LEARNING KNOWLEDGE TO SUPPORT
DOMAIN-INDEPENDENT NARRATIVE INTELLIGENCE
A Thesis
Presented to
The Academic Faculty
by
Boyang Li
In Partial Fulfillment
of the Requirements for the Degree
Doctor of Philosophy in the
School of Interactive Computing
Georgia Institute of Technology
May 2015
Copyright 2015 by Boyang Li
LEARNING KNOWLEDGE TO SUPPORT
DOMAIN-INDEPENDENT NARRATIVE INTELLIGENCE
Approved by:
Associate Professor Mark O. Riedl,
Advisor
School of Interactive Computing
Georgia Institute of Technology
Assistant Professor Brian Magerko
School of Literature, Media, and
Communication
Georgia Institute of Technology
Assistant Professor Jacob Eisenstein
School of Interactive Computing
Georgia Institute of Technology
Associate Professor Stacy Marsella
College of Computer and Information
Science
Northeastern University
Professor Ashok Goel
School of Interactive Computing
Georgia Institute of Technology
Date Approved: November 14 2014
ACKNOWLEDGEMENTS
Behind every successful dissertation and its author, there is a group of wise people.
First and foremost, I would like to thank my advisor Mark Riedl, who I have been
fortunate enough to work with since the beginning of my Ph.D. program. Without
his guidance, encouragement and support, this dissertation would never have been
completed. I must also thank every member of my dissertation committee, whose
expertise and experience in multiple research fields have been tremendously beneficial.
I am really glad to have had the opportunity to learn from each of you.
Many friends and colleagues have helped me in the writing of this dissertation and
other research projects I carried out at Georgia Tech. Many thanks go to Stephen Lee-
Urban, Mohini Thakkar, Yijie Wang, Yangfeng Ji, Brian O’Neill, Alexander Zook,
Hong Yu, Nicholas Davis, Rania Hodhod and George Johnston.
I benefited from the advice from and discussions with many good friends and great
teachers in the research community. For this, I would like to thank Ian Horswill,
Emmett Tomai, Jonathan Rowe, Fox Harrell, Wei Lu, Arnav Jhala, David Roberts,
and Leslie Kaelbling.
Last but certainly not the least, I must thank my parents and family for their
love and support, whom I did not have a lot of time to visit while working on this
dissertation.
iii
TABLE OF CONTENTS
ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . iii
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
SUMMARY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
I INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Narratives and Narratology . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Narrative Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Knowledge Intensity of Narrative Intelligence . . . . . . . . . . . . . 6
1.4 Open Narrative Intelligence . . . . . . . . . . . . . . . . . . . . . . . 8
1.4.1 Necessity of Script-Like Knowledge . . . . . . . . . . . . . . 11
1.4.2 The Scheherazade System . . . . . . . . . . . . . . . . . . 12
II RELATED WORK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1 Narratives and Narrative Structures . . . . . . . . . . . . . . . . . . 15
2.1.1 Tiered Models of Narratives . . . . . . . . . . . . . . . . . . 17
2.2 Computational Narrative Intelligence . . . . . . . . . . . . . . . . . 21
2.2.1 Fabula Generation . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.2 Sjuzhet Generation . . . . . . . . . . . . . . . . . . . . . . . 26
2.2.3 Story Text Generation . . . . . . . . . . . . . . . . . . . . . 28
2.2.4 Story Understanding . . . . . . . . . . . . . . . . . . . . . . 29
2.3 Detecting Lexical Sentiments . . . . . . . . . . . . . . . . . . . . . . 32
2.4 Automatically Acquiring Script-Like Knowledge . . . . . . . . . . . 35
2.4.1 Learning from General-Purpose Corpora . . . . . . . . . . . 35
2.4.2 Crowdsourcing Knowledge . . . . . . . . . . . . . . . . . . . 38
2.5 Open Narrative Intelligence . . . . . . . . . . . . . . . . . . . . . . . 39
III LEARNING DOMAIN MODELS . . . . . . . . . . . . . . . . . . . 43
3.1 The Plot Graph Representation . . . . . . . . . . . . . . . . . . . . 44
iv
3.2 Collecting the Exemplar Stories . . . . . . . . . . . . . . . . . . . . 46
3.3 Learning Primitive Events . . . . . . . . . . . . . . . . . . . . . . . 49
3.3.1 Sentence Similarity from Syntactic Structures . . . . . . . . . 50
3.3.2 OPTICS for Small Data Sets . . . . . . . . . . . . . . . . . . 53
3.3.3 Mixtures of Dirichlet Distributions . . . . . . . . . . . . . . . 56
3.4 Evaluating the Learned Events . . . . . . . . . . . . . . . . . . . . . 58
3.5 Improving Learned Events with Crowdsourcing . . . . . . . . . . . . 61
3.6 Learning precedence relations . . . . . . . . . . . . . . . . . . . . . . 63
3.6.1 Method 1: Smart Thresholding . . . . . . . . . . . . . . . . . 65
3.6.2 Method 2: Integer Quadratically Constrained Programming . 69
3.6.3 Practical Concerns In Learning Precedence Relations . . . . 72
3.7 Learning Mutual Exclusion Relations . . . . . . . . . . . . . . . . . 73
3.8 Learning Optional and Conditional Events . . . . . . . . . . . . . . 74
3.9 Evaluating the Learned Graphs . . . . . . . . . . . . . . . . . . . . . 79
3.9.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.9.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.9.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.10 Limitations and Future Work . . . . . . . . . . . . . . . . . . . . . . 90
3.11 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
IV GENERATING AND TELLING STORIES . . . . . . . . . . . . . 93
4.1 Fabula Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.1.1 Legal Passages Through a Plot Graph . . . . . . . . . . . . . 94
4.1.2 Generating Passages With a Search Algorithm . . . . . . . . 98
4.2 Evaluating the Generated Fabula . . . . . . . . . . . . . . . . . . . . 101
4.2.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.2.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.3 Sjuzhet Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
v
4.3.1 PageRank and Stationary Distributions of Digraphs . . . . . 110
4.3.2 The EventRank Algorithm . . . . . . . . . . . . . . . . . . 113
4.3.3 Generating Different Sjuzhets . . . . . . . . . . . . . . . . . . 114
4.4 Textual Realization . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
4.4.1 Textual Interestingness . . . . . . . . . . . . . . . . . . . . . 118
4.4.2 Smooth SentiWordNet . . . . . . . . . . . . . . . . . . . . . 121
4.4.3 Connecting Sentences . . . . . . . . . . . . . . . . . . . . . . 126
4.5 Evaluating the Generated Story Texts . . . . . . . . . . . . . . . . . 128
4.5.1 Crowdsourcing Colorful Textual Descriptions . . . . . . . . . 128
4.5.2 Evaluating Generated Story Texts . . . . . . . . . . . . . . . 129
4.5.3 Evaluating Smooth SentiWordNet . . . . . . . . . . . . . . . 132
4.6 Limitations and Future Work . . . . . . . . . . . . . . . . . . . . . . 134
4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
V UNDERSTANDING STORIES . . . . . . . . . . . . . . . . . . . . . 137
5.1 The Story Understanding Problem . . . . . . . . . . . . . . . . . . . 138
5.2 NP-Hardness of the Story Understanding Problem . . . . . . . . . . 141
5.3 Simplifying Plot Graphs . . . . . . . . . . . . . . . . . . . . . . . . . 145
5.3.1 Cause-for-Removals . . . . . . . . . . . . . . . . . . . . . . . 147
5.3.2 Implied Temporal Relations . . . . . . . . . . . . . . . . . . 155
5.3.3 Implied Preconditions . . . . . . . . . . . . . . . . . . . . . . 162
5.3.4 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
5.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
5.4.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
VI FUTURE WORK AND CONCLUSIONS . . . . . . . . . . . . . . 169
6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
vi
6.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
6.3 Potential Applications . . . . . . . . . . . . . . . . . . . . . . . . . . 171
6.4 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
6.4.1 Creativity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
6.4.2 Multiplicity of Plot Graph Levels . . . . . . . . . . . . . . . 174
6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
APPENDIX A SMOOTH SENTIWORDNET . . . . . . . . . . . 177
APPENDIX B GENERATED STORY TEXTS . . . . . . . . . . 181
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
vii
LIST OF TABLES
1 Questions answered by SAM to illustrate story understanding . . . . 29
2 Two crowdsourced exemplar narratives in the bank robbery situation 48
3 Statistics of the crowdsourced corpora . . . . . . . . . . . . . . . . . 59
4 Precision, recall, F1, and purity of the identified event clusters . . . . 62
5 Accuracy of the learned precedence relations by the smart thresholding
method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6 Accuracy of the learned precedence relations by the IQCP method . . 85
7 Differences in accuracy of the learned precedence relations by the two
methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
8 Statistics of human edits made to the generated fabulas . . . . . . . . 106
9 Selecting events by typicality from a fabula to create sjuzhets . . . . 116
10 Fictionality of example words computed from the Gooogle N-Gram
corpus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
11 Sentences selected from events using different narration styles . . . . 120
12 Examples of fictional books obtained from Project Gutenberg . . . . 122
13 Some most positive and most negative words in Smooth SentiWordNet 124
14 Example sentences selected with positive and negative sentiments . . 125
15 Statistics of the additionally crowdsourced stories for interestingness . 129
16 Accuracy of the detected story-level textual interestingness, concise-
ness, and sentiments . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
17 Accuracy of the detected sentence-level sentiments . . . . . . . . . . . 133
18 Acceleration obtained by simplifying plot graphs based on mutual ex-
clusion relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
viii
LIST OF FIGURES
1 A sample plot graph for the pharmacy situation. . . . . . . . . . . . . 13
2 An example story contrasting the fabula, sjuzhet, and textual media . 20
3 Parsing result produced by the Stanford Parser . . . . . . . . . . . . 51
4 A reachability plot produced by OPTICS . . . . . . . . . . . . . . . . 54
5 Clusters extracted from a reachability plot . . . . . . . . . . . . . . . 55
6 The size of gold standard clusters in the gas pumping situation . . . . 60
7 Restoring low-confidence precedence relations . . . . . . . . . . . . . 68
8 Organizing vertices in a directed graph into multiple layers . . . . . . 71
9 Identifying optional events . . . . . . . . . . . . . . . . . . . . . . . . 75
10 Identifying optional events with clear paths . . . . . . . . . . . . . . . 76
11 A plot graph for the movie date situation, created by smart thresholding 80
12 A plot graph for the movie date situation, created by the IQCP method 81
13 Contrasting passages in plot graphs with finite-state machines . . . . 95
14 The user interface for editing fabula generated by Scheherazade . . 102
15 A plot graph for the bank robbery situation, created by the smart
thresholding method . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
16 The typicality of events in the bank robbery situation . . . . . . . . . 115
17 Reducing a 3-SAT problem to a plot graph . . . . . . . . . . . . . . . 143
18 An example of plot graph simplification due to mutual exclusion . . . 146
19 Examples of causes for removal . . . . . . . . . . . . . . . . . . . . . 149
20 A race condition during the detection of Cause-for-Removals . . . . . 152
21 Another race condition during the detection of Cause-for-Removals . 153
22 Simplifying Causes-for-Removals . . . . . . . . . . . . . . . . . . . . . 157
23 An example of adding a temporal relation during mutual exclusion
analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
24 A race condition during mutual exclusion analysis concerning the ad-
dition of temporal relations . . . . . . . . . . . . . . . . . . . . . . . 160
25 An example of implied co-occurrence . . . . . . . . . . . . . . . . . . 162
ix
SUMMARY
Narrative Intelligence is the ability to craft, tell, understand, and respond appro-
priately to narratives. It has been proposed as a vital component of machines aiming
to understand human activities or to communicate effectively with humans. How-
ever, most existing systems purported to demonstrate Narrative Intelligence rely on
manually authored knowledge structures that require extensive expert labor. These
systems are constrained to operate in a few domains where knowledge has been pro-
vided.
This dissertation investigates the learning of knowledge structures to support Nar-
rative Intelligence in any domain. I propose and build a system that, from an corpus
of simple exemplar stories, learns complex knowledge structures that subsequently en-
able the creation, telling, and understanding of narratives. The knowledge represen-
tation balances the complexity of learning and the richness of narrative applications,
so that we can (1) learn the knowledge robustly in the presence of noise, (2) generate
a large variety of highly coherent stories, (3) tell them in recognizably different narra-
tion styles and (4) understand stories efficiently. The accuracy and effectiveness of the
system have been verified by a series of user studies and computational experiments.
As a result, the system is able to demonstrate Narrative Intelligence in any domain
where we can collect a small number of exemplar stories. This dissertation is the first
step toward scaling computational narrative intelligence to meet the challenges of the
real world.
x
CHAPTER I
INTRODUCTION
Our dreams and stories may contain implicit aspects of our lives even
without our awareness.
Daniel J. Siegel
The long history of narratives in human civilizations can be at least traced back
to early myths found in diverse geographical regions, which are believed to define
collective identities and justify social institutions [95, 104]. Indeed, narratives in
various forms play multiple important functions in human cultures, both for the
cognition of individuals and the collective cognition of a society. As a result, Narrative
Intelligence (NI), or the ability to create, understand and respond to narratives,
is believed to be crucial for human intelligence, and by extension, crucial for the
simulation of human intelligence [20, 40, 57, 113, 203].
In the past few decades, attempts to computationally simulate Narrative Intelli-
gence has been limited by the reliance on manually authored knowledge, whose cost
of authoring limits the scalability of computational systems. In this dissertation, I
investigate the problem of learning knowledge in support of Narrative Intelligence
tasks. My approach learns narrative knowledge from a corpus of simple stories and
applies the learned knowledge to accomplish story generation, storytelling, and story
understanding.
This introductory chapter defines Narrative Intelligence and motivates the prob-
lem of learning knowledge in support of NI tasks.
1
1.1 Narratives and Narratology
Before discussing Narrative Intelligence, it is necessary to first define the notion of a
narrative. Prince [141] contrasted several definitions of narrative. Here I take a broad
definition:
Definition 1 (Narrative). A narrative is two or more conceptually and temporally
related events, which are presented through a particular medium to its audience.
In this definition, narratives include “novels and romances, novellas and short sto-
ries, history, biography and autobiography, epics, myths, folktales, legends and bal-
lads, news reports, spontaneous accounts in ordinary conversation, and so on” [141].
A narrative may be told orally, written in text, performed on stage, presented as audio
and video clips, or communicated through other media. Throughout this dissertation,
I use the word “story” interchangeably with the word “narrative”.
The prototype theory of categorization [157] argues that a concept is defined not
by clear-cut rules but by prototypical examples. In this sense, prototypical examples
of narrative would include Homer’s Iliad, Shakespeare’s Hamlet, and CBS’s The Big
Bang Theory. We could list features for prototypical narratives, such as involving the
same human or anthropomorphic characters, consisting of a beginning, a middle and
an ending, containing events related by causal relations, etc. Having these features
would make a narrative more prototypical and more easily recognizable as a narrative,
but in my definition these features are not strictly necessary.
In the study of narratives, the structuralist school has proposed models for break-
ing down a narrative into multiple levels. In particular, Mieke Bal [5] proposed a
three-tier model, the three tiers being denoted as the fabula, the sjuzhet
1
, and the
media in this dissertation. The fabula includes every event that actually happened
to and around characters in the narrative, which may happen simultaneously. The
1
Alternative spellings include sjuˇzet, syuzhet, suzhet, suzet, sujet, etc.
2
sjuzhet includes the events that appear in the narrative. That is, the sjuzhet is a
subset of the events in the fabula, which have been rearranged into a linear sequence.
The ordering of events in the sjuzhet may differ from their ordering in the fabula.
Finally, the sjuzhet is realized in the media tier as oral narration, text, video, etc.
Bal’s model is not meant to be a cognitive account about how humans create
or process narratives, but is used to an analytical tool to delineate the complexity
of a narrative and show different layers of construction. I will return to the three
tiers when defining Narrative Intelligence. For more details on narratives and tiered
narrative structures, see Section 2.1.
1.2 Narrative Intelligence
Cognitive research shows that narratives play important roles in human cognition and
communication. Narratives help us communicate complex situations and understand
intentions of other people [21, 135]. Storytelling can persuade and impart values [74],
create emotional bonding between teller and listener [45], help the construction of self
identities [167], and facilitate language learning among children [83].
Given the important role of narratives in human culture and human communica-
tion, many AI researchers and cognitive scientists (e.g. [20, 40, 57, 113, 203]) believe
that Narrative Intelligence
2
or the ability to craft, tell, understand and respond ap-
propriately to narratives, is crucial for human intelligence, and by extension, crucial
for the simulation of human intelligence. Thus, research into Narrative Intelligence
should bear significance in the quest for human-level AI and AIs capable of commu-
nicating with humans.
Definition 2 (Narrative Intelligence). Narrative Intelligence is the ability to craft,
tell, understand and respond appropriately to narratives.
2
The term Narrative Intelligence is said to have been coined at MIT in 1990 by Marc Davis and
Michael Travers [41]
3
Recent years have seen the emergence of real-world applications of Narrative In-
telligence. As consumers often consume digital content, such as video games and
virtual training scenarios, faster than the content can be produced, automated nar-
rative generation and rendering techniques may help to satiate the appetite for new
content. We may, for example, generate new plotlines for games to adapt to player
preferences and improve replayability [97]. Similarly, we can create new training
scenarios catering to different individual needs [123] and automatically produce cine-
matographic rendering of 3D scenes [151]. Interactive narratives have been deployed
in a number of science, archaeology, and theatrical museums (e.g. [159, 170, 194]). A
storytelling robot receptionist was installed at Carnegie Mellon University [70]. Story
understanding systems have been used to retrieve cyberbulling stories similar to a
user’s story [102], to retrieve local viewpoints for news events [107], and to retrieve
previous baseball games to be used as commentary [93].
In order to understand the complex information-processing ability that is Narra-
tive Intelligence, it is helpful to study NI as several interrelated components. I propose
NI contains a generative aspect and a comprehension aspect. Borrowing from Bal’s
three-tiered model discussed in Section 1.1, we can further break down both aspects
into several components. First, generative Narrative Intelligence can be broken down
into the following major components:
[G1] The ability to create a number of events that actually happen in the narrative
world, so as to achieve certain aesthetic or communicative goals, such as
coherence, suspense or persuasion.
[G2] The ability to create the sjuzhet based on the fabula, so as to achieve certain
aesthetic or communicative goals. Common techniques used to create the
sjuzhets include, but are not limited to, selecting events from the fabula,
rearranging selected events, and projecting the events to a particular point
of view (e.g. a personal perspective from a story character).
4
[G3] The ability to describe or perform the sjuzhet in an acceptable format, such
as texts, movies, comic strips, or theatrical performances, to the audience, so
as to achieve certain aesthetic or communicative goals. This process creates
the media.
In other words, the first three components constitute the process of creating the
fabula, transforming the fabula to the sjuzhet, and transforming the sjuzhet to the
media. The comprehension aspect of Narrative Intelligence can be similarly broken
down to include:
[U1] The ability to infer what events are described by the media. This ability
infers the sjuzhet from the media, and is the reverse process of G3.
[U2] The ability to infer what events have happened from the events being de-
scribed. This ability infers the fabula from the sjuzhet, and is the reverse
process of G2.
[U3] The ability to understand the aesthetic and communicative purposes of the
narrative and produce appropriate affective responses. Possible responses
include common affects, such as happiness or disgust, and narrative-specific
affects, such as identification with characters or suspense (cf. [47, 112, 126]).
Thus, narrative comprehension involves backward inference. By reading the text or
watching the movie, we are supposed to infer the sjuzhet from the media, and the
fabula from the sjuzhet, and interpret their implications. Such backward inference will
inevitably encounter uncertainty, so we must allow it to make reasonable mistakes.
Despite being listed separately, the generative and comprehensive aspects of Nar-
rative Intelligence do not work in isolation for the following reasons. First, the two
aspects share knowledge structures and operations on the knowledge structures. Sec-
ond, narrative creation requires the help of narrative understanding. Sharples [163]
5
propose a cognitive theory that creative writing alternates between an engagement
stage of intensive writing and a reflection stage where the results are reevaluated.
That is, a good story writer probably needs good comprehension skills, so that she
can continuously evaluate her own work, predict the audience’s affective responses,
and revise accordingly. Similar models are proposed by Flower and Hayes [58] and
Gerv´as and Le´on [68]. Finally, story understanding also requires generative abili-
ties, as understanding often requires drawing connections between elements presented
separately. The ISSAC story understanding system [120], for example, is capable of
creating novel analogies in order to understand novel concepts in science fictions.
I do not claim that the above component-based characterization of Narrative In-
telligence accurately captures how the human brain creates, understands or responds
to narratives. The main purpose of this description is computational. It provides one
possible division of labor for the complex processes underlying narrative creation and
comprehension. It may help AI researchers create modular computational systems,
or focus on one ability without losing sight of other challenges. Several computational
NI systems [4, 33, 119] adopted the division from Bal’s three-tier model in order to
create a pipelined process. In this dissertation, I implement components G1, G2, G3,
and U2 (See Section 1.4).
1.3 Knowledge Intensity of Narrative Intelligence
How can we build computational systems that can demonstrate Narrative Intelligence
as characterized above? Evidence from both cognitive science and artificial intelli-
gence systems suggests that in order to achieve Narrative Intelligence, we need a great
amount of knowledge about the world and people described in the narratives.
Cognitive studies suggest that in achieving Narrative Intelligence, humans make
use of extensive knowledge (e.g. [16, 202]) , which often needs to be acquired and
developed over an extended period of time (e.g. [2, 193]). Hudson and Shapiro [81]
6
listed four types of knowledge needed for narrative intelligence: (a) content knowl-
edge, or knowledge about events, event sequences, characters and social interactions,
(b) structural knowledge, or knowledge about the order and manner in which a story
is narrated, (c) microlinguistic knowledge, or syntactic knowledge for the understand-
ing and production of textual narrative representation, and (d) contextual knowledge
about pragmatic functions that stories serve for tellers and listeners. The categoriza-
tion of knowledge by Hudson and Shapiro is consistent with different components of
Narrative Intelligence defined earlier. The first three types of knowledge roughly cor-
respond to the creation and understanding of fabula, sjuzhet, and media respectively.
The last type of knowledge corresponds to the ability to understand storytellers’
intentions and audience’s responses (U3).
Existing computational NI systems are also heavily reliant on knowledge. From
the knowledge perspective, there are mainly two approaches toward computational
Narrative Intelligence: case-based reasoning (CBR) systems and planners. The case-
based reasoning approach (e.g. [37, 200, 190, 67]) builds Narrative Intelligence on
known stories. A new story can be understood in the context of old stories, and
snippets of known stories can be combined to create new stories. As one of the
earliest research into narrative understanding, Schank and Abelson [161] proposed
scripts, or the knowledge of typical events organized in typical temporal and causal
relations, as central to the understanding of narratives. Most CBR systems adopt
some variants of script. A second type of narrative intelligence systems (e.g. [115,
92, 154] employ knowledge of individual actions with preconditions and effects. By
matching preconditions with effects, a story planner can combine actions in any
causally consistent way to generate stories it has not known before. Notably, some
hybrid systems [136, 153, 97] utilize both script-like knowledge and knowledge about
individual actions (See Section 2.2 for a more complete review of computational NI
systems).
7
Regardless of the exact representation of knowledge, existing NI systems mostly
rely on human experts to manually author the knowledge content, which is usually an
expensive process. Although learning is a major component in the 4-phase CBR cy-
cle, most existing CBR systems require complex knowledge that are difficult to learn,
so the learning phase is usually ignored. Consequently, most computational systems
purported to demonstrate Narrative Intelligence are restricted to micro-worlds where
background knowledge has been provided a priori. These systems can generate and
understand narratives well in a few domains, but their Narrative Intelligence dimin-
ishes once the domain of application changes. This knowledge bottleneck has been
widely recognized. Nevertheless, the problem of automatically acquiring necessary
knowledge to support NI has not been thoroughly investigated.
1.4 Open Narrative Intelligence
In order to tackle the knowledge bottleneck that has troubled computational NI sys-
tems, Open Narrative Intelligence systems have begun to attract research interest
(e.g. [177, 166, 114]). Open NI Systems are not restricted to a predefined set of
domains because they are capable of learning knowledge necessary for Narrative In-
telligence in unknown domains.
In this dissertation, I aim to systematically tackle the knowledge bottleneck issue
by developing an Open NI system that supports story generation, telling, and under-
standing. More specifically, it attempts to answer the following research questions:
How can a computational system acquire structured knowledge in support of
narrative intelligence?
What is a good computational representation for such knowledge?
How can the learned knowledge be utilized efficiently in story creation, story-
telling, and story understanding?
8
The answer to these questions is summarized as the thesis statement of this Ph.D.
dissertation:
A computational system can acquire knowledge in the form of plot graphs,
which accurately describe common situations in the real world, to sup-
port Narrative Intelligence in a broad range of domains. The learned
plot graphs can support the generation of a large variety of coherent nar-
ratives, the telling of narratives in distinct narrator styles, and efficient
comprehension of narratives.
This dissertation will demonstrate a system that, from an input corpus of simple
exemplar stories, learn complex knowledge structures that subsequently enable the
creation, telling, and understanding of stories. The proposed knowledge representa-
tion aims to strike a balance between the complexity of learning and the richness of
narrative applications, so that we can accomplish the following goals:
Learn the knowledge robustly in the presence of noise.
By utilizing the learned knowledge, generate a large variety of coherent fabula
that can reasonably happen in the real world. This corresponds to the first
generative NI component, G1.
By utilizing the learned knowledge, select events from the fabula to generate
sjuzhet, which corresponds to the NI component G2.
By utilizing the learned knowledge, tell the sjuzhet with natural language, pro-
ducing interesting stories. This corresponds to the NI component G3.
By utilizing the learned knowledge, efficiently infer fabula from a given sjuzhet.
This corresponds to the comprehensive NI component U2. and interesting sto-
ries based on the learned representation.
9
The realization of NI capabilities G1, G2, G3, and U2 shows that the learned
knowledge representation can effectively and efficiently support computational Nar-
rative Intelligence, in both story generation and story understanding. To the best of
my knowledge, this is the first computational system capable of learning knowledge
to support both aspects of NI. I implement all story generation capabilities but not
all story understanding capabilities. The capability U2 is selected because inferring
fabula from sjuzhet is the most relevant to the plot graph representation, which for
the most part represents story structures in terms of events and their interrelation-
ships. Implementing U2 hence helps us to examine if story understanding can be
efficiently performed on this representation.
As a result, my system will be able to demonstrate NI in any domains where a small
number of exemplar stories can be collected. The collection process is easy because
writing the stories does not require training in computer science. One inexpensive
way (but not the only way), as I demonstrate, is to crowdsource the exemplar stories
from Amazon Mechanical Turk.
In addition to long-term scientific contributions, the development of the Scheher-
azade system brings about immediate benefits in terms of real-world applications.
The system’s capability to learn socio-cultural conventions can be employed for build-
ing virtual training scenarios that familiarize users with foreign cultures or guide
children with autism to adhere to social conventions. It could be expensive to hire
experts to author materials to cover many social situations in a foreign culture, such
as interviewing for a job in Japan or greeting the locals in Samoa. In comparison, the
Scheherazade system only requires a small number of exemplar stories written by
English-speaking non-experts who have some experience with the foreign culture and
that particular social situation. It was shown that crowdsourced data can help the
creation of intelligent systems help people with autism to learn about social situa-
tions [14]. Moreover, the Scheherazade system can provide background stories and
10
diverse narration styles to support believable virtual characters who talk about their
daily activities in the virtual world. Believable virtual characters find applications
in games, virtual training environments and healthcare. Section 6.3 contains a more
detailed discussion on the applications of Scheherazade.
1.4.1 Necessity of Script-Like Knowledge
As discussed earlier, current Narrative Intelligence systems make use of two major
forms of knowledge: scripts and action templates, which brings up the question which
knowledge representation should we learn. While both types of knowledge are useful,
I choose to learn script-like knowledge for the reasons explained below.
Despite the flexibility of action templates, I argue that action templates by them-
selves are not sufficient for the task of generating and understanding complex narra-
tives. There are two major reasons: First, reasoning about everything from first prin-
ciples is computationally slow. Even simple daily situations such as buying medicine
from a pharmacy store can involve complex reasoning: the pharmacist performs
information-seeking actions to check if the customer has the necessary prescription;
the customer requests the receipt as a contingency plan in case the medicine needs to
be returned; both parties may need to reason about the intentions of the other. Al-
though each problem can be solved computationally (e.g. with techniques discussed
in [52, 96, 152]), solving them every time is not as efficient as executing a known script.
Second, in many social situations, we need to respect social conventions, which may
not have resulted from clear rational reasoning at the present time. For example,
the choice of handshaking versus bowing as greetings may have important historic
reasons (such as hygiene conditions), but those reasons may not be present at our
time. Thus, respecting social conventions requires us to follow existing scripts.
My approach makes use of a script-like representation called a plot graph. A plot
graph representation contains vertices and edges. Vertices represent events that can
11
occur in the given situations. Some events are optional, indicating that their occur-
rence is not required for the legality of the event sequence. There are two types of
edges in a plot graph. A unidirectional precedence relation signifies that one event
strictly occurs before another event. The precedence relations form a partial-order
model, which allows for variations in legal event sequences for the situation. Prece-
dence relations coincide with causal and temporal precedence information, which are
important for narrative comprehension (cf. [56, 72, 187]). A bidirectional mutual ex-
clusion relation between two events signifies that the two events cannot occur together
in any story. Mutual exclusion relations provide support for situational variations and
alternatives.
An example plot graph portraying a pharmacy situation is shown in Figure 1.
The unidirectional precedence relations are shown as arrows, and the bidirectional
mutual exclusion relations are shown as dashed lines. The graph also includes an
optional event ”Customer takes change”. A formal definition of this representation
can be found in Section 3.1.
1.4.2 The Scheherazade System
This section provides an overview of the Scheherazade system
3
, and lays out the
organization of this dissertation.
The Scheherazade system learns the structure of events in a given situation
from crowdsourced exemplar stories describing that situation. The systems acquires
a number of simple exemplar stories for a particular social or procedural situation
from Amazon Mechanical Turk (AMT).
After a corpus of stories are acquired, the learning of the plot graph proceeds in
four steps. First, we cluster sentences with similar semantic meaning from exemplar
stories, each cluster becoming one event in the plot graph. In order to reduce the
3
Different from the story annotation scheme by David Elson [50] with the same name.
12
Pharmacist asks
for prescription
Customer
produces
prescription
Customer orders
drugs
Customer can’t
produce
prescription
Pharmacist
checks
prescription
Pharmacist
refuses to sell
Pharmacist
delivers drugs
Customer
swipes a card
Customer pays
cash
Customer
leaves
Customer takes
the receipt
Customer takes
change
Legend
Event
Mutual exclusion
Optional event
Precedence relation
Figure 1: A sample plot graph for the pharmacy situation.
13
difficulty in natural language processing, we have asked crowd workers from AMT
to use simple language, i.e., using one sentence with a single verb to describe one
event, avoiding pronouns, etc. In the second step, we identify precedence relations
between the events. The third step learns mutual exclusion relations using mutual
information. The final step identifies certain events as optional. Chapter 3 describes
the learning of plot graph and the evaluation of learned graphs.
Story generation in Scheherazade is the process of generating a linear sequence
of events while respecting the constraints posed by the precedence relations, mutual
exclusion relations, and event optionality. This linear sequence contains all events
that are presumed to have happened in the virtual world, which can be seen as a
linearized fabula. Scheherazade further selects a subset of fabula events in order
to create an interesting sjuzhet, and transform the sjuzhet into a textual description.
Chapter 4 details the generation of stories from the learned representation and their
evaluation.
Chapter 5 tackles the problem of efficient story understanding. When we infer
fabula from sjuzhet, it may be computationally expensive to estimate the probability
of a certain event happening in the fabula. Although the general problem is shown
to be NP-hard, I present methods to reduce its computational cost for commonly
encountered plot graphs. Finally, Chapter 6 summarizes and concludes this disserta-
tion.
14
CHAPTER II
RELATED WORK
What scientists do when they look at the line of bubbles on the screen is
work out the story of the particle that made them: what sort of particle it
must have been, and what caused it to move in that way, and how long it
was likely to continue.
Philip Pullman
This chapter starts by reviewing theories of narrative and narrative structures
to build a theoretical background and introduce vocabulary for subsequent discus-
sions. After that, I review traditional Narrative Intelligence systems based on man-
ually coded knowledge, as well as existing work for learning script-like information.
Combining the strengths of those work, Open Narrative Intelligence systems learn
knowledge in order to support Narrative Intelligence. These systems are discussed
and compared in the last section.
2.1 Narratives and Narrative Structures
On the definition of narratives, there are two major schools of thoughts. In classic
narrative theory (e.g. [61, 171]), a narrative is defined as the opposite of drama. A
narrative is communicated by a narrator whereas a drama presents the actions and
scenes directly. This school of thought emphasizes this mediacy as the most important
feature of narrative.
Despite the apparent differences, from an information-processing perspective the
understanding of an oral narrative, a theatrical drama, and a motion picture likely
share some common processes. For example, there is likely a process to reconstruct a
15
timeline of events, as events may have been presented out of order and simultaneous
events have been fit into a linear sequence. There is likely a process that infers events
that have happened but not presented in order to understand the events that are
presented. Although the drama and the motion picture seem more direct than the
oral narrative, they are not identical to the real world. The theatrical stage, for
example, may represent an apartment, a garden, or even a battlefield. The audience
usually have to understand certain stage and cinematographic conventions in order
to understand dramas or movies (cf. [11, 109]).
AI Researchers looking for an information-processing account of narratives turn to
the structuralist school (e.g. [164, 182, 185]), which defines narratives around events
that make change to a fictional world and relate to one another. Structuralists treat
oral narratives, novels, dramas, and movies all as legitimate narratives. Nevertheless,
it can be difficult to establish a detailed consensus definition for narrative, as struc-
turalists’ definitions differ in basic features such as the minimum number of events
required and if causal relationships are necessary in additional to temporal relation-
ships. E. M. Forster [59] famously claimed that ”the king died and then the queen
died” is a story, whereas ”the king died, and then the queen died of grief” qualifies as
a plot for it introduces a causal relationship. To Forster, a story needs at least two
events, and the plot presented to the audience should contain causal relationships
(see Section 2.1.1 for the difference between story and plot). Genette [64] argued that
a single event ”the king died” can qualify as a minimal story. Tomashevsky [185]
believed causality is necessary for stories. In contrast, Schmid [162] considers the dif-
ficulty of unambiguously determining if causal relations exist in a story or any forms
of text, and argues temporal relations should be sufficient for the definition of stories.
The interested reader is referred to Schmid [162] and Prince [141] for further details
on this subject.
Aiming at achieving Narrative Intelligence computationally, I attempt to avoid
16
this debate by taking a broad and practical definition:
Definition 1 (Narrative). A narrative is the telling of two or more conceptually
and temporally related events, which are presented through a particular medium to an
audience.
Further clarification is required for this definition. The medium of a story can
take many forms, such as text, comic, audio or video. By this definition, narratives
include “novels and romances, novellas and short stories, history, biography and au-
tobiography, epics, myths, folktales, legends and ballads, news reports, spontaneous
accounts in ordinary conversation, and so on” [141]. The telling of a narrative re-
quires a narrator and one or more narratees, which may be made explicit or implicit.
A narrative may also be told to oneself, either spoken out loud or silently. I exclude
a story with one event to distinguish a story from an event description. After all,
a story with only one event is uncommon in most daily scenarios. This dissertation
treats the words story and narrative as synonyms.
The prototype theory of categorization [157] argues that a concept is defined not
by clear-cut rules but by prototypical examples. In this sense, prototypical examples
of narrative would include Homer’s Iliad, Shakespeare’s Hamlet, and CBS’s The Big
Bang Theory. We could list features for prototypical narratives, such as involving the
same human or anthropomorphic characters, consisting of a beginning, a middle and
an ending, containing events related by causal relations, etc. Having these features
would make a narrative more prototypical and more easily recognizable as a narrative,
but in my definition these features are not strictly necessary.
2.1.1 Tiered Models of Narratives
In the structuralist tradition, we can further break down a narrative into a tiered
generative model of narrative. The model is not meant to be an accurate and plausible
cognitive model about how stories are actually generated by human writers, but a tool
17
for analytical purposes. Structuralists such as Tomashevsky [185] and Bal [5] used this
model as a way to analyze narratives and reader’s understanding of narratives. Such
models also enable Artificial Intelligence programs to generate stories in a pipelined
fashion (e.g. [4, 33, 150]).
The idea that a narrative can be analyzed in terms of two tiers, a fabula and a
sjuzhet, can be traced back to theorists earlier than Tomashevsky. However, Toma-
shevsky [184, 185] probably provided the earliest clear statement of their differences
and purposes. The fabula refers to the events that happen in the narrative-depicted
world. The sjuzhet refers to the description of events from which a reader learns
about the fabula. As various artistic techniques are used to produce the sjuzhet, such
as describing events out of order (e.g. flashbacks), describing events from different
perspectives, or omitting events, the reader must reconstruct the fabula from the
sjuzhet in order to understand the narrative [63]. In English, the fabula is referred
to as the story and the sjuzhet is often referred to the plot or discourse, which is
not dissimilar to the dichotomy by Forster [59] described earlier. French theorists
Todorov [182] and Genette [64] describe similar dichotomy as histoire and discours
or histoire and ecit respectively.
The concept of sjuzhet in the above two-tier model may be still considered am-
biguous. Bal [5] proposed a three-tier model that further differentiates between the
text in which the events are presented (which she called the text) and the events
being presented in the text (which she called the story). As narrative analysis is not
limited to the textual form, in this dissertation, I denote the three layers as fabula,
sjuzhet, and media.
Schmid [162] recently proposes a four-tier model. Schmid notes that the real world
contains infinite amount of details. Thus, there must be a process where the author
selects what constitute an event that is eligible to be included in the fabula. This
selection process is different from sjuzhet composition. Concepts of events must be
18
abstracted from artificially segmented perceptual inputs of the human senses.
1
This dissertation follows the three-tier model proposed by Bal [5]. Thus, I consider
a narrative to contain three tiers, as defined below:
Definition 3 (Fabula). A fabula is the complete set of events that include but are not
limited to all events presented in the narrative, that happen to and around all story
characters within a continuous time period, and that are temporally partially ordered
as multiple events can happen simultaneously.
Definition 4 (Sjuzhet). A sjuzhet is a linear sequence of events that are selected
from the fabula to constitute the narrative, possibly in a different ordering from the
fabula and with possible repetition of events. One event may be repeated from multiple
viewpoints, such as viewpoints of an omniscient narrator or different story characters.
Definition 5 (Media). The media is the form of representation in which the sjuzhet
is actualized and presented to the audience, including but not limited to oral narration,
audio recordings, video footages, theatrical performances, text, multimedia hypertext,
and so on.
An example story that illustrates the differences between the three tiers of fabula,
sjuzhet, and media is shown in Figure 2. Many different literary techniques may be
used to transform a fabula into a sjuzhet, including but not limited to: linearization,
which turns partial-ordered events into a single sequence that is suitable to telling;
reorganization, which orders the events differently from the order in which they ac-
tually occur (e.g. flashbacks); compression and expansion, which selects more or less
events from the fabula, creating a slower or faster pace; focalization, which tells events
from different points of view.
This dissertation follows this three-tier model and proposes methods to generate
fabula, sjuzhet, and text respectively. For the purpose of this dissertation, the task
1
Other complex models have been proposed. See, for example, Branigan’s 8-tier model [16],
whose structure is vastly different from those reviewed.
19
Fabula
Alice and Bob entered a
movie theater
Alice and Bob sat down
in the auditorium
...
Alice and Bob watched
the movie
Alice and Bob kissed
The movie ended
Alice and Bob stood up
...
The next day, Bob
thought about the
movie date during class
The teacher asked Bob a
question
Alice and Bob entered a
movie theater together
Alice and Bob watched
the movie
Alice and Bob kissed
Sjuzhet
Bob thought about the
movie date last night
during his class.
The teacher asked Bob a
question
Media(Text)
Bob just couldn’t stop thinking
about last night. What a wonderful
night! Alice and he went to watch a
movie together. At the end of the
movie, they kissed. Deeply.
Can this gentleman answer my
question? Bob’s thought was
suddenly interrupted by Mr. Smith.
Figure 2: An example story contrasting the fabula, sjuzhet, and textual media
of event segmentation and abstraction, as described by Schmid [162], is performed by
human crowd workers from Amazon Mechanical Turk and is not modeled directly by
AI.
Although the three-tier model was not proposed as a model of how human writers
actually write stories, some reasoning about the fabula, sjuzhet, and media is likely to
have happened during writing. For example, to write a detective novel, one probably
has to first think of a criminal plot, which is the fabula. The detective usually explains
this plot, which is realized as the sjuzhet and the text. For computational systems
that aim to generate narratives, this three-tier model provides a useful division of
labor and a modular system design. The next section reviews those systems.
20
2.2 Computational Narrative Intelligence
In this section, I review traditional story generation systems that rely on manually
coded knowledge. In line with the definition of Narrative Intelligence in Chapter 1,
I categorize generative systems into fabula generation, sjuzhet generation, and text
generation. Story understanding systems are categorized into systems that make
factual inferences and systems that predict or simulate human readers’ responses to
narratives.
2.2.1 Fabula Generation
As defined previously, fabula refers to the events that actually happen in the world,
not the events presented to the audience. Thus, strictly speaking, one cannot really
present a fabula. However, many computational systems do not explicitly reason
about or make changes to sjuzhet structures, and can be thought of as employing a
simple means of presenting a linear sjuzhet that includes everything in the generated
fabula or directly adopting the sjuzhet existing in a retrieved story case. To highlight
the fact that these systems do not reason directly about the sjuzhet, I call these fabula
generation systems.
There are two major approaches for fabula generation and Narrative Intelligence in
general: case-based reasoning and planning. These systems utilize different knowledge
structures and different operations on these structures. Some recent systems make
use of both case-based reasoning and planning.
Computational case-based reasoning (CBR) [87] is motivated by the observation
that humans use past experiences to guide decision making. Hence, Narrative In-
telligence can be built on top of a collection of known stories or story snippets. A
new story can be understood in the context of old stories, and known story segments
can be combined to create new stories. The notion of script was proposed as one
of the earliest research into narrative understanding. Minstrel [190] is an elaborate
21
CBR story generation system that reuses known stories and adapts them using rules
(called TRAMs) that recognizes story snippets similar to a query and transform the
snippet to satisfy the query. The story structure is constructed by filling in templates.
Minstrel utilizes three types of knowledge, including known stories, recipes to achieve
story themes and dramatic effects, and methods for adapting stories. All three types
of knowledge are manually authored.
2
Another inspiration of the case-based approach comes from the narratologist
studies that found many narratives contain combinations of similar building block.
Propp [142] found that the majority of Russian folklores can be explained by the
combination of 31 high-level narrative functions, which are semantic units that serve
similar purposes in the narrative, such as the hero receiving an interdiction and the
hero subsequently disobeying the interdiction. Studies of North American Indian
and Eskimo folktales yielded similar results [35, 46]. In the book Plotto published in
1928, the prolific novelist William Cook summarized popular plots as numerous in-
terconnected recipes. Gerv´as et al. [67] proposed a case-based reasoning system that
retrieves and combines Propp’s functions into stories. Characters in the functions can
be replaced for coherence and similar functions can be substituted.
The Riu system [129] generates steam-of-consciousness stories. The system repre-
sents cases as a temporal sequence of actions organized as a semantic network, where
each action is associated with a textual template. The retrieval of cases is performed
by matching identical text and structural analogy computed by the Structural Map-
ping Engine [54]. Similarly, Riedl and Le´on [149] transform a story analogically from
one domain to another and uses planning to fix incoherence.
The complete CBR cycle contains four phases: Retrieve, Reuse, Revise, and Re-
tain. The Retrieve phase locates an existing case, which is modified to solve a new
2
A recent reconstruction of Minstrel [181] can be downloaded from http://minstrel.soe.ucsc.
edu
22
problem in the Reuse phase. The Revise phase checks the newly formed solution
against reality. The Retain phase keeps good solutions for future use. Interestingly,
most CBR systems for story generation do not have full-fledged Revise and Retain
phases. This is probably because it is difficult to validate fictional stories, so it may
be undesirable to keep unvalidated stories. The complexity of knowledge used by
such systems also makes learning difficult.
If we have a very large source of story cases and organize them effectively in
memory, the CBR approach could theoretically work well in many domains. However,
due to the lack of the Retain phase, and the cost of manually authoring cases, CBR
systems can suffer from the knowledge bottleneck. Due to the lack of cases, we may
not find a case that is close enough to the story we wish to generate or understand.
When the target story and the retrieved case differ too much, substantial adaptations
are required, and too many changes to the case risk making it incoherent. The lack
of the Revise phase makes it difficult to detect incoherence, aggregating the problem.
As an example, Pe´rez y Pe´rez and Sharples [137] pointed out Minstrel may replace a
girlfriend with a horse because both can be kissed by a knight.
A second type of Narrative Intelligence systems are the story planners. Story
planners utilize a world model consisting of a number of action templates. An action
template contains a number of precondition propositions, which must be true for this
action to happen, and a number of effect propositions, which will become true after
this action happens. An action template may be lifted, i.e. containing variables to be
instantiated. For example, the eat(?person, ?food) action contains two variables
?person and ?food that can bind to different entities like Adam, Betty, noodle or
pizza. In order to achieve a given goal state, a planner instantiates lifted actions
and links individual actions into a complete sequence, where the preconditions of all
actions are satisfied by effects of earlier actions and the initial state.
Causal structures in a plan enable complex reasoning about story structures and
23
aesthetics, so a large number of story planning systems have followed this line of
research. The planner could be used to achieved goals of individual story characters,
or author goals about how the story should end. Talespin [115] is the first applica-
tion of planners in story generation and plans the actions for individual characters
in the hope that some interesting stories will emerge. Riedl and Young’s Intentional
Partial-Order Causal-Link Planner (IPOCL) [154] considers both character intentions
and author goals. Each character intention creates an intentional frame, and each
intentional frame contains actions that achieves that intention and is motivated by
another action. This allows IPOCL to create both coherent characters and satisfy
authorial intent. The Virtual Storyteller [179] system split each character into two
planners with different mindsets. The first one is an in-story character who is not
aware of anything outside direct perception, and the second is is an out-of-story actor
who is intentionally aiming for achieving story goals. Porteous et al. [139] explicitly
considered time during story planning. Ware and Young [196] extended IPOCL to
handle possible intentional conflicts in multiagent narratives, but the algorithm does
not automatically create or avoid conflicts. Brenner [17] proposed narrative genera-
tion with multiagent planning and limited look ahead. Si et al. [165] used partially
observable Markov processes, which might be considered as planning under uncer-
tainty, to model characters in interactive narratives. As an exception to the rule,
Cavazza et al. [24] presented a Hierarchical Task Network (HTN) planner to create
character behaviors. A HTN planner possesses the knowledge of how each higher-
level action can decompose into a sequence of lower-level actions. This knowledge is
probably more similar to scripts than to action templates.
By breaking stories into small building blocks, the planning approach is capable
of generating novel stories while maintaining the story coherence, which is usually
defined as characteristics of story plans (e.g. actions being justified by proper in-
tentions [154] or not having dead ends [97]). However, there are two reasons why
24
action templates cannot completely replace script knowledge. First, scripts provide
condensed knowledge that can be slow to compute directly from action templates.
For example, reasoning about characters’ intentions can be very expensive, as every
character may be reasoning recursively about other characters’ reasoning. Second,
scripts do not capture social conventions that are not completely results of logical
deductions. Therefore, social conventions need to be encoded as scripts.
As action templates and scripts have different strengths, some recent systems uti-
lize both types of knowledge representation. The Universe system [92] uses decompo-
sitions of high-level actions into low-level actions as well as actions’ preconditions and
effects to guarantee the coherence of the generated story. The planning mechanism in
Universe appears to lack backtracking and is a precursor to the modern Decomposi-
tional Partial-Order Causal-Link Planner (DPOCL) [205]. Mexica [136] models cre-
ating writing according to the model of two alternating modes by Sharples [163]. The
engagement mode uses case-based reasoning to project the next major action for char-
acters, whereas the reflection mode uses planning to fill in missing details. Flaiclough
and Cunningham [53] used case-based planning with Propp functions to generate sto-
ries in a multi-player game. Riedl and Sugandh [153] proposed a system similar to
multi-case based planning, which links multiple small story vignettes with planning
in order to maintain coherence. Li and Riedl [97] adapt human-authored plotlines in
games, which are represented as decompositional partial-order plans. They extend
the DPOCL algorithm to remove certain undesirable causal structures, such as ac-
tions that do not contribute to the main storyline (i.e. dead ends). The Doraemon
gadget generator [99] creates high-tech or magical gadgets that achieve impossible
deeds in fictions by analogically modifying the behaviors of common objects which
are represented as partial-order plans. The ability to create highly imaginative gad-
gets illustrates the power of hybrid systems.
25
Like CBR systems, story planners are also significantly restricted by the avail-
ability of manually authored knowledge. Hybrid systems build their strengths from
the utilization of both scripts and action templates, but the need to author both
types of knowledge aggravates the problem of knowledge bottleneck. Due to the sub-
stantial time and financial cost of authoring knowledge, the knowledge available to
those systems usually can describe only a few domains. As a result, these NI systems
can only operate in these domains. Some of those systems, such as PAM or ISSAC,
can produce results almost rivaling human intelligence in a known domain, but they
cannot operate outside such domains at all.
The proposed dissertation aims to address this knowledge bottleneck by devel-
oping methods to acquire structured knowledge in support of Narrative Intelligence
in arbitrary domains. Complex representations are usually powerful but difficult to
learn. By adopting the plot graph representation, my approach aims to strike a
balance between the complexity of representation and the ease of learning.
2.2.2 Sjuzhet Generation
A few computational systems have focused on sjuzhet techniques. Bae and Young [4]
explored the use of sjuzhet techniques, including flashbacks and foreshadowing, based
on causal structures in plans. They detected significant events that directly con-
tribute to goal states, and initiating events that are on causal chains leading to the
significant events. An initiating event that is not causally linked to any other event is
called separable, as the removal of this event will not affect the rest of the narrative
other than the corresponding significant event. This initiating event is then omitted
from the narrative, and presented after the significant event using flashback. Fore-
shadowing could be used to allude to the Initiating Event without completely giving
the event away.
Cheong and Young [34] created suspenseful sjuzhet based on a given fabula. They
26
first detected a core set of events, or kernels, without which the story could not be
understood. The rest of events are called satellites. Drawing from Trabasso and
Sperry [187], events that have a large number of incoming and outgoing causal links
and are close to the goal state are recognized as kernels. They followed the proposal of
Gerrig and Bernardo [65] and argued that overall suspense increases as the number of
plans in which the protagonist achieves her goals decrease, and the number of plans in
which the protagonist fails increase. Therefore, the potential suspense of each event
is computed from how many of its effects can threaten the protagonist’s goal and
how much it can support the protagonist’s goal. A suspenseful story is created by
adding events outside the kernel set with high potential suspense to the kernel events.
O’Neill and Riedl [128] provided another interpretation of Gerrig and Bernardo [65]
for the detection of suspense, which is reviewed in Section 2.2.4.
Another interesting problem in sjuzhet generation is focalization (cf. [3, 206]),
which refers to the technique of storytelling from multiple viewpoints, such as view-
points of an omniscient narrator or any story characters. In addition to the omniscient
foci, storytelling could be told from an external foci, where we know the external be-
haviors but not the mental activities of a character, or an internal foci, where we know
both the external behaviors and the mental activities of a character. Bae et al. [3]
propose a framework for generating different internal focalization using different li-
braries of actions templates. Porteous et al. [138] use a planning approach where
stories under different viewpoints are turned into constraints for planning.
The Curveship system [119] has an elaborate Teller module that supports sjuzhet
techniques including reordering the events, changing the frequency, speed, and focal-
ization of narration. The fabula events may be ordered in chronological order, reverse
chronological order, or in a zigzag form that interleaves the present and the past. It
also support flashback, flashforward, categorical and random ordering of events. The
27
system contains an advanced system of text templates for realizing textual descrip-
tions for events.
Most of the sjuzhet generation systems are based on planners, possibly due to the
rich set of information that causal structures of plans can provide. In Section 4.3, I
will discuss the detection of how typical an event is to a learned plot graph, and how
to use the typicality values to create sjuzhets. The task is similar to the detection
and use of kernel and satellite events in [34].
2.2.3 Story Text Generation
There are also research effort on creating the media layer of narratives. The AU-
THOR system [22] translates a symbolic plan generated by a story planner into a
textual story with several major steps: (1) segmenting the story into sentences and
paragraphs, (2) scanning from the beginning to end to determine articles, pronouns,
and contextual references, (3) making lexical choices to avoid repetition, and (4) ag-
gregating and reordering clauses to create complex and compound sentences. After
that, the intermediate product goes through a surface realizer to produce the final
story text.
Based on the Big Five model of personality [188, 124], Mairesse and Walker cre-
ated the Personage model [111] that maps the Big Five dimensions (Extraversion,
Emotional Stability, Agreeableness, Conscientiousness, Openness to Experience) to
a wide variety of language features including semantic content, syntactic structure,
sentimental polarity, vocabulary, and pragmatic markers such as “kind of and “you
know”. The computational implementation transforms a symbolic discourse plan
with precise semantic meaning into text according to the mapping in order to portray
different personality. Although the Personage model covers broad ground, the con-
tent of the discourse plan could limit the generated text. For example, if the system
28
Table 1: Examples of questions answered by SAM to illustrate its capability of story
understanding, produced by SAM after reading a story about John went to a restau-
rant and received a burnt hamburger. Adapted from [38].
Q: Did John sit down in the restaurant?
A: Probably.
Q: What did the waiter serve John?
A: The waiter served John a hamburger.
Q: Why didn’t John eat the hamburger?
A: Because the hamburger was overdone.
Q: Did John pay the check?
A: No. John was angry because the hamburger was overdone and
so he left the restaurant.
is not aware of possible paraphrases, it may not find the best paraphrase that char-
acterize a given personality. Rishes et al. [155] used Personage to create different
tellings of stories generated from a semantic representation consisting of events and
character intentions [50]. The generated linguistic styles differ mostly in pragmatics
rather than content.
Instead of generating from symbolic representations with precise semantic mean-
ing, I generate story text by selecting from existing sentences that are similar but
not strictly synonymous to describe an event (i.e. sentences may differ in content).
I consider parameters directly related to word choices: degree of details, fictionality,
and sentiments. Section 4.4 contains a detailed description.
2.2.4 Story Understanding
The comprehension aspect of Narrative Intelligence, as defined in Chapter 1, includes
the ability to (1) infer facts (from text to sjuzhet, and sjuzhet to fabula) and (2)
understand purposes of narratives and respond to narratives affectively. Story un-
derstanding systems can be categorized according to which of the two abilities they
achieve.
The first generation of story understanding systems mainly focus on the factual
29
inference problem. Schank and Abelson [161] proposed that scripts, containing the
knowledge of typical events organized in typical temporal and causal relations, are
central to the understanding of narratives. Following their lead, a number of systems
aim to understand stories by matching stories with known scripts consisting of slots to
be filled. SAM [37] applies known scripts to match a given story, so that it can fill in
empty slots in the script, resolve coreference, and predict events.
3
PAM [200] extends
SAM to make use of plans and causal relations between events to infer characters’
intentions. PAM considers the relationships between different characters’ intentions,
such as conflicts and coordination. Both SAM and PAM use question and answering
to demonstrate their story understanding capabilities. Table 1 shows some questions
and answers given by SAM in [38]. Recognizing the difficulty of manually authoring
scripts that are sufficiently numerous and detailed to handle the complexity of real-
world stories, the AQUA system [145] aims to learn from stories to improve scripts.
Section 2.5 contains a more detailed review of AQUA.
The ISSAC story understanding system [120] is notable for it transcends the
boundary between story generation and story understanding. Utilizing a vast set
of knowledge, ISSAC is capable of understanding novel concepts in science fiction by
creating analogies. In today’s terms, these analogies are conceptual blends [189], such
as the blending of robotic manufacturing machines with human intelligence to create
a robot commonly seen in science fictions.
Several other work aim to model the human readers’ cognitive process and to pro-
duce similar responses. Niehaus [122] and Cardona-Rivera et al. [156] build situation
models for narrative comprehension. These models determine salience of the events
in human readers’ memory, as they read a narrative sequentially.
3
In a candid historical account, Wendy Lehnert notes that due to the limited amount of knowledge
that SAM possesses, its input has to be “reverse-engineered to stay inside the limitation of SAM’s
available scripts” [94], although it was only meant to be a prototype.
30
O’Neill and Riedl [128] designed the Dramatis system for simulating how hu-
mans detect suspense in stories. They offer another interpretation of Gerrig and
Bernardo [65] that is different from Cheong and Young’s [34]. O’Neill and Riedl ar-
gue generating all possible plans to see if the number of successful and failed plans
increase or decrease is computationally impractical and cognitively implausible. In-
stead, they model the audience as searching for the most likely plan in which the
protagonist avoids a negative outcome, where likelihood is correlated with ease of
cognitive retrieval. Thus, a plan that refers to actions that have been activated or
primed previously is considered more likely. When the audience find it difficult to
think of an avoidance plan, suspense ensures.
The EA NLU system developed by Tomai [183] is mainly concerned with the
communicative purposes of narratives. The system can perform three understanding
tasks: (1) what is a proper moral decision in the situation described by a narrative,
(2) who is to blame for a negative outcome, and (3) what is the moral of a story.
EA NLU is based on ResearchCyc [39], which expresses over 5 million facts in higher
order logical propositions, but narrative-specific extensions still have to be made at
times. Tomai defines a set of narrative functions that readers expect from a story.
The natural language processing leverages this knowledge to abduct the meaning of
simple sentences.
In addition to the above symbolic approaches for story understanding, recent years
have seen the emergence of machine learning approaches toward limited semantic un-
derstanding of narratives, such as extraction of characters, events, and emotional
trajectories. Bamman et al. [6, 7] learn simple character types from novels. Valls-
Vargas et al. [192] detect heros, villains, and other characters by comparing extracted
characters’ actions against a predefined character interaction matrix. Following the
narrative analysis by Labov and Waletzky [90, 91], who classified several types of nar-
rative clauses, Ouyang and McKeown [132] and Swanson et al. [178] detect the major
31
actions in narratives. Mohammad [117] visualize the emotional trajectories of nov-
els using the number of emotional words. Elsner [49] computes similarities between
novels using character mentions and character emotions. While being able to cap-
ture some structural information, Elsner notes the technique is not yet sophisticated
enough to produce role-based narrative summaries.
2.3 Detecting Lexical Sentiments
In Section 4.4, the Scheherazade system recombines crowdsourced sentences into
a natural language presentation of a story. In order to generate stories with positive
or negative sentiments, I detect the sentiment of sentences by aggregating sentiments
of individual words. In the current section, I briefly review relevant work in the
detection of sentiments and emotions in natural language text. Sentiment analysis
has gathered significant research interests over the past few decades (e.g. [133, 169].
This review only concerns work that detect sentiments and emotions of individual
words.
Work in this area start with having linguistic experts annotating the emotions
associated with English words (e.g. [172]). Similar to the situation of narrative knowl-
edge, expert annotations for a large number of words are accurate but expensive to
acquire. Thus, there is generally a trade-off between the amount of annotation and
their accuracy, leading to a spectrum of dictionaries with varying degrees of coverage
and accuracy.
Expert annotations are usually the most accurate type of emotional dictionaries
but usually are limited to a small number of words and positive/negative polarity.
The General Inquirer lexicon [172] contains 11788 words where 1915 are labeled as
positive and 2291 are labeled as negative. It also contains additional sentiments cate-
gories include hostile, strong, weak, etc. Cerini et al. [25] developed the Micro-WNOp
32
corpus
4
containing 1105 WordNet synsets annotated with their positivity and neg-
ativity with real numbers in the [0,1] interval. As an extension of WordNet [116],
where word senses are organized as synsets, WordNet-Affect [173, 174] systemati-
cally selected words and labeled if the word describes emotion, mood, cognitive state,
emotional response and other aspects of affects. When a word is labeled as an emo-
tional word, it is further classified into positive, negative, ambiguous and neutral.
WordNet-Affect contains 1637 words and 918 synsets that are labeled as emotional.
The emerging technique of crowdsourcing also provides an cost-effective way to
annotate words’ emotions. Approaches such as the NRC Emotion Lexicon [118] and
SentiSense [43] can cover more words and more emotional categories with reduced
cost, but may not come with the same accuracy, compared to expert annotation. For
example, the NRC lexicon contains crowdsourced binary annotations of 8 emotion
categories for 24,200 word senses, about 14 times more than the number of emotion
words in WordNet-Affect. Both the NRC annotations and SentiSense contain binary
labels instead of real numbers.
Finally, there are automatic approaches that determines sentiments or emotions
of words based on their relations or associations with known words. Turney and
Littman [191] measure the association between words using pointwise mutual infor-
mation (PMI) and latent semantic analysis (LSA). First, an appropriate size of word
neighborhood is selected, which could be a document, a paragraph, a sentence, or a
fixed window of words. All words in the neighborhood are considered to be related.
We can compute the PMI based on the probabilities of two words appearing in the
same neighborhood. The semantic orientation, or valence of a word, is computed as
its association with a given set of positive words (W
+
) minus its association with a
4
Available at http://www.unipv.it/wnop
33
given set of negative words (W
).
SO-PMI(word) =
X
w
+
W
+
PMI(word, w
+
)
X
w
W
PMI(word, w
) (1)
A second measure computes a singular value (SVD) decomposition of a td-idf (Term
Frequency Inverse Document Frequency) matrix where rows represent words and
columns represent neighborhoods. A low rank approximation of the matrix can then
be obtained. The similarity between two words (LSA-Sim) is computed as the co-
sine of the angles between two word vectors in the U matrix of the SVD low rank
approximation.
SO-LSA(word) =
X
w
+
W
+
LSA-Sim(word, w
+
)
X
w
W
LSA-Sim(word, w
) (2)
Lu et al. [108] compute context-sensitive sentiment of words from user reviews.
They point out the word “unpredictable”, as an example, is usually negative in the
domain of electronic appliances, but may carry positive sentiment in movie reviews.
They consider multiple criteria, such as ratings on user reviews, linguistic features,
and general-purpose sentiment values, and framed the problem as a linear integer
programming problem. However, the NP-hardness of integer programming suggests
this approach may not scale well to very large corpus.
SentiWordNet [51] is another automatic approach that propagates known senti-
ments of a few words to other words along inter-synset relations in WordNet. Senti-
WordNet 3.0 has very broad coverage, providing sentiment labels for 117,659 synsets,
about 5 times as large as the NRC lexicon. Each synset is labeled with its positivity,
negativity, and objectivity as a point on the 3-dimensional simplex. That is, the three
dimensions must sum to 1. However, our experiments indicate SentiWordNet con-
tains substantial amount of inaccuracy, hindering accurate detection of sentiments.
To overcome this issues, Section 4.4 describes a corpus-based smoothing technique,
called Smooth SentiWordNet, to alleviate the inaccuracy of SentiWordNet.
34
Compared to Turney and Littman, my corpus-based method takes the distance
between words into account. The PMI and LSA methods consider all words in the
same neighborhood to have an equal degree of relatedness. I use Gaussian kernel
functions to assign less influence to words farther away. Given that sentiments may
change depending on the context, I used a corpus containing only fictions, so my
approach may be considered to be a narrative-specific expansion of the SentiWordNet
lexicon. Mohammad [117] computes emotions associated with fictional characters
using 5-grams from the Google N-Gram dataset. In comparison, I used 100-word
neighborhood windows to capture longer-distance influences.
2.4 Automatically Acquiring Script-Like Knowledge
The learning of discourse structure and relationships between events has been studied
for a long time. Several annotated corpora [23, 140, 144] have been created and often
used in supervised approaches. As annotated corpora are difficult to create and
can overfit to specific domains, in recent years there has been increasing interest in
learning script-like knowledge directly with little or no supervision. In this section, I
review unsupervised approaches for learning script-like knowledge.
2.4.1 Learning from General-Purpose Corpora
The earliest work on automatically extracting script-like knowledge that I am aware
of is by Fujiki et al. [62], who worked with Japanese news corpora. They proposed
three principles: (1) an action is a triplet of subject, verb, and object; (2) shared
subjects and objects indicate relations between actions; (3) the significant relations
can be identified by their frequency. Fujiki et al. used the first paragraph of news
articles because they are likely to be in chronological order. Although they extracted
only a small number of actions (41 pairs from 11 years of newspaper articles), the
three principles are adopted by many later work.
Brody [19] proposed a similar approach for learning actions (which Brody referred
35
to as clauses) and relations between actions. Instead of using a thesaurus like Fujiki
et al., Brody identified similar subjects, verbs and objects by clustering them based on
co-occurrence, which can be considered as distributional similarity. Further, Brody
proposed filtering valid relations between actions using the technique of hypothesis
testing.
Chambers and Jurafsky [26, 27] aim at learning Schankian scripts with events and
temporal partial orderings. Extending Brody’s work, they identified co-occurrence of
events based on pointwise mutual information and overlapping participants. Schema
learned are used for the task of understanding by extracting the slot-fillers in these
schema [28].
Talukdar et al. [180] learned clauses such as actedIn(actor, film) and wonPrize(film,
award) and their relative orderings. The orderings are based on their position in the
documents. As propositions in documents are not always in chronological order,
they focus on macro-level propositions, which are less likely to be described in a
non-chronological order.
Several work [69, 71] aimed at mining causal relations instead of entire scripts.
Girju [69] used supervised learning to identify linguistic rules that indicate causal
relations between noun phrases. The rules are based on WordNet categories that the
noun phrases belong to and the choice of the verb. Gordon et al. [71] describe an
approach to mining causal relations from millions of personal webblog stories. They
note the challenges associated with extracting causal, commonsense information from
such a corpus and also note that increasing the size of the corpus from one million to
ten million produced statistically insignificant improvements. Gordon et al. further
suggest that causal information in stories from these sources is best left implicit, and
that the ability to select between causal relations does not constitute a full solution
to open-domain commonsense causal reasoning.
Learning from general-purpose corpora has the advantage of large data sets, but
36
there are also several challenges that are often encountered by learning from general-
purpose corpora, which deserve some consideration. These challenges include
1. The content of general-purpose corpora may not align with the purpose of learn-
ing. These corpora may not contain the information needed, but also contain a
plethora of irrelevant information. Depending on the nature of the corpora, they
may be biased towards some topics and disregard other topics. For example, a
corpus of news articles will contain plenty information about politics and the
economy, but little about knights and goblins. Aiming at adult readers with suf-
ficient background knowledge, news articles may not contain typical procedures
for dining at a restaurant, but may contain a large number of rare incidents
at restaurants. Kasch and Oates [85] attempted to tackle this challenge by re-
trieving documents relevant to a given topic from the Internet. Words highly
correlated with the given topic are found by Latent Semantic Analysis and used
to retrieve relevant documents.
2. Documents are not always written in chronological order, creating obstacles
for the identification of temporal sequences. News articles usually start with a
summary, and lay out the details later. They may also refer to earlier events
at any time. Flashbacks and foreshadowing are common literary devices. To
alleviate this problem, Fujiki et al. [62] only used the first paragraph of news
reports. Talukdar et al. [180] focused on macro-level propositions.
3. Finally, the complex natural language in general-purpose corpora is difficult
for current technologies to process. Although some aspects of language, such
as part-of-speech, can be computationally recognized with extremely high ac-
curacy, many important tasks such as semantic role labeling and coreference
resolution remain challenging at this time.
A possible method to circumvent or alleviate these issues is to create specialized
37
corpora with crowdsourcing. In this dissertation, the specialized corpora is created
by asking crowd workers to write stories segmented into chronological sequences of
events and use simple language. Such a corpus contains the information we need and
relatively little irrlevant information. Learning from such specialized corpora could
provide bootstrapping for learning from more complex and noisy corpora. Incremental
learning has been shown to be beneficial in both cognitive science [48, 88] and deep
learning research [10].
My approach adopted some statistical methods proposed by earlier work, such as
the use of hypothesis testing for identifying temporal relations. This work has two
specific contributions: (1) The learned knowledge is put to the test of applications
of story generation and understanding, where previous work did not test the learned
knowledge with extensive applications. (2) This work identifies mutual exclusion
relations, which none of the above work recognizes. Mutual exclusion relations allows
us to incorporate multiple possibilities in one situation, thereby creating a compact
representation. For the purpose of storytelling, mutual exclusions play an important
role in maintaining story coherence. Empirical evaluation of the importance of mutual
exclusion relations can be found in Section 4.2.
2.4.2 Crowdsourcing Knowledge
Crowdsourcing has also been used as an effective method of knowledge acquisition.
The MakeBelieve system [106] extracts commonsense rules about action sequences
from the OpenMind knowledge base [168], which was built purely by crowdsourcing.
However, the knowledge focuses on declarative facts about non-fictional topics, which
limits their applicability to storytelling.
Regneri et al. [147] acquire scripts from crowdsourced bullet-list style instructions
entered by anonymous individuals and then manually converted to a canonical form.
They used an algorithm for aligning multiple sequences, yielding a total ordering
38
of events. In contrast, my approach builds a partial order of events with different
branches (as indicated by mutual exclusion relations). One partial order can describe
multiple total orders, allowing the story generation and understanding process to
dynamically adapt ordering of events depending on input or environmental changes.
Weltman et al. [198] designed a user interface aimed to collect highly detailed
step-by-step description of pictorial stories as well as causal interpretation. However,
their initial user study indicates the process to be complex and time-consuming. After
two hours of training, two of the ten participants could not finish the task, and four
other participants “agreed to do the bare minimum”. Though detailed knowledge
can support powerful NI, it is arguable that this knowledge representation errs on the
side of being too difficult to acquire.
2.5 Open Narrative Intelligence
If we can supplement Narrative Intelligence systems with the ability to learn knowl-
edge, we obtain Open Narrative Intelligence systems. These systems tackle the knowl-
edge bottleneck that has troubled traditional approaches for decades and extend
Narrative Intelligence into previously unknown domains and situations. This section
reviews these systems.
As a precursor to today’s Open NI systems, AQUA was proposed [145] to al-
leviate the knowledge bottleneck faced by story understanding systems like SAM.
At a time when machine learning was still in its infancy, Ram points out that it is
inherently difficult to encode sufficient knowledge for story understanding in the real
world. Instead of answering questions raised by users, AQUA is capable of raising
questions whose answer can improve its scripts. AQUA first tries to apply a script to
understand a story. When the script contains gaps or the story contradicts the script,
AQUA creates a question and stores it in memory, which may be resolved when new
information comes in. AQUA can thus learn new knowledge from stories to improve
39
the scripts it possesses. Taking a symbolic approach, AQUA is still heavily reliant on
manually coded knowledge to raise appropriate questions.
SayAnything [177] is a textual case-based reasoning system that is capable of
taking turns to write a story together with a user, each writing one sentence at a
time, without human-authored domain knowledge except some training data. Trained
on a human-annotated corpus, the system mines more than 1.6 millions stories with
an average length of 26.24 sentences from Weblog entries released by spinn3r.com.
After a user inputs a sentence, the n most similar sentences are retrieved using the
PL2 scoring function in the Terrier Information Retrieval Toolkit [131]. A neural
network was trained to rank the n sentences and the best is selected. During the
training process, the entity grid feature for local coherence [9] is shown to be the best
feature for training the neural network. The retrieved sentence is adapted to make
well-formed and coherent sentences and make pronouns agree between sentences.
The coherence of the generated story is maintained by both the sentence-retrieval
mechanism and the human input.
Another textual case-based reasoning approach has been presented by Sina et
al. [166], who modify crowdsourced semi-structured stories to create alibi for virtual
suspects for the purpose of training police officers. Each crowd worker is instructed
to write a three-part story, including an introduction of who, when, and where, the
body of the story, and the author’s overall opinion of the entire experience (e.g. of
a restaurant she visited). Asking the crowd to provide semi-structured data is an
effective strategy also used by my approach. The retrieval of cases depend on a
manually designed feature vector. Some features, such as number of children, appear
specific to the domain of application. Manually authored domain knowledge also has
been used in the adaptation, such as replacing daughters with nieces.
In a work similar to this dissertation, McIntyre and Lapata [114] learned schema,
or linear sequences of events that share the same entity from fairy tale texts. Each
40
sentence is considered to be one event. Sentences with the same verb and similar
arguments can be merged into one event, where similarity is computed with the
Wu and Palmer’s similarity measure [204] based on WordNet. The user specifies
required entities in a story, and schema associated with these entities are merged into
a single plot graph, where each path through the graph constitutes a story. A genetic
algorithm is used to optimize for the entity grid coherence measure [9], similar to the
SayAnything system.
My plot graph used in this dissertation assumes a partial order between events,
and different alternatives are explicitly represented using mutual exclusion relations.
McIntyre and Lapata’s plot graph contain implied mutual exclusion relations between
different paths in the graph. With their representation, we may generate a story
containing parts of two schema only if the two sequences share an event. The increased
separation is useful for learning from text corpus containing very different texts that
are likely to be incompatible. In comparison, Scheherazade learns from specialized
corpora containing texts that describe similar situations and are mostly compatible,
allowing a greater degree of recombination.
The Restaurant Game [130] uses human playing traces in a virtual restaurant to
learn a probabilistic model of restaurant activity. As the Restaurant Game is an
existing virtual game, there is an a priori known set of actions that can occur in
restaurants (e.g., sit down, order, etc.). Nevertheless, building a virtual environment
manually for each possible domain in order to gather knowledge about the domain
is very labor-intensive and unrealistic. One advantage of the approach developed in
this dissertation is that it does not require any 3D virtual environment.
Although the above systems aim to reduce the need for manually authored knowl-
edge, arguably none really eliminates the need completely. Manually authored and
41
designed domain-specific knowledge appears in the form of training annotations, fea-
ture design, 3D virtual worlds, etc. Neither does the system presented in this disser-
tation eliminate such needs. In order to crowdsource stories pertaining to a particular
social situation, a human-readable description of the situation must be written, which
requires at least some knowledge about the social situation.
Lastly, it is worth mentioning some large-scale efforts for data annotation that
aim to support Narrative Intelligence. Elson [50] developed an annotation scheme for
stories. Analogies and similarities can be identified from annotated stories, which was
shown to be a better measure of similarity compared to semantic similarities based on
logic propositions. Finlayson [55] learned Propp’s narrative functions from a corpora
of fully annotated Russian folklore by merging story paths in a Markovian model.
The system successfully identifies major narrative functions, but also misses many
transitions between the events. The system has not been used to generate stories or
understand unannotated stories.
To my best knowledge, this work is the first system that learns structured knowl-
edge for any domains that can support story generation, storytelling and story un-
derstanding.
42
CHAPTER III
LEARNING DOMAIN MODELS
Listen to great storytellers; slowly, you will learn about voice, timing,
tension, structure, climax—all the things you need to tell stories that will
capture the imagination of your audience.
Carmen Agra Deedy
Narrative Intelligence is highly knowledge intensive. Most existing computational
NI systems rely on manually coded knowledge and are restricted to operate in a
few domains where knowledge is available. The Scheherazade system is capable of
learning knowledge needed for computational NI from crowdsourced exemplar stories.
In this chapter I present the crowdsourcing procedures and learning algorithms.
Before any learning can take place, we collect some exemplar stories from human
workers on Amazon Mechanical Turk (AMT). Collecting stories from human writers
has the following advantages: First, these stories provide access to distributed memory
of real-world experiences and cultural conventions as consensus among members of
a society. Thus, they allow us to learn about social and cultural norms in addition
to procedural knowledge. Second, it taps collective creativity that can often produce
more diverse stories than an individual writer [110]. Third, we can be confident that
the corpus contains highly specialized information about a specific situation. Four, we
can guide the human writers to present the information in a form easy for computers
to process.
This chapter provides a detailed procedure for learning domain models, or plot
graphs, that describe how events typically unfold in social or procedural situations.
I first introduce the formal representation of plot graphs in Section 3.1. After that,
43
I describe how to acquire corpora of exemplar stories in Section 3.2. Subsequently,
Section 3.3 explains how we identify primitive events in a social situation by clustering
sentences with similar meaning. Section 3.6, 3.7, and 3.8 explain the learning of plot
graph structures, including (1) precedence relations, (2) mutual exclusion relations
and (3) optional and conditional events. The learn plot graphs are evaluated in
Section 3.4 and 3.9.
3.1 The Plot Graph Representation
A plot graph describes how social and procedural situations unfold as sequences of
events. In other words, a plot graph indicates what events can happen in a situation,
and constrains the way in which the events can happen. In a perfect model, only event
sequences that are legal according to the plot graph can happen in the situation. In
learning the plot graph, we aim to accurately model real-world situations as reflected
in the exemplar stories and filter out noises contained in the exemplar stories.
Definition 6 (Plot Graph). A plot graph G is a tuple hE, P, M
x
, E
o
, E
c
i. E is the
set of events. T E × E is a set of precedence relations between events. Mutual
exclusions relations between the events belong to the set M
x
E × E. Finally,
E
o
E is a set of optional events and E
c
E is a set of events that are conditioned
on the optional events.
In the graphic representation, an event is represented as a vertex in the graph, a
precedence relation is represented as a directed edge, and a mutual exclusion relation
is a bidirectional edge. The precedence relations and the events form a directed acyclic
graph (DAG).
The precedence relations and mutual exclusion relations put constraints on how
a situation may unfold. A precedence relation from event e
i
E to event e
j
E
indicates that event e
i
always happens before event e
j
in all legal event sequences
according to plot graph G. A mutual exclusion relation between event e
i
and event
44
e
j
indicates that events e
i
and e
j
never happen in the same legal event sequences
according to plot graph G.
Some graph notations and terminologies will become useful later in this chapter
and are worth introducing here. We use the notation (e
i
, e
j
) T to denote the fact
that there is a precedence relation from event e
i
E to e
j
E. We use the notation
(e
i
, e
j
) M
x
to denote the fact that there is a bidirectional mutual exclusion relation
between event e
i
E to e
j
E, so (e
i
, e
j
) M
x
(e
j
, e
i
) M
x
. We further define
the following:
Definition 7 (Parent). If and only if there is a precedence relation from event e
i
E
to event e
j
E (i.e. (e
i
, e
j
) T ), event e
i
is called a parent of event e
j
.
Definition 8 (Child). If and only if event e
i
is a parent of event e
j
, event e
j
is called
a child of event e
i
.
Definition 9 (Predecessor). The predecessor relation is defined as the transitive clo-
sure of the parent relation. That is, if e
i
E is a parent of event e
j
E, e
i
is a
predecessor of event e
j
. If e
k
E is a parent of event e
i
and e
i
is a predecessor of
event e
j
, e
k
is also a predecessor of event e
j
.
Definition 10 (Successor). Similar to the predecessor relation, the successor relation
is defined as the transitive closure of the child relation.
Definition 11 (Path). A path in a plot graph from event e
i
to event e
j
is defined as a
sequence of events he
1
, e
2
, ..., e
p
i such that every pair of adjacent events are connected
by a precedence relation, i.e. q [1, n 1], (e
q
, e
q+1
) T and e
1
= e
i
and e
p
= e
j
.
On a historic note, The partial-order DAG representation has been employed in
interactive narrative systems like Weyhrauch’s Tea for Three [199], and Nelson and
Mateas’s Anchorhead [121] to represent plots. Both systems use the directed edges to
denote necessary preconditions, which means a parent must happen before its children
45
are allowed to happen. Nelson and Mateas introduce an additional AND/OR label
that covers all parents for one event. If several parents are OR-ed together, only one
is needed before the child can happen. If they are AND-ed, all are required. In this
dissertation, I introduce the mutual exclusion relation, which serves a similar function
as the AND/OR label but is more expressive.
The main purpose of this representation is to bridge the learning process with
applications of Narrative Intelligence. This representation does not model causal
relations commonly used by story planners (e.g. [99]) because computationally iden-
tifying these relations remain difficult. Compared to simpler representation such as
used by Chambers and Jurafsky [26], I represent mutual exclusion relations, which
are important for maintaining story coherence during generation.
3.2 Collecting the Exemplar Stories
We first present the protocols for collecting exemplar stories from human workers
on Amazon Mechanical Turk. Each human worker is given a short description of a
social or procedure situation, such as “John and Sally went on a date at a movie
theater”, and are asked to write a story containing 10-20 sentences describing the
entire sequence of events. A compensation from 0.60 to 1 was paid out for each
accepted story.
For humans, writing stories is a natural form of communication. Telling stories
is found to be an effective means for human experts to share tacit knowledge [77],
which could be difficult to articulate otherwise. This is in contrast to, for example,
letting workers write production rules or manipulate probabilistic graphical models
and semantic networks. Turning knowledge engineering into story writing simplifies
the task and helps to increase the number of potential participants and lower the cost
of acquiring the corpus (i.e. the cost of hiring crowd workers to create the corpus).
This allows us to easily crowdsource exemplar stories. Although the AI techniques in
46
this dissertation are not tied to a specific story collection procedure, crowdsourcing
provides a practical and cost-effective means for collection.
Understanding free-form natural language is still largely an open problem. Thus,
we reduce the difficulty of natural language processing by asking the writers to use
simple language. Specifically, we require the AMT crowd workers to:
Segment the narrative into events such that each sentence contains a single
activity.
Make sure to include a complete sequence of events from the beginning to the
end.
Use proper names for all the characters in the task. This allows us to avoid
co-reference resolution altogether. We provide a cast of characters for common
roles, e.g., for the task of going to a fast-food restaurant, we provide named
characters in the role of the restaurant patron, the waiter, and so on. Currently,
these roles must be hand-specified, although future work could extract those
from commonsense knowledge bases.
Use simple natural language such as using one verb per sentence, avoiding con-
ditionals, complex, and compound sentences. Direct speech, such as “John said
’pass me the salt’ ”, usually contains more than one verbs and is advised against.
Use the past tense. Some verbs in the present tense, such as ”purchase”, can
sometimes be misclassified as nouns by the syntactic parser. The past tense
alleviates this problem.
We do not expect every human writer to follow these instructions precisely. Mi-
nor issues, such as a few misspellings or erroneous use of pronouns, are manually
corrected. If the story contains too many violations of the requirements, it is rejected
and no compensation is paid. A crowsourced correction is possible, as much research
47
Table 2: Two crowdsourced exemplar narratives in the bank robbery situation
Story 1 Story 2
John walked into the bank. Sally stood behind a counter.
John went up to the counter. John entered the bank.
John pulled out his gun. Sally saw John.
John asked Sally for the money. John approached Sally.
Sally started to cry. Sally said hello.
The police arrived. John handed Sally a note.
The police handcuffed John. John showed Sally a gun.
The police took John away. Sally read the note.
Sally opened the cash register.
Sally put money into a bag.
Sally triggered the silent alarm.
Sally handed John the bag.
John thanked Sally.
John exited the bank.
on crowdsourcing (e.g. [82, 127, 197]) have focus on this problem. However, an auto-
mated correction is out of the scope of this dissertation. At this time, lexical errors
in the crowdsourced corpora are manually corrected. Due to the existence of idiosyn-
cratic events, this correction does not completely eliminates noise in the data sets.
See Section 3.3 for how the Scheherazade system handles noise in the learning the
primitive events.
Experience indicates about 60-80 stories are sufficient for learning a plot graph
for a situation with 40-50 events. Each story can be acquired for 60 cents to 1 dollar,
so the total cost for one plot graph is around 36 to 80.
We refer to each segmented activity as a step. Table 2 shows two sample crowd-
sourced examplar stories for the situation of bank robbery. We can observe that each
48
story mentions only some events that may happen in this situation, and they often de-
scribe the same activity using different language. In addition, the stories can describe
different variations of the same situation. For example, In Story 1, John was captured
before he could get the money, but in Story 2 he managed to leave the bank. These
difficulties are handled by the next few steps of the learning process, starting with
learning primitive events, each of which may be described using different language.
3.3 Learning Primitive Events
The event learning process discovers the primitive units of event to be included in the
plot graph. This process learns salient concepts used by a community to represent
and reason about common situations. This phase of learning is based on a simple as-
sumption: salient concepts are mentioned more frequently than non-salient concepts.
Therefore, if multiple sentences describe the same underlying event, the event is likely
to be important for the situation. A clustering algorithm is used to seek clusters of
sentences semantically similar to each other. These clusters then become events in
the plot graph.
This clustering process faces two challenges. First, the algorithm must cope with
small data sets with complex structures. As each crowdsourced story has a monetary
cost, it is desirable to limit the number of stories that must be crowdsourced. In this
dissertation, 60-80 stories are usually acquired for one social or procedural situation.
Each social situation contains 30-60 event clusters. Many clusters contains only 4-5
sentences, which means the clustering algorithm must be sensitive to small clusters.
Second, the algorithm must work under a substantial amount of noise. The crowd
workers are not trained before the data collection, so they tend to include in their
stories non-salient, idiosyncratic events, as well as non-events such as descriptions of
the environments. The clustering algorithm hence must be able to filter out these
sentences as noises. Please refer to Table 3 in Section 3.4 for statistics of the data
49
sets acquired through crowdsourcing.
To cater to this highly challenging task, I adapted the OPTICS [1] clustering
algorithm to extract clusters from small data sets. This method derives similarity
between sentences by comparing syntactic dependency graphs, and benefits from the
noise resistance of the OPTICS algorithm. This method is compared with a prob-
abilistic method for detecting mixtures of Dirichlet distributions. Empirical results
indicate that though the probabilistic method provides a principled mathematical
formulation, the OPTICS-based method works better for most data sets.
In Section 3.3.1, I describe how I compute the similarity between any two sentences
with their syntactic dependency graphs. In Section 3.3.2, I describe the OPTICS
clustering algorithm and a new method for extracting clusters from the reachability
plot it produces. In Section 3.3.3, I describe the clustering method for learning
mixtures of Dirichlet distributions.
3.3.1 Sentence Similarity from Syntactic Structures
The similarity between two sentences is computed as a weighted sum of their gram-
matical similarity, based on their syntactic structures and similarities between the
words, and temporal similarity, or how similar the two sentences are based on the
relative location in the story.
I first compute the syntactic structural similarity between two sentences following
Lintean and Rus [105]. The similarity between two sentences is computed as the
aggregated similarities between the syntactic dependencies, which are further aggre-
gated from similarities between words.
The Stanford parser [86] is used to extract the syntactical structure of a sentence
as a directed graph (by collapsing and propagating dependencies in the basic tree
structure). Each edge on the graph describes a syntactical dependency involving two
words. The word at the tail of the directed edge (i.e. the word that the edge originates
50
John
handed
Sally
a
note
iobj
det
Figure 3: Parsing result of the sentence ”John handed Sally a note.” produced by
the Stanford Parser. The grammatical relations between the words are noun subject
(nsubj), direct object (dobj), indirect object (iobj) and determinant (det) respectively.
The arrows point from governors to dependents.
from) is called the governor and the word at the head of the directed edge is called the
dependent. A syntactic graph has one root word that has no incoming edges. Each
dependency belongs to a syntactic type, such as ”nsubj” or ”dobj”. Figure 3 shows
a sample parsing result. For more information on Stanford Typed Dependencies, the
interested reader is referred to [44].
When two syntactical dependencies are of different types, their similarity is zero.
When the two dependencies belong to the same type, I compute the average of the
word similarity between the governors and that between the dependents. The simi-
larity between two words can be computed based on WordNet [116]. Empirically, we
found the Resnik’s word similarity function [148] to be the most effective.
Resnik [148] uses only the is-a relations in WordNet. The similarity between
two synsets/concepts in WordNet is computed by locating their common ancestor
with the most information content, or the least probability. Between a synset s
and its hypernym h, we have P (s) P (h). The probabilities were found from the
Brown Corpus [60]. Using the information content, as argued by Resnik, is better
than simply counting edges (e.g. [204]) because different edges are not of the same
semantic distance. The Resnik similarity is non-negative but does not have an upper
51
bound. I normalize the similarities to between 0 and 1 for each data set separately.
My algorithm also caters to phrases consisting of two nouns, such as “movie
theaters”, which frequently appear in the data sets. Noun compound phrases are
indicated by a syntactic dependency of the type nn, or noun compound modifier.
When such a dependency is detected, the algorithm checks if the noun compound
exists as a phrase in WordNet. If it does, the noun compound is merged into one
word for the purpose of similarity calculation. Treating the two words as one phrase
allows the system to compute similarities at a finer granularity.
The dependencies are weighed by their distance from the root word of the sentence.
Words and phrases that describe unimportant details tend to be far away from the
root word. The depth of a dependency is computed as the number of edges from the
dependency’s dependent word to the sentence’s root word. Thus, the minimum depth
of a dependency is 1. An exponentially diminishing function of depth is used as the
weight. Formally, the similarity between two syntactic dependency of the same type
is
SyntSim(d
i
, d
j
) = w(d
i
, d
j
) s(d
i
, d
j
) (3)
where the function s(·, ·) computes the average of Resnik similarities between the
governors, gov(·), and the dependents, dep(·), of the two dependencies:
s(d
i
, d
j
) =
ResnikSim(gov(d
i
), gov(d
j
)) + ResnikSim(dep(d
i
), dep(d
j
))
2
(4)
and w(d
i
, d
j
) is the weight of this pair of dependencies:
w(d
i
, d
j
) = exp (0.2(depth(d
i
) + depth(d
j
) 2)) (5)
After computing the similarities between all possible pairs of syntactical depen-
dencies from the two sentences, I find the maximum matching between the depen-
dencies using the Hungarian algorithm [89]. Suppose sentence A and sentence B
are described with a set of syntactic dependencies D
A
= {a
1
, a
2
, ..., a
n
} and D
B
=
52
{b
1
, b
2
, ..., b
m
}. A matching M D
A
× D
B
pairs one element in A with at most
one element in B. The Hungarian algorithm finds the maximum matching M
=
argmax
M
P
(a
i
,b
j
)M
SyntSim(a
i
, b
j
). The similarity between the two sentences is then
computed as a weighted average
P
(a
i
,b
j
)M
SyntSim(a
i
, b
j
)/
P
(a
i
,b
j
)M
w(a
i
, b
j
).
The second component of sentence similarity is temporal similarity. We rely on
event location—a step’s location as the percentage of the way through a narrative—
to disambiguate syntactically similar steps that happen at different times. This is
especially useful when a situation is highly linear with little variation. For example,
when going to a movie theater, one will “wait in line” to buy tickets and then may
“wait in line” to buy popcorn. These two activities may share many syntactical
similarities, but will differ in their locations in the narrative. I use a weighted sum
of the two components of the syntactic similarity and the temporal similarity, as the
overall similarity between the two sentences.
TotalSim(d
i
, d
j
) = SyntSim(d
i
, d
j
) + λ TempSim(d
i
, d
j
) (6)
The parameter λ typically ranges from 0.3 to 0.15.
3.3.2 OPTICS for Small Data Sets
OPTICS
1
[1] is one of the most popular density-based clustering algorithms with
the advantage of being resistant to noise and being able to detect clusters of different
shapes and densities. The algorithm maps data points onto a sequence of reachability
values where dents and valleys indicate clusters. In this section, I introduce the OP-
TICS algorithm and propose a new heuristic for extracting clusters from reachability
plots that work well on small data sets.
The underlying intuition of OPTICS is that a cluster is formed when a number
of points are close to one another. The algorithm takes as input a distance matrix
1
In the tradition of having good acronyms for clustering algorithms, OPTICS stands for Ordering
Points To Identify the Clustering Structure.
53
Figure 4: A reachability plot produced by OPTICS, where valleys represent clusters.
Reproduced from [1]. The parameter MinP ts = C
m
1
that contains the distance between any two data points. It has two paramters: the
minimum number of points in a cluster, denoted as C
m
, and the maximum distance
between two points that we will consider, denoted as . OPTICS allows to be set to
infinity, which means we will consider two points however far apart they are. OPTICS
computes a core distance for each data point, which is the radius of the circle that
contains the nearby C
m
1 number of points. A smaller core distance means the data
point is closer to a cluster.
OPTICS defines the reachability distance between two points o and p as
reachability-distance(o, p) = max(core-distance(p), distance(o, p)) (7)
OPTICS can start its processing from any data point. It keeps a priority queue of
all unprocessed data points, ordered from low to high reachability distance. As a
new data point is processed, the reachability of all other points in the queue are
updated as the reachability to the current point. OPTICS always process the data
points with the smallest reachability in the queue. Therefore, when OPTICS finishes
processing all data points in one cluster, which have low reachability to each other,
54
Figure 5: Clusters extracted from a reachability plot for the gas pumping situation.
The recognized clusters are shown in various colors, and sentences not belonging to
any clusters are shown in black.
it will encounter a data point with high reachability since this new point is far away
from the current cluster. This data point could signal a new cluster or could be noise,
depending on whether it is followed by other points of low reachability. When we
visualize the reachability distances as a bar chart, we obtain a so-called reachability
plot. One such reachability plot is shown as Figure 4. Regions with low reachability
are clusters separated by regions of high reachability.
The next task is to extract clusters from the reachability plot. The original OP-
TICS paper [1] provides a method for extracting clusters from reachability plots.
However, the small size of data sets we have tend to produce rather crooked reacha-
bility plots on which the technique do not work well as in more smooth plots.
In order to extract clusters from reachability plots, I first use the method by
Sander et al. [160] to convert a reachability plot into a hierachy of clusters. I then
collect all leaf nodes. Each leaf node contains a number of data points. A leaf node
is considered to contain a valid cluster if (1) there is a sufficient drop in reachability
between the highest point and the lowest point in the node and (2) there is a relatively
55
flat region in the node. I then extract all regions that are relatively flat from a valid
leaf node as a cluster. An example of extracted clusters is shown in Figure 5.
3.3.3 Mixtures of Dirichlet Distributions
In the second method, we represent each sentence as a bag of n-grams. I use the
n-gram data produced by Lin et al. [103], where each n-gram is represented as a
1000-dimension vector. I normalize the original data so that each vector sums to
1. A cluster is represented as a Dirichlet distribution that all the n-grams of all the
sentences in the cluster are drawn from. Formally, we parameterize each cluster with
a vector θ
j
, j [1, N]. We have a number of sentences s
i
, i [1, M], each consisting
of n-grams x
i
1
, . . . , x
i
h
. Both θ
j
and x
i
h
are of dimension D. We maximize the overall
likelihood function:
Λ(x
x
x;θ
θ
θ) =
M
Y
i=1
N
X
j=1
p
j
K
Y
h=1
P (x
i
h
| θ
j
) (8)
where
P (x
i
h
| θ
j
) =
Γ
P
D
r=1
θ
j,r
Q
D
r=1
Γ(θ
j,r
)
D
Y
r=1
(x
i
h,r
)
θ
j,r
1
(9)
is the probability density function of the Dirichlet distribution.
This is a classic formulation where the entire data set is a mixture of distributions.
The clustering problem is equivalent to finding the probabilistic model that assigns
maximum probability to the observed sentences (in other words, finding the maximum
likelihood estimate). This can be solved by the expectation-maximization algorithm.
To simplify the problem we introduce a set of auxiliary variables z
i,j
{0, 1}, i
[1, M], j [1, N]. z
i,j
= 1 denotes that the sentence s
i
is assigned to the j
th
cluster.
In the E-step, with θ fixed, we want to find the optimal assignment for z:
ˆz
ˆz
ˆz = argmax
z
z
z
P (x
x
x,z
z
z | θ
θ
θ) = argmax
z
M
Y
i=1
p
j
K
Y
k=1
P (x
i
k
|θ
j
)
!
δ(z
i,j
=1)
= argmax
z
M
X
i=1
δ(z
i,j
= 1)
log(p
j
) +
K
X
k=1
log(P (x
i
k
|θ
j
))
!
(10)
56
where the delta function δ(z
i,j
= 1) equals 1 when z
i,j
= 1 and equals 0 otherwise.
Noting that sentences from different stories are independent, and their assignment
does not affect each other given θ is fixed, we can consider the assignments of z for
each story separately. Observing that two sentences in one stories rarely describe the
same event, we further enforce the constraint that no sentences in one story are put
into the same cluster. That is, each cluster contains one or zero sentences from each
story:
X
i
z
i,j
1, j (11)
We also enforce the constraint that every sentence is assigned to exactly one cluster:
X
j
z
i,j
= 1, i (12)
This is solved as a 0-1 integer linear programming problem for each story.
In the M-step, the assignment of z is fixed, and we compute p
p
p and θ
θ
θ that maximizes
the data likelihood. p
p
p is computed with Laplace smoothing:
ˆp
j
=
P
N
i=1
δ(ˆz
i,j
= 1) + 1
N + M
(13)
The the maximum likelihood estimator for θ
θ
θ can be sovled for individual θ
j
ˆ
θ
j
= argmax
θ
j
P (x
x
x,z
z
z | θ
θ
θ) = argmax
θ
j
M
Y
i=1
K
Y
k=1
P (x
i
k
|θ
j
)
!
δ(ˆz
i,j
=1)
= argmax
θ
j
M
X
i=1
δ(ˆz
i,j
= 1)
K
X
k=1
log(P (x
i
k
|θ
j
))
!
(14)
However, the maximization does not have a closed-form solution and must be solved
numerically [80].
We need to select D, the number of dimensions for x
x
x and θ
θ
θ, so that many clusters
can be sufficiently differentiated, and each θ
j
can be reliably estimated from limited
data. I experimentally determined D = 150, so that the 1000-dimension vectors
provided by Lin et al. [103] have been reduced to 150 dimensions using principal
57
component analysis. Each sentence was broken into the fewest number of n-grams
using an A* algorithm.
In order to deal with noisy sentences, I use a large number of clusters (M typically
set to about 130), and discard from the final results any clusters with less than the
minimum number of sentences (i.e. C
m
). This approach was shown to be effective.
I also experimented with adding Markov transitions between the learned clusters to
the probabilistic model. However, experiments indicate that that doing so actually
decreases performance.
3.4 Evaluating the Learned Events
Corpora of exemplar stories have been collected for the following situations:
Going to a fast food restaurant.
Taking a date to a movie theater.
Robbing a bank.
Pumping gas into a car
Making coffee
Going through airport procedures before boarding an airplane
Buying medicines at a pharmacy
A wife catching her husband having an affair
A wedding proposal
Table 3 shows the attributes of these crowdsourced corpora for each situation,
including the number of stories crowdsourced, the average number of sentences in
each crowdsourced story, the number of unique verbs and nouns in the entire corpus,
58
Table 3: Statistics of the crowdsourced corpora
Situation
Stories
Avg.
Steps
Unique
Verbs
Unique
Noun
Gold Std.
Events
Noisy Sentences
% Noisy Sen-
tences
Fast Food Restaurant
30 6.91 41 29 21 6 2.71
Movie Date
78 11.24 108 104 56 183 20.86
Bank Robbery
60 11.78 142 115 44 160 22.63
Airport
61 15.08 130 182 49 251 27.28
Coffee Making
36 11.89 72 98 35 180 42.06
Gas Bumping
59 14.25 105 113 41 56 6.66
Pharmacy
69 12.56 146 145 58 147 16.96
Extramarital Affairs
79 14.95 248 267 59 475 40.22
Wedding Proposal
37 13.00 134 136 26 239 49.69
59
Figure 6: The size of gold standard clusters in the gas pumping situation, as percent-
age in the entire corpus.
and the number of events in a manually created gold standard. For each corpus,
I manually created a gold standard set of clusters against which the automatically
discovered clusters are evaluated against. It is worth noting that most data sets
contain a large number of clusters (40-60) with only a few sentences in each cluster.
Figure 6 shows the size of gold standard clusters in the gas pumping situation as their
percentage in the entire data set. Only three clusters contain more than 42 sentences,
or 5% of the entire corpus. On the other hand, 12 clusters contain less than 1% of
sentences. Further, data sets typically contain a large percentage (about 20% to 40%)
of noisy sentences that do not belong to any cluster. The compound effects of small
clusters and the presence of substantial noise make the task of event learning difficult.
We compare the two learning techniques on all situations. Table 4 presents the
results of event learning on our two crowdsourced corpora, using the MUC6 cluster
scoring scheme [195] to match actual cluster results against the gold standard. The
purity of a cluster measures intra-cluster homogeneity. Higher purity indicates higher
cluster quality. For each automatically generated cluster C
i
, we find a gold standard
cluster G
j
such that the maximum number of sentences in C
i
belongs to G
j
. The
60
overall purity aggregates the overlap between C
i
and G
j
for all clusters. Given a set
of gold standard clusters G = {G
1
, G
2
, ..., G
n
}, and a set of automatically generated
cluster C = {C
1
, C
2
, ..., C
m
}, we define
purity(G, C) =
1
N
m
X
i=1
max
j
|C
i
G
j
| (15)
where N is the total number of sentences.
We can observe that the best precision values for most data sets are in the range of
high 70s to 80s. Two data sets, coffee making and wedding proposal, are more difficult
than the rest, with precision in the 60s. The recall values are lower than precision, but
it is arguable that precision is more important than recall in this task. Most purity
values are in the 70s or higher, except three data sets: coffee making, gas pumping,
and wedding proposal. The results indicate some data sets are more difficult than
others. The two most difficult data sets, coffee making and wedding proposal, have
the most noise sentences, and the second and the third least stories, which contributed
to their difficulty. Performance can probably be improved by collecting more stories
for the two data sets.
The first method, based on OPTICS and similarities computed from syntactic
structures, consistently outperforms the second method, which models clusters as
Dirichlet mixtures. This can be attributed to the several reasons: (1) the first method
utilizes syntactic structures, which are more informative than n-grams, (2) the sim-
ple sentence structures resulted from crowdsourced stories helped us in computing
grammatical similarities between them, and (3) OPTICS provides a robust method
for dealing with noise sentences that do not belong to any clusters.
3.5 Improving Learned Events with Crowdsourcing
Although the event learning process achieves acceptably high accuracy rates, the
errors at this stage can be amplified in later stages of learning, resulting in low-
quality plot graphs. One possible solution is to make use of crowdsourcing again to
61
Table 4: Precision, recall, F1, and purity of the identified event clusters
Situation
OPTICS Dirichlet Mixitures
Precision Recall F1 Purity Precision Recall F1 Purity
Fast Food 0.880 0.688 0.772 0.836 0.753 0.738 0.745 0.642
Movie Date 0.908 0.590 0.715 0.731 0.699 0.738 0.718 0.672
Bank Robbery 0.851 0.586 0.694 0.722 0.564 0.576 0.570 0.552
Airport 0.798 0.437 0.562 0.736 0.492 0.585 0.534 0.453
Coffee Making 0.633 0.362 0.460 0.625 0.404 0.443 0.423 0.458
Gas Pumping 0.672 0.422 0.519 0.573 0.493 0.439 0.464 0.443
Pharmacy 0.778 0.512 0.617 0.705 0.532 0.526 0.528 0.484
Extramarital Affairs 0.815 0.404 0.541 0.765 0.320 0.516 0.395 0.290
Wedding Proposal 0.637 0.541 0.585 0.634 0.378 0.498 0.429 0.426
62
improve the clustering accuracy.
The following is one possible procedure for crowdsourcing the clustering of sen-
tences. After the algorithm decides on a set of clusters of sentences, the clusters are
shown to a number of crowd workers who are tasked to inspect the cluster members
and pick out sentences that are dissimilar to other sentences in the same cluster.
Under sufficient agreement, a particular sentence can be removed from its cluster.
The removed sentences are put into a trash can. In the next step, we ask the crowd
workers (possibly different workers) to pick out one sentence as a summary of this
cluster. Finally, workers are asked to restore sentences in the trash back into the
clusters wherever possible, with the help of the cluster summaries.
Preliminary results in the bank robbery domain show such a procedure may im-
prove the purity of clusters to 89.8%, approximating agreement between humans.
However, determining the best method for crowdsourcing clustering is out of the
scope of this dissertation.
3.6 Learning precedence relations
Once we have discovered events that can happen in a given situation, the next stage
is to identify the structure of the plot graph that most accurately explains the set of
crowdsourced exemplar stories. The process begins with the identification of prece-
dence relations, which creates a partial ordering among the events. A partial order
does not require every event to be ordered with respect to every other event. Some
events may be unordered with respect to each other. When a story is generated, a
total order of the events consistent with the specified partial order can be selected
dynamically. Hence, a partial order allows sequences of events to change according
to the need of story generation or understanding and is less rigid compared to a total
order. The assumption of partial order is different from approaches that learns total
orders, such as Regneri et al. [147].
63
Chambers and Jurafsky [26, 27] also tried to infer temporal relations in the models
them learn. They train their models on the Timebank corpus [144], which uses
temporal signal words. Since our corpora are highly specialized to our task, the
precedence relations can be inferred directly from the sentence orders in the exemplar
stories.
I present two methods for learning the precedence relations. The first method
applies a global threshold to the confidence of all possible precedence relations, and
then attempts to restore precedence relations that might have been left out by a
conservative threshold. The second method keeps as many precedence relations as
possible, while maintaining the acyclicity of the graph. In practice, I find both meth-
ods have their merits. The first method helps to show the major relations in the
graph, allowing us to identify important graph structures. Graphs learned by the
first method can produce a greater variety of stories. The second method eliminates
cycles while preserving more precedence constraints. It produces more linear stories
with less variety, which helps to reduce errors in generated stories.
The precedence learning process selects valid precedence relations among all pos-
sible relations. I present two different methods for learning precedence relations, and
both relies on the notion of the confidence of each relations, as defined below.
For every pair of events (e
i
, e
j
), there are three possible ordering: e
i
is ordered
before e
j
; e
i
is ordered after e
j
; or the order between e
i
and e
j
is not important for the
situation. The three cases are denoted using the following mathematical notations:
e
i
e
j
, e
i
e
j
, and e
i
k e
j
. The reader may have noticed that the notation e
i
e
j
is equivalent to e
j
e
i
. When e
i
is ordered before e
j
, there may still be other events
ordered between e
i
and e
j
. We can choose to accept or reject the two hypotheses
e
i
e
j
and e
i
e
j
only one of the two can be accepted. If both are rejected, we
accept e
i
k e
j
.
We count the amount of evidence for and against the two hypotheses. Let s
l
be
64
a sentence in the event cluster e
i
, and s
m
a sentence in the event cluster e
j
. If s
l
and
s
m
both appear in the same exemplar story, and s
l
is ordered before s
m
, we count
one observation in support of e
i
e
j
. Conversely, if s
l
and s
m
appear in the same
exemplar story and s
m
appears before s
l
, we count one observation in support of
e
i
e
j
.
The confidence of e
i
e
j
is computed by a one-tailed hypothesis testing based on
the binomial distribution as
confidence(i, j) =
k
ij
1
X
i=1
n
ij
i
1
2
n
ij
(16)
where n
ij
is the total number of observations we have, and k
ij
is the observations
that support e
i
e
j
. Similarly, the confidence of e
i
e
j
is computed based on the
number of observations that support e
i
e
j
.
3.6.1 Method 1: Smart Thresholding
The first method accepts all hypotheses e
i
e
j
whose confidence exceeds a threshold
T
p
[0.5, 1). The threshold T
p
applies to the entire graph and allow us to generate an
initial estimate of the graph structure through simple counts. In practice, we find that
the graph quality is sensitive to the selection of these parameters. Selecting a set of
parameters that always work for the entire graph is often impossible as different parts
of the graph may respond well to different parameters. Thus, it is desirable for graph
estimation to be robust against parameter selection, and to locally relax the global
thresholds for some relations. We achieve this goal by using a high threshold for T
p
,
which filters out many potential precedences, and then restoring missing relations
back to minimize a measure of graph error, effectively relaxing the global threshold
locally.
Since a plot graph encodes event ordering, we introduce an error measure based
on the expected number of interstitial events between any pair of events. The error
is the difference between two distance measures, D
G
(e
i
, e
j
) and D
N
(e
i
, e
j
). D
G
(e
i
, e
j
)
65
is the number of events on the shortest path from e
i
to e
j
on the graph (e
i
excluded).
In contrast, D
N
(e
i
, e
j
) is the normative distance from e
i
to e
j
averaged over the entire
corpus of exemplar stories.
Formally, we find the shortest path from e
i
to e
j
and denote it as sp
i,j
. Recall
a path is a sequence of events he
i
, e
1
, e
2
..., e
j
i starting from event e
i
, ending in event
e
j
, and every adjacent pair is connected by a precedence relation. Similarly, we find
the shortest path from e
j
to e
i
and denote it as sp
ji
. If e
i
is ordered with respect to
e
j
, one of sp
ij
and sp
ji
exists. If e
i
is parallel to e
j
, neither sp
ij
nor sp
ji
exists. The
graph distance between e
i
and e
j
is thus defined as
D
G
(e
i
, e
j
) =
|sp
ij
|−1, if sp
ij
−|sp
ji
|+1, if sp
ji
0, otherwise
(17)
Since the shortest path contains at least two events e
i
and e
j
, D
G
(e
i
, e
j
) 1 if there is
at least one path from e
i
to e
j
. D
G
(e
i
, e
j
) = 0 if e
i
and e
j
are unordered with respect
to each other on the graph. D
G
(e
i
, e
j
) 1 if there is at least one path from e
j
to e
i
.
Having defined the graphical distance D
G
(e
i
, e
j
), we formally define the normative
distance D
N
(e
i
, e
j
) between two events e
i
and e
j
, as the average distance between the
two events over all exemplar stories. An exemplar story N
k
is a sequence of sentences
hs
k
1
, s
k
2
, ...i. Let the notation s
k
1
e
i
denote that sentence s
k
1
belongs to the event
cluster e
i
. The definition of D
N
relies on the definition of d
N
, simply the number of
interstitial sentences between any two sentences in the same narrative. That is, for
any s
k
a
, s
k
b
N
k
,
d
N
(s
k
a
, s
k
b
) = b a (18)
Next, we define the set S
ij
as the set of pairs of sentences that appear in the same
exemplar story and belong to e
i
and e
j
respectively:
S
ij
= {(s
k
a
, s
k
b
)|s
k
a
e
i
s
k
0
b
e
j
} (19)
66
Now D
N
(e
i
, e
j
) is just the average of d
N
(·) over the set S
ij
:
D
N
(e
i
, e
j
) =
1
|S
ij
|
X
(s
k
a
,s
k
b
)S
ij
d
N
(s
k
a
, s
k
b
) (20)
As the distance on the plot graph should approximate the average distance be-
tween the corresponding sentences in the exemplar stories, we aim to minimize the
mean squared graph error (MSGE), which is defined as the average squared distance
between the graph distance D
G
from the ground truth D
N
:
MSGE =
1
|P |
X
e
i
,e
j
P
(D
G
(e
i
, e
j
) D
N
(e
i
, e
j
))
2
(21)
where P is the set of all ordered event pairs (e
i
, e
j
) such that e
j
is reachable from e
i
or e
i
and e
j
are parallel on the graph (i.e. D
G
(e
i
, e
j
) 0):
P = {(e
i
, e
j
)|D
G
(e
i
, e
j
) 0} (22)
We note that for some data sets, removing any outlier sentences that do not belong
to any clusters from the stories before counting the interstitial sentences tends to
produce better results.
We utilize the MSGE error measure to improve the graph based on the belief
that D
N
represents the normative distance we expect between events in any narrative
accepted by the plot graph. That is, typical event sequences in the space of narratives
described by the plot graph should have D
G
(e
i
, e
j
) D
N
(e
i
, e
j
) for all events. A
particularly large deviation from the norm may indicate that some edges with low
confidence could be included in the graph to make it closer to user inputs and reduce
the overall error.
We implement a greedy, iterative improvement procedure that reduces mean
square graph error in a plot graph (Algorithm 1). For each pair of events (e
i
, e
j
)
such that e
j
is reachable from e
i
or the two events e
i
and e
j
are parallel, we compute
a set of potential predecessor events U:
U = {e
k
|e
k
successors(e
i
), D
G
(e
i
, e
k
) = D
N
(e
i
, e
j
) 1}
67
a
c
b
(a) Before
a
b
c
(b) After
Figure 7: Restoring low-confidence precedence relations: (a) The precedence relation
(shown as dashed arrow) from b to c has a confidence below the global threshold and
is initially rejected. (b) Adding the precedence back to the graph creates the desired
separation between a and c.
where successors(e
i
) is the set of all direct and indirect successors of e
i
. In other words,
if any e
k
U becomes the immediate predecessor of e
j
, D
G
(e
i
, e
j
) = D
G
(e
i
, e
k
) + 1
will be equal to D
N
(e
i
, e
j
). Starting from the pairs of events (e
i
, e
j
) with the largest
deviation from the norm, computed as D
N
(e
i
, e
j
) D
G
(e
i
, e
j
), we check if adding
an edge e
i
e
j
will create any cycles or increases the graph error MSGE. If not,
the edge e
i
e
j
is added to the graph. This intuition is illustrated in Figure 7
where the edge (dashed arrow) from event b to event c was originally rejected due to
insufficient confidence; adding the edge to the graph creates the desired separation
between events a and c.
Adding an edge may increase overall graph error. However, in some domains,
adding all edges that do not create cycles, regardless if adding them increases MSGE,
tends to produce graphs correspond better to human intuition. This is due to factors
not captured by the MSGE heuristic. If we skip that MSGE test, we obtain an
aggressive version of graph improvement.
68
Algorithm 1 Plot graph improvement
procedure ImproveGraph(G = hE, P, M
x
, E
o
, E
c
i)
P all event pairs (e
i
, e
j
) such that e
j
successor(e
i
) or that e
i
and e
j
are
unordered, i.e. D
G
(e
i
, e
j
) 0.
P Sort P in decreasing order of D
N
(e
i
, e
j
) D
G
(e
i
, e
j
).
for each (e
i
, e
j
) in P do
U successors e
k
of e
i
such that D
G
(e
i
, e
k
) = D
N
(e
i
, e
j
) 1
for each e
k
in U do
if edge he
i
, e
k
i / T then add new edge
Build a new graph G
0
= hE, T he
i
, e
k
i, M
x
, E
o
, E
c
i
if G0 does not contain cycles, and MSGE(G
0
) < MSGE(G) then
G G
0
Aggressive version skips the MSGE test
end if
end if
end for
end for
end procedure
We find a relatively high T
p
( 0.7) combined with the graph improvement step
leads to robust graph estimation. A conservative T
p
initially discards many edges in
favor of a lower-diameter graph with many unordered events. After that, the improve-
ment algorithm opportunistically restores the relations of different levels of evidence
as long as graph error can be reduced. This effectively relaxes the threshold locally.
Rare events are automatically excluded from the graphs because their relations to all
other events do not meet our probability and confidence thresholds.
3.6.2 Method 2: Integer Quadratically Constrained Programming
Even with the graph improvement step, which relaxes the global threshold somewhat,
sometimes it can be difficult to find a good value for the parameter T
p
. This issue is
further complicated by the requirement that the plot graph must be acyclic. To solve
these two issues, I propose the second method for learning precedence relations using
integer programming.
To avoid looping infinitely back to the same event, any plot graph must be a
directed acyclic graph (DAG). By setting T
p
> 0.5, we can make sure to reject one of
69
e
i
e
j
and e
j
e
i
, hence eliminating cycles of two events. There are no self cycles
because no event can precede itself. However, the graph may still contain cycles that
involve three or more events.
Cycles can always be eliminated by increasing the global threshold. Cheng et al. [32]
discussed two simple methods to find the minimum threshold to eliminate all cycles
from the graph. Nevertheless, increasing the global threshold affects the entire graph
and may discard good precedence relations that are not involved in any cycles. Thus,
it is difficult to learn an acyclic graph by only adjusting the global threshold.
In order to eliminate cycles in the plot graph and preserve as many precedence
relations as possible, I formulate the learning problem as an integer quadratically
constrained program (IQCP). IQCP is an NP-hard problem, but very efficient off-
the-shelve solvers, such as the Gurobi solver [75], are available. The general case of
cycle elimination (i.e. minimum feedback edge set) is also NP-hard and APX-hard
(i.e. difficult to approximate) [84]. Thus, formulating precedence learning as IQCP
is appropriate.
The formulation of the IQCP problem is based on the insight that vertices in a
DAG can be arranged into a number of layers where all directed edges go from a
higher layer to a lower layer. An illustrative example is shown in Figure 8, which
shows a directed graph containing 9 vertices organized into 5 layers. All edges go
from a higher layer (denoted with a smaller number on the vertices) to a lower layer
(denoted with a greater number on the vertices) except the red edge at the bottom
right, which creates a cycle. Eliminating edges going from lower layers to higher
layers guarantees that the graph is acyclic.
Consequently, we can formulate the IQCP problem as follows. We attach a level
variable l
i
, which can take any positive integer value, to each vertice e
i
. We also
attach a binary variable x
ij
{0, 1} for any precedence hypothesis e
i
e
j
. x
ij
= 1
if and only if we accept the hypothesis. We want to preserve as many precedence
70
1
22
333
44
5
Figure 8: Organizing vertices in a directed graph into multiple layers. All edges go
from a higher layer (denoted by a smaller number) to a lower layer (denoted with a
greater number) except the cycle-creating red edge at the bottom right.
relations as possible and prioritize hypothesis of higher confidence, so we maximize
the following sum of logarithm.
x
x
x = argmax
x
x
x
X
i
X
j
(log(conf(i, j)) log(T
p
)) x
ij
(23)
The reason that we subtract log(T
p
) is the following: log(x) is negative when 0 <
x < 1, so a direct sum of logs will be 0 or negative. However, we know T
p
is a lower
bound on the confidences of hypotheses that we accept. By subtracting log(T
p
), we
ensure log(conf(i, j)) log(T
p
) > 0, for all hypotheses we may accept. It is worth
noting that adding a constant C to log(x) shifts the function upwards and does not
change the shape of the function. We typically set T
p
to 0.5, so any precedences with
50% or lower confidence are rejected, but a higher global threshold may also be used.
In order to ensure the plot graph is acyclic, we need to respect the constraint:
l
j
x
ij
l
i
1, i, j (24)
In plain English, if we accept the hypothesis e
i
e
j
, x
ij
= 1, then the level variable l
j
must be greater than the level variable l
i
, so that the directed edge goes from a lower
level to a higher level. If we reject the hypothesis e
i
e
j
, the variable x
ij
will be set
71
to 0, and the constraint is always satisfied since l
j
is a positive integer. It is clear
that the objective function is linear but the constraints are quadratic with respect to
the variables x
x
x and l
l
l.
3.6.3 Practical Concerns In Learning Precedence Relations
It is worthwhile to compare the two methods for learning plot graphs. The smart
thresholding method first filters all precedence hypotheses with a hard global thresh-
old and then tries to compensate for it. The IQCP method typically uses a lower
hard threshold and removes precedence relations relatively lower confidence whose
inclusion will creates cycles. Therefore, the IQCP method takes a “softer” and more
flexible approach which retains more precedence relations. This is advantageous when
data are sparse. However, when data are sufficient, the IQCP method may appear
too lenient and keeping incorrect precedence relations that are only slightly above the
lower threshold. This may be compensated by a higher hard threshold.
In general, we face the classic trade-off between overfitting and underfitting, or bias
and variance, when choosing the correct filtering threshold. A low global threshold or
an aggressive precedence restoration algorithm, can accept more precedence relations,
and may potentially leads to overfitting. On the other hand, if a high threshold is
adopted, we may learn too few precedence relations and the model can underfit.
An overfit model possessing many precedence constraints will produce very limited
variations in story generation, whereas an underfit model will produce variations that
are too wild to make sense. As with most machine learning algorithms, it is generally
difficult to presciently determine the correct trade-off that works on all data sets and
all situations.
Since IQCP is NP-hard, it may not scale well to very large problems. However,
it works well for the size of the problems I deal with. As I showed in Table 3, all
situations have less than 60 gold standard events. Counting one level variable l for
72
each event and one acceptance variable x for each potential edge, there are at most
3660 variables, which are well within the capability of modern solvers and hardware.
On an Intel i5 (Haswell architecture) processor with 16 GB RAM, the Gurobi solver
[75] takes less than 1 second to solve the IQCP problems for every situation I have.
3.7 Learning Mutual Exclusion Relations
After identifying precedence relations, we further identify mutual exclusion relations
between the events. Mutual exclusion relations are a generalization of the AND/OR
labels introduced by Nelson and Mateas [121]. Many social and procedural situations
contain several possible outcomes or alternative ways to perform an activity. For
example, a bank robber may escape or be caught by police. A restaurant patron
may choose to dine-in or order to-go meals. Due to the presence of alternative or
branching event sequences, different exemplar stories may describe different events,
and the plot graph may contain several paths that are incompatible. In order to
separate incompatible events and create coherent legal event sequences, the learning
process identifies mutual exclusion relations.
In the plot graph representation, a bidirectional mutual exclusion relation between
two events indicates that the two events cannot both happen in a single story. Mu-
tual exclusion relations pose constraints on valid sequences of events, which help to
maintain story coherence during story generation.
Mutual exclusion relations are identified based on the mutual information between
events, a measure of their interdependence. Suppose E
i
{0, 1} is a random variable
indicating if event e
i
exists in an input narrative. The mutual information between
two events e
i
and e
j
is:
MI(E
i
, E
j
) =
X
E
i
∈{0,1}
X
E
j
∈{0,1}
p(E
i
, E
j
) log
p(E
i
, E
j
)
p(E
i
)p(E
j
)
(25)
where p(·) denotes the probability of the random variables. For example, p(E
i
= 1)
is the probability that event e
i
happens in a narrative, estimated as the ratio of input
73
narratives containing e
i
to the total number of narratives in the corpus. p(E
i
=
1, E
j
= 1) is the probability that e
i
and e
j
happen in the same narrative, etc. We
can also write the mutual information as:
MI(E
i
, E
j
) = C(0, 0) + C(0, 1) + C(1, 0) + C(1, 1) (26)
where
C(a, b) = p(E
i
= a, E
j
= b) log
p(E
i
= a, E
j
= b)
p(E
i
= a)p(E
i
= b)
(27)
The partial sum C(0, 1)+C(1, 0) expresses the tendency for the two random variables
to take on different values, or the tendency that the presence of one event predicts the
absence of the other event. We recognize two events to be mutually exclusive when
C(1, 0)+C(0, 1) > 0, which suggests mutual exclusion, and MI(E
i
, E
j
) is greater than
a threshold T
m
, which suggests strong interdependence between the two variables.
3.8 Learning Optional and Conditional Events
To explain the underlying rationale for having optional and conditional events, we
must jump forward a little in content to explain one of the several rules in story
generation (the topic is addressed in Chapter 4). The one rule that is relevant here
is that we assume that precedence constraints are equivalent to necessary conditions.
That is, if event a temporally precedes event b, event b can only happen after event
a has happened because event a establishes a necessary condition for event b.
Optional events are introduced to make sure every event in the plot graph has a
chance of appearing in at least one legal event sequence that may be generated from
the graph. Given two events a and b, there are several possible configurations:
1. Events a and b are temporally ordered but not mutually exclusive.
2. Events a and b are not temporally ordered but mutually exclusive.
3. Events a and b are both temporally ordered and mutually exclusive.
74
a c
b
(a)
a c
b
(b)
Figure 9: Identifying optional events: (a) Event a is optional and b is conditional.
(b) Event a is not optional due to the mutual exclusion relation between a and b.
The first two cases can be easily handled, but the third case creates a paradox.
Without loss of generality, let us assume event a is ordered before event b, or (a, b)
E. Due to this precedence relation, event b cannot happen until event a has happened,
but after event a has happened, event b will be excluded by the mutual exclusion
relation and still cannot happen.
To make sure event b can happen in some legal event sequences, we must relax our
assumption of necessity Instead of requiring event a to be a precondition for event
b, we do not consider event a to be a precondition for any event. That is, event a
becomes optional in the plot graph. If event a occurs in a narrative, event b will be
excluded, but if event a does not occur, then event b can happen and is still necessary
for its successor events. Event b becomes conditioned event a.
We need to be careful not to make events unnecessarily optional. One interesting
scenario is illustrated in Figure 9. On the left, we recognize event a to be optional
and event b to be conditioned on event a. On the right, however, event a is mutually
exclusive to another event c that is mutually exclusive to event a and is a predecessor
for event b. When event c happens, event a will be excluded from the plot graph,
and event b may be included in an event sequence. Thus, in this scenario, we do not
recognize event a to be optional or event b to be conditional.
Now let me introduce the notion of a clear path. Recall that a path from event e
i
to event e
j
in a plot graph is a sequence of events he
i
, e
1
, e
2
, . . . , e
j
i, where adjacent
75
a c
b
d
e
f
Figure 10: Identifying optional events with clear paths. Although event a is mutually
exclusive to event c, a predecessor of event b, there is no clear path from event c to
event b. Therefore, we must recognize event a as optional and event b as conditional.
events are connected by a precedence relation. The path is clear if no two events on
the path are mutually exclusive to each other. Moreover, the path is also clear if two
events e
p
and e
q
(p < q) on the path are mutually exclusive to each other, but e
p
has
been recognized as optional.
In general, when we have two events a and b, in order to recognize event a as
optional and event b as conditional, all of the following conditions must be satisfied:
Event a precedes event b, or (a, b) T
Event a is mutually exclusive to event b, or (a, b) M
x
There does not exist event c such that event c is mutually exclusive to event a,
and there is a clear path going from event c to event b, and a is not included in
this path. Or formally,
6∃ c E s.t. (a, c) M
x
ClearP ath(c, b) a 6∈ ClearP ath(c, b) (28)
It should be noted that this set of conditions are sufficient, but not necessary, to make
sure every event can happen in at least some event sequences.
To see why we need the notion of a clear path, let us consider the scenario in
Figure 10. We notice that event f is not optional because event f provides an
76
alternative path to event e. We also notice that event a is mutually exclusive to
event c, a predecessor of event b. However, the path from event c to event b is not
clear because it goes through the pair of events d and e. When event c happens, event
e cannot happen, so there is effectively no path from event c to event b. Thus, we
must recognize event a as optional and event b as conditional.
A polynomial-time algorithm for detecting optional and conditional events can be
built using the adjacency matrix representation of graphs. The adjacency matrix M
is constructed as follows:
M
ij
=
1 if there is precedence relation from vertex i to vertex j
0 if there is no precedence relations from vertex i to vertex j
(29)
Let M
2
= M × M, and M
3
= M × M × M, and so on. It is easy to show that if
and only if there is a path from vertex i to vertex j that contains one intermediate
vertex between i and j, M
2
ij
> 0. Otherwise, M
2
ij
= 0. This is because if there is
another vertex k such that M
ik
= 1 and M
kj
= 1, then M
2
ij
=
P
k
0
M
ik
0
M
k
0
j
> 0.
More generally, we can show M
n
ij
> 0 if and only if there is a path from vertex i to
vertex j that contains n 1 intermediate vertices.
The fact that the plot graph is acyclic helps us reduce the computation needed
because the length of longest path in the graph cannot exceed the number of vertices,
n. Therefore, to find if there is a path between vertex i and vertex j, it suffices to
check the entry at row i and column j in the matrices M
2
, M
3
, . . . , M
n1
. A path
exists if and only if the entry in any of these matrices is non-zero. We do not need to
compute matrices to the n
th
or higher powers, because a path containing more than
n 2 intermediate vertices cannot exist in the acyclic graph. To make sure we find
clear paths only, for each mutual exclusion relation between vertices e
i
0
and e
j
0
, we
set the entries at (i
0
, j
0
) and (j
0
, i
0
) to zero after each matrix multiplication.
The algorithm for identifying optional and conditional events is shown as Al-
gorithm 2. The most time consuming step in the algorithm is the computation of
77
Algorithm 2 Identifying optional and conditional events
procedure IdentifyOptional(G = hE, P, M
x
, E
o
, E
c
i)
M
1
an n-by-n zero matrix
for each (e
i
, e
j
) T do
M
1
ij
1
end for
for k 2 to n 1 do
M
k
M × M
k1
end for
P a empty list of pairs of events
for each (e
i
, e
j
) M
x
do
for k 1 to n 1 do
if M
k
ij
> 0 then
P P append (i, j)
break
else if M
k
ji
> 0 then acyclicity: M
k
ij
and M
k
ji
cannot both > 0
P P append (i, j)
break
end if
end for
end for
for each (i, j) P do
if 6∃ e
q
E, s.t. (e
q
, e
i
) M
x
and q 6= j and ClearPath(G, e
i
, e
q
, e
j
) then
O O e
i
C C e
j
end if
end for
end procedure
function ClearPath(G = hE, P, M
x
, E
o
, E
c
i, e
i
, e
q
, e
j
)
M
1
an n-by-n zero matrix
for each (e
a
, e
b
) T, a 6= i, b 6= i do the path cannot go through e
i
if (e
a
, e
b
) 6∈ M
x
then
M
1
ab
1
end if
end for
for k 2 to n 1 do
M
k
M × M
k1
for each (e
a
, e
b
) M
x
do
M
k
ab
0
end for
end for
return k, s.t. M
k
qj
> 0
end function
78
n 2 matrices M
2
, M
3
, . . . , M
n1
. The simplest matrix multiplication algorithm has
time complexity of O(n
3
). The widely used Strassen algorithm [175] has reduced the
complexity to O(n
2.807355
), even though more complex and faster methods have been
invented (e.g. the Coppersmith-Winograd algorithm [36]). Since there are O(n
2
) pairs
of vertices in the graph, we need to run the algorithm no more than O(n
2
) times.
Hence, it only takes polynomial time to find all optional and conditional events.
3.9 Evaluating the Learned Graphs
Before applying learned plot graphs in tasks of Narrative Intelligence, it is desirable to
check if the graphs make sense to human judges. In this section, I perform a quantita-
tive evaluation on some of the learned plot graphs. The learned precedence relations
in the plot graphs are evaluated by crowd workers from AMT. The mutual exclusion
relations and optional events are not quantitatively evaluated, as their meaning can
become opaque when evaluated without a proper context. However, they will be
indirectly evaluated when we evaluate the generated stories in Section 4.2.
Figure 11 and Figure 12 show the plot graphs learned by the smart thresholding
and the IQCP method respectively (referred to as the ST graph and the IQCP graph
hereafter). These plot graphs are learned from the gold standard clusters under the
assumption that we can further crowdsource the clustering of sentences and achieve
near perfect clusters. The event labels are English interpretations of each event for
presentation purposes only, based on manual inspection of the sentences in each event.
For clarity, edges that do not affect the partial ordering are omitted from the figure.
Close inspections of the plot graphs show that the two graphs do not differ in
the sequence of major events. The ordering of the key events, such as “buy tickets”,
“find seats”, “watch movie”, and “go home”, are mostly correct in both graphs.
Most precedence relations seem to be reasonable, although a few can be added or
removed to make the graphs to capture the ideal “date at a movie theater” situation
79
Sally enters car
John drives to
Sally’s
Buy tickets
Buy popcorn
and soda
Find seats
Enter theater
Drive to
theater
Buy
popcorn
Pick Sally up
Buy drinks
Lege nd
Normal event
Optional event
Conditional event
Precedence relations
Mutual exclusion
Buy
refreshments
Sit down
Talk
Movie begins
Eat popcorn
Drink sodas
Watch movie
Enjoy movie
Put arms
around
Hold hands
Movie ends
Stand up
John meets
Sally at the
theater
Leave theater
Walk to car
Use
bathroom
Kiss
Go home
Say goodbye
Park car
Figure 11: A plot graph for the movie date situation, created by smart thresholding.
80
Sally enters car
John drives to
Sally’s
Show tickets
Buy tickets
Buy popcorn
and soda
Find seats
Enter theater
Drive to
theater
Buy
popcorn
Pick Sally up
Buy drinks
Lege nd
Normal event
Optional event
Conditional event
Precedence relations
Mutual exclusion
Park car
Arrive at
theater
Buy
refreshments
Sit down
Talk
Movie begins
Eat popcorn
Drink sodas
Watch movie
Enjoy movie
Put arms
around
Hold hands
Movie ends
Stand up
John meets
Sally at the
theater
Leave theater
Walk to car
Use
bathroom
Hug
Kiss
Go home
Say goodbye
Figure 12: A plot graph for the movie date situation, created by the IQCP method.
81
more accurately. In general, we tend to see ordered relations when we expect causal
necessity, and we see unordered events when ordering variations are supported by the
data.
Comparing the two graphs, The IQCP graph seems to be more reasonable. For ex-
ample, the IQCP graph positions the event “park car” correctly but the smart thresh-
olding method does not. The IQCP graph also correctly positions “enjoy movie” be-
tween “movie begins” and “movie ends”, and positions “buy drinks” correctly before
“find seats”. The ST graph correctly identifies that John and Sally could put their
arms around each other most of the time during the movie, and they can drink the
soda anytime, whereas the IQCP method restrains the two events “put arms around”
and “drink sodas” more narrowly. However, the IQCP graph correctly identifies John
and Sally can talk during most time of the movie, but the ST graph puts the “talk”
event after “stand up” and before “kiss”. The ST graph also contains fewer events
than the IQCP graph because events not temporally related to any other events are
not included in the graph.
In general, the smart thresholding method errs on the side of omitting precedences
with low probability, and the IQCP graph errs on the side of including too many
precedences, so the ordering of events may be overly constrained. In the movie date
situation, whereas the storyline is mostly linear without significant branches and
contingencies, the IQCP method appears to work better.
The mutual exclusions detected are correct, but some are missing. The mutual
exclusion between “John meets Sally at the theater” and “go home” is because when
the two people meet at the theater, they usually say goodbye at the theater and the
story ends. If John picked up Sally, however, human writers tend to describe how
John and Sally go home together. We miss some mutual exclusion relations, such as
the one between “John drives to Sally’s” and “pick Sally up” because they are two
different descriptions of the same event. The missing mutual exclusions exist in both
82
graphs, as they use the same mechanism for detecting mutual exclusion relations.
3.9.1 Methodology
We evaluate the two graphs in the movie date situation empirically with human judges
recruited from Amazon Mechanical Turk. The learned precedence relations as well as
the absence of such relations are checked. From the ST graph, we randomly sampled
30 pairs of adjacent events, i.e. events in the automatically generated plot graph that
are ordered by a before relation without any interstitial events. We also randomly
sampled 29 pairs of parallel events, i.e. events for which the plot graph indicates no
necessary ordering relative to one another. From AMT, we recruited 144 workers.
Each worker was paid 0.120.20 to check seven pairs of events.
Each worker was instructed to consider each pair of events in the context of going
on a date to the move theater. Each pair of events (A, B) was presented to a worker
in a randomized order (50% of workers saw A before B and 50% of workers saw the
opposite) and workers were asked whether (a) it is more likely that A comes before B,
(b) it is more likely that B comes before A, or (c) that they are unable to tell which
should come first. In order to detect cheating or randomly clicking, two of the seven
pairs were designed as validation questions. These two pairs of events do not appear
anywhere in the plot graph, but were manually written and have obvious orderings.
If a worker provided a wrong answer on either of two pairs, all of his or her answers
were considered invalid and discarded. Each worker was allowed to participate in the
study only once.
3.9.2 Results
With the gold standard created by the crowd workers, we evaluate the two methods
for learning precedence relations. The results for the smart thresholding (ST) graph
and the IQCP graph are shown in Table 5 and Table 6 respectively. The differences
between the two tables, resulted from subtracting the ST results from the IQCP
83
Table 5: Accuracy of the learned precedence relations by the smart thresholding method
entropy= 0 entropy < 0.4 entropy< 0.6 entropy< 0.8 entropy<
acc. % pairs acc. % pairs acc. % pairs acc. % pairs acc. % pairs
All 0.76 29 0.64 42 0.66 54 0.54 73 0.53 100
Adjacent 1.00 40 0.93 50 0.90 67 0.82 73 0.70 100
Parallel 0.20 17 0.20 34 0.25 41 0.22 79 0.28 100
All-sans-ends 0.80 24 0.73 37 0.68 46 0.48 76 0.49 100
Adjacent-sans-ends 1.00 35 0.50 50 0.83 60 0.69 80 0.60 100
Parallel-sans-ends 0.33 14 0.40 24 0.43 33 0.25 76 0.33 100
84
Table 6: Accuracy of the learned precedence relations by the IQCP method
entropy= 0 entropy < 0.4 entropy< 0.6 entropy< 0.8 entropy<
acc. % pairs acc. % pairs acc. % pairs acc. % pairs acc. % pairs
All 0.88 29 0.76 42 0.72 54 0.64 73 0.57 100
Adjacent 1.00 37 0.94 47 0.88 60 0.79 76 0.70 100
Parallel 0.33 14 0.29 33 0.31 43 0.33 67 0.34 100
All-sans-ends 0.91 26 0.86 37 0.78 49 0.66 72 0.57 100
Adjacent-sans-ends 1.00 31 0.91 45 0.85 55 0.75 76 0.65 100
Parallel-sans-ends 0.50 14 0.64 21 0.53 36 0.44 64 0.40 100
85
Table 7: Differences in accuracy of the learned precedence relations by the two methods. The ST scores are subtracted from
the QICP scores.
entropy= 0 entropy < 0.4 entropy< 0.6 entropy< 0.8 entropy<
All 0.12 0.12 0.06 0.1 0.04
Adjacent 0 0.01 -0.02 -0.03 0
Parallel 0.13 0.09 0.06 0.11 0.06
All-sans-ends 0.11 0.13 0.1 0.18 0.08
Adjacent-sans-ends 0 0.41 0.02 0.06 0.05
Parallel-sans-ends 0.17 0.24 0.1 0.19 0.07
86
results, are summarized in Table 7.
The rows in those tables indicate subsets of the data. The first three rows show
the results from all sampled pairs, all sampled adjacent pairs, and all sampled parallel
pairs (the remaining rows are explained later). The columns measure accuracy—the
percentage of time human workers agree with our plot graphs—at different levels
of worker agreement. We measure human agreement on each pair of events as the
entropy of their answers. The entropy for the j
th
pair of events (a
j
, b
j
) is defined as:
entropy(X
j
) =
3
X
i=1
P (x
ji
) lnP (x
ji
) (30)
where x
ji
{before(a
j
, b
j
), before(b
j
, a
j
), parallel(a
j
, b
j
)}. The probability distribu-
tion P (X
j
) is observed directly from human responses for the pair (a
j
, b
j
). The
columns of Table 5 show statistics for event pairs with increasing entropy from left
to right (i.e. decreasing worker agreement). For example, the first column include
only pairs where workers unanimously agree (entropy = 0), which are 29% of all pairs
evaluated (row “All”), and of those 29%, workers agreed with the ordering in the ST
graph 76% of the time. Lowering the entropy threshold filters out pairs of events with
low agreement from consideration.
3.9.3 Discussion
We draw four sets of observations about our plot graph learning algorithm in the
movie situation:
Overall accuracy. The smart thresholding method yields an overall accuracy
over 53%, whereas the IQCP method has a higher accuracy of 57%. Both are
well above the purely random baseline of 33%. When we examine only pairs
for which workers perfectly agree with each other our accuracy is as high as
76% for smart thresholding, and 88% for IQCP, although this only accounts for
about 29% of our total sampled pairs. We found that when humans could not
87
reach consensus on a pair of events, they tend to also disagree with our system.
The IQCP method consistently outperforms the smart thresholding method in
overall accuracy.
Adjacent events. Our system is very accurate when it comes to determining
when a before relation should exist between a pair of events. Workers agree
with our before relations at about 90% of the time when they can reach good
consensus (entropy < 0.6). This suggests our algorithm is a good model of the
ground truth. Accuracy decreases as workers begin to disagree but remains
high (0.7-0.8) in the worst case. The IQCP method shows slightly lower
performance in the conditions of entropy < 0.6 and entropy < 0.8, suggesting
it recognizes precedence relations more aggressively than the other method.
Parallel events. For all parallel pairs, workers agreed with our system only
28-33% of the time. However, workers agreement is generally lower for parallel
events than adjacent events. Unanimous agreement can be reached on only 14-
17% of all pairs, in contrast to 37-40% for adjacent pairs. The lack of agreement
on many of these pairs suggests insufficient collective social expectation of the
orderings. The reason that individual worker may prefer one ordering to another
may be attributed to the way questions were asked (which ordering is more
likely). Even though one ordering is likely, the other ordering may be also
possible. Our results suggest that although we are missing before relations that
would eliminate parallel events our system may be correctly placing events as
parallel in the graph when there is very little agreement on ordering.
The IQCP method generally performs better on parallel events than the smart
thresholding method. As the IQCP method recognizes more precedences, it is
more selective in recognizing the lack of precedence relations, and thus produces
better results. However, the success of the IQCP method is also attributed to
88
the fact that the movie date situation contains a mostly linear storyline, without
significant branches and contingencies.
Removing events with sparse data. During the crowdsourcing process, I
observe that people start and end the exemplar stories at different points. Thus,
data about events at the beginning and the end are more sparse than rest of
the plot graph. It is reasonable to postulate that data sparsity has affected
the accuracy of learned precedences at the two ends of the graph. To test this
hypothesis, the last three rows of result tables show the results when we remove
all pairs involving events before “buy tickets” in the ST graph and three events
at the end: “go home”, “walk to car”, and “say goodbye” from the data. 43
pairs of events remain after these events are removed.
The removal significantly improves the accuracy for the detection of parallel
events, where improvements are seen in most conditions of human worker agree-
ments. Accuracy of parallel events improves by 3% to 13% in the ST graph,
and 6% to 35% in the IQCP graph. The overall accuracy improves when human
workers reach unanimous agreement by 3% to 4%. We do not see improvements
when human workers disagree with each other, but again it may be particularly
difficult, even for humans, to determine the correct relations between those
events.
The results suggest the data sparsity is a real concern for graph learning, and
there is room for the story acquisition procedure to address this issue. After the
removal of these events, the IQCP method outperforms the smart thresholding
method by an even higher margin, further suggesting the strength of the IQCP
method.
Overall, this evaluation demonstrates the system is capable of learning plot graphs
that is consistent with human intuition to an encouraging degree. When humans can
89
reach an agreement, they usually agree with the system. Our results show our accu-
racy percentage reaches high 80s to low 90s when humans reach unanimous agreement
among themselves. Even when we include all pairs of sentences where humans have
little consensus among themselves, our accuracy is still around 60%, well above the
purely random baseline of 33%. When humans cannot agree with each other, they
tend not to agree with the system, but that is more or less expected. Scarcity of data,
especially at the beginning of the story, can negatively affect the result of learning.
3.10 Limitations and Future Work
Many story generation and story understanding systems (e.g. [99, 128, 154, 201])
have demonstrated the utility of causal relations between events for tasks of Narrative
Intelligence. However, in this dissertation, I do not discover causal relations from the
exemplar stories. One difficulty I face in discovering causal relations is people tend
to omit very obvious causes and effects when writing the exemplar stories, even when
instructed not to do so. For example, in preliminary experiments people tend to
mention either “finding a table” and “sitting down” at a restaurant, because the two
actions are immediately causally related and almost always follow each other. Since
we only observe second-hand information, the most obvious causal relations tend to
be missing in the data and become difficult to detect
2
.
Ideally, one can learn causal conditions so they can be used to create plan-like
structures as in aforementioned Narrative Intelligence systems. However, the human
notion of causality remains vague. Some theories [31, 125] postulate that causal re-
lations can be captured by simple statistics, such as in the comparison between the
probability of an event e given a cause c, P (e|c), and the probability of the same
event in the absence of a cause, P (ec). Such causal relations may be learned by
controlled and unbiased experiments of counterfactuals [134]. On the other hand,
2
In the final experiment, we used those two events as an example in our instructions to workers,
telling them not to omit events, but the effects were limited.
90
other theorists [79] argue causality is perceived as the abnormal necessary condition.
For example, a train wreck has several necessary conditions, including that the train-
ing is traveling at high speed, that the rail is faulty, and that the train is heavy enough
to break out of the rail. Most people would consider the faulty rail as the cause of the
train wreck, as this condition deviates from what is believed as normal. This view
suggests learning a normality model, as done in this dissertation, may facilitate the
identification of causal relations.
The current model contains two types of relations: precedence and mutual exclu-
sion. Although the two types can capture a large number of situational variations, it
does distinguish between events that are happening simultaneously and events that
should happen sequentially with an indeterminate ordering. For example, the two
events “John and Amy lay in bed” and “John’s wife Sally opened the door” in the
extra-marital affairs situation happen simultaneously. The two events “John paid for
the food” and “The cashier passed John the food” in the fast food restaurant situ-
ation can happen in any order. The difference is that, in the latter case, the events
must happen in an particular order, but this order can be arbitrary. The reason my
approach cannot detect this difference is that I only observe exemplar stories that are
linear sequences. Simultaneous events and events with no specific orderings all have
to be linearized in the exemplar stories, and there is little observed information to
differentiate the two types of relationships. Additional annotations may be required
to learn simultaneity between events.
Clustering quality can further be improved by recruiting crowd workers to modify
the automatically identified clusters. The system also cannot identify events that
may happen in multiple locations and for more than once. These improvements are
left for future work.
91
3.11 Summary
In this section, I have presented a formal definition of plot graphs, which extends
traditional plot graphs used in interactive narrative systems, and includes events in
a situation, precedence relations and mutual exclusions between events, and events
deemed to be optional and conditional. I have also presented a crowdsourcing proce-
dure for acquiring suitable exemplar stories and algorithms for learning each compo-
nents of the plot graph from those exemplars.
The learned plot graphs have been evaluated in two ways: whether the learned
events match manually created gold standards, and whether the learned precedence
relations (or the lack of them) match the intuition of crowd workers recruited from
Amazon Mechanical Turk. The evaluation shows the learning algorithms achieve
satisfactory performance. The precedence learning algorithm achieves an accuracy
percentage in high 80s and low 90s, when human judges can reach unanimous agree-
ment.
However, the utility of learned plot graphs can only be tested by applying them
in tasks of Narrative Intelligence. In the next two chapters, I will present methods
for generating and understanding stories based on the learned graphs.
92
CHAPTER IV
GENERATING AND TELLING STORIES
No story comes from nowhere; new stories are born from old—it is the
new combinations that make them new.
Salman Rushdie
The utility of learned plot graphs can only be demonstrated by their use in sup-
porting various Narrative Intelligence tasks. In this chapter, I demonstrate story
generation and storytelling based on learned plot graphs. The pipelined generation
process covers all three tiers in Bal’s model [5]: fabula, sjuzhet, and text (or media).
The story generation/telling pipeline starts with the generation of a linearized
fabula. As noted in the previous chapter, my representation does not distinguish
between simultaneous events and sequential events whose order is determined at run
time. Thus, we directly generate linearized fabulas whose events are linearly se-
quenced. This generation is described in Section 4.1. Afterwards, we select some
events from the fabula in order to create interesting stories in Section 4.3. Finally,
in Section 4.4, we tackle the problem of generating story text for a given sjuzhet by
reusing crowdsourced sentences in the learned plot graphs. The generated fabula and
story tests are evaluated by human judges recruited on Amazon Mechanical Turk.
The evaluations are described in Sections 4.2 and 4.5 respectively.
4.1 Fabula Generation
In this section, I describe how the Scheherazade system generates a linearized
fabula from a plot graph. By showing what events may happen in in a social or
procedural situation and how these events are involved in a number of precedence
93
relations and mutual exclusion relations, a plot graph defines a space of linear event
sequences. Each sequence can be considered as a linearized fabula.
4.1.1 Legal Passages Through a Plot Graph
In a finite-state machine (FSM) defined by a directed graph, a valid sequence of states
is a walk through the graph that starts and ends in the correct states. A walk in a
directed graph is defined as a sequence of vertices where vertices adjacent in the walk
are also adjacent in the direct graph.
Unlike a finite-state machine, the directed edges or precedence relations in a plot
graph define a partial order, rather than transitions, between the events. Therefore,
in a plot graph, valid sequences of events are not walks. This difference is illustrated
in Figure 13. Figure 13(a) shows a finite-state machine where we can visit state
d from either state b or state c, so valid sequences include abd and acd. In a plot
graph with the same layout (Figure 13(c)), however, we can only visit event d after
both event b and c have been visited, creating two valid sequences abcd and acbd.
The partial-order graph compactly captures variations in the fabula caused by events
happening in different orders. To represent the two sequences in 13(c) as a FSM,
we would need to replicate the vertices b and c, as illustrated in Figure 13(b). In
contrast, we can easily represent the two sequences in Figure 13(a) as a partial plot
graph by simply adding a mutual exclusion relation, as illustrated in Figure 13(d).
The compactness of partial-order plot graph facilitates learning, as we do not need
to replicate the learned events in the graph. However, this representation does make
graph traversal more complex. In this section, I explain the procedures for traversing
plot graphs.
I call a valid sequence of event through the plot graph a passage. Since each
vertex in a plot graph represents an event, we refer to the action of visiting a vertex
as executing an event. To generate valid fabula is to generate passages through a plot
94
a
b
c
d
(a) Two walks (abd, acd)
in a FSM
a
b
c
b
c
d
(b) Two walks (abcd,
acbd) in a FSM
a
b
c
d
(c) Two passages (abcd,
acbd)
a
b
c
d
(d) Two passages (abd,
acd)
Figure 13: Contrasting passages in plot graphs with walks in finite-state machines.
The circles are vertices, and black solid arrows are directed edges. Red dashed lines
are mutual exclusion relations. The valid walks and passages are shown as dashed
blue arrows.
95
graph.
A valid passage is defined by the following rules that determine when an event
can be executed.
1. Incoming precedence relations to an event are treated as necessary conditions
for the execution of that event. Therefore, an event e can be executed if and
only if (1) all of its parents meet any one of the four conditions:
the event has no parents.
the parent has been executed.
the parent is optional.
the parent has been removed from the graph by a mutual exclusion relation.
and (2) that none of e’s children has been executed, as explained below.
2. Optional events may be skipped and are not considered to be necessary condi-
tions of any events. However, once an optional event e has been skipped and
one of its children has been executed, event e cannot no longer be executed.
This is because executing event e after one of its children violates a precedence
relation. Those events are called expired events.
3. When an event e is executed, by the definition of mutual exclusion relations,
all events that are mutually exclusive to e are removed from the plot graph.
Removed events cannot be executed.
4. Event removals by mutual exclusions are recursive. If all parents of an event e
have been removed from the plot graph because of mutual exclusion relations,
event e must also be removed.
As noted previously, mutual exclusion relations may be inadvertently omitted
due to the application of a universal threshold T
m
. This recursive deletion prop-
agates mutual exclusion from parents to children and compensates for omitted
96
mutual exclusion relations. This recursive removal continues until no events in
the plot graph meet the criterion for removal.
The recursive deletion is equivalent to adding mutual exclusion relations in the
plot graph. We determine the events being deleted at run time instead of adding
the relations when learning the plot graphs. This is because it is not always
clear what mutual exclusion relation should be added to the plot graph, due
to problems such as race conditions. Section 5.3 describes these problems in
details.
5. If an event e has more than one parents, and some, but not all, of the parents
have been removed, event e will not be removed from the plot graph. In this
case, parents of the removed events become parents of event e in order to avoid
losing structural information.
6. When an optional or conditional event e is removed by a mutual exclusion
relation, its children will not be removed. Parents of event e will become direct
parents of event e’s children.
7. We stop executing events when no events may be executed, or when we reach
one ending event in the plot graph. An ending event is an event with no children.
The algorithms for maintaining the legality of passages are shown as Algorithm 3
and Algorithm 4. By implementing rule 1, the ExecutableEvents function finds
the set of all executable events at any time. The set of all executable events at any
time is termed the fringe. The passage generation algorithm (which I will explain
in the next section) executes one vertex e
step
at one time. After e
step
is added to
the passage, the UpdateGraph function is called to bring the graph up to date.
The ExcludeEvents function computes events that should be removed by mutual
exclusion relations and their transitive closure. These events are directly removed
from the graph and will no longer be considered. To preserve structural information,
97
Algorithm 3 Graph Maintenance 1
function ExecutableEvents(G = hE, P, M
x
, E
o
, E
c
i, history, E
expired
)
fringe
for each event e E, e / E
expired
do
if (e
p
, e) P, e
p
history e
p
E
o
then
fringe fringe {e}
end if
end for
end function
function UpdateGraph(G = hE, P, M
x
, E
o
, E
c
i, history, e
step
)
E
expired
ExpiredEvents(G, e
step
, history)
E
excluded
ExcludedEvents(G, e
step
)
AddLinksAcross(G, E
excluded
)
E E \ (E
excluded
E
expired
)
return G
end function
procedure AddLinksAcross(G = hE, P, M
x
, E
o
, E
c
i, E
x
)
for each event e E
x
, if e / E
o
e / E
c
do
for each (e, e
c
) P do Add links between all predecessors
for each (e
p
, e) P do of e and successors of e
P = P {(e
p
, e
c
)}
end for
end for
end for
end procedure
according to the generation rule 5, parents of removed events are linked to the children
of removed events with precedence relations, as performed by the AddLinksAcross
function.
4.1.2 Generating Passages With a Search Algorithm
The fabula generation algorithm executes one event (or one vertex) at a time. During
this process, we add the executed events to the generated passage. Algorithm 5 shows
the fabula generation algorithm. In order to implement rule 2 and 6, the Gener-
ateFabula function first adds precedence relations from parents of optional events
to children of optional events, and precedence relations from parents of conditional
events to children of conditional events. This ensures the removal of any optional
98
Algorithm 4 Graph Maintenance 2
function ExcludedEvents(G = hE, P, M
x
, E
o
, E
c
i, e
step
E)
E
excl
for each mutual exclusion he
step
, e
x
i in M
x
do
E
excl
E
excl
e
x
end for
for each mutual exclusion he
x
, e
step
i in M
x
do
E
excl
E
excl
e
x
end for
repeat
E
old
E
excl
for each event e E do
parents direct parents(e)
if parents E
excl
then if all direct parents have been excluded
E
excl
E
excl
e then exclude e as well
end if
end for
until E
excl
== E
old
repeats until convergence
return E
excl
end function
function ExpiredEvents(G = hE, P, M
x
, E
o
, E
c
i, e
step
E, history)
E
expr
for each temporal order he, e
step
i in P do
if e / history then  e
i
has not been executed, but
E
expr
E
expr
e executing it now will violate a temporal order
end if
end for
end function
or conditional events by mutual exclusion relations will not recursively remove any
children. The function then generates n fabulas stochastically and selects the best
according to a given fitness function in EvaluateFitness.
The PassThruGraph function is responsible for generating one legal passage
through the plot graph. Selection from the fringe is performed by the function Pick-
Event. The selection could be deterministic or stochastic. When the heuristic is
stochastic, it selects events from a distribution over the fringe, which is the case shown
in Algorithm 5. Taken together, the stochastic heuristic can provide a local estimate
of which event might lead to a good fabula (however goodness is defined), whereas
99
Algorithm 5 Fabula Generation
function GenerateFabula(G = hE, P, M
x
, E
o
, E
c
i)
AddLinksAcross(G, E
o
)
AddLinksAcross(G, E
c
)
fabula
best −∞
for i = 1 n do Find the best of the n stories, only if
new-fabula WalkGraph(G) events are stochastically selected
value EvaluateFitness(new-story)
if value > best then
fabula new-fabula, best value
end if
end for
return fabula
end function
function PassThruGraph(G = hE, P, M
x
, E
o
, E
c
i)
fabula hi
while not IsCompleteStory(G, f abula) do
fringe ExecutableEvents(G, fabula)
e PickEvent(fringe) Select an event e according to a heuristic
fabula fabula + e Append event e to the story
G UpdateGraph(G, fabula)
end while
return fabula
end function
function IsCompleteStory(G = hE, P, M
x
, E
o
, E
c
i, history E)
E
end
find ending events
for each event e in E do
if @e
s
E, he, e
s
i T then
E
end
E
end
e Events without successors are ending events
end if
end for
return (history E
end
6= )
end function
100
the fitness function (i.e. EvaluateFitness) provides a global evaluation of the en-
tire fabula. The stochastic method of generation is computationally less efficient, but
may be convenient if it is difficult to provide an estimate of story quality based on
only part of the fabula. On the other hand, if we employ a deterministic mechanism
for event selection, it suffices to generate one story. If the plot graph is completely
correct, every legal event sequence should be coherent, but other aesthetics measures,
such as novelty, may differ for different event sequences.
The generation algorithm produces over a million different legal passages through
the plot graph for the bank robbery situation (Figure 15). The authorial leverage
[30]—the ratio of possible narratives to authoring effort—of our system is high con-
sidering the input contains only 60 examplar stories. The quality of generated fabulas
is assessed in Section 4.2.
4.2 Evaluating the Generated Fabula
This section describes a large-scale evaluation of the coherence of fabula generated
from the bank robbery plot graph as shown in Figure 15. As all legal event sequences
should be coherent in a completely correct plot graph, this evaluation provides a mea-
sure for the quality of the learned plot graph (including precedence, mutual exclusion,
and optional events) in addition to the quality of the generation procedure.
4.2.1 Methodology
This evaluation focuses on the coherence of generated fabula. Although coherence
is arguably the most important measure, it is certainly not the only possible quality
measure for a story. In a traditional approach for story generation, a knowledge
engineer may carefully design the knowledge to guarantee the coherence of the story.
In a machine learning approach, however, I do not have much freedom to adjust the
learned representation. Thus, it is important to test if it is possible to create coherent
stories from the learned representation.
101
Figure 14: The user interface that human judges used to edit fabula generated by the
system.
102
Although the notion of coherence can be subjective, I attempt to measure it
quantitatively as the minimum number of edits that human judges need in order to
make a fabula coherent. Human judges read the fabulas as a list of sentences and
make changes to render the stories more coherent. Judges could delete or reorder
using a drag-and-drop interface. They can also write up to three new events to be
inserted into the story, but could not specify the location of insertion. To detect
random clicking and cheating, one obviously incorrect event was inserted in every
fabula. A judge’s response is only accepted if that event is correctly deleted. The
web interface used by the human judges is shown in Figure 14.
450 people were recruited on AMT and paid 20-30 cents as compensation. 60
participants edited the 60 stories from the crowdsourced corpus. 100 additional stories
were uniformly sampled from all possible fabulas that could be generated by the plot
graph. No fitness functions to filter the generated fabulas were used. 300 participants
edited generated fabulas such that three judges saw each fabula.
To establish a baseline, another 30 fabulas were generated by uniformly sampling
from all events in the plot graph. That is, no learned graph structures, including any
precedence relations, mutual exclusion relations or optionality of events were used.
Therefore, these 30 fabulas were most likely to be illegal according to the plot graph.
The length of the random fabulas were set to 23, which is approximately the average
length of fabulas generated from the plot graph. Each of the 30 fabulas were edited by
3 human workers, resulting in 90 edited fabulas. For all computer-generated fabulas,
the most frequent sentence of the underlying natural language cluster was selected to
describe each event in the text presented to the human workers.
The number of added events and deleted events from the original fabula presented
to the human worker can be easily obtained from data. However, obtaining the num-
ber of movements is a little more complex, as counting the number of mouse clicks
or drag-and-drops will inevitably include human operations that were performed by
103
John enters
bank
John drives to
bank
Sally screams
John points gun
at Sally
John
approaches
Sally
John hands Sally
a note
Sally puts
money in bag
John shows
gun
John pulls out
gun
Sally is
scared
John waits
in line
John scans
bank
Sally reads
note
John opens
bank door
The note
demands money
Sally opens cash
drawer
John demands
money
John covers
face
John sees
Sally
Sally greets
John
John gives
Sally bag
Sally collects
money
Sally gives
John bag
John takes
bag
Sally gives John
money
John collects
money
Sally cries
Police arrests
John
John gets in
car
John drives
away
John leaves
bank
Sally calls
police
Sally presses
alarm
Police arrives
Lege nd
Normal event
Optional event
Conditional event
Precedence relations
Mutual exclusion
Figure 15: A plot graph for the bank robbery situation, created by the smart thresh-
olding method. T
p
= 0.75, T
m
= 0.05
104
mistake and later undone. To avoid this problem, an algorithm is used to determine
the number of movements required to change the original event sequence to the edited
sequence submitted by a human worker. The algorithm computes the number of in-
verted pairwise orderings in the original event sequence associated with each event. It
then picks the event with most inverted pairwise orderings, places it in a position that
minimizes inverted orderings, and repeats until the two sequences become identical.
The number of movements performed by this algorithm is recorded.
4.2.2 Results
As each edit suggests a problem with the fabula, less edits indicate better quality. The
number of events added, deleted, or moved are shown in Table 8. Average numbers of
additions, deletions, and reorderings after normalizing for story length are also shown.
The Welsh t-test was used to determine if the difference between human-authored and
computer generated fabulas is statistically significant at p < 0.05. In the Computer
column, I denote whether the differences between the human-written fabulas and the
computer-generated fabulas are statistically significant. In the Random column, the
first symbol (i.e. before the comma) denotes statistical significance between the ran-
dom fabulas and the human-written stories, and the second symbol denotes statistical
significance between the random fabulas and the computer-generated fabulas.
I find pronounced differences between the random fabulas and other two condi-
tions. For most measures, the differences are strongly statistically significant with the
only exception in the number of added events after normalizing for story length. On
average 2.29 events were deleted from the random fabulas, which is three times the
number of deleted events in computer-generated fabulas. The number of movements
for the random fabulas is also roughly three times as the AI condition.
Compared to the random baseline, for most edit metrics, the human and the
AI conditions appear to be similar. No significant differences exist in number of
105
Table 8: Statistics of edits made by human judges to the generated fabulas.
*
indicates
the difference between the two columns is statistically significant. indicates the
difference between the two columns is not statistically significant.
Human Computer Random
Mean original length 12.78 23.14
*
23.00
Mean final length 11.82 21.89
*
20.71
Mean events added 0.33 0.49 0.80
*,*
Mean events added (normalized) 0.028 0.021 0.035
,*
Mean events deleted 0.30 0.76
*
2.29
*,*
Mean events deleted (normalized) 0.02 0.03 0.10
*,*
Mean events deleted (2 events withheld) 0.28 0.27 -
Mean events deleted (2 withheld, norm.) 0.02 0.01 -
Mean events moved 0.57 4.88
*
14.67
*,*
Mean events moved (normalized) 0.04 0.21
*
0.64
*,*
added events between the two conditions. After normalizing for length, the computer
condition appears to have less added events than the human condition. The exception
is in the number of moved events, where the differences across all three conditions
are large and statistically significant. However, some reorderings could be consistent
with the plot graph, as discussed in the next section.
I find a small but statistically significant difference in the number of events deleted
between computer-generated fabulas and human-written fabulas. Although statisti-
cally significant, the mean difference between conditions is less than half of an event.
The significance vanishes when two events “wait in line” and “open door” are with-
held. The two events account for 64.5% of deletes, and occurred rarely in the corpus
(6 and 9 times out of 60), which explains why the system has less certainty about
their inclusion. After normalizing for length and withholding, the AI has less deletions
than human authors.
There is a clear and statistically significant difference in the length of the stories
106
for the two conditions. The computer generated stories contains on average 23.14
sentences, which is almost twice as long as the human-written story. This is because
the learning algorithm of plot graphs merge events from many exemplar stories, so
the plot graph contains more events than any single exemplar stories. I will discuss
and address this issue in Section 4.3.
4.2.3 Discussion
With the help of the random baseline, I conclude that this novel method of measuring
story coherence can indeed capture differences in story coherence. The number of
movements appear to be the most sensitive measures. In comparison to deletion and
movements, the number of added events is less sensitive, although is still reflective of
the overall coherence. There are several possible reasons for this observation. First, it
could be due to the fact that the user interface allows easy drag-and-drop operations
for re-ordering and deletion, but no drag-and-drop for adding events. Second, only
three blanks were provided for addition, which restricted editing but also prevented
human workers from deviating from the typical story. Third, as the random fabulas
were fairly long with 23 events, the human workers could have found the events they
needed and did not need to add new events.
I conclude that the story generation algorithm does not omit essential events any
more often than human authors. I conclude that despite the existence of multiple
incompatible alternatives throughout the plot graph, the system was able to differen-
tiate between them and does not add events that contradict human intuition. This
result is attributed to the use of mutual exclusion relations to successfully separate
incompatible events.
There are several reasons why a judge may have reordered events. The fact that
these events are not deleted indicates that they contribute to the story but are not
ideally positioned. The precedence relations may be under- or over-constraining the
107
events, leading to incoherent fabula. Alternatively, I also find some edits not to be
strictly necessary for story coherence, indicating judges may prefer one ordering to
another out of aesthetic concerns rather than strictly coherence reasons.
Overall, we find 32.3% of moves to be consistent with the plot graph, indicat-
ing that the reordering exists in another legal story and the occurrence of under-
constraining. The rest of the moves violate temporal constraints in the plot graph,
indicating events being over-constrained. However, the changes may be due to aes-
thetic rather than coherence reasons. For example, “pressing the alarm” is over-
constrained to occur in the second half of the story. Two moved events account for
a plurality of inconsistencies: “get in car” is uniformly moved from the end to the
beginning, but both positions seem reasonable; “Sally cries” is a rare event in the
corpus. Removing these two events reduces the inconsistencies to 44% of generated
stories.
Leslie Kaelbling (personal communication) proposed a hypothesis that if a story
is excessively incoherent, human judges may find it difficult to make changes at all.
However, our observations suggest the human judges tend to make more changes
than necessary. This may be explained by the special incentive provided by the
AMT, where the employer has total control over whether to pay a worker or not.
Thus, workers may be motivated to do a little more work than they consider to be
bare minimum, just in order to appeal to the employer.
In conclusion, the fabulas generated from learned plot graph is largely on par with
human-written stories in the aspect of added and removed events. The plot graph
likely contains some errors that over- or under-constraints the ordering of events, al-
though we should not interpret all reorderings as errors. To my knowledge, this is a
first study showing computer-generated stories may match human-written stories in
some ways and is thus very encouraging. Considering the fact that these stories were
108
generated as random passages through the plot graph, this indicates the structures, in-
cluding the precedence relations, mutual exclusion relations, and optional/conditional
events, in learned plot graphs are of high accuracy and quality.
4.3 Sjuzhet Generation
In Section 4.2, we have seen that the generated fabula are usually longer than human-
written stories. The reason that humans do not include every event in their story is
probably that some events are too obvious or too mundane to be told. From the events
being told, a human reader is able to infer that those untold events have happened.
For example, if a storyteller says she found a seat in the movie theater, in the lack of
contradictory evidence, the audience would assume that she sat down. This inference
process is critical for story understanding [72, 207], and successful storytellers know
how to make use of this inference to their advantage. However, Scheherazade
learns scripts that include most events in the situation. By following the scripts, the
system tends to generates overly verbose stories.
Filtering out some of the less interesting events in the fabula is a commonly used
sjuzhet technique. Narratologists Barthes [8] and Chatman [29] both noted that
events in a narrative carry different importance. Barthes used the term nuclei to
describe important, plot-driving events, and the term catelyzers to describe unimpor-
tant events adding to the aesthetics of the story. Chatman used the term kernels and
satellites to refer to the same concepts. The ability to differentiate important events
from others has been considered as an integral part of Narrative Intelligence. For
example, a comprehensive questionnaire for measuring narrative abilities of children
developed by Heilmann et al. [78] includes the evaluation if the child includes critical
events and deemphasizes minor events. Cheong and Young [34] use the kernel/satellite
differentiation to manipulate sjuzhets and generate suspenseful stories.
In this work, I generate sjuzhets consisting of both kernel events and satellite
109
events. Following Bathes and Chatman, I postulate that an effective storytelling
strategy is to tell a set of kernel events that establish the situation and major storyline,
plus a set of interesting satellite events. Kernel events set up the situation and form
a major causal chain in the story. They create a context without which the story
would become difficult to understand. Example of kernel events are “we went to a
restaurant”, “we ordered steaks” and “we finished the meal”. On the other hand, we
can create interesting satellite events as infrequent and atypical events in a situation,
such as “the waiter complained the tip was too little”. Therefore, the central issue in
sjuzhet generation becomes the computation of how typical an event is to a particular
situation. My strategy for generate interesting sjuzhet selects the k most typical and
k least typical events to produce a story of 2k length.
I introduce an algorithm called EventRank, which determines the typicality of
events in a plot graph, taking into consideration of the size of the event cluster, the
structure of temporal orderings, as well as mutual exclusion relations. The algorithm
is inspired by Personalized PageRank [76], which computes the importance of vertices
contained in a strongly connected directed graph structure, as explained in the next
two sections.
4.3.1 PageRank and Stationary Distributions of Digraphs
The PageRank algorithm [18] was developed to measure the importance of webpages
on the Internet, which is modeled as a directed graph with webpages and hyperlinks.
Imagine a random algorithm that visits one webpage at one time epoch and uni-
formly randomly follows hyperlinks on the current page. The webpages then form a
Markov chain, where the webpage visited in the current epoch depends only on the
page visited in the previous epoch. In a network of n webpages, the webpage visited
at time epoch t is a categorical distribution, parameterized by a n-dimensional vector
x
x
x
(t)
. x
(t)
i
is the probability that we visit page i at time t. By measure theory, we
110
require that the sum of all entries of x
x
x
(t)
equal to 1 for all t.
N
X
i=1
x
(t)
i
= 1, t (31)
The change in the page being visited over time can be captured by a transition matrix
A, whose entry A
ij
denotes the probability of transiting from webpage j to webpage
i. Again, the probability of transiting from webpage j to all webpages must be equal
to 1.
N
X
i=1
A
ij
= 1, j (32)
Given no prior preferences for any single outgoing link to others, we can assign equal
weight to each outgoing link.
A
ij
=
1/d
j
if there is a directed edge from vertex j to vertex i
0 if there is no directed edges from vertex j to vertex i
(33)
where d
j
is the out-degree of vertex j. It is obvious that the webpage being visited
at the next time epoch can be computed as multiplication by A.
x
x
x
(t+1)
= Ax
x
x
(t)
(34)
If the graph represented by A is irreducible and aperiodic, it can be shown that
as t approaches infinity, the distribution of the current webpage will approach a
distribution x
x
x
()
, regardless of her starting position x
x
x
(0)
.
x
x
x
()
= lim
c→∞
A
c
x
x
x
(0)
, x
x
x
(0)
(35)
x
x
x
()
is called the equilibrium distribution of the Markov chain formed by the web-
pages. In an irreducible graph, we can find a path from any vertex to any other
vertex. In an aperiodic graph, we return to the same vertex at irregular intervals.
For a formal discussion of stationary distributions of Markov Chains, the reader is
referred to textbooks on Markov Chain Monte Carlo such as [158].
111
Given A, x
x
x
()
can be easily computed as the eigenvector of A corresponding to the
eigenvalue 1, which is also the biggest eigenvalue A. Thus, x
x
x
()
is a property of the
network structure captured by transition matrix A and is independent of the initial
distribution x
x
x
0
. Intuitively, when we have randomly wandered in the network for
sufficient time, where we started off would have little bearing on our current position.
PageRank takes x
x
x
()
i
to be the importance of the i
th
webpage.
It is important to maintain the property of irreducibility and aperiodic. For exam-
ple, if the network graph is not strongly connected and contains several components,
we will not reach one component from the other. Thus, where we start becomes
important. If the graph is periodic, we will always return to the same webpage after
a period T , so where we start will also matter. To avoid these situations, PageRank
allows random jumps from any page to every other page with a small probability.
That is, we add another uniform matrix B to A.
˜
A = λA + (1 λ)B (36)
where
B
ij
=
1
n
(37)
and take the eigenvector of
˜
A, which is guaranteed to have a unique equilibrium
distribution. The constant λ is usually set to 0.85.
The Personalized PageRank [76] is a variant of PageRank that utilizes a simple
insight: The added matrix B (in Equation 37) does not have to be uniform. We
can bias the matrix and increase the importance of certain webpages, in order to
incorporate information other than the graph structure. The next section describes
how we can incorporate extra information of event frequency and mutual exclusion
relationships to determine the typicality of events in a situation.
112
4.3.2 The EventRank Algorithm
The precedence relations and vertices in a plot graph form a directed acyclic graph,
which is not strongly connected. To satisify the requirement of strong connectivity,
precedence relations are added between each ending event (events with no outgoing
links) and each starting event (events with no incoming links). The transition matrix
A can then be constructed based on the precedence relations according to Equation 33.
In addition to the precedence relations, we want to consider two pieces of informa-
tion when computing the typicality of events in a situation. The first is the frequency
of events being mentioned in crowdsourced exemplar stories, since one story typi-
cally only mentions a subset of all events. Events that are mentioned more often are
probably more typical to a situation. Second, mutual exclusion represents alternative
branches in a plot graph, and they should have effects on typicality values. Specifi-
cally, we would like mutual exclusive events to weaken each other, and symmetrical
branches to have similar typicality values. The information is incorporated into the
PageRank algorithm by modifying the matrix B.
As we favor events that are mentioned more frequently in the exemplar stories,
when events i and j are not mutually exclusive, we let
B
ij
f
i
(38)
where f
j
is the frequency of event i being mentioned in the exemplar stories. If events
i and j are mutually exclusive, however, we penalize event i (and by symmetry, event
j) by letting
B
ij
f
i
τo
j
(39)
where
o
i
=
X
k /X
i
f
k
card(X
i
)
(40)
and X
i
is the set of vertices mutually exclusive to vertex i, and card(X
i
) is its car-
dinality. τ is a value in [0, 1), and I typically set it to between 0.3 and 0.5. The
113
rationale behind the above formula is that if an event e
i
has fewer mutually exclusive
events, it is more likely to be included in a story and hence more powerful in weak-
ening other events. The power of weakening is evenly distributed among all events
mutually exclusive to e
i
.
Finally, we normalize B so that each column sum up to 1. We again compute
the stationary distribution x
x
x
()
by finding the eigenvector with the eigenvalue 1 of
the matrix
˜
A = λA + (1 λ)B. Figure 16 shows the bank robbery plot graph with
the typicality values of each event. λ is set to 0.7 and τ set to 0.5. The structure
of the graph is generated by the smart thresholding method. For the purpose of
visualization, the median of the typicality values has been normalized to 1.
We can make two observations from the diagram. First, structural features of
the graph are reflected by the typicality values. For example, vertices that multiple
paths converge onto, such as ”John approaches Sally”, and ”John leaves bank” are
considered to be typical. Second, comparable alternative events also receive simi-
lar typicality values. The vertices ”John pulls out gun” and ”John hands Sally a
note” denote two major situational variations and receive similar values. The two
alternative endings ”John drives away” and ”Police arrests John” also receive similar
values. Our observation suggests the EventRank algorithm does capture the struc-
tural information in plot graphs, including both the precedence relations and mutual
exclusion relations.
4.3.3 Generating Different Sjuzhets
Table 9 shows events in a fabula and the corresponding typicality values, generated
from the bank robbery plot graph. Employing the typicality values in the selection of
events, we can create different sjuzhets. For example, we can create a short summary
by picking some of the most typical events from a fabula. Shown in bold in Table 9,
the top five most typical events in the bank robbery situation are “John approaches
114
John enters
bank 2.55
John drives to
bank 0.85
Sally screams
0.40
John points gun
at Sally 0.73
John
approaches
Sally 4.09
John hands Sally
a note 1.04
Sally puts
money in bag
2.78
John shows
gun 0.40
John pulls out
gun 1.13
Sally is
scared 1.61
John waits
in line 1. 57
John scans
bank 0.74
Sally reads
note 0.51
John opens
bank door 0.93
The note
demands money
0.27
Sally opens cash
drawer 0.69
John demands
money 1.37
John covers
face 0.48
John sees
Sally 0.78
Sally greets
John 1.89
John gives
Sally bag 0.69
Sally collects
money 2.03
Sally gives
John bag 1.05
John takes
bag 1.02
Sally gives John
money 0.57
John collects
money 1.23
Sally cries
0.58
Police arrests
John 1.76
John gets in
car 1.36
John drives
away 1.28
John leaves
bank 2.74
Sally calls
police 0.76
Sally presses
alarm 1.23
Police arrives
0.98
Legend
Normal event
Optional event
Conditional event
Precedence relation
Mutual exclusion
Figure 16: The typicality of events in the bank robbery situation. τ is set to 0.5.
115
Table 9: Selecting events by typicality from a fabula to create sjuzhets. The five most
typical events are shown in bold, and the five least typical events are underlined.
Event Typicality
John drives to bank 0.85
John opens bank door 0.93
John enters bank 2.55
John scans bank 0.74
John waits in line 1.57
John sees Sally 0.78
John approaches Sally 4.09
Sally greets John 1.89
John pulls out gun 1.13
John points gun at Sally 0.73
Sally screams 0.4
Sally is scared 1.61
John demands money 1.37
John gives Sally bag 0.69
Sally collects money 2.03
Sally puts money in bag 2.78
John collects money 1.23
Sally presses alarm 1.23
John leaves bank 2.74
Sally cries 0.58
Police arrests John 1.76
Sally”, “Sally puts money in bag”, “John leaves bank”, “John enters bank”, “Sally
collects money”. Ordering the five events in the order they appear in the fabula, they
form a mostly intelligible summary of the bank robbery situation. The summary
does not mention John demanding money or collecting money, but those might be
inferred.
In the terminology of Chatman [29], the most typical events can be considered
as kernel events, and the least typical events as satellite events. Therefore, we may
make the sjuzhet more interesting sjuzhet by adding some of the least typical events
116
to the most typical events. The most typical events, or the kernels, establish the
social situation, so the readers can understand the context. The least typical events,
or the satellites, make the story more interesting. The top 5 least typical events in
this fabula include: “Sally screams”, “Sally cries”, “John points gun at Sally”, “John
scans bank”, and “John drives to bank”, which are underlined in Table 9. Take the
summary as the baseline, adding the top 5 least events to the top 5 most typical events
reveals emotional states of Sally and arguably makes a more interesting sjuzhet.
4.4 Textual Realization
The final step of story generation is to describe the events in a selected medium. In
this dissertation, the medium of choice is text. This section discusses how to generate
the textual realization of distinct styles for a given sjuzhet by selecting sentences
from event clusters. There are several motivations for generating text with different
styles. First, a narrative may be told to achieve different communicative goals. For
example, if a storyteller wants to cheer up her audience, she may employ a positive
tone. In contrast, the telling of horror stories may benefit from a negative storytelling
style making more use of words inciting fear and suspense. Second, a narrative
can be told by different narrators in different styles. Consider the application of
virtual characters. Depending on the circumstances, mood and personality, a virtual
character may choose to speak very succinctly, or to be very talkative. The technique
of focalization utilizes multiple viewpoints in telling one narrative (see Section 2.2.2
for a review). Therefore, developing parameterized methods for telling stories in
different styles can provide practical benefits.
In this section, I first describe an approach for crowdsourcing interesting sentences.
After that, I discuss criteria for selecting textual description of individual events, fol-
lowed by an Viterbi-style algorithm that aligns adjacent sentences to improve textual
coherence.
117
4.4.1 Textual Interestingness
We investigate two aspects of language that affect the interestingness of generated
stories. The first is the amount of details provided, and the second is the degree that
the story language resembles the language used in fictions.
We first model the amount of details as the probability of a sentence in English,
as Information Theory suggests a less likely sentence should contain more informa-
tion and therefore more details. We compute the probability of an English word as
its frequency in the Google N-Gram corpus. Due to the large size of the corpus,
these frequencies approximate word probabilities in general English. We compute the
probability of a sentence using the bag-of-word model where each word is indepen-
dently drawn from a multinomial distribution. Thus, the probability of sentence S
containing words w
1
, w
2
, . . . , w
k
each appearing x
1
, x
2
, . . . , x
k
times is
P (S) =
P
k
i=1
x
i
!
Q
k
i=1
(x
i
! )
k
Y
i=1
P (w
i
)
x
i
(41)
where P (w
i
) is the probability of the word w
i
. In this experiment, I used the average
frequency over the 10-year period of 1991 to 2000 in the “English 2012” corpus to
compute P (w
i
). Stop words are removed before computation.
We further consider the style of language is as how much it resembles fictional
novels. The language used in fictions has distinctive word choice as fictions tend to
accurately describe actions (e.g. “snatch” instead of “take”), emotions, and make less
use of formal words (e.g. “facility”, “presentation”). If a word appears more often in
fiction books than in all books, we can presume that its use may create a sense that
a story is being told in a literary manner. Therefore, the fictionality of a word w is
the ratio
f
w
= P
fic
(w)/P (w) (42)
where P (w) is the probability of a word computed previously and P
fic
(w) is the
probabilities of a world appearing in the “English Fiction 2012” corpus from the
118
Table 10: Fictionality of example words computed from the Gooogle N-Gram corpus
Word POS Fictionality
goodbye NN 11.24
sugary JJ 10.04
hobbit NN 8.31
laughing NN 8.19
softly RB 7.02
feature NN 0.23
menu NN 0.23
presentation NN 0.23
restoration NN 0.23
produce VB 0.20
appropriate JJ 0.20
Google N-Gram corpus. See Table 10 for ficitionality for some example words. Words
with low fictionality, such as “feature” and “restoration”, tend to appear in technical
or business documents.
The fictionality of a sentence is aggregated from fictionality values of individual
words as an exponentiated average:
fic(S) =
P
wW
exp(αf
w
)
card(W )
(43)
where W is the set of words in sentence S, and card(W ) is its cardinality. α is a
scaling parameter. The exponential function puts more weights on more fictional
words so that a few highly fictional words are not canceled off by many words with
low fictionality.
Table 11 shows some results of our heuristics for determining the interestingness of
sentences. We observe that the most probable sentence (MostProb) usually provides
a concise summary for the event. The most fictional (MostFic) sentence usually
contains more subjective emotions and character intentions, as indicated by the words
“smirk”, “nervously” and so on. The least probable (LeastProb) sentence is usually
longer and contains more objective details.
119
Table 11: Sentences selected from event clusters using the probability criterion, the
fictionality criterion, and their harmonic mean as the MID criterion.
Example event 1: John covers face
·
MostProb: John put on a fake mustache.
·
LeastProb: John kept his head down as he pulled open the outer
door and slipped his Obama mask over his face.
·
MostFic: MF: John looked at his reflection in the glass of the door,
gave himself a little smirk and covered his face.
·
MID: John kept his head down as he pulled open the outer door
and slipped his Obama mask over his face.
Example event 2: Sally puts money in bag
·
MostProb: Sally put 1,000,000 in a bag.
·
LeastProb: Sally put the money in the bag, and collected the money
from the 2 tellers next to her.
·
MostFic: Sally quickly and nervously stuffed the money into the
bag.
·
MID: Sally quickly and nervously stuffed the money into the bag.
Example event 3: John drives away
·
MostProb: John drove away.
·
LeastProb: John pulled out of the parking lot and accelerated,
thinking over which route would make it easier to evade any police
cars that might come along.
·
MostFic: John sped away, hoping to get distance between him and
the cops.
·
MID: John sped away, hoping to get distance between him and the
cops.
As a balance between the most fictional and the least probable, we combine the
two selection criteria and create the heuristic for the most interesting details (MID)
by using the harmonic mean. We first rank each sentence under the least probable
and the most fictional criteria: r
LP
and r
MF
. That is, the least probable sentence has
r
LP
= 1 and so on. The mean rank is:
r
MID
=
2 r
LP
r
MF
r
LP
+ r
MF
(44)
The sentence with the lowest r
MID
is picked as the sentence with the most interesting
120
details.
4.4.2 Smooth SentiWordNet
A story may be told in positive or negative tones in order to achieve different com-
munication goals, or to tell the story from the perspective of characters with different
mood. For instance, a positive tone may be used to cheer up the audience, whereas
a negative tone may be suitable for telling horror stories.
To detect sentiments of natural language, in this section I describe a corpus-based
technique for detecting the sentiment of English words and sentences. The technique
builds off SentiWordNet [51], which tags each synset (word sense) in WordNet [116]
with three values: positivity, negativity, and objectiveness, the three summing to
1. SentiWordNet was created by propagating known sentiments of a few seed words
along word relationships in WordNet to provide good coverage of words. While this
automatic approach creates good coverage, I also find that it produces many erroneous
values, resulting in unreliable sentiment judgments.
I propose an unsupervised, corpus-based technique to correct errors found in the
original library and expand its coverage beyond words appearing in WordNet. The
underlying intuition is that words in the same neighborhood, including adjacent words
and words in the same sentences and the same paragraph, should share similar sen-
timents, allowing us to automatically “smooth” any errors in the original sentiment
library. In addition, words closer should have a stronger influence than words farther
away.
A review of similar lexicon expansion techniques is provided in Section 2.3. In
comparison to existing techniques, my method makes use of full narrative text to
consider long-distance influences and considers the fact that word association tends
to diminish as we move farther away from a word. The use of a fiction-only corpus
produces a narrative-specific lexicon, as we know sentiments can change depending
121
Table 12: Examples of fictional books obtained from Project Gutenberg
Gutenberg ID Book Title
11 Alice’s Adventures in Wonderland
12 Through the Looking-Glass
15 Moby Dick
16 Peter Pan
24 O Pioneers!
27 Far from the Madding Crowd
32 Herland
33 The Scarlet Letter
35 The Time Machine
36 The War of the Worlds
41 The Legend of Sleepy Hollow
42 The Strange Case of Dr. Jekyll and Mr. Hyde
44 The Song of the Lark
45 Anne of Green Gables
46 A Christmas Carol in Prose; Being a Ghost Story of Christmas
47 Anne of Avonlea
51 Anne of the Island
54 The Marvelous Land of Oz
55 The Wonderful Wizard of Oz
on the context [108].
We obtained 9108 English books from Project Gutenberg (http://www.gutenberg.
org) that are labeled as fiction. Some of the book titles are shown in Table 12. The
complete list of books can be downloaded at http://boyangli.co/SSWN/booklist.
txt. These books are tagged with parts of speech (POS) and lemmatized with the
Stanford POS Tagger [186]. Each pair of lemma and POS is considered a unique
word. For every occurrence of a target word we want to compute sentiment value for,
we consider a neighborhood of 100 words, i.e. 50 to the left and the right of the target
word. The target word is at position 0. The words to its immediate left and right
are at position -1 and 1, and so forth. Only nouns, verbs, adjectives and adverbs in
122
complete sentences in a neighborhood can influence the target word. The index set
W includes their position. For a word w
i
at position i W , we place a Gaussian
kernel function g
i
centered at its position, which indicates the influence of word w
i
on a word at position j:
g
i
(j) = exp
(i j)
2
d
(45)
where the parameter d determines how fast the function diminishes with distance,
and is empirically set to 32. In the k
th
neighborhood, the sentiment s
k
w
0
of the target
word is computed as a weighted average of all kernel functions at position 0:
s
k
w
0
=
P
iW
k
s
swn
w
i
g
i
(0)
P
iW
k
g
i
(0)
(46)
where s
swn
w
i
is the sentiment retrieved from SentiWordNet, i.e. the difference between
the positive and negative sentiments for w
i
. The SentiWordNet value for any word
has no influence on itself, i.e. 0 / W
k
, k. The final sentiment value for the target
word s
w
0
is the average of all its occurrences in the corpus.
s
w
0
=
P
i=1...K
s
k
w
0
K
(47)
We aggregate sentiments of individual words in sentence S, again using the exponen-
tial average:
fic(S) =
P
wU
sign (f
w
) exp (β|f
w
|)
card(U)
(48)
where card(U) is the cardinality of set U, which contains any noun, verb, adjective or
adverb in that sentence. β is a scaling parameter. The exponential function ensures
that words expressing strong sentiments are weighted more heavily than words with
weak sentiments.
Using this corpus-based technique, a new sentiment dictionary is constructed,
which I call Smooth SentiWordNet (SSWN). I selected a subset of English words that
are of interests to the storytelling task to be included in the dictionary. The exemplar
stories in two previously crowdsourced social situations—dating at the movie theater
123
Table 13: Some most positive and most negative words in Smooth SentiWordNet
Word POS Sentiment
bone-chilling JJ -8.004
shriek NN -6.738
scop VB -6.734
wail VB -6.301
panic VB -5.679
sob NN -5.642
scream NN -5.316
nightmare NN -4.954
terror NN -4.261
panic NN -4.048
weep VB -3.840
enjoyable JJ 1.821
deal NN 1.974
beam VB 1.984
admire VB 2.005
immensely RB 2.028
sexy JJ 2.139
captivate VB 2.393
small-framed JJ 2.992
nacho NN 3.149
snack VB 3.277
tryed (sic.) JJ 4.621
and bank robbery—contain 1001 unique nouns, verbs, adverbs and adjectives. From
the corpus of fiction books, the aggregated influences of each adjectives and adverbs
on their neighbors are measured. Highly influential adjectives and adverbs were added
to the dictionary, producing a total of 7559 words. After computing the raw sentiment
values for these words, we normalize the values so that 1 percentile and 99 percentile
of the values fall in the range of [1, 1], to account for outliers. Twenty-two most
positive and negative words in SSWN are shown in Table 13. For the two hundred
most positive and negative words, See Appendix A. The entire dictionary can be
124
Table 14: Example sentences selected from event clusters with positive and negative
sentiments
Example event 1: Sally puts money in bag
·
Positive: Sally continued to cooperate, putting the money into the
bag as ordered.
·
Negative: Sally’s hands were trembling as she put the money in the
bag.
Example event 2: Sally cries
·
Positive: Sally cried, somewhat relieved it may be over soon.
·
Negative: Sally felt tears streaming down her face as she let out
sorrowful sobs.
Example event 3: Sally calls police
·
Positive: Sally described John as best as she could to the police.
·
Negative: Still shaken, Sally reached for the phone and in a pan-
icked manner called the police.
Example event 4: John opens bank door
·
Positive: John took a deep breath and opened the bank door, letting
an elderly woman exit before he entered himself.
·
Negative: John opened the bank door while his heart was beating
fast.
Example event 5: John pulls out gun
·
Positive: John pulled out the gun, still smiling.
·
Negative: John reached behind his back and withdrew his pistol.
downloaded at http://boyangli.co/SSWN/dictionary.txt.
We can observe that most of these words are put into the correct category. The
order of negative words are also fairly accurate, e.g. “wail” is more negative than
“sob” and “weep”. The positive words contain a few more misclassifications such as
“small-framed” and “beam” than negative words. Noises resulted from typos in the
original text (e.g. “tryed”) and errors in lemmatization (e.g. “captivate” was probably
lemmatized from “captivating”). Overall, the most positive and most negative words
seem qualitatively largely accurate.
125
Table 14 shows some of the most positive and most negative sentences. We find the
results to reflect the valences of individual words. In example events 1-3, individual
words like trembling or relieve dominate the entire sentence, and we can correctly
identify positive and negative sentences. In example event 4, elderly and woman have
positive valences, which coincide with the semantic meaning of the sentence. However,
there are also cases where the aggregation of individual words valences deviates from
the semantic meaning of the sentence. In example 5, the positive value of smile is the
main reason for selecting the positive sentence, but smiling criminals may appear even
scarier than usual. Due to language phenomena such as sarcasm and readers’ personal
judgments, the aggregated sentiment for a sentence is not 100% accurate. I describe
a human study that quantitatively evaluates Smooth SentiWordNet in Section 4.5.3.
4.4.3 Connecting Sentences
For each event, we can find individual sentences ranked highest for any criterion or
combinations of criteria using the harmonic mean. However, this selection does not
consider the coherence between sentences and may results in incoherent texts due to
two major problems: (1) previously mentioned objects can suddenly disappear and
previously unmentioned objects can appear, and (2) a sentence can repeat actions in
the previous sentence.
To address these problems, we propose a Viterbi-style algorithm, which consid-
ers both selection criteria for individual sentences and the interconnection between
sentences. In a hidden Markov model, the Viterbi algorithm generates a sequence
of hidden variables that best explains a sequence of observed random variables. The
algorithm relies on two things specified in a hidden Markov model: One, the prob-
abilities of a hidden variable generating any observation. That is, the observation
indicates preference over the values of a hidden variable. Two, the probabilities of
any hidden variable transiting to the hidden variable in the next time slice. That is,
126
Algorithm 6 A Viterbi-Style Algorithm for Generating Story Text
function GenerateText(event sequence hc
1
, c
2
, . . . , c
n
i)
for each sentence s
k
{s
1
, s
2
, . . . , s
m
} in event cluster c
n
do
(seq
k
, score(seq
k
)) BestSeqEndingIn(s
k
, c
n
)
end for
return the highest scored sequence from seq
1
, seq
2
, . . . , seq
m
end function
function BestSeqEndingIn(s
i
, c
j
)
for each sentence s
p
{s
1
, s
2
, . . . , s
m
} in event cluster c
j1
do
(seq
p
, score(seq
p
)) BestSeqEndingIn(s
p
, c
j1
) stored previously
new seq
p
seq
p
+ s
i
score(new seq
q
) score(seq
q
) + score(s
q
, s
i
) + score(s
i
)
end for
best seq the highest scored sequence from new seq
1
, . . . , new seq
m
return (best seq, score(best seq))
end function
we have preferences over pairs of values for adjacent variables.
Our problem is similar as we want to find the highest scored sentence sequence
based on preferences over sentences in each event cluster, as well as preferences on
how adjacent sentences connect. We do not consider connection between non-adjacent
sentences. Specifically, we score the connection between any two sentences s
i
, s
j
as
score(s
i
, s
j
) = log
shared nouns(i, j) + 1
shared verbs(i, j) + 1
(49)
where shared nouns(i, j) is the number of nouns shared by the two sentences, and
shared verbs(i, j) is the number of verbs shared by the two sentences. Similarly, we
score individual sentences as the reciprocal of their ranks according to any selection
criterion c:
score(s
i
) =
1
rank
c
(s
i
)
(50)
Our algorithm is shown as Algorithm 6. The BestSeqEndingIn function is
recursive, because in order to find the best sequence ending in a given sentence s
j
i
from the j
th
event cluster c
j
, we need to consider the scores of best sequences ending
in every sentence from the previous cluster c
j1
, in addition to the connection between
127
every sentence from cluster c
j1
and s
j
i
. Due to the Markov property, we do not need
to consider previous clusters c
1
, . . . , c
j2
. We can then iterate over every sentence from
cluster c
j
to find the best sequence ending in cluster c
j
. A dynamic programming
approach can be used to store every sequence ending in every sentence from every
cluster and their scores.
Computing the time and space complexity is straightforward. Suppose we have
a sequence of n clusters and m sentences in each cluster. We first need to compute
the score of each sentence, which takes O(mn) time. In addition, we need to consider
the connection between the m sentences in one cluster and the m sentences in the
next cluster, which takes O(m
2
) time. This computation is repeated for the n 1
connections. Therefore, the total time complexity is O(mn)+O(m
2
(n1)) = O(m
2
n).
At each iteration, we need to store the best sequence ending in each of the m sentences
in the current cluster. Each sequences contains at most n sentences, one from each
cluster. Therefore, the total space complexity is O(mn).
4.5 Evaluating the Generated Story Texts
In this section, I present user studies aimed to evaluate the stories generated using
different narrator styles and sentiments. The user study also evaluates the accuracy
of Smooth SentiWordNet.
4.5.1 Crowdsourcing Colorful Textual Descriptions
Before generating stories, I performed a second round of crowdsourcing as an attempt
to collect interesting event descriptions for each learned event cluster. In Section 3.2, I
presented the protocol for crowdsourcing exemplar stories from AMT. For the purpose
of learning plot graphs, crowd workers are instructed to write stories in simple and
bland language. Though simplified language facilitates plot graph learning by side-
stepping many hard natural language processing problems, it is not conducive to
generating vivid or sentimental speech. Therefore, I need to perform a second round
128
Table 15: Statistics of the additionally crowdsourced stories to enhance the interest-
ingness of storytelling
Movie Date Bank Robbery
# Stories 20 10
# Sentences 470 210
# Words per sentence 14.53 13.7
# Verbs per sentence 2.36 2.6
of crowdsourcing to acquire sentences that would interest human readers.
Once again, I recruited crowd workers on Amazon Mechanical Turk. Each worker
was shown the events that constitute a complete story to help them understand the
story context. For the compensation of 1, they are asked to write detailed de-
scriptions for each of these events. They were cued, but not required, to describe
characters’ intentions, facial expressions and actions. All answers that describe the
correct events were accepted. No filtering based on subjective judgment of the inter-
estingness of the answers were performed. The crowdsourcing was performed for the
movie date situation and the bank robbery situation.
Some statistics of these newly crowdsourced stories are shown in Table 15. The
statistics show that the newly crowdsourced sentences are longer and use more verbs
than the sentences in the original corpora intended to simplify learning. Some of
these sentences are shown in Table 11 and Table 14. In Table 11, the least probable
and most fictional sentences are mostly from this data set, whereas the most probably
sentence typically comes from the original, simplified language exemplars.
4.5.2 Evaluating Generated Story Texts
With the newly crowdsouced colorful sentences added to the event clusters, I now
describe an human study that quantitatively evaluates whether the narration styles
I defined is consistent with human readers’ intuition.
129
4.5.2.1 Methodology
A total of 52 undergraduate, masters, and doctoral students participated in our study.
Each participants read two groups of stories in both the movie date situation and the
bank robbery situation.
The two groups of stories were generated with the Viterbi-style algorithm with
different sentence selection criteria. The newly crowdsourced sentences were added to
the event clusters. As revealed in the previous study on story coherence (Section 4.2),
Scheherazade may generate stories with coherence issues such as incorrectly or-
dered events. In order to avoid the influence of story coherence, I manually edited the
event sequence to maintain its coherence. Discourse planning was not used to avoid
confounding factors. All stories used in this user study can be found in Appendix B.
The first group of stories includes stories generated from the the most interesting
details (MID) criterion, the most probable (MostProb) criterion, and a story where
we use the MID criterion but penalize long sentences. After reading the stories,
participants are asked to select the most interesting story, the most detailed story
and the most concise story. Our hypothesis is that human readers will select the MID
story as containing the most details and the most interesting, and the MostProb story
as the most concise. We set α to 12. The third story, where we use the MID criterion
but penalize long sentences, is used as a control condition. The least probable criterion
tends to favor long sentences. One hypothesis is that by penalizing long sentences,
we could describe similar amount of details in fewer words. The third story is added
to test this alternative hypothesis.
The second group of stories includes a story with the most positive sentiments, and
a story with the most negative sentiments. We set β to 16 and 2 for the movie data
and bank robbery situation respectively. After reading the second group, participants
are asked to select a positive and a negative story. We hypothesize human readers
will agree with the algorithm’s sentiment judgments.
130
Table 16: Accuracy of the detected story-level textual interestingness, conciseness,
and sentiments. denotes p < 0.0001. * denotes p < 0.0005.
Participant Agreement %
Test Movie Date Bank Robbery
Most Concise Story 90.38 75.00
*
Most Detailed Story 97.92 100.00
Most Interesting Story 88.46 80.77
Positive/Negative Stories 86.54 92.31
4.5.2.2 Results
Table 16 shows the percentage of human participants that agree with the algorithm.
All results are predominantly positive and consistent with my initial hypotheses. In
the movie date situation, the participants agreements are in upper 80s and 90s. In
the bank robbery situation, there is lower agreement on which story is most concise,
but unanimous agreement on which story is the most detailed.
The hypotheses are tested using a one-tailed hypothesis testing based on the multi-
nomial/binomial distribution. For the three stories in group 1, the null hypothesis
that human judges select the stories randomly suggests that the participants agree-
ment should be close to 1/3. The second group contains only two stories, so the
random baseline is 1/2. The statistical tests find we can reject the null hypotheses
at very high confidence levels of p < 0.0001 and p < 0.0005. The alternative hypoth-
esis that penalizing for length may maintain the level of interestingness has not been
supported.
4.5.2.3 Discussion
In conclusion, the heuristics I developed for creating diverse narrator styles were
shown to accurately capture the human intuition of interestingness, conciseness, and
positive versus negative sentiments. The learning of narrator styles from large data
131
sets can be considered successful.
However, it is arguably easier to detect the sentiment of an entire story than to
detect the sentiment of individual sentences, because a few sentences labeled with
wrong sentiments, when mixed together with many correctly labeled sentences in a
long story, may be overlooked by the participants in the study. To further evaluate
the constructed sentiment lexicon, Smooth SentiWordNet (SSWN), I also conducted
a sentence-level study, as detailed in the next section.
4.5.3 Evaluating Smooth SentiWordNet
This user study aims to evaluate whether I can predict the sentiments of individual
sentences using the sentiment lexicon Smooth SentiWordNet.
4.5.3.1 Methodology
From 45 event clusters taken from both situations, we first compute the top 3 most
positive sentence and top 3 most negative sentences. Each participant saw a pair
of positive and negative sentence randomly selected from the top 3, and was asked
to choose the positive and negative sentence. A total of 52 undergraduate, masters,
and doctoral students participated in our study. They performed 4678 comparisons
of 265 unique pairs of sentences. SSWN labels a sentence as positive if it has higher
sentiment than the median sentence in a cluster, and negative if it is lower.
The participants’ responses created a gold standard for the sentences we selected.
Based on these labels, I compared SSWN with the original SentiWordNet lexicon
as well as the the sentiment detection technique by Socher et al. [169] from Stanford
University. To compare with SentiWordNet, the word sentiment values in Equation 48
were substituted with values directly taken from SentiWordNet. A positive or negative
label can then be produced for each sentence. I tuned β to maximize performance.
To compare with Socher et al.’s method, I input the sentences into their web demo
and took the overall label. The results are summarized in Table 17.
132
Table 17: Accuracy of the detected sentence-level sentiment, with comparison to
SentiWordNet and the technique by Socher et al.. denotes p < 0.0001.
Participant Agreement %
Test Smooth SWN SentiWordNet Socher et al.
Sentence Sentiments (Individual) 70.76 59.60 35.91
Sentence Sentiments (Majority) 80.75 64.53 39.25
4.5.3.2 Results
Results are shown as Table 17. Overall, 70.76% of participants’ decisions agree with
the results produced by SSWN. The majority opinion on each pairs of sentences
agree with our algorithm for 80.75% of the time. In addition, SSWN outperforms
SentiWordNet by a margin of 11.16% to 16.22%, and outperform Socher et al.’s
technique by 34.85% to 41.5%, but it is worth noting Socher et al.’s algorithm targets
movie reviews and has not been tuned on our data set. A Chi-Square test shows the
difference between the two conditions to be extremely statistically significant at the
level of 0.0001.
4.5.3.3 Discussion
Results of this experiment indicates that SSWN can predict the overall sentiment of
sentences much more accurately than the original SentiWordNet. The results strongly
suggest that the corpus-based technique used to create SSWN is able to correct er-
rors in SentiWordNet. The automatic method used by SentiWordNet propagates
sentiment values along relations between words in WordNet, whereas my technique
propagates sentiment values along neighborhoods in texts in a selected corpus. These
two techniques can complement each other in creating a lexicon of high pratical utility.
133
4.6 Limitations and Future Work
While there are many sjuzhet creation techniques, in this dissertation I only consider
one of the most important techniques: the ability to omit certain events in order to
improve the interestingness of generated stories. I have not addressed the question
whether the ordering of events could be re-arranged to achieve certain aesthetic effects
such as suspense.
When creating an interesting sjuzhet, I consider how typical an event is to a
situation. Typical events are used to establish a situation. However, the proposed
EventRank algorithm does not consider if an event appears in many different sit-
uations. If an event is typical to many different situations, it may not be able to
correctly establish the intended situation. A probabilistic formulation may differen-
tiate P (event|situation) and P (situation|event). The evaluation of the EventRank
algorithm is future work.
I proposed an algorithm for textual coherence based on noun agreements and verb
differentiation. Text generation using more sophisticated coherence measures such as
that proposed by Barzilay and Lapata [9] is also left for future work.
4.7 Summary
In this section, I present algorithms for generating stories based on learned plot
graphs. The ability to generate stories and tell stories demonstrates the utility of the
learned plot graphs.
Following the pipelined three-tier narrative model, the Scheherazade system
sequentially generates fabulas, sjuzhets, and natural language texts of a story. I
first define several graph-walking rules for generating fabulas, and the EventRank
algorithm for evaluating the typicality of an event in a situation. After that, I pro-
pose heuristics for generating different storytelling styles by selecting sentences from
crowdsourced sentences and build a sentiment dictionary, Smooth SentiWordNet,
134
based on a large corpus of fiction books. Finally, I propose a Viterbi-style algorithm
for considering both individual sentences and connection between sentences to create
a coherence text.
I performed user studies to evaluate if the generated fabulas are coherent and if
the natural language texts can express different narration styles, including the inter-
estingness of the text, the degree that it resembles languages in fiction, and positive or
negative sentiments. The user studies show that (1) the fabulas generated are mostly
coherent, event matching human-written stories on some measures of coherence, and
(2) the sentence selection heuristics and the sentiment decisions made based on SSWN
strongly correlate with human intuition. These positive results demonstrate the high
quality of the learn knowledge representation and the versatile storytelling capabilities
of the Scheherazade system.
The field of computational story generation and storytelling has been associated
with the field of computational creativity (cf. [42, 66, 100]). Boden [13] has proposed a
classification of three types of creativity: combinational, exploratory, and transforma-
tional. If we consider the plot graph as defining a space for all legal event sequences,
Scheherazade explores sequences in this space and may be considered to be ex-
ploratory creativity. On the other hand, the plot graphs combines stories from differ-
ent exemplar stories. Scheherazade selects events from a plot graph, and describes
these events by selecting from crowdsourced sentences. Therefore, Scheherazade
may also be considered to possess combinational creativity.
The work in this chapter has made important contributions to generative Narra-
tive Intelligence. To the best of my knowledge, this is the first Narrative Intelligence
system that can generate all three tiers of narratives, including fabula, sjuzhet, and
natural language text, from automatically learned knowledge. The user studies high-
light the quality and variety of generated stories, suggesting that the learning algo-
rithm for plot graphs is effective. The system provides a framework that integrates
135
both plot-level and language-level generation of stories and provides flexible controls
for the style of narration to be adjusted according to the needs of applications.
136
CHAPTER V
UNDERSTANDING STORIES
The ninety percent of human experience that does not fit into established
narrative patterns falls into oblivion.
Mason Cooley
In previous chapters, I demonstrate the utility of the learned knowledge represen-
tation (i.e. plot graphs) by implementing and evaluating techniques that generate
and tell stories based on the learned knowledge. In this chapter, I further demon-
strate that the Scheherazade system can understand stories based on plot graphs,
which capture patterns of common situations. In particular, I focus on the problem
of inferring the fabula from a given sjuzhet, defined as the capability U2 in narrative
intelligence in Section 1.2. As discussed earlier, when a human tells a story, only a
selected set of events that happened in the narrative world are told. The events being
told constitute the sjuzhet. As an audience trying to fully understand the story, an
AI system needs to infer the events that happened, which constitute the fabula.
I frame story understanding in terms of questions and answers. Answering ques-
tions has been a popular method for computational systems to demonstrate story
understanding (c.f. [37, 73]). The Scheherazade system answers questions regard-
ing events that have not been told to the system based on events that haven been told
to the system. I first formally define the story understanding problem, and then prove
its computational complexity to be NP-hard. After that, I present methods to reduce
the computational complexity and solve the problem efficiently. The performance
gains are evaluated against different sets of random graphs.
137
5.1 The Story Understanding Problem
As explained in Section 2.1, the story being told by human storytellers is a sjuzhet,
not a fabula. Humans rarely tell every event that happened in the narrative world,
but only those are needed for an interesting story. Thus, a common type of narrative
understanding problem is to figure out what really happened in the narrative world
based on what has been told to the AI. This ability to infer events that happened
in the past and events that will happen in the future based on limited perception is
important for making sense of the world. Consider the following example questions:
Sally loved John. John asked Sally to marry him. Did Sally say yes?
John covered his face and entered a bank. Later, John ran away from the bank.
Did John demand money from the bank teller?
John demanded money from the bank teller. The bank teller pressed the silent
alarm. Did John escape from the bank?
In a restaurant, John ordered food. After a while, John paid and left. Did John
get food?
The above questions first posit that some events in the story domain have happened.
The system needs to determine the probability of some other events happening, and
should provide answers such as “the probability of John demanding money from the
bank teller is 98%”, or “It is very likely that John got food”. It is worth emphasizing
that these questions are about if an event happens in the fabula, not if the event is
included in the sjuzhet. The ability to answer those questions demonstrate that the
system possesses Narrative Intelligence about the domain.
Recall our definition for a plot graph G = hE, P, M
x
, E
o
, E
c
i, where E, T, M
x
, E
o
, E
c
are the sets of events, precedence relations, mutual exclusion relations, optional
events, and conditional events respectively. As explained in the previous chapter,
138
a legal event sequence is constructed by walking the plot graph according to these
simple rules:
1. An event e is eligible to be added to the sequence (or simply, event e “happens”)
if it has not parents or all of its parents meet one of the conditions: (1) the
parent is optional (2) the parent has been deleted from the graph, or (3) the
parent has been added to the sequence.
2. When an event e is added to the sequence, all other events that are mutually
exclusive to e are removed from the plot graph.
3. Direct parents of removed events become direct parents of direct children of
removed events.
4. If all parents of an event have been removed from the plot graph, that event is
also removed and cannot be added to the sequence.
5. We stop adding events to the sequence when no events may be added, or when
we reach one ending event in the plot graph.
Let S(G) denote all legal event sequences that the graph G can generate based
on the above rules. Each sequence S S(G) is a sequence of events he
1
, e
2
, . . . , e
k
i
where e
1
, e
2
, . . . , e
k
E. When we have a set of events F E, and the sequence S
contains all events in F , we write F S.
As explained earlier, in this chapter, I focus on story understanding as the infer-
rence of probability of events that are not mentioned in a story, given the knowledge
that some other events in the situation have happened. I now define the Story Un-
derstanding Problem (SUP) formally:
Definition 12 (The Story Understanding Problem). Given a plot graph G = hE,
T, M
x
, E
o
, E
c
i, one set of events E
required
E, and a query event e
query
, the Story
139
Understanding Problem determines the probability of e
query
occurring in an event se-
quence s randomly drawn from S
G
that includes all events in E
required
.
The SUP is equivalent to finding the conditional probability P (e
query
S|E
required
S, S S(G)). For simplicity, we write it as P (e
query
|E
required
, G).
Events in a plot graph can have complex interactions, and it is not straightforward
to compute this probability directly from the graph (it will become clearer when we
consider some of these interactions in Section 5.3). One method for solving this
problem is to generate all possible legal event sequences, and count the number of
sequences containing E
required
and e
query
. Let S
r
(G) denote the set of legal event
sequences that contain E
required
, and S
q
(G) denote the set of legal event sequences
that contain E
required
{e
query
}. It is obvious that S
r
(G) and S
q
(G) are subsets of
S(G). Assuming each event sequence is equally likely, we can compute the required
probability as the ratio of the set cardinalities.
P (e
query
|E
required
, G) =
card(S
q
(G))
card(S
r
(G))
(51)
where card(·) is the cardinality function. When there are no legal sequence containing
all events in E
required
, the conditional probability is undefined.
Equation 51 assumes each event sequence is equally likely in the real world, which
admittedly may not always hold. Suppose we can obtain preferences over the events
sequences as a weight function w(S) 0 for every event sequence S S(G), such
that P (S) w(S), the probability may be computed as
P (e
query
|E
required
, G) =
P
SS
q
(G)
w(S)
P
S
0
S
r
(G)
w(S
0
)
(52)
The weight function can take any form provided that w(s) 0, s S(G). Learning
probabilities of event sequences directly from crowdsourced corpora may be difficult
as the crowdsourced stories are sjuzhets rather than fabulas. This problem is out of
the scope of the current dissertation.
140
As the next section shows, the SUP problem is NP-Hard, so it is unlikely that we
will find a polynomial time algorithm for the SUP problem. Generating all possible
sequences in S
G
is expensive, as the number of all possible sequences is exponential
to the number of all vertices in a graph in the worst case. However, it is still possible
to improve the performance of story understanding over the brute force method by
utilizing structures of the plot graphs. Such methods are developed in Section 5.3.
5.2 NP-Hardness of the Story Understanding Problem
In order to prove SUP is NP-hard, we first consider a simplified version of SUP.
Definition 13 (The Simplified Story Understanding Problem). Given a plot graph
G containing n events and one set of events E
required
E, the Simplified Story Un-
derstanding Problem (SSUP) determines if there a legal event sequence that contains
all events in E
required
.
Instead of finding a conditional probability (i.e. a ratio of two probabilities), the
SSUP finds a true/false answer. When E
required
= , the SSUP is trivially true.
The Simplifed Story Understanding Problem (SSUP) is a weaker form of the
SUP, since if we can determine in SUP that the probability of a legal event sequence
containing some required events is defined and non-zero, we can conclude there must
be at least one such event sequence for SSUP.
More specifically, we can perform a simple reduction from any SSUP to an SUP
as follows: Given a SSUP with E
required
6= , choose any event e E
required
. Let
E
0
req
= E
required
\ e. The new SUP determines the probability of e occurring in legal
event sequences containing all events in E
0
req
. The SSUP has an affirmative answer if
and only if the answer to the constructed SUP is defined and non-zero.
Consequently, if SSUP is NP-complete, it follows that solving the Story Under-
standing Problem is NP-hard. Now we show the NP-completeness of the Simplified
141
Story Understanding Problem. The proof contains two steps: (1) SSUP is in NP and
(2) SSUP is NP-hard.
NP is the set of computational problems that can be verified in polynomial time.
When presented with any event sequence S = he
1
, e
2
, . . . , e
k
i, we can verify if S is
legal and contains all events in E
required
. We need to do the following:
1. Check if k n. If k > n, the sequence s must not be legal. The time complexity
is O(1).
2. Check if e
1
, e
2
, . . . , e
k
E. This time complexity of this step, using a simple
scanning algorithm, is O(kn).
3. Check if the sequence contains all events in E
required
, i.e. E
required
s. Similar
to the previous step, the time complexity of this step is O(kn).
4. We construct another legal event sequence S
0
by adding events in S one at a
time, from the beginning to the end of S. When adding each event, we check
if the event being added is eligible according to the story generation rules. It
is obvious that each event can be checked in O(n) time, and the total time
complexity for this step is O(n
2
).
As k n, the totally time complexity of checking a solution is O(n
2
). Therefore, the
Simplified Story Understanding Problem is in NP.
We now show the 3-Satisfiability problem (3-SAT) can be reduced to the Simplified
Story Understanding Problem in polynomial time. As 3-SAT is NP-complete, if any
3-SAT problem can be reduced to a SSUP in polynomial time, then SSUP must be
at least as difficult as 3-SAT. Thus, SSUP is NP-hard. Since we have shown SSUP is
in NP, it must also be NP-complete.
Definition 14 (3-Satisfiability). Consider a boolean formula which can be written as
a number of conjunctive clauses, each containing exactly three boolean variables, such
142
x
1
x
2
x
3
x
4
x
5
x
6
¬x
1
¬x
2
¬x
3
¬x
4
¬x
5
¬x
6
c
1
c
2
c
3
Figure 17: Reducing a 3-Satisfiability problem f = (x
1
x
2
x
3
) (x
2
¬x
3
x
4
)
(¬x
4
x
5
¬x
6
) to a plot graph.
as (x
1
x
2
x
3
) (x
2
¬x
3
x
4
) (¬x
4
x
5
¬x
6
). Is there an boolean assignment
that makes the formula true?
Given a 3-SAT boolean formula, we can construct the following plot graph: For
each boolean variable in the formula, we create two vertices on the graph corre-
sponding to the variable and its negate. Take the example of the formula f =
(x
1
x
2
x
3
) (x
2
¬x
3
x
4
) (¬x
4
x
5
¬x
6
), we would create 12 vertices
for x
1
, ¬x
1
, x
2
, ¬x
2
, x
3
, ¬x
3
, x
4
, ¬x
4
, x
5
, ¬x
5
, x
6
, and ¬x
6
. For a formula involving n
variables, we will have 2n vertices. We create a mutual exclusion relation between
each variable vertex and its negation, e.g. between the vertices x
1
and ¬x
1
. Sub-
sequently, we create one vertex for one clause in the formula, and add a precedence
relation from each variable vertex to every clause that variable appears in. In this
example, we will create three clause vertices c
1
, c
2
and c
3
. Figure 17 shows the plot
graph we constructed. For a boolean formula with n variables and m conjunctive
clauses, we have 2n variable vertices, n mutual exclusion relations between the vari-
able vertices, m clause vertices, and 3m precedence relations, so the construction
has time complexity O(n + m). The constructed SSUP problem asks if a legal event
sequence can contain all the clause variables. In our example, E
required
= {c
1
, c
2
, c
3
}.
We now establish correspondence between the 3-SAT problem and the constructed
143
Simplified Story Understanding Problem: There is a valid variable assignment satis-
fying the boolean formula if and only if there is a legal event sequence in the plot
graph that includes all the clause vertices. That is,
The 3-SAT has a Yes answer The constructed SSUP has a Yes answer
First, we study the forward direction: for each valid variable assignment satisfying
f, we can find a legal event sequence including all the clause vertices. For f to be
true, each conjunctive clause must be true, and at least one of the three variables
in each clause must be true. For each variable x
i
that is true, we add the vertex
corresponding to x
i
to the event sequence. For each variable x
i
that is false, we
add the corresponding negated vertex ¬x
i
to the event sequence. Since the variable
vertices do not have parents, and we never add both x
i
and ¬x
i
, the additions are
always possible. After adding these variables, vertices corresponding to variables
taking false values will be removed from the plot graph. For each clause vertex, at
least one parent vertex will have been added to the sequence, so that clause vertex
will be kept in the plot graph, rather than deleted recursively. Thus, we can add all
clause vertices to the event sequence, and produce a legal event sequence satisfying
the SSUP.
Next, we study the reverse direction: for a legal event sequence including all
the clause vertices, we can find a valid variable assignment satisfying f. Since the
event sequence includes all clause vertices, it must also include at least one of its
three parents. Setting that variable to true will make that conjunctive clause to true.
Thus, we can find a variable assignment that make each conjunctive clause true, and
thereby satisfying the entire boolean formula.
Having established the equisatisfiability of 3-SAT and SSUP on the newly con-
structed plot graph, we conclude the SSUP is NP-complete and SUP is NP-hard.
144
5.3 Simplifying Plot Graphs
One way to solve the Story Understanding Problem is to generate all possible stories,
or event sequences, allowed by the plot graph G, and count how many sequences
contain all events in the set E
required
(i.e. the cardinality of S
r
(G)), and how many
sequences contain all events in E
required
as well as the event e
query
(i.e. the cardinality
of S
r
(G)). The ratio of the two counts is the required probability P (e
query
|E
required
, G).
However, generating all stories in a plot graph is a very expensive operation. If the
plot graph is large, sequences containing the set E
required
may be only a small fraction
of all possible stories, so much of the computation is wasted. We would like to cut
the number of stories we generate and reduce the overall computational cost. That
is, we want to create a new plot graph G
simplified
from the original plot graph G
original
such that
P (e
query
|E
required
, G
simplified
) = P (e
query
|E
required
, G
original
), (53)
and
card(S(G
simplified
)) < card(S(G
original
)). (54)
Assuming P 6= NP, we will not be able to find a polynomial time algorithm for
the Story Understanding Problem. However, we can still simplify the plot graph
and reduce the amount of computation. If an event a E
required
is involved in
a mutual exclusion relation with another event b, in many cases we can remove b
from the plot graph as b will never happen in event sequences we care about. One
such example is shown in Figure 18. In this particular case, E
required
= {x}, so
vertex b, being mutually exclusive to x, is removed from the plot graph. Children
of b, including vertices d, e, and h, are also removed due to transitive closure. The
simplified graph avoids the generation of event sequences abdeh and abedh, thereby
saving computation.
When we remove unreachable vertices from the original plot graph G
original
to
145
a
b
x
d
e
f
g
h
i
(a) Original
a
x
f
g
i
(b) Simplified
Figure 18: An example of plot graph simplification due to mutual exclusion relations.
E
required
= {x}, as shown in red. We know vertex x must exist in the event sequence,
so we can remove vertex b. Children of b are also removed due to transitive closure.
create the simplified plot graph G
simplified
, we must make sure (1) G
simplified
can
generate every event sequence that contain E
required
from G
original
, or more formally,
S S(G
original
), E
required
S S S(G
simplified
) (55)
and (2) every event sequence that G
simplified
can generate is legal according to G
original
,
i.e.
S(G
simplified
) S(G
original
) (56)
These two conditions guarantee Equation 53 is satisfied. We do not require every
event sequence in S(G
simplified
) to contain E
required
.
Several considerations exist when we perform static analysis of plot graphs based
on mutual exclusion relations. The first is what events can be removed given the
execution of some other events. The second is how we regularize the event sequences
in the graph where some events have been removed, so we do not create illegal event
sequences. In both issues, we also need to consider race conditions that cannot be
determined in static analysis. Some of these computations may take exponential time
in the worse case, but their running times for typical plot graphs are usually short. In
146
addition, the results of these computation can be stored together with the learned plot
graphs, so story understanding can be performed quickly when called upon. Details
of these computation will be discussed in the three subsections below.
5.3.1 Cause-for-Removals
The first task we need to do is to find out what events in the plot graph are capable
of deleting other events, taking transitive closure into consideration. We define the
notion of a Cause-for-Removal.
Definition 15 (Cause-for-Removal). Given a plot graph G = hE, T, M
x
, Oi, a Cause-
for-Removal (CfR) for a vertex e E is a minimal set of vertices E
c
E such that
when all vertices in E
c
are executed, e will be removed from the plot graph. As E
c
is
minimal, there is no vertex d E such that the execution of events E
c
\ d, i.e. all
events in E
c
except d, will remove d from the plot graph.
This definition is straightforward. For example, if two vertices u and v are mu-
tually exclusive to each other, u is a CfR for v and v is a CfR for u. If we denote
the set of vertices mutually exclusive to u as M
x
(u), the condition is equivalent to
v M
x
(u). Below are the scenarios that vertex u is a CfR for vertex v:
1. when u and v are directly involved in a mutual exclusion relation. That is,
v M
x
(u)
2. when u is mutually exclusive to all parents of v. If we denote the set of parent
vertices of vertex v as P a(v), this condition can be written as P a(v) M
x
(u).
These two conditions seem obvious. Figure 19 shows some examples. In Figure 19(a),
vertex a is a CfR for both vertices b and c, and vertex d is a CfR for vertex c. In
Figure 19(b), we note vertex e is not a Cause-for-Removal of vertex f, because e
cannot remove both parents of f at the same time. In fact, the existence of e creates
a race condition. When e is executed before d, vertex b is deleted and a precedence
147
from a to f is created. Due to this new precedence relation, the execution of d cannot
remove all parents of f. I will discuss more about race conditions in Section 5.3.2.1.
However, a small modification, shown in Figure 19(c), produces a completely
different result. Neither vertex e or d can delete all parents of vertex f, but we made
two key changes: Vertex b does not have any parents, so deleting b will not create a
new precedence relation. Vertex e is ordered before d, so that vertex c will be removed
after vertex b. If c is removed before b, it will create a new precedence from a to f,
which prevents f from being removed. Note that if e is not ordered before d, a race
condition between e and d can occur.
Intuitively, in order to recursively remove a vertex v from the plot graph, we must
remove all of its parents and prevent any new precedence relations from its grandpar-
ents. When vertex v has grandparents, we must cut any ties with its grandparents
by executing a single event. Otherwise, new precedence relations will be created and
the removal of v is prevented.
We first consider CfRs containing only a single vertex. Based on the above in-
tuition, we can put these vertices into two categories. Category-A (Cat-A) vertices
remove a vertex and all of its predecessors. Category-B (Cat-B) vertices remove a
vertex but not all of its predecessors. In Figure 19(d), vertex d is a CfR for vertex
a. Since a does not have any predecessors, d is a Category-A vertex for a. Vertex
d removes f and its parent b, so d is a Cat-A vertex for f and b. Vertex h removes
g but not its parent c, so vertex h is a Cat-B vertex for g. The reader may have
noticed that the definition of Cat-A and Cat-B vertices depends on the vertices being
removed.
The division between Cat-A and Cat-B vertices is the most important insight
in detecting Cause-for-Removals. It is critical for combining multiple vertices into
a single CfR, as discussed below, and for detecting race conditions, as discussed in
Section 5.3.1.1.
148
a
b
c
d
.
.
.
(a) The CfR of b is a and the
CfRs of c are {a} and {d}
a
b
c
d
e
f
.
.
.
.
.
.
(b) The CfR of f is {d}
a
b
c
d
e
f
.
.
.
(c) The CfR of f is {d, e}.
a
b
c
d
e
f
g
h
i
.
.
.
(d) The CfR of i is {d, e, h}.
a
b
c
d
e
f
g
h
.
.
.
.
.
.
(e) The CfRs of h are {d, e}
and {g}.
Figure 19: Examples of causes for removal
149
When we combine single-vertex CfRs to create a multi-vertex CfR by removing
all parents of a vertex, we need to note that in a multi-vertex CfR, there can be many
Cat-A vertices but only one Cat-B vertex, and this Cat-B vertex must execute after
all Cat-A vertices. Since Cat-A vertices remove a vertex and all of its predecessors,
we do not worry about cutting anything from its predecessors. On the other hand,
Cat-B vertices does cut a vertex off from its predecessors. So we must finish this
cutting with one single vertex. To form a valid Cause-for-Removal, the Cat-B vertex
must be preceded by all Cat-A vertices. If some Cat-A vertices are preceded by the
Cat-B vertex, the vertices cannot form a CfR. If some Cat-A vertices are parallel to
the Cat-B vertex, we have a race condition. I will discuss race conditions in Section
5.3.2.1.
We can now examine Figure 19(d) and 19(e). In Figure 19(d), when considering
CfRs for vertex i, we see it does not have direct mutually exclusive vertices. Hence,
the only way to remove i is to remove all of its parents. Parent vertices a and f can be
removed by Cat-A vertices, and parent vertex g can only be removed by h, a Cat-B
vertex. Since h is preceded by both d and e, we can take the union set {d, e, h}m
which forms a valid CfR for vertex i. In Figure 19(e), vertex f can be removed by
vertex g, which is a Cat-B vertex. In addition, its parent b can be removed by d, a
Cat-A vertex. Another parent c can be removed by vertex e, a Cat-B vertex. As d
precedes e, the union {d, e} is another valid CfR for f. Vertex f is the only parent
of vertex h, so CfRs of f are also CfRs of h.
5.3.1.1 Race Conditions
Race conditions happen during the simplification of plot graphs because although
we know which events are required in the event sequences, we do not know their
relative order. This lack of information may lead to some indeterminacy. A race
condition in the detection of Cause-for-Removals can happen between a group of
150
Cat-A vertices and a Cat-B vertex, as well as between two Cat-B vertices. There are
no race conditions between Cat-A vertices because any two Cat-A vertices can work
together in any order to remove other vertices.
We first examine the competition between Cat-A and Cat-B vertices. With some
modification to Figure 19(d), we produce an example shown as Figure 20(a). Figure
20(a) removes the precedence relation from vertex e to vertex h in Figure 19(d). Thus,
the order of execution between e and h is left unspecified. If we execute the vertices
in the order hdehi or hedhi, the vertex i will be removed as all of its parents have been
removed. However, if we executes the events in the order hdhei, a precedence relation
will be created between c and i immediately after executing h. The execution of vertex
e will not remove vertex i from the graph. In summary, the ordering of execution can
affect whether a vertex is removed or not. Figure 20(c) contains no race conditions.
Since h always executes before e, vertex i will never be removed. There is neither a
CfR nor a race condition.
Thus, a race condition can happen for a set of vertices U if the following conditions
are satisifed:
U contains some Cat-A vertices and one Cat-B vertex
The order of execution is unspecified between some Cat-A vertices and the
Cat-B vertex.
If the Cat-B vertex is executed after all Cat-B vertices, some other vertices in
the graph will be removed. That is, when properly ordered, U is a CfR for some
vertices.
Race conditions can happen between Cat-B vertices. We have seen one such
example in Figure 19(b), which is reproduced as Figure 21(a). Vertex d is a CfR for
both vertices b and c, so it is CfR for their common child, f. Vertex e can only remove
the vertex b but not c, so it is not a CfR for f. However, if e is executed before d,
151
a
b
c
d
e
f
g
h
i
.
.
.
.
.
.
(a) A race condition involving d, e and h.
b
c
d
e
f
h
i
.
.
.
.
.
.
(b) When h happens before e.
a
b
c
d
h
e
f
g
i
.
.
.
(c) No race conditions nor CfRs.
Figure 20: A race condition during the detection of Cause-for-Removals, resulting
in the indeterminacy whether a vertex can be removed. This race condition hap-
pens between Cat-A vertices and Cat-B vertices.Vertices that have been executed are
shaded.
152
a
b
c
d
e
f
.
.
.
.
.
.
(a) Original
a
c
d
e
f
.
.
.
.
.
.
(b) e executes first. f will be kept
when d subsequently executes
Figure 21: Another race condition during the detection of Cause-for-Removals, re-
sulting in the indeterminacy whether a vertex can be removed. This race condition
happens between two Cat-B vertices. Vertices that have been executed are shaded.
a new precedence relation will be created from a to f, preventing f to be removed
later. Note vertex d is a Cat-B vertex for vertices b and c, and vertex e is a Cat-B
vertex for vertex b.
Stating the above intuition formally, given a vertex v, and two Cat-B vertices a
and b that remove parents of v, the two vertices a and b are in a race condition if
Neither a or b are in direct mutual exclusion with vertex v.
One of a and b removes some parents of v but not all parents of v. Without
loss of generality, we assume a satisfies this condition.
The other vertex, b, removes v by removing all parents of v.
The order of a and b are unspecified. Suppose a b, and a cannot be skipped
in the graph (i.e. a 6∈ E
o
E
c
), b is not a valid Cause-for-Removal.
a and b are not involved in the same mutual exclusion relation.
When a race condition exists in a plot graph, the correct simplification will depend
on the order of occurrence of competing vertices. The order of competing vertices
can determine whether the children of the removed vertex can still happen, and when
153
they can happen. Instead of creating two different plot graphs, I do not remove the
vertex being contended in the race condition.
The algorithm for detecting CfR is sketched in Algorithm 7. The algorithm pro-
cesses vertices in the order of a topological sort of the graph (function TopoSort),
which guarantees we process parents before children. Clearly, CfRs of one vertex may
affect the CfR of its children, but not the other way around. The algorithm first
handles vertices without parents. These nodes cannot be removed by removing all
of their parents, and every vertex mutually exclusive to them is necessarily a Cat-A
vertex. If a vertex has some parents, we first try to find Cause-for-Removals that
remove all of its parents. After that, vertices that are directly mutually exclusive are
recorded as Cat-B CfRs.
The MergeParentsCfR attempts to find all CfRs that can remove the parents
of a given vertex v. It starts with CfRs of one parent, and checks if this CfR is com-
patible with other CfRs that removes other parents. If the two CfRs are compatible,
they are merged. If they form a race condition, the race condition is also noted. This
function may potentially perform a combination of all CfRs of all parents, and the
number of combinations is exponential to the number of parents in the worst case.
However, most CfRs are not compatible and most vertices in realistic plot graphs
do not have more than 4 or 5 parents. Thus, this algorithm terminates quickly in
practice.
The conditions for two CfRs to be compatible are stated as follows: for two CfRs
c
1
and c
2
that removes some parents of vertex v, they are compatible if all of the
following are true.
Either c
1
or c
2
contains no Cat-B vertices, or they contain the same Cat-B
vertex.
If a Cat-B vertex exists in either c
1
or c
2
, it must be preceded by all other
vertices in both c
1
and c
2
.
154
No two vertices in c
1
and c
2
are involved in the same mutual exclusion relation.
The conditions for two CfRs to be in a race condition are stated as follows: for
two CfRs c
1
and c
2
that removes some parents of vertex v, they are in a race condition
if all of the following are true.
Either c
1
or c
2
contains a Cat-B vertex, or they contain the same Cat-B vertex.
The Cat-B vertex is parallel to some other vertices in c
1
or c
2
, but never precedes
any vertices in c
1
or c
2
.
When vertices c
1
and c
2
are executed in the correct order, they can remove v.
No two vertices in c
1
and c
2
are involved in the same mutual exclusion relation.
Another possible speed-up comes from the simplification of Cause-for-Removals
by casting them to boolean formulae. Figure 22 shows two examples where the CfR
for vertex g can be simplified. In Figure 22(a), the vertex f has two CfRs: {c} and
{d}. The vertex e has one CfR: {d}. When we combine the CfRs of the parents
of vertex g, we obtain {d} and {c, d}. It is worth noting that a list of CfRs can
understood as a boolean formula, where the vertices in each CfR are conjunctive
because all must be executed to remove a vertex, and different CfRs are disjunctive
because any, if executed, may remove the vertex. With slight abuse of notation, we
can write the above list of CfR as (cd)(d). We can easily see that (cd)(d) = d.
This simplification process reduces the number of CfRs that need to participate in
the combination and accelerates computation. The simplification algorithm is shown
in Algorithm 8.
5.3.2 Implied Temporal Relations
The removal of a vertex from a plot graph removes a precondition for all its children
that remain in the plot graph. Such removals may lead to the generation of event
155
Algorithm 7 Detecting Causes for Deletion
function DetectCfR(G = hE, P, M
x
, E
o
, E
c
i)
order TopoSort(G)
create a dictionary D that maps each vertex to a set of CauseForRemovals
for each v order do Processing in topological order
if ParentsOf(v) = then
if ¬v E
o
then
for each m = (u, v) M
x
do skip optional events with no parents
create a new CfR c, c.CatA = {u}
add c to D(v)
end for
end if
else
pd MergeParentsCfR(G, D, v) A set of CfRs
add pd to D(v)
for each m = (u, v) M
x
do
if u has not been used in any CfR in D(v) then
create a new CfR c, c.CatB = u
add c to D(v)
end if
end for
end if
end for
return D
end function
function MergeParentsCfR(G, D, v)
parents ParentsOf(G, v)
active D(parents.head)
while active 6= do
parents parents.tail
next D(parents.head)
for each c active, each d next do
if Compatible(c, d) then
e Merge(c, d)
next-active next-active {e}
else if RaceCondition(c, d) then
create a new race condition rc, rc.focus v, rc.foes = {c, d}
add rc to the global list of race conditions
end if
end for
active next-active
end while
return active
end function
156
a
b
c
e
f
d
g
.
.
.
(a) The CfRs of g are {d} and
{c, d}, which is equivalent to {d}
Figure 22: Simplifying Cause-for-Removals
Algorithm 8 Simplifying CfR Boolean Formulas
function Simplify(CL = {{v
1
1
, v
1
2
, . . . , v
1
k
}, {v
2
1
, v
2
2
, . . . , v
2
k
} . . .})
for each clause c
i
= {v
1
1
, v
1
2
, . . . , v
1
k
} CL do
Remove duplicate elements in clause
if clause c
j
CL, i 6= j, c
j
c
i
then
CL CL \ {c
i
}
end if
end for
Remove duplicate clauses in CL
return CL
end function
sequences that are not legal in the original plot graph. To solve this problem, we
need to explicitly represent precedence relations that are implicit in the original plot
graph.
Let us consider the situation shown in Figure 23(a). Suppose E
required
= {x}, so
we can remove vertex b from the plot graph. The child of b, vertex e, will not be
removed from the plot graph as one of its parents, vertex c, still remains in the plot
graph.
If we simply remove b from the plot graph, we obtain an incorrect simplification
shown in Figure 23(c). In the original graph (Figure 23(a)), vertex e can only happen
after either (1) that both parents b and c happen or (2) that c and x happen, since
x removes b as a precondition for e. In the incorrect simplification in Figure 23(c),
vertex e can happen before vertex x.
157
a
b
c x
e
.
.
.
(a) Original
a
cx
e
.
.
.
(b) Simplified
a
cx
e
.
.
.
(c) Incorrect
Figure 23: An example plot graph where a temporal relation must be added to ensure
the correctness of the simplified graph. E
required
= {x}, shown in red. When vertex
x happens, vertex b will be removed, so vertex e will have one less precondition. To
ensure that e only happens after x, we must add an explicit temporal relation from
vertex x to vertex e. The naive simplification in (c) allows vertex e to happen before
vertex x.
The problem is that the original graph contains an implied precedence relation
from x to e because x removes one precondition for e. In general, an implied temporal
relation exists between a vertex e, and a group of vertices U = {u
1
, . . . , u
k
} when all
of the following conditions are satisfied:
U is a Cause-for-Removal of at least one parent p of e, and p is not optional
or conditional. Conditional or optional events are not preconditions for any
vertices and can be removed without affecting its children.
U is not a Cause-for-Removal of e. That is, ¬CfR(U, e).
All vertices in U are unordered w.r.t. e, i.e. u
i
U, u
i
k e
To represent this implied precedence relation, when U only contains one single
vertex, we can explicitly add a precedence relation to the graph. This newly added
directed edge is shown as the dashed arrow in Figure 23(b). This added edge would
prevent the generation algorithm from generating an illegal sequence. It is impor-
tant to note a difference between this newly added precedence relation and explicit
precedence relations existing in the original plot graph: the newly added precedence
158
relations cannot be used to determine transitive closure of deletion. If for some rea-
sons we need to delete vertex c, we must also delete e due to transitive closure; we
cannot treat x as a parent of e. To differentiate this newly added relation and the
traditional precedence relations that can be used to determine transitive closure of
deletion, we call this new type of relations temporal relations.
However, when U contains more than one vertex, we do not have a general method
to represent this implied precedence relation on the graph. Hence, we perform a
filtering step after generating all event sequences. This filtering removes all sequences
where e executes before any vertices in U.
5.3.2.1 Race Conditions
Race conditions can also happen in the detection of implied temporal relations. Figure
24 shows an example where we cannot decide if and where we need to add a temporal
relation. When either vertex c or vertex d executes, vertex b will be removed. Between
vertex c or vertex d, it is the earlier vertex that removes a precondition for vertex e.
As we discussed in Section 5.3.2, we must add a temporal relation to the graph, but
we do not know which vertex, c or d, is the earlier vertex. If vertex c executes before
vertex d, the new temporal relation should go from c to e, as shown in Figure 24(b).
If vertex d executes before vertex c, the new temporal relation must go from d to e,
yielding the simplified graph in Figure 24(c). If we do not add the temporal relation,
we may generate the sequence haexyi or haeyxi, which are not allowed in the original
graph.
This indeterminacy of temporal relation ceases to exist when either c or d is
ordered before e and is not optional. Without loss of generality, Figure 24(d) shows
the case where vertex d is ordered before vertex e. In this case, c is free to happen
before or after d. This is because when b does not happen, d is guaranteed to happen
before e. Illegal sequences haecdi and haedci are not possible. The indeterminacy also
159
a
b
c
d
e
.
.
.
.
.
.
(a) Original
a
c
d
e
.
.
.
.
.
.
(b) c executes first
a
c
d
e
.
.
.
.
.
.
(c) d executes first
a
b
c
d
e
.
.
.
.
.
.
(d) No indeterminacy
a
b
c
d
e
.
.
.
.
.
.
(e) A race condition caused by
optionality.
Figure 24: An example of race condition during mutual exclusion analysis concerning
the addition of temporal relations. Depending on the order of occurrence between
vertex c and vertex d and if they are optional, we may have to add temporal relations
in the simplified plot graphs. Vertices that have executed are shaded.
160
does not exists when either c or d is ordered after e. If e precedes c, for example, we will
know that we must add a temporal relation from d to e, resolving the indeterminacy.
If both vertices c and d are ordered w.r.t. e, it is evident that no temporal relations
are needed and no race conditions exist.
However, the race condition comes back when the vertex (either c or d) preceding
e is made optional, as shown in Figure 24(e). Vertex d is ordered before vertex e
but is also optional. Since d is optional, it is not required to execute. We again
have two correct but incompatible cases: When c executes before d, it will remove b,
and we do not have to add any temporal relations. This may generate the sequence
hadeci. When we decide to skip d, b is kept in the graph and is removed by c, so a
temporal relation from c to e is needed. This may generate the sequence hacei. The
temporal relation is needed here to avoid the illegal sequence haeci. Therefore, the
race condition reappears.
Stating the above intuition, the conditions for having a race condition during the
detection of implied temporal relations are:
Two CfRs, c
1
and c
2
can remove one parent p of vertex v, but neither c
1
and c
2
can remove v.
The order of execution between c
1
and c
2
are not specified. That is, a c
1
, 6∃
b c
2
, a b and b c
1
, 6∃ a c
2
, a b.
The order of execution between c
1
and v are unspecified. That is, a c
1
, ¬a
v ¬v a. The orders of execution between all vertices in c
2
and v are also
unspecified. Without loss of generality, If the order between c
1
and v is specified,
then all vertices in c
1
must be optional or conditional.
No vertices in c
1
and c
2
are involved in a mutual exclusion relation.
161
a
c
b d
e
.
.
.
(a) Original
a
b d
e
.
.
.
(b) when e is in
E
required
Figure 25: An example of of implied co-occurrence in the static analysis of mutual
relations. In the original graph (a), vertex e can only execute when b executes. In
the simplified graph (b), we can execute e without b.
5.3.3 Implied Preconditions
A third concern in the static analysis of mutual exclusion relations is implied pre-
conditions that must also be made explicit after simplification. One example of this
situation is shown in Figure 25(a). Note in the graph vertex c and e are ordered by a
precedence relation and are mutually exclusive, but they are not identified as optional
or conditional. This is because the existence of vertex b provides a path to vertex e,
which gives e a chance to execute. In this plot graph, vertex e can only execute when
vertex b executes. When we know vertex e must execute, we can remove vertex c from
the plot graph, resulting in Figure 25(b). However, removing vertex c allows vertex e
to execute without vertex b. For example, we can generate the sequence hadei where
d removes b.
The problem is that vertex b is an implied precondition in the original graph,
which is lost during our simplification. However, we do not have a graph construct
that would guarantee a precondition in plot graphs. To solve this issue, I perform an
extra filtering step that discards all sequences where the implied preconditions are
violated.
The conditions for the existence of an implied precondition between vertex a and
162
b are:
Vertex c is ordered before vertex b, and b and c are mutually exclusive.
Vertices c and b are not recognized as optional and conditional because of vertex
a, which provides an alternative path to b.
5.3.4 Algorithm
In this section, I outline the overall algorithm for the static analysis of mutual ex-
clusion relations (shown as Algorithm 9). The function SimplifyGraph takes five
parameters: G, C, R, IT, IP, E
required
and E
query
. G = hE, P, M
x
, E
o
, E
c
is the origi-
nal plot graph. D is a dictionary that maps each vertex in E to a set of Cause-for-
Removals. R is a list of race conditions. Each race condition contains two sets of
vertices (R.set1 and R.set2) that compete for one vertex (R.focus). IT is a set of
implied temporal relations. Each temporal relation contains a set of vertices E
prior
and another vertex E
later
, where vertices in E
prior
is constrained to execute before
E
later
. IP is a set of implied preconditions.
The algorithm contains many bookkeeping steps and requires some explanation.
The second line finds E
deferred
. When a CfR contains a Cat-B vertex and one or
more Cat-A vertices, the Cat-B vertex must execute later. Therefore, if E
required
does
not contain all Cat-A vertices in the CfR, we must not remove vertices that can be
removed by the Cat-B vertex alone. E
deferred
contains these Cat-B vertices.
Line 4 finds all vertices that may be removed from the plot graph, captured by
E
possible
. Line 5 finds E
remaining
. If one vertex is being contended in a race condition,
this vertex must not be removed from the plot graph. E
remaining
contains these
vertices. Therefore, E
possible
\ E
remaining
are the vertices to be removed from the plot
graph. Line 8, 9, and 10 handle the insertion of implied temporal relations.
The Count function generates all possible vertex sequences from the simplified
graph G
simplified
. Note that a valid sequence must begin with a valid source vertex
163
Algorithm 9 The Algorithm for Mutual Exclusion Analysis
function SimplifyGraph(G, D, R, IT, IP, E
required
, E
query
)
all-cfr all CfRs in D
E
deferred
{cfr.CatB all-cfr | cfr.CatA 6⊆ E
required
}
E
active
E
required
\ E
deferred
E
possible
vertices that can be removed by E
active
E
remaining
{r.focus | r R (r.set1 E
required
r.set2 E
required
)}
E
removed
E
possible
\ E
remaining
G
simplified
RemoveEvents(G, E
removed
)
T {t IT | t.E
prior
E
required
}
T
added
{(t.Eprior(0), t.E
later
) | t T, |t.E
prior
| = 1}
G
simplified
.P G
simplified
.P T
added
E
sources
source nodes in G
E
ends
end nodes in G
return Count(G
simplified
, E
sources
, E
ends
, IT, IP, E
required
, E
query
)
end function
function Count(G
simplified
, E
sources
, E
ends
, IT, IP, E
required
, E
query
)
seqs all possible vertex sequences starting from E
sources
and ends in E
ends
or
ends when no more vertices may be executed
valid-seqs vertex sequences where IT and IP constraints are respected
s
1
number of sequences in valid-seqs containing E
required
s
2
number of sequences in valid-seqs containing E
required
and E
query
return s
1
/s
2
end function
in the original graph G, and it must end with a valid end vertex in the original
graph G, or when no more vertices may be executed. The generation algorithm has
been shown as Algorithm 5. The generated sequences are further filtered with the
implied temporal relations and implied precondition constraints. Finally, we count
the sequences and compute the ratio.
5.4 Evaluation
I evaluate the effectiveness of the algorithm for plot graph simplification by test-
ing them against random plot graphs with random precedence relations and mutual
exclusion relations.
164
5.4.1 Methodology
The algorithm for generating random plot graphs takes three parameters: the total
number of vertices, total number of precedence relations, and total number of mutual
exclusion relations. As the plot graph is a directed acyclic graph, the vertices are first
numbered from 1 to n. Precedence relations are uniformly drawn from all possible
precedence relations, which go from a vertex with a smaller number to a vertex with
a greater number. After the precedences are created, we omit precedences that are
implied by transitive closure (e.g. a b and b c implies a c.) Similarly, to gener-
ate the mutual exclusion relations, I uniformly drawn from all possible vertex pairs.
Finally, optional and conditional vertices are detected. For each set of parameters,
10,000 random plot graphs are generated.
In order to test the effectiveness of the analysis based on mutual exclusion rela-
tions, from each random plot graph, I pick one vertex involved in at least one mutual
exclusion relation as E
required
. Another vertex is uniformly drawn from all vertices as
E
query
.
For each randomly generated plot graph, I record the total number of generated
sequences and the amount of computation time for the original graph and the sim-
plified graph respectively. For the original plot graph, all computation time is spent
on generating event sequences. For the simplified plot graph, the computation time
contains the time spent on graph simplification and generating the event sequences
for the simplified plot graph. The averages over 10,000 graphs are computed as the
sum from all simplified graphs divided by the sum from all original graphs.
5.4.2 Results
The acceleration obtained for random plot graphs generated by different sets of pa-
rameters are shown in Table 18. The sets of parameters were selected so that the
random graphs resemble the actual plot graphs we have learned. As the number of
165
Table 18: Acceleration obtained by simplifying plot graphs based on mutual exclusion
relations
Configuration Savings
Error Rate
Vertices Precedences Mutual Exclusions Sequences Time
10 20 3 29.47% 27.03% 1.56%
10 25 3 30.19% 23.78% 0.63%
10 30 4 30.34% 22.74% 0.66%
12 25 4 31.04% 32.97% 2.86%
12 35 4 31.31% 31.34% 1.05%
14 40 4 31.84% 32.93% 1.75%
14 50 6 32.32% 33.34% 2.09%
16 60 6 32.85% 36.29% 2.61%
16 70 6 32.62% 32.34% 1.52%
16 70 8 33.06% 34.03% 2.54%
vertices grows, the number of precedence relations and mutual exclusion relations
are also increased to maintain a similar level of parallelism and mutually exclusive
alternatives in the plot graphs. The configurations are generally sorted in increasing
order of complexity.
The results indicate that the simplification of plot graphs based on mutual ex-
clusion relations can achieve substantial increase in performance. On average, the
knowledge of one event involved in a mutual exclusion relation can reduce the total
number of sequences generated by more than 30%. The average saving on computa-
tion time grows from a low of 22.74% to a high of 36.78% when the number of vertices
grows from 10 to 16. The error is greater than zero across all tests, indicating the
algorithm does not always simplify plot graphs correctly. However, the error rate
never exceeded 3%.
166
5.4.3 Discussion
One important purpose of the graph simplification algorithm is to effectively perform
story understanding in large, complex plot graphs. As the number of legal vertex
sequences grows superlinearly with the number of vertices, it is important to reduce
the time of computation when the number of vertices becomes large.
Observing Table 18, we note the simplification algorithm scales well. As plot
graphs become more complex, the saving in number of generated vertex sequences
tends to increase, although the speed of increment is slow. The saving on the actual
computation time portrays a more optimistic picture. The saving on time increases
by 7% to 14% when the number of vertices increases from 10 to 16. This indicates
the simplification algorithm only incurs a small computation overhead, and the over-
head is dominated by the growth of vertices. Overall, this indicates the simplification
algorithm can effectively reduce the computation needed to answer the Story Under-
standing problem.
It must be acknowledged that the current simplification algorithm does not sim-
plify all plot graphs correctly. It is likely there are other forms of race conditions or
implied temporal relations that I have not discovered. However, the error rate does
not always increase as the plot graph grows more complex. Moreover, mistakes are
only made on less than 3% of all plot graphs. In other words, given a random plot
graph, we are at least 97% confident that the simplification algorithm will produce a
correct answer. Other approximation algorithms, such as Monte Carlo methods that
sample sequences from plot graphs, are left for future work.
5.5 Summary
In this chapter, I have investigated the Story Understanding Problem and its math-
ematical properties. Noting that the sjuzhet of a story often hides details in the
fabula from the audience, I examined the inference of unmentioned events based on
167
explicitly mentioned events in the story, by utilizing the plot graphs Scheherazade
has learned earlier.
I provided a mathematical definition for the Story Understanding Problem and
further showed that it is NP-hard. However, by utilizing properties of the plot graph
representation, it is possible to reduce the needed computation time by simplifying
the plot graph, if we know some events that exist in the story are also involved in some
mutual exclusion relations. Employing this information, we can remove vertices from
the plot graph, thereby reducing the amount of computation spent on generating all
possible event sequences. I have discussed and solved many issues in performing this
type of static analysis of plot graphs, including finding Cause-for-Removals, race con-
ditions, implied temporal relations and preconditions. Experimental results suggest
the simplification algorithm scales well when the plot graph grows more complex. On
average, simplifying plot graphs can reduce the total number of event sequences by
29.47% to 33.06%, and the total computation time by 27.03% to 36.29%.
In conclusion, the simplification algorithm demonstrates good scalability that al-
lows us to efficiently answer the Story Understand Problem. The ability to perform
story understanding based on automatically learned knowledge holds the promise of
scaling computational Narrative Intelligence to solving diverse and large real-world
problems.
168
CHAPTER VI
FUTURE WORK AND CONCLUSIONS
If I’ve vividly laid out the narrative, the reader will come to his own con-
clusions.
Rick Atkinson
In this concluding chapter, I summarize this dissertation and highlight its contri-
butions. I then discuss two major unsolved problems that lie ahead on the road to
strong computational Narrative Intelligence, and speculate about possible solutions.
6.1 Summary
Despite their differences, most existing computational Narrative Intelligence systems
rely on hand-crafted knowledge, which requires considerable amount of expert labor.
As a result, these systems are confined to a few domains where knowledge has been
provided. This knowledge bottleneck has been widely recognized, but its solution has
not been thoroughly investigated.
In this dissertation, I propose procedures and techniques for learning knowledge
to support narrative intelligence, including story generation, storytelling and story
understanding, in domains previously unknown to the system. The Scheherazade
system is capable of extending its own knowledge when it encounters unfamiliar sto-
rytelling domains. I collected a number of simple exemplar stories in a given socio-
cultural or procedural situation (e.g. a date at a movie theater, pumping gas into a
car, etc.) from workers on Amazon Mechanical Turk. The crowdsourcing approach
allows us to circumvent natural language challenges and directly tap the collective
social norms from the minds of members of a community.
169
A situational model, called a plot graph, can be learned robustly from crowd-
sourced stories. A plot graph describes how a situation typically unfolds in terms
of events, precedence relations, mutual exclusion relations, and optional/conditional
events. The plot graph representation has proven to be versatile in supporting tasks
of narrative intelligence. I have created a complete algorithmic pipeline for generating
the fabula (i.e. all events that happen in the narrative world), the sjuzhet (i.e. the
events actually being told) and the media (i.e. the final story in text). I also devel-
oped algorithms for efficiently inferring if an untold event has happened, given the
events being told to the AI. Evaluations demonstrate that the system can (1) learn
plot graphs accurately (Section 3.4 and 3.9), (2) generate stories that match human-
written stories in some measures of coherence, marking an important milestone for
computational Narrative Intelligence (Section 4.2), (3) tell the stories in distinctive
narration styles (Section 4.5), and (4) reduce the computation time by approximately
30% or more in story understanding (Section 5.4). Consequently, the system can
demonstrate narrative intelligence in any situation where a small number of exemplar
stories may be collected.
6.2 Contributions
In this dissertation, I made the following contributions:
The system addresses the knowledge bottleneck that has troubled Narrative
Intelligence systems for decades. By learning from crowdsourced stories, the
system is able to demonstrate Narrative Intelligence in any situation where a
small number of stories can be collected. This is the first computational system
that employ learned knowledge to perform both story generation and story
understanding.
I provide a novel knowledge representation, called a plot graph. This represen-
tation can be accurately learned using statistical methods and supports many
170
Narrative Intelligence tasks, thereby effectively bridging learning and applica-
tion. This representation incorporates natural language processing into an NI
framework.
In the evaluation of fabula coherence, I demonstrate for the first time that
computer-generated stories can approximate human-written stories in some as-
pects of coherence. This is also the first time that a computational NI system
is directly compared to humans.
The Scheherazade system learns diverse narration styles from large data sets,
which enables applications in storytelling virtual characters.
6.3 Potential Applications
In addition to long-term scientific contributions, the development of the Scheher-
azade system provides immediate benefits in supporting several potential applica-
tions, including virtual training environments and virtual characters.
The Scheherazade system was initially inspired by the need to develop an
intelligent and interactive training application for preparing children with autism for
navigating everyday social situations, as autistic children often experience difficulties
in understanding and following social conventions. During the development, it became
apparent that the manual authoring approach will not scale up to the complexity of
real-world scenarios. Therefore, it is necessary for a computational system to learn the
knowledge by itself. Inspired by Boujarwah et al. [14]’s initial work on crowdsourcing
scripts for the training environment, in this dissertation, I develop an entire procedural
pipeline from collecting exemplar stories to the automatic learning of the knowledge
and its applications.
The Scheherazade system’s capability to learn socio-cultural conventions is not
limited to the the mainstream culture where autistic children need to cope with, but
can also extend to foreign cultures. For example, training scenarios can be developed
171
to prepare users for job interviews in Japan or respecting local customs in Samoa.
It could be expensive to hire experts on foreign cultures to manually author suffi-
cient knowledge in support of virtual training environments. If the culture is not well
studied, finding an expert on the subject could be a difficult task by itself. In com-
parison, Scheherazade only requires a small number of exemplar stories written by
English-speaking non-experts who have some experience with the foreign culture and
that particular social situation. No training in computer science or AI is required.
Moreover, the Scheherazade system enables the creation of diverse virtual char-
acters in games or virtual worlds that can tell stories. For entertainment purposes,
oftentimes we would like to create an illusion that virtual characters lead their own
lives and are not just part of a show that disappears when the player looks away.
Thus, we would like to give each character a background story or an alibi, which can
explain where they have been and what they have done while they are not with the
player [176]. Virtual characters should be able to tell these background stories and
recall details to substantiate their stories when asked to. Therefore, the knowledge of
common social situations, the ability to identify prototypical events, and the ability
to tell stories in diverse narration styles can help the creation of believable virtual
characters.
Virtual characters with background stories also have applications in the health-
care industry. For example, Bickmore et al. [12] found that the ability to tell autobi-
ographical stories increases the likelihood that human users will interact with virtual
characters over an extended period. This can be advantageous when we need to en-
courage users to keep interacting with some programs or electronic devices, such as
educational software or medical devices for self-monitoring.
172
6.4 Future Work
The limitations of each individual components of the system have been discussed
throughout this dissertation. In this section, I will discuss two major future directions:
how to understand and generate creative stories, and how to further increase the
flexibility of the knowledge representation.
6.4.1 Creativity
It can be argued that the stories currently generated by Scheherazade are not
qualitatively creative, although the stories generated by the system can be differ-
ent from any input story. As the plot graphs capture typical behaviors in typical
situations, the stories generated by following the plot graphs are prototypical and
lack surprises. Therefore, an important future direction is to extend the creative
capabilities of Scheherazade.
It is possible to specifically crowdsource exemplar stories that break the norm. In
an earlier, unsuccessful attempt to crowdsource stories for the movie date situation,
for example, one creative story introduced a third character, a friend of John who
insisted in sitting with the couple and disrupted their date. In later attempts, the
crowdsourcing instructions were changed to be explicit that we only wanted stories
about the most typical or mundane stories in a situation. However, the statistical
nature of learning makes it difficult to learn from those creative stories. Among the
stories that are very dissimilar from other stores, some are coherent and interesting,
whereas others are actually nonsensical. From a pure statistical point of view, it is
difficult to separate those two cases.
As a consequence, a major challenge for learning from creative exemplar stories is
to understand creative stories. One possibility is to crowdsource the understanding.
The crowd workers may be asked to vote for stories that are both coherent and
creative, or repair nonsensical stories using a mechanism similar to the story repairing
173
experiment in Section 4.2.
AI systems have attempted to understand creative stories. Ram [146] pointed
out that most, if not all, creative story variations are derived by combining and
modifying known behaviors and scripts. My survey of the manga Doraemon provides
evidence the creative process of most science-fictional gadgets is based on modifying
and blending ordinary objects we encounter in our life [98, 101]. This process is
similar to conceptual blending [189, 15], a cognitive paradigm for human creativity.
Therefore, given enough plot graphs, it is plausible that a computational system can
make sense of a creative exemplar story by combining different plot graphs. For
example, if the system possesses a plot graph about a third wheel situation, it may
understand the aforementioned creative story as a conceptual blend of the third wheel
graph and the movie date graph. Those stories that the system cannot make sense of
will be regarded as nonsensical. The ISSAC system [120] understands creative story
based on a large amount of manually authored knowledge. It is an promising research
direction to build such a system based on purely learned knowledge.
6.4.2 Multiplicity of Plot Graph Levels
Being able to move between different levels of abstraction is a strength of human
cognition and provides benefits to Narrative Intelligence. A high-level representation
of a situation provides a concise and intelligible summary. A low-level representation
contains many details, such as movements of body parts and facial muscles, moment-
to-moment thought processes, and so on. A skillful story writer use different levels
of abstraction to zoom in the important details and fast-forward unimportant events,
creating variations in the tempo of the narrative. These techniques are known as
compression and expansion (See Section 2.1.1). One vivid example of the expansion
technique is the so-called “bullet time”. First showcased in the movie The Matrix,
bullet time shows bullet dodging actions in hyper slow motion in order to achieve
174
special artistic effects [143]. A high-level representation can capture commonalities
between many seemingly different scenarios, such as representing that the gas pump-
ing situation, the pharmacy situation, and the movie date situation all involve the
purchase of merchandise. A high-level understanding can facilitate the technique of
compression.
The current plot graph representation contains basically two levels: the events
and their natural language descriptions. In Section 4.3, I provided the EventRank
algorithm that identifies the most and the least prototypical events in the plot graph.
A higher-level representation, which sits on top of the event level, can thus be par-
tially simulated by extracting most prototypical events. However, the EventRank
algorithm does not identify a hierarchical relationship between the higher level and
the event level, which potentially harms the quality of story summaries it produces.
For example, in the bank robbery situation, it is currently impossible to create a
higher level “demand money” event that encompasses both the event “John demanded
money” and the event “John handed Sally a note”. This high-level event should be
assigned a higher weight than any of those two events and be recognized as a very
prototypical event.
Future work is also required to learn a lower level of representation that describes
body movements. With this level, the system will be able to create detailed descrip-
tions such as “John set his left foot into the bank’s lobby”. Understanding the body
movements associated with each action may also allow us to reason about unintended
consequences, such as John stepping on banana peels and falling.
6.5 Conclusions
Narrative Intelligence is a vital component of human intelligence. Computational
replication of Narrative intelligence has important implications for the development
of human-level AI and numerous commercial applications. Decades of research made
175
it clear that computational NI systems that aim to scale beyond a few simple domains
must be able to learn the needed knowledge by itself. In recent years, we have
witnessed the emergence of several Open Narrative Intelligence systems [177, 166,
114]. The Scheherazade system presented in this dissertation is the first system
that can utilize the learned knowledge in both tasks of story generation and story
understanding, and the first system that can generate stories that approximate the
coherence of human-written stories. Its capabilities have been tested and proved in
several user studies and evaluations. As such, I conclude the thesis statement, first
proposed in Section 1.4 have been achieved.
The Scheherazade system marks an important milestone for scaling computa-
tional Narrative Intelligence to meet the challenges of the real world. Its development
helped us understand important problems and identify possible solutions in the sci-
entific pursuit of an Artificial Intelligence capable of crafting, telling, understanding,
and responding appropriately to narratives.
176
APPENDIX A
SMOOTH SENTIWORDNET
This Appendiex shows the 100 most positive and 100 most negative words in Smooth
SentiWordNet. The negative words are shown on the left column and the positive
words are shown on the right column.
Lemma / POS Sentiment Lemma / POS Sentiment
bone-chilling/JJ -8.00399 silky/JJ 0.90291
shriek/NN -6.73801 kiss/VB 0.90774
scop/VB -6.7339 fulfil/VB 0.90887
wail/VB -6.30141 information/NN 0.91023
panic/VB -5.67925 purchase/VB 0.91234
sob/NN -5.64176 arrange/VB 0.91776
scream/NN -5.31574 cooperate/VB 0.92015
nightmare/NN -4.95367 fully/RB 0.92036
terror/NN -4.26091 wine/NN 0.92326
panic/NN -4.04759 better/JJ 0.93538
weep/VB -3.84039 satisfy/VB 0.9364
fright/NN -3.7388 view/NN 0.93922
sob/VB -3.56282 comfortable/JJ 0.93955
terrify/VB -3.51428 topic/NN 0.93957
anger/NN -3.4937 new/JJ 0.94163
victim/NN -3.28723 friend/NN 0.94739
scream/VB -3.28343 smile/VB 0.94977
screech/VB -3.22727 pretty/JJ 0.95263
sweaty/JJ -3.22524 dame/NN 0.96067
hysterically/RB -3.01847 politics/NN 0.9745
177
flee/VB -2.83545 interact/VB 0.97931
hate/VB -2.80631 sport/NN 0.98264
bark/VB -2.72129 lady/NN 0.98368
collapse/VB -2.70325 present/VB 0.99649
dread/VB -2.63414 important/JJ 1
yell/VB -2.62926 menu/NN 1.00148
restroom/NN -2.61856 stoplight/NN 1.00301
frightened/JJ -2.56544 young/JJ 1.00905
stumble/VB -2.52733 favorite/JJ 1.01332
freeze/VB -2.51679 particularly/RB 1.01674
fear/NN -2.503 yes/NN 1.01886
thirst/NN -2.47606 best/RB 1.02546
thunder/NN -2.42492 best/JJ 1.02959
death/NN -2.23877 good/JJ 1.03473
dismay/NN -2.17959 expensive/JJ 1.03554
clutch/VB -2.09228 interest/NN 1.04276
violently/RB -2.07236 blond/JJ 1.05009
struggle/VB -2.02189 special/JJ 1.0502
scared/JJ -2.00743 comfy/JJ 1.05107
scurry/VB -1.95179 like/VB 1.06617
shocking/JJ -1.93507 restoration/NN 1.0752
darkness/NN -1.89049 bright/JJ 1.09547
sorrowful/JJ -1.83753 favor/NN 1.09927
cause/VB -1.68308 baseball/NN 1.10351
worst/JJ -1.66493 glory/NN 1.10587
alarm/NN -1.61991 appropriate/JJ 1.123
suspenseful/JJ -1.61467 nameplate/NN 1.12736
burst/NN -1.60499 credit/NN 1.13414
rush/VB -1.58249 ecstatic/JJ 1.13983
heat/NN -1.55584 love/NN 1.14522
178
ordeal/NN -1.55112 mirror/VB 1.14923
frighten/VB -1.42775 satisfactory/JJ 1.14993
convict/VB -1.41959 music/NN 1.15872
darken/VB -1.41634 greet/VB 1.16503
twisted/JJ -1.35634 queue/VB 1.16565
wind/NN -1.33197 sweet/JJ 1.17272
nervous/JJ -1.32785 mutual/JJ 1.19994
tremble/VB -1.31654 character/NN 1.2009
rush/NN -1.30267 perfectly/RB 1.21689
sound/NN -1.26086 chat/VB 1.22213
sunglass/NN -1.22531 beverage/NN 1.22738
throat/NN -1.21807 thorough/JJ 1.22821
tear/NN -1.16682 impress/VB 1.23884
mortally/RB -1.14252 delicious/JJ 1.2429
slam/VB -1.13843 remarkable/JJ 1.24377
pound/VB -1.12395 tasty/JJ 1.2681
dash/VB -1.09977 initiate/VB 1.27167
llama/NN -1.07874 coyly/RB 1.29285
darkened/JJ -1.06664 epic/NN 1.3139
pretzel/NN -1.05842 interesting/JJ 1.31427
sad/JJ -1.05651 happy/JJ 1.34045
excruciating/JJ -1.03506 perfect/JJ 1.34832
ice-cold/JJ -1.02378 facility/NN 1.35465
rip/VB -1.01851 rehearsed/JJ 1.35902
whip/VB -1.0169 harmony/NN 1.39367
audible/JJ -1 goer/NN 1.39632
robbery/NN -0.98525 enthral/VB 1.40056
coarse/JJ -0.95885 wonderful/JJ 1.41497
pack/NN -0.95334 applaud/VB 1.46886
prison/NN -0.94258 enjoy/VB 1.47755
179
distract/VB -0.91207 romantic/JJ 1.51518
tense/JJ -0.88605 blonde/JJ 1.51591
fall/VB -0.88226 classic/JJ 1.52007
level/VB -0.86619 presentation/NN 1.52359
escape/VB -0.86512 clunker/NN 1.62548
leg/NN -0.85497 flirtatious/JJ 1.66228
unburied/JJ -0.85305 richly/RB 1.6989
rob/VB -0.83481 delightful/JJ 1.78992
body/NN -0.82224 preview/NN 1.82059
devour/VB -0.79887 enjoyable/JJ 1.82074
loud/JJ -0.79637 deal/NN 1.97416
irremediable/JJ -0.77608 beam/VB 1.9844
ignition/NN -0.77341 admire/VB 2.00495
heart-rending/JJ -0.74582 immensely/RB 2.02764
fetid/JJ -0.74403 sexy/JJ 2.13925
shoot/VB -0.72868 captivate/VB 2.39255
nylon/NN -0.7126 small-framed/JJ 2.99211
mangled/JJ -0.70885 nacho/NN 3.14946
grievously/RB -0.70578 snack/VB 3.2766
dank/JJ -0.68328 tryed/JJ 4.62135
180
APPENDIX B
GENERATED STORY TEXTS
This section shows the stories generated for the purpose of the user study discussed
in Section 4.5.2.
B.1 Stories in the Bank Robbery Situation
B.1.1 The Positive Story
John drove his grandmothers borrowed old clunker to the bank on Main St.
John took a deep breath and opened the bank door, letting an elderly woman exit
before he entered himself.
John did not want to be recognized.
John entered the Yes bank at 1 pm.
John gave a thorough look around the bank to see how many people were inside.
John spotted a young blond teller, Sally, behind the counter.
John stood behind the lady and toddler and politely waited his turn, noticing the
name plate on the counter... Sally’.
John walked carefully up to the counter and interacted with Sally.
Sally started feeling the hairs on the back of her neck stand up.
John pulled out the gun, still smiling.
John pointed his gun at Sally to push the point of the seriousness.
Upon seeing the gun, Sally yelled out in fear.
John wanted more money.
John presented Sally with a bag in which to put the money.
Sally continued to cooperate, putting the money into the bag as ordered.
181
John pocketed the money.
Sally cried, somewhat relieved it may be over soon.
Sally described John as best as she could to the police.
John ran out of the bank and got in his grandmothers car.
John drove away, sure the police were behind him all the way back to his grandmothers
house in the country.
B.1.2 The Negative Story
John drove to the bank, with a nervous look on his face.
John opened the bank door while his heart was beating fast.
John put on sunglasses.
John walked into the bank with a handgun underneath his jacket.
John looked around the bank, scoping out security cameras or guards.
John noticed one of the tellers named Sally seemed bored and distracted.
John stood in line.
John approached Sally naturally as to not raise alarm.
Sally saw Obama standing in front of her and she felt her whole body tense up as her
worst nightmare seemed to be coming true.
John pulled out a gun.
John leveled the gun at Sally and kept it on her.
Sally let out a bone-chilling scream.
John barked his orders at Sally, demanding she put the money in the bag.
John forced the bag into Sallys hands.
Sallys hands were trembling as she put the money in the bag.
John then grabbed the bag of money out of Sallys nervous hands.
Sally felt tears streaming down her face as she let out sorrowful sobs.
Still shaken, Sally reached for the phone and in a panicked manner called the police.
182
John quickly fled the bank and entered into his car.
John escaped in the car.
B.1.3 The Concise Story
John drove to the bank.
John opened the bank door.
John did not want to be recognized.
John went into the bank.
John looked around the bank.
John recognized Sally.
John stood in line.
John walked to Sally.
Sally was scared.
John took out a gun.
John pointed gun at Sally.
Sally screamed.
John demanded 1,000,000 from Sally.
John gave a bag to Sally.
Sally put money into the bag.
John took the money.
Sally cried.
Sally called the police.
John got in his car.
John drove away in the car.
183
B.1.4 The Story with Most Interesting Details
John got into his car with his disguise, gun and note in his knapsack and headed
towards the Old Second in the next town over, repeating his rehearsed demands
silently over and over in his head.
John watched while a little old lady left the bank and walked to her car and then
slipped on his gloves, slipped his gun into his coat pocket, grabbed his mask and
strode determinedly to the lobby door and pulled it open.
John looked at his reflection in the glass of the door, gave himself a little smirk and
covered his face.
John took another deep breath as he wondered if this was really a good idea, and
entered the bank.
John looked around the bank, making sure his timing was right.
John spotted a young blond teller, Sally, behind the counter.
John stood behind the lady and toddler and politely waited his turn, noticing the
nameplate on the counter... “Sally”.
When it was his turn, John, wearing his Obama mask, approached the counter.
Sally saw Obama standing in front of her and she felt her whole body tense up as her
worst nightmare seemed to be coming true.
Once Sally began to run, John pulled out the gun and directed it at the bank guard.
John wore a stern stare as he pointed the gun at Sally.
Sally screamed hysterically which alerted other people in the bank.
John demanded Sally to give her all of the money she had in her drawer, and all the
money that was close that she could get to quickly.
John tossed the bag he had brought for the money at Sally.
Sally put the money in the bag, and collected the money from the 2 tellers next to
her.
John struggled to stuff the money in his satchel.
184
Sally was quietly sobbing as John grabbed the bag full of money.
Sally called the cops.
John strode quickly from the bank and got into his car tossing the money bag on the
seat beside him.
The car drove away from the bank.
B.1.5 The Story with Most Interesting Details and Length Penalty
John calmly drove to the bank to avoid drawing attention to himself.
John opened the door.
John covered his face.
John walked into the front door of the bank.
John looked around, scanning the bank for anything nearby.
John spotted a young blond teller, Sandy, behind the counter.
John stood behind the lady and toddler and politely waited his turn, noticing the
nameplate on the counter... “Sally”.
Quietly and calmly John walked up to Sally’s window.
Sally saw the smile and got scared.
John reached behind his back and withdrew his pistol.
John wore a stern stare as he pointed the gun at Sally.
This caused Sally to let out an audible shriek.
John asked for all the money in the drawer.
John then handed Sally an empty sack.
Sally stuffed the money into John’s bag.
John pocketed the money.
Sally continued to sob hysterically as she sat down on her stool.
Sally called the cops.
John got into his red pickup truck and slammed the door.
185
John sped away, hoping to get distance between him and the cops.
B.2 Stories in the Movie Date Situation
B.2.1 The Positive Story
John drove his red car to Sally’s house to pick her up for their movie date.
Sally, in her sexy yellow dress climbed into the sleek sports car.
John and Sally bought their movie tickets at The Dame”, a local theater that had
been restored to its former 1940s glory.
John and Sally showed their tickets to the attendant who smiled and waited for them
at the entrance to the theater.
Sally bought herself a small soda, while John got nachos and some sour patch kids.
John and Sally entered the theater and admired the big screen.
John and Sally bought delicious drinks.
John and Sally scoured the theater for the best seats in the house, and found a perfect
pair right in the center of the theater.
John and Sally sat down in the velvet covered seats, and took in the richly appointed
theater.
John and Sally chatted during the previews.
After the lengthy and irritating previews, the movie finally began.
Sally watched the opening scenes intently while she snacked on her popcorn.
John and Sally took a drink of their beverages to wash down the popcorn.
During the romantic parts of the movie, John pulled the classic “yawn and put arms
around” move.
John and Sally sat captivated by the movie.
During the movie, the drinks really went through John and he needed to excuse
himself.
John and Sally felt endlessly fulfilled by the movie and enjoyed it immensely.
186
John gave Sally a soft hug during the movie, and Sally returned the favor.
John and Sally felt romantic holding hands throughout the movie.
The movie had a delightful and surprising ending.
John and Sally stood up from their seats during the ending credits.
John and Sally walked through the theater and then through the parking lot to get
to their car.
John and Sally left the theater and enjoyed the fresh night air.
John kissed Sally good night.
B.2.2 The Negative Story
With sweaty palms and heart racing, John drove to Sallys house for their first date.
Sally was glad to get out of the heat and get into the air conditioned car.
John and Sally arrived at the theater just before the movie was scheduled to start
and rushed to buy their tickets for the movie.
John and Sally quickly whipped out their tickets for the ticket checker.
John and Sally decided that they wanted some popcorn.
John and Sally entered the darkened theater and paused letting their eyes adjust to
the dimness.
John realized they had not bought drinks for themselves, which caused them both to
laugh before they went back to the snack bar.
Stumbling through the darkness of the theater, John and Sally eventually found their
seats.
John awkwardly grabbed Sallys coat and then they both sat down.
John made nervous small talk to Sally before the movie started.
The movie began with a loud sound of thunder.
The couple devoured the popcorn they ordered.
Because Johns throat became dry from the popcorn, John drank his soda.
187
John tried to put his arms around Sally, but she dodged him.
John and Sally watched the film.
John and Sally walked to the restrooms.
John and Sally sat with big smiles on their faces as they watched the movie.
After a sad moment in the movie, John reached out and hugged Sally.
Sally held Johns sweaty hands throughout the movie.
Much to Johns dismay, the movie ended.
Sally stood up first, followed by John, and they both stretched quickly to revive their
bodies after a long time of sitting.
John and Sally slowly walked to the car and got inside, dreading the end of the date.
John drove the car out of the movie theater parking lot and turned toward Sallys
house.
At the front door of Sallys house, John leaned in and gave Sally a peck on the lips.
B.2.3 The Concise Story
John drove his car to Sallys house.
Sally got into Johns car.
John got the tickets.
John and Sally gave the tickets to the usher.
Sally paid for the snacks.
Sally and John went into the theater.
John bought drinks.
Sally found seats.
John and Sally sat.
John and Sally talked about the movie.
The movie began.
John shared his popcorn with Sally.
188
Sally took a drink.
John puts his arm around Sally.
John and Sally watched the film.
John went to the bathroom.
John and Sally enjoyed the movie.
John hugged Sally during the movie.
John and Sally held hands during the movie.
The movie ended.
Sally got up.
John and Sally ran to their car.
John and Sally went outside the theater.
Sally kissed John.
B.2.4 The Story with Most Interesting Details
John drove his red car to Sally’s house to pick her up for their movie date.
Sally carefully entered the car with a smile.
John and Sally arrived at the theater just before the movie was scheduled to start
and rushed to buy their tickets for the movie.
The tattooed, male attendant gave John a smile and a wink as he whispered the
words what a catch,” while John showed the tickets at the entrance.
John bought a large tub of popcorn and two boxes of malted milk balls, Sally’s
favorite.
John and Sally entered the darkened theater and paused letting their eyes adjust to
the dimness.
John decided that he wanted to buy a beer, so he asked Sally if she wanted a glass
of red or white wine and John proceeded to buy them.
John and Sally found seats in the back.
189
It was still a few minutes before show time so the theater was still empty, making it
easier for John and Sally to move down the row to their seats.
Having sat down, John asked Sally about her day and they enjoyed a bit of small talk
while waiting for the show to begin.
Just as Sally had finished relating the events of that day to John, the theater fully
darkened, the screen was unveiled, and the show began.
John and Sally shared popcorn during the movie, feeding each other and laughing.
John and Sally drank the ice-cold sodas, which went perfectly with the buttered
popcorn.
Finally working up the courage to do so, John extended his arm to embrace Sally. He
was relieved and ecstatic to feel her move closer to him in response.
Both feeling unsure of the others’ next move, they watched the movie in silence.
Sally stood up to use the restroom during the movie, smiling coyly at John before
that exit.
John and Sally sat with big smiles on their faces as they watched the movie.
John hugged Sally during the movie, and she returned the hug.
Sally held John’s sweaty hands throughout the movie.
The movie ended and John and Sally remained seated for a few minutes, waiting for
the crowd to leave the theater.
Sally stood up first, followed by John, and they both stretched quickly to revive their
bodies after a long time of sitting.
Still holding hands, John walked Sally back to his car through the maze of people all
scurrying out of the theater.
John and Sally carefully left the movie theater parking lot, not wanting to be hit by
another car.
John let go of Sally’s hand and opened the passenger side door of his car for her but
instead of entering the car, she stepped forward, embraced him, and gave him a large
190
kiss.
B.2.5 The Story with Most Interesting Details and Length Penalty
John drove to Sally’s house, and nervously checked himself out in the rearview mirror
before ringing her doorbell.
Sally, waiting for John on the front porch, slid into the passenger seat with a smile.
John paid for Sally at the ticket booth.
John and Sally showed their tickets to the attendant who smiled and waited for them
at the entrance to the theater.
John and Sally were surprised how expensive the snacks were.
John and Sally entered the darkened theater and paused letting their eyes adjust to
the dimness.
Drinks in hand, John and Sally found two seats near the back row.
John and Sally sat in chairs.
John and Sally spoke curtly to each other about politics.
Once the movie began, John and Sally sat quietly.
John thought the popcorn was extra buttery.
John and Sally drank the sugary sodas.
John nervously put his arm around Sally and she snuggled into him.
John and Sally watched the movie silently.
Sally stood up to use the restroom during the movie, smiling coyly at John before
that exit.
John and Sally laughed and smiled at the movie plot.
John hugged Sally during the movie, and she returned the hug.
John and Sally also held hands throughout the movie, even though John’s hands were
sweaty.
John and Sally smiled as the movie ended and the lights came on.
191
John and Sally slowly got up from their seats.
John and Sally walked to the car in the parking garage.
John held the door open for Sally as they exited the theater.
John and Sally very softly kissed.
192
REFERENCES
[1] Ankerst, M., Breunig, M. M., Kriegel, H.-P., and Sander, J., “OP-
TICS: Ordering points to identify the clustering structure,” in Proceedings of the
ACM SIGMOD International Conference on Management of Data, pp. 49–60,
1999. 3.3, 3.3.2, 4, 3.3.2
[2] Applebee, A., The Child’s Concept of Story: Ages 2 to 17. University of
Chicago Press, 1978. 1.3
[3] Bae, B.-C., Cheong, Y.-G., and Young, R. M., “Automated story gen-
eration with multiple internal focalization,” in Proceedings of the 2011 IEEE
Conference on Computational Intelligence and Games, pp. 211–218, 2011. 2.2.2
[4] Bae, B.-C. and Young, R. M., “A use of flashback and foreshadowing for
surprise arousal in narrative using a plan-based approach,” in Proceedings of the
1st Joint International Conference on Interactive Digital Storytelling (Spier-
ling, U. and Szilas, N., eds.), pp. 156–167, 2008. 1.2, 2.1.1, 2.2.2
[5] Bal, M., Narratology: Introduction to the Theory of Narrative. Toronto,
Canada: University of Toronto Press, 1997. 1.1, 2.1.1, 4
[6] Bamman, D., O’Connor, B., and Smith, N. A., “Learning latent personas
of film characters,” in Proceedings of the 51st Annual Meeting of the Associa-
tion for Computational Linguistics (Volume 1: Long Papers), (Sofia, Bulgaria),
pp. 352–361, Association for Computational Linguistics, August 2013. 2.2.4
[7] Bamman, D., Underwood, T., and Smith, N. A., “A bayesian mixed ef-
fects model of literary character,” in Proceedings of the 52nd Annual Meeting of
the Association for Computational Linguistics (Volume 1: Long Papers), (Bal-
timore, Maryland), pp. 370–379, Association for Computational Linguistics,
June 2014. 2.2.4
[8] Barthes, R., “Structural analysis of narratives,” in Image, Music, Text, New
York: Hill and Wang, 1977. 4.3
[9] Barzilay, R. and Lapata, M., “Modeling local coherence: An entity-based
approach,” in Proceedings of the 43rd Annual Meeting on Association for Com-
putational Linguistics, ACL ’05, (Stroudsburg, PA, USA), pp. 141–148, Asso-
ciation for Computational Linguistics, 2005. 2.5, 4.6
[10] Bengio, Y., Louradour, J., Collobert, R., and Weston, J., “Curricu-
lum learning,” in Proceedings of the 26th Annual International Conference on
Machine Learning, 2009. 2.4.1
193
[11] Berliner, T. and Cohen, D. J., “The illusion of continuity: Active percep-
tion and the classical editing system,” Journal of Film and Video, vol. 63, no. 1,
pp. 44–63, 2011. 2.1
[12] Bickmore, B., Schulman, D., and Yin, L., “Engagement vs. deceit: Virtual
humans with human autobiographies,” in Proceedings of the 9th International
Conference on Intelligent Virtual Agents, 2009. 6.3
[13] Boden, M., The Creative Mind: Myths and Mechanisms. Weidenfeld/Abacus
& Basic Books, 1990. 4.7
[14] Boujarwah, F. A., Kim, J. G., Riedl, M. O., Arriaga, R. I., and
Abowd, G. D., “Building a knowledge base to support the authoring of social
skills instructional modules,” in Proceedings of the 2001 International Meeting
for Autism Research, (San Diego, CA), 2011. 1.4, 6.3
[15] Brandt, L. and Brandt, P. A., “Making sense of a blend,” Apparatur,
vol. 4, pp. 62–71, 2002. 6.4.1
[16] Branigan, E., Narrative Comprehension and Film. Routledge, 1992. 1.3, 1
[17] Brenner, M., “Creating dynamic story plots with continual multiagent plan-
ning,” in Proceedings of the 24th National Conference on Artificial Intelligence,
2011. 2.2.1
[18] Brin, S. and Page, L., “The anatomy of a large-scale hypertextual web search
engine,” Computer Networks and ISDN Systems, vol. 30, pp. 107–117, 1998.
4.3.1
[19] Brody, S., “Clustering clauses for high-level relation detection: an
information-theoretic approach,” in Proceedings of the 43rd Annual Meeting
of the Association for Computational Linguistics, 2007. 2.4.1
[20] Bruner, J., Acts of Meaning. Cambridge, MA: Harvard University Press. 1,
1.2
[21] Bruner, J., “The narrative construction of reality,” Critical Inquiry, vol. 18,
pp. 1–21, 1991. 1.2
[22] Callaway, C. B. and Lester., J. C., “Narrative prose generation,” Artificial
Intelligence, vol. 139, pp. 213–252, 2002. 2.2.3
[23] Carlson, L., Marcu, D., and Okurowski, M. E., “Building a discourse-
tagged corpus in the framework of rhetorical structure theory,” in Proceedings
of the 2nd SIGDIAL Workshop on Discourse and Dialogue, pp. 1–10, 2003. 2.4
[24] Cavazza, M., Charles, F., and Mead, S. J., “Generating content for
scenario-based serious-games using crowdsourcing,” in Proceedings of the 1st
International Joint Conference on Autonomous Agents and Multiagent Systems,
2002. 2.2.1
194
[25] Cerini, S., Compagnoni, V., Demontis, A., Formentelli, M., and Gan-
dini, C., “Micro-WNOp: A gold standard for the evaluation of automatically
compiled lexical resources for opinion mining,” in Language resources and lin-
guistic theory: Typology, Second Language Acquisition, English Linguistics, Mi-
lan, Italy: Franco Angeli Editore, 2007. 2.3
[26] Chambers, N. and Jurafsky, D., “Unsupervised learning of narrative event
chains,” in Proceedings of the 46th Annual Meeting of the Association for Com-
putational Linguistics: Human Language Technologies, pp. 789–797, 2008. 2.4.1,
3.1, 3.6
[27] Chambers, N. and Jurafsky, D., “Unsupervised learning of narrative
schemas and their participants,” in Proceedings of the Joint Conference of the
47th Annual Meeting of the ACL and the 4th International Joint Conference on
Natural Language Processing of the AFNLP, ACL ’09, (Singapore), pp. 602–610,
Association for Computational Linguistics, 2009. 2.4.1, 3.6
[28] Chambers, N. and Jurafsky, D., “Template-based information extraction
without the templates,” in Proceedings of the 49th Annual Meeting of the Asso-
ciation for Computational Linguistics: Human Language Technologies, (Port-
land, Oregon), pp. 976–986, Association for Computational Linguistics, 2011.
2.4.1
[29] Chatman, S. B., Story and Discourse: Narrative Structure in Fiction and
Film. Ithaca and London: Cornell University Press, 1978. 4.3, 4.3.3
[30] Chen, S., Nelson, M., and Mateas, M., “Evaluating the authorial leverage
of drama management,” in Proceedings of the 5th AAAI Conference on AI and
Interactive Digital Entertainment, AAAI Press, 2009. 4.1.2
[31] Cheng, P., “From covariation to causation: A causal power theory,” Psycho-
logical Review, vol. 104, pp. 367–405, 1997. 3.10
[32] Cheng, W., Rademaker, M., Baets, B. D., and H
¨
ullermeier, E., “Pre-
dicting partial orders: Ranking with abstention,” in Machine Learning and
Knowledge Discovery in Databases (Balc
´
azar, J., Bonchi, F., Gionis, A.,
and Sebag, M., eds.), pp. 215–230, Springer, 2010. 3.6.2
[33] Cheong, Y.-G., A Computational Model of Narrative Generation for Sus-
pense. Ph.D. dissertation, North Carolina State University, Department of
Computer Science, 2007. 1.2, 2.1.1
[34] Cheong, Y.-G. and Young, R. M., “Suspenser: A story generation system
for suspense,” IEEE Transactions on Computational Intelligence and AI in
Games, 2014. 2.2.2, 2.2.4, 4.3
[35] Colby, B. N., “A partial grammar of eskimo folktales,” American Anthropol-
ogist, no. 75, pp. 645–662, 1973. 2.2.1
195
[36] Coppersmith, D. and Winograd, S., “Matrix multiplication via arithmetic
progressions,” Journal of Symbolic Computation, vol. 9, no. 3, pp. 251–280,
1990. 3.8
[37] Cullingford, R. E., Script Application: Computer Understanding of News-
paper Stories. Ph.D. dissertation, Yale University, Department of Computer
Science, 1978. Technical Report 116. 1.3, 2.2.4, 5
[38] Cullingford, R. E., “SAM,” in Inside Computer Understanding: Five Pro-
grams Plus Miniatures (Schank, R. C. and Riesbeck, C. K., eds.), 1981. 1,
2.2.4
[39] Curtis, J., Baxter, D., Wagner, P., Cabral, J., Schneider, D., and
Witbrock, M. J., “Methods of rule acquisition in the textlearner system,” in
AAAI Spring Symposium: Learning by Reading and Learning to Read, pp. 22–
28, AAAI. 2.2.4
[40] Dautenhahn, K., “The narrative intelligence hypothesis: In search of the
transactional format of narratives in humans and other social animals,” in
Cognitive Technology: Instruments of Mind (Beynon, M., Nehaniv, C. L.,
and Dautenhahn, K., eds.), vol. 2117 of Lecture Notes in Computer Science,
pp. 248–266, Springer Berlin Heidelberg, 2001. 1, 1.2
[41] Davis, M. and Travers, M., “A brief overview of the narrative intelligence
reading group,” in Narrative Intelligence (Mateas, M. and Sengers, P.,
eds.), pp. 27–38, John Benjamins, 2003. 2
[42] Davis, N., Li, B., O’Neill, B., Riedl, M., and Nitsche, M., “Distributed
creative cognition in digital filmmaking,” in Proceedings of the 8th ACM con-
ference on Creativity and Cognition, pp. 207–216, ACM, 2011. 4.7
[43] de Albornoz, J. C., Plaza, L., and Gerv
´
as, P., “SentiSense: An easily
scalable concept-based affective lexicon for sentiment analysis,” in Proceedings
of the 8th International Conference on Language Resources and Evaluation,
LREC ’12, 2012. 2.3
[44] de Marneffe, M.-C. and Manning, C. D., “Stanford typed dependen-
cies manual,” 2008. Revised 2013. Last retrived from nlp.stanford.edu/
software/dependencies_manual.pdf. 3.3.1
[45] Dunbar, R., Grooming, Gossip and the Evolution of Language. Harvard Uni-
versity Press, 1997. 1.2
[46] Dundes, A., The Morphology of North American Indian Folktales. Folklore
Fellows Communications, 1964. 2.2.1
[47] Dyer, M. G., “The role of affect in narratives,” Cognitive Science, vol. 7,
pp. 211–242, 1983. [U3]
196
[48] Elman, J. L., “Learning and development in neural networks: The importance
of starting small,” Cognition, vol. 48, pp. 781–799, 1993. 2.4.1
[49] Elsner, M., “Character-based kernels for novelistic plot structure,” in Pro-
ceedings of the 13th Conference of the European Chapter of the Association for
Computational Linguistics, EACL ’12, (Stroudsburg, PA, USA), pp. 634–644,
Association for Computational Linguistics, 2012. 2.2.4
[50] Elson, D., Modeling Narrative Discourse. Ph.D. dissertation, Columbian Uni-
versity, Department of Computer Science, 2012. 3, 2.2.3, 2.5
[51] Esuli, A., Baccianella, S., and Sebastiani, F., “SentiWordNet 3.0: An
enhanced lexical resource for sentiment analysis and opinion mining,” in Pro-
ceedings of the Seventh International Conference on Language Resources and
Evaluation (LREC’10) (Calzolari, N., Choukri, K., Maegaard, B.,
Mariani, J., Odijk, J., Piperidis, S., Rosner, M., and Tapias, D.,
eds.), (Valletta, Malta), pp. 19–21, European Language Resources Association
(ELRA), 2010. 2.3, 4.4.2
[52] Etzioni, O., Hanks, S., Daniel Weld, D. D., Lesh, N., and
Williamson, M., “An approach to planning with incomplete information,”
in Proceedings of the 3rd International Conference on Principles of Knowledge
Representation and Reasoning, pp. 115–125, 1992. 1.4.1
[53] Fairclough, C. and Cunningham, P., “A multiplayer case based story en-
gine,” Tech. Rep. TCD-CS-2003-43, Department of Computer Science, Trinity
College Dublin, 2003. 2.2.1
[54] Falkenhainer, B., K., F., and Gentner, D., “The structure-mapping en-
gine: Algorithm and examples,” Artificial Intelligence, vol. 20, no. 41, p. 163,
1989. 2.2.1
[55] Finlayson, M., Learning Narrative Structure from Annotated Folktales. Ph.D.
dissertation, Massachusetts Institute of Technology, Department of Electrical
Engineering and Computer Science, 2011. 2.5
[56] Fivush, R., Kuebli, J., and Clubb, P. A., “The structure of events and
event representations: A developmental analysis,” Child Development, vol. 92,
pp. 188–201, 1992. 1.4.1
[57] Flanagan, O., Consciousness Reconsidered. Bradford, 1993. 1, 1.2
[58] Flower, L. and Hayes, J., “A cognitive process theory of writing,” College
Composition and Communication, vol. 32, no. 4, pp. 365–387, 1981. 1.2
[59] Forster, E. M., Aspects of the Novel. Edward Arnold, 1927. 2.1, 2.1.1
[60] Francis, W. N. and Ku
ˇ
cera, H., Frequency Analysis of English Usage.
Houghton Mifflin, 1982. 3.3.1
197
[61] Friedemann, K., Die Rolle des Erz¨ahlers in der Epik. 1910. Reprint: Darm-
stadt 1965. 2.1
[62] Fujiki, T., Nanba, H., and Okumura, M., “Automatic acquisition of script
knowledge from a text collection,” in Proceedings of the Tenth Conference on
European chapter of the Association for Computational Linguistics, EACL ’03,
(Budapest, Hungary), pp. 91–94, Association for Computational Linguistics,
2003. 2.4.1, 2
[63] Garc
´
ıa Landa, J. A., Structural Narratology: An Introduction. Longman,
1994. 2.1.1
[64] Genette, G., Nouveau discours du ecit. Seuil, 1983. 2.1, 2.1.1
[65] Gerrig, R. J. and Bernardo, A. B., “Readers as problem-solvers in the
experience of suspense,” Poetics, vol. 22, pp. 459–472, 1994. 2.2.2, 2.2.4
[66] Gerv
´
as, P., “Computational approaches to storytelling and creativity,” AI
Magazine, vol. 30, no. 3, pp. 49–62, 2009. 4.7
[67] Gerv
´
as, P., D
´
ıaz-Agudo, B., Peinado, F., and Herv
´
as, R., “Story
plot generation based on CBR,” Journal of Knowledge-Based Systems, vol. 18,
pp. 235–242, 2005. 1.3, 2.2.1
[68] Gerv
´
as, P. and Le
´
on, C., “Reading and writing as a creative cycle: The need
for a computational model,” in Proceedings of the 5th International Conference
on Computational Creativity, (Ljubljana, Slovenia), 2014. 1.2
[69] Girju, R., “Automatic detection of causal relations for question answering,”
in Proceedings of the ACL 2003 Workshop on Multilingual Summarization and
Question AnsweringMachine Learning and Beyond, 2003. 2.4.1
[70] Gockley, R., Bruce, A., Forlizzi, J., Michalowski, M., Mundell,
A., Rosenthal, S., Sellner, B., Simmons, R., Snipes, K., Schultz,
A. C., and Wang, J., “Designing robots for long-term social interaction,” in
Proceedings of 2005 IEEE/RSJ International Conference on Intelligent Robots
and Systems, pp. 2199–2204, 2005. 1.2
[71] Gordon, A., Bejan, C., and Sagae, K., “Commonsense causal reasoning
using millions of personal stories,” in Proceedings of the 25th AAAI Conference
on Artificial Intelligence, (San Francisco, CA), AAAI Press, 2011. 2.4.1
[72] Graesser, A., Singer, M., and Trabasso, T., “Constructing inferences
during narrative text comprehension,” Psychological Review, vol. 101, pp. 371–
395, 1994. 1.4.1, 4.3
[73] Graesser, A. C., Lang, K. L., and Roberts, R. M., “Question answering
in the context of stories,” Journal of Experimental Psychology:General, vol. 120,
pp. 254–277, 1991. 5
198
[74] Grisham, T., “Metaphor, poetry, storytelling and cross-cultural leadership,”
Management Decision, vol. 44, pp. 486–503, 2006. 1.2
[75] Gurobi Optimization, Inc., “Gurobi optimizer reference manual,” 2014.
3.6.2, 3.6.3
[76] Haveliwala, T. H., “Topic-sensitive PageRank,” in Proceedings of the 11th
International Conference on World Wide Web, pp. 517–526, 2002. 4.3, 4.3.1
[77] Hedlund, J., Antonakis, J., and Sternberg, R., “Tacit knowledge and
practical intelligence: understanding the lessons of experience,” Tech. Rep. ARI
Research Note 2003-04, United States Army Research Institute for the Behav-
ioral and Social Sciences, Fort Belvoir, VA, 2002. 3.2
[78] Heilmann, J., Miller, J., Nockerts, A., and Dunaway, C., “Properties
of the narrative scoring scheme using narrative retells in young school-age chil-
dren,” American Journal of Speech-Language Pathology, vol. 19, no. 2, pp. 154–
166, 2010. 4.3
[79] Hilton, D. J. and Slugoski, B. R., “Knowledge-based causal attribution:
The abnormal conditions focus model,” Psychological Review, vol. 93, no. 1,
pp. 75–88, 1986. 3.10
[80] Huang, J., “Maximum likelihood estimation of dirichlet distribution parame-
ters,” 2005. 3.3.3
[81] Hudson, J. A. and Shapiro, L. A., “From knowing to telling: The devel-
opment of children’s scripts, stories, and personal narratives,” in Developing
narrative structure, pp. 89–136, Hillsdale, NJ: Lawrence Erlbaum, 1991. 1.3
[82] Ipeirotis, P. G., Provost, F., and Wang, J., “Quality management on
Amazon Mechanical Turk,” in Proceedings of the ACM SIGKDD Workshop on
Human Computation, pp. 64–67, 2010. 3.2
[83] Johnston, J., “Narratives: Twenty-five years later,” Topics in Language Dis-
orders, vol. 28, pp. 93–98, 2008. 1.2
[84] Kann, V., On the approximability of NP-complete optimization problems.
Ph.D. dissertation, Royal Institute of Technology, Department of Numerical
Analysis and Computing Science, 1992. 3.6.2
[85] Kasch, N. and Oates, T., “Mining script-like structures from the web,” in
Proceedings of the NAACL-HLT 2010 Workshop on Formailism and Methodol-
ogy for Learning by Reading, 2010. 1
[86] Klein, D. and Manning, C., “Accurate unlexicalized parsing,” in Proceedings
of the 41st Meeting of the Association for Computational Linguistics, pp. 423–
430, 2003. 3.3.1
199
[87] Kolodner, J. L., “An introduction to case-based reasoning,” Artificial Intel-
ligence Review, vol. 6, no. 1, pp. 3–34, 1992. 2.2.1
[88] Krueger, K. A. and Dayan, P., “Flexible shaping: how learning in small
steps helps.,” Cognition, vol. 110, pp. 380–394, 2009. 2.4.1
[89] Kuhn, H. W., “The Hungarian Method for the assignment problem,” Naval
Research Logistics Quarterly, vol. 2, pp. 83–97, 1955. 3.3.1
[90] Labov, W., The Language of Life and Death. Cambridge, U.K.: Cambridge
University Press, 2013. 2.2.4
[91] Labov, W. and Waletzky, J., “Narrative analysis: Oral versions of personal
experience,” in Essays on the Verbal and Visual Arts (Helm, J., ed.), pp. 12–
44, 1967. 2.2.4
[92] Lebowitz, M., “Planning stories,” in Proceedings of the 9th Annual Confer-
ence of the Cognitive Science Society, 1987. 1.3, 2.2.1
[93] Lee, G., Bulitko, V., and Ludvig, E., “Sports commentary recommenda-
tion system (SCoReS): Machine learning for automated narrative,” in Proceed-
ings of the 8th AAAI AIIDE Conference on Artificial Intelligence and Interac-
tive Digital Entertainment, 2012. 1.2
[94] Lehnert, W., “Cognition, computers and car bombs: How Yale prepared me
for the 1990s,” in Beliefs, reasoning and decision making (Schank, R. and
Langer, E., eds.), Hillsdale, NJ: Erlbaum, 1994. 3
[95] Leonard, S. A. and McClure, M., Myth and Knowing: An Introduction to
World Mythology. McGraw-Hill, 2003. 1
[96] Levesque, H. J., “What is planning in the presence of sensing?,” in Pro-
ceedings of the 13th National Conference on Artificial Intelligence, (Portland,
Oregon), pp. 1139–1146, AAAI Press, 1996. 1.4.1
[97] Li, B. and Riedl, M. O., “An offline planning approach to game plotline
adaptation,” in Proceedings of the 6th Conference on Artificial Intelligence for
Interactive Digital Entertainment, (Palo Alto, CA), pp. 45–50, AAAI Press,
2010. 1.2, 1.3, 2.2.1
[98] Li, B. and Riedl, M. O., “Creative gadget design in fictions: Generating novel
object types in analogical spaces,” in Proceedings of the 8th ACM Conference
on Creativity and Cognition, (Atlanta, Georgia), pp. 41–50, 2011. 6.4.1
[99] Li, B. and Riedl, M. O., “A phone that cures your flu: Generating imagi-
nary gadgets in fictions with planning and analogies,” in Proceedings of the 4th
Workshop on Intelligent Narrative Technologies, (Palo Alto, CA), pp. 41–48,
AAAI Press, 2011. 2.2.1, 3.1, 3.10
200
[100] Li, B., Zook, A., Davis, N., and Riedl, M. O., “Goal-driven conceptual
blending: A computational approach for creativity,” in Proceedings of the 2012
International Conference on Computational Creativity, 2012. 4.7
[101] Li, B., Zook, A., Davis, N., and Riedl, M. O., “Goal-driven conceptual
blending: A computational approach for creativity,” in Proceedings of the 3rd
International Conference on Computational Creativity, (Dublin, Ireland), 2012.
6.4.1
[102] Lieberman, H., Dinakar, K., and Jones, B., “Crowdsourced ethics with
personalized story matching,” in CHI ’13 Extended Abstracts on Human Factors
in Computing Systems, CHI EA ’13, (New York, NY, USA), pp. 709–714, ACM,
2013. 1.2
[103] Lin, D., Church, K., Ji, H., Sekine, S., Yarowsky, D., Bergsma,
S., Patil, K., Pitler, E., Lathbury, R., Rao, V., Dalwani, K., and
Narsale, S., “New tools for web-scale n-grams,” in Proceedings of the 7th
International Conference on Language Resources and Evaluation, (Valleta,
Malta), 2010. 3.3.3, 3.3.3
[104] Lincoln, B., Theorizing myth: Narrative, ideology, and scholarship. Univer-
sity of Chicago Press, 1999. 1
[105] Lintean, M. C. and Rus, V., “Paraphrase identification using weighted de-
pendencies and word semantics,” Informatica, vol. 34, pp. 19–28, 2010. 3.3.1
[106] Liu, H. and Singh, P., “MAKEBELIEVE: Using commonsense to generate
stories,” in Proceedings of the 18th National Conference on Artificial Intelli-
gence, 2002. 2.4.2
[107] Liu, J. and Birnbaum, L., “Localsavvy: aggregating local points of view
about news issues,” in Proceedings of the 1st International Workshop on Loca-
tion and the Web, 2008. 1.2
[108] Lu, Y., Castellanos, M., Dayal, U., and Zhai, C., “Automatic con-
struction of a context-aware sentiment lexicon: An optimization approach,” in
Proceedings of the International World Wide Web Conference, 2011. 2.3, 4.4.2
[109] Magerko, B., Manzoul, W., Riedl, M., Baumer, A., Fuller, D.,
Luther, K., and Pearce, C., “An empirical study of cognition and the-
atrical improvisation,” in Proceedings of the ACM Conference on Creativity
and Cognition, pp. 117–126, ACM, 2009. 2.1
[110] Maher, M. L., “Computational and collective creativity: Who’s being cre-
ative?,” in Proceedings of The 3rd International Conference on Computational
Creativity, (Dublin, Ireland), pp. 67–71, 2012. 3
201
[111] Mairesse, F. and Walker, M., “Towards personality-based user adapta-
tion: Psychologically informed stylistic language generation,” User Modeling
and User-Adapted Interaction, vol. 20, no. 3, pp. 227–278, 2010. 2.2.3
[112] Mar, R. A., Oatley, K., Djikic, M., and Mullin, J., “Emotion and narra-
tive fiction: Interactive influences before, during, and after reading,” Cognition
and Emotion, vol. 25, pp. 818–833, 2011. [U3]
[113] Mateas, M. and Sengers, P., “Narrative intelligence,” in Proceedings of
the AAAI Fall Symposium on Narrative Intelligence, (North Falmouth, MA),
pp. 1–10, AAAI, 1999. 1, 1.2
[114] McIntyre, N. and Lapata, M., “Plot induction and evolutionary search for
story generation,” in Proceedings of the 48th Annual Meeting of the Association
for Computational Linguistics, pp. 1562–1572, 2010. 1.4, 2.5, 6.5
[115] Meehan, J., The Metanovel: Writing Stories by Computers. Ph.D. disserta-
tion, Yale University, 1976. 1.3, 2.2.1
[116] Miller, G., “WordNet: A lexical database for english,” Communications of
the Association for Computing Machinery, vol. 38, pp. 39–41, 1995. 2.3, 3.3.1,
4.4.2
[117] Mohammad, S., “From once upon a time to happily ever after: Tracking emo-
tions in novels and fairy tales,” in Proceedings of the 5th ACL-HLT Workshop on
Language Technology for Cultural Heritage, Social Sciences, and Humanities,
(Portland, OR, USA), pp. 105–114, Association for Computational Linguistics,
June 2011. 2.2.4, 2.3
[118] Mohammad, S. M. and Turney, P. D., “Emotions evoked by common words
and phrases: Using Mechanical Turk to create an emotion lexicon,” in Pro-
ceedings of the NAACL-HLT 2010 Workshop on Computational Approaches to
Analysis and Generation of Emotion in Text, CAAGET ’10, (Stroudsburg, PA,
USA), pp. 26–34, Association for Computational Linguistics, 2010. 2.3
[119] Montfort., N., Generating Narrative Variation in Interactive Fiction. Ph.D.
dissertation, University of Pennsylvania, 2007. 1.2, 2.2.2
[120] Moorman, K. and Ram, A., A Functional Theory of Creative Reading. Tech-
nical report GIT-CC-94/01, College of Computing, Georgia Institute of Tech-
nology, 1994. 1.2, 2.2.4, 6.4.1
[121] Nelson, M. J. and Mateas, M., “Search-based drama management in the
interactive fiction Anchorhead,” in Proceedings of the 1st Artificial Intelligence
and Interactive Digital Entertainment Conference (AIIDE), (Marina del Rey,
California), June 2005. 3.1, 3.7
202
[122] Niehaus, J., Cognitive Models of Discourse Comprehension for Narrative Gen-
eration. Ph.D. dissertation, North Carolina State University, Department of
Computer Science, 2009. 2.2.4
[123] Niehaus, J., Li, B., and Riedl, M. O., “Automated scenario adaptation in
support of intelligent tutoring systems,” in Proceedings of the 24th Conference
of the Florida Artificial Intelligence Research Society, (Palm Beach, FL), 2011.
1.2
[124] Norman, W. T., “Toward an adequate taxonomy of personality attributes:
Replicated factor structure in peer nomination personality ratings,” Journal of
Abnormal and Social Psychology, vol. 66, no. 6, p. 574583, 1963. 2.2.3
[125] Novick, L. and Cheng, P. W., “Assessing interactive causal influence,” Psy-
chological Review, vol. 111, no. 2, pp. 455–485, 2004. 3.10
[126] Oatley, K., “A taxonomy of the emotions of literary response and a theory
of identification in fictional narrative,” Poetics, vol. 23, pp. 53–74, 1994. [U3]
[127] Oleson, D., Sorokin, A., Laughlin, G. P., Hester, V., Le, J., and
Biewald, L., “Programmatic gold: Targeted and scalable quality assurance
in crowdsourcing,” in Proceedings of the 3rd Human Computation Workshop,
pp. 64–67, 2010. 3.2
[128] O’Neill, B. and Riedl, M. O., “Dramatis: A computational model of sus-
pense,” in Proceedings of the 28th AAAI Conference on Artificial Intelligence,
2014. 2.2.2, 2.2.4, 3.10
[129] Onta
˜
n
´
on, S. and Zhu, J., “On the role of domain knowledge in analogy-based
story generation,” in Proceedings of the 22nd International Joint Conferences
on Artificial Intelligence, (Barcelona, Spain), pp. 1717–1722, 2011. 2.2.1
[130] Orkin, J. D., Collective Artificial Intelligence: Simulated Role-Playing from
Crowdsourced Data. Ph.D. dissertation, Massachusetts Institute of Technology,
Department of Electrical Engineering and Computer Science, 2012. 2.5
[131] Ounis, I., Lioma, C., Macdonald, C., and Plachouras, V., “Research
directions in Terrier,” Novatica/UPGRADE Special Issue on Web Information
Access, 2007. 2.5
[132] Ouyang, J. and Mckeown, K., “Towards automatic detection of narrative
structure,” in Proceedings of the 9th International Conference on Language
Resources and Evaluation (LREC’14) (Calzolari, N., Choukri, K., De-
clerck, T., Loftsson, H., Maegaard, B., Mariani, J., Moreno, A.,
Odijk, J., and Piperidis, S., eds.), (Reykjavik, Iceland), European Language
Resources Association (ELRA), May 2014. 2.2.4
[133] Pang, B. and Lee, L., “Opinion mining and sentiment analysis,” Foundations
and trends in information retrieval, vol. 2, pp. 1–135, 2008. 2.3
203
[134] Pearl, J., “The foundations of causal inference,” Sociological Methodology,
vol. 40, p. 75149, 2010. 3.10
[135] Pennington, N. and Hastie, R., “The story model for juror decision mak-
ing,” in Inside the Juror: The Psychology of Juror Decision Making (Hastie,
R., ed.), pp. 192–221, Cambridge, UK: Cambridge University Press, 1993. 1.2
[136] P
´
erez y P
´
erez, R. and Sharples, M., “MEXICA: A computer model of a
cognitive account of creative writing,” Journal of Experimental and Theoretical
Artificial Intelligence, vol. 13, pp. 119–139, 2001. 1.3, 2.2.1
[137] P
´
erez y P
´
erez, R. and Sharples, M., “Three computer-based models
of storytelling: BRUTUS, MINSTREL and MEXICA,” Knowledge-Based Sys-
tems, vol. 17, no. 1, pp. 15–29, 2004. 2.2.1
[138] Porteous, J., Cavazza, M., and Charles, F., “Narrative generation
through characters point of view,” in The SIGCHI Conference on Human Fac-
tors in Computing Systems, 2010. 2.2.2
[139] Porteous, J., Teutenberg, J., Charles, F., and Cavazza, M., “Control-
ling narrative time in interactive storytelling,” in Proceedings of the 10th Inter-
national Conference on Autonomous Agents and Multiagent Systems (Tumer,
Yolum, Sonenberg, and Stone, eds.), (Taipei, Taiwan), pp. 449–456, 2011.
2.2.1
[140] Prasad, R., Dinesh, N., Lee, A., Miltsakaki, E., Robaldo, L., Joshi,
A., and Webber, B., “The Penn Discourse Treebank 2.0,” in Proceedings of
the 6th International Conference on Language Resources and Evaluation, 2008.
2.4
[141] Prince, G., A Dictionary of Narratology. Lincoln, NE: University of Nebraska
Press, 1987. 1.1, 1.1, 2.1, 2.1
[142] Propp, V. Y., Morphology of the Folktale. Austin, TX: University of Texas
Press, 1968. Originally published in Russian in 1928. 2.2.1
[143] Purse, L., “The new spatial dynamics of the bullet-time effect,” in From
Hollywood to “Reality” TV and Beyond, pp. 151–160, 2005. 6.4.2
[144] Pustejovsky, J., Hanks, R., P. S., See, A., Gaizauskas, R., Setzer,
A., Radev, D., Sundheim, B., Day, D., Ferro, L., and Lazo, M., “The
TIMEBANK corpus,” in Proceedings of Corpus Linguistics, pp. 647–656, 2003.
2.4, 3.6
[145] Ram, A., Question-Driven understanding: An integrated theory of story un-
derstanding, memory, and learning. Ph.D. dissertation, Yale University, De-
partment of Computer Science, 1989. Research Report No. 710. 2.2.4, 2.5
204
[146] Ram, A., “Creative conceptual change,” in Proceedings of the 15th Annual
Conference of the Cognitive Science Society, (Boulder, CO), p. 1726, 1993.
6.4.1
[147] Regneri, M., Koller, A., and Pinkal, M., “Learning script knowledge
with web experiments,” in Proceedings of the 48th Annual Meeting of the As-
sociation for Computational Linguistics, 2010. 2.4.2, 3.6
[148] Resnik, P., “Using information content to evaluate semantic similarity in a
taxonomy,” in Proceedings of the 14th International Joint Conference on Arti-
ficial Intelligence, pp. 448–453, Morgan Kaufmann Publishers Inc, 1995. 3.3.1
[149] Riedl and Le
´
on, “Generating story analogues,” in Proceedings of The 5th
Conference on Artificial Intelligence in Interactive Digital Entertainment, (Palo
Alto, CA), 2009. 2.2.1
[150] Riedl, M. O., Narrative Generation: Balancing Plot and Character. Ph.D.
dissertation, North Carolina State University, Department of Computer Science,
2004. 2.1.1
[151] Riedl, M. O., Rowe, J. P., and Elson, D. K., “Toward intelligent support
of authoring machinima media content: Story and visualization,” in Proceedings
of the 2nd International Conference on Intelligent Technologies for Interactive
Entertainment (INTETAIN), (Playa del Carmen, Cancun, Mexico), 2008. 1.2
[152] Riedl, M. O., Stern, A., Dini, D., and Alderman, J., “Dynamic experi-
ence management in virtual worlds for entertainment, education, and training,”
International Transactions on Systems Science and Applications, vol. 3, no. 1,
pp. 23–42, 2008. 1.4.1
[153] Riedl, M. O. and Sugandh, N., “Story planning with vignettes: Toward
overcoming the content production bottleneck,” in Proceedings of the 1st Joint
International Conference on Interactive Digital Storytelling, (Erfurt, Germany),
2008. 1.3, 2.2.1
[154] Riedl, M. and Young, R., “Narrative planning: Balancing plot and charac-
ter,” Journal of Artificial Intelligence Research, vol. 39, pp. 217–268, 2010. 1.3,
2.2.1, 3.10
[155] Rishes, E., Lukin, S., Elson, D., and Walker, M., “Generating different
story tellings from semantic representations of narrative,” in Proceedings of the
6th International Conference on Interactive Storytelling, 2013. 2.2.3
[156] Rogelio E. Cardona-Rivera, Bradley A. Cassell, S. G. W. and
Young, R. M., “Indexter: A computational model of the event-indexing situ-
ation model for characterizing narratives,” in the Proceedings of the 2012 Work-
shop on Computational Models of Narrative, 2012. 2.2.4
205
[157] Rosch, E., “Cognitive representations of semantic categories,” Journal of Ex-
perimental Psychology: General, vol. 104, no. 3, p. 192233, 1975. 1.1, 2.1
[158] Ross, S. M., Introduction to Probability Models. Academic Press, 11th ed.,
2014. 4.3.1
[159] Rowe, J., Lobene, E., Mott, B., and Lester, J., “Play in the museum:
Designing game-based learning environments for informal education settings,”
in Proceedings of the 9th International Conference on the Foundations of Digital
Games, (Fort Lauderdale, FL), 2014. 1.2
[160] Sander, J., Qin, X., Lu, Z., Niu, N., and Kovarsky, A., “Automatic ex-
traction of clusters from hierarchical clustering representations,” in Proceedings
of the 7th Pacific-Asia Conference on Advances in Knowledge Discovery and
Data Mining, PAKDD ’03, (Berlin, Heidelberg), pp. 75–87, Springer-Verlag,
2003. 3.3.2
[161] Schank, R. and Abelson, R., Scripts, plans, goals, and understanding: An
inquiry into human knowledge structure. Hillsdale, NJ: Lawrence Erlbaum As-
sociates, 1977. 1.3, 2.2.4
[162] Schmid, W., Narratology: An Introduction. Walter de Gruyter, 2010. 2.1,
2.1.1, 2.1.1
[163] Sharples, M., How we write: Writing as creative design. London, UK: Rout-
ledge, 1999. 1.2, 2.2.1
[164] Shklovsky, V., Theory of Prose. Walter de Gruyter, 1990. Originally pub-
lished in 1929 in Russian. Translated by B. Sher. 2.1
[165] Si, M., Marsella, S. C., and Pynadath, D. V., “Thespian: Using multi-
agent fitting to craft interactive drama,” in Proceedings of the 4th International
Joint Conference on Autonomous Agents and Multiagent Systems, pp. 21–28,
2005. 2.2.1
[166] Sina, S., Rosenfeld, A., and Kraus, S., “Generating content for scenario-
based serious-games using crowdsourcing,” in Proceedings of the 28th AAAI
Conference on Artificial Intelligence, 2014. 1.4, 2.5, 6.5
[167] Singer, J. A., “Narrative identity and meaning making across the adult lifes-
pan: An introduction,” Journal of Personality, vol. 72, no. 3, pp. 437–460, 2004.
1.2
[168] Singh, P., Lin, T., Mueller, E. T., Lim, G., Perkins, T., and Zhu,
W. L., “Open Mind Common Sense: Knowledge acquisition from the gen-
eral public,” in Confederated International Conferences DOA, CoopIS and
ODBASE, pp. 1223–1237, Springer-Verlag, 2002. 2.4.2
206
[169] Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning, C., Ng, A.,
and Potts, C., “Recursive deep models for semantic compositionality over
a sentiment treebank,” in Proceedings of the 2013 Conference on Empirical
Methods in Natural Language Processing, 2013. 2.3, 4.5.3.1
[170] Sparacino, F., “Museum intelligence: using interactive technologies for effec-
tive communication and storytelling in the puccini set designer exhibit,” 2004.
1.2
[171] Stanzel, F. K., A Theory of Narrative. Cambridge, England: Cambridge
University Press, 1984. Translated by C. Goedsche. 2.1
[172] Stone, P., Dunphry, D. C., Smith, M. S., and Ogilvie, D. M., The
General Inquirer: A computer approach to content analysis. Cambridge, MA:
MIT Press, 1966. 2.3
[173] Strapparava, C. and Valitutti, A., “WordNet-Affect: an affective ex-
tension of WordNet,” in Proceedings of the 4th International Conference on
Language Resources and Evaluation, 2004. 2.3
[174] Strapparava, C., Valitutti, A., and Stock., O., “The affective weight
of lexicon,” in Proceedings of the 5th International Conference on Language
Resources and Evaluation, 2006. 2.3
[175] Strassen, V., “Gaussian elimination is not optimal,” Numerische Mathe-
matik, vol. 13, no. 4, pp. 354–356, 1969. 3.8
[176] Sunshine-Hill, B. and Badler, N., “Perceptually realistic behavior through
alibi generation,” in Proceedings of the 6th AAAI Conference on Artificial In-
telligence for Interactive Digital Entertainment, 2010. 6.3
[177] Swanson, R. and Gordon, A., “Say anything: Using textual case-based
reasoning to enable open-domain interactive storytelling,” ACM Transactions
on Interactive Intelligent Systems, vol. 2, 2012. 1.4, 2.5, 6.5
[178] Swanson, R., Rahimtoroghi, E., Corcoran, T., and Walker, M.,
“Identifying narrative clause types in personal stories,” in Proceedings of the
15th Annual Meeting of the Special Interest Group on Discourse and Dialogue
(SIGDIAL), (Philadelphia, PA, U.S.A.), pp. 171–180, Association for Compu-
tational Linguistics, June 2014. 2.2.4
[179] Swartjes, I., Whose story is it anyway? How improv informs agency and
authorship of emergent narrative. Ph.D. dissertation, University of Twente,
2010. 2.2.1
[180] Talukdar, P. P., Wijaya, D., and Mitchell, T., “Acquiring temporal con-
straints between relations,” in Proceedings of the 21st ACM international con-
ference on Information and knowledge management, CIKM ’12, (Maui, Hawaii,
USA), pp. 992–1001, ACM, 2012. 2.4.1, 2
207
[181] Tearse, B., Mawhorter, P., Mateas, M., and Wardrip-Fruin, N.,
“Skald: Minstrel reconstructed,” IEEE Transactions on Computational Intelli-
gence and AI in Games, vol. 6, pp. 156–165, June 2014. 2
[182] Todorov, T., “Categories of the literary narrative,” Film Reader, vol. 2,
pp. 19–37, 1977. 2.1, 2.1.1
[183] Tomai, E., A Pragmatic Approach to Computational Narrative Understand-
ing. Ph.D. dissertation, Northwestern University, Department of Elecrtrical
Engineering and Computer Science, 2009. Tech. Rep. No. NWU-EECS-09-17.
2.2.4
[184] Tomashevsky, B., “Thematics,” in Russian Formalist Criticism: Four Es-
says, pp. 62–95, University of Nebraska, 1965. 2.1.1
[185] Tomashevsky, B. V., Teoriya literatury. Poetika. 1925. Reprint: Letchworth
1971. 2.1, 2.1.1
[186] Toutanova, K., Klein, D., Manning, C., and Singer, Y., “Feature-rich
part-of-speech tagging with a cyclic dependency network,” in Proceedings of the
Human Language Technologies Conference, NAACL-HLT ’03, 2003. 4.4.2
[187] Trabasso, T. and Sperry, L., “Causal relatedness and importance of story
events,” Journal of Memory and Language, vol. 24, pp. 595–611, 1985. 1.4.1,
2.2.2
[188] Tupes, E. C. and Christal, R. E., “Recurrent personality factors based
on trait ratings,” Tech. Rep. USAF ASD No. 61-97, U. S. Air Force, Lackland
Airforce Base, TX, 1961. 2.2.3
[189] Turner, M. and Fauconnier, G., The Way We Think. Conceptual Blending
and the Mind’s Hidden Complexities. New York: Basic Books, 2002. 2.2.4, 6.4.1
[190] Turner, S. R., Minstrel: A Computer Model of Creativity and Storytelling.
Ph.D. dissertation, University of California at Los Angeles, 1993. 1.3, 2.2.1
[191] Turney, P. D. and Littman, M. L., “Measuring praise and criticism: In-
ference of semantic orientation from association,” ACM Transactions on Infor-
mation Systems, vol. 21, pp. 315–346, Oct. 2003. 2.3
[192] Valls-Vargas, J., Onta
˜
n
´
on, S., and Zhu, J., “Toward character role as-
signment for natural language stories,” in Proceedings of the 6th Workshop on
Intelligent Narrative Technologies, 2013. 2.2.4
[193] van den Broek, P., “Discovering the cement of the universe: The develop-
ment of event comprehension from childhood to adulthood,” in Developmental
spans in event comprehension: Bridging fictional and actual events, pp. 321–
342, Mahwah, NJ: Lawrence Erlbaum, 1997. 1.3
208
[194] Vayanou, M., Katifori, A., Karvounis, M., Kyriakidi, M., Roussou,
M., Tsangaris, M., Boile, M., Ioannidis, Y., Rennick-Egglestone,
S., and Pujol, L., “The impact of interactive digital storytelling in cultural
heritage sites,” in Proceedings of the International Digital Storytelling Confer-
ence: Digital Storytelling in Times of Crisis, (Athens, Greece), 2014. 1.2
[195] Vilain, M., Burger, J., Aberdeen, J., Connolly, D., and Hirschman,
L., “A model-theoretic coreference scoring scheme,” in Proceeding of the 6th
Conference on Message Understanding, pp. 45–52, 1995. 3.4
[196] Ware, S. G. and Young, R. M., “CPOCL: A narrative planner support-
ing conflict,” in Proceedings of The 7th Conference on Artificial Intelligence
in Interactive Digital Entertainment, (Palo Alto, CA), pp. 212–221, Oct 2011.
2.2.1
[197] Wauthier, F. and Jordan, M., “Bayesian bias mitigation for crowd-
sourcing,” in Advances in Neural Information Processing Systems 24 (NIPS)
(Shawe-Taylor, J., Zemel, R., Bartlett, P., Pereira, F., and Wein-
berger, K., eds.), pp. 1800–1808, 2011. 3.2
[198] Weltman, J. S., Iyengar, S. S., and Hegarty, M., “Mind the gap: Col-
lecting commonsense data about simple experiences,” in Proceedings of the 2013
International Conference on Intelligence User Interfaces, (Santa Monica, CA),
pp. 41–48, AAAI Press, 2013. 2.4.2
[199] Weyhrauch, P., Guiding Interactive Drama. Ph.D. dissertation, Carnegie
Mellon University, Department of Computer Science, 1978. Technical Report
CMU-CS-97-109. 3.1
[200] Wilensky, R., Planning and Understanding: A Computational Approach to
Human Reasoning. Reading, MA: Addison-Wesley Publishing Co., 1983. 1.3,
2.2.4
[201] Wilensky, R., “Why John married Mary: Understanding stories involving
recurring goals,” Cognitive Science, vol. 2, pp. 235–266, 1978. 3.10
[202] Williams, J. P., “Comprehension of students with and without learning dis-
abilities: Identification of narrative themes and idiosyncratic text representa-
tions,” Journal of Educational Psychology, vol. 85, pp. 631–641, 1993. 1.3
[203] Winston, P. H., “The strong story hypothesis and the directed perception
hypothesis,” in The AAAI Fall Symposium, 2011. 1, 1.2
[204] Wu, Z. and Palmer, M., “Verb semantics and lexical selection,” in Proceed-
ings of the 32nd Annual Meeting of the Association for Computational Linguis-
tics., p. 133138, 1994. 2.5, 3.3.1
209
[205] Young, R. M., Pollack, M. E., and Moore., J. D., “Decomposition and
causality in partial-order planning,” in The Proceedings of the Second Interna-
tional Conference on Artificial Intelligence and Planning Systems, 1994. 2.2.1
[206] Zhu, J., Onta
˜
n
´
on, S., and Lewter, B., “Representing game characters’ in-
ner worlds through narrative perspectives,” in The 6th International Conference
on Foundations of Digital Games, pp. 204–210, 2011. 2.2.2
[207] Zwaan, R. A. and Radvansky, G. A., “Situation models in language com-
prehension and memory,” Psychological Bulletin, vol. 123, no. 2, pp. 162–185,
1998. 4.3
210