Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
SlideShare a Scribd company logo

Doust

•
•
A
anesah

This document provides a summary of Richard Doust's PhD research which aims to investigate narrative structures that produce effects like suspense, curiosity, and surprise. The research will develop a formal approach to modeling the inferences made during story comprehension and use this to measure suspense levels. A "storybase" method is proposed to systematically generate variant tellings of the same story in order to test models of narrative effects.

1 of 4
2010 CRC PhD Student Conference Investigating narrative ‘effects’: the case of suspense Richard Doust, richard.doust@free.fr Supervisors Richard Power, Paul Piwek Department/Institute Computing Status Part-time Probation viva Before Starting date October 2008 1 Introduction Just how do narrative structures such as a Hitchcock film generate the well-known feeling known as suspense ? Our goal is to investigate the structures of narratives that produce various narrative effects such as suspense, curiosity, surprise. The fundamental question guiding this research could be phrased thus: What are the minimal requirements on formal descriptions of narratives such that we can capture these phenomena and generate new narratives which contain them ? Clearly, the above phenomena may depend also on extra-narrative features such as music, filming angles, and so on. These will not be our primary concern here. Our approach consists of two main parts: 1. We present a simple method for defining a Storybase which for our purposes will serve to produce different ‘tellings’ of the same story on which we can test our suspense modelling. 2. We present a formal approach to generating the understanding of the story as it is told, and then use the output of this approach to suggest an algorithm for measuring the suspense level of a given telling of a story. We can thus compare different tellings of a story and suggest which ones will have high suspense, and which ones low. 2 Suspense 2.1 Existing definitions Dictionary definitions of the word ’suspense’ suggest that there really ought to be several different words for what is more like a concept cluster than a single concept. The Collins English dictionary gives three definitions: 1. apprehension about what is going to happen. . . 2. an uncertain cognitive state; "the matter remained in suspense for several years" . . . 3. excited anticipation of an approaching climax; "the play kept the audience in suspense" anticipation, ex- pectancy - an expectation. Gerrig and Bernardo (1994) suggest that reading fiction involves constantly looking for solutions to the plot-based dilemmas faced by the characters in a story world. One of the suggestions which come out of this work is that suspense is greater the lower the number of solutions to the hero’s current problem that can be found by the reader. Cheong and Young’s (2006) narrative generating system uses the idea that a reader’s suspense level depends on the number and type of solutions she can imagine in order to solve the problems facing the narrative’s preferred character. Generally, it seems that more overarching and precise definitions of suspense are wanting in order to connect some of the above approaches. The point of view we will assume is that the principles by which literary narratives are designed are obscured by the lack of sufficiently analytical concepts to define them. We will use as our starting point work on stories by Brewer and Lichtenstein (1981) which seems fruitful in that it proposes not only a view of suspense, but also of related narrative phenomena such as surprise and curiosity. Page 15 of 125 1
2010 CRC PhD Student Conference 2.2 Brewer and Lichtenstein’s approach In Brewer and Lichtenstein (1981) propose that there are three major discourse structures which account for the enjoyment of a large number of stories: surprise, curiosity and suspense. For suspense, there must be an initiating event which could lead to significant consequences for one of the characters in the narrative. This event leads to the reader feeling concern about the outcome for this character, and if this state is maintained over time, then the reader will feel suspense. As Brewer and Lichtenstein say, often ‘additional discourse material is placed between the initiating event and the outcome event, to encourage the build up of suspense’ (Brewer and Lichtenstein, 1981, p.17). Much of the current work can be seen as an attempt to formalise and make robust the notions of narrative understanding that Brewer laid out. We will try to suggest a model of suspense which explains, for example, how the placing of additional material between the initiating event and the outcome event increases the suspense felt in a given narrative. We will also suggest ways in which curiosity and surprise could be formally linked to suspense. We also hope that our approach will be able to shed some light on the techniques for creating suspense presented in writer’s manuals. 3 The storybase 3.1 Event structure perception Our starting point for analysing story structure is a list of (verbally described) story events. Some recent studies (Speer, 2007) claim that people break narratives down into digestible chunks in this way. If this is the case, then there should expect to discover commonalities between different types of narrative (literature, film, storytelling) especially as regards phenomena such as suspense. One goal of this work is to discover just these commonalities. 3.2 Storybase : from which we can talk about variants of the ’same’ story. One of the key points that Brewer and Lichtenstein make is that the phenomena of suspense depends on the order in which information about the story is released, as well as on which information is released and which withheld. One might expect, following this account, that telling ‘the same story’ in two different ways might produce different levels of suspense. In order to be able to test different tellings of the same story, we define the notion of a STORYBASE. This should consist of a set of events, together with some constraints on the set. Any telling of the events which obeys these constraints should be recognised by most listeners as being ‘the same story’. We define four types of link between the members of the set of possible events: • Starting points, Event links, Causal constraints, Stopping points. The causal constraints can be positive or negative. They define, for example, which events need to have been told for others to now be able to be told. Our approach can be seen as a kind of specialised story-grammar for a particular story. The grammar generates ‘sentences’, and each ‘sentence’ is a different telling of the story. The approach is different to story schemas. We are not trying to encode information about the world at this stage, any story form is possible. With this grammar, we can generate potentially all of the possible tellings of a given story which are recognisably the same story, and in this way, we can test our heuristics for meta-effects such as suspense on a whole body of stories. 4 Inference 4.1 Inference types To model the inferential processes which go on when we listen to or read a story, or watch a film, we define three types of inference: 1. Inference of basic events from sensory input : a perceived action in the narrative together with an ‘event classifier module’ produces a list of ordered events. 2. Inferences about the current state of the story (or deductions). 3. Inferences about the future state of the story (or predictions). Page 16 of 125
2010 CRC PhD Student Conference Clearly these inferential processes also rely on general knowledge about about the world or the story domain, and even about stories themselves. So, for each new story event we build up a set of inferences STORYSOFAR of these three types. At each new story event, new inferences are generated and old inferences rejected. There is a constant process of maintenance of the logical coherence of the set of inferences as the story is told. To model this formally, we create a set of ‘inferential triples’ of the form: “if X and Y then Z” or X.Y->Z, where X, Y, and Z are Deductions or Predictions. 5 Measuring suspense 5.1 A ‘suspense-grammar’ on top of the storybase To try to capture phenomena such as suspense, curiosity and surprise, we aim to create and test different algorithms which take as their input the generated story, together with the inferences generated by the triples mentioned above. A strong feature of this approach is that we can test our algorithms on a set of very closely related stories which have been generated automatically. 5.2 Modelling conflicting predictions Our current model of suspense is based on the existence of conflicting predictions with high salience. (This notion of the salience of a predicted conflict could be defined in terms of the degree to which whole sets of following predictions for the characters in the narrative are liable to change. For the moment, intuitively, it relates to how the whole story might ‘flow’ in a different direction.) For the story domain, we construct the set INCOMP of pairs of mutually conflicting predictions with a given salience: INCOMP = { (P1,NotP1,Salience1), (P2,NotP2,Salience2), . . . } We can now describe a method for modelling the conflicting predictions triggered by a narrative. If at time T, P1 and NotP1 are members of STORYSOFAR, then we have found two incompatible predictions in our ‘story-so-far’. 5.3 The predictive chain We need one further definition in order to be able to define our current suspense measure for a story. For a given prediction P1, we (recursively) define the ’prediction chain’ function C of P1: C(P1) is the set of all predicted events P such that P.y -> P’ where P’ is a member of C(P1) for some y. 5.4 Distributing salience as a rough heuristic for modelling suspense in a narrative Suppose we have a predicted conflict between predictionA and predictionB which has a salience of 10. In these circumstances, it would seem natural to ascribe the salience of 5 to each of the (at least) two predicted events predictionA and predictionB which produce the conflict. Now suppose that leading back from predictionA there is another predictionC that needs to be satisfied for the predictionA to occur. How do we spread out the salience of the conflict over these different predicted events ? 5.5 A ’thermodynamic’ heuristic for creating a suspense measure A predicted incompatibility as described above triggers the creation of CC(P1,P2,Z), the set of two causal chains C(P1) and C(P2) which lead up to these incompatible predictions. Now, we have : CC(P1,P2,Z) = C(P1) + C(P2) To determine our suspense heuristic, we first find the size L of CC(P1,P2,Z). And at each story step we define the suspense level S in relation to the conflicting predictions P1 and P2 as S = Z / L. Intuitively, one might say that the salience of the predicted incompatibility is ’spread over’ or distributed over the relevant predictions that lead up to it. We can call this a ‘thermodynamic’ model because it is as if the salience or ‘heat’ of one predicted conflicting moment is transmitted back down the predictive line to the present moment. All events which could have a bearing on any of the predictions in the chain are for this reason subject to extra attention. Page 17 of 125
2010 CRC PhD Student Conference If the set of predictions stays the same over a series of story steps, and in a first approximation, we assume that the suspensefulness of a narrative is equivalent to the sum of the suspense level of each story step, then we can say that the narrative in question will have a total suspense level S-total relative to this particular predicted conflict of S-total = Z/L + Z/(L-1) + Z/(L-2) + . . . + Z/L as the number of predictions in CC(P1,P2,Z) decreases each time a prediction is either confirmed or annulled. To resume we can a working definition of suspense as follows: 5.6 Definition of suspense Definition : the suspense level of a narrative depends on the salience of predicted con- flicts between two or more possible outcomes and on the amount of story time that these predicted conflicts remain unresolved and ‘active’. From this definition of suspense we would expect two results: 1. the suspense level at a given story step will increase as the number of predictions necessary to be confirmed leading up to the conflict decreases, and 2. the way to maximise suspense in a narrative is for the narrative to ‘keep active’ predicted incompatibilities with a high salience over several story steps. In fact, this may be just how suspenseful narratives work. One might say, suspenseful narratives engineer a spreading of the salience of key moments backwards in time, thus maintaining a kind of tension over sufficiently long periods for emotional effects to build up in the spectator. 6 Summary We make two claims: 1. The notion of a storybase is a simple and powerful to generate variants of the same story. 2. Meta-effects of narrative can be tested by using formal algorithms on these story variants. These algorithms build on modelling of inferential processes and knowledge about the world. 7 References • Brewer, W. F. (1996). The nature of narrative suspense and the problem of rereading. In P. Vorderer, H. J. Wulff, and M. Friedrichsen (Eds.), Suspense: Conceptualizations, theoretical analyses, and empirical explorations. Mahwah, NJ: Lawrence Erlbaum Associates. 107-127. • Brewer, W.F., and Lichtenstein, E. H. (1981). Event schemas, story schemas, and story grammars. In J. Long and A. Baddeley (Eds.), Attention and Performance IX. Hillsdale, NJ: Lawrence Erlbaum Associates. 363-379. • Cheong, Y.G. and Young, R.M. 2006. A Computational Model of Narrative Generation for Suspense. In Computational Aesthetics: Artificial Intelligence Approaches to Beauty and Happiness: Papers from the 2006 AAAI Workshop, ed. Hugo Liu and Rada Mihalcea, Technical Report WS-06-04. American Association for Artificial Intelligence, Menlo Park, California, USA, pp. 8- 15. • Gerrig R.J., Bernardo A.B.I. Readers as problem-solvers in the experience of suspense (1994) Poetics, 22 (6), pp. 459- 472. • Speer, N. K., Zacks, J. M., & Reynolds, J. R. (2007). Human brain activity time-locked to narrative event boundaries. Psychological Science, 18, 449-455. Page 18 of 125

More Related Content

Doust

  • 1. 2010 CRC PhD Student Conference Investigating narrative ‘effects’: the case of suspense Richard Doust, richard.doust@free.fr Supervisors Richard Power, Paul Piwek Department/Institute Computing Status Part-time Probation viva Before Starting date October 2008 1 Introduction Just how do narrative structures such as a Hitchcock film generate the well-known feeling known as suspense ? Our goal is to investigate the structures of narratives that produce various narrative effects such as suspense, curiosity, surprise. The fundamental question guiding this research could be phrased thus: What are the minimal requirements on formal descriptions of narratives such that we can capture these phenomena and generate new narratives which contain them ? Clearly, the above phenomena may depend also on extra-narrative features such as music, filming angles, and so on. These will not be our primary concern here. Our approach consists of two main parts: 1. We present a simple method for defining a Storybase which for our purposes will serve to produce different ‘tellings’ of the same story on which we can test our suspense modelling. 2. We present a formal approach to generating the understanding of the story as it is told, and then use the output of this approach to suggest an algorithm for measuring the suspense level of a given telling of a story. We can thus compare different tellings of a story and suggest which ones will have high suspense, and which ones low. 2 Suspense 2.1 Existing definitions Dictionary definitions of the word ’suspense’ suggest that there really ought to be several different words for what is more like a concept cluster than a single concept. The Collins English dictionary gives three definitions: 1. apprehension about what is going to happen. . . 2. an uncertain cognitive state; "the matter remained in suspense for several years" . . . 3. excited anticipation of an approaching climax; "the play kept the audience in suspense" anticipation, ex- pectancy - an expectation. Gerrig and Bernardo (1994) suggest that reading fiction involves constantly looking for solutions to the plot-based dilemmas faced by the characters in a story world. One of the suggestions which come out of this work is that suspense is greater the lower the number of solutions to the hero’s current problem that can be found by the reader. Cheong and Young’s (2006) narrative generating system uses the idea that a reader’s suspense level depends on the number and type of solutions she can imagine in order to solve the problems facing the narrative’s preferred character. Generally, it seems that more overarching and precise definitions of suspense are wanting in order to connect some of the above approaches. The point of view we will assume is that the principles by which literary narratives are designed are obscured by the lack of sufficiently analytical concepts to define them. We will use as our starting point work on stories by Brewer and Lichtenstein (1981) which seems fruitful in that it proposes not only a view of suspense, but also of related narrative phenomena such as surprise and curiosity. Page 15 of 125 1
  • 2. 2010 CRC PhD Student Conference 2.2 Brewer and Lichtenstein’s approach In Brewer and Lichtenstein (1981) propose that there are three major discourse structures which account for the enjoyment of a large number of stories: surprise, curiosity and suspense. For suspense, there must be an initiating event which could lead to significant consequences for one of the characters in the narrative. This event leads to the reader feeling concern about the outcome for this character, and if this state is maintained over time, then the reader will feel suspense. As Brewer and Lichtenstein say, often ‘additional discourse material is placed between the initiating event and the outcome event, to encourage the build up of suspense’ (Brewer and Lichtenstein, 1981, p.17). Much of the current work can be seen as an attempt to formalise and make robust the notions of narrative understanding that Brewer laid out. We will try to suggest a model of suspense which explains, for example, how the placing of additional material between the initiating event and the outcome event increases the suspense felt in a given narrative. We will also suggest ways in which curiosity and surprise could be formally linked to suspense. We also hope that our approach will be able to shed some light on the techniques for creating suspense presented in writer’s manuals. 3 The storybase 3.1 Event structure perception Our starting point for analysing story structure is a list of (verbally described) story events. Some recent studies (Speer, 2007) claim that people break narratives down into digestible chunks in this way. If this is the case, then there should expect to discover commonalities between different types of narrative (literature, film, storytelling) especially as regards phenomena such as suspense. One goal of this work is to discover just these commonalities. 3.2 Storybase : from which we can talk about variants of the ’same’ story. One of the key points that Brewer and Lichtenstein make is that the phenomena of suspense depends on the order in which information about the story is released, as well as on which information is released and which withheld. One might expect, following this account, that telling ‘the same story’ in two different ways might produce different levels of suspense. In order to be able to test different tellings of the same story, we define the notion of a STORYBASE. This should consist of a set of events, together with some constraints on the set. Any telling of the events which obeys these constraints should be recognised by most listeners as being ‘the same story’. We define four types of link between the members of the set of possible events: • Starting points, Event links, Causal constraints, Stopping points. The causal constraints can be positive or negative. They define, for example, which events need to have been told for others to now be able to be told. Our approach can be seen as a kind of specialised story-grammar for a particular story. The grammar generates ‘sentences’, and each ‘sentence’ is a different telling of the story. The approach is different to story schemas. We are not trying to encode information about the world at this stage, any story form is possible. With this grammar, we can generate potentially all of the possible tellings of a given story which are recognisably the same story, and in this way, we can test our heuristics for meta-effects such as suspense on a whole body of stories. 4 Inference 4.1 Inference types To model the inferential processes which go on when we listen to or read a story, or watch a film, we define three types of inference: 1. Inference of basic events from sensory input : a perceived action in the narrative together with an ‘event classifier module’ produces a list of ordered events. 2. Inferences about the current state of the story (or deductions). 3. Inferences about the future state of the story (or predictions). Page 16 of 125
  • 3. 2010 CRC PhD Student Conference Clearly these inferential processes also rely on general knowledge about about the world or the story domain, and even about stories themselves. So, for each new story event we build up a set of inferences STORYSOFAR of these three types. At each new story event, new inferences are generated and old inferences rejected. There is a constant process of maintenance of the logical coherence of the set of inferences as the story is told. To model this formally, we create a set of ‘inferential triples’ of the form: “if X and Y then Z” or X.Y->Z, where X, Y, and Z are Deductions or Predictions. 5 Measuring suspense 5.1 A ‘suspense-grammar’ on top of the storybase To try to capture phenomena such as suspense, curiosity and surprise, we aim to create and test different algorithms which take as their input the generated story, together with the inferences generated by the triples mentioned above. A strong feature of this approach is that we can test our algorithms on a set of very closely related stories which have been generated automatically. 5.2 Modelling conflicting predictions Our current model of suspense is based on the existence of conflicting predictions with high salience. (This notion of the salience of a predicted conflict could be defined in terms of the degree to which whole sets of following predictions for the characters in the narrative are liable to change. For the moment, intuitively, it relates to how the whole story might ‘flow’ in a different direction.) For the story domain, we construct the set INCOMP of pairs of mutually conflicting predictions with a given salience: INCOMP = { (P1,NotP1,Salience1), (P2,NotP2,Salience2), . . . } We can now describe a method for modelling the conflicting predictions triggered by a narrative. If at time T, P1 and NotP1 are members of STORYSOFAR, then we have found two incompatible predictions in our ‘story-so-far’. 5.3 The predictive chain We need one further definition in order to be able to define our current suspense measure for a story. For a given prediction P1, we (recursively) define the ’prediction chain’ function C of P1: C(P1) is the set of all predicted events P such that P.y -> P’ where P’ is a member of C(P1) for some y. 5.4 Distributing salience as a rough heuristic for modelling suspense in a narrative Suppose we have a predicted conflict between predictionA and predictionB which has a salience of 10. In these circumstances, it would seem natural to ascribe the salience of 5 to each of the (at least) two predicted events predictionA and predictionB which produce the conflict. Now suppose that leading back from predictionA there is another predictionC that needs to be satisfied for the predictionA to occur. How do we spread out the salience of the conflict over these different predicted events ? 5.5 A ’thermodynamic’ heuristic for creating a suspense measure A predicted incompatibility as described above triggers the creation of CC(P1,P2,Z), the set of two causal chains C(P1) and C(P2) which lead up to these incompatible predictions. Now, we have : CC(P1,P2,Z) = C(P1) + C(P2) To determine our suspense heuristic, we first find the size L of CC(P1,P2,Z). And at each story step we define the suspense level S in relation to the conflicting predictions P1 and P2 as S = Z / L. Intuitively, one might say that the salience of the predicted incompatibility is ’spread over’ or distributed over the relevant predictions that lead up to it. We can call this a ‘thermodynamic’ model because it is as if the salience or ‘heat’ of one predicted conflicting moment is transmitted back down the predictive line to the present moment. All events which could have a bearing on any of the predictions in the chain are for this reason subject to extra attention. Page 17 of 125
  • 4. 2010 CRC PhD Student Conference If the set of predictions stays the same over a series of story steps, and in a first approximation, we assume that the suspensefulness of a narrative is equivalent to the sum of the suspense level of each story step, then we can say that the narrative in question will have a total suspense level S-total relative to this particular predicted conflict of S-total = Z/L + Z/(L-1) + Z/(L-2) + . . . + Z/L as the number of predictions in CC(P1,P2,Z) decreases each time a prediction is either confirmed or annulled. To resume we can a working definition of suspense as follows: 5.6 Definition of suspense Definition : the suspense level of a narrative depends on the salience of predicted con- flicts between two or more possible outcomes and on the amount of story time that these predicted conflicts remain unresolved and ‘active’. From this definition of suspense we would expect two results: 1. the suspense level at a given story step will increase as the number of predictions necessary to be confirmed leading up to the conflict decreases, and 2. the way to maximise suspense in a narrative is for the narrative to ‘keep active’ predicted incompatibilities with a high salience over several story steps. In fact, this may be just how suspenseful narratives work. One might say, suspenseful narratives engineer a spreading of the salience of key moments backwards in time, thus maintaining a kind of tension over sufficiently long periods for emotional effects to build up in the spectator. 6 Summary We make two claims: 1. The notion of a storybase is a simple and powerful to generate variants of the same story. 2. Meta-effects of narrative can be tested by using formal algorithms on these story variants. These algorithms build on modelling of inferential processes and knowledge about the world. 7 References • Brewer, W. F. (1996). The nature of narrative suspense and the problem of rereading. In P. Vorderer, H. J. Wulff, and M. Friedrichsen (Eds.), Suspense: Conceptualizations, theoretical analyses, and empirical explorations. Mahwah, NJ: Lawrence Erlbaum Associates. 107-127. • Brewer, W.F., and Lichtenstein, E. H. (1981). Event schemas, story schemas, and story grammars. In J. Long and A. Baddeley (Eds.), Attention and Performance IX. Hillsdale, NJ: Lawrence Erlbaum Associates. 363-379. • Cheong, Y.G. and Young, R.M. 2006. A Computational Model of Narrative Generation for Suspense. In Computational Aesthetics: Artificial Intelligence Approaches to Beauty and Happiness: Papers from the 2006 AAAI Workshop, ed. Hugo Liu and Rada Mihalcea, Technical Report WS-06-04. American Association for Artificial Intelligence, Menlo Park, California, USA, pp. 8- 15. • Gerrig R.J., Bernardo A.B.I. Readers as problem-solvers in the experience of suspense (1994) Poetics, 22 (6), pp. 459- 472. • Speer, N. K., Zacks, J. M., & Reynolds, J. R. (2007). Human brain activity time-locked to narrative event boundaries. Psychological Science, 18, 449-455. Page 18 of 125