s
Interestingly, the accuracy was high only for
the traditional rhyme types, but lower on the as-
sonance and consonance. Even though our rules
can easily and objectively measure the existence
of these rhyming types, it is interesting to see that
people’s judgment deviates from the values out-
put by the aesthetic functions. Especially reveal-
ing is the low accuracy on consonance. Our sys-
tem sees consonance whenever two words have the
same consonants in the same positions such as in
jo (already) and ja (and) or en (I don’t) and on (is).
Even though these words do exhibit consonance,
it seems that people do not find such consonance
perceivable. This being said, the mere existence of
rhyming is not enough, but it should also be per-
ceivable. Just what this perceivability entails is an
interesting question left for future research.
For semantics it is difficult to draw any mean-
ingful conclusions as more often than not, peo-
ple simply did not know whether to agree or not.
However, the results do seem promising for the
correctness semantic clusters (60 % of the time)
and the furthest clusters (72% of the time). At any
rate, semantics calls for further qualitative analy-
sis in the future as it seems to be a difficult thing
to assess for people.
In the case of imagery, it seems that people
agreed on the concreteness 39% of the time, al-
though the score might seem low, it is to remem-
ber that all the concrete words were presented as a
list in the framing. If even one of the words was
not perceived as concrete, people were likely to
disagree. Sentiment, on the other hand, resulted in
mixed accuracies; the accuracy for positive sen-
timent was 69% whereas for negative sentiment
the accuracy was 28%. As the sentiment analysis
was based on an existing state-of-the-art method,
this result is surprising. However, it is very likely
the case that negativity in poetry is expressed in
a very different way than in other text types. In
other words, there is a need for a sentiment anno-
tated corpus consisting of poetry and other liter-
ary texts for better predicting the sentiment in po-
ems. All in all, the prediction for concrete words
could also benefit from a dataset authored specifi-
cally for Finnish.
The accuracy for the metaphorical words was
high, 73%. However, the interpretation provided
for one of the metaphorical words gave inconclu-
sive results, as people either did not know or had
very mixed judgments. This part as well calls for
qualitative analysis in the future.
5.3 Expressions
Finally we evaluate the expressions of the master
and the apprentice in relation to each other. For
this evaluation we treat both of the masters as one,
and we evaluate the best apprentice according to
the masters’ liking. We sample randomly 10 po-
ems from the corpus for which both the master
and the apprentice had produced altered poems.
We evaluate these poems by asking people which
one of the generated poems from the same orig-
inal one they prefer, that of the master or that of
the apprentice. We present the two poems on the
same page, shuffling their order for each printout.
We also shuffle the order of the poems. We ask 10
people to rate the 20 poems, 10 master generated
and 10 apprentice generated ones.
Figure 2: People’s preference for each poem
Figure 2shows the preference of the people per
poem. The poetry generated by the apprentice was
most often preferred by the judges. The master
generated poetry did not reach to a majority in
preference for any poem. The interesting question
of what happens in the poems that result in a tie
in people’s preference calls for a future qualitative
study to understand better the phenomenon of the
6 Conclusions
We have shown our novel method for generating
poetry in Finnish. With the help of the FACE
model, we were able to conduct evaluation on the
aesthetics and framing that was revealing of the
shortcomings of our system. Framing made it pos-
sible to assess the core functionality better by min-
imizing the room for people reading more into
the poem than what was there. Having the op-
tion for people to say that they do not know rather
than forcing them to either agree or disagree re-
vealed the difficulty of assessing semantics and
metaphors even for people. We propose for the fu-
ture to conduct evaluation on such high level fea-
tures of language on a qualitative fashion to better
understand how people perceive these in generated
As a vast majority of the NLP research focuses
on English, we had to deal with the practical is-
sue of the scarce annotated resources for Finnish
to capture the high level features such as con-
creteness, sentiment and metaphor. As a result
we ended up developing useful resources for the
aesthetic functions which we have made publicly
available on Github6.
Rodrigo Agerri, Montse Cuadros, Sean Gaines, and
German Rigau. 2013. Opener: Open polarity en-
hanced named entity recognition. Procesamiento
del Lenguaje Natural, (51).
Khalid Alnajjar, Hadaytullah Hadaytullah, and Hannu
Toivonen. 2018. “Talent, Skill and Support. A
method for automatic creation of slogans. In Pro-
ceedings of the 9th International Conference on
Computational Creativity (ICCC 2018), pages 88–
95, Salamanca, Spain. Association for Computa-
tional Creativity.
Khalid Alnajjar and Mika H¨
ainen. 2018. A
master-apprentice approach to automatic creation of
culturally satirical movie titles. In Proceedings of
the 11th International Conference on Natural Lan-
guage Generation, pages 274–283.
Leo Breiman. 2001. Random forests. Machine learn-
ing, 45(1):5–32.
Marc Brysbaert, Amy Beth Warriner, and Victor Ku-
perman. 2014. Concreteness ratings for 40 thousand
generally known english word lemmas. Behavior
research methods, 46(3):904–911.
Janet Burroway. 2007. Imaginative Writing: The Ele-
ments of Craft. Pearson.
David J Chalmers. 1995. Absent qualia, fading qualia,
dancing qualia. Conscious experience, pages 309–
John William Charnley, Alison Pease, and Simon
Colton. 2012. On the notion of framing in computa-
tional creativity. In ICCC, pages 77–81.
Simon Colton, John William Charnley, and Alison
Pease. 2011. Computational creativity theory: The
FACE and IDEA descriptive models. In ICCC,
pages 90–95.
Simon Colton, Jacob Goodwin, and Tony Veale. 2012.
Full-FACE poetry generation. In ICCC, pages 95–
Michael Cook, Simon Colton, Alison Pease, and
Maria Theresa Llano. 2019. Framing in computa-
tional creativity a survey and taxonomy. In The
proceedings of the tenth international conference on
computational creativity, pages 156–163.
K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. 2002.
A fast and elitist multiobjective genetic algorithm:
Nsga-ii.Trans. Evol. Comp, 6(2):182–197.
Yanlin Feng and Xiaojun Wan. 2019. Learning bilin-
gual sentiment-specific word embeddings without
cross-lingual supervision. In Proceedings of the
2019 Conference of the North American Chapter of
the Association for Computational Linguistics: Hu-
man Language Technologies, Volume 1 (Long and
Short Papers), pages 420–429, Minneapolis, Min-
nesota. Association for Computational Linguistics.
Brendan J Frey and Delbert Dueck. 2007. Clustering
by passing messages between data points. science,
Pablo Gerv´
as. 2001. An expert system for the compo-
sition of formal Spanish poetry. In Applications and
Innovations in Intelligent Systems VIII, pages 19–32.
Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Ar-
mand Joulin, and Tomas Mikolov. 2018. Learning
word vectors for 157 languages. In Proceedings
of the International Conference on Language Re-
sources and Evaluation (LREC 2018).
Mika H¨
ainen. 2018. Harnessing NLG to create
Finnish poetry automatically. In Proceedings of the
Ninth International Conference on Computational
Creativity, pages 9–15.
Mika H¨
ainen. 2019. UralicNLP: An NLP library
for Uralic languages.Journal of Open Source Soft-
ware, 4(37):1345.
Mika H¨
ainen and Jack Rueter. 2018. Develop-
ment of an Open Source Natural Language Gener-
ation Tool for Finnish. In Proceedings of the Fourth
International Workshop on Computational Linguis-
tics for Uralic Languages, pages 51–58.
Katri Haverinen, Jenna Nyblom, Timo Viljanen,
Veronika Laippala, Samuel Kohonen, Anna Missil¨
Stina Ojala, Tapio Salakoski, and Filip Ginter. 2014.
Building the essential resources for Finnish: the
Turku dependency treebank.Language Resources
and Evaluation, 48(3):493–531.
Kyle E. Jennings. 2010. Developing Creativity: Arti-
ficial Barriers in Artificial Intelligence.Minds and
Machines, 20(4):489–501.
Anna Jordanous. 2012. A standardised procedure for
evaluating creative systems: Computational creativ-
ity evaluation based on what it is to be creative.Cog-
nitive Computation, 4(3):246–279.
Tuomas Juntunen. 2012. Kirjallisuudentutkimus.
In Genreanalyysi: tekstilajitutkimuksen k ¨
pages 528—-536.
Jenna Kanerva, Juhani Luotolahti, Veronika Laippala,
and Filip Ginter. 2014. Syntactic n-gram collection
from a large-scale corpus of internet Finnish. In Hu-
man Language Technologies-The Baltic Perspective:
Proceedings of the Sixth International Conference
Baltic HLT, volume 268, pages 184–191.
Mervi Kantokorpi, Lyytik¨
ainen Pirjo, and Viikari Auli.
1990. Runousopin perusteet. Gaudeamus.
Justine Kao and Dan Jurafsky. 2012. A computational
analysis of style, affect, and imagery in contem-
porary poetry. In Proceedings of the NAACL-HLT
2012 Workshop on Computational Linguistics for
Literature, pages 8–17.
Guillaume Klein, Yoon Kim, Yuntian Deng, Vincent
Nguyen, Jean Senellart, and Alexander Rush. 2018.
OpenNMT: Neural machine translation toolkit. In
Proceedings of the 13th Conference of the Associa-
tion for Machine Translation in the Americas (Vol-
ume 1: Research Papers), pages 177–184, Boston,
MA. Association for Machine Translation in the
Carolyn Lamb and Daniel G. Brown. 2019. Twit-
Song 3.0: towards semantic revisions in computa-
tional poetry. In Proceedings of the Tenth Inter-
national Conference on Computational Creativity,
pages 212–219.
Carolyn Lamb, Daniel G Brown, and Charles LA
Clarke. 2018. Evaluating computational creativity:
An interdisciplinary tutorial. ACM Computing Sur-
veys (CSUR), 51(2):28.
Juntao Li, Yan Song, Haisong Zhang, Dongmin Chen,
Shuming Shi, Dongyan Zhao, and Rui Yan. 2018.
Generating classical Chinese poems via conditional
variational autoencoder and adversarial training.
In Proceedings of the 2018 Conference on Em-
pirical Methods in Natural Language Processing,
pages 3890–3900, Brussels, Belgium. Association
for Computational Linguistics.
Juri Lotman. 1974. Den poetiska texten. Stockholm.
Minh-Thang Luong, Hieu Pham, and Christopher D
Manning. 2015. Effective approaches to attention-
based neural machine translation. arXiv preprint
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Cor-
rado, and Jeff Dean. 2013. Distributed representa-
tions of words and phrases and their compositional-
ity. In Advances in neural information processing
systems, pages 3111–3119.
Joanna Misztal and Bipin Indurkhya. 2014. Poetry
generation system with an emotional personality. In
ICCC, pages 72–81.
Hugo Gonc¸alo Oliveira. 2017. A survey on intelligent
poetry generation: Languages, features, techniques,
reutilisation and evaluation. In Proceedings of the
10th International Conference on Natural Language
Generation, pages 11–20, Santiago de Compostela,
Spain. Association for Computational Linguistics.
Hugo Gonc¸alo Oliveira, Raquel Herv´
as, Alberto D´
and Pablo Gerv´
as. 2017. Multilingual extension and
evaluation of a poetry generator. Natural Language
Engineering, 23(6):929–967.
Tommi A Pirinen, Inari Listenmaa, Ryan Johnson,
Francis M. Tyers, and Juha Kuokkala. 2017. Open
morphology of finnish. LINDAT/CLARIN digital
library at the Institute of Formal and Applied Lin-
guistics, Charles University.
Ivor Armstrong Richards. 1936. The Philosophy of
Rhetoric. Oxford University Press, London, United
Yuki Tanigaki, Kaname Narukawa, Yusuke Nojima,
and Hisao Ishibuch. 2014. Preference-based nsga-
ii for many-objective knapsack problems. In 2014
Joint 7th International Conference on Soft Comput-
ing and Intelligent Systems (SCIS) and 15th Interna-
tional Symposium on Advanced Intelligent Systems
(ISIS), pages 637–642. IEEE.
Jukka Toivanen, Hannu Toivonen, Alessandro Valitutti,
and Oskar Gross. 2012. Corpus-Based Generation
of Content and Form in Poetry. In Proceedings
of the Third International Conference on Computa-
tional Creativity.
Tony Veale. 2016. The shape of tweets to come: au-
tomating language play in social networks. Multiple
Perspectives on Language Play. Mouton De Gruyter,
Language Play and Creativity series, pages 73–92.
Ping Xiao, Khalid Alnajjar, Mark Granroth-Wilding,
Kathleen Agres, and Hannu Toivonen. 2016.
Meta4meaning: Automatic metaphor interpreta-
tion using corpus-derived word associations. In
Proceedings of the 7th International Conference
on Computational Creativity (ICCC 2016), Paris,
France. Sony CSL, Sony CSL.
Cheng Yang, Maosong Sun, Xiaoyuan Yi, and Wenhao
Li. 2018. Stylistic Chinese poetry generation via un-
supervised style disentanglement. In Proceedings of
the 2018 Conference on Empirical Methods in Nat-
ural Language Processing, pages 3960–3969, Brus-
sels, Belgium. Association for Computational Lin-
Xiaoyuan Yi, Maosong Sun, Ruoyu Li, and Wenhao
Li. 2018. Automatic poetry generation with mutual
reinforcement learning. In Proceedings of the 2018
Conference on Empirical Methods in Natural Lan-
guage Processing, pages 3143–3153, Brussels, Bel-
gium. Association for Computational Linguistics.
... The computational linguistic research community should be interested in identifying for various reasons, including the need to evaluate machine-generated poetry [Agarwal and Kann, 2020;Van de Cruys, 2020;Li et al., 2020a;Hopkins and Kiela, 2017], translating poetry [Yang et al., 2019a;Ghazvininejad et al., 2018;, to gain a deeper understanding of poetic language [Haider, 2021;Hämäläinen and Alnajjar, 2019;Waltz, 1975], and to develop recommender systems [Zhang et al., 2020;Foley IV, 2019] for readers interested in poetry. By analyzing large corpora of poetry [Haider, 2021;Haider et al., 2020;Gopidi and Alam, 2019;Fairley, 1969] and identifying the most outstanding examples, researchers can improve evaluation metrics for machine-generated poetry [Yi et al., 2018;Liu et al., 2019c;Ghazvininejad et al., 2017;Gonçalo Oliveira, 2017], develop new models and techniques for natural language processing [Hu and Sun, 2020], and promote and preserve cultural heritage. ...
Full-text available
The primary focus of this thesis is to make Sanskrit manuscripts more accessible to the end-users through natural language technologies. The morphological richness, compounding, free word orderliness, and low-resource nature of Sanskrit pose significant challenges for developing deep learning solutions. We identify four fundamental tasks, which are crucial for developing a robust NLP technology for Sanskrit: word segmentation, dependency parsing, compound type identification, and poetry analysis. The first task, Sanskrit Word Segmentation (SWS), is a fundamental text processing task for any other downstream applications. However, it is challenging due to the sandhi phenomenon that modifies characters at word boundaries. Similarly, the existing dependency parsing approaches struggle with morphologically rich and low-resource languages like Sanskrit. Compound type identification is also challenging for Sanskrit due to the context-sensitive semantic relation between components. All these challenges result in sub-optimal performance in NLP applications like question answering and machine translation. Finally, Sanskrit poetry has not been extensively studied in computational linguistics. While addressing these challenges, this thesis makes various contributions: (1) The thesis proposes linguistically-informed neural architectures for these tasks. (2) We showcase the interpretability and multilingual extension of the proposed systems. (3) Our proposed systems report state-of-the-art performance. (4) Finally, we present a neural toolkit named SanskritShala, a web-based application that provides real-time analysis of input for various NLP tasks. Overall, this thesis contributes to making Sanskrit manuscripts more accessible by developing robust NLP technology and releasing various resources, datasets, and web-based toolkit.
... The computational linguistic research community could be interested in identifying fine poetry for various reasons, including the need to evaluate machine-generated poetry (Agarwal and Kann, 2020;Van de Cruys, 2020;Hopkins and Kiela, 2017), translating poetry Ghazvininejad et al., 2018;Krishna et al., 2019), to gain a deeper understanding of poetic language (Haider, 2021;Hämäläinen and Alnajjar, 2019;Waltz, 1975), and to develop recommender systems Foley IV, 2019) for readers interested in poetry. By analyzing large corpora of poetry (Haider, 2021;Haider et al., 2020;Gopidi and Alam, 2019;Fairley, 1969) and identifying the most outstanding examples, researchers can improve evaluation metrics for machine-generated poetry (Yi et al., 2018;Liu et al., 2019;Ghazvininejad et al., 2017;Gonçalo Oliveira, 2017), develop new models and techniques for natural language processing (Hu and Sun, 2020), and preserve cultural heritage. ...
Full-text available
Sanskrit poetry has played a significant role in shaping the literary and cultural landscape of the Indian subcontinent for centuries. However, not much attention has been devoted to uncovering the hidden beauty of Sanskrit poetry in computational linguistics. This article explores the intersection of Sanskrit poetry and computational linguistics by proposing a roadmap of an interpretable framework to analyze and classify the qualities and characteristics of fine Sanskrit poetry. We discuss the rich tradition of Sanskrit poetry and the significance of computational linguistics in automatically identifying the characteristics of fine poetry. The proposed framework involves a human-in-the-loop approach that combines deterministic aspects delegated to machines and deep semantics left to human experts. We provide a deep analysis of Siksastaka, a Sanskrit poem, from the perspective of 6 prominent kavyashastra schools, to illustrate the proposed framework. Additionally, we provide compound, dependency, anvaya (prose order linearised form), meter, rasa (mood), alankar (figure of speech), and riti (writing style) annotations for Siksastaka and a web application to illustrate the poem's analysis and annotations. Our key contributions include the proposed framework, the analysis of Siksastaka, the annotations and the web application for future research. Link for interactive analysis:
... These vectors can be utilized to measure the semantic similarity between words, find analogous concepts, cluster words (Hämäläinen and Alnajjar, 2019;Stekel et al., 2021) and more. In this work, we use English and Finnish as the big languages that facilitate aligning and classifying words and sentences for the endangered languages. ...
Full-text available
In this paper, we present an approach for translating word embeddings from a majority language into 4 minority languages: Erzya, Moksha, Udmurt and Komi-Zyrian. Furthermore, we align these word embeddings and present a novel neural network model that is trained on English data to conduct sentiment analysis and then applied on endangered language data through the aligned word embeddings. To test our model, we annotated a small sentiment analysis corpus for the 4 endangered languages and Finnish. Our method reached at least 56\% accuracy for each endangered language. The models and the sentiment corpus will be released together with this paper. Our research shows that state-of-the-art neural models can be used with endangered languages with the only requirement being a dictionary between the endangered language and a majority language.
... Word embeddings capture the semantic and syntactic links between words by constructing vector representations of words. These vectors can be utilized to measure the semantic similarity between words, find analogous concepts, cluster words (Hämäläinen and Alnajjar, 2019;Stekel et al., 2021) and more. In this work, we use English and Finnish as the big languages that facilitate aligning and classifying words and sentences for the endangered languages. ...
Conference Paper
Full-text available
In this paper, we present an approach for translating word embeddings from a majority language into 4 minority languages: Erzya, Moksha, Udmurt and Komi-Zyrian. Furthermore, we align these word em-beddings and present a novel neural network model that is trained on English data to conduct sentiment analysis and then applied on endangered language data through the aligned word embeddings. To test our model, we annotated a small sentiment analysis corpus for the 4 endangered languages and Finnish. Our method reached at least 56% accuracy for each endangered language. The models and the sentiment corpus will be released together with this paper. Our research shows that state-of-the-art neural models can be used with endangered languages with the only requirement being a dictionary between the endangered language and a majority language.
... Two expert annotators went through the video files and annotated metaphors by surrounding them with v tags for vehicles and t tags for tenors. The use of experts is motivated by the fact that previous research has found that non-expert annotators struggle with metaphors (Hämäläinen and Alnajjar, 2019). ...
Full-text available
We present the first openly available multimodal metaphor annotated corpus. The corpus consists of videos including audio and subtitles that have been annotated by experts. Furthermore, we present a method for detecting metaphors in the new dataset based on the textual content of the videos. The method achieves a high F1-score (62\%) for metaphorical labels. We also experiment with other modalities and multimodal methods; however, these methods did not out-perform the text-based model. In our error analysis, we do identify that there are cases where video could help in disambiguating metaphors, however, the visual cues are too subtle for our model to capture. The data is available on Zenodo.
... Two expert annotators went through the video files and annotated metaphors by surrounding them with v tags for vehicles and t tags for tenors. The use of experts is motivated by the fact that previous research has found that non-expert annotators struggle with metaphors (Hämäläinen and Alnajjar, 2019). ...
Conference Paper
Full-text available
We present the first openly available multi-modal metaphor annotated corpus. The corpus consists of videos including audio and subtitles that have been annotated by experts. Furthermore , we present a method for detecting metaphors in the new dataset based on the tex-tual content of the videos. The method achieves a high F1-score (62%) for metaphorical labels. We also experiment with other modalities and multimodal methods; however, these methods did not out-perform the text-based model. In our error analysis, we do identify that there are cases where video could help in disambiguat-ing metaphors, however, the visual cues are too subtle for our model to capture.
... In the field of poem generation, meter is a feature that is very often considered in generated poetry (Colton, Goodwin, and Veale 2012;Lau et al. 2018;Hämäläinen and Alnajjar 2019b;Zugarini, Melacci, and Maggini 2019;Lewis, Zugarini, and Alonso 2021). By incorporating meter, people can be more forgiving when evaluating the output of the system as it is known that people are ready to interpret more into the content of the output of a computationally creative system if the form is correct (Veale 2016). ...
Conference Paper
Full-text available
We present a novel neural model for modern poetry generation in French. The model consists of two pretrained neural models that are fine-tuned for the poem generation task. The encoder of the model is a RoBERTa based one while the decoder is based on GPT-2. This way the model can benefit from the superior natural language understanding performance of RoBERTa and the good natural language generation performance of GPT-2. Our evaluation shows that the model can create French poetry successfully. On a 5 point scale, the lowest score of 3.57 was given by human judges to typ-icality and emotionality of the output poetry while the best score of 3.79 was given to understandability.
... Word Sense Induction and Word Sense Disambiguation provide fertile ground for researchers, starting from very early attempts to tackle these non-trivial tasks, such as "simulated annealing" according to human-edited dictionary Cowie et al. [1992] and employing the "conceptual distance" between contexts Agirre and Rigau [1996], going through later unsupervised methods, that use patterns of word co-occurrence Bordag [2006] or bigrams of web search results Udani et al. [2005], continuing with "hidden concepts" of the contextual words, that not necessarily overlap with the sense of the ambiguous word Chang et al. [2014], and ending with the most recent solutions like Eyal et al. [2021], that uses word substitutions of modern Masked Language Models, such as Google BERT MLM. In addition, the technique of pipelining a clustering algorithm and cosine distance (note that we do this in reverse order) is used by Hämäläinen and Alnajjar [2019] for dividing words (not necessarily polysemous or even similar) into semantic fields. This idea is not directly related to the Word Sense Induction task but it is another approach to dividing a given text into different semantic groups by using clustering and cosine distance. ...
Full-text available
This paper presents ACCWSI (Attentive Context Clustering WSI), a method for Word Sense Induction, suitable for languages with limited resources. Pretrained on a small corpus and given an ambiguous word (a query word) and a set of excerpts that contain it, ACCWSI uses an attention mechanism for generating context-aware embeddings, distinguishing between the different senses assigned to the query word. These embeddings are then clustered to provide groups of main common uses of the query word. We show that ACCWSI performs well on the SemEval-2 2010 WSI task. ACCWSI also demonstrates practical applicability for shedding light on the meanings of ambiguous words in ancient languages, such as Classical Hebrew and Akkadian. In the near future, we intend to turn ACCWSI into a practical tool for linguists and historians.
Full-text available
Sentiment analysis and opinion mining are essential tasks with many prominent application areas, e.g., when researching popular opinions on products or brands. Sentiments expressed in social media can be used in brand name monitoring and indicating fake news. In our survey of previous work, we note that there is no large-scale social media data set with sentiment polarity annotations for Finnish. This publication aims to remedy this shortcoming by introducing a 27,000-sentence data set annotated independently with sentiment polarity by three native annotators. We had three annotators annotate the whole data set, which provides a unique opportunity for further studies of annotator behavior over the sample annotation order. We analyze their inter-annotator agreement and provide two baselines to validate the usefulness of the data set.
Conference Paper
Full-text available
This paper presents a new, NLG based approach to poetry generation in Finnish for use as a part of a bigger Poem Machine system the objective of which is to provide a platform for human computer co-creativity. The approach divides generation into a linguistically solid system for producing grammatical Finnish and higher level systems for producing a poem structure and choosing the lexical items used in the poems. An automatically extracted open-access semantic repository tailored for poem generation is developed for the system. Finally , the resulting poems are evaluated and compared with the state of the art in Finnish poem generation.
Full-text available
In the past years the natural language processing (NLP) tools and resources for small Uralic languages have received a major uplift. The open-source Giellatekno infrastructure has served a key role in gathering these tools and resources in an open environment for researchers to use. However, the many of the crucially important NLP tools, such as FSTs and CGs require specialized tools with a learning curve. This paper presents UralicNLP, a Python library, the goal of which is to mask the actual implementation behind a Python interface. This not only lowers the threshold to use the tools provided in the Giellatekno infrastructure but also makes it easier to incorporate them as a part of research code written in Python.
Conference Paper
Full-text available
We present an open source Python library to automatically produce syntactically correct Finnish sentences when only lemmas and their relations are provided. The tool resolves automatically morphosyntax in the sentence such as agreement and government rules and uses Omorfi to produce the correct morphological forms. In this paper, we discuss how case government can be learned automatically from a corpus and incorporated as a part of the natural language generation tool. We also present how agreement rules are modeled in the system and discuss the use cases of the tool such as its initial use as part of a computational creativity system, called Poem Machine.
This article is a tutorial for researchers who are designing software to perform a creative task and want to evaluate their system using interdisciplinary theories of creativity. Researchers who study human creativity have a great deal to offer computational creativity. We summarize perspectives from psychology, philosophy, cognitive science, and computer science as to how creativity can be measured both in humans and in computers. We survey how these perspectives have been used in computational creativity research and make recommendations for how they should be used.