No source: created in electronic format.
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.
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
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 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>
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,
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 daughter is the 12th token in the text).
<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
<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 Interdiction function label used earlier in the
text.
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.
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,