<<

INTERNATIONAL INSTITUTE OF INFORMATION TECHNOLOGY, HYDERABAD

Generating Non-Linear

by Vinay Chilukuri

A thesis submitted in partial fulfillment for the degree of Masters by Research in Computer Science.

Cognitive Science Lab

June 2011

INTERNATIONAL INSTITUTE OF INFORMATION TECHNOLOGY, HYDERABAD

Certificate

It is certified that the work contained in this thesis, titled ‘Generating Non-Linear Narratives’ by Vinay Chilukuri has been carried out under my supervision and is not submitted elsewhere for a degree.

Date: Advisor: Prof. Bipin Indurkhya

iii “A story should have a beginning, a middle, and an end... but not necessarily in that order.”

Jean-Luc Godard. INTERNATIONAL INSTITUTE OF INFORMATION TECHNOLOGY, HYDERABAD

Generating Non-Linear Narratives by Vinay Chilukuri Advisor: Dr. Bipin Indurkhya

Abstract

Stories a crucial role in our day-to-day lives. Apart from the entertainment sector, they are also being utilized in different areas of games, education and training. The ability to generate rich stories similar to human authored ones is of significance in story generation research. Along with the content of the story the manner in which it is presented is an important factor in effective . Though a lot of development happened in auto- matic story generation, most of it ignored the way in which the story could be presented. This thesis deals with a presentation aspect of the story, the order in which the events could be narrated, in order to make it more engaging than that of the chronological versions generated by many story generation systems.

We hypothesize that cognitive engagement is manifested when a is deviated from its chronological order without affecting the ease of comprehension. Narratives with events presented in out of chronology are termed non-linear. From a computational perspective, given n events of a story, the total orderings possible are n!. Generating an order in random would affect the comprehension of the material. We have studied how the comprehension of the viewers was affected upon changing the presentation order of the events, by employing a commercial feature film. Using these experimental results along with previous theory from narrative comprehension, we arrive at an algorithm that generates different presentation orders of a story by assessing the amount of processing load required for processing an order of events. Using the principles of the event-indexing model, a cognitive model of narrative comprehension, a function was devised that calculates the processing load, by taking input, the desired amount of non-linearity that could be imparted into the story.

An experimental evaluation is undertaken in order to test the generated narratives by our system, corresponding to the different input levels of non-linearity, in terms of engagement. We have obtained results that, for the story employed, increasing the non-linearity up to a medium level did not produce any change in the difficulty in comprehension and thus contributing to engagement. However, it was observed that higher levels of non-linearity led to a decrease in engagement and that, inconsistencies present in the story had a negative affect on the engagement of the material. Acknowledgements

I would like to whole heartedly thank my research advisor, Prof. Bipin Indurkhya, for supporting my idea since its inception. Without his consent, this idea of mine would not have materialized into a full blown research project. He has given me great independence in choosing and pursuing my topic of interest and carrying it forward. Under his supervision, I have learned about what is research and how to do it. Sincerely, this research in an interesting field would not have been possible without his support.

I like to thank Amitash Ojha for constantly helping me out with any problem I had. His experience and knowledge were immensely useful for my study. Specifically, his help on experiment design and statistical analysis facilitated my work. Without his questioning, I might not have been able to identify the finer aspects and problems involved with my work. I cherish the philosophical talks, the metaphoric language and the plight that we shared before the night of paper submission deadlines. He has been a mentor and more than that a great friend.

I would like to thank Saraschandra Karanam for his immense help and advice on the prob- lems that I presented to him. Because of his advice on many matters related to both my work and the issues of the University, my work here was completed in a smooth manner. He urged me to complete my thesis as soon as possible and get over with it. We shared many good moments on discussing a number of topics ranging from research to culture and to films. I also cherish the fun table-tennis matches, the dinners and the late night movie goings that we had together.

I’d like to sincerely thank Apara Ranjan for constantly questioning me and enabling me to bring out a research problem from an inquisitive idea that I had. Discussions with her introduced me to behavioral research and helped me to streamline my thoughts.

There will be dark times and there will be light. I had, perhaps, the darkest phase of my life. And then, there were Pradeep and Pujitha. I never believed that one could make best friends after completing college and while doing their Masters. I was so wrong. These amazing friends helped me to come out of a black hole that I was experiencing during my stay at this Institution. A meeting in a creating writing class brought me two friends for life. All the moments with Pradeep about films, culture, behavior and life are so enriching that I think that he is a comrade-in-arms for me. I feel short of words to describe the amazing support Pujitha has given me through these years of my Masters. A thank you seems pale to what she has done. Each and everything that we shared together has to be cherished. I have realized how important it is to be around the right set of people. It is amazing when some one brings a great enthusiasm in you and make you work better. Mahesh is an

vi awesome friend, with whom I share a great bond talking about gaming, music and movies. I appreciate his encouragement for my research when I used to discuss it with him. Mahesh, Pradeep and Pujitha gave me the best of times, making my mind saner in an environment to concentrate on my work.

Without my fellow batchmates, I would not have come out of this quagmire. The amazing fun times I had in the shady restaurants with Vijay, Sai Deepak and Sandeep were great. I would specially want to thank Vijay for putting up with all my crazy rants right from the start of entering this Institution. I’d like to thank Yogi for experiencing the hazy nights with me and sharing all the problems we experienced as postgraduate students. He is a great friend.

It is so good to meet people who have the same passion as yours. Vamshi is a fellow lab member but a good friend with whom I passionately talk about films and life. There are few people who understand my obsessions with films and he is also in the same boat with me. His constant rants to quickly complete my thesis and get into making a film were energizing.

I would also like to specifically thank my dear friends, Sharmi, GSK, Vedanth, Harshita, Koumudhi and Meher without whom there would be a part missing out of my life here. I would like to extend my thanks to all my fellow members of the Cognitive Science Lab for supporting me in my research. Without the environment of all these people, this thesis would not have been possible. I cannot imagine myself completing my work in any other environment as good as this.

Last but not the least and most importantly, I would like to thank Hans Zimmer for com- posing those beautiful soundtracks. I would definitely not have completed a single report, a paper or my thesis without listening to those amazing scores. I strongly feel that his scores talk about the ability of a man to achieve anything. I used to feel that I am solving an important problem for the world while listening to his music. It has become an integral part of me that I cannot do without it anymore. I owe my productivity to Hans Zimmer’s music. Thank you very much.

Vinay Chilukuri.

Contents

Declaration of Authorship iii

Abstract v

Acknowledgements vi

List of Figures xi

List of Tables xiii

1 Introduction1 1.1 Problem Statement & Approach ...... 3 1.2 Contributions ...... 4 1.3 Layout of the Thesis ...... 5

2 Background Study7 2.1 Narrative Theory ...... 7 2.2 Engagement ...... 10 2.3 Narrative Generation ...... 12 2.3.1 Story Generation: Computational perspective ...... 12 2.3.2 Sjuzhet Generation/Narrative Generation ...... 14 2.3.2.1 Prevoyant ...... 15 2.3.2.2 Montfort’s Research in Interactive - The nn system . 17 2.3.3 Discourse Generation ...... 18 2.4 Narrative Comprehension ...... 19 2.4.1 Event Indexing Model ...... 20 2.5 Our work ...... 21

3 Experimental Study of Narrative Comprehension 23 3.1 Theory ...... 23 3.2 Method ...... 25 3.2.1 Experiment 1 ...... 25 3.2.2 Results ...... 28 3.2.3 Experiment 2 ...... 30 3.2.4 Results ...... 31

ix Contents x

3.3 Discussion ...... 32

4 Generating Non-linear Narratives 35 4.1 Problem Revisited ...... 35 4.2 Obtaining Engagement by Narrative Variation in Order ...... 36 4.3 Defining Non-linearity ...... 37 4.4 Design of the Algorithm ...... 38 4.4.1 Calculating the Degree of Non-linearity ...... 39 4.5 Working Example ...... 42 Calculation of the degree of causality ...... 43 4.6 Implementation ...... 44 4.7 Evaluation ...... 45 4.7.1 Method ...... 45 4.7.2 Results ...... 47 4.7.3 Discussion ...... 49

5 Conclusion 53

Bibliography 57 List of Figures

2.1 Different planes of a Narrative ...... 8 2.2 Factors affecting Narrative Engagement ...... 11 2.3 Process of Computational Narrative Generation ...... 14

3.1 Average recall scores(in percentages) of the participants across the groups . . 29 3.2 Average recall scores(in percentages) of the participants across the groups, when material is scrambled at a micro-level...... 31

4.1 Algorithm for event selection and presentation ordering based on the degree of non- linearity ...... 42 4.2 An input Fabula adopted from Bae and Young (2008) ...... 43 4.3 Graph showing the causal links among the events of the fabula. Adopted from Bae and Young (2008) ...... 44 4.4 Low level non-linear narrative (2nd select) obtained after running the algorithm .. 46 4.5 High level non-linear narrative (minimal level) obtained after running the algorithm 46 4.6 Difficulty in comprehension as measured by the Narrative Understanding scale of Busselle (2009) over various non-linear narratives ...... 48

xi

List of Tables

3.1 Overview of the original (Non-linear) version of the film...... 26 3.2 Mean recall scores of the participants (in percentages) in different categories of the questions across the groups ...... 29 3.3 Overview of the linear version of the segment, The Gold Watch from the film. 30 3.4 Motivations of the main character in each segment of the film considered in our study...... 33

4.1 PL-values of sequence pairs in the fabula calculated at the first run of the Algorithm ...... 42 4.2 Scale used to assess the Engagement of the participants after reading the non-linear narratives ...... 47 4.3 Post-hoc analysis of the values of difficulty in comprehension across the non- linear narratives ...... 49

xiii

To my Parents

xv

Chapter 1

Introduction

Stories are ubiquitous. The way in which a story is told is an important aspect in storytelling. Besides the content of the story, the manner in which it is presented is also effective in engaging the . Writers and filmmakers use narrative techniques like flashbacks, in media res, flash-forwards etc. to effectively convey the content of their material. Generating these kind of narratives that engage the audience are of high value in effective communication and in the experience of the story.

Story generation systems find applications in many areas. These include the creation and control of user interfaces, entertainment software, educational software applications and corporate training tools as suggested by Mateas and Sengers(2003). In education, narrative centered learning environments can enable the user to actively obtain knowledge through a form of story (Mott and Lester, 2006, Mott et al., 1999). Narratives have been also employed in training scenarios to impart and improve the skills of the participants. (Hill et al., 2003).

Lot of AI research in story generation was concentrated on generating the content of the story over the decades (Meehan, 1976, Nelson and Mateas, 2005, Riedl and Young, 2004). Much less focus has been given to the way in which information is conveyed to the audience which is also fundamental for its appreciation by the readers/viewers. The outcome of most of the story generation systems is a series of events connected in their chronological order of occurrence. However, the narratives we experience via text or film are hardly linear. This research addresses the issue of the presentation aspect of the story generated by such systems.

1 Chapter 1. Introduction 2

Authors rearrange the events of the story to convey them in an effective and engaging manner. Given the events of a story, identifying the order in which these events have to be presented depends on the causal structure of the story and the effect that the author wants to convey. For example, if the intention of the author is to generate , as in the mystery genre, the ordering of the events might be different than that of, if the author’s desire is to convey information, as in a drama genre. Presenting the events of a story in simple chronological fashion might not always be effective in conveying the information, like the current story generation systems do. The intention of this research is that it might serve as an extension to the functionality of a Story Generation system. That is it takes the output of the story generation system as input and generates a narratives by rearranging the events in order to achieve better engagement than that of the chronological narratives.

Rearranging the order in which the events are presented might have an effect in conveying the information in an engaging manner by introducing dramatic effects such as suspense, surprise etc. (Brewer and Lichtenstein, 1982). In this thesis, we term narratives that present their events in chronological order as linear, while narratives that have their events presented in out-of-chronological order are termed non-linear. Given that the subject of the story world is familiar, a story presented in a simple chronological order might bore the viewer and he might guess in advance what is going to happen. This might be because a typical viewer must be familiar with a lot of narratives and a narrative schema would be developed over the years. On the other hand, if the story was presented in a highly scrambled manner, it would be difficult for the viewer to comprehend because of his/her limited processing capabilities and memory.

One could generate non-linear narratives by randomly reordering the events of the narra- tive (Montfort, 2007b). But generating such narratives by randomly reordering the events is confusing for practical purposes. Previous studies report a decrease in comprehension when the events of a narrative are scrambled randomly (Cowen, 1984, Ohtsuka and Brewer, 1992). Therefore, we need to find a subset of narratives that do not significantly affect the comprehension of the material even when changing the order of presentation. From a nar- rative generation perspective, one needs to find out how much could a narrative be deviated from its chronological order because a large amount of deviation would result in a loss of comprehension of the content and also might take away the experience of a story.

I hypothesize that, if the content of the story is familiar to the audience, narratives that Chapter 1. Introduction 3 present their events in a different order are more engaging than that of a simple chronological (or linear) order. The engagement referred here is the Cognitive engagement that results from how the information is processed by a person as opposed to the Emotional Engagement (Kintsch, 1980).

Automatically generating a narrative for a narratee (person who consumes the narrative) to comprehend requires knowledge of the process of Narrative Comprehension. Specifically, one needs to understand how comprehension is affected by varying the order in which the events are presented in a narrative. Previous research (Cowen, 1984, Ohtsuka and Brewer, 1992) indicates that comprehension decreased with increasing the scrambling of the events in a narrative. Most of this research was performed by employing narratives designed for experimental purposes and with narratives of very short duration (Cowen, 1984, Furman et al., 2007, Ohtsuka and Brewer, 1992). At the time of this writing, no study exists, to the best of my knowledge, that evaluates the order effects of comprehension on a large scale narrative, like a or a feature-film. In this thesis, I study the effect of different orderings on events on the comprehension of the material in the context of a feature-film.

Knowledge of the effect of the presentation order of events of a story on the comprehension of the reader is vital in order to generate narratives of various presentation orders.

1.1 Problem Statement & Approach

Given n events comprising a story, the number of possible orderings are n!. Out of these possibilities, we need to find the narratives that are comprehendable to the viewers, with their limited processing capabilities. Though a high-level non-linear narrative might be understood by a few viewers, there is a danger that an attempt to comprehend such material might lead a viewer’s focus to frustration.

Also, we hypothesize that narratives that are in different order without requiring a lot of processing load are more engaging than that of the narratives that require more processing load or no processing load. Therefore, the problem is to find the amount of non-linearity that could be introduced in the story without compromising on the viewers comprehension.

This problem is approached by first studying how narrative comprehension is affected with the change in order of events in a narrative. A feature-film, that is non-linear in nature, is Chapter 1. Introduction 4 edited in different ways resulting in varying orders of non-linearity. Each of the narratives are shown in an experiment and the recall of information to various questions pertaining to the narrative was measured. Also, a segment from the same film was edited in different ways and similar procedure was performed. Results were obtained that state that comprehension is not affected when the goal-hierarchy of the narrative is maintained, while re-arranging the narrative.

The results obtained in my experiments along with previous research, were leveraged for the process of achieving narrative variation in the order of events. My module takes the output generated by a story generator, a story world or a fabula, which is a simple chronological order of events, represented by a partial order plan, and generates an order of events based on a user input of the level of non-linearity (low, mid and high), a sjuzhet. Combining the principles of the Event Indexing model (Zwaan et al., 1995) and the partial order plan representation, I have designed a function that outputs a value representing the degree of non-linearity between two events. This function, based on the Processing Load Hypothesis of the Event-Indexing model, is the core of our system. Using this function the degree of non-linearity is assessed among all the pairs of sequences present in the story, and a presentation order is generated according to the desired input of non-linearity.

Empirical evaluation is conducted on the generated narratives on the basis of engagement and the degree of non-linearity one could impart into a story to make it more engaging.

1.2 Contributions

• The foremost contribution of this work is the design of a system and its implementation that generates narratives of different presentation orderings according to the user input of desired non-linearity. We were also interested to find the amount of non-linearity that one could introduce in a story in order to make it more engaging and without affecting its comprehension. The results of the study state that narratives of high-level non-linearity are not engaging than that of their chronological counterparts.

• This study addresses the problem of generating non-linear narratives. My system does not randomly generate an order, but using the principles of the Event-Indexing model, it decides which event to present next in order to maintain the desired level of non- linearity in a narrative. A fundamental question of narrative generation is to know Chapter 1. Introduction 5

what event to present next in a narrative. We designed a function that determines what event has to be presented next and that keeps a check on the degree of non- linearity, and thus the comprehensibility of the order that is being generated.

• We studied the order effects on narrative comprehension using actual film material (a feature length film). No study exits that employs large scale narratives in film. Results of this experiment support the Event-Indexing Model and extend it, which was originally postulated for text narratives, to the domain of film as well. Also the results emphasize the importance of the motivation of the principal character of a story. This refines the model’s dimensional equality assumption. This research provides evidence that non-linearity is not only dependent on rearranging the events but the level of granularity at which the events are reordered by maintaining the structure of the story.

• My system that deals with the presentation aspect of a story, extends the functionality of a story generation system, by generating narratives which are close to the real world.

1.3 Layout of the Thesis

The rest of the thesis explains our approach to develop the algorithm that could generate different orders for a given story, without affecting the comprehension of the content. Chap- ter 2 gives the background information to understand the context of my work. Overview on topics of Narrative theory, Narrative Engagement are provided. The chapter also talks about the computational generation of stories and discusses two systems that are more relevant to my work on Narrative generation. The chapter ends with the overview of Cognitive models of Narrative Comprehension and explains the Event-Indexing Model. Chapter 3 talks about my experimental study of narrative comprehension with respect to varying orders of non- linearity. Chapter 4 presents my theory in arriving at the algorithm for narrative variation, the algorithm, its working and evaluation. Chapter 5 concludes this thesis by providing further research questions that could be taken up in the future.

Chapter 2

Background Study

2.1 Narrative Theory

This section introduces a few basic terms in to set the context of this thesis.

In terms of narratology, story and narrative are two different things. There are two levels when discussing a narrative. The content level and the expression/presentation level. A story refers to the content, the characters, the events and objects that inhabit the story world while a narrative also deals with the expression of this content to the narratees (the audience) (Genette et al., 1980). Content and Expression might be generic terms. According to Russian formalism, the set of all the objects (including characters), actions and events comprising the story world in their chronological occurrence is termed as fabula. The way in which the events are presented to the narratees is termed sjuzet. Many researchers refer the fabula as the story and the sjuzet as the Narrative. In this thesis we use the terms fabula and sjuzet with story and narrative interchangeably.

The way in which a story is narrated is characterized in Narratology by the following param- eters: Order, Speed, Frequency and Voice (Chatman, 1980, Genette et al., 1980). The order in which the events of the fabula are narrated is termed as the order of . The story events that occur in chronological order could be varied, for aesthetic or dramatic purposes when they are narrated. For example, in mystery stories, an important piece of information, though happens in the beginning of the story, is often revealed to the readers/viewers at the end for achieving the dramatic effect of suspense. Our research deals specifically with this variation in Narrative Order. 7 Chapter 2. Background Study 8

Figure 2.1: Different planes of a Narrative

Prince (1973) states that “Ordering of events in time is one of the most fundamental char- acteristics of any story”. Given the events of a fabula, that are in chronological order, there are many ways of ordering them in presentation of a narrative. Specifically, given n events in chronological order, the total number of orderings are n!. However, if we look at the event orderings in a narrative, they fall into certain classes that have been there since tradi- tional narratives. Genette et al.(1980) identifies these conventional orderings prevalent in narratives over the years and categorizes them into following:

Chronicle A narrative in which the events are ordered according to their chronological occurrence in the story. That is there is no change in the order of the events of the story when presented.

Retrograde The events are ordered in the reverse of their chronological occurrence in the story. Examples of this type of narration are the films: Irreversible (2002), the main sequence in Memento (1999).

ZigZag In these narratives events from different time frames are presented together. The events that are paired must be similar in some way.

Analepsis An event or a sequence of events from the “past” are presented along with the events that are being narrated in the “now”. Also, the event that from the past must be related in some fashion to the event that has just finished. Example of this type are the flashbacks found in many stories.

Proplepsis Similar to analepsis, the events from the “future” are presented along with the events in the “now”. Though the similarity between the current event and the event of Chapter 2. Background Study 9

prolepsis might not be evident to the narratee while experiencing the narrative, it would be clear in retrospect. These type of narratives are called flashforwards.

Syllepsis The order of events is based on some grouping that is not chronological. For instance, in recounting a stereotypical adventure, all the encounters with monsters might be narrated, all the arrivals in new places, and then all the acquisitions of treasures.

Achrony Narratives that have their events ordered in an out-of-chronological fashion randomly are of this type. In this case, the relationship between the order in which events are narrated and the order in which they occur is impossible to establish, or is exceedingly difficult to establish and seems arbitrary. We term narratives that are achronous to be non-linear. Example: Tarantino’s film Pulp Fiction (1994). The interesting thing with these type of narratives is that thought it is almost impossible for the researcher to establish the motivation behind the order of events, there might be some motivation for the author/filmmaker to order these narratives in the way they were ordered. That is we speculate that the order of events in these narratives are seemingly random.

The at which an event is narrated is termed as Speed. Each event of a narrative could be narrated with different Speed. It can be either included or omitted in the narrative. If it is included, then it could be narrated using more descriptive mechanisms or in a simple fashion. For example, an event in the story world that takes place in a duration of 5 minutes could be narrated to the viewer in a span of a minute or 15 minutes. Though Genette et al. (1980) has initially referred to this parameter as the narrative duration of the event, he later corrected it as the pace or the speed of an event, as it appropriately captures the pace of narrating and the pace of the narrated events.

Frequency determines how a set of events will be narrated. An event in the story could be narrated several times through out the narrative for aesthetic reasons. An example of this kind of narration is the film, Vantage Point (2008), where the same event is repeated through several times in the film albeit through different characters’ point of views.

Narrative Voice concerns with the Narrator, his relationship with the narratee and what is told (Genette et al., 1980). If the narrator is omniscient of everything in the story world, the voice is said to be an omniscient narration. If the narrator is one of the characters of the story, then it is termed as a Subjective Narration. If the narrator also includes his thoughts Chapter 2. Background Study 10 on the happenings of the story, then it is said to be a first person narrative. A narrative could be subjective but from a third-person point of view.

Though the sjuzet or the narrative plane deals with the expression aspect of the story, another plane has been identified that deals with the way the story is expressed in a medium like text or film. This form of expression in a medium is called the Discourse. For example, in the case of verbal narratives, rhetorical language could be used to narrate a story, while the same story could be narrated using simple language sentences. In filmic representation, an event could be narrated using long continuous shots and the same could be narrated using rapid succession of cuts. This form of expression of the story realized in a medium, is termed as the Discourse.

2.2 Engagement

Engagement is the quality of a narrative, the experience a narrative imparts in the audience. While it is difficult to formulate a functional definition to an experience, researchers have identified the factors that might lead to this experience through a narrative. Experiencing these factors, one might say that a person is engaged.

Busselle and Bilandzic(2009) has identified factors leading to Engagement that are at a fun- damental level with respect to the interaction between a person and a narrative. These are broadly classified into Narrative Understanding, Emotional Engagement, Attentional Focus and Narrative Presence. A scale was also developed based on these factors to measure the overall engagement of a narrative. We look at each of these factors.

Narrative Understanding The dimension of Narrative Understanding deals with the ease of comprehending a narrative or more specifically the lack of difficulty in comprehending it. To comprehend a narrative, audience members construct mental models of meaning to represent a story (Graesser et al., 1994, Van Dijk et al., 1983, Zwaan et al., 1995). From a mental model theory of comprehension (We cover comprehension in Section 2.4), this dimension deals with the ease of constructing mental models in order to make meaning out of the narrative. This dimension is measured using scale-items such as, “At points, I had a hard time making sense of what was going on in the program”. This sub-scale of Narrative Understanding is presented in Table 4.2. If a narrative is said to highly engage Chapter 2. Background Study 11

Figure 2.2: Factors affecting Narrative Engagement the audience, then according to this dimension, audience members should be unaware when comprehension progresses smoothly and should become aware only when it falters.

Emotional Engagement concerns with emotions viewers have toward the characters in the narrative, like empathy, sympathy etc. This dimension is related to the emotional arousal component of narrative engagement. Narratives arouse a lot of emotions in the viewers/readers, though it is difficult to measure the degree of emotions manifested in the audience members. According to McKee(1998), emotions contribute to a key role in engag- ing the audience.

Attentional focus describes the viewer’s focus on or distraction from the material being presented. Essentially one should not be aware that one is not distracted when experiencing a high level of engagement. A truly engaged viewer should be unaware of focused attention, and should become aware only if attention drifts or must be refocused.

Narrative Presence deals with the transportation of the reader/viewer to the story world of the narrative. This concerns the process of deixis - a process where the audience members locate themselves in the story world in order to make meaning of the narrative (Galbraith, 1995). As stated in Busselle and Bilandzic(2009), Chapter 2. Background Study 12

As suggested by Deictic Shift Theory (Segal, 1995a,b) audience members switch to the time and location of the narrative, and to the subjective world of the characters. This is necessary because some information makes sense only from the deictic center of the story.

Deictic shift can be seen as a cognitive process necessary for understanding and for emotional perspective taking processes, such as identification (Cohen, 2001) or empathy (Zillmann, 1995).

Narrative Presence might occur either because of this inherent deictic activity, which is unique to narratives, or because of an intense focus on the narrative that results in the loss of self-awareness. Narratives that engage the audience might have a high diectic activity.

Based on the above factors, we see that Engagement might result from the processing of a story (Narrative Understanding dimension) or from the emotions the narratee experiences while processing the story (Emotional Engagement or Narrative Presence) or any other subjective preferences. This distinction of Engagement that occurs based on different aspects of the narrative and the reader is important.

Kintsch(1980) categorized the interest manifested through a narrative to be of two kinds, cognitive interest and emotional interest. The former arises from a well-organized discourse structure; the latter from emotional context in the story. That is the emotional interest depends on the content of the narrative. We might also add that it depends on the subject’s background knowledge and his preferences towards the content of the narrative. Cognitive interest on the other hand occurs due to the processing of the narrative and the way it is structured. This research deals with the cognitive aspect of engagement manifested by a narrative.

2.3 Narrative Generation

2.3.1 Story Generation: Computational perspective

Computational Story generation is based on a basic assumption that a story can be ap- proximated by the help of a plan. A plan is a sequence of actions that move an AI agent Chapter 2. Background Study 13 from an initial state to a goal state. Likewise, a story is sequence of actions the takes from his initial conditions in order to achieve his/her goal. There has been a lot of research (Lebowitz, 1985, Meehan, 1976, Riedl and Young, 2004, Turner, 1993) based on the planning approach to story generation over the years. Other methods like Search-based Drama Management (Nelson and Mateas, 2005), case-based approach to creativity (Turner, 1993) were also developed.

In planning, an taken by the agent (or a protagonist) drives the agent to a state. Based on the state the agent is now in, it needs to choose the necessary action to arrive at its goal state. A precondition is the condition that specifies what must be true in a state before an action can be executed. An effect is the state in which the agent is in after the action has been executed. A good primer on planning is provided in Russell and Norvig (2009). If an effect of an action leads the agent to a goal state then the sequence of actions that were taken by the agent until then is the plan of the agent. Comparing this with a story, a series of actions leads a protagonist to reach the goal.

Like a planning algorithm, a story generation system is given the following as input: initial world description, set of actions available for the protagonist and the goal. The system finds the sequence in which these actions have to be taken by the protagonist (agent) in order to achieve the goal. The planning descriptions are usually in STRIPS specification. A fabula could be best approximated by using a partial-order plan (Young et al., 1994). Sec. 4.4 shows how the components of a partial-order plan represent various aspects of the events of a story.

Computational Story Generation could be broadly categorized into two approaches: an Autonomous Agent based approach and an Author-based one. In the former, a story is obtained by the interactions among the agents (or characters) inhabiting the story world. Each of the agents is given specific goals to achieve. The agent has to generate its own actions and execute them in order to realize its goal. The interaction of the actions among such events leads to a story. This approach has a long history and is first used by the system TALESPIN (Meehan, 1976). The advantages of using this approach is that it is relatively easy to generate a story. Also, the story generated might be believable (Cheong, 2007) as the agents plan their own actions in order to achieve their individual goals. But obtaining a story from the interactions among the agents in this simple way might not yield a story that is dramatically satisfying. Use of a story manager that places the agents in dramatic Chapter 2. Background Study 14 situations might yield rich stories. According to Cheong(2007), However, it is less likely that the generated story would be interesting without a story manager that is in charge of creating dramatic situation such as posing a global goal that needs the collaboration between the agents or arranging goals for agents that conflict each other. On the other hand, the author-centric approach provides plot coherency, since a global planning process is used to construct the actions of all characters in a story over the story’s entire duration. In this approach, however, it is difficult to ensure that each character acts according to its own internal nature, since actions are prescribed by a central planning system.

A lot of research has been undertaken over the years in Automatic Story Generation. Fab- ulist attempts at generating stories with characters that are more believable. Search based drama managers (Moe) (Nelson and Mateas, 2005) attempt at generating a story by search- ing through a plot-graph. Other attempts at story generation include systems that deal with generating stories in order to produce a dramatic effect (Cheong and Young, 2006, Turner, 1993). Recent works in this field include the works of Niehaus(2009) and Porteous et al. (2010). It is not in the scope of this thesis to look at these systems of story generation. A good review of the systems could be found at Cheong(2007), Niehaus(2009).

Figure 2.3: Process of Computational Narrative Generation

Traditional story-generation systems have concentrated on generating the content of the story rather than presentation. The above mentioned systems also deal primarily with generating the content of a story. Focus on the presentation aspect of a story has been slowly increased during the past five years. Our work is in this area and we look at a few systems that take a set of events in chronological order and generate a sjuzet.

2.3.2 Sjuzhet Generation/Narrative Generation

There has been relatively very less research in the area of presentation aspect of the story as opposed to content generation. In this section we look at systems that focus on the expression plane of the story and in particular that deal with Narrative variation in the Order of presentation of events of the story. We present two systems, Prevoyant (Bae and Young, 2008) and Nick Montfort’s nn system (Montfort, 2007a), that deal with temporally Chapter 2. Background Study 15 altering the order of events of a story and presenting them in a narrative for various effects. These works are relevant to our research.

2.3.2.1 Prevoyant

Prevoyant is a system that aims to achieve surprise at a point in the story by temporally altering the order of events by the means of flashbacks and . The authors state that a reader could be engaged with the result of a surprise occurrence in the story that he is reading. The surprise effect that Prevoyant tries to bring about is a Cognitive response than that of an emotional one. The basis of this system is that if certain information is hidden from the reader, and when an unexpected outcome occurs, it surprises the reader. The explanation of this unexpected occurrence is then presented as a flashback to the reader, who then finds that the story is coherent in retrospect. This attribute is termed as the postdictablity of the story.

Prevoyant takes in a source story (a fabula) in its partial-order plan representation and a Generator module finds out the points where a flashback or a could be inserted. It also finds out the content that needs to be inserted at these points. Once a deci- sion about the temporal re-ordering of the elements has been reached, an Evaluator module checks if the inserted flashback/foreshadowing has the potential to surprise the reader and the postdictability of the presented narrative. The Generator and Evaluator modules are explained in a more detailed fashion in the following paragraphs.

Reader Model A narrative generation system needs to know how a reader comprehends a narrative. Prevoyant’s reader model simulates the reasoning process of an implied reader. This reader model is implemented by a plan-based reader model that is a part of the Longbow planning system. Employing the reader model using a planning system is based on the assumption that human planning process can be approximated by a partial-order planner. The Longbow system is based on the DPOCL (Decompositional Partial Order Causal Link) planning algorithm (Young et al., 1994). The reader’s background knowledge and the current knowledge in the story characterizing what he has read so far is represented by a plan library.

The working of Prevoyant is carried by two modules, the Generator and the Evaluator. The Generator creates potential flashbacks and foreshadowings, while the Evaluator checks for Chapter 2. Background Study 16 the unexpectedness of the flashbacks and the postdictability of the story. The function of the Prevoyant is explained in four phases:

• Generation of a flashback: is a certain information in the story that is presented as the cause of an event after the event took place. According to the authors, if the reader is unaware of the absence of the initiating events leading to an event, then surprise occurs. Based on this principle, the Generator module identifies a potential set of events that could be considered for flashbacks. The goal literals in the causal plan structure of the fabula are treated as significant events (SE). The series of events leading to these goal literals (connected by causal links in the plan structure) are termed as Initiating Events (IE). For each pair of SE and IE, Prevoyant checks if the omission of the IE leads to any disruption in the causality of the story as experienced by the reader. This check is made with the help of the plan-based reader model of Prevoyant. When a significant event that meets the requirements is found, it is termed to be the point of insertion of the flashback and its corresponding Initiating Events are presented as a flashback, soon after the occurrence of the SE.

• Evaluation of unexpectedness of the flashback: The aim of the system is to evoke surprise in the reader. Hence, when the generator selects events in the stories as potential candidates for flashbacks, the Evaluator checks to see if these flashbacks cause unexpectedness in the reader. The level of unexpected- ness is not measured but the Evaluator only judges whether a flashback is unexpected or not. If the Reader model (the comprehension process of the Reader) is able to find a complete plan albeit the omission of the series of events leading to the significant event, then there is expectedness in the flashback. The evaluator discards this IE, SE pair and considers the next pair until a desired pair, that possibly evokes surprise, is found.

• Generation of a flashforward: The system does not explicitly generate a flashforward. As of current reading, it treats the non-flashback part of the narrative and considers it as a flash-forward.

• Evaluation for Postdictability: Postdictability is the ability of the reader to understand the macro-structure of the story in retrospect, without any conflicts among the story Chapter 2. Background Study 17

elements. According to the authors, since the source story is generated from a sound planner, the temporally reconstructed story in retrospect ensures postdictability.

Based on the selection of flashbacks and flashforwards, Prevoyant rearranges the events of the story and presents them in a narrative with the aim of evoking surprise. The partial-order plan structure of the narrative is realized in media with the help of a Zocalo environment.

2.3.2.2 Montfort’s Research in Interactive Fiction - The nn system

Though Montfort’s research is in the area of Interactive Fiction, its principles are strongly obtained from traditional narratology. Montfort’s nn system generates the world of the story along with how the events of the world are communicated to the player. The user/player takes the place of one of the characters of the world and he is allowed to perform a set of actions. He enters the actions at a user prompt. The system outputs the effect of his actions in the story world and its characters. This is one system that implements Narrative variation at different levels.

The nn system has a knowledge base that models the story world of the narrative, the world of the characters as seen by them and discourse level representations. It also has various modules (Simulator, Narrator and Recognizer) that interact with this knowledge base in generating the narrative and presenting it to the user. The set of possible objects, events and actions are modeled in the IF Actual World component. According to the author,

The main motivation for the Focalizer worlds is their use in narrating. They allow the generation of a narrative from a particular actor’s perspective.

Using these Focalizer worlds, it is possible to see how two different characters recount the story, including some overlapping events and some that are different. The Discourse model of the system deals with the interactor. It has to process the text entered by the user and print out the text to the user based on his previous actions and the current action. A Simulator Module interacts with the IF Actual World, the Focalizer World to generate an instance of the world of the story to the player. The Narrator module takes in this story world and according to a author-specified plan generates a representation to the input story world, before presenting it to the end user. Chapter 2. Background Study 18

The Narrator module in this system incorporates narrative variations of different types. They are variation in the order of events, variation in the speed (or duration) of the events, in frequency and in Narrative voice. We mainly look at how variation in Narrative Orders are achieved in this system because of its relevance to our work.

Given a set of events in chronological order, according to Genette et al.(1980), there are many ways of ordering them. The orderings found in narratives could be categorized at a high-level into: Chronicle, retrograde, zigzag, analepsis, prolepsis, syllepsis and achrony (as explained in Sec. 2.1). The nn system achieves these orderings by various processes as described in Montfort(2007b). However, it states that, in the case of generation of achrony, the type of narratives our research is focused on, ordering the events of a narrative at random is a possible way to produce achrony.

Randomly ordering events of a narrative would require a lot of the reader’s cognitive pro- cessing than originally required. Such narratives might cause frustration to the viewer. As the work of Montfort(2007b) put it, “A narrator that orders events uniformly at random is probably confusing for all practical purposes”. Comprehending such complex narratives is relatively difficult on the part of the reader/viewer. In order to generate narratives that have their events in an achronous order and yet are comprehensible, we need to look at the process of comprehension while processing a narrative.

Montfort(2007b) implements a mechanism that recounts a story by altering the presentation order of the events in the story.

2.3.3 Discourse Generation

Discourse generation systems primarily deal with how a story is realized with respect to a medium. (Though some authors mean discourse to encapsulate both sjuzet and the medium of representation, there is no stringent distinction between the two. In this paper we term discourse to mean the representational aspects of the story that are related to the medium). Textual narratives depend on the aspects of language for expression, whereas visual narratives use film grammar, to convey the story. Discourse generation deals with automatically realizing the story in a medium, according to the goals of an author.

Given the fabula as the input, if the fabula has to be realized in textual format, the discourse generator takes the role of a Natural Language Generator (NLG) that concerns with task Chapter 2. Background Study 19 of representing the events of the fabula in sentences. This task is based on the parameters of the narrative prose namely, person (should the narrative be presented in a first person, second or a third person narration), focalization (should it be a subjective narration or an omniscient narration), order of narration etc. (Loenneker, 2005). STORY BOOK (Callaway and Lester, 2002) is one system that addressed the issues of Narrative prose in medium realization. However, STORY BOOK does not implement the above discourse parameters. It expects the story that is input to it to be equipped with them. Loenneker(2005) addressed this gap and provided emphasis on the discourse plane of the narrative. Narrative levels that deal with the level of the presence of the narrator were put forth in his research. Montfort(2007a) also significantly deals with issues of discourse and their implementation in his system on interactive fiction.

Discourse in the visual medium would deal with decisions regarding placement of the camera, the orientation of the characters, camera moves etc. depending on the events of the story. This intelligent camera control mechanism for narrative discourse was implemented by Jhala and Young(2005). Intelligent cinematography is also used in the field of machinima as a storyboarding and visualizing tool (Jhala et al., 2008, Riedl et al., 2008) and also in production (Elson and Riedl, 2007).

2.4 Narrative Comprehension

Comprehension is an essential activity of the human brain. It could be termed as the process of making meaning of certain information or could also be termed the output of this process.

The process of comprehension could be defined as the interaction between incoming informa- tion and the information already processed (Branigan, 1992). When a reader reads a text, it is processed into propositional constructs called as the textbase. Further processing of these constructs results in the formation of a mental model called the situation model. This incoming information has to be fused with the information already present in the reader’s memory. This process of inference involves finding the relevant information with respect to the incoming information, activating it and then forming associations between them. Infer- ences are categorized into two types in the process of Comprehension. When the incoming information has to be inferred with respect to the information that has been read previously, they are called bridging-inferences. Comprehension process also involves making meaning of Chapter 2. Background Study 20 the incoming information with the reader’s prior experiences or the background knowledge. The inferences of such kind are called knowledge-based inferences or elaborations.

Several models have been put forth to address this complex process of comprehension that study the construction of mental models in memory, their representation, how previous information is activated in memory etc. Though the general aim of all these models is to explain the process of comprehension, each of them deals with specific issues pertaining to the process. A good review of the prominent models of comprehension is provided by McNamara and Magliano(2009)

In this section we look at a model called the Event-Indexing model proposed by Zwaan et al. (1995), that is most relevant to the field of Narrative Comprehension. The model’s focus is the construction of situation models, and how inferences are generated based on the these mental representations, while processing a story. We base our work using the principles of this model due to its relevancy to narratives and the results that corroborate it.

2.4.1 Event Indexing Model

All of the models of comprehension agree on the assumption that the process of comprehen- sion involves the construction of mental models called Situation Models. The Event-Indexing model proposed by Zwaan et.al garnered a lot of interest over the years, in the context of narrative comprehension. According to the Event-Indexing Model, a situation or an event of a narrative is mentally represented by five dimensions namely:

• Time: the time at which the event took place

• Space: The location at which the event occurred.

• Characters/Protagonist: The central character in the event or the characters present in the event.

• Causality: If the event is causally related to the one that occurred previously

• Motivation/Intention: The motivation of the action performed by the protagonist or a principal character in the event.

When an event is first processed, its five indices are stored and a node is created in memory. When the next event occurs, its indices are compared with that of the previous node. If Chapter 2. Background Study 21 there is a difference in any of the index, it is updated and a new node is created in the memory. If the new event being presented has an overlap with any of the indices of the representation of the event already in the memory, a link on this index is formed between them in memory. As the events are presented, links are formed among them based on this index or dimensional overlap.

According to the EI model, the current model is the model currently under construction in working memory. This constitutes the construction of the situation model of the current event or the incoming information. The information that has already been processed and integrated with previous information from the narrative or the background knowledge of the comprehender constitutes the integrated model. The integrated model is the global model that was constructed by integrating, one at a time, the models that were constructed at times t1 to tn − 1 while the person processes information. Comprehension is the process of integrating the incoming information with the one that has already been integrated. The current model has to be connected to the integrated model for the reader to make meaning of the story. The connections are made based on the similarity on the dimensional indices of the two models. Finally, The complete model is the model that is stored in long-term memory after all the textual input has been processed.

Processing Load Hypothesis According to the Event-Indexing model, the ease of inte- grating two representations of the events are dependent on the degree of their overlap on the mentioned five dimensions. More the number of common indices between two representa- tions of the events, the easy it is to integrate the current node into a coherent structure. This is called the Processing load hypothesis. It concerns the updating process of the situation models in memory.

2.5 Our work

Our work addresses the issue of generating a non-linear narrative (Achrony) without affecting the comprehension of the material. Instead of randomly reordering events of a narrative, we devise a function that gives an order of events based on the user-desired level of processing complexity. We believe that non-linear style of narration or ‘Achrony’ is the generic style of narrative generation. Our question is to find what event has to come next after a certain event, without affecting comprehension of the readers/viewers. In this regard, we develop Chapter 2. Background Study 22 an algorithm that estimates the degree of processing load between any two events of a story. We use this function to make decisions about the narrative generation process. Unlike almost all of the narrative generation systems, we do not use any plan-based reader model that approximates the reader’s reasoning process. Rather, we adopt the principles of the Event-Indexing model, a cognitive model of Narrative Comprehension, in this regard. To facilitate the process of Narrative generation using a cognitive model, we need to understand how different orderings of events in narratives affect comprehension. We study the process experimentally in the next chapter. Chapter 3

Experimental Study of Narrative Comprehension

Generating narratives of varying orders of non-linearity without affecting the viewer compre- hension, warrants the study of Narrative comprehension. This chapter studies how narrative comprehension is affected by varying the order of information being presented. Previous studies in this topic employed narratives of small duration and none of them have tested on commercial feature-films. We use the domain of film because it is more sensual and is more direct than that of a text. The results of this study are analyzed using the framework of the Event-Indexing model. Though the event-indexing model had been originally postulated for text comprehension, studies have proved that it could be extended for film as well (Magliano et al., 1996, 2001).

3.1 Theory

Film is a temporal art. The events comprising the story in a film medium are portrayed with respect to time. When the timeline of a chronological ordering of events has been disrupted by the use of narrative methods, there might be a disruption in causality that the events convey. Such a narrative that has its events disrupted in their chronology of occurrence is termed as non-linear. For example, let E1, E2, E3, E4 be the events of a story in which E2 follows E1 in time, E3 follows E2 and so on. Consider that the protagonist is killed in E4. Employing a flash-forward in the narrative might give us the order of E1-E4-E2-E3. In E4

23 Chapter 3. Experimental Study of Narrative Comprehension 24 we see that the protagonist has been killed, but in E2, whose screen order now is after E4, he is alive and performing some actions. Such a change in the order of the events might affect the ongoing causality of the content being presented.

Considering this relation of causality with change in the temporal order of the events, we define a non-linear narrative as one in which the events are not-chronological and causality cannot be inferred by any other attributes, at the outset.As more information is revealed, the causality between the initial two events might become evident. Without additional information it is difficult to process or integrate these two events. In the context of the event-indexing (EI) model, this notion of non-linearity implies that there are lesser mental connections between the events in a non-linear narrative than that of a chronological one. And according to the processing load hypothesis, lesser the number of links common between two events, the more difficult it is to integrate them into coherent structure. Using the framework of the EI model, we say that the degree of linearity or non-linearity of two events is defined by the number of links common between them. Lesser the number of connections, greater is the non-linearity of the material.

An important aspect to consider when reordering the events of a story is the level of granu- larity one needs to consider an event to be. According to film grammar, a shot is the basic unit in a narrative. Series of shots comprise a scene and series of scenes result in a sequence. One could consider an event in the film to be at a micro-level of a scene or at a macro-level of a sequence, or a series of sequences. Hierarchically, a sequence conveys more information or meaning than that of a scene and a scene conveys more meaning than a shot. When the material has been rearranged randomly, there is a danger that it would impede ongoing action or a dialogue in a shot because there is no notion of where a would occur. Such random reordering of the film would definitely decrease comprehension. Therefore, besides the order of presentation, one also needs to consider the amount of material that is consid- ered as an event (level of granularity) and then perform the rearranging of these events in order to study the effect of non-linearity.

Though empirical studies were done in the past to analyze the order effects on narrative comprehension, none of them had laid an emphasis on how the material has been scrambled. Cowen(1984) used an experimental film of 2 minutes to test the impact of different ordering of the scenes in the narrative. The study showed that recall of information in the narrative decreased with an increase in the non-linearity of the content. Also recall was found to Chapter 3. Experimental Study of Narrative Comprehension 25 decline in the narrative when it employed the use of temporal shifts like flashbacks and flash-forwards, than with respect to the simpler linear version (Ohtsuka and Brewer, 1992). On the contrary, in their research on memory of an audio-visual narrative, Furman et al. (2007) have taken a situational-comedy episode of 27 minutes and reported a peripheral result that scrambling the order of the events at random in the video did not have any impact on the recall of content with respect to the original version. Though fair amount of research has been done to study the order effects on narrative comprehension in the past, none has been tested on actual film material and none of them have tested order effects in the context of the recently developed and much studied Event-Indexing model. Our interest was to study the order effects at a large scale by employing a full-length commercial feature film and observe if there would be any change in the comprehension of the viewer by using the theory posited by the Event-Indexing model. Though empirical studies were done in the past to analyze the order effects on narrative comprehension, none of them laid an emphasis on how the material has been scrambled or at what level they introduce a narrative discontinuity.

Using the processing-load hypothesis, we assess the non-linearity of the material by the number of common links between two consecutive scenes or sequences. We hypothesize that narrative comprehension decreases with increasing non-linearity of the material and that non-linearity is dependent not only on the order in which the information is presented but the level at which it is scrambled. The lesser the level at which information is scrambled the more is the non-linearity of the material.

3.2 Method

We have taken the film, Pulp Fiction (Bender and Tarantino, 1994) for our study and have designed two experiments that would test our hypothesis. Experiment 1 introduces narrative discontinuities or time shifts at a sequence level while the study in Experiment 2 uses content from the movie where time shifts are introduced at a scene level.

3.2.1 Experiment 1

The film was edited in two different ways to increase the degree of non-linearity in the nar- rative. Chapter 3. Experimental Study of Narrative Comprehension 26

Participants Thirty-nine participants consisting of postgraduates and undergraduates from the IIIT-Hyderabad volunteered for the experiment. All of them are in the age range of 21 to 29 years. They are of Indian nationality and were comfortable with watching En- glish films since at least three years.

Materials Participants viewed one of the three versions of the film, Pulp Fiction. This could be classified as in the crime and the genre. Table 3.1 shows the brief structure of the original film without any manipulations. This non-linear version of the film has been

Table 3.1: Overview of the original (Non-linear) version of the film.

Episode 1 : Pumpkin and Honeybunny dis- cussing to rob a restaurant. Episode 2 : Jules and Vincent go to a house to retrieve their boss’ briefcase. Episode 3 : Jules and Vincent come to a bar to return the briefcase. Their boss gives Butch some money and asks him to lose a boxing match. Vincent takes his boss’ wife on a date. Episode 4 : Butch has violated an agreement with his boss and have to escape. He meets Vin- cent in his apartment and kills him. Episode 5 : After collecting the boss’ briefcase, Jules and Vincent get into a situation where Vincent accidentally kills a person in their car. They go to a friend’s place to clean up. Episode 6 : After getting out of the situation they go to the Diner to have breakfast. Pumpkin and Honeybunny initiate the robbery. They get threatened by Jules and leave the place. Jules and Vincent leave the Diner.

taken and was edited in such a way to produce two other versions: A linear counterpart (events that follow the chronological timeline of the story) and a highly non-linear or jum- bled up version of the film. Given a non-linear version, it was easy to know where to cut the film and join the segments in order to make events to follow a linear timeline. Arranging the episodes of the original film in the order of 2-5-1-6-3-4 would yield the linear version of the film. A highly non-linear version of the film (the jumbled version) was also created by edit- ing the original. The order of the events of this version with respect to the original version is 1-2-4-5-3-6. The deviation from the linearity in each of these versions was assessed by the Chapter 3. Experimental Study of Narrative Comprehension 27 theory of the Processing Load hypothesis. That is the number of common indices between consecutive events of the jumbled version is lower than that for the non-linear version. Our intention was to make the segments of the movie as incomprehensible as possible. So, for the jumbled version, we have achieved a degree of non-linearity that was higher than that of the original, non-linear version. With the help of editing, we had obtained three versions of the film (including the original) with varying linearity for carrying out our experiment.

A questionnaire was designed to assess the recall of the participants on various aspects of the movie. The questions covered all the segments of the movie and were broadly categorized into three categories: Questions that test the Details of the film, that ask for the description of the events of the film, questions that test the reasoning of the events and the actions of the characters in the film. Six independent raters were asked to categorize the questions into the three mentioned groups. A high agreement of 82% was obtained among the raters. Three questions that were rated ambiguous by a majority of the raters have been omitted from the questionnaire.

Procedure Students at the University were informed about an experiment on movie-watching and that there would be monetary incentive for those who participate in it. Participants were told that there would be two parts to the experiment that they have to watch a movie and have to answer a few questions based on the movie. The movie was shown in a quite room with each participant watching the movie in his own monitor and with a pair of head- phones. A maximum of three participants watched the movie in the room at any time. Subtitles were not included for any participant as they might bring any external variables into the experiment. Care was taken to select the participants for the experiment who had not seen the movie previously. All the participants who started the experiment finished it successfully and were given an incentive of Rs. 50 for their time.

Participants were assigned, randomly to each of the three groups that correspond to the three versions of the film and they had no knowledge of which version they were watching. Participants were not given any instructions while starting the experiment other than asking them to watch the movie with concentration and not taking any breaks. Once the movie was finished, participants were asked to take a break of 5-10 minutes if they wished before getting started on the questionnaire aspect of the experiment. The questionnaire phase of the experiment has been started within a maximum of 20 minutes from the time the movie had finished. Participants were given instructions on how to answer the questions and were Chapter 3. Experimental Study of Narrative Comprehension 28 specifically reminded not to consider any spelling mistakes, grammatical errors when they were entering the answers. They were also informed to enter the answers as accurately as possible and were asked to enter ‘NA’ if they did not know the answer to any question. The questionnaire consisted of two types of questions: Multiple choice type where the participant had to just enter the option corresponding to the correct answer, and, descriptive questions such as, “Describe what happened in the Diner?” for which the participant had to type the whole response. Each question was accompanied by a picture of the character or characters in the movie, which the question was referring to. Also, the order of the questions presented in the questionnaire was in accordance with the order of the scenes in the version of the movie they had seen. Questions were presented to the user with the help of a software, PsychoPy (Peirce, 2007) that aided in designing the questionnaire.

Participants were asked to rate the movie on a scale of 10 once they have completed all the questions. Recall was measured by considering the individual score of the participant over the total score of the questions comprising the questionnaire. Multiple choice questions were scored directly based on the option entered by the participant. A rating of 1 was given for a correct answer and a zero for a wrong answer. Also, answers for which ‘NA’ was entered were also given zero marks. The descriptive questions for which the participant had to type an answer were coded based on the keywords comprising the answer. Each of these keywords were given a rating of 1. These keywords expressed the answer to the question being asked at an essence. The answers were manually scrutinized to see for the presence of the keywords. If the response of a participant conveyed the same meaning as that of the intended keywords, marks were awarded without any reduction.

3.2.2 Results

All the data obtained were analyzed by one-way analysis of variance with the order of the scenes of the narrative as the Independent variable.

We wished to see if people would remember more content when the information presented to them is in an out-of-chronological fashion instead in a canonical linear manner. In this context, we hypothesized that the recall of information from the participants would decrease as the degree of non-linearity in the narrative is increased. From the results we found out that there was no significant difference in the recall scores of the participants in any of the groups. The mean recall scores for the linear, non-linear and jumbled groups were 58.846% Chapter 3. Experimental Study of Narrative Comprehension 29

(SEM = 2.219), 54.903% (SEM = 5.310) and 56.442% (SEM = 2.311) respectively. The

Table 3.2: Mean recall scores of the participants (in percentages) in different categories of the questions across the groups

Descriptive Detail Reasoning Linear 58.78 72.00 42.50 Non-Linear 52.93 67.43 42.14 Jumbled 55.36 70.57 40.36

difference among the means of the recall scores of the three groups was found not significant with changing the order of the scenes. The mean recall scores for the different categories of questions (Descriptive, Detail and Reasoning type) across the groups are presented in Table 3.2. The F values for the descriptive, detail and reasoning questions across the groups are F (2, 27) = 0.450, ns, F (2, 36) = 0.277, ns and F (2, 27) = 0.103, ns respectively. That is we have to accept the Null hypothesis that there is no significant difference in the mean recall scores of each of the groups and also among the different categories of questions across these groups. Fig. 3.1 shows the graph of the means of the total Recall scores across the groups. Mean recall scores in each of the categories of Detail, Descriptive and Reasoning questions

Figure 3.1: Average recall scores(in percentages) of the participants across the groups across the groups were considered and no significant difference was reported. The difference among the means of these scores across the groups among the types of questions was not significant with changing the order of the scenes. Thus in all of the categories, there was no significant effect on the order of the scenes or non-linearity on the recall scores. Chapter 3. Experimental Study of Narrative Comprehension 30

3.2.3 Experiment 2

This experiment was conducted to study the effect of changing order of the scenes in a small segment of the movie. The average time until the next discontinuity occurs is much lesser than that of the full-length movie. An unruptured segment from Pulp Fiction has been taken and is scrambled for this experimental purpose.

Participants Forty-five people comprising of graduate students and undergraduates who have not seen the film before, from the IIIT-Hyderabad have volunteered in this experiment. The participants are in the age group of 21-35 years.

Materials A segment (‘The Gold Watch’) of length 42 minutes from the movie has been taken and has been scrambled. As with the material in the previous experiment, there were five discontinuities we also decided to have five discontinuities in this segment. We asked six judges who have already seen the movie to identify scenes in this segment. Four scenes were identifed with an agreement of 73% among them. We have introduced a fifth time shift into the narrative and care was taken that this timeshift was not interrupting an on-going dialogue or an action. Once we identified that a certain action was completed we have introduced the timeshift. Table 3.3 presents an overview of the actions in the chronological version of this segment.

Table 3.3: Overview of the linear version of the segment, The Gold Watch from the film.

Episode 1 : Marsellus tells Butch to lose the boxing fight. Episode 2 : Butch goes to the boxing fight, violates his boss’ agreement and then escapes in a taxi. Marsellus sends men after Butch. Episode 3 : Butch leaves to a motel to meet his girlfriend, takes a nap and discovers that his watch is missing. Episode 4 : Butch goes to his apartment to get his watch. He kills a gangster there. On his way back, Butch encounters Marsellus and they two get in a fight in a parlor. A cop ties them down. Episode 5 : Butch escapes and also saves Marsellus from the cop. Episode 6 : Marsellus asks him to leave town and that everything is settled. Butch agoes to his motel and lraves town with his girl friend.

The segment has been edited in two different ways with the help of the processing-load of the Event-Indexing Model and three versions of the segment with increasing degree of non- linearity were obtained. We call them the Linear, Quasi-Linear and the Non-Linear groups. Chapter 3. Experimental Study of Narrative Comprehension 31

The Quasi-Linear group is deviated from the chronological version but its deviation is lesser than that of the Non-Linear group. Arranging the episodes of the original film in the order of 1-2-4-6-3-5 would yield the quasi-linear version of the film, while the order 1-3-6-4-2-5 would give the non-linear version.

Procedure The procedure was identical to that of Experiment 1.

3.2.4 Results

We wanted to know if there was any change in the Recall scores of the participants when the order-effects are introduced in a shorter movie version with the motivation index being scrambled across the episodes comprising the segment of the movie. The data from the experiment after being standardized/quantized has been analyzed by a one-way analysis of variance. A statistical significance of F (2, 42) = 6.845, p < .005, ω = 0.61 has been obtained. This implies that there is a significant difference in the recall scores among the different versions of the movies. Fig. 3.2 shows the graph of the means of the total Recall scores across the groups. Post-hoc tests were conducted to investigate deeper in to the

Figure 3.2: Average recall scores(in percentages) of the participants across the groups, when material is scrambled at a micro-level. relationships among the groups. It was found that there was a significant difference between the Linear - Quasi-Linear group (MD = 14.190, SEM = 4.025, p < 0.01) and also between Linear-Non-Linear groups (MD = 15.299, SEM = 5.106, p < 0.05). The results state that Chapter 3. Experimental Study of Narrative Comprehension 32 the recall scores in the linear version of the movie are significantly better than that of the scores in the quasi-linear and the non-linear versions. However there was no significance observed between the quasi-linear and the non-linear groups; (MD = 1.108, SEM = 4.638, ns)

3.3 Discussion

In support of our hypothesis, there was a decline in recall values when the material was scrambled at a scene level, but no significance was observed in the values of the participants who watched the full-movie. The results indicate that both the order in which the infor- mation is presented as well as the level at which the information has been scrambled might have an effect on the recall scores of the participants.

Each macro level segment of the movie has a structure of events that could be considered as having a setup, complication, confrontation and resolution. The segments of the movie were rearranged without tampering this internal structure of the segments. Research indicates that when a well structured story is scrambled, comprehension of the readers is not affected as judged by the similarity in summaries produced by them. (Kintsch et al., 1977). In the experiment where the structure of the sequence was tampered by rearranging the comprising scenes, comprehension was decreased. This result is in accordance with the previous research of (Cowen, 1984, Ohtsuka and Brewer, 1992) where the material was scrambled randomly.

From an Event-Indexing model standpoint, in Experiment 1, there was no change in the dimensions of time and motivation through out a sequence as the events of the sequence were kept intact. Once a segment/sequence has been completed, a discontinuity occurs in the narrative which results in a change of the dimensional values. Where as in the material that was scrambled at a much smaller level, there is also change in the time and motivation dimensions.

The task of narrative comprehension would require the viewer to understand the character motivations, the actions performed in accordance with the motivations, the causality of the events occurred in the narrative etc. In Experiment-1 the viewer would require to understand how the sequences are related to each other as a whole. That is one has to infer the causality of the sequences. A sequence culminates in the completion of a major event (usually in the realization of a sub-goal) of a story. The viewer would know the goal Chapter 3. Experimental Study of Narrative Comprehension 33 or sub-goal of the character while processing a sequence as the events of the sequence are intact. The motivations of the in each segment are presented in Table 3.4. In Experiment-2, as the sequence is ruptured, the viewer would have to infer why a character is performing an action (the character’s intention) besides how those actions are related.

Table 3.4: Motivations of the main character in each segment of the film considered in our study.

Segment Main character Motivation 1 Pumpkin, Honeybunny To rob a restaurant. 2 Jules, Vincent To retrieve their boss’ brief- case. 3 Vincent To take care of the boss’ wife. 4 Butch To escape the city 5 Jules, Vincent To get out of the situation of the deadbody in the car 6 Jules To return the briefcase to their boss.

Generally, actions of a character follow from his goals. But here, as the goal/intention itself is not clear in the first place, the viewer has to infer it from the different actions performed by the character across scenes. The viewer might have to see a certain amount of information in order to infer the primary motivation of the character. The processing load to required to monitor the goal information in order to infer it would increase with the amount of content that needs to be processed. This increase in processing load is the additional processing overhead that is manifested when the goal information is not clear, as in Experiment-2. We propose that the decline in the recall of information in the non-linear version of the film is due to this additional processing overhead that comes into play when the motivation of the characters is not clear. Besides, the film presents the stories of three people and how their lives intertwine. Due to the multiple protagonists in the story, viewers have to monitor multiple goals (Magliano et al., 2005) and there is an additional overhead in this aspect.

Immediate Integration Principle states that a new event in a narrative has to be immediately integrated with old information, otherwise there would be a decline in the comprehension of the narrative (Ohtsuka and Brewer, 1992). This principle however was not observed on the material that was scrambled at a sequence level. Theoretically, there were absolutely no common indices among the first three segments of the film in the jumbled version. There Chapter 3. Experimental Study of Narrative Comprehension 34 was a delay of almost about 58 minutes for the segment that could be integrated to occur in the narrative.

Inability to form a chronological version of the narrative with increasing non-linearity has been reported by the work of Ohtsuka and Brewer(1992), Speer and Zacks(2005). Our results have showed that the viewers in the quasi-linear and the non-linear groups of Ex- periment 2 were unable to infer events of the narrative in their chronological fashion. This is established by the poor responses to the questions that dealt with the temporal aspect of two events there were separated by a discontinuity. For example, From where did Butch come to the motel?, Where did Butch go after escaping from the parlor. The participants in the Linear group answered these questions with an average score of 88.62% while the average score in the quasi-linear and the non-linear groups is 36.35% and 41.45% respectively. Based on our result we suggest that, with increasing non-linearity it is difficult to draw inferences on the temporal aspects of the narrative like what event happened before and what after. This result establishes the limitation of memory recall upon processing information in a non-linear manner.

The EI model is based on the assumption of dimensional equality that all the dimensions that represent the mental model have equal importance in narrative comprehension. This study refines this assumption by identifying the importance of the motivation/goal dimension in narrative comprehension. We arrive at a conclusion that monitoring the goal is much difficult when the material is scrambled at a scene level than at a sequence level. We affirm the previous findings on narrative comprehension and extend it to the context of film narrative and add that not only is narrative comprehension dependent on the degree of non- linearity in the material but also in the manner in which the material has been scrambled. This research on the effect of presentation order on narrative comprehension is useful for our study on computational narrative generation. In order to generate narratives with different presentation orders, we need to understand how presentation order affects comprehension. Chapter 4

Generating Non-linear Narratives

This chapter talks about the development of an algorithm for generating narratives by achieving narrative variation in order. We leverage the results obtained from our experimen- tal study on Narrative Comprehension along with previous research to generate narratives that have their events presented in out of chronology.

4.1 Problem Revisited

From a narrative generation viewpoint, a fundamental question is what event has to be presented next in a story. This problem could be looked as, given a set of events in a story, what is the order in which the events have to be presented to the viewer that would meet the goals of the author (e.g generating suspense, effectively conveying a point of the story etc.). Our goal here is to generate narrative order that is more engaging than its chronological counterpart.

There would be many ways in which the events of the story could be presented, in terms of order. From a computational perspective, if a story comprises of n events, the number of possible ways of ordering them is n!. Randomly generating an order from the n events and presenting it as a narrative, is not practical. The ordering generated might be too unconventional for the author and also incomprehensible for the audience. Also, out of all the possible orders, most of them might require a lot of processing load on the reader for comprehension. An attempt in comprehending such highly scrambled narratives might lead

35 Chapter 4. Generating Non-linear Narratives 36 the comprehender to frustration as he/she is unable to make meaning of the story with his/her limited processing capabilities and memory.

As addressed by the research of Montfort(2007b), generating a narrative order to achieve achrony could be done by randomly ordering the events by placing a uniform probability distribution on it. Also quoted in the research,

A narrator that orders events uniformly at random is probably confusing for all practical purposes.

Previous studies report a decrease in comprehension when the events of a narrative are scrambled randomly (Cowen, 1984, Ohtsuka and Brewer, 1992). Therefore, we have to find the subset of narratives that do not significantly affect the ease of comprehension of the material even with changing the order of presentation of events.

4.2 Obtaining Engagement by Narrative Variation in Order

Cognitive engagement could be manifested by altering the manner in which information is processed (Brewer and Lichtenstein, 1982, Busselle and Bilandzic, 2009). A change in the order of events from chronological order might increase cognitive engagement.

Assuming that the content of the story is familiar to the audience (for e.g a household drama), in a non-linear narrative the viewer has to stay more focused than that of a linear version, in order to understand what is going on in the story. This is because the causality of the structure of events in a non-linear narrative has been broken with rearrangement of events. If the material is more scrambled then there is a danger that the focused viewer might lose interest or get frustrated with the presentation and stop watching it. On the other hand, though presenting the same information in a linear order will take away the load, it might become too easy for the comprehender to process and his attention towards the material is not fully used. One aim of this work is to find out the set of those narratives that demand more processing load of the viewer and yet not affecting the ease of comprehension. That is the viewers should not feel any difficulty in comprehending such narratives.

The cognitive load of processing the narrative could be increased by changing the order in which the information is being presented to the reader or a viewer. Events narrated Chapter 4. Generating Non-linear Narratives 37 in a chronological order have the least cognitive load for processing because there are no discontinuities in the causality of the actions, as each action follows from a previous action. Temporal rearrangement of the events alters the causality of the actions (Branigan, 1992) or more specifically might delay the causal inferences one has to make for comprehending the story. Therefore, deviating the presentation of events from their chronological order of occurrence will require more processing on the viewer to make inferences. That is, by varying the presentation of events, we are manipulating the germane cognitive load (Paas et al., 2003) hoping to achieve a gain in the focus a person puts towards comprehending the narrative. In effect the engagement in our case is derived from the process of comprehension.

4.3 Defining Non-linearity

While there is no functional definition of non-linearity, narratives that have their events presented in out of chronology are termed non-linear (or achronous). Authors may use this style of narration to effectively convey a story or as a stylistic means. As mentioned in the previous chapter, we define a non-linear narrative as one in which the events are not-chronological and causality cannot be inferred by any other attributes, at the outset. This is a useful and a generic definition of non-linearity. Narratives that have flashbacks or flash-forwards, syllepsis etc. can all be derived from this definition. For example, a flashback narration has also certain events presented in out of chronological order, however we could infer a relation between the current event and the event of flashback. If we could solve the problem of generation at a generic level of narrative variation, I believe that similar principles could be applied to generate narratives of specific variation.

Degree of non-linearity could be termed as the amount of deviation of a given order of events from its chronological order. We assess the degree of non-linearity that is incurred by rearranging the order of events in terms of the processing load that one requires to comprehend the material. Generating an order involves knowing what event has to be presented next in a narrative. The degree of non-linearity is increased if the next event being presented is not relevant at all to the previously presented events, at the outset. The narrative in retrospect might make sense to a reader. Ability to generate an order of events corresponding to a desired non-linearity requires calculating the processing load among the events. Chapter 4. Generating Non-linear Narratives 38

Processing load between two events, according to the EI model’s Processing Load hypothesis, is defined as the function of the number of dissimilar dimensional links. That is, more the number of dissimilar dimensional links between two events, the more difficult it is to process them. We use this hypothesis to define degree of non-linearity between two events as the number of dissimilar dimensional values between them. We extend this definition that the degree of non-linearity of a certain order of events (or a narrative) as the cumulative sum of the processing load values among the pairs of events, that occur in order in the narrative.

4.4 Design of the Algorithm

This section presents an algorithm for computational generation of non-linear orderings in a narrative. The algorithm takes in a chronological order of events, a fabula in its partial- order plan representation, and outputs a narrative that has its event order according to a user specified level of non-linearity (low, mid or high).

Unlike many SGAs, a planning-based reader model is not employed for approximating the reasoning process of the narratee. Instead, the principles of Event-Indexing model are used to assess the comprehension of the narratee by calculating the ease of integrating an incoming event with one already in memory. The Processing Load Hypothesis is used to measure this level of difficulty in comprehending a narrative.

A narrative has a goal hierarchy (Chatman, 1980). A protagonist in order to realize his goal of the story has to first achieve his sub-goals. These sub-goals are hierarchically related to the final goal of the story. There is also an event hierarchy in a narrative that is related to the goals of the protagonist. A sequence is a series of events that culminates when the sub- goal of a protagonist has been met. According to the results (cf. Sec. 3.2.2) obtained in our experiments on Narrative Comprehension, we group the events of a story into sequences and re-order these sequences by keeping the events in them intact. We adopt such a grouping because results indicate that the comprehension was not affected when the events were grouped at a sequence level, by maintaining the structure of the story to an extent.

Also, by grouping the narrative into sequences, the search space to obtain the narrative of the desired level of non-linearity is reduced. If a narrative has n events and is grouped into m sequences (where m < n), then the possible number of orderings are m! which would be less than the initial n! orderings. The reduction in search space depends on the number Chapter 4. Generating Non-linear Narratives 39 of events in each of the sequences. In our research, we assume that the sequences in the input fabula are already marked or the user can specify what events constitute a sequence. Parsing the fabula and identifying the sequences is not in the scope of this work.

4.4.1 Calculating the Degree of Non-linearity

The algorithm calculates the degree of non-linearity among all the sequence pairs and decides on the order in which the sequences have to be presented according to the input of degree of non-linearity. We define the degree of non-linearity as the Processing-Load value (PL value) which is the number of dissimilar dimensional links among the events. When operating at the sequence level we calculate the PL-values when there is a narrative shift or when a sequence ends. If M and N are two consecutive sequences of a narrative, the processing load PL(M, N) between them is calculated between the last event of sequence M and the first event of sequence N.

Calculation of dimensional shifts

When the EI model was first proposed, the dimensions were described to be dichotomous (Zwaan, 1999). For any two events A and B in the fabula F, we however consider a dimen- sional shift to occur only if the following happens:

1. Time: Time is a continuous entity. If A and B occur in the same time frame, then they are related on the time dimension. Also, if B follows A in time, then they are related. This is calculated from the temporal ordering information, T from the fabula F.

2. Space: We consider this to be a dichotomous entity. Either the event takes place at one place or not. This is obtained from the fabula’s action description. The notion of the concept of a location is employed. If two events occur in the same building, for instance and one event occurs in the hall and the other event occurs in the bedroom, the events are considered to be occurring in different spaces (McKee, 1998).

3. Protagonist: From the plan step we could infer if the protagonist is present in both A and B. If he is not, then there is a shift in the protagonist index. Chapter 4. Generating Non-linear Narratives 40

4. Causality: This is treated as a continuous dimension. Events in story are causally connected. Though two events do not have a direct causal relation, they might be related through other events of the story. Consider, A and B to be two events separated by an event E and that A achieves E and E achieves B. In such a case, though A and B are not directly related on the causal dimension, they are still causally related by means of the event E.

We define the degree of causality between two events A, B as the minimal number of events the comprehender needs to know in order to form a causal inference between A and B. This is obtained by the length of the shortest path in the causal graph of F with A and B as the terminal points. A path connecting two events in the causal graph assesses the amount of information the reader needs to know along with the information already present in memory to form a causal inference. Though there may be many paths connecting A and B that contribute to the causal strength, the nodes on the shortest path are common among all these paths. That is in order to make a causal connection between A and B, the reader has to know at least the events along the shortest path connecting them. If l is the length of the shortest path, then the degree of causal relatedness is 1/l.

While calculating the shortest path between two events, there are nodes that are al- ready in memory of the reader and that are not yet presented. From the reader’s viewpoint it is easier to form an inference with an event already in memory rather than not forming one at all due to lack of information. In order to incorporate this we say that during inference, if a path that leads to an event in memory is encountered, it is given an edge-weight of 0.5. If the event is not yet presented to the reader, it has an weight of 1. Assigning these values as edge-weights is a methodological convenience rather than a theoretical concern. Summing up all the edge weights along the shortest path gives its length.

Reverse chronology and the effect of Causality: During the design of a narrative, considering to place events in reverse order, that follow each other chronologically in the story, leads to an additional overhead on the viewer when making a causal inference between them (Ghislotti, 2004). For example when we show an effect first and its cause later, the viewer has to re-arrange these events mentally in-order to infer causality between them (Branigan, 1992). Due to this additional processing overhead, Chapter 4. Generating Non-linear Narratives 41

we add a penalty in such cases, by saying that in order to present events in reverse order the degree of causal unrelatedness is double than that if presented in a natural way. Calculating the degree of causality is mentioned in Section 4 of this paper by giving an example.

5. Intention/Motivation: As we have grouped the events of the fabula into sequences and each sequence has a specific goal, the goal or the motivation dimension of the protagonist is constant through out the sequence. We consider the motivation changed when a new sequence begins. As the algorithm rearranges sequences, the motivation is always different for any two sequences of the fabula.

The dimensions along which A and B are related are given a value of zero and the others given a value of one. The causal dimension ranges from 0 to 2 and is a continuous entity. Summing the values of all the five dimensions gives the PL value or the degree of non- linearity between two events. We use this value of an event pair and the input degree of non-linearity to decide on what event to present next to the viewer so that level of non- linearity is maintained at a threshold, the specified input level.

The algorithm calculates the PL-values among all pairs of sequences and ranks them in ascending order. The interval of non-linearity is taken from low to high where low is defined to be middle-rank/2, mid is middle-rank and high is middle-rank + high-rank / 2. After obtaining the PL-values of all pairs of sequences in the fabula F, the algorithm selects the pair of sequences that correspond to the input level of non-linearity and is added to the sjuzhet S. After selecting this pair, PL values of the sequences remaining in F are recalculated with respect to the starting and ending sequences of S. This recalculation is done because the degree of causality changes given the events in S are in reader’s memory. The sequence e corresponding to the lowest values among PL(e, Head(S)) and PL(Tail(S), e) is then selected and added to S. The algorithm could also select other levels of PL-values instead of the lowest value. This selection effects the degree of non-linearity produced for the given level of low, mid or high. For the moment, the algorithm selects the lowest PL-values among the sequence pairs. Making such selection will output narratives of minimal level of non- linearity whose order is defined by the first sequence pair that is included in S, which is selected based on the input level of non-linearity. The above process is repeated until all the sequences of F are added to S. The algorithm is presented in 4.1. Chapter 4. Generating Non-linear Narratives 42

Initialization: fabula F, level of non-linearity v (low, mid or high).

Termination: If F is empty or has only one sequence.

1. For all sequence pairs in F, calculate the PL values and rank them.

2. Select the sequence pair that corresponds to the input level of non-linearity, v. low is defined as mid-rank/2, mid is mid-rank and high is defined as mid-rank + high-rank / 2.

3. Add the sequence pair to S.

4. While F 6= 0; do

• F ← F − S; • For all events e in F, calculate PL(e, Head(S)) and PL(Tail(S), e) where Head(S) gives the first sequence of S and Tail(S) gives the last sequence of S. comment: If S=ABCD, Head(S) is A and Tail(S) is D. • Select the event e that has the lowest PL-value among these values and add it to S.

Figure 4.1: Algorithm for event selection and presentation ordering based on the degree of non-linearity

4.5 Working Example

In this section we present the working of our algorithm that generates a narrative order with a mid-level degree of non-linearity by taking the fabula as input. The fabula F we consider is taken from (Cheong and Young, 2006) and is presented in 4.2. We also present the structure of causal links among the plan steps of the fabula in 4.3 adopted from (Bae and Young, 2008). According to Bae and Young(2008) Plan steps in the fabula are denoted by circles. Step 1 corresponds to the starting event in the story while step 20 is the ending event. The directed arcs between two steps denote the causal link between the effect of the first step and the pre-condition of the next.

Table 4.1: PL-values of sequence pairs in the fabula calculated at the first run of the Algorithm

- A B C D A - 1 3.5 3.5 B 5.42 - 1 3 C 5.7 4.33 - 2 D 5.43 4.33 4.2 -

According to the fabula F, we assume that the sequences in the story end at events 6, 9, 16 and 20. We denote these four sequences by A, B, C and D respectively. Given this Chapter 4. Generating Non-linear Narratives 43

Background: The lunatic supervillain known as Jack has been developing biological weapons of devastating proportions. To accomplish the final stages of weapon development, he kidnapped the famous scientist, Dr. Cohen, and brought him to his private fortress on Skeleton Island. Jack expected that the FBI would soon send Smith, their top agent, to rescue Dr. Cohen. To keep the troublesome Smith out of his hair, Jack ordered his own agent, Erica, to monitor Smith and capture him if he is assigned to Dr. Cohen’s rescue operation.

Story: (1) Erica installs a wiretap in Smith’s home while he is away. (2) Erica eavesdrops on the phone conversation in which Smith is given the order to rescue Dr. Cohen. (3) Erica meets with Smith. (4) Erica tells Smith that her father was kidnapped by Jack and taken to Skeleton Island, and she asks Smith to save her father. (5) Erica gives Smith the blueprints of Jack’s fortress, with her father’s cell marked. (6) Erica provides Smith with a boat for transportation to Skeleton Island. (7) Before going to the island, Smith hides a diamond in his shoe. (8) Smith goes to the port containing Erica’s boat. (9) Smith rides the boat to Skeleton Island. (10) Smith sneaks into the cell marked on the map containing Erica’s father. (11) Jack and his guard capture Smith as he enters the cell. (12) The guard disarms Smith. (13) The guard locks Smith into the cell. (14) Smith bribes the guard with the diamond in his shoe. (15) The guard unlocks the door. (16) Smith leaves the cell. (17) Smith sneaks into the lab where Dr. Cohen is captured. (18) Smith fights the guards in the lab. (19) Smith takes Dr. Cohen from the lab. (20) Smith and Dr. Cohen ride the boat to the shore.

Figure 4.2: An input Fabula adopted from Bae and Young (2008) fabula, our algorithm calculates the PL-values for all the sequence pairs as presented in Table 1. Ranking the PL-values and calculating the mid-rank by taking the floor value of low-rank+high-rank/2 gives 5 and 4.2 is the PL-value corresponding to it. The sequence pair corresponding to this mid-rank is DC. We add this sequence pair to the sjuzhet being generated, S and remove the events D, C from F. When we have more than one sequence corresponding to a rank, we select at random any sequence and add it to S. For the remaining events e in the fabula, the algorithm now calculate the PL values with respect to the head and tail sequences of S, given S. That is, PL(e, D) and PL(C, e) are calculated for all the remaining events in F. Ranking these PL-values and selecting the corresponding PL- value value with least event, we have PL(B, D). Therefore B is now removed from F and added it to S at the beginning of D. S now is BDC. By repeating the above procedure, we obtain the completed narrative order as ABDC which corresponds to the mid-level degree of non-linearity.

Calculation of the degree of causality We present here the calculation of the degree of causality between two events that are being considered for presentation in reverse order, for example, BA. When we consider this order, a causal shift occurs between the last event of B, step 9 and the first event of A, step 1. The algorithm checks to see how we could infer 1 from 9, given the events in S. Since the sequences are temporally reversed and as the causal structure is a directed graph, we cannot reach step 1 from step 9. We therefore consider the Chapter 4. Generating Non-linear Narratives 44

Figure 4.3: Graph showing the causal links among the events of the fabula. Adopted from Bae and Young (2008) reverse of reaching step 9 from step 1. If l denotes the length of the shortest path from step 1 to step 9, then the degree of causality between two steps is 1 / l. In this case 1/3.5 (0.5 is the edge-weight for edges connecting the nodes that are already in memory)/. Or, the degree to which the events are causally unrelated is 1 - 1/3.5, which is 0.71 approximately. But as the events were reversed for calculating the shortest path, we add the penality of this action by doubling the degree of unrelatedness. Therefore our value of causal unrelatedness for the order BA is 1.42.

4.6 Implementation

A prototypical implementation of the algorithm was performed. Users have to mark events of their choice, in the fabula, as sequences and have to input the level of non-linearity (low, mid or high). The system considers these event markers as the sequences of the story and performs re-arrangement of these sequences. Finally, the system outputs an order of these sequences according to the level of non-linearity entered. Chapter 4. Generating Non-linear Narratives 45

4.7 Evaluation

In order to evaluate the narratives generated by our system on the basis of engagement, an experimental study was undertaken. Also of interest, was the amount of non-linearity that could be introduced without increasing the difficulty in comprehension.

4.7.1 Method

Participants 170 participants from IIIT-Hyderabad and the neighbouring colleges were considered for the experiment. Participants age ranged from 19 to 45 years.

Materials A fabula that was generated by a planning-system, was adopted for our exper- iment. We applied our algorithm to vary the order of events of the story in several ways. Six independent raters were asked to read the story and identify the sequences of the story. Four sequences with an agreement of 78% were identified. The story grouped by the iden- tified sequences, we ran the algorithm to obtain narrative variation in orders of low, mid and high. We considered six different type of narratives for testing. Three of the narratives correspond to the minimal levels of low, mid and high level non-linearity. The other three are obtained by selecting the minimal level of non-linearity according to the input (low, mid or high) and while generating the order of next sequences, the sequence pairs with the 2nd lowest Processing-Load values were select. We call these to be narratives of 2nd-selected non-linearity. The low level non-linear narrative generated by our system turned out to be the chronological ordering of the story.

The plans generated by the Crossbow system were converted into plain English. We have adopted the language of the story as-is. The stories that were re-ordered by the algorithm were slightly modified by the experimenters to include discourse elements. For example, words like ‘Earlier’ and ‘Later’ were used before the sentences in order to convey a temporal shift in the story, whenever one occurred due to the re ordering of the sequences. Sample narratives obtained from the output of our system are presented in Figures 4.4 and 4.5

Procedure Students at the University were informed about the experiment that they would have to read a story and answer a few questions. Participants accessed a web page to take part in the experiment. The web-site provided instructions on reading the story. Instructions were also provided that there will not be any questions on comprehension of the story content Chapter 4. Generating Non-linear Narratives 46

Background: The lunatic supervillain known as Jack has been developing biological weapons of devastating proportions. To accomplish the final stages of weapon development, he kidnapped the famous scientist, Dr. Cohen, and brought him to his private fortress on Skeleton Island. Jack expected that the FBI would soon send Smith, their top agent, to rescue Dr. Cohen. To keep the troublesome Smith out of his hair, Jack ordered his own agent, Erica, to monitor Smith and capture him if he is assigned to Dr. Cohen’s rescue operation.

Story: While making arrangements to go to the Island, Agent Smith hides a diamond in his shoe. Smith goes to the port containing a boat and rides the boat to get to Skeleton Island. He sneaks into a cell marked on his map containing the location of Erica’s father. Jack and his guard capture Smith as he enters the cell. The guard disarms Smith and he locks him into the cell. Smith bribes the guard with the diamond in his shoe. The guard unlocks the door and Smith leaves the cell. Smith sneaks to the lab where Dr. Cohen is captured. He fights the guards in the lab and Jack. Smith takes Dr. Cohen out from the lab and they both ride the boat to the shore. Earlier when Jack orders Erica to monitor Smith, she installs a wiretap in Smith’s home while he is away. She eavesdrops on the phone conversation in which Smith is given the order to rescue Dr. Cohen. Erica meets Smith. She tells him that her father was kidnapped by the evil Jack and taken to Skeleton Island. She pleads Smith to save her father and he agrees. Erica gives him the blueprints of Jack’s fortress, with her father’s cell marked. She also provides Smith with a boat for transportation to Skeleton Island. Figure 4.4: Low level non-linear narrative (2nd select) obtained after running the algo- rithm

Background: The lunatic supervillain known as Jack has been developing biological weapons of devastating proportions. To accomplish the final stages of weapon development, he kidnapped the famous scientist, Dr. Cohen, and brought him to his private fortress on Skeleton Island. Jack expected that the FBI would soon send Smith, their top agent, to rescue Dr. Cohen. To keep the troublesome Smith out of his hair, Jack ordered his own agent, Erica, to monitor Smith and capture him if he is assigned to Dr. Cohen’s rescue operation.

Story: In the Skeleton Island, Smith sneaks to the lab where Dr. Cohen is captured. He fights the guards in the lab and Jack. Smith takes Dr. Cohen out from the lab. They both ride the boat to the shore. Earlier, Erica installs a wiretap in Smith’s home while he is away. She eavesdrops on the phone conversation in which Smith is given the order to rescue Dr. Cohen. Erica meets Smith. She tells him that her father was kidnapped by the evil Jack and taken to Skeleton Island, and she asks Smith to save her father. Erica gives him the blueprints of Jack’s fortress, with her father’s cell marked. She also provides Smith with a boat for transportation to Skeleton Island. While making arrangements to go to the Island, Smith hides a diamond in his shoe. Smith goes to the port containing Erica’s boat. Smith rides the boat to get to Skeleton Island. He sneaks into the cell marked on the map containing the location of Erica’s father. Jack and his guard capture Smith as he enters the cell. The guard disarms Smith and he locks him into the cell. Smith bribes the guard with the diamond in his shoe. The guard unlocks the door and Smith leaves the cell.

Figure 4.5: High level non-linear narrative (minimal level) obtained after running the algorithm Chapter 4. Generating Non-linear Narratives 47 and that they are to read the story freely or casually but with concentration. Six of the narratives generated by our systems were randomly presented to the participants accessing the web-page. They were unaware that a narrative has been re-ordered for experimental purpose and was being presented to them. A subject had to read the narrative that was presented and answer a few questions on his/her reading experience. The questions were taken from the Narrative Understanding sub-scale of the overall Narrative Engagement scale developed by (Busselle and Bilandzic, 2009). This scale measures the participants’ subjective experience of the ease of comprehension (or the lack of difficulty in comprehension) of the narrative presented to them. The engagement scale measures the participants’ degree of agreement towards a scale-item. We have modified the scale-item labels to measure the degree of difficulty instead.

• The narrative engaged me at a level 1-Low; 5-High

Questions adopted from the Narrative Understanding sub-scale of Busselle (2009): • At points, I had a time making sense of what was going on in the story 1-Easy; 5-Hard

• My understanding of the characters’ actions is 1-Clear; 5-Unclear

• I had a time recognizing the thread of the story 1-Easy; 5-Hard

Table 4.2: Scale used to assess the Engagement of the participants after reading the non-linear narratives

Engagement was measured by means of two variables. People were asked to rate their level of overall engagement as they read through the narrative. Also their experience of the difficulty in comprehension of the narrative was measured by the Narrative Understanding scale. A remarks section was provided at the end of the experiment to record any feedback from the participants. The scale used for the experiment is presented here in table 4.2.

4.7.2 Results

All the variables were standardized before performing any analysis. The Narrative Under- standing sub-scale used was found reliable (cronbach’s α = 0.703). Participants’ ratings Chapter 4. Generating Non-linear Narratives 48 to the scale items were combined by calculating the mean. The scale data, which is non- parametric in nature, was analyzed using the Kruskal-Wallis test to find any significance present, with the order of the sequences as the independent variable.

We wanted to know the amount of non-linearity that could be introduced without increas- ing the difficulty in comprehension of the narratives and if, increasing the non-linearity produced any change in engagement. According to our hypothesis cognitive engagement, which is associated with the process of comprehension, is achieved if the processing load required for comprehension is increased to a threshold and decreases if the processing load crosses this threshold. We assess cognitive engagement using the parameter of difficulty in comprehension as measured by the Narrative Understanding scale. According to our ex-

Figure 4.6: Difficulty in comprehension as measured by the Narrative Understanding scale of Busselle (2009) over various non-linear narratives periment, in narratives that were generated with a high-level non-linearity (high and high 2nd select groups), readers reported a subjective increase in the difficulty of comprehension with respect to the chronologically ordered narrative, as measured by the scale-items. A significant change was observed in the values measured by the 3-item scale, that measures lack of difficulty in comprehension H (5) = 15.599, p < .008. Chapter 4. Generating Non-linear Narratives 49

Table 4.3: Post-hoc analysis of the values of difficulty in comprehension across the non- linear narratives

Order of Events Mean Diff. Std. Error Sig Chronological/Low- Mid non-linearity -0.433 0.185 0.195 level non-linearity High non-linearity -0.808∗ 0.207 0.004 Low non-linearity (2nd select) -0.583 0.198 0.052 Mid non-linearity (2nd select) -0.537 0.201 0.098 High non-linearity (2nd select) -0.698∗ 0.195 0.010

Post-hoc analysis of the data is presented in table 4.3. No significant difference in the experience of the level of difficulty in comprehension was observed in narratives of low and middle-level groups. That is readers felt that high level non-linear narratives were difficult to comprehend while the participants who read the other level narratives did not report any such difficulty.

We also tested if the effect of the order of events produced a change in engagement. This was measured by the scale-item, ‘Overall engagement’. No significant difference in this measure was observed among the six narratives with varying orders of linearity H (5) = 4.169, ns.A change in order of events did not produce any change in the engagement.

It was observed that character inconsistencies in the story affected the overall engagement of the readers. Readers who rated their overall engagement as low reported that their understanding of the character actions was unclear. A strong negative correlations was observed between the two scale items of ‘Overall engagement’ and ‘Understanding of the character actions’, r s = -.183, p < 0.01.

4.7.3 Discussion

Despite the difficulty introduced in processing the low level and mid level non-linear nar- ratives, readers did not report any difficulty in comprehension. High non-linear narratives (high level and high level 2nd select) were rated to be difficult in processing than that of the chronological version. According to Busselle and Bilandzic(2009), if viewers are truly engaged then they should not experience any difficulty in processing a story despite its dif- ficulty. Based on these results we can claim that, for the story employed, increasing the non-linearity up to the mid-level produced no change in engagement. Chapter 4. Generating Non-linear Narratives 50

The results of the experiment do not tell if any group of the narrative is cognitively more engaging than the other. A reason that we did not find any significant difference in the overall engagement values in the narratives might be the aspects of the content of the story and also the way in which the story was presented using language, the discourse.

The elements of discourse used here were minimal. Medium plays an important part in conveying the story to the reader/viewer. In order to reduce any external effects due to the language, we have adopted a minimal use of discourse. Ideally, the generated narrative order should be passed through a discourse generator so that some richness in the might be obtained as evident in the narratives authored by humans.

Readers are usually familiar with a lot of stories. In this experiment, they remarked that the story was ‘routine’ and cheesy. They also said that the character actions are incomplete. Such a behavior might be due to the fact that the story itself is not human authored and is generated by a planning system. These issues definitely affect the engagement of the readers. The negative correlation between the scale items measuring ‘overall engagement’ and ‘understanding character actions’ also reflects this. These aspects of the content of the story concern the Story Generator but not of our system.

The scale item that measures the overall engagement might not be valid measure to test the effect of our system because the narrative variation generated by the system achieves only cognitive engagement but not an emotional one. Overall engagement of the reader incorpo- rates both cognitive and emotional aspects of engagement. A reader’s rating of engagement would therefore include the subjective experience of both cognitive and emotional aspects of engagement and is difficult to know if there was an effect of narrative variation on the cognitive engagement using the values of this scale item.

The subjective experience of the overall engagement may be affected by both the aspects of form and content. As our system tries to achieve narrative engagement by manipulating the form of a narrative, we adopt the Narrative Understanding scale that also deals with aspects of the form, while the other sub-scales deal with the aspects of content, emotions etc. As the content of the story is unchanged and only a change in the form (order of events) is generated, employing this scale is appropriate. In general it is difficult to measure the effect of variation in the form of a narrative because form cannot be isolated from the content of a story. Chapter 4. Generating Non-linear Narratives 51

We observed that by increasing the non-linearity of the narrative up to the mid-level, there was no significant change in the difficulty of comprehension. Thus our system could generate narratives of varying non-linearity without compromising on the subjective effect of difficulty in comprehension.

According to the results we cannot claim that mid-level non-linearity or a particular level of non-linearity is optimum for all narratives. Comprehension depends on the content of the narrative, the background knowledge of the readers among a few others. The level of non-linearity that could be introduced without compromising on the ease of comprehension is dependent on the structure of that story. Our algorithm considers this structure of the story and decides the order in which these events could be presented.

Chapter 5

Conclusion

This paper studied how narratives could be automatically generated from a given story, with varying degree of non-linearity. The emphasis of narrative generation was on generating an order of representation for a given story, without affecting the ease of comprehension of the content. Cognitive engagement could be obtained by varying the order in which information is presented to the readers. Finding the amount of deviation that could be administered, without affecting comprehension, to a story in chronological ordering was an important task of our study. Experimental study was conducted to study the effect of presentation order of events on narrative comprehension. Results obtained along with the principles of the Event-Indexing model were leveraged to arrive at a computational system towards narrative generation. Empirical evaluation concluded that readers experienced difficulty in comprehending narratives of high level non-linearity than that of low and mid-level non- linear narratives.

In the context of generating narrative variation in order, this research addressed the issue of achrony. Specifically, this research addressed the problem of generating non-linear narratives posed by Montfort (2007) that generating such narratives would be confusing for practical usage. Unlike other systems that use a plan based reader model that approximates the reader’s reasoning process, we use the principles of the Event-Indexing model, a cognitive model of narrative comprehension in our approach at narrative generation.

Unlike previous studies, we studied the order effects on comprehension by employing a commercial film narrative. The results corroborate the Event-Indexing model, which was

53 Chapter 5. Conclusion 54 initially proposed for text narratives, in the domain of film narrative as well. Our exper- imental results on order effects on narrative comprehension emphasized the importance of the goal dimension of the characters in an event, thus refining the assumption of the model’s dimensional equality hypothesis (Zwaan, 1999). Also we have considered that the spatial dimension to be a dichotomous entity, whereas in many cases it is perceived as a contin- uous entity. A proper analysis on dimensional relevance would strengthen the notion and the calculation of degree of non-linearity. We propose a refinement of the degree of causal relatedness between two events to employ the theory of spreading activation as a future study.

We have considered two notions of non-linearity in narratives. One definition is a generic level definition that a narrative that has its events presented in out of chronological order is termed non-linear. The system generates narratives of these kinds with desired amount of non-linearity. We cannot comment on the generation of narratives that are termed non-linear in our current literature or film, because a non-linear narrative is not just a rearranging of events from their chronology, but often the stylistic choice of an author. There is no functional definition that captures the intent of the author in structuring the narrative the way it is.

Evaluation of the effect of presentation order on engagement is difficult because engagement depends significantly on the content of the narrative and also the background knowledge of the person who is processing it. Therefore we cannot say that a narrative of a specific order is more engaging than the other, in isolation. Irrespective of the manifestation of engagement in the reader, we propose an algorithm that considers the structure of the story, its events and fundamentally decides which event to be presented next in the narrative. We believe that this ability to decide of our system is fundamental to the process of Narrative Generation.

In our current implementation, the user has to specify certain events as sequence markers and then proceed with the system. Also, instead of grouping the story into a set of sequences, the author of the story could identify key events in the story and designate them as sequences and run the algorithm. Given this control of designating events, though certain automation has been taken away, depending on the user input, our system could then generate a story that is more conventional of the stories in the realm of flash-backs. This power of designating what events have to be presented out of their chronological order in a narrative would facilitate Chapter 5. Conclusion 55 the author to see various ways in which the story could be told without being too complex for the readers/viewers. We hope that this system of ours could contribute as a tool for the authors in deciding the order in which the events of a story have to be presented.

Though our intention of the system to as a plug-in module for a story generation system that extends its functionality, we hope that it could be used on any story provided the story be in a partial order plan representation. As a further direction towards this research, evalu- ation of the system on large scale narratives is suggested. However, testing the algorithm on such narratives is a tedious process because one needs to code the causality of the numerous events by hand and obtain a plan representation. Also, evaluation of the system should be conducted by employing narratives that have rich discourse.

Though our work focused on traditional narratives, we hypothesize that the principles of degree of non-linearity could be used in deciding what event has to be presented next, in interactive narratives like games as well.

Solving the problem of narrative generation at a general level might help us in its appli- cation to generate narratives of specific variation like flashbacks, in media res etc. If we could generate non-linear narratives that engage the audience without being too much of a distraction then they may provide an interesting experience to the audience.

Bibliography

B.C. Bae and R. Young. A use of flashback and foreshadowing for surprise arousal in narrative using a plan-based approach. Interactive Storytelling, pages 156–167, 2008.

L. Bender and Q. Tarantino. Pulp fiction, 1994.

E. Branigan. Narrative comprehension and film. Routledge, 1992.

W.F. Brewer and E.H. Lichtenstein. Stories are to entertain: A structural-affect theory of stories. Journal of Pragmatics, 6(5-6):473–486, 1982. ISSN 0378-2166.

R. Busselle and H. Bilandzic. Measuring narrative engagement. Media Psychology, 12(4): 321–347, 2009.

C.B. Callaway and J.C. Lester. Narrative prose generation. Artificial Intelligence, 139(2): 213–252, 2002. ISSN 0004-3702.

S.B. Chatman. Story and discourse: in fiction and film. Cornell Uni- versity Press, 1980.

Y.G Cheong. A Computational Model of Narrative Generation for Suspense. PhD thesis, North Carolina State University, 2007. Ph.D Dissertation.

Y.G. Cheong and R.M. Young. A computational model of narrative generation for suspense. In AAAI 2006 Computational Aesthetic Workshop, Boston, MA, USA, 2006.

J. Cohen. Defining identification: A theoretical look at the identification of with media characters. Mass Communication and Society, 4(3):245–264, 2001. ISSN 1520-5436.

P.S. Cowen. Film and text: Order effects in recall and social inferences. Educational Technology Research and Development, 32(3):131–144, 1984.

57 Bibliography 58

D.K. Elson and M.O. Riedl. A lightweight intelligent virtual cinematography system for machinima production. Defense Technical Information Center, 2007.

O. Furman, N. Dorfman, U. Hasson, L. Davachi, and Y. Dudai. They saw a movie: Long- term memory for an extended audiovisual narrative. Learning & Memory, 14(6):457, 2007.

M. Galbraith. Deictic shift theory and the poetics of involvement in narrative. Deixis in narrative: A cognitive science perspective, pages 19–59, 1995.

G. Genette, J.E. Lewin, and J. Culler. Narrative discourse: An in method. Cornell University Press Ithaca, NY, 1980. ISBN 0801410991.

S. Ghislotti. Do you remember sammy jankins? film narration and spectator’s memory. Moving Image Studies, 4, 2004.

A.C. Graesser, M. Singer, and T. Trabasso. Constructing inferences during narrative text comprehension. Psychological Review, 101(3):371–395, 1994.

R. Hill, J. Gratch, S. Marsella, J. Rickel, W. Swartout, and D. Traum. Virtual humans in the mission rehearsal exercise system. K ”unstliche Intelligenz, 4(03):5–10, 2003.

A. Jhala and R.M. Young. A discourse planning approach to cinematic camera control for narratives in virtual environments. In Proceedings of the National Conference on Artificial Intelligence, volume 20, page 307. Menlo Park, CA; Cambridge, MA; London; AAAI Press; MIT Press; 1999, 2005.

A. Jhala, C. Rawls, S. Munilla, and R.M. Young. Longboard: A sketch based intelligent storyboarding tool for creating machinima. In Proceedings of the Florida Artificial Intel- ligence Research Society Conference (FLAIRS), 2008.

W. Kintsch. Learning from text, levels of comprehension, or: Why anyone would read a story anyway. Poetics, 9(1-3):87–98, 1980.

W. Kintsch, T.S. Mandel, and E. Kozminsky. Summarizing scrambled stories. Memory and Cognition, 5(5):547–552, 1977.

M. Lebowitz. Story-telling as planning and learning. Poetics, 14(6):483–502, 1985. Bibliography 59

B. Loenneker. Narratological knowledge for natural language generation. In Proceedings of the 10th European Workshop on Natural Language Generation (ENLG-05), pages 91–100, 2005.

J.P. Magliano, K. Dijkstra, and R.A. Zwaan. Generating predictive inferences while viewing a movie. Discourse Processes, 22(3):199–224, 1996.

J.P. Magliano, J. Miller, and R.A. Zwaan. Indexing space and time in film understanding. Applied Cognitive Psychology, 15(5):533–545, 2001.

J.P. Magliano, H.A. Taylor, and H.J.J. Kim. When goals collide: Monitoring the goals of multiple characters. Memory & cognition, 33(8):1357, 2005.

M. Mateas and P. Sengers. Narrative intelligence. J. Benjamins Pub., 2003. ISBN 1588112748.

R. McKee. Story: Substance, Structure, Style, and the Principles of . Methuen, 1998. ISBN 0413715507.

D.S. McNamara and J. Magliano. Toward a Comprehensive Model of Comprehension. Psy- chology of Learning and Motivation, 51:297–384, 2009.

J.R. Meehan. The metanovel: writing stories by computer. PhD thesis, Yale University, 1976.

N. Montfort. Generating narrative variation in interactive fiction. PhD thesis, University of Pennsylvania, 2007a.

N. Montfort. Ordering events in interactive fiction narratives. In Proceedings of the AAAI Fall Symposium on Intelligent Narrative Technologies, pages 07–05, 2007b.

B. Mott and J. Lester. Narrative-centered tutorial planning for inquiry-based learning en- vironments. In Intelligent Tutoring Systems, pages 675–684. Springer, 2006.

B.W. Mott, C.B. Callaway, L.S. Zettlemoyer, S.Y. Lee, and J.C. Lester. Towards narrative- centered learning environments. In Proceedings of the 1999 AAAI Fall Symposium on Narrative Intelligence, pages 78–82, 1999.

M.J. Nelson and M. Mateas. Search-based drama management in the Interactive Fiction Anchorhead. In Proceedings of the First Annual Conference on Artificial Intelligence and Interactive Digital Entertainment (AIIDE-05), 2005. Bibliography 60

J Niehaus. Cognitive Models of Discourse Comprehension for Narrative Generation. Phd dissertation, North Carolina State University, 2009.

K. Ohtsuka and W.F. Brewer. Discourse organization in the comprehension of temporal order in narrative texts. Discourse Processes, 15:317–317, 1992.

F. Paas, J.E. Tuovinen, H. Tabbers, and P.W.M. Van Gerven. Cognitive load measurement as a means to advance cognitive load theory. Educational psychologist, 38(1):63–71, 2003.

J.W. Peirce. PsychoPy–Psychophysics software in Python. Journal of neuroscience methods, 162(1-2):8–13, 2007.

J. Porteous, M. Cavazza, and F. Charles. Narrative generation through characters’ point of view. In Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems: volume 1-Volume 1, pages 1297–1304. International Foundation for Autonomous Agents and Multiagent Systems, 2010.

M.O. Riedl and R.M. Young. An intent-driven planner for multi-agent story generation. 2004.

M.O. Riedl, J.P. Rowe, and D.K. Elson. Toward intelligent support of authoring machinima media content: story and visualization. In Proceedings of the 2nd international conference on INtelligent TEchnologies for interactive enterTAINment, pages 1–10. ICST (Institute for Computer Sciences, Social-Informatics and Telecommunications Engineering), 2008.

S.J. Russell and P. Norvig. Artificial intelligence: a modern approach. Prentice hall, 2009. ISBN 0136042597.

E.M. Segal. A cognitive-phenomenological theory of fictional narrative. Deixis in narrative: A cognitive science perspective, pages 61–78, 1995a.

E.M. Segal. Narrative comprehension and the role of deictic shift theory. Deixis in narrative: A cognitive science perspective, pages 3–17, 1995b.

N.K. Speer and J.M. Zacks. Temporal changes as event boundaries: Processing and memory consequences of narrative time shifts. Journal of Memory and Language, 53(1):125–140, 2005.

S.R. Turner. MINSTREL: A Computer Model of Creativity and Storytelling. 1993. Bibliography 61

T.A. Van Dijk, W. Kintsch, and T.A. Van Dijk. Strategies of discourse comprehension. Academic Press New York, 1983. ISBN 0127120505.

R.M. Young, M.E. Pollack, and J.D. Moore. Decomposition and causality in partial-order planning. In Proceedings of the Second International Conference on AI and Planning Systems, volume 188, page 193. Citeseer, 1994.

D. Zillmann. Mechanisms of emotional involvement with drama. Poetics, 23(1-2):33–51, 1995. ISSN 0304-422X.

R.A. Zwaan. Five dimensions of narrative comprehension: The event-indexing model. Nar- rative comprehension, causality, and coherence: Essays in honor of Tom Trabasso, 1999.

R.A. Zwaan, M.C. Langston, and A.C. Graesser. The construction of situation models in narrative comprehension: An event-indexing model. Psychological Science, pages 292–297, 1995.