N4S Research - Story Comparison Adam Amos-Binks 1 1 Department of - PowerPoint PPT Presentation

N4S Research - Story Comparison Adam Amos-Binks 1 1 Department of Computer Science North Carolina State University LAS Presentation, Sept 2, 2015

Outline Motivation 1 Background 2 Contribution 3 Validation 4 Summary and Next Steps 5 References 6

Story Plan Comparison Plan based representation of a story Given a generative model of story plans, what sense can it generate? Compare two story plans − → quantifying a story solution space

Story Plan Comparison Plan based representation of a story Given a generative model of story plans, what sense can it generate? Compare two story plans − → quantifying a story solution space

Story Plan Comparison Plan based representation of a story Given a generative model of story plans, what sense can it generate? Compare two story plans − → quantifying a story solution space

Narrative for Sensemaking (N4S)

Narratology: Narrative Properties Narrative - cognitive tool[5] has a c ausal structure is a problem solving strategy leads to enhanced cognition Narrative - structure[3] Story Discourse

Narratology: Narrative Properties Narrative - cognitive tool[5] has a c ausal structure is a problem solving strategy leads to enhanced cognition Narrative - structure[3] Story Discourse

Cognitive Psych.: Causal Networks Figure : Causal network of a narrative[10]

Cognitive Psych.: Causal Networks Figure : Narrative events with high recall [10]

Cognitive Psych.: Intention What is Intention?[1] Intention related to other psych. states: Beliefs and desires Theory of Intention: address problem of Intending to act ◮ Future looking: commit now, act later Definition of Intention: commensense version tied to plans and planning Humans are rational agents ◮ Capacity 1: form and execute plans ◮ Capacity 2: act purposively ◮ partial plans

Rise of....

AI: Planning Overview Intelligent agent action execution World represented in propositional logic Sound and complete planning algorithms: ◮ input: domain and problem ◮ output: solution plans Figure : Traditional application of planning

AI: Planning Domain Predicates Action Schemata ◮ Preconditions ◮ Parameters ◮ Effects Figure : Dock domain action schemata

AI: Planning Problem Objects/literals Initial state Goal state Figure : Problem Init Figure : Problem Goal

AI: Planning Solution Plan Generated by a planning algorithm: partial-order, state-space Has no open pre-conditions or threatened causal links Is part of a solution space (can be huge ) Figure : Dock Solution Plan P

AI: Story Planning History Long standing goal of AI to produce coherent narrative Bi-partite model of computational narrative[14] � � POCL story plans, P : S , B , O , L ◮ Enables possible story world chronology ◮ Explicitly model action, temporality and causality ◮ Contains a beginning & end ◮ Requires a story domain , problem , solutions Figure : Story planning and ROI

AI: Story Domain Figure : Space domain schemata sample ( : types c r e a t u r e landform s h i p − p l a c e ) ( : p r e d i c a t e s ( a l i v e ? c r e a t u r e − c r e a t u r e ) ( stunned ? c r e a t u r e − c r e a t u r e ) ( h a b i t a b l e ? p l a c e − p l a c e ) ( s a f e ? p l a c e − p l a c e ) ( s a f e ? var ) ( s a f e ? c r e a t u r e − c r e a t u r e ) ( e r u p t i n g ? landform − p l a c e ) ( at ? c r e a t u r e − c r e a t u r e ? p l a c e − p l a c e ) ( f i g h t i n g ? c r e a t u r e 1 − c r e a t u r e ? c r e a t u r e 2 − c r e a t u r e ) ( f r i e n d s ? c r e a t u r e 1 − c r e a t u r e ? c r e a t u r e 2 − c r e a t u r e ) ( c a p t a i n ? c r e a t u r e − c r e a t u r e ? s h i p − s h i p ) ( guardian ? c r e a t u r e − c r e a t u r e ? p l a c e − p l a c e )) Figure : Space domain predicates

AI: Story Problem ( d e f i n e ( problem e x p l o r e ) ( : o b j e c t s zoe − c r e a t u r e l i z a r d − c r e a t u r e s h i p − s h i p cave − p l ace s u r f a c e − landform ) ( : i n i t ( h a b i t a b l e s h i p )( h a b i t a b l e cave )( h a b i t a b l e s u r f a c e ) ( s a f e s h i p )( s a f e s u r f a c e )( s a f e cave )( s a f e zoe )( s a f e l i z a ( a l i v e zoe )( a l i v e l i z a r d ) ( at zoe s h i p )( at l i z a r d cave ) ( captain zoe s h i p ) ( guardian l i z a r d s u r f a c e ) ( i n t e n d s zoe ( f r i e n d s zoe l i z a r d )) ( i n t e n d s zoe ( s a f e zoe )) ( i n t e n d s zoe ( a l i v e zoe )) ( i n t e n d s l i z a r d ( s a f e l i z a r d )) ( i n t e n d s l i z a r d ( a l i v e l i z a r d ) ) ) ( : goal ( not ( h a b i t a b l e s u r f a c e ) ) ) ) Figure : Space Exploration Problem

AI: Story Plan Solution Semantic Properties Character intentions (IPOCL)[8] ◮ POCL story plan P : � � S , B , O , L , I ◮ Intention Frame: � � c , g , m , σ , T ◮ � � zoe , safe , begin − erupt , teleport , { 2 , 3 } Conflict (CPOCL)[11] ◮ non-executed steps represent foiled plans ◮ � � zoe , safe , begin − erupt , teleport , { 2 , 3 } in conflict with � � zoe , friends , init , teleport , { 0 , 4 , 5 } Figure : Space Exploration Solution

AI: “Everyone has plan.......”

Approach: Story Plan Comparison Plan Comparison Domain independent vs specific[2] Set theoretic - Jaccard similarity[9] ◮ δ A ( p , p ′ ) = 1 − | S ( p ) ∩ S ( p ′ ) | | S ( p ) ∪ S ( p ′ ) | ◮ δ C ( p , p ′ ) = 1 − | L ( p ) ∩ L ( p ′ ) | | L ( p ) ∪ L ( p ′ ) | Figure : Jaccard similarity

Approach: Story Plan Comparison Story Comparison Causal structure [4] Know exemplars apriori, TTD-MDP[6] Levenshtein edit distance[7] Story Plan Comparison Story vs plan comparison Domain specific distance metric (all stories) Need to summarize story semantics!

Approach: Story Plan Comparison Story Comparison Causal structure [4] Know exemplars apriori, TTD-MDP[6] Levenshtein edit distance[7] Story Plan Comparison Story vs plan comparison Domain specific distance metric (all stories) Need to summarize story semantics!

Approach: Story Plan Summary Important Steps in Story Plans[13] Identify causal chain Identify highest causal degree: preconditions + effects (used) E ( P ) Figure : Important events of solution plan P

Approach: Story Plan Summary Intention Frame Summary � � c , g , m , σ ◮ � � � � zoe , friends , init , teleport , { 0 , 4 , 5 } − → zoe , friends , init , teleport ◮ � � zoe , safe , begin − erupt , teleport , { 2 , 3 } − → � � zoe , safe , begin − erupt , teleport J ( P ) Story Plan Summary ψ ISIF ( P ) � � = E , J (1)

Approach: Story Plan Summary Intention Frame Summary � � c , g , m , σ ◮ � � � � zoe , friends , init , teleport , { 0 , 4 , 5 } − → zoe , friends , init , teleport ◮ � � zoe , safe , begin − erupt , teleport , { 2 , 3 } − → � � zoe , safe , begin − erupt , teleport J ( P ) Story Plan Summary ψ ISIF ( P ) � � = E , J (1)

Approach: Story Plan Distance Metric ISIF Distance Metric � E ( ψ ISIF � ) ∩ E ( ψ ISIF � � � J ( ψ ISIF ) ∩ J ( ψ ISIF � ) � + ) � δ ISIF ( ψ ISIF , ψ ISIF 1 2 1 2 ) = 1 − (2) 1 2 � E ( ψ ISIF ) ∪ E ( ψ ISIF � J ( ψ ISIF ) ∪ J ( ψ ISIF � � � � ) � + ) � 1 2 1 2

Preliminary Results: Solution Space Diversity Figure : Glaive[12] − 10 000 POCL solution plans, Π

Preliminary Results: Solution Space Properties (a) Distinct steps (b) Distinct causal links (c) Distinct important (d) Distinct intention steps frames Figure : Distinct plan elements

Preliminary Results: Comparing two plans (a) Plan π 1 (b) Plan π 2 Figure : Example story plan structures

Preliminary Results: Comparing two plans π i S ( π i ) L ( π i ) I ( π i ) E ( π i ) π 1 8 (6 executed, 2 non-executed) 35 2 2 π 2 8 (5 executed, 3 non-executed) 34 1 1 Table : Story plan properties π i , π k δ A ( π i , π k ) δ C ( π i , π k ) δ ISIF ( π i , π k ) π 1 , π 2 0.11 0.03 0.25 Table : Distance metric results

Summary Story plan representation Story plan summary ISIF story plan distance metric Figure : generative model vs classification model

Research Next Steps More extensive evaluation Alternate story plan distance metrics Quantifying solution space diversity Figure : Story planning and ROI

Questions and DO 6 Discussion Bardic integration? Author cyber domain? Figure : Cyber attack lifecycle Figure : Anticipatory event metric

References I Michael E Bratman. Intentions in Communication. In Philip R. Cohen, Jerry Morgan, and Martha E. Pollack, editors, Intentions in Communication , chapter What Is In, pages 15–30. The MIT Press, 2003. Alexandra Coman and Héctor Muñoz avila. Generating Diverse Plans Using Quantitative and Qualitative Plan Distance Metrics. In AAAI Conference on Artificial Intelligence , volume 18015, pages 946–951, 2011. Tom Conley and Seymour Chatman. Story and Discourse , volume 8. 1979.