Propp Revisited: Integration of Linguistic Markup into Structured Content Descriptors of Tales

1
Propp Revisited:
Integration of Linguistic
Markup into Structured
Content Descriptors of
Tales
Lendvai, Piroska
piroska@nytud.hu
Research Institute for Linguistics, Hungarian
Academy of Sciences, Budapest, Hungary
Declerck, Thierry
declerck@dfki.de
Language Technology Lab, DFKI GmbH,
Saarbrücken, Germany
Darányi, Sándor
Sandor.Daranyi@hb.se
Swedish School of Library and Information
Science, University College Boras/Göteborg
University, Sweden
Malec, Scott
malec@andrew.cmu.edu
Carnegie Mellon University, USA
Metadata that serve as semantic markup, such
as conceptual categories that describe the
macrostructure of a plot in terms of actors
and their mutual relationships, actions, and
their ingredients annotated in folk narratives,
are important additional resources of digital
humanities research. Traditionally originating
in structural analysis, in fairy tales they are
called functions (Propp, 1968), whereas in
myths – mythemes (Lévi-Strauss, 1955); a
related, overarching type of content metadata is
a folklore motif (Uther, 2004; Jason, 2000).
In his influential study, Propp treated a corpus
of tales in Afanas'ev's collection (Afanas'ev,
1945), establishing basic recurrent units of the
plot ('functions'), such as
Villainy, Liquidation
of misfortune, Reward,
or
Test of Hero
,
and the combinations and sequences of
elements employed to arrange them into
moves.
1
His aim was to describe the DNA-
like structure of the magic tale sub-genre as
a novel way to provide comparisons. As a
start along the way to developing a story
grammar, the Proppian model is relatively
straightforward to formalize for computational
semantic annotation, analysis, and generation
of fairy tales. Our study describes an effort
towards creating a comprehensive XML markup
of fairy tales following Propp's functions, by
an approach that integrates functional text
annotation with grammatical markup in order to
be used across text types, genres and languages.
The Proppian fairy tale Markup Language
(PftML) (Malec, 2001) is an annotation scheme
that enables narrative function segmentation,
based on hierarchically ordered textual content
objects. We propose to extend PftML so
that the scheme would additionally rely on
linguistic information for the segmentation
of texts into Proppian functions. Textual
variation is an important phenomenon in
folklore, it is thus beneficial to explicitly
represent linguistic elements in computational
resources that draw on this genre; current
international initiatives also actively promote
and aim to technically facilitate such integrated
and standardized linguistic resources. We
describe why and how explicit representation of
grammatical phenomena in literary models can
provide interdisciplinary benefits for the digital
humanities research community.
In two related fields of activities, we address
the above as part of our ongoing activities in
the CLARIN
2
and AMICUS
3
projects. CLARIN
aims to contribute to humanities research by
creating and recommending effective workflows
using natural language processing tools and
digital resources in scenarios where text-based
research is conducted by humanities or social
sciences scholars. AMICUS is interested in motif
identification, in order to gain insight into
higher-order correlations of functions and other
content units in texts from the cultural heritage
and scientific discourse domains. We expect
significant synergies from their interaction with
the PftML prototype.
1. Proppian fairy tale Markup
Language (PftML)
Creating PftML was based on the insight that
Propp's functions – organized in tables to
categorize his observations – were analogous
to metadata, and as such renderable by
hierarchically arranged elements in eXtensible
Markup Language (XML) documents. A tale

2
consists of one or more moves and on a
lower level of functions which are modeled as
elements. Function elements themselves have
XML attributes that allow for the efficient
extraction of data from the text using XQuery
from within a native XML database. The
embedded structure of Proppian functions as
represented by PftML markup is illustrated in
Fig. 1 by an annotated excerpt from the English
translation of the Russian fairy tale
The Swan-
Geese
.
Note that Proppian functions are applied
to relatively long, semantically coarse-grained
textual chunks, i.e. sentences, but linguistic
elements that convey a function actually
encompass a shorter sequence of words; e.g.
contrary to the markup in the example, both
Command
and
Execution
only pertain to
linguistic units smaller than full sentences.
<Corpus>
<Folktale Title="The Swan-Geese" AT="480"
NewAfanasievEditionNumber="113"
ProppConformity="Yes">
<Move>
<Preparation>
<InitialSituation> Once upon a time a man
and a woman lived with their daughter and small
son. </InitialSituation>
<CommandExecution>
<Command subtype="Interdiction">
"Dearest daughter," said the mother, "we are
going to work. Look after your brother! Don't
go out of the yard, be a good girl, and we'll
buy you a handkerchief." </Command>
<Execution subtype="Violated"> The
father and mother went off to work, and the
daughter soon enough forgot what they had told
her. She put her little brother on the grass
under a window and ran into the yard, where she
played and got completely carried away having
fun.</Execution>
</CommandExecution>
</Preparation>
<Villainy subtype="Kidnapping"> In swooped the
swan-geese, snatched up the little boy, and
flew away with him. </Villainy>
<ConsentToCounteraction> When the girl came
back inside, her brother was missing! "Oh
no!" she cried. She dashed here and there, but
there was no sign of him. She called for him,
cried, and wailed how angry mother and father
would be, but her brother did not answer. </
ConsentToCounteraction>
2. Integration of PftML with
linguistic annotation
We propose to combine PftML with a stand-
off, multi-layered linguistic markup scheme
to ensure modularity and reusability of
linguistic information associated with textual
elements, supporting interoperability of fairy
tales annotation in different languages and
versions. As seen in Fig. 1, PftML is interleaving
the Proppian annotation with the text. This in-
line annotation strategy has some drawbacks:
a text can hardly be annotated in fine-grained
manner without losing readability, or with
information originating from different sources
e.g. indicating different views on narrative
functions.
Stand-off annotation strategy, following
the standardization initiatives for language
resources conducted within ISO,
4
stores
annotation separately from the original text,
linking these by referencing mechanisms.
We adopt the ISO multi-layered annotation
strategy, representing linguistic information
on the following levels: segmentation of the
text in tokens; morpho-syntactic properties
of the tokens; phrasal constituencies;
grammatical dependencies; semantic relations
(e.g. temporal, co-referential), cf. (Ide and
Romary, 2006).
We illustrate how the linguistic annotation
layers can be combined with the PftML
annotation in one stand-off annotation file,
showing here only the morphosyntactic and
constituency annotation, as they are applied
to the first five tokens of the sub-sentence
annotated with the 'Violated Execution' function
in Fig. 1. In the morphosyntactic annotation, the
value of the
TokenID
of the 12th word is pointing
to the original data (e.g.,
daughter
is the 12th
token in the text).
5
<wordForms>
<W ID="w11" POS="ART" LEMMA="the" MORPH="Sg"
tokenID="t11">the</W>
<W ID="w12" POS="NN" LEMMA="daughter"
MORPH="Sg" tokenID="t12">daughter</W>
<W ID="w13" POS="ADV" LEMMA="soon"
tokenID="t13">soon</W>
<W ID="w14" POS="ADV" LEMMA="enough"
tokenID="t14">enough</W>
<W ID="w15" POS="VVFIN" LEMMA="forget"
MORPH="Past" tokenID="t15">forgot</W>
...
</wordForms>
In the constituency annotation level displayed
below, words are grouped into syntactic
constituents (e.g. the nominal phrase
the
daughter
). The span of constituents is marked
by the value of the features
from
and
to
, which

3
are pointing to the previous morpho-syntactic
annotation layer.
<phrases>
<phrase id="p4" from="w11" to="w12"
type="NP">the daughter</phrase>
<phrase id="p5" from="w13" to="w14"
type="ADVP">soon enough</phrase>
<phrase id="p6" from="w15" to="w15"
type="VG">forgot</phrase>
<phrase id="p7" from="w16" to="w20"
type="REL_COMP">what they had told her
</phrase> ...
</phrases>
PftML and (for example) word-level annotation
can be combined in one stand-off XML element,
where each specific PftML annotation receives a
span of textual segments associated with it:
<Execution subtype="Violated" id="e1"
inv_id="Command1" from="w11" to="w21"> </
Execution>
The values
w11
and
w21
are used for defining a
region of the text for which the Propp function
holds;
Command1
refers to the
Interdiction
function label used earlier in the text.
6
3. Benefits for humanities
research
The integrated annotation scheme enables
narrative segmentation enhanced by additional
information about the linguistic entities
that constitute a given function. A folklore
researcher might be interested in which
natural language expressions correspond to
which narrative function: in the <
Execution
subtype="Violated"
> example,
forgot
can be an
indicator of this function. In fact, it is also
relevant to signal that
forgot
is a verb, and
to reduce the strings
forgets, forgot, forgotten
to one lemma (i.e. base form), so that all
morphological forms are retrieved when any of
these variants is queried.
Navigating through the different types of IDs
included in the multilayered annotation, a
researcher can obtain statistics over linguistic
properties of fairy tales. For example, the
grammatical subject of a function can be
extracted, e.g. to see which characters
participate in commands and their violation.
Note that if – according to the current scheme –
the narrative function boundaries are imprecise,
the <
Execution subtype="Violated"
> function
in our example sentence would incorrectly
contain two grammatical – and three semantic
– subjects (
father and mother
, and
daughter
).
Linguistic information will enable detecting
functions that refer to each other, as syntax and
semantics of sentence pairs in such relations
mirror – at least partly – each other, e.g.
Don't
go out of the yard
and
ran onto the street
.
Detecting cross-reference in turn contributes to
identifying a function's core elements, which is
a crucial step in understanding the linguistic
vehicles by which motifs operate and the degree
of variation and optionality they allow.
4. Concluding remarks
Since the content descriptors in PftML might
pertain to textual material on the supra-
or subsentential level, there is a need to
investigate the mechanisms underlying the
assignment of a function to a span of
words. We propose to tackle this issue
based on linguistic analysis, hypothesizing that
boundaries of certain linguistic objects overlap
with boundaries of Proppian functions. A direct
consequence of more precise segmentation of
functions is that linguistic characterization,
retrieval, and further computational processing
of texts from the folktale genre will improve, and
facilitate detecting higher-level, domain-specific
cognitive phenomena. It would also become
feasible to detect from corpus evidence if
there exist additional functions beyond Propp's
scheme.
Integration along the above lines with
ontological resources of fairy tales is described
in a separate study by us (Lendvai et al., 2010).
We expect from our strategy – applied to tales
in different versions in different languages –
to lead to the generation of a multilingual
ontology of folktale content descriptors, which
would be extending the efforts of the MONNET
project,
7
originally focussing on financial and
governmental issues. In future work we plan
to address embedding our annotation work
into the TEI framework,
8
and extend the ISO
strategy on using well-defined data categories
for linguistic annotation labels
9
to those of
functions corresponding to PftML labels, to
facilitate porting our approach to other literary
genres.

4
References
Afanas'ev, A.
(1945).
Russian fairy tales.
New
York: Pantheon Books.
Ide, N. and Romary, L.
. 'Representing
linguistic corpora and their annotations'.
Proc.
of LREC.
2006.
Jason, H.
(2000).
Motif, type and genre.
A manual for compilation of indices and a
bibliography of indices and indexing.
Helsinki:
Academia Scientiarum Fennica.
Lendvai, P., Declerck, T., Darányi, S.,
Gervás, P., Hervás, R., Malec, S., and
Peinado, F.
. 'Integration of linguistic markup
into semantic models of folk narratives: The
fairy tale use case'.
In Proc. of LREC.
2010.
Lévi-Strauss, C.
(1955). 'The structural study
of myth'.
Journal of American Folklore.
68
:
428–444.
Malec, S. A.
. 'Proppian structural analysis and
XML modeling'.
In Proc. of CLiP.
2001.
Propp, V. J.
(1968).
Morphology of the
folktale.
Austin: University of Texas Press.
Uther, H. J.
(2004).
The types of international
folktales: a classification and bibliography.
Based on the system of Antti Aarne and Stith
Thompson.
Helsinki: Academia Scientiarum
Fennica.
Notes
1.
The full list of functions is available at
http://clover.sla
vic.pitt.edu/sam/propp/praxis/features.html
2.
http://www.clarin.eu
3.
http://ilk.uvt.nl/amicus
4.
http://www.tc37sc4.org
5.
Note that the original string is normally not present in these
layers but is displayed in annotation examples for readability's
sake.
6.
We started to implement this work within the D-SPIN
project (see
http://www.sfs.uni-tuebingen.de/dspi
n
), which is the German complementary project to CLARIN.
7.
Multilingual ONtologies for NETworked Knowledge,
see
http://cordis.europa.eu/fp7/ict/languagetec
hnologies/project-monnet_en.html
8.
http://www.tei-c.org/index.xml
9.
see
http://www.isocat.org/