ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying...

30
ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Ying Feng Robert Goldstone Robert Goldstone Vladimir Menkov Vladimir Menkov Computer Science Department Computer Science Department Psychology Department Psychology Department Indiana University Indiana University

Transcript of ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying...

Page 1: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation

Ying FengYing FengRobert GoldstoneRobert GoldstoneVladimir MenkovVladimir Menkov

Computer Science DepartmentComputer Science DepartmentPsychology DepartmentPsychology Department

Indiana UniversityIndiana University

Page 2: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

How concepts get their meaningsHow concepts get their meanings

Conceptual webConceptual web– A concept’s meaning comes from its A concept’s meaning comes from its

connections to other concepts in the connections to other concepts in the same conceptual systemsame conceptual system

External groundingExternal grounding– A concept’s meaning comes from its A concept’s meaning comes from its

connection to the external worldconnection to the external world

Page 3: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Conceptual WebConceptual Web

PhilosophyPhilosophy– Conceptual role semanticsConceptual role semantics– Conceptual incommensurabilityConceptual incommensurability

PsychologyPsychology– Isolated and interrelated conceptsIsolated and interrelated concepts– Latent semantic analysisLatent semantic analysis

Computer ScienceComputer Science– Semantic networksSemantic networks– Intrinsic meaning in large databasesIntrinsic meaning in large databases

Page 4: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Externally Grounded ConceptsExternally Grounded Concepts

PhilosophyPhilosophy– The symbol grounding problemThe symbol grounding problem

PsychologyPsychology– Perceptual symbol systemsPerceptual symbol systems

Computer scienceComputer science– Embodied cognitionEmbodied cognition

Page 5: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Translation Across Conceptual SystemsTranslation Across Conceptual Systems

How can we determine that two people share a How can we determine that two people share a matching concept of something (such as matching concept of something (such as MushroomMushroom)?)?– The publicity of concepts: we want to say that two The publicity of concepts: we want to say that two

people both have a concept of people both have a concept of MushroomMushroom even though even though they know different things (Fodor, 1998)they know different things (Fodor, 1998)

Cross-person translation as a challenge to Cross-person translation as a challenge to conceptual web accounts of meaning (Fodor & conceptual web accounts of meaning (Fodor & Lepore, 1992)Lepore, 1992)– If a concept’s meaning depends on its role in its system, If a concept’s meaning depends on its role in its system,

and if two people have different systems, then they and if two people have different systems, then they can’t have the same meaningcan’t have the same meaning

Page 6: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Problem DefinitionProblem Definition

Given Conceptual Systems Given Conceptual Systems AA and and BB– Each has a set of conceptsEach has a set of concepts– Both have the same set of relation typesBoth have the same set of relation types

Match the concepts in Match the concepts in AA with those in with those in B B based onbased on– Internal relations between concepts in each Internal relations between concepts in each

systemsystem– Partially known (external) correspondence Partially known (external) correspondence

between the two systems between the two systems

Page 7: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Graph RepresentationGraph Representation

Represent conceptual systems as Represent conceptual systems as graphsgraphs– Concepts --> nodes in graphConcepts --> nodes in graph– Relations --> links in graphRelations --> links in graph

Featured graphFeatured graph– Directed vs undirectedDirected vs undirected– Labeled vs unlabeled Labeled vs unlabeled – Weighted vs unweightedWeighted vs unweighted

Page 8: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Example of Conceptual SystemsExample of Conceptual Systems

seal reindeer moose

polar bear

wolf

Coexists (weighted)

Hunts (unweighted)

0.9 0.1 0.6 0.8

0.4

A sample of Manitoba wildlife

phoque renne élan

l’ours polaire

loup

0.9 0.3 0.8

0.6

Un sample de fauna Québécoise

Page 9: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Graph MatchingGraph Matching

InputsInputs– Graphs representing Graphs representing AA and and BB– External correspondence matrix External correspondence matrix EE

OutputOutput– Correspondence matrix Correspondence matrix C (m x n)C (m x n)– 0 ≤ C(A0 ≤ C(Aq, q, BBxx) ≤ 1) ≤ 1 indicating the correspondence between indicating the correspondence between

concept concept AAqq andand BBxx

Principle of AlignmentPrinciple of Alignment– Aligning nodes so that the relations between Aligning nodes so that the relations between

each pair of nodes in one graph are similar to each pair of nodes in one graph are similar to those between the aligned nodes in the other those between the aligned nodes in the other graphgraph

Page 10: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Measuring Similarity of Relation BundlesMeasuring Similarity of Relation Bundles

Relation bundleRelation bundle– All relations between a given pair of nodes represented All relations between a given pair of nodes represented

as an M-dimensional vectoras an M-dimensional vector Measuring similarity between Relation BundlesMeasuring similarity between Relation Bundles

Sim(ASim(Aqq,A,Arr;B;Bxx,B,Byy) = 1 - Diff(A) = 1 - Diff(Aqq,A,Arr;B;Bxx,B,Byy))

Diff(ADiff(Aqq,A,Arr;B;Bxx,B,Byy) = ∑|w) = ∑|wii(A(Aqq,A,Arr)-w)-wii(B(Bxx,B,Byy)| / M)| / M 0 ≤ Sim ≤ 10 ≤ Sim ≤ 1

Aq

Ar

0.2 1.0

Bx

By

0.6 0.5

Page 11: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Correspondence MatrixCorrespondence Matrix

C = C =

00000011AA44

00110000AA33

11000000AA22

00001100AA11

BB44BB33BB22BB11

A permutation matrix A permutation matrix

Describing a one-to-one matchingDescribing a one-to-one matching

Page 12: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Correspondence MatrixCorrespondence Matrix

C = C =

0.40.40.70.70000AA44

0.70.70.40.40000AA33

00000011AA22

00001100AA11

BB44BB33BB22BB11

Not a permutation matrixNot a permutation matrix

Used to find the best one-to-one matchingUsed to find the best one-to-one matching

Page 13: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

ABSURDIST II: An Optimization AlgorithmABSURDIST II: An Optimization Algorithm

Match quality of a permutation: global Match quality of a permutation: global edge similarity measure to maximizeedge similarity measure to maximizeGlobalEdgeSim(P) = GlobalEdgeSim(P) =

ββ∑∑q,rq,rSim(ASim(Aqq,A,Arr;B;BP(q)P(q),B,BP(r)P(r)) + α∑) + α∑q,rq,rE(AE(Aqq,B,BP(q)P(q)))

AB

Page 14: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Energy FunctionalEnergy Functional Generalization to an arbitrary correspondence Generalization to an arbitrary correspondence

matrix matrix CC: energy functional to maximize: energy functional to maximize

K(C) = α (E ∙ C) + β Excitation(C) - χ Inhibition(C)K(C) = α (E ∙ C) + β Excitation(C) - χ Inhibition(C) External( External( CC ) = ( ) = (E ∙ CE ∙ C)) Reward forReward for CC matching matching EE Excitation(Excitation( CC ) = ) =

∑∑q,r,x,y q,r,x,y Sim(ASim(Aqq,A,Arr;B;Bxx,B,Byy) C(A) C(Aq, q, BBxx) C(A) C(Ar, r, BByy) / (n-1)) / (n-1)

Reward for internal similarity matchingReward for internal similarity matching InhibitionInhibition( ( CC ) = ) =

(∑(∑q,r,xq,r,xC(AC(Aq,q,BBxx)C(A)C(Ar,r,BBxx) + ∑) + ∑q,x yq,x yC(AC(Aq,q,BBxx) C(A) C(Aq,q,BByy) )/(2(n-1))) )/(2(n-1))

Penalty for non-orthogonality of rows or columnsPenalty for non-orthogonality of rows or columns

Page 15: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Iterative Optimization ProcessIterative Optimization Process

Maximize Maximize K(C)K(C) on cube on cube QQ in in correspondence matrix spacecorrespondence matrix spaceCC00 = starting point = starting pointNNtt = = grad K(Cgrad K(Ctt) = net input) = net inputVVtt = = Damp( LNDamp( LNtt, C, Ct t )) = update = updateCCt+1t+1 = = CCt t + V + Vtt

Damp()Damp() = damping function= damping functionKeepsKeeps CCt+1t+1 inin QQ

LL = learning rate= learning rate Net input has external similarity, Net input has external similarity,

excitation, and inhibition termsexcitation, and inhibition terms

Page 16: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

ConvergenceConvergence

Similar to steepest descent, but Similar to steepest descent, but damping ensures convergence to a damping ensures convergence to a maximum on cube maximum on cube QQ

Always converges to a maximum Always converges to a maximum (not necessarily global)(not necessarily global)

Alternatives: quadratic programming Alternatives: quadratic programming methods (NP-hard)methods (NP-hard)

Page 17: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Choosing ParametersChoosing Parameters

Choose learning rate Choose learning rate LL – Stay in cube Stay in cube QQ– Ensure convergence Ensure convergence – Maximize convergence speedMaximize convergence speed

Page 18: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Choosing ParametersChoosing Parameters

Influence of Influence of χ /βχ /β on the convergence on the convergence points of points of CC– Low Low χχ: progress along the principal : progress along the principal

eigenvector, eventualy converge to a matrix of eigenvector, eventualy converge to a matrix of all 1’s all 1’s

– High High χχ: converge to the closest permutation : converge to the closest permutation matrix matrix

Heuristic solution: choose Heuristic solution: choose χ /βχ /β to better to better balance excitation and inhibitionbalance excitation and inhibition- Start with a lower - Start with a lower χχ, adaptively vary (generally, , adaptively vary (generally,

increase) it to avoid convergence to a matrix of increase) it to avoid convergence to a matrix of all 1’sall 1’s

Page 19: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Iteration CostsIteration Costs

O(nO(n44)) terms in the formula for terms in the formula for NNtt Sparse conceptual systems: average Sparse conceptual systems: average

degree of node is degree of node is d << nd << n Exploiting sparsity to compute Exploiting sparsity to compute

update in update in O(nO(n2 2 dd 2 2)) operations operations

Page 20: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Testing ABSURDIST IITesting ABSURDIST II• Create conceptual systems for two peopleCreate conceptual systems for two people

– Create a set of Create a set of NN concepts in Person concepts in Person AA– Define relations radnomly between concepts of various Define relations radnomly between concepts of various

labels and with random weightslabels and with random weights– Copy these concepts to Person Copy these concepts to Person BB– Add noise to Add noise to BB's relations and/or their weights's relations and/or their weights

• Measure ABSURDIST II’s ability to recover true Measure ABSURDIST II’s ability to recover true alignmentsalignments– 200 separate runs200 separate runs– Initialize each correspondence unit to a certain value (Initialize each correspondence unit to a certain value (0.5)0.5)– Activation passing for a set number of iterationsActivation passing for a set number of iterations– Any concepts connected by a unit with more than a Any concepts connected by a unit with more than a

threshold (threshold (0.80.8) activity are assumed to be aligned) activity are assumed to be aligned

Page 21: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

ABSURDIST II GUIABSURDIST II GUI

Page 22: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Noise Tolerence vs Graph SizeNoise Tolerence vs Graph Size

Page 23: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Noise Tolerence vs Graph DensityNoise Tolerence vs Graph Density

Page 24: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Coverage Noise vs Intensity NoiseCoverage Noise vs Intensity Noise

Page 25: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Iteration Steps vs Graph SizeIteration Steps vs Graph Size

Page 26: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Iteration Steps vs Graph DensityIteration Steps vs Graph Density

Page 27: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

External Similarity SeedingExternal Similarity Seeding

Page 28: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Potential Applications of ABSURDIST IIPotential Applications of ABSURDIST II

• Object recognitionObject recognition– Within-object relations provide a strong Within-object relations provide a strong

constraint for aligning objectsconstraint for aligning objects• Analogical reasoning and similarityAnalogical reasoning and similarity

– Most models of analogy require highly Most models of analogy require highly structured, propositional representations structured, propositional representations ABSURDIST II can be applied when similarities ABSURDIST II can be applied when similarities are known, but structured representations are are known, but structured representations are hard to find: pictures, words, etc.hard to find: pictures, words, etc.

• Translating across large databasesTranslating across large databases – Ontologies, dictionaries,thesauri, etc.Ontologies, dictionaries,thesauri, etc.

Page 29: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

ConclusionsConclusions• Connecting concepts to both the world and Connecting concepts to both the world and

each other is an attractive optioneach other is an attractive option– These connections are mutually supportive, not These connections are mutually supportive, not

antagonisticantagonistic– Within-system relational information may have Within-system relational information may have

surprisingly large influencessurprisingly large influences

• Absurdist II improvements on Absurdist IAbsurdist II improvements on Absurdist I– Graph representation of arbitrary conceputal Graph representation of arbitrary conceputal

systemssystems– Optimization frameworkOptimization framework– Better convergence behaviorBetter convergence behavior– Cost reduction from O(nCost reduction from O(n44) down to O(n) down to O(n2 2 dd 2 2))

Page 30: ABSURDIST II: A Graph Matching Algorithm and its Application to Conceptual System Translation Ying Feng Robert Goldstone Vladimir Menkov Computer Science.

Thank YouThank You

ABSURDIST II websiteABSURDIST II websitehttp://www.cs.indiana.edu/~yingfeng/ABSURDIST/http://www.cs.indiana.edu/~yingfeng/ABSURDIST/

Contact InfoContact Info– Ying Feng: Ying Feng: [email protected]@cs.indiana.edu

– Robert Goldstone: Robert Goldstone: [email protected]@indiana.edu

– Vladimir Menkov: Vladimir Menkov: [email protected]@cs.indiana.edu