Computational approaches to complex cultural objects such as narrative have tended to focus on the complete modelling problem, proposing high-level concepts that are manipulated to generate or analyse stories. For example, Tale-Spin [13] uses the characters' goals and Mexica [14] uses a tension curve to represent love, emotion and danger in order to drive the generation process. In this paper, we put forward a di erent view based on the idea of a fundamental element that di erentiates narrative from other forms of human cognition and that we call a narrative thread. Furthermore, this paper supports the position that splitting phenomena such as narrative into constituent parts and researching each one separately, thereby postponing research into their precise interaction, is a pertinent and fertile approach. Our narrative model focusses speci cally on suspense and relies on the assumption that the empathetic e ects of a given narrative can be isolated from its plot-level description. This assumption has enabled us to build and get feedback from a model that should increase our understanding of the di erent roles played by event order and empathy in story-telling. We implemented our model computationally for some simple textual narratives and obtained promising evaluation results (N=46) (see [5] and [6]).
If we have a pile of bricks, we know that if we put them together in a few standard ways, we can end up building a wall and maybe even a whole building. But what are the bricks of narrative? For sentence understanding, the bricks may be words or perhaps morphemes. For vision, there are edges and surfaces from which we derive the existence of objects. But, what is the right elemental object to capture the `storiness' of narratives?
In this paper, we explore the potential explanatory power of one answer to that question, and show that our particular narrative brick can at least be used to build a wall, and perhaps soon, a whole building. Of course, like all scienti c processes, the real world will give our model feedback on the accuracy of the analysis it proposes, pointing to faults and weaknesses and enabling new constraints and modi cations to be imposed.
Additionally, our focus for now will be on one ingredient which we believe
is constitutive of what makes a good story, namely: suspense. As discussed in
[
We frame this article by generalising the development steps leading to our model with the aim of encouraging similar `fundamental' formal and computational approaches aimed at understanding complex human cognition. The rest of this paper details these steps.
The rst step we follow is to look for something similar to a syntax/semantics distinction. Secondly, we collect the constraints on a potential fundamental element that could generate the syntax observed. We then create an element that potentially satis es them and with it, derive a variety of theoretical behaviours and e ects which can be tested and modi ed or refuted. Once the model achieves a degree of explanatory power for some very speci c human experiences, we can test it further and also use it generatively to create new cultural artefacts.
1.1
Before we can start our process, we must make the initial step of dividing up the
`stu ' we have into its basic parts. Stories and the real world have in common that
human cognition divides them up into events. [
In `Why anyone would read a story anyway', [
This distinction is, of course, analoguous to the syntax/semantics distinction prevalent in language understanding. To emphasise the parallels, we can say the following: syntax/plot is the collection of rules that govern how words/ events are assembled into meaningful sentences/narrative sequences, whilst semantics/empathy is concerned with the meaning/importance of the words/events themselves.
Of course, in stories, creating empathy is essentially linked to the concept of a character. In our view, perhaps 70% of the sentences of a novel are concerned primarily with creating and maintaining the basic empathetic link between the reader and the main characters. However, we postpone work on story semantics by assuming the existence of an `empathy module' that produces a certain output. Our focus will be on the narrative parsing of the story plot.
We next postulate the psychological existence and computational utility of a fundamental element with which we can capture all the narrative-speci c processing that is needed to generate and understand the `story syntax', or plot. To create such an element, we rst identify the available and derivable constraints such an element must satisfy.1 2.1
Here we combine research in psychology on the requirements for knowledge
structures used for sentence understanding (see for example [
{ The element must have a clear starting point and end-point.
This is derived from Brewer and Lichtenstein's psychological theory of narrative understanding which suggests that three major discourse structures account for the enjoyment of a large number of stories: surprise, curiosity and suspense. Their approach is based on the existence of Initiating Events (IE) and the corresponding Outcome Events (OE) that are triggered by them.
To produce suspense, the IE and OE must be ordered chronologically. In addition, `often additional discourse material is placed between the initiating event and the outcome event, to encourage the build-up of suspense' ([1, p. 17]).
{ The element must have a clear linear path to completion.
This is derived from the constraints on causal consequences grounded in
psychological research such as [
We interpret these conditions together as indicating that our element must consist of a linear series of events.
{ The element must be internally consistent: no event can contradict any other event in the same narrative thread.
These constraints are inspired by the Glaive project ([
{ The element must be interruptible.
Much as events in the same fundamental element should not be in con ict, events in di erent elements must have some potential for con ict in order to produce the uncertainty necessary for suspense. We therefore need to model information about the following type of interaction between events: if Event E occurs in a story, then Event F can no longer occur in this story. This is analogous to syntactic rules that exclude certain category sequences in sentence grammar. 2.2
We now use all these constraints to deduce the fundamental element that will form the basis of our model of narrative. We call this element a narrative thread.
Combining constraints from 2.1 and 2.1, we extend the simple IE-OE pair to obtain the following:
IE ! Event1 ! Event2 ! Event3 ! ::: ! OE (1)
This simple script-form has the necessary clear nal result or outcome in the storyworld that we need for suspense and, according to the constructionist model, is also habitually and easily generated during narrative comprehension.
Using the constraints in 2.1, we can allow our narrative threads to be built by combining both causal chains and intentional paths whilst checking for internal consistency.
Finally, to create the necessary interruptibility mentioned in 2.1, we postulate the existence of a set of disallowing pairs (E; F ) such that an event E in one narrative thread can disallow an event F in a di erent thread.
Here is an example of a narrative thread taken from [
Narrative threads thus model a reader's expectations about what might happen next in a given story. For a given storyworld, there will be multiple narrative threads in competition with each other. Their interaction models what goes on when a reader experiences stories in that storyworld.
threads are simply lists of events that are likely to follow each other. In general, threads are informed by a variety of inferential and associative mechanisms: scripts, models of story characters involving beliefs, goals and desires, principles of naive physics and even previous narrative experiences2.
Our claim is that it is not necessary to know exactly how di erent information sources were used to construct narrative threads for them to be used successfully to model narrative. The available storyworld information and inferences can always be translated into the linear structure that we call a narrative thread. We thus postulate a separation between the inferential sources that build narrative threads, from the structure of the ongoing suspense processes they trigger and maintain.
Checking for generality We must distinguish certain typical narrative
instantiations from the fundamental narrative processes that underpin them. This is
the token/type distinction. Thus, for example, the fairytale models developed in
[
Similarly, the structures used to model narrative should not be entangled with the speci c types of narrative for which they create narrative e ects. Just as with syntax, we should be able to create two narratives that have the same formal structure even though they exist in two completely di erent storyworlds.
We now create a theoretical sca old for the story-telling phenomena upon which we can formalise a variety of narrative phenomena. 3.1
Firstly, to model a given storyworld, we derive the set of narrative threads that model the causal and intentional links it contains or suggests. We also create a set of disallowing event-pairs that de ne the interactions between the threads3. 3.2
The next step is to model the input to the storyworld. In this case, a story is just one particular sequence of events chosen from all possible events, that is, from all the events in the narrative threads used to model the storyworld. Telling a story, under this account, is then simply to evoke one by one this sequence of events. 3.3
When an event E is told in the story, any narrative thread T that contains event E is activated. Intuitively, this means that the reader predicts that the events that follow E in thread T will be told in the story. As a story progresses, narrative threads are activated and de-activated. Also, some pairs of upcoming events in di erent active narrative threads may be in con ict with each other, that is, they may belong to the set of disallowing event-pairs.
Using our basic building block and this theoretical story-telling sca olding, we
can now derive some structurally re ned de nitions of a number of narrative
concepts. In our case, we make the claim that the following three fundamental
mechanisms occur in stories to create suspense and are usefully de ned as distinct
narrative phenomena:
3 For the moment, the narrative threads used in our computational model are derived
by hand. However, our narrative thread is very similar to the narrative schemas
developed by [
Translated into a context of goals and plans, completion-based suspense would correspond to the achievement of a goal. An example would be the captain of a football team walking slowly towards the tribune to receive and raise in the air the prize cup her team has just won. As the outcome event becomes ever more imminent the suspense increases; we can talk of completion imminence. There may be absolutely no expectation that the achievement of this goal will be interrupted and yet we experience a kind of suspense in such cases4.
Completion imminence is a function of the number of events in the thread that remain to be told before it is complete. Figure 1 shows a thread with a completion imminence of 4 events at this stage in the story. We have tagged these events to show how the cup-raising episode could appear. This occurs when at least two active narrative threads contain upcoming incompatible events of which only one can actually occur. Furthermore, a big di erence in story outcomes is expected depending on which event actually does occur. An example would be a chase between a policeman and a thief who is trying to cross a border. As the policeman gets closer and closer to catching the thief, the imminence of the catching event increases; we can talk of interruption imminence.
The con ict-based suspense of a thread A is related to the smallest number
of events still to be told in some other thread B before an event can be told
which interrupts A. To illustrate con ict-based suspense between two threads,
in Figure 2, we show thread A which has an interruption imminence of 3 events
due to thread B. We have tagged these events to show how the chase episode
described could appear.
4 Completion-based suspense may be linked to a phenomenon discovered by Lithuanian
psychologist Bluma Zeigarnik who observed the e ect of interruption on memory
processing. [
To model this, we imagine that in such situations a common event is present in two di erent narrative threads as illustrated in Figure 3. When the common event is told in the story, both threads become activated as candidates to explain the event's presence in the story. Subsequent story events may disallow one of the threads leaving one thread as the correct `explanation' of the common event. This epistemological gap- lling process is suspenseful in itself: we know that every time we nd out new information, it is likely to have a major e ect on the set of active narrative threads. Relative importance At any one moment in a given story, a variety of suspenseful situations may be present. Relatively unimportant suspenseful situations may coexist with life-or-death situations. The importance of these situations will also often depend on the reader's emotional involvement with them and this could be low or high and positively or negatively valenced.
In our model, we presuppose the existence of a `empathy module' that ascribes a relative emotional importance value to all the narrative threads. The value ascription of a narrative thread is related to the reader's appraisal of the state of the storyworld when the last event in the thread has been told. This use of a single value to encompass a multitude of factors enables us to take into account the modelling of emotions but at the same time keep our focus on the structure of the information ow and its relation to suspense.
Often, the importance of events is related to the fate of some story character. In this case, we can base the importance values on the i) the current level of sympathy (or antipathy) towards a character involved in an event, and ii) the perceived desirability (or undesirability) of the event in relation to that character. Foregroundedness If a given narrative thread T is not evoked for several story steps during the telling of a story, we assume its e ect on suspense will drop because it is less present in the reader's mind. At the same time, other narrative threads are of course active and competing for the reader's attention. Of course, as soon as a new event belonging to thread T is told, it comes again to the foreground and regains its full potential for suspense creation. Therefore, in addition to a measure of the relative importance of the narrative threads, we also need a measure for the degree of foregrounding of a thread.5 4.5
With a suitable heuristic, we can now combine our measures for the three types of suspense: completion-based, con ict-based and revelatory with the additional contextual measures of importance and foregroundedness to obtain values for the suspense contributions of all active threads. We then combine these values to create an overall suspensefulness value for each story-step. We now have the apparatus needed to predict a step-by-step suspense pro le for a given story.
To test this conceptual framework, we rst implemented our narrative model
computationally. We then derived storyworld-speci c information in the form
5 In our model, foregroundedness is roughly equivalent to recency of mention. Recency
has been extensively researched in the psychological eld as an important factor
in uencing memory (see for example [
Analysis of the results using a variety of statistical tests show that our narrative model predicts step-by-step uctuations in suspense levels for short stories that have a high correlation with average self-reported human suspense judgements6.
Our computational implementation is fully independent of the storyworld information. We thus claim to have created a model of suspense that could be used for any media or genre.
We see this work as a signpost towards further development of models of narrative based on what we see as its fundamental ingredients. Further work should explain how higher-level narrative concepts, such as certain types of plot and central characters, naturally include the key ingredients of what is needed to build a successful story. We now examine other directions for further research. 6.1
Scene-switching can be de ned as the alternation between di erent narrative view-points that show di erent sequences of events belonging to the same story. It is ubiquitous in suspenseful lms. One possible explanation for its use is that scene-switching increases the length of time that the suspense generated by a particular narrative thread is present in the story. If we have two narrative threads A and B, and we show rst A then switch to B, then as long as A does not get forgotten, its suspensefulness can continue to a ect the reader or viewer, even as we are watching events in narrative thread B7. 6.2
The inspiration for our model came from work on suspense, curiosity and surprise
introduced in [
Our narrative model can also model surprise. If thread A has many told events and thread B has no told events, then surprise will occurs if an event from thread B occurs in the story which suddenly disallows thread A.
In our view, the key moments of many narratives in lm or text combine all three of these entertaining narrative e ects in what we dub the `narrative breathing cycle', illustrated in Figure 4 on page 11. In this cycle, { rst a suspenseful situation is interrupted by a surprising event { this event sets up a new phase of revelatory or con ict-based suspense { another surprising event occurs which interrupts this suspense phase
Of course, stories di er in the amounts of con ict-based and revelatory suspense they evoke and also in the length of time that suspense is maintained before a new surprising event occurs. We can say that stories have di erent suspense and surprise pro les. 6.3
Earlier work on a net-linguistic implementation of an Earley Parser (see [
In this light, our research raises the possibility of constructing a complete functional theory of narrative. Such a theory would postulate that all the story steps that an author produces must modify, disallow or activate at least one narrative thread and thus have some e ect on the surprise, curiosity or suspense of the story at that point. Thus, just as we identify the function that each word plays in a sentence, we could identify the function that each story step plays in a story. We could derive a functional content-independent map of a story, much like the syntactic analysis of a sentence.
In our quest for a fundamental element underpinning narrative, we have been able to separate out suspense as an essential ingredient of narrative and also given a possible theoretical analysis of how it functions. We have also proposed a typology of suspense. This not only enables us to predict the suspense pro les of textual or lmic stories, but potentially also of other narrative-like artefacts such as music, advertising and humour. We think that it should be possible to use our model to generate suspenseful stories from a pre-de ned storyworld ex nihilo.
In making these contributions, we have followed the position that splitting phenomena such as narrative into its constituent parts is an approach that allows a high degree of portability into a variety of domains. We believe that the increased creativity and inter-disciplinarity that this way of proceeding fosters is a good measure of its potential.
Furthermore, as readers of suspenseful novels know, suspense makes us focus.
If the balance between skill and challenge is just right for a reader constantly
attempting to understand and predict what will happen next in a suspenseful
story, they may enter into the creativity-boosting state of Flow (see [
Most of this research was undertaken as part of a Ph.D. program at the Open University, UK under the supervision of Richard Power and Paul Piwek.