Simulating the Everyday Creativity of Readers
Brian O’Neill and Mark Riedl
School of Interactive Computing
Georgia Institute of Technology
{boneill, riedl}@cc.gatech.edu
Abstract
Sense-making is an act of everyday creativity. Research
suggests that comprehending the world is an act of
story construction. Story comprehension, the process of
modeling the world of a fictional narrative, thus involves
creative story construction ability. In this paper,
we present an intelligent system that “reads” a story and
incrementally builds a model of the story world. Based
on psychological theories of story comprehension, our
system computationally simulates the everyday creative
process of a human reader with a combination of story
generation search and strategies for inferring character
goals. We describe the work in the context of a Synthetic
Audience – a system that assists amateur storywriters
by reading the story, analyzing the resultant
sense-making models, and providing critique.
Introduction
Humans exhibit creativity in a vast range of domains such
as music, art, dance, and storytelling. On a regular basis,
humans exhibit creativity in ways that are overlooked:
problem-solving, inference, and, in general, sense-making
are all creative acts that humans carry out on a regular, and
sometimes unconscious, basis. Sense-making is an act of
human cognitive creativity: the construction of a narrative
that explains what is happening around us (Bruner 1990;
Gerrig 1994). We call this everyday creativity. Consider
the following example:
It’s nighttime; Jesse is standing above Marlow, a gun to
his head. The trigger slowly squeezes… The next morning,
we see William, Marlow’s brother, digging a shallow
grave. He drops a body into the hole in the ground.
It’s Jesse’s lifeless body…
What happened that night? Who are these characters and
what were their goals? These are questions that arise in the
mind of the reader. A reader must effectively reconstruct
the narrative, and infer concepts and events not explicitly
read, causal relationships between events, and the goals
and motivations of characters (Graesser et al 1994).
Boden (2009) suggests that not all creativity is high art.
If this everyday creativity could be computationally harnessed
what could a system do? Systems could employ
human-analogous sense-making processes in order to build
a model of the world – the real world or a fictional world
observed in a book or movie. This model can be used to
simulate human responses to stimuli. With respect to everyday
creativity in story comprehension, we could “read”
stories or “watch” movies and produce cognitive and affective
responses equivalent to those of a human audience.
Those responses can, in turn be used to provide feedback
to amateur human creators that need assistance with storytelling
ability by simulating the responses of a human
reader/viewer receiving it for the first time. With that final
concept in mind, this paper describes initial steps towards a
“synthetic audience.” A synthetic audience aims to assist
creators by modeling the cognitive processes of recipients
of a creative artifact and providing feedback. Feedback
from a synthetic audience could be given to the human
creator faster and more frequently than feedback from another
human source. The synthetic audience must have
sufficiently robust ability in everyday creativity. For the
purposes of the synthetic audience, we computationally
model human creative sense-making processes in the context
of story comprehension, focusing on the ability of a
system to reconstruct a narrative from the events it reads.
 Readers/viewers (we will use the media-agnostic term
“audience”) are actively engaged in cognition when reading/viewing
a narrative (Gerrig 1993). The audience engages
in problem solving from the perspective of story
world characters, attempts to resolve (intentionally or unintentionally
placed) gaps in the narrative (called ellipses),
and forecasts future events. The inference processes applied
by the audience provide an explanation for what has
been observed, but unlike conventional problem-solving,
the results of these processes cannot be declared right or
wrong. These inference processes can be exploited by
authors to enable many of the more interesting cognitive
phenomena of storytelling: ellipses, suspense, surprise, and
genre expectations, among others.
In this paper, we present a component of the synthetic
audience: a model builder that constructs a possible mental
model for a human reader. The model builder “reads” the
story as it is authored by the human creator. After each
read event, it builds or revises its mental model by hypothesizing
the goals of the story characters using a number
of strategies, and reconstructs the story using a narrative
generation planner. The model is then a source of
feedback for the larger synthetic audience system. Different
model structures are indicative of possible comprehenProceedings
of the Second International Conference on Computational Creativity 153
sion issues, such as changing character goals, diverging
storylines, or unmotivated actions by the characters.
In the remainder of this paper, we discuss related work
in the psychology of narrative comprehension, story understanding,
story generation, and creativity support. We then
describe our model builder, in the context of a synthetic
audience. Finally, we will show that our model builder
produces a cognitively plausible mental model of the storyso-far,
improving on approaches that do not model human
creative processes.
Background
Reader Inference
While reading a story, readers continuously make inferences
about aspects of the story that have not been explicitly
stated in order to make sense of the narrative (Graesser
et al. 1994). Some inferences can be made with little effort
while reading, while other types of inference occur only
when the audience has been given time to reason. The former
group, described as online inferences, includes:
• Superordinate goals – Inferring the overarching goal
motivating a particular character’s actions.
• Causal antecedents – Inferring the causal relationship
between the current action and information that appeared
previously in the text.
Conversely, offline inferences, those made when the audience
is afforded time to reason, are as follows:
• Subordinate goals – Inferring the lesser goals or plan of
action used to achieve the current event or state.
• Causal consequences – Inferring the effects of the current
action.
In particular, the online processes drive the creative search
for a narrative explanation of observed events. That is, the
inference of a character’s superordinate goal is a projection
of that character’s actions into the future, resulting in the
construction of narrative structure explaining why a character
has performed observed actions. Likewise, inference of
causal antecedents results in the construction of narrative
structure that fills in the gaps between observed events in
order to explain how particular events came to pass.
How does one represent a mental model of a story?
Graesser and Franklin (1990) developed QUEST to model
the human question-answering process as a theory of
sense-making. QUEST was demonstrated in the context of
story comprehension (Graesser et al. 1991). Stories are
represented as directed graphs, where nodes represent story
events, character goals, and world states. Edges represent
relationships such as causality or the formation of goals.
Traversing the arcs in a QUEST diagram allows one to
answer questions, such as what enabled an event to come
to pass, or why a character performed an action. Chains of
causally-linked goals and events are called goal hierarchies,
in which the last goal in the chain is the superordinate
goal, the motivating goal for the entire sequence.
Figure 1 shows a QUEST structure for a story. Event
nodes E3 and E4, and goal nodes G3 and G4 make up a goal
hierarchy. The superordinate goal, node G4, is Initiated (I)
by the state node E2. That is, because of E2 the hero has the
indicated goal. Causality is shown in QUEST by Consequence
(C) arcs between event nodes. In the example, E5
occurred as a consequence of both E3 and E4. Goal hierarchies
are formed by chains of Reason (R) arcs, indicating
subordinate/superordinate relationships between goals. In
the example, Goal G3 is subordinate to Goal G4.
Related Work
Story Understanding. Sense-making is an example of
everyday creativity, a process that humans use on a daily
basis to explain the world around them. Sense-making in
the context of stories shares many similarities to story understanding,
a process by which a computational system
extracts knowledge from a narrative text. Mueller (2002)
summarizes many of the approaches taken in story understanding,
including the application of known scripts and
plans, the inclusion of plot units, and connectionist approaches.
Other approaches include generating questions
while reading and attempting to answer them with subsequent
text (Ram 1994). In general, story understanding
systems extract knowledge from complete texts, whereas
we infer through abduction events and character goals in
incomplete stories in order to assist amateurs.
Creativity Support. There are two general approaches to
computational creativity support. The first are tools that
assist creators by providing an appropriate interface for
complicated creation processes without using AI (e.g.,
Skorupski et al. 2007). The second approach leverages AI
to form a team made of the human creator and the computational
tool. These mixed-initiative approaches (e.g., Si et
al. 2008) lead to artifacts that have equal contributions
from both the human and the AI. With the synthetic audience
agent, we aim to leverage AI, as mixed-initiative tools
do, without having explicit involvement in the creation of
the product. In our approach, which we describe as “computer-as-audience”
(Riedl and O’Neill 2009), the agent
provides feedback to a human creator from the perspective
of the recipient of a creative artifact.
Story Generation. Computational approaches to narrativegeneration
typically address the problem of creating content
as either a search problem or an adaptation problem.
See Gervás (2009) for a history of story generation research.
Our system uses elements of both of these apFigure
1. A QUEST structure for a story. Nodes represent
states (S), character goals (G), and events (E). 1: Villain wants
to be powerful. 2: Villain coerces Hero into agreeing to help.
3: Hero robs a bank. 4: Hero gives money to Villain. 5: Villain
bribes the president.
S1 G2,vil
E2,vil
G3,h
E3,h
G4,h
E4,h
G5,vil
E5,vil
I
I
C C
R
R
Proceedings of the Second International Conference on Computational Creativity 154
proaches, incorporating a case-based reasoner to infer
character goals and a search-based story generator to construct
narratives that explain the inferred goals.
We utilize the IPOCL narrative generation algorithm
(Riedl and Young 2010), a refinement search approach to
constructing novel narrative structures that are both causally
coherent and believable, as part of our system. IPOCL
requires character actions to be justified both causally and
by motivations and intentions. Specifically, it utilizes special
data structures called frames of commitment to enforce
the constraint that all events must be motivated by some
preceding event. That is, every frame, representing a character
goal, must be caused by some event as means of explaining
why that character has the goal in question. As a
refinement search process, IPOCL works backwards from
a goal state, using causal and intentional requirements to
guide the selection and instantiation of new events. Figure
2 shows the frames and events of an IPOCL plan, describing
the same story as the QUEST diagram in Figure 1.
Events are rectangles, while rounded boxes represent
frames of commitment. Solid lines between events indicate
a causal relationship. Dashed lines indicate that the event
was carried out in service of that frame of commitment –
as part of the character’s attempt at achieving its goal.
IPOCL shares similarities to the above psychological
theories of narrative comprehension. IPOCL narrative
plans can be converted to QUEST structures, and vice
versa. Christian and Young (2004) present an algorithm for
converting partial-order plans into QUEST structures. This
algorithm has been updated to IPOCL plans (Riedl and
Young 2010), specifically translating frames of commitment
into goal hierarchies.
Synthetic Audience
The goal of a synthetic audience is to provide an amateur
storywriter with feedback from the perspective of a recipient.
That is, the synthetic audience “reads” the story as it is
being written and produces cognitive and emotive responses
based on theories of human story comprehension.
A synthetic audience is able to provide feedback faster, and
with greater frequency, than a human critic. In order to
provide such feedback, it is necessary to model human
responses to creative artifacts. When working with stories
as they are being written, the system must derive a mental
model of the story-in-progress based solely on what has
been authored. The synthetic audience has to make inferences
about events that are missing from the story, the
causal relationships between events, and character goals.
These inferences are comparable to human gap-filling and
sense-making processes, both of which are carried out during
reading comprehension. Thus, the synthetic audience
system derives a mental model of the story as it is written,
based on recognized human creative processes.
 A storywriter using the synthetic audience authors states
– facts and descriptions – and events by selecting event
templates from a list of options. These templates allow the
author to fill in the specifics of the state or event, such as
people, locations, or objectives. Authors can add states or
events at any point in the story, regardless of chronology.
The synthetic audience continually re-reads the story as it
is authored, constructing a model from the audience’s perspective
and uses that model to provide feedback. When
there is feedback from the synthetic audience, it is displayed
to the author in a non-intrusive manner; the author
may respond to the feedback or ignore it.
Knowledge Representation
The synthetic audience requires knowledge about the semantics
of events in order to make inferences. That is, the
synthetic audience requires a domain theory – a description
of how the story can change. We use a domain theory
comprised of STRIPS-like event templates that provides
information about preconditions and effects of any event
the human authors. This is the same representation used by
IPOCL. Because we employ a narrative planner, we also
require every story to have a precondition. The user declares
the facts of the initial state, as expository states
where it makes sense to do so. If the initial state is incomplete,
the Synthetic Audience uses a special operator, Assert,
that causes facts to be inferred as true in the initial
state, as a last resort to avoid failure. The use of Assert
operator to modify the initial state is equivalent to the
technique described by Riedl and Young (2006).
Model Builder Algorithm
The core of the synthetic audience is the Model Builder.
The purpose of the Model Builder is to construct a possible
mental model of a human audience, revising this mental
model as each event is “read” in chronological order. The
model is constructed by hypothesizing the goals of each
character in the story, and constructing a narrative that explains
what has been read. Once the story is read, the
model is used to generate feedback in the form of critique.
The Model Builder starts by reading the newest event. If
the newest event is not at the end of the story, the Model
Builder rewinds the construction process to the latest point
before the newly authored event and processes the remainder
of the story as if encountered for the first time.
Figure 2. An IPOCL narrative plan corresponding to the QUEST structure from Figure 1.
Coerce (vil, hero, has(a, money))
Give (hero, vil, money)
Bribe (vil, prez)
Initial state
Villain intends to be powerful
Hero intends that Villain has money
Rob-Bank (hero)
Goal situation:
corrupt (prez)
Proceedings of the Second International Conference on Computational Creativity 155
The search for superordinate character goals drives the
Model Building process because of their importance in
story comprehension and sense-making (Graesser et al.
1994). The Model Builder uses four strategies to hypothesize
the superordinate goal for each character that is actively
engaged in the current event. When characters are
not actively engaged, previously inferred goals for those
characters are retained from earlier iterations.
Character Goal Inference Strategies. The model builder
uses the following four strategies, in the order given, to
infer goals for characters actively engaged in the event.
1. Declared Goals (D). The Declared Goals strategy hypothesizes
goals that are explicitly declared in the new
event for other characters. For example, if a character
states its intention, then that goal is accepted at face
value. Likewise, if a character instructs a subordinate
character to do something, that goal is accepted for the
latter character.
2. Existing Goals (E). The Existing Goals strategy tracks
goals that were hypothesized at an earlier point and remain
unresolved based on the authored story. This strategy
merely tries to place the new event into the hypothesized
mental model from the previous iteration.
3. Proposed Goals (P). The Proposed Goals strategy uses
a case-based goal recognizer to infer character goals,
based on that character’s existing goal hierarchies. The
recognizer is given a QUEST model containing only the
acting character’s goal hierarchy, contextual state nodes,
and the most recently added event. The recognizer
searches its case library for a QUEST model of a story
with a similar chain of events as those in the given event
and hierarchy. For the best match, the recognizer returns
the goal at the top of the relevant goal hierarchy.
4. Top-of-Hierarchy (T). The final strategy is the Top-ofHierarchy
strategy, which assumes that the most recently
authored event is the goal of the characters involved.
Top-of-Hierarchy is a last-resort strategy that is
tantamount to “wait and see what happens next.” The
name of the strategy refers to the notion that the new
event is the top of a QUEST goal hierarchy.
When more than one character is actively engaged in an
event, the goal inference process is applied to each character,
one at a time, in arbitrary order. The hypothesized
goals and authored events are given to the IPOCL planner
in order to test the goal hypothesis by generating a narrative
sequence that explains the goals.
Testing the Hypothesis. Once the model builder has identified
the goals of the characters, it tests the hypothesis by
generating a narrative that links together all authored
events to the hypothesized goals of the characters. This is
achieved as follows. First, an instance of an IPOCL plan is
created by instantiating every authored event in the
QUEST model. Narratives generated during the prior iteration
may have events that were generated but not written
by the human author; these events are discarded. Temporal
constraints enforce chronological ordering of authored
events. The model builder constructs a goal situation consisting
of newly hypothesized goals for the current character
as well as un-realized goals for other characters carried
over from prior iterations. Additionally, a frame of commitment
is created for each un-realized hypothesized character
goal across all characters.
A modified IPOCL planner is instructed to satisfy the
preconditions for each event and each proposition of the
goal situation, and to find a motivating event for each
frame. As a refinement search algorithm, IPOCL takes a
plan in any stage of completeness, finds a flaw – a reason
why the plan is not complete – and resolves it. In the process,
other flaws may be created, resulting in an iterative
process. In this case, unresolved preconditions are solved
by instantiating a new event, reusing an existing event, or
by the initial world state. If IPOCL fails to find a plan, or if
a plan is found that does not link all events in causal chains
terminating with character goals, the current hypothesis is
rejected and the model builder tries the next strategy.
We modified the IPOCL algorithm as follows. First, we
bound the search depth to approximate cognitive limitations.
Second, we provide a heuristic that strongly prefers
to reuse authored events. Third, we add the Assert operator
described above to declare unstated facts to be part of the
initial state. Finally, we provide a special event, Decide,
which has no preconditions and has the effect of giving a
character an intention. Decide is equivalent to the system
admitting that it does not know why a character performed
an action without failing the search. The Model Builder
heuristic highly penalizes the inclusion of Decide events in
the narrative, thus relegating its use to a last resort.
When the hypothesis is accepted, the plan generated by
IPOCL is converted to a QUEST model using the previously
mentioned IPOCL-to-QUEST algorithm.
Characters with Multiple Goals. It is possible for characters
to pursue multiple goals throughout the course of a
story. If goals hypothesized by the model builder using the
P or T strategies are not accepted, then the Model Builder
attempts to create a new goal hierarchy with any events
that were not linked. The hypothesis is retested using the
above technique.
Synthetic Audience Feedback
The synthetic audience generates feedback based on the
QUEST model that resulted from hypothesis testing. Various
QUEST structures indicate potential comprehension
problems including:
• Diverging storylines – Implied by disjoint causal chains.
• Unmotivated goals – Indicated by the need to use Decide
events to construct missing Initiates arcs.
• Unexplained events or motivations – Indicated by the
use of Assert operators to modify the initial state because
missing information must be inferred.
• Sudden shifts in model – Indicated by a sudden change
in goals hypothesized by the Model Builder from the
reading of one event to the next.
If any potential mental model indicates possible reader
comprehension issues, then this is feedback that we would
aim to provide to the creator. The mental model conProceedings
of the Second International Conference on Computational Creativity 156
structed by the Model Builder is only one possible model.
However, this single model remains useful in the context
of a larger synthetic audience system, as it can be instantiated
with different domain theories and background
knowledge to explore a variety of audiences.
Example
Consider the following scenario, selected to illustrate the
goal inference strategies, in which a human author inputs a
story with gaps. The story involves Aladdin, Jasmine, a
genie, and a king. For purposes of illustration of the Model
Builder, we will assume that the author has declared the
characters up front, and provided various additional facts
about the story world. The facts include: the genie is
trapped in a magic lamp; a dragon possesses the lamp; and
the king hates and fears the genie. The author writes the
first few events:
1. The King orders Aladdin to retrieve the magic lamp.
2. Aladdin travels from the palace to the mountains.
3. Aladdin gives the magic lamp to the King.
The Synthetic Audience “reads” the events in order. The
first event, Order, involves multiple characters. It arbitrarily
chooses to attempt to infer Aladdin’s goals first. Using
the Declared Goal (D) strategy, and based on the semantics
of the order event, it hypothesizes that the orderee – Aladdin
– will adopt the given goal: that the King should have
the magic lamp. Next the Model Builder processes the
King. Strategies D and E are not applicable as the event
does not declare a goal for the King, nor is there a prior
hypothesis about his goal. Invoking the Proposed Goal (P)
strategy, the case-based goal recognizer hypothesizes that
the King’s goal could be to kill the Genie. This is because
our case-base includes a story in which one character hires
another to kill someone he hates. Hypothesis testing produces
a plausible narrative in which Aladdin slays the
dragon, takes the lamp, and gives it to the King who destroys
it, killing the Genie. The resultant narrative is converted
to a QUEST model and stored for later reference.
The Model Builder processes the second and third
events involving Aladdin traveling to the mountains and
then giving the lamp to the King. In both cases, the E strategy
verifies that this is consistent with the previously hypothesized
goal for Aladdin. The Model Builder infers that
the dragon lives in the mountains. In each case, the King’s
goal is retained from the first iteration because he is not an
active character in either of the second or third events. Figure
3(a) shows the QUEST structure of the model constructed
after the first three events.
Now suppose that the author adds one last event:
4. The King commands the Genie to make Jasmine love
him.
The Model Builder arbitrarily decides to process the King
first. The D strategy does not apply. The model builder
attempts the (E) strategy, re-using the King’s goal of killing
the Genie. However, it cannot find any plausible narrative
in which the King commands the Genie as part of a
goal hierarchy resulting in the Genie’s death. Therefore,
the (E) strategy fails. The model builder again tries the (P)
strategy. The case-based goal recognizer hypothesizes that
the King’s goal could be to marry Jasmine, replacing the
earlier hypothesized goal. The model builder processes the
Genie’s involvement in event 4, and using the (D) strategy,
determines that the Genie will adopt his given goal: to
make Jasmine love the King. Hypothesis testing produces a
narrative in which the King falls in love with Jasmine, and
sends Aladdin to retrieve the lamp. The Genie, under the
influence of the King, casts a love spell on Jasmine. Finally,
Jasmine and the King get married. Figure 3(b) shows
the QUEST model for the new model.
Discussion
The Synthetic Audience is a cognitively plausible process
for computationally constructing a model of the story-sofar.
The system employs the everyday creativity of inference
and future prediction. Graesser et al. (1994) are vague
about the exact inference process used by human readers,
proposing spreading activation; we assert that IPOCL,
which reasons over representations that are analogous to
QUEST structures, is a plausible substitution.
For any domain in which there are gaps in the events
(a) Model after three events.
(b) Model after four events.
Figure 3. QUEST models of the example story. Numbers correspond
to numbered events in the text. Nodes with dashed
lines were inferred during narrative reconstruction.
S0
G1,K E1,K
G2,A
G3,A
E2,A
E3,A
Gdead,K Edead,K
C
C
C
I
I
R
R
R
Aladdin's
inferred quest
G1,K E1,K
G2,A
G3,A
E2,A
E3,A
E4, K G4,K
C
C
C
I
R
R
R
Gmarry Emarry
Gspell,G Espell,G
C R
C
Elove,K
I
Aladdin's
inferred quest
Guse,K Euse,K
C R
I
Proceedings of the Second International Conference on Computational Creativity 157
that can be observed, it is potentially non-trivial to create a
well-formed sense-making model in which the relationships
between adjacent and non-adjacent events are found.
This is true of observations of the noisy world around us,
and also true of stories authored by amateur storywriters.
Searching for the connections is a creative act because a
complete narrative explanation is created as a by-product.
We believe that the approaches taken by the Model
Builder are applicable in many domains, so long as the
planner and case-based reasoner contain appropriate domain
knowledge. Our approach works with stories that
have (a) strong causal relationships (e.g., few non-sequiturs
or random events), and (b) highly goal-driven character
behavior – characters have a few top-level goals that are
motivated by prior world events and not arbitrarily
adopted. While not appropriate for all genres of story,
these properties are common enough in popular massconsumption
stories and video games.
Is it enough just to employ a story planner or other
search process to fill gaps? If one were to employ a story
generator such as IPOCL after each event read, one would
fill gaps between events; this is equivalent to the Model
Builder’s Top-of-Hierarchy strategy. However, such an
approach would miss opportunities. First, any such model
would not be representative of human models because the
inference of superordinate character goals is one of the
foremost online processes of an active reader. There are
some events that are rarely ever superordinate goals, and
thus rarely ever tops of goal hierarchies. Second, inference
of superordinate character goals is a form of future prediction.
By looking into the future and tracking back, one can
often find connections between seemingly disparate causal
chains; well-formed stories frequently tie plotlines together
and human readers expect it. Of course, how closely the
Model Builder matches human audience performance depends
on the case library.
Third, and most significantly, the Model Builder constructs
the sense-making model incrementally. One could
wait until the story is complete, in which case a naïve gapfiller
and the Model Builder would likely produce the same
result. By reading the story one event at a time and building
the model incrementally, the Synthetic Audience can
trace changes in the model over the course of the story,
thus providing feedback about surprises, suspense, and
other cognitive and emotive effects on audiences that result
from drastic revisions of the model. We conclude that the
Model Builder’s character goal inference strategies are
critical. The D and P strategies provide superordinate goals
that drive the creative explanation process. The E strategy
provides continuity. The T strategy is a “catch all” when
all the audience can do is wait and see.
The synthetic audience models human everyday creative
processes. The synthetic audience reconstructs the narrative,
making inferences about event causality and character
goals, the same kinds of inferences that human readers
make while reading a story. The synthetic audience performs
incrementally, revising the model as the story is
authored, rather than comprehending only complete stories,
like typical story understanding systems. This model of
human everyday creative processes can be applied to recipients
of creative artifacts, allowing feedback to be offered
to creators from the audience perspective.
<references_biblio/>
References
Boden, M. A. 2009. Computer models of creativity. AI
Magazine 30(3): 23–34. AAAI Press.
Bruner, J. 1990. Acts of Meaning. Cambridge, MA: Harvard
University Press.
Christian, D., and Young, R. 2004. Comparing cognitive
and computational models of narrative structure. In Proc.
of the 19th National Conference on Artificial Intelligence.
Gerrig, R. J. 1993. Experiencing Narrative Worlds: On the
Psychological Activities of Reading. Yale University Press.
Gerrig, R. J. 1994. Narrative thought?. Personality and
Social Psychology Bulletin 20(6): 712–715.
Gervás, P. 2009. Computational approaches to storytelling
and creativity. AI Magazine 30(3): 49–62.
Graesser, A.; Lang, K. L.; and Roberts R. M. 1991. Question
answering in the context of stories. Journal of Experimental
Psychology: General 120(3): 254–277.
Graesser, A. C., and Franklin, S. P. 1990. QUEST: A cognitive
model of question answering. Discourse Processes
13: 279–303.
Graesser, A. C.; Singer, M.; and Trabasso, T. 1994. Constructing
inferences during narrative text comprehension.
Psychological Review 101(3): 371–395.
Mueller, E.T. 2002. Story understanding. In Encyclopedia
of Cognitive Science, London: MacMillan Reference.
Ram, A. 1994. AQUA: Questions that drive the explanation
process. In Schank, R.; Kass, A; and Riesbeck, C.,
eds., Inside Case-Based Explanation, Hillsdale, NJ: Lawrence
Erlbaum Associates. 207–261.
Riedl, M., and O’Neill, B. 2009. Computer as audience: A
strategy for artificial intelligence support of human creativity.
In Proc. of the CHI 2009 Workshop on Computational
Creativity Support.
Riedl, M.O. and Young, R.M. 2006. Story Planning as
Exploratory Creativity: Techniques for Expanding the Narrative
Search Space. New Generation Computing 24(3).
Riedl, M. O., and Young, R. M. 2010. Narrative planning:
Balancing plot and character. Journal of Artificial Intelligence
Research 39: 217–267.
Si, M., Marsella, S.C., and Riedl, M.O. 2008. Integrating
story-centric and character-centric processes for authoring
interactive drama. In Proc. of the 4th Conference on Artificial
Intelligence and Interactive Digital Entertainment.
Skorupski, J.; Jayapalan, L.; Marquez, S.; and Mateas, M.
2007. Wide ruled: A friendly interface to author-goal based
story generation. In Proc. of the 4th International Conference
on Virtual Storytelling.
Proceedings of the Second International Conference on Computational Creativity 158