=Paper=
{{Paper
|id=Vol-2347/paper12
|storemode=property
|title=Functional Unit Analysis: Framing and Aesthetics for Computational Storytelling
|pdfUrl=https://ceur-ws.org/Vol-2347/paper12.pdf
|volume=Vol-2347
|authors=Sven Wilke,Leonid Berov
|dblpUrl=https://dblp.org/rec/conf/c3gi/WilkeB18
}}
==Functional Unit Analysis: Framing and Aesthetics for Computational Storytelling==
https://ceur-ws.org/Vol-2347/paper12.pdf
Functional Unit Analysis: Framing and
Aesthetics for Computational Storytelling
Sven WILKE, Leonid BEROV 1
Institute for Cognitive Science, University of Osnabrück, Osnabrück, Germany
Abstract. The plots of narratives can be analyzed by decomposition into functional
units (FUs): complex structural elements that carry semantic meaning. A manual
approach to the identification of FU using a graph-representation of plot has been
proposed for human-made narratives, and on this basis used for plot summarization.
A computational utilization, however, has remained inviable due to the complexity
of the involved interpretative tasks. The present paper explores the use of the FU
approach for the analysis of computationally generated narratives, because in this
case interpretation can be evaded. It demonstrates that an agent-based approach to
computational storytelling can be used to generate and aggregate all necessary plot
phenomena in order to identify FUs. This is validated in-vivo by demonstrating that
the extracted FU structure can be used to summarize, and aesthetically evaluate a
plot; as has been predicted by the underpinning narrative theory. We argue that such
an analysis is beneficial because it provides the system with an abstract, functional
understanding of the generated artifacts, a step towards more-general intelligence.
For instance, in the computational creativity domain it allows a system that has been
formerly only capable of expression-type generative acts to also perform aesthetic-
and framing-type generative acts.
Keywords. computational creativity, aesthetics, framing, computational storytelling,
summarization
1. Introduction
Apart from generating novel artistic artefacts, computational systems need to perform
other tasks in order to be deemed creative [7]: Such systems strongly benefit from being
able to evaluate their output and from presenting it in a special light. This means that
besides purely generative tasks they also need to perform analytical tasks. The present
paper suggests that in the context of computational plot generation this can be achieved
through the use of functional-unit-based plot analysis.
1.1. Functional Unit Based Plot Analysis
The functional unit (FU) model was proposed as a tool for the computational summa-
rization of existing narratives by Lehnert [13]. It operates on a graph representation of
plot and works by identifying strategically significant portions of the plot called com-
plex FU, which are points of high relevance for summaries. Plot graphs, in this model,
1 Corresponding Author: lberov@uos.de
Figure 1. A sample of primitive units adopted from [13]. I denotes an intention, + and - are vertices of positive
and negative affect.
can contain three different types of vertices, representing affect states of characters’ per-
ceptions of the events of the plot, and connected by different types of edges. The affect
states that can be contained in a FU are: positive (denoted: +), negative (−) and neutral
affect (I). Positive states describe any event which is appraised with positive emotions
by a character. Inversely, negative states describe events which are negatively appraised.
States of neutral affect indicate “mental states”; in this formalization always intentions.
Edges describe how these states are related to each other, and can be of the following
types: motivation, actualization, termination, equivalence or inter-character edges. Based
on this formalism, [13] defines “15 legal pairwise configurations” (primitive FUs, e.g.
‘motivation’ or ‘loss’, see Fig. 1) that act as an alphabet, as well as a potentially open set
of complex units (e.g. ‘denied request’ or ‘retaliation’, see Fig. 2) which are assembled
from them and were defined based on a narratological analysis. An existing story is anal-
ysed by transforming the story-text into the introduced graph-representation, and then
detecting all FUs contained in it. When this is done, a connectivity graph is built by using
the different instances of FUs as vertices, and connecting them with edges wherever two
unit instances share one or more vertices in the plot graph.
Lehnert proposed a procedure for generating summaries based on connectivity
graphs [13]. For that, the units contained in the graph get translated into natural language
by using template-like generational frames which are supplied to the program for each
unit type. Into the frames, information about the specific instance of a unit is fed, al-
lowing the frames to generate text including the characters involved in the unit, and the
concrete context of the instance.
An attempt at implementing this procedure was recently made [12] with modest re-
sults, which demonstrate the complexity involved in translating narrative text into the
proposed graph representation. The main problem being that the required natural lan-
guage processing includes complex interpretative tasks like event-based discretization,
intention and emotion detection as well as the identification of causality relations—not
commonly addressed in research.
1.2. Use in Computational Storytelling
In a computational storytelling system, plots are generated by the system itself so that
all required phenomena can potentially be inferred from computational representations
instead of being laboriously gleaned from a text. Ideally, the system’s internal represen-
tation of events already uses a graph-based format, which can lend itself for the identi-
fication of FUs. This is the case with the simulation-based storytelling system presented
Figure 2. Examples of complex FUs adopted from [13]. ‘?’ represents wildcard vertices.
in [1], and indeed a theoretical reformulation of the concepts required for plot unit anal-
ysis in terms of the system’s architecture has been proposed in prior work [2].
Aside from their original purpose, FUs can also provide valuable information about
the aesthetics of a generated plot. As proposed by Ryan [16], the strategic significance
captured by FUs is important for the so-called tellability measure of plot quality. Tella-
bility, here, is understood as the suitability of a set of events to be rendered as a story,
i.e. a pre-textual measure of a plot’s aesthetics. In its complete version it covers different
aspects like motivic, contextual and formal properties. Ryan’s focus is mainly on the last,
which again consists of multiple factors such as semantic opposition, semantic symmetry
but also functional polyvalence. Functional polyvalence is given, when the same event
fulfills several functions in the overall plot. It can be identified through a FU analysis, by
simply searching for events that are part of several complex units. Therefore, identifying
the amount of overlap of FUs in a story allows an insight into the quality of its plot.
The goal of this paper is to present our implementation of FU based plot analysis
in the above mentioned storytelling system. We argue for the usefulness of this analysis-
method by demonstrating with a case study that it can be used to perform a cautious first
aesthetic evaluation of plots and to automatically generate functional summaries, which
can be used as framing. For the proof-of-concept in this paper most parts of the natural
language generation is hard-coded, however, nothing precludes an extension with more
linguistic knowledge at a later stage of the project.
2. Related Work
To outline the motivation for this project related work has to be introduced on the topic
of computational creativity, while the specific technical contribution has to be contextu-
alized in the field of textual summarization.
2.1. Computational Creativity
Computational creativity research is concerned with “computational systems which, by
taking on particular responsibilities, exhibit behaviours that unbiased observers would
deem to be creative” [8]. The different tasks (“generative acts”) creative systems can per-
form have been analysed to fall into four categories: creating exemplars, creating con-
cepts, executing an aesthetic evaluation and creating a framing for the generated artefact
(“the FACE model”) [7].
Framing is contextual information that is created by an artist to put their work in a
certain context, and has been known to increase an artifact’s aesthetic value as well as
the creativity that is ascribed to the generating process. In ‘the wild’ it has been known
to vary wildly in form [4]. For computational systems, [4] postulated that framing can
be an outline of the system’s motivation, describe its intention when creating the art-
work, or provide insights into how the artwork was created. We argue that, additionally,
framing may also be concerned with presenting the artwork itself in a special light, that
is, framing-as-something (the most notable example probably being Duchamp’s piece
‘Fountain’, a ready-made urinal framed as a fountain). Based on this extension we sug-
gest that in the case of computational storytelling, one potential piece of framing infor-
mation could be a short interpretative summary of the plot, outlining abstract details like
formal structure or theme-like higher meanings. An early example for work towards this
direction can be found in the system Minstrel [19] that provided a moral for each story.
Our work follows in this vain but differs in that it focuses its summary on laying out the
functional structure of the plot (i.e. form) instead of its meaning (i.e. theme).
It has been argued that aesthetic appreciation is one of the three principles that guide
observers in their ascription or non-ascription of creativity [6]. Indeed, an aesthetic eval-
uation of generated artifacts is important for a system in order to guide its generation
and perform a selection of the valuable pieces from the bulk of its generation. This is a
hard problem, since in most artistic domains no universally accepted aesthetic theories
prevail, let alone a quantifiable evaluation function. Some systems have instead opted
to quantitatively approximate reader-oriented notions like suspense [5] or tension [15].
Our work instead focuses only on the structure of the generated artefact and attempts to
capture a notion of aesthetics brought forward in the context of narrative theory in order
to capture the elegance and cohesion of a plot.
2.2. Text Summarization
Two main types of tasks exist for text summarization: extractive summarization aims at
extracting the main information-bearing sentences in the source text, while abstractive
summarization generates text not contained in the input based on some sort of reason-
ing about the content (see e.g. [11]). Most work performed in this context is concerned
with the analysis of existing, argumentative text. Presently, the state of the art is achieved
through the use of deep sequence-to-sequence neural networks, which are trained on
large corpora of text in a supervised way [17]. Our approach differs radically in that it
operates not on the surface realization of text, but rather a graph-based deep structure
with clear functional semantics. Furthermore, it does not extrapolate summarization rules
based on a corpus of existing summaries, but instead by detecting the presence of a fixed
set of FUs and the way they are connected. Because this implies an act of functional
analysis, this approach is located in the abstractive realm of summarization. It is further-
more based on narratological analysis and thus more appropriate for the present domain
than models trained on news texts.
3. Implementation
As previously mentioned, the storytelling system’s architecture was already put into the
context of the FU unit model. Three steps were needed from that state to generate sum-
maries based on the FUs of the plot: (1) The instances of FUs had to be identified in the
plot graph, (2) The connectivity graph had to be built from the identified instances and
(3) Natural language had to be generated from the created connectivity graph.
3.1. Functional Unit Identification
The best way to enable the detection of arbitrary complex FUs is to ensure that the
whole alphabet of primitive units can be detected. For this, the capturing of vertices of
intentions, positively or negatively appraised events was implemented in the storytelling
system as envisioned in [2]. With the addition of motivation and actualization edges in
the way [2] suggested, these vertices build most of the primitive plot units. Termination
edges were implemented differently, as not only events that cause a change in the be-
lief base (i.e. positive or negative trade-off) can be the source of a termination edge, but
also the change of the belief base itself (i.e. loss and resolution, see fig. 1). The authors
also predicted no need for equivalence edges between intention vertices, but in order to
capture the “perseverance” primitive unit along with the related complex unit “starting
over”, these connections needed to be established. No equivalence edges were imple-
mented between non-neutral affect states. Instead, they were introduced as “causality”
edges, which depicts the semantics of the primitive plot units “complex positive event”
and “complex negative event”, as found in [13], more closely.
The identification of FUs in a plot graph can then be performed through application
of the subgraph-isomorphism algorithm to the plot graph and each FU graph. In our
implementation, the VF2 algorithm was used [9]. Since for the connectivity graph not
only complex, but also primitive FUs are required, the isomorphism search is conducted
for both types of units.
3.2. Constructing the Connectivity Graph
For each identified unit a respective vertex is added to the connectivity graph, and edges
are added between each vertex pair that represents two FU instances with overlap in at
least one vertex in the plot graph. As suggested in [13], the complexity of the graph is
reduced by removing all units that are entailed by another unit, which leaves all top-level
FUs in the graph.
While this helped to reduce the number of instances immensely, resulting graphs
still contained small numbers of FU clusters with no interconnection. This arises from
the properties of the storytelling system, which generates actions for each character even
‘outside of the main story line’. Thus, all instances are removed from the graph, which
are not in a cluster with (i.e. directly or indirectly connected to) at least one complex FU
instance.
In the original proposal, the connectivity graphs were undirected. We found it helpful
to make the edges directed and let them represent temporal precedence. This allows
summarizing in a temporally ordered fashion. As FU instances usually cover a time span,
the execution step at which the first vertex of an instance occurred is chosen to temporally
compare instances to each other (see Fig. 4 in the Appendix).
Figure 3. An example of a complex FU (nested subgoal) that matches a set of vertices in the plot. The corre-
sponding template extracts information from the plot graph, combines it with the FU’s semantics and results in
a textual representation of the match.
3.3. Natural Language Summary Generation
The natural language generation algorithm was implemented prototypically, with ori-
entation towards realizing the surface form of our case study. Better generalization to
other graphs can be achieved by replacing instances of hard-coded text with linguistic
knowledge.
First of all, the system selects which FU instances are to be included in the sum-
mary. For this, all instances of complex FUs are ordered with regard to the number of
incident vertices (neighbors) in the connectivity graph. Then, the instance of each unique
FU that has the most neighbors is chosen as the most relevant instance of that unit. Af-
terwards, these instances get ordered temporally, such that the resulting summary will
follow (roughly) the same order as the story.
After these units have been selected, they are each translated into natural language.
For this, a natural language template has been setup for each FU. The templates are able
to include semantic information which was gathered from the plot graph. This includes
information about which characters participated in a FU, as well as the content of one of
the vertices which belongs to the unit. Which vertex’ content gets extracted depends on
the FU and was determined manually beforehand.
An example for how a FU maps to vertices of the plot graph is shown in Figure 3.
Together with the natural language template for the unit the system can create a natural
language snippet for this instance of the FU.
With this in place, the full text for each FU can be generated. The last step consists
of connecting the natural language snippets of all translated units into a single sentence,
by connecting them via and-conjunction. Example summaries in the context of our case
study will be presented in the Case Study section.
3.4. Tellability Computation
As outlined above, functional polyvalence (an event fulfilling several functions for the
plot) is taken to be one factor of tellability. This property is given, when a vertex belongs
to more than one complex FU. We can therefore compute a plot’s functional polyvalence
f p by counting the number of vertices which belong to more than one complex FU, and
dividing this number by the total number of vertices in the plot:
∑ p(v)
v∈V
f p(G) =
|V |
with G = (V, E) denoting the plot-graph and the function p (for polyvalence) returning
1 iff vertex v is part of more than one FU, and 0 otherwise. The normalization allows a
comparison of plots of different length with regard to tellability.
Since the underlying narrative theory does not provide information on the precise
quantification of tellability, this computation is subject to further empirical evaluation.
However, such a measure already allows a rudimentary generate-and-test exploration of
a plot space. For instance, in the employed storytelling system plot can be changed by
altering the characters’ personalities, and functional polyvalence can be used to compare
the resulting plots. The results of such an exploration in the context of our case study
will be discussed in the next section.
4. Case Study
The usefulness of the above implementation is explored via a case study on the story
“The Little Red Hen”. Our storytelling system is capable of recreating the plot of this
fable by simulating the affect and interactions of the involved characters, and can explore
the plot space spanned by this narrative universe by changing the characters’ personality
parameters (see [2,3]).
Three variants of the story were chosen as exemplary versions2 . The original fable
achieved the highest tellability score of the three with 0.059. In the second variant, where
the hen does not punish the other animals for rejecting her requests for help but instead
shares the bread, the computed tellability was 0.023. The last version contains no inter-
action between the animals: The hen bakes the bread alone, and eats it alone, without
ever asking for help. This version scored lowest of the three, with a tellability score of
0.005. This ordering is compatible with our intuitions.
The tellability measure was also used in a generate-and-test search, where person-
ality parameters were varied systematically and the resulting plot’s tellability was mea-
sured. For computability reasons the search space had to be reduced by assuming that
the hen has a distinct personality, but the other animals have the same. This revealed that
tellability shows strong signs of locality, with hard boundaries. We assume that these
boundaries arise from the discretization of the mood and personality traits in the agent’s
plan library, where behavior changes qualitatively only among the discrete values of
high, medium, low, positive and negative.
Generating summaries from the connectivity graphs of the fables required further
pre-processing. As the stories contain symmetry between the characters, it was more
difficult for the algorithm to create a summary from the graph, because of the amount
of similar but distinct vertices. We hence implemented a fairy-tale specific symmetry-
detector which detects and merges vertex-pairs of the same FU belonging to different
characters. Essentially, we are assuming that these FU (and character-) instances are
equivalent for the summary. This resulted in the following summaries for the three story
variants:
2 A full natural language version, along with each story variant’s plot graph, can be found at http://www.
home.uni-osnabrueck.de/leberov/tlrh_versions.htm
1. I wanted to write a story in which the hen takes up a complex plan to create bread,
the pig, the cow and the dog deny the hen’s request for help and the hen retaliates
against the pig, the cow and the dog by punishing them.
2. I wanted to write a story in which the hen takes up a complex plan to create bread
and the pig, the cow and the dog deny the hen’s request for help.
3. I wanted to write a story in which the hen takes up a complex plan to create bread.
5. Conclusion
This paper has presented a prototypical implementation of FU analysis in a computa-
tional storytelling system. The system was enabled to create a graph-representation of
plot that captures events representing intentions, positive and negative affects as well
as their interrelations of the types: motivation, actualization and termination. Based on
this representation, the detection of primitive and complex FUs (structurally significant
building blocks of the plot) was enabled using a subgraph-isomorphism based search. A
case study was employed for in-vivo evaluation, where the presented system was applied
to the plot of a fable.
Several difficulties were uncovered. A first was that inter-character edges are re-
quired not only for units that cover speech acts, but also events that are simultaneously
perceived by several characters. This functionality is not easily enabled in the underlying
storytelling system, which resulted in the inability to detect significant complex units.
More relevantly, the case study raised our suspicion about the robustness of the approach.
The complex FUs defined by Lehnert were not in all instances capable of matching the
intended semantics in the automatically generated plot graphs, and hence had to be re-
designed. This is not per se problematic, since their original creation was arbitrary in
nature. Yet, it demonstrates that the same semantic meaning can be matched by a mul-
titude of possible FUs, a clear sign of poor generalization. To combat this problem we
implemented several modifications, like polyemotional vertices and wild-card edges, yet
generally suspect that a more flexible formalism is advisable.
In general we evaluate the case study positively. FU analysis enabled the system to
create plausible abstract summaries of story plots, and an aesthetic analysis of these plots
based on functional polyvalence returned results that are compatible with our intuition.
We thus suggest, that FU based analysis can be gainfully employed in the computational
creativity context to perform framing and aesthetic evaluation tasks. To be able to ben-
efit from this method in a comparable way, other storytelling systems need to represent
characters’ affect, intentions, and the causal connections between them. However, we do
not consider this a drawback since the importance of these properties for understanding
plot is grounded in narrative theory (apart from Ryan also compare [14,10]).
Future work is in preparation to empirically evaluate the viability of the generated
summaries for framing, as well the influence of this framing on a systems perceived
creativity. Furthermore, a complete implementation of the tellability measure will be
performed in order to serve as quality measure in an Engagement-Reflection type [18]
creative cycle. The suitability of this measure for guiding the generative process can be
then empirically evaluated by comparing the resulting stories’ quality, and will hopefully
shed some light on Ryan’s take on this elusive phenomenon.
6. Acknowledgment.
The second author is grateful for support for this work provided by an Alexander von
Humboldt Ph.D. fellowship.
References
[1] Berov, L.: Steering plot through personality and affect: An extended BDI model of fictional characters.
In: German Conference on Artificial Intelligence. pp. 293–299 (2017)
[2] Berov, L.: Towards a computational measure of plot tellability. In: Proceedings of the 10th INT. pp.
169–175 (2017)
[3] Berov, L., Kühnberger, K.U.: An evaluation of perceived personality in fictional characters generated by
affective simulation. In: Proceedings of the 9th ICCC. pp. 24–31 (2018)
[4] Charnley, J.W., Pease, A., Colton, S.: On the notion of framing in computational creativity. In: Proceed-
ings of the Third ICCC. pp. 77–81 (2012)
[5] Cheong, Y.G., Young, M.: Suspenser: A story generation system for suspense. IEEE Trans. Comput.
Intell. AI Games 7(1), 39–52 (2015)
[6] Colton, S.: Creativity versus the perception of creativity in computational systems. In: AAAI Spring
Symposium: Creative Intelligent Systems. pp. 14–20 (2008)
[7] Colton, S., Charnley, J.W., Pease, A.: Computational creativity theory: The FACE and IDEA descriptive
models. In: Proceedings of the Second ICCC. pp. 90–95 (2011)
[8] Colton, S., Wiggins, G.A.: Computational creativity: the final frontier? In: Proceedings of the 20th
ECAI. vol. 12, pp. 21–26 (2012)
[9] Cordella, L.P., Foggia, P., Sansone, C., Vento, M.: A (sub) graph isomorphism algorithm for matching
large graphs. IEEE Trans. Pat. Anal. Mach. Intell. 26(10), 1367–1372 (2004)
[10] Dannenberg, H.: Coincidence and Counterfactuality. University of Nebraska Press (2008)
[11] Gambhir, M., Gupta, V.: Recent automatic text summarization techniques: a survey. Artif. Intell. Rev.
47(1), 1–66 (2017). https://doi.org/10.1007/s10462-016-9475-9
[12] Goyal, A., Riloff, E., others: A computational model for plot units. Comput. Intell. 29(3), 466–488
(2013)
[13] Lehnert, W.G.: Plot units and narrative summarization. Cog. Sci. 5(4), 293–331 (1981)
[14] Palmer, A.: Fictional Minds. University of Nebraska Press (2004)
[15] Pérez y Pérez, R., Sharples, M.: Mexica: A computer model of a cognitive account of creative writing.
J. Exp. Theor. Artif. Intell. 13(2), 119–139 (2001)
[16] Ryan, M.L.: Possible Worlds, Artificial Intelligence, and Narrative Theory. Indiana University Press
(1991)
[17] See, A., Liu, P.J., Manning, C.D.: Get to the point: Summarization with pointer-generator networks.
ArXiv e-prints (2017), https://arxiv.org/abs/1704.04368
[18] Sharples, M.: How we Write: Writing as Creative Design. Psychology Press (1999)
[19] Turner, S.R.: Minstrel: a computer model of creativity and storytelling. Ph.D. thesis, University of Cali-
fornia at Los Angeles, Los Angeles, CA (1993)
7. Appendix
Figure 4. The connectivity graph of the original version of “The Little Red Hen”. NG is a nested goal unit,
DR a denied request unit, Ret a retaliation unit and Mot and Ena are the primitive units of motivation and
enablement. The directed edges indicate temporality. The unit with the bold label has the most neighbors in the
graph.