HOMER: A Creative Story Generation System Dimitrios N' Konstantinou Supervisor: Prof' Paul Mc Kevitt

1
HOMERA Creative Story Generation
SystemDimitrios N. KonstantinouSupervisor
Prof. Paul Mc KevittSchool of Computing
Intelligent SystemsFaculty of EngineeringUnivers
ity of Ulster at Magee
2
Objectives of HOMER

• To build a StoryTelling agent that generates
• style-constrained stories
• stories with a point-of-view
• natural language output
• domain-independent stories

3
Related Work

• TALESPIN a StoryTelling agent
• (Meehan, 1976)
• MINSTREL a creative StoryTeller
• (Turner, 1990)
• JOSEPH a rational StoryTelling agent
• (Lang, 1997)
• STORYBOOK an NLG-based system
• (Callaway Lester, 2001)

4
Architecture of HOMER
Story Outliner
Natural Language Generator
user input
schema output
story output
5
HOMER Modules

• Story Outliner
• Case-Based Reasoning Paradigm
• Input via User Interface
• Case Libraries
• Implements Adaptation Strategies
• Natural Language Generator (NLG)
• Generates sentences
• Composes sentences into story discourse

6
User Interface types of input

• Matching Knowledge
• Type of information requested of user categorial
• Aim to classify, infer and retrieve
• End-result instantiation of kernel story
  fragment
• Idiosyncratic Knowledge
• Type of information entered by user instantial
• Aim to adapt retrieved story fragment and
  further
• instantiate kernel story fragment

7
Case Libraries types of cases

• Archetypical story segments
• Human-authored story segments (e.g. HAMLET ghost
  scene)
• Mode-based segmentation
• Default values
• Instantiated story outlines
• Story outlines output by Adaptation Strategies
• Stereotypical scripts (à la Schank)
• Scripts based on common-sense reasoning

8
Case Library Organisation

• Types of Modes based on Narrative Reasoning
• Romance mode
• Ironic mode
• Tragic mode
• Comedy mode

9
Case Segmentation

• Criteria for Segmentation
• Semantic
• Mode-based
• Example
• Tragic-mode segment
• High-Status - Flaw - Novel-Experience
• Downfall - New-State Demonic -Vision

10
Configuration of Segments
CRITERIA semantics- discourse-based Example
rhetorical-relation
additive rhetorical-relation antithesis
rhetorical-relation causality

HIGH-STATUS SEGMENT
FLAW SEGMENT
DOWNFALL SEGMENT
11
Internal Case ConfigurationOrganizational
Principle Semantic Representation gt Domain
Knowledge Representation

• Example-Case Hamlet
• CIRCUMSTANCES ..
• SEMANTIC
• action material
• actor Hamlet
• actor-type human
• actee1 ghost
• actee1Type superhuman

• Segment Initiation
• NARRATIVE
• action-role sub-event
• action-category encounter
• action-instance meet
• actor-role protagonist
• actee1-role catalyst

12
Adaptation Strategies

• Type-1 Contextual Adaptation
• Type-2 Intra-modal Adaptation
• Type-3 Cross-modal Adaptation
• Type-4 Point-of-view Adaptation
• Type-4 Point-of-view Adaptation

13
Contextual Adaptation
Example Trigger attribute gt
era CIRCUMSTANCES era medieval gt
era modern location castle
location mansion weapon spear
weapon gun vehicle horse
vehicle car
14

• Type-2
• segments from different story-cases of SAME
  mode
• TRIGGER rhetorical relations
• Type-3
• segments from different cases of DIFFERENT
  mode
• TRIGGER Semantic Similarities
• Type-4
• based on Semantics-Narrative Knowledge mapping
• TRIGGER Narrative roles

15
Comparison with other StoryTelling Systems

• MINSTREL (Turner, 1990)
• Commonalities
• CBR paradigm
• Some Adaptation
  strategies
• Differences
• Strictly
  domain-dependent
• Limited Semantic
  Representation
• No discourse
  strategies

16
STORYBOOK (Callaway Lester, 2001)

• Commonalities
• Natural Language
  Module
• Grammar Formalism
• Differences
• Lacks Story
  Creativity
• Ad-hoc discourse
  strategies
• Strictly
  domain-dependent

17
Natural Language Generation Module

• Work in tandem with the construction of Story
  Outliner
• Use of Functional Unification Formalism
• - not difficult to extend with discourse
  facilities
• - can be used as media-allocator in
  multi-modal story-systems

18
Current State of Implementation

• Programming Language Allegro Common Lisp
• Knowledge elicitation 30 narratives
• Number of story segments 15
• User Interface menu-based and inferencing
• Adaptation Strategies 4 types

19
(No Transcript)
20
the story-kernel so far is S(KERNEL1
SEGMENT_TYPE HIGH-STATUS ERA ANCIENT LOCATION
EUROPE PLACE .PALACE WEAPON SWORD VEHICLE
CHARIOT) the retrieved story-segment
(S(STORY CIRCUM S(CIR ERA ANCIENT TIME NIL
PLACE .PALACE LOCATION EUROPE WEAPON SWORD
VEHICLE CHARIOT POLARITY POSITIVE RELATEE
PROTAGONIST RELTYPE HUMAN RELNROLE NIL
EVENTUALITY CREATE) SEMAN S(SEM
PROC_ACTION ATTRIBUTIVE ACTOR OEDIPUS
ACTORTYPE HUMAN ACTEE1 KING ACTEE1TYPE
ATTRIBUTE ACTEE2 NIL ACTEE2TYPE NIL)
NARRAT S(NAR ACTNROLE 3 ACTORNROLE
PROTAGONIST ACTORNTYPE HERO ACTOR1NROLE NIL
ACTEE1NROLE STATUS ACTEE1NTYPE NIL
ACTEE2NROLE NIL ACTEE2NTYPE NIL
PROC_CATEGORY CHANGE CAT_INSTANCE BECOME)))
21
Conclusion Future Work

• HOMER is a creative story generation system
• Uses CBR paradigm to model creativity
• Extension and fine-tuning of Adaptation
  strategies
• Development of Schema-based output to NLG module
• Introduction of Discourse strategies in NLG
  module