The present paper proposes a common representation model that allows the free exchange of knowledge between di erent story generation systems as a base for a collaborative environment to run an enhanced process of literary creation. In addition to this objective, this model aims at the development of a story representation formalism for creating a common knowledge base that can be fed in the future with the outcomes of new storytelling systems, without the need to adapt it to every system-speci c representation model.
Automatic story generation is a part of a wider research area in Arti cial
Intelligence named Computational Creativity (CC), which is the pursuit of creative
behaviour in machines [
Story generation systems are faced with a signi cant challenge of acquiring
knowledge resources in the particular representation formats that they use. They
face an inherent di culty when using formal languages in the detachment
between the formulation of the needs in the real world and its representation in
a formal construction. These di culties are increasingly greater when the
systems participate in a co-creation model with humans [
The present paper introduces a model for expressing knowledge related to the domain of narrative. The goal of this formalism is to ease the interchange of information between di erent story generation systems operating in a collaborative ecosystem. Its purpose is to establish a co-creation cycle inside an architectural model that allows the involved systems to take advantage of the shared knowledge model and use it for enhancing the quality of the generated texts.
This section reviews the way story generation systems represent the knowledge they need to create stories. The existence of commonalities across the existing systems is necessary to establish a shared model for knowledge representation.
TALE-SPIN [
Author [
Mexica [
Thespian [
Fabulist [
STellA (Story Telling Algorithm) [
PropperWryter [
Charade [
In addition to these single storytelling systems, Slant [
Slant represents the story in a shared data structure, according to the blackboard architectural pattern. This resource allows the participating storytelling systems and components to create stories collaboratively. The goal is to allow the systems to in uence each other for generating an enhanced result. The blackboard architecture and the Slant story XML format that is used, open up new possibilities for collaboration between creative literary systems, allowing models of creativity to be developed and added in di erent con gurations.
The process for generating stories in Slant begins with minimal, partial
proposals from a simple unit, named the Seeder [
To enable a better understanding of the problem, let us consider how to develop
an automatic story generation system that could generate a story like \War and
peace" [
In the light of the above, creating such a system would be a formidable challenge. Considering the capabilities of the existing storytelling systems, the target generator should be able to operate at di erent levels of detail and resolve many interwoven subplots. For example, when generating the narration of the war, the system should create a detailed description of the movements carried out by the Napoleonic army on Russian territory, and the corresponding response of the Russian generals. To do this, it would need to focus on strategy, with a discourse aimed at describing the movements of troops, the relationships between the actions of the Napoleonic army and the countermeasures of the Russian army, the activity of the convoys of military supplies, the terrain di culties, and nally, a detailed description of the battles fought during the campaign. In such type of narration there is no room for long dialogues, nor for focusing on individual thoughts or feelings. Narrative resources are aimed at representing the complete picture of events. The detail is used in the description of the facts and the scenarios in which they take place.
On the other hand, to develop the story of the relationships between the characters that take place during the war, it would require the system to be able to simulate the various characters involved, their feelings, their intentions, their social norms and their actions. The interaction between the characters would lead to changes in mutual perception and feelings. To convey all this evolution, it would be necessary to resort to witty dialogues, in which the characters reveal their feelings and intentions, or to minutely narrated scenes that show how they behave. Of course, a believable love story can only be achieved by employing a good amount of knowledge about every single character. This knowledge includes a complete psychological portrait, which allows a deep understanding of its motivations, a proper representation of the historical context, and many other aspects related to the characters' consistency.
The above examples do not exhaust the list of requirements. In a novel like \War and peace" there are multiple aspects that would require of a speci c processing in a storytelling system. To mention some of them: the thematic aspects require a simulation model centred on the intentions of the author; the description of a duel requires a simulator of physical world; and a philosophical re ection about the status of the servants requires a reasoning engine that includes knowledge concerning the social context in nineteenth-century Russia.
It seems quite di cult to generate such a novel with a single storytelling system. The prior example just tries to pose a question about the complexity of working at di erent levels of knowledge representation. The simplest way of creating rich and complex stories is using di erent systems, generating di erent types of content according to their capabilities.
The existence of di erent story generators, each with a di erent approach not only to how to identify the next action to continue an existing draft, but also to how to evaluate the quality of a partial draft requires the de nition of a speci c interaction protocol to guide the way they collaborate among them to create a single consensus draft. The context of the interaction would be a set of story drafts that are being developed over a sequence of interactions. As such, every draft contains a fusion of the information provided by the various generators, represented in the shared format. 4
From a methodological point of view, this paper focuses on identifying the
dimensions considered by the knowledge managed by storytelling systems. The
strategy being followed was originally outlined in a previous paper [
The representation model is strongly in uenced by the components of
narrative identi ed in the classic Narratology [
In order to clarify the developed model, it is necessary to set out the main
concepts that come into play. The following de nitions are based on those
established by the traditional literature [
Narrative is considered as a story articulated in a discourse. It is a complex concept, which involves several components itself.
Story is the content of the narration; it includes what happens, namely, the plot, and the space in which the action occurs.
Space includes the characters, settings, props and anything which is present either physically or abstractly in the space of the narrative |that is to say, the existents. Because existents change and evolve over time, the space also consists of an initial state which contains the set of all existents' states as they exist before the start of the plot.
Plot is the basic structure of any story. A plot is a set of events with an overall structure which represents both the temporal ordering, and the causal relations between the events. Events typically consist of one or more low-level actions, instigated by and/or a ecting a number of entities in the space. The concept of plot encompasses the ways in which the events and characters' actions in a story are arranged and how this arrangement in turn facilitates identi cation of their motivations and consequences.
Discourse is the particular mean of telling a story. This may include several aspects such as the narrator's perspective, the ordering and duration of the events in the plot, etc.
Apart from these high level concepts, that constitute the roots of the model,
there are other relevant concepts that have been considered due to its importance
in this representation model. The representation of the characters is one of them.
According to Jannidis [
The proposed representation model considers these three levels, covering totally the two rst, and partially the third one.
Another key concept is the narrative space, that goes further than the
representation of a world in terms of a simple container for existents and a location
for events. Resorting again to narratology, we have considered ve levels [
If we analyse in detail the existing storytelling systems, two conclusions can be
reached regarding knowledge representation [
plot features behaviour features relationships features cognition function cultural physical cultural physical cultural physical
The model has been designed in a hierarchical structure, in which the root concept is the story. A story represents what both intuitively and narratologically can be considered a story, that is, a narration of events happening in a setting. It is composed by the two classic narratological components: the plot and the space.
The plot is represented as a sequence of scenes. A scene is conceptually related to the division of a play, that represents a single episode inside the plot. It is clearly conditioned by the time division, which means that is a sequence of events that happen during a time frame. From a spatial point of view, it is also constrained to take place in a single spatial frame |considering the spatial frame de nition mentioned before. So, the scene is composed by a sequence of events, that can be actions or happenings. An action is an act performed by one or more characters in the story, generating consequences. The resulting consequences of every action are expressed as a modi cation in the global state of the space |considering it as the whole setting and the existents. A happening is an event that happens in the plot, as an accident or as a consequence of a prior action or happening. A happening can be natural |it rains| or arti cial |a car accident.
The space encompasses the whole universe in which the plot is developing.
It is composed by the setting and the existents. The existents are the whole
set of actors that take a part in the story. They can be characters, living beings
|an animal|, and an object in the setting. The two last types are mainly
de ned by their physical features and their cultural signi cance in the story.
The characters are the most relevant, and also the most complex to represent,
elements in the story. The proposed model considers not only their physical,
psychological and social features, but also their cognitive-related characteristics.
The cognition of the characters is represented in a very detailed manner due
to its importance for ensuring story consistency and characters liability. The
aspects considered have been chosen after analysing those used by the existing
storytelling systems [
They refer to facts about the world which the character considers as axioms, regardless of they are true. They can be part of the character's cultural or religious code, or simply originate in a particular misconception of the world. { Dreams: The dreams represent the unconscious aspirations of the character.
He/she may not be aware of them, but they can operate at a subconscious level and inspire his/her intentions. { Fantasies: The fantasies are product of characters' imagination. They are beliefs or notions based on no solid foundation, a fact which the character is perfectly aware of. They represent aspirations that the character considers unreachable, but he/she enjoys thinking about them. { Emotions: The emotions are related to the feelings of the character. They are usually in uenced by the relationships that the character establishes with the others, and the evolution of them during the story.
Another relevant element of character's representation is the function.
Although many approaches treat characters as entities subject to speci c rules that
interact in a simulated story world, there is another important line of thought in
the treatment of characters: the functional view. In this perspective, pioneered
by Aristotle [
The setting is a combination of a set of physical |or virtual| locations in which the action of the story takes place, and the set of cultural and physical rules that govern the story world. The locations can be considered the scenario in which every scene that composes the plot takes place. So, as shown in the model, every scene links to its corresponding location. 6
The model presented before aims at providing a reliable way of interchanging knowledge between di erent story generation systems, each of them using its own inner representation model.
The application of this model must necessarily rely on an architecture which
combines di erent storytelling systems. For that purpose, it has been proposed
a service-oriented architecture[
Every service will understand and generate messages for communicating with the others. The inner story generation processes of every system are unlikely to be interchangeable, but not the nal product: the story. To this end, the proposed representation model is centred on the story, that is the common element among the various systems. Therefore, by means of the common model it is possible to establish an iterative cycle of story generation, in which every service contributes to the story by generating a part of it. This way, the ability of PropperWryter is to develop the general scheme of the story, while STellA can provide a detailed simulation of the di erent scenes, and Charade describes the evolution of the relationships between the characters. All the required information for achieving this can be represented by means of the common model. 7
The model presented along this paper is based on the analysis of the knowledge required by existing storytelling systems. It tries to establish a baseline for make easier the interchange of knowledge between di erent story generation systems, and make possible the generation of richer and higher-quality stories by combining the capabilities of several systems.
In order to avoid to a ect the internal story generation process of every participating system, the proposed model focuses on the product, that is the story. This approach seems to be the most convenient for keeping the common representation agnostic from the speci cs of each system.
This knowledge representation model requires to be complemented by a architectural model which allow the development of a collaborative generation process involving various storytelling systems. Future work involves the development of a test eld for this model, which allow to re ne it and overcome its aws. Currently, the authors are working in the target architecture, which involves the development of a storytelling ecosystem based on the referred systems |STellA, PropperWryter and Charade| as a benchmark for applying the common representation model.
Some of the essential points that will need to be evaluated are those related to knowledge consistency between systems, the possible degradation of the semantics, and the loss of relevant information during the communication between the participating systems. We considered that the future work on the proposed storytelling ecosystem will allow a complete study of these concerns, and also the improvement of the model to guarantee that the nal result covers all the speci ed objectives.
This paper has been partially funded by the project IDiLyCo: Digital Inclusion, Language and Communication, Grant. No. TIN2015-66655-R (MINECO/FEDER).