Model-based Story Summary

Patrick Henry Winston

Computer Science and Artificial Intelligence Laboratory

Massachusetts Institute of Technology
Cambridge, MA, USA
phw@mit.edu

Abstract
A story summarizer benefits greatly from a reader model because a reader model enables the story
summarizer to focus on delivering useful knowledge in minimal time with minimal effort. Such a
summarizer can, in particular, eliminate disconnected story elements, deliver only story elements
connected to conceptual content, focus on particular concepts of interest, such as revenge, and
make use of our human tendency to see causal connection in adjacent sentences. Experiments
with a summarizer, built on the Genesis story understanding system, demonstrate considerable
compression of an 85-element précis of the plot of Shakespeare’s Macbeth, reducing it, for example,
to the 14 elements that make it a concise summary about Pyrrhic victory. Refocusing the
summarizer on regicide reduces the element count to 7, or 8% of the original.

1998 ACM Subject Classification I.2.0 General/Cognitive simulation

Keywords and phrases story telling and summarization, story understanding, cognitive modeling

Digital Object Identifier 10.4230/OASIcs.CMN.2015.157

1 Vision
Suppose you want a program to summarize a story. How should your program decide what
to include and what to leave out? I suggest that people read summaries mainly to acquire
useful knowledge in minimal time with minimal effort. Thus, a summary program should
focus on knowledge useful as precedent, exclude obvious inferences, but include reflective
inferences that help the reader understand how the key elements are connected. Accordingly,
a summary program should adhere to several principles reminiscent of the maxims of Grice
[5], and in so adhering, a summary program must have an understanding of human story
understanding in general and of the summary reader in particular. My students and I have
built such an understanding into our Genesis story-understanding system, and we can adjust
Genesis to model the knowledge and interests of particular summary readers.

2 Genesis models aspects of story understanding by humans

Much recent work has focused on applications that digest large amounts of data so as to
exhibit a kind of intelligence. Google’s caption generator [14], for example, is no doubt an
engineering marvel, but it sheds little or no light on our human visual faculty. Likewise,
IBM’s Watson [1] is no doubt intelligent in some ways, but it does not think as we think.
Work on Genesis goes in a different direction. Genesis was developed in the belief that
story understanding and telling is the distinguishing feature of human intelligence [15, 16, 17].
The aim in building Genesis is to model aspects of that story understanding and telling
feature at the expense of working with story summaries written in simple English of the kind
we can get through the START parser [6] and into Genesis’s inner language of relations and
events.
© Patrick Henry Winston;
licensed under Creative Commons License CC-BY
6th Workshop on Computational Models of Narrative (CMN’15).
Editors: Mark A. Finlayson, Ben Miller, Antonio Lieto, and Remi Ronfard; pp. 157–165
OpenAccess Series in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
158 Model-based Story Summary

One such simple Genesis-readable story is the following précis, which is based loosely
on Shakespeare’s play, Macbeth. It is itself a summary, but it is also an anvil on which to
hammer out principles that enable further compression and clarification.

Macbeth précis
Scotland and England are countries. Dunsinane is a castle and Birnam Wood is a forest.
Macbeth, Macduff, Malcolm, Donalbain, Lady Macbeth, Lady Macduff, Cawdor, and
Duncan are persons. Lady Macbeth is Macbeth’s wife. Lady Macduff is Macduff’s
wife. Lady Macbeth is evil and greedy. Duncan is the king, and Macbeth is Duncan’s
successor. Duncan is an enemy of Cawdor. Macbeth is brave. Macbeth defeats
Cawdor. Duncan becomes happy because Macbeth defeats Cawdor. The witches are
weird. The witches meet at night. The witches danced and chanted. Macbeth tells
witches to speak. Macbeth talks with the witches. Birnam Wood is a forest. Witches
predict that Birnam Wood will go to Dunsinane. The witches predict that Macbeth
will become Thane of Cawdor. The witches predict that Macbeth will become king.
The witches astonish Macbeth. Duncan executes Cawdor. Macbeth becomes Thane of
Cawdor. Duncan rewarded Macbeth because Duncan became happy. Lady Macbeth
wants Macbeth to become king. Macbeth is weak and vulnerable. Lady Macbeth
persuades Macbeth to want to become the king because Lady Macbeth is greedy.
Macbeth loves Lady Macbeth. Macbeth wants to please lady Macbeth. Macbeth
wants to become king because Lady Macbeth persuaded Macbeth to want to become
the king. Lady Macbeth plots to murder the king with Macbeth. Macbeth invites
Duncan to dinner. Duncan compliments Macbeth. Duncan goes to bed. Duncan’s
guards become drunk and sleep. In order to murder Duncan, Macbeth murders the
guards, Macbeth enters the king’s bedroom, and Macbeth stabs Duncan. Macbeth
becomes king. Malcolm and Donalbain become afraid. Malcolm and Donalbain flee.
Macbeth’s murdering Duncan leads to Macduff’s fleeing to England. In order to
flee to England, Macduff rides to the coast and Macduff sails on a ship. Macduff’s
fleeing to England leads to Macbeth’s murdering Lady Macduff. Macbeth hallucinates
at a dinner. Lady Macbeth says he hallucinates often. Everyone leaves because
Lady Macbeth tells everyone to leave. Macbeth’s murdering Duncan leads to Lady
Macbeth’s becoming distraught. Lady Macbeth has bad dreams. Lady Macbeth
thinks she has blood on her hands. Lady Macbeth tries to wash her hands. Lady
Macbeth kills herself. Birnam Wood goes to Dunsinane. Macduff’s army attacks
Dunsinane. Macduff curses Macbeth. Macbeth refuses to surrender. Macduff kills
Macbeth.

Given the Macbeth précis, Genesis notes and infers several kinds of causal connections.
Connections noted are those signaled by the word because, the words leads to, and the
words in order to in stories. Because signals a direct cause between story elements (Duncan
becomes happy because Macbeth defeated Cawdor); leads to indicates there is a chain
of unstated causes connecting two story elements (Macbeth’s murdering Duncan leads to
Macduff’s fleeing to England); in order to explains how something is done (In order to murder
Duncan, Macbeth murders the guards, Macbeth enters the king’s bedroom, and Macbeth
stabs Duncan).

2.1 Genesis deploys various kinds of common-sense rules

In addition to noting explicit causal connections, Genesis produces other causal connections
using inference rules, including deduction rules, abduction rules, explanation rules, and
 11:07:21 EDT 24-Apr-2015

Demonstrations Library Read Record Rerun About

| P Vie Con Start Exp Elabora Insp Sou ResSum Ret ||| Pop Views Controls Start viewer Experts Elaboration graph Inspector Sources Results Summary Retelling
Lady Macbeth is greedy. Macbeth murders Duncan. Macduff flees to England. Macbeth angers Macduff. Macduff kills Macbeth.

Random Unabridged Connected Concept centered Dominant concept centered Unresolved questions Table
Duncan is a king.

Macbeth is Duncan's

successor.

P. H. Winston 159
Lady Macduff is Macduff's wife.

Macbeth
Inspector Summary

||| Pop Views Controls Start viewer Experts Elaboration graph Inspector Sources Results Summary Retelling

Macbeth

Rules: 42 Lady Macbeth Macduff is Lady Macbeth Duncan Duncan Cawdor Duncan Duncan Lady Macbeth
persuades that
Macbeth Macbeth Macbeth Duncan Lady Macbeth Macbeth Macbeth
is Macbeth's Macduff's defeats becomes executes becomes harms rewards Macbeth wants
murders murders becomes becomes becomes becomes harms
wife. relation. Cawdor. happy. Cawdor. dead. Cawdor. Macbeth. guards. Duncan. king. dead. queen. happy. Duncan.
to become king.
Inferences: 54

Concepts: 16 Macbeth is Lady Lady Macduff Macbeth Macbeth Guards Macbeth

Macbeth's is Macduff's wants to enters become harms
husband. relation.
Discoveries: 11 become king. bedroom. dead. guards.

Explicit elements: 71
Macbeth is Lady Macbeth
Macbeth's stabs
Inferred elements: 31 relation. Duncan.

Total elements: 102

Lady Macbeth
Story reading time: 7.0 sec is Macbeth's
relation.

Total time elapsed: 8.2 sec

Lady Macduff
is Macduff's
wife.

Macduff is Lady
Macduff's
husband.

Lady
Macbeth is
greedy.

Duncan
is a king.

Macbeth is
Duncan's
successor.

Lady Macbeth
tells everyone
to the leave.

Macbeth Macbeth Macbeth Macduff Macbeth Everyone Lady Macbeth Lady Macbeth Macbeth Lady Macbeth Macduff
Macduff flees Lady Macbeth Lady Macbeth Macduff kills
murders Lady harms Lady harms becomes angers becomes harms becomes angers harms
to England.
Macduff. Macduff. Macduff. unhappy. Macduff. leaves. distraught.
kills herself. harms herself.
Macbeth. unhappy. Macbeth.
Macbeth.
Macbeth.

Macduff Macbeth Macduff

Lady Macduff Lady Macbeth
rides to becomes harms Lady
becomes dead. becomes dead.
coast. dead. Macbeth.

Macduff sails
on ship.

England is Dunsinane Birnam Duncan is Cawdor is Lady Malcolm is Macduff is Macbeth is Lady Duncan is Macbeth Witches
Scotland is a Lady Macduff Donalbain is
Wood is a Macbeth is a Macbeth is Cawdor's
a country. country. is a castle. a person. a person. is a person. a person. a person. a person. a person. is brave. are weird.
forest. person. evil. enemy.

Witches meet
Witches Witches Macbeth tells Macbeth Witches predict
that Birnam
Witches predict Witches predict Witches Macbeth Lady Macbeth Macbeth is Macbeth Macbeth Macbeth wants
Lady Macbeth
plots to murder
witches to the talks with that Macbeth that Macbeth astonish becomes wants Macbeth loves Lady to please Lady
at night. dance. chant. speak. witches.
Wood goes to
Dunsinane.
becomes thane. becomes king. Macbeth. thane. to become king. vulnerable. is weak. Macbeth. Macbeth.
king with
Macbeth.

Macbeth invites Duncan

Duncan goes
Guards Guards Malcolm Donalbain Malcolm Donalbain Macbeth Lady Macbeth
says that
Lady Macbeth Lady Macbeth Lady Macbeth Birnam Wood
Army attacks
Duncan to compliments become becomes becomes hallucinates has bad thinks she has tries washing goes to
dinner. Macbeth.
to bed.
drunk. sleep. afraid. afraid. flees. flees. at dinner.
Macbeth
dreams. blood on hands. hands. Dunsinane.
Dunsinane.
hallucinates.

Macduff Macbeth
curses refuses to
Macbeth. surrender.

Analysis

Revenge
Figure 1 Elaboration graph generated by the Macbeth précis. Connections are color coded:
Pyrrhic victory Pyrrhic victory Mistake because har... Mistake because har... Mistake because unh... Suicide Success Regicide Answered prayer Revenge

Elaboration graph
deduction rules and explicit because connections produce black lines; explicit leads to connections 100%

produce blue lines; explanation rules produce orange connections. You can expand the diagram if
you are using a PDF viewer.

presumption rules. Deduction rules, such as If x kills y, then y becomes dead, make connections
whenever all their antecedents are in a story. Abduction rules make connections between
elements and presumed antecedents. For example, Genesis’s reader model may include the
abduction rule If x kills y, then x must be insane. Explanation rules make connections only
when there is no other known way to explain an element. For example, Macduff kills Macbeth
is explained by the explanation rule If x angers y, then y may kill x and the previously
inferred element Macbeth angers Macduff. Presumption rules, like abduction rules, make
connections between elements and presumed antecedents, but only when there is no other
known way to explain an element. Presumption rules, unlike explanation rules, do not require
antecedents to be already in place. Abduction rules, explanation rules, and presumption
rules are ranked, so that the highest ranking rule dominates in the event multiple rules are
available for explaining an unexplained event. We intend to develop a more sophisticated,
context-sensitive process.
The noted and inferred causal connections constitute the elaboration graph of causally
connected elements as shown in Figure 1.

2.2 Genesis discovers concepts by searching for connections

Genesis finds concepts in the elaboration graph by searching for elements that instantiate
concept patterns. In general, concept patterns include specifications for sequences of causal
relations that start and end with particular, specified elements. The concept pattern for

CMN’15
 13:25:25 EST 01-Mar-2015

Demonstrations Library Read Record Run About

| P Vi Co Star Ex Elabor Ins So Re SumRet ||| Pop Views Controls Start viewer Experts Elaboration graph Inspector Sources Results Summary Retelling
Concept analysis Escalation analysis
Speech Predictions
160 Model-based Story Summary

Macbeth
murders Lady Macbeth Macbeth Macduff Macduff
harms angers kills harms
Macduff leads Macduff. Macduff. Macbeth. Macbeth.
to Macduff kills
Figure 2 The instantiated revenge concept pattern found in the Macbeth précis.
Macbeth.
I note that revenge, for example, is just a single such sequence described by x’s harming y leads to y’s
harming x. An instantiated revenge pattern is shown in Figure 2.
murder and Remarkably, the elaboration graph, augmented by discovered concept patterns, provides
Results Inspector
the substrate for developing models of many kinds of story understanding and telling,
||| Pop Views Controls including
Start viewer Experts question answering,
Elaboration graph Inspector cultural bias in
Sources Results interpretation,
Summary Retelling instructional telling with a
learner model, persuasive telling with a listener model, precedent-based prediction, and as
described here, summary. Macbeth/revenge
Macbeth is Lady Macbeth's Cawdor Macbeth Duncan
Lady Macbeth is Macbeth's wife. Macbeth defeats Cawdor. Duncan becomes happy. Duncan executes Cawdor. Lady Macbeth becomes queen. Duncan rewards Macbeth. Macbeth murders guards. Macbeth murders Duncan. Guards become dead. Macbeth becomes happy. Macbeth harms Duncan.
relation.
becomes dead. becomes thane. becomes dead.

Rules: 39 Macbeth is Lady Macbeth's

husband.
Lady Macbeth is Macbeth's

relation.
Duncan harms Cawdor. Macbeth wants to become king. Macbeth enters bedroom.
Macbeth
becomes king.

Inferences: 49 2.3 We provide common-sense rules and concept patterns in English

Lady Macduff is Macduff's wife. Macduff is Lady Macduff's relation. Macbeth stabs Duncan. Macbeth harms guards.

Macduff is Lady Macduff's husband. Lady Macduff is Macduff's relation.

Concepts: 15
My students and I provide Genesis with common-sense rules, concept patterns, and stories;
Lady Macbeth is greedy.

Discoveries: 11 Duncan is a

all rules, patterns, and stories are provided in English as indicated in the examples. Our
king.

Explicit elements: 70
Macbeth is Duncan's successor.

purpose is to establish, by telling, what Genesis needs to know to exhibit a kind of humanlike
Inferred elements: 33
understanding.
Total elements: 103 Everyone Macduff kills

We think it reasonable, at this stage, to tell Genesis what it needs to know. One reason is
Macduff flees to England. Macbeth murders Lady Macduff. Macbeth harms Lady Macduff. Macbeth harms Macduff. Macduff becomes unhappy. Macbeth angers Macduff. Lady Macbeth becomes distraught. Lady Macbeth kills herself. Lady Macbeth harms herself. Lady Macbeth harms Macbeth. Macbeth becomes unhappy. Lady Macbeth angers Macbeth. Macduff harms Macbeth.
leaves. Macbeth.

Lady Macduff Macbeth

Macduff rides to coast. Lady Macbeth becomes dead. Macduff harms Lady Macbeth.

Story reading time: 6.4 sec

becomes dead. becomes dead.

that much of what we know we learn by being told. Few would have the concept of Pyrrhic
Macduff sails
on ship.

Total time elapsed: 9.0 sec

victory, for example, without being told. Another reason is that much of what we tell Genesis
I am I am Thane is England is a Duncan is a Cawdor is a Macduff is a Macbeth is a Malcolm is a Macbeth is a Macduff is a

in experimenting with one story finds use in other stories. Revenge, for example, is revenge
Dunsinane is a castle. Birnam Wood is a forest. Lady Macduff is a person. Lady Macbeth is a person. Donalbain is a person.
eastern. machiavellian. noble. country. person. person. person. person. person. thane. thane.

Witches Witches have Macbeth

Lady Macbeth is evil. Duncan is Cawdor's enemy. Macbeth talks with witches. Witches astonish Macbeth. Macbeth loves Lady Macbeth. Macbeth plans to murder king. Lady Macbeth plans to murder king. Macbeth invites Duncan to dinner. Duncan complements Macbeth.
dance. visions. becomes king.

not only in Macbeth, but also in fairy tales and international conflicts. Yet another reason is
Duncan goes to bed. Guards become drunk.
Guards
sleep.
Malcolm
becomes afraid.
Donalbain
becomes afraid.
Malcolm
flees.
Donalbain
flees.
Macbeth hallucinates at dinner. Lady Macbeth has bad dreams. Birnam Wood goes to Dunsinane. Birham wood is a forest. Burnham wood goes to Dunsinane.
Army attacks
Dunsinane.

that we have done research on learning concept patterns from ensembles of stories [2, 7], and
Macduff curses Macbeth. Macbeth refuses surrendering.

Analysis we are engaged in research on learning common sense by mining various textual sources.
Revenge Revenge Pyrrhic vi... Pyrrhic vi... Mistake ... Mistake ... Mistake ... Suicide Success Regicide Answere...

3 The Genesis model enables principle-based story summary 100%

Elaboration graph

Genesis, as a model of story understanding by humans, suggests several principles for

summary. Some compress the story provided; others expand the story by adding helpful
explanations. All work toward helping the reader to focus on the elements that convey useful
knowledge and to grasp how the useful story elements are connected.
In the following, I articulate several such principles, and I explain how those principles
are reflected in a model of story summarization by humans. I also show how the Genesis
story summarizer, based on that model, performs on a test case.

3.1 The principle of connection

Good precedents exhibit causal connections between events that are likely to be seen again in
future situations, thereby enabling understanding, prediction, and control. Accordingly, the
Genesis story summarizer preserves those explicit story elements that are involved in causal
connections, where the causal connections are either explicit or inferred. Genesis filters out
explicit story elements that are neither an antecedent nor a consequent in any kind of causal
connection.
P. H. Winston 161

The Macbeth précis contains 55 sentences, which, when understood by Genesis, expand
to 85 explicit story elements, with the expansion caused by separately counting elements
that are embedded in compound sentences and explicit causal connections and by adding
one to the element count for each explicit causal connection. In what follows, I compare the
number of summary elements with the number of explicit story elements for various versions
of the Genesis summarizer.
Many of the explicit elements are not involved in causal connections of any kind, explicit
or inferred, and thus offer little or nothing by way of constraining precedent. Keeping only
those explicit elements that are causal connections and explicit elements that are embedded
in Genesis’s inferred causal connections produces the following summary in which the START
system produces the English, with occasional awkwardness, from Genesis’s inner language of
relations and events:

Macbeth, with principle of connection

Lady Macbeth is Macbeth’s wife. Lady Macduff is Macduff’s wife. Duncan is a
king. Macbeth is Duncan’s successor. Duncan becomes happy because Macbeth
defeats Cawdor. Duncan executes Cawdor. Duncan rewards Macbeth because Duncan
becomes happy. Lady Macbeth persuades that Macbeth wants to become king because
Lady Macbeth is greedy. Macbeth wants to become king because Lady Macbeth
persuades that Macbeth wants to become king. In order to murder Duncan, Macbeth
murders guards; in order to murder Duncan, he enters bedroom; in order to murder
Duncan, he stabs Duncan. Donalbain is Duncan’s son. Malcolm is Duncan’s son.
For Macbeth to murder Duncan leads to Macduff’s fleeing to England. In order to
flee to England, Macduff rides to coast; in order to flee to it, he sails on ship. For
Macduff to flee to England leads to Macbeth’s murdering Lady Macduff. Everyone
leaves because Lady Macbeth tells everyone to the leave. For Macbeth to murder
Duncan leads to Lady Macbeth’s becoming distraught. Lady Macbeth kills herself.
Macduff kills Macbeth.

Thus, the principle of connection allows the Genesis summarizer to reduce the number of
summary elements to 34, 40% of the 85 explicit story elements.

3.2 The principle of concept focus

Good precedents tend to be told in a manner that focuses attention on conceptual content
because associating a story with its conceptual content is part of what separates novices
from domain experts [3, 4]. Accordingly, another version of the Genesis story summarizer
includes only explicit elements that lead eventually—via a chain of inferred connections—to
an element lying in an instantiated concept pattern.
The elaboration graph plays a central role in this kind of summary because searches
in the elaboration graph discover concepts and because searches in the elaboration graph
determine which explicit elements are connected to those concepts. Filtering out other
elements produces the following Macbeth summary:

Macbeth, with principle of concept focus added

The story is about Regicide, Mistake because unhappy, Answered prayer, Revenge,
Suicide, Mistake because harmed, Success, and Pyrrhic victory. Lady Macbeth is
Macbeth’s wife. Lady Macduff is Macduff’s wife. Lady Macbeth persuades that
Macbeth wants to become king because Lady Macbeth is greedy. Macbeth wants to
become king because Lady Macbeth persuades that Macbeth wants to become king.

CMN’15
162 Model-based Story Summary

In order to murder Duncan, Macbeth murders guards; in order to murder Duncan,

he enters bedroom; in order to murder Duncan, he stabs Duncan. Macbeth murders
Duncan, probably because Macbeth wants to become king, Duncan is a king, and
Macbeth is Duncan’s successor. For Macbeth to murder Duncan leads to Macduff’s
fleeing to England. In order to flee to England, Macduff rides to coast; in order to flee
to it, he sails on ship. For Macduff to flee to England leads to Macbeth’s murdering
Lady Macduff. For Macbeth to murder Duncan leads to Lady Macbeth’s becoming
distraught. Lady Macbeth kills herself, probably because Lady Macbeth becomes
distraught. Macbeth becomes unhappy. Macduff kills Macbeth, probably because
Macbeth angers Macduff.

Now the summary contains only 30 of the 85 explicit story elements or 35%. Excluded are
elements such as Duncan becomes happy because Macbeth succeeded, and Duncan rewarded
Macbeth because Duncan becomes happy. None of the elements involved leads to an element
in an instantiated concept.

3.3 The principle of dominant concept focus

Good precedents tend to have a particular purpose and focus attention on one or a few key
concepts. Accordingly, yet another version of the Genesis story understander retains an
explicit story element only if that element is connected via a chain of inferences to a key
concept.
Which of the discovered concepts are the key concepts? There are several reasonable
possibilities with which we propose to experiment once we have a large enough corpus of
Genesis-readable stories, including concepts that cover a lot of the elements of the story over
a long time span, concepts that involve violent acts, such as murder, concepts that excite big
emotional reaction, concepts that indicate a dramatic situation, such as those identified by
Polti, concepts that the summarizer wants the reader to note, concepts that the summarizer
knows the reader wants to note, concepts that are rarely observed, and concepts that involve
memorable elements.
For example, in the Macbeth précis, Pyrrhic victory dominates all other concepts in the
sense that it incorporates the most story elements. Using Pyrrhic victory to summarize,
rather than all concepts, Genesis produces the following:

Macbeth, with principle of dominant concept focus added

The story is about Pyrrhic victory. Lady Macbeth is Macbeth’s wife. Lady Macduff is
Macduff’s wife. Lady Macbeth persuades that Macbeth wants to become king because
Lady Macbeth is greedy. Macbeth wants to become king because Lady Macbeth
persuades that Macbeth wants to become king. In order to murder Duncan, Macbeth
murders guards; in order to murder Duncan, he enters bedroom; in order to murder
Duncan, he stabs Duncan. Macbeth murders Duncan, probably because Macbeth
wants to become king, Duncan is a king, and Macbeth is Duncan’s successor. For
Macbeth to murder Duncan leads to Macduff’s fleeing to England. In order to flee to
England, Macduff rides to coast; in order to flee to it, he sails on ship. For Macduff to
flee to England leads to Macbeth’s murdering Lady Macduff. Macduff kills Macbeth,
probably because Macbeth angers Macduff.

The elements that deal with Lady Macbeth’s suicide drop out; the number of summary
elements is 25, 29% of the explicit story elements.
P. H. Winston 163

Memorable elements, incidentally, are readily captured in simple concept patterns that
may involve no leads to elements, such as this Memorable event pattern: a woman becomes
the bishop. Of course, what constitutes a memorable event may not be so memorable at a
different time or place.

3.4 The principle of interpretation transparency

Good summaries do not require readers to guess how the summarizer has reasoned. Ac-
cordingly, the Genesis story summarizer is explicit about the assumptions it makes. In
particular, the Genesis story summarizer includes not only the consequents of explanation
rules, which are explicit in the story, but also the fully instantiated explanation rule, even
though the antecedents themselves may be the consequents of deduction rules and not
ordinarily included.
For example, the previous two summaries include Macduff kills Macbeth, probably because
Macbeth angers Macduff. The rationale is that the summarizer, in eagerness to create a
more coherent and easily understood story, has added something not completely obvious
about how the summarizer has interpreted the story. Thus the summarizer’s reasoning is
transparent and the reader is relieved of reasoning effort.

3.5 Compression by eliminating details of how actions are performed

Good summaries stick to essentials. Accordingly, the Genesis story summarizer can be
directed to eliminate details of how actions are performed, providing further compression.
Impatient readers will not care, for example, about exactly how Macbeth murders Duncan,
so the Genesis story summarizer suppresses details about the guards, the bedroom, and
stabbing:

Macbeth, with detail suppression added

The story is about Pyrrhic victory. Lady Macbeth is Macbeth’s wife. Lady Macduff is
Macduff’s wife. Lady Macbeth persuades that Macbeth wants to become king because
Lady Macbeth is greedy. Macbeth wants to become king because Lady Macbeth
persuades that Macbeth wants to become king. Macbeth murders Duncan, probably
because Macbeth wants to become king, Duncan is a king, and Macbeth is Duncan’s
successor. For Macbeth to murder Duncan leads to Macduff’s fleeing to England. For
Macduff to flee to England leads to Macbeth’s murdering Lady Macduff. Macduff
kills Macbeth, probably because Macbeth angers Macduff.

With means deleted, the number of summary elements is further reduced to 18, 21% of
the explicit story elements.

3.6 Compression using the post hoc ergo propter hoc assumption
Good summaries refrain from making natural inferences explicit because making them explicit
is unnatural and annoying. Accordingly, the Genesis story summarizer supposes the reader
will instinctively find plausible causal connections between adjacent events.
After this does not mean because of this in logic, but we use it nevertheless in telling
stories smoothly, dropping explicit cause when proximity makes the cause apparent:

Macbeth, with post hoc ergo propter hoc processing added

The story is about Pyrrhic victory. Lady Macbeth is Macbeth’s wife. Lady Macduff is
Macduff’s wife. Lady Macbeth persuades that Macbeth wants to become king because

CMN’15
164 Model-based Story Summary

Lady Macbeth is greedy. Macbeth wants to become king. Macbeth murders Duncan,
probably because Duncan is a king, and Macbeth is Duncan’s successor. Macduff
flees to England. Macbeth murders Lady Macduff. Macduff kills Macbeth, probably
because Macbeth angers Macduff. Macduff.

Processing with post hoc ergo propter hoc transforms Macduff’s fleeing to England leads
to Macbeth murders Lady Macduff to Macbeth murders Lady Macduff. With post hoc ergo
propter hoc in play, the number of summary elements is 15, 18% of the explicit story elements.

4 Experiments

Using Genesis to summarize Shakespearian play summaries and cyberwar summaries produced
the following percentages of summary elements relative to total elements. The Connected
column reports the fraction of the explicit story elements that are reported when reporting
all and only the elements in the story that are causally connected; the All-methods column
reports the fraction of the explicit story elements reported when all of the principles here
described are engaged.

Connected All methods

Macbeth 40% 18%
Hamlet 41% 14%
Estonia vs. Russia 40% 60%
Georgia vs. Russia 26% 19%

The compression numbers are not dramatic because the test stories are already summaries.
The numbers generally drop when limiting the summary to elements that lead eventually
to one or more instantiated concept patterns. One exception is Estonia vs. Russia. In this
summary, one concept pattern is Aggression of a bully, a concept pattern that looks for
which side the reader is friendly with: x is my friend. x’s angering y leads to y’s harming x.
Instantiating that concept pattern brings in I am Estonia’s friend, a disconnected element,
but an element that corresponds to an element in the concept pattern. If the reader happens
to be Russia’s friend, the concept pattern triggered is Teaching a lesson and I am Russia’s
friend is included.

5 Contributions

Work on the Genesis story understanding and telling system has been inspired, in part, by
the pioneering work of Roger Shank and his students [8, 9, 10, 11]. Work on Genesis has
also been inspired, in part, by paleoanthropologist Ian Tattersall’s reflections on what makes
us human [12, 13], which led me to the conclusion that story understanding and story telling
plays a major role. I have focused here on principles of story summary and shown how those
principles are reflected the Genesis story summarizer. In particular, I have:
Argued that a reader model is a necessary foundation for good story summary
Identified the principles of connection, concept focus, dominant concept focus, and
interpretation transparency.
Suggested means compression and introduced post hoc ergo propter hoc processing.
Exhibited an implemented, principle-based summarizer at work on a representative story
from the Genesis library, a précis of Macbeth, showing a compression of 84%.
P. H. Winston 165

References
1 David Ferrucci, Eric Brown, Jennifer Chu-Carroll, James Fan, David Gondek, Aditya
Kalyanpur, Adam Lally, J. William Murdock, Eric Nyberg, John Prager, Nico Schlaefer,
and Chris Welty. The AI behind watson—the technical article. AI Magazine, 2010.
2 Mark A. Finlayson. Learning Narrative Structure from Annotated Folktales. PhD thesis,
Electrical Engineering and Computer Science Department, MIT, Cambridge, MA, 2012.
3 Mark A. Finlayson and Patrick Henry Winston. Intermediate features and informational-
level constraint on analogical retrieval. In Proceedings of the 27th Annual Meeting of the
Cognitive Science Society, pages 666–671, 2005.
4 Dedre Gentner and Arthur B Markman. Structure mapping in analogy and similarity.
American Psychologist, 52(1):45–56, 1997.
5 H. Paul Grice. Studies in the Way of Words. Harvard University Press, Cambridge, MA,
1989.
6 Boris Katz, Gary Borchardt, and Sue Felshin. Syntactic and semantic decomposition strate-
gies for question answering from multiple resources. In Proceedings of the AAAI 2005
Workshop on Inference for Textual Question Answering, 2005.
7 Caryn Krakauer and Patrick Henry Winston. Story retrieval and comparison using concept
patterns. In Mark Alan Finlayson, Pablo Gervas, Deniz Yuret, and Floris Bex, editors, Pro-
ceedings of the 3rd Workshop on Computational Models of Narrative (CMN’12), volume 3,
pages 119–124. European Language Resources Association (ELRA), 2012.
8 Wendy Lehnert. Plot units and narrative summarization. Cognitive Science, 5(4):293–331,
1981.
9 Roger C. Schank. Conceptual dependency: A theory of natural language understanding.
Cognitive Psychology, 3(4):552–631, 1972.
10 Roger C. Schank and Robert P. Abelson. Scripts, Plans, Goals, and Understanding: An
Inquiry into Human Knowledge Structures. Lawrence Erlbaum Associates, Mahwah, NJ,
1977.
11 Roger C. Schank and Christopher K. Riesbeck. Inside Computer Understanding: Five
Programs Plus Miniatures. Lawrence Erlbaum Associates, Mahwah, NJ, 1981.
12 Ian Tattersall. Becoming Human. Harcourt, Orlando, FL, 1998.
13 Ian Tattersall. Human evolution and cognition. Theory in Biosciences, 123(4):371–379,
2010.
14 Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neu-
ral image caption generator. Cornell University Library digital repository, arXiv:1411.4555
[cs.CV], 2014.
15 Patrick Henry Winston. The strong story hypothesis and the directed perception hypothesis.
In Pat Langley, editor, Technical Report FS-11-01, Papers from the AAAI Fall Symposium,
pages 345–352, Menlo Park, CA, 2011. AAAI Press.
16 Patrick Henry Winston. The next 50 years: a personal view. Biologically Inspired Cognitive
Architectures, 1:92–99, 2012.
17 Patrick Henry Winston. The right way. Advances in Cognitive Systems, 1:23–36, 2012.

CMN’15