神经符号人工智能:黑盒模型时代的任务导向调研 Neuro-Symbolic Artificial Intelligence A Task-Directed Survey in the Black-Box Models Era_第1页
神经符号人工智能:黑盒模型时代的任务导向调研 Neuro-Symbolic Artificial Intelligence A Task-Directed Survey in the Black-Box Models Era_第2页
神经符号人工智能:黑盒模型时代的任务导向调研 Neuro-Symbolic Artificial Intelligence A Task-Directed Survey in the Black-Box Models Era_第3页
神经符号人工智能:黑盒模型时代的任务导向调研 Neuro-Symbolic Artificial Intelligence A Task-Directed Survey in the Black-Box Models Era_第4页
神经符号人工智能:黑盒模型时代的任务导向调研 Neuro-Symbolic Artificial Intelligence A Task-Directed Survey in the Black-Box Models Era_第5页
已阅读5页,还剩13页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

10418

ProceedingsoftheThirty-FourthInternationalJointConferenceonArtificialIntelligence(IJCAI-25)

SurveyTrack

Neuro-SymbolicArtificialIntelligence:

ATask-DirectedSurveyintheBlack-BoxModelsEra

GiovanniPioDelvecchio,LorenzoMolfetta,GianlucaMoro

DepartmentofComputerScienceandEngineering,UniversityofBologna,CesenaCampus

Viadell’Universit50,I-47522Cesena,Italy

{g.delvecchio,lorenzo.molfetta,gianluca.moro}@unibo.it

Abstract

Theintegrationofsymboliccomputingwithneu-ralnetworkshasintriguedresearcherssincethefirsttheorizationsofArtificialintelligence(AI).

TheabilityofNeuro-Symbolic(NeSy)methodstoinferorexploitbehavioralschemahasbeenwidelyconsideredasoneofthepossibleproxiesforhuman-levelintelligence.However,thelim-itedsemanticgeneralizabilityandthechallengesindecliningcomplexdomainswithpre-definedpatternsandruleshindertheirpracticalimple-mentationinreal-worldscenarios.Theunprece-dentedresultsachievedbyconnectionistsystemssincethelastAIbreakthroughin2017haveraisedquestionsaboutthecompetitivenessofNeSyso-lutions,withparticularemphasisontheNaturalLanguageProcessingandComputerVisionfields.

Thissurveyexaminestask-specificadvancementsintheNeSydomaintoexplorehowincorporat-ingsymbolicsystemscanenhanceexplainabil-ityandreasoningcapabilities.Ourfindingsaremeanttoserveasaresourceforresearchersex-ploringexplainableNeSymethodologiesforreal-lifetasksandapplications.Reproducibilitydetailsandin-depthcommentsoneachsurveyedresearchworkaremadeavailableat

/

disi-unibo-nlp/task-oriented-neuro-symbolic.git

.

1Introduction

“Stackedneurallayersisallweneed”.Thisstatementsum-marizesmostofthecurrentresearcheffortsintheArtificialIntelligence(AI)field,especiallyinNaturalLanguagePro-cessing(NLP)andComputerVision(CV).Theadvancementsandachievementsofthelatestneuralmodelsaremarvelous,buttheyconcealfundamentaldrawbacksregardingdataef-ficiencyandexplainability.TheneedtoresorttoNeuro-Symbolic(NeSy)componentsnaturallyarisesfromthene-cessityfortrustworthyandefficientsolutions.Thedatatypeof“thoughts”intheconnectionistapproaches—alsoknownastensors—andthehiddenunsupervisedmanipulationsofsuchinformationconstituteaphysicalbarriertohierarchicalandabstractplanning,whoseachievementcannotbereachedviamereinput-outputrelationships[

Marcus,2018

].Ifwe

18

AAAI50(20)

rt

IJCAI31(11)

o

h

S

20

NeurIPS28(17)

c

i

g

o

L

ICLR17(10)

y

ICML17(14)

N

e

S

B

l

92

EMNLP9(6)

a

c

k

-

B

42

CVPR7(4)

o

x

ACL7(5)

ICCV2(0)

JMLR2(2)

NAACL1(1)

TACL1(1)

ConsideredDiscarded

Figure1:DistributionofNeSypeer-reviewedpapersoverthepe-riod2017-2024.Theinnermostringdelineatestheinclusioncriteria.Greyscaleslicesdenotestudiesexcludedforrelyingonblack-boxmethods,exclusivelylogicalapproaches,orbrevity,whilethecol-oredslicecomprisesthesurveyedones.Theoutermostringrepre-sentsthenumberofresearchworksfromeachselectedvenue,withexactcountsreportedinthelegend.Thenumberofpapersconsid-eredforeachtrackisenclosedinparentheses.

envisionanot-too-distantfuturewheredataavailabilityandqualitybecomecriticalconcerns,mainlyduetothepoten-tialbiasintroducedbysyntheticinformation,weurgefindingandexploringresearchpathwayswithcompositionalabilitiesweaklyrelatedtothetrainingset’sdimension[

Giunchiglia

etal.,2022

].Weupholdthatneuro-symbolicmethodscanprovideanalternativetobreakthecurse-of-dimensionalitymodernmethodologiessufferfrom.Whileeffective,spar-sityenforcementregularizationmechanisms[

Bach,2017

]re-quirestrongassumptionsaboutthenatureofthetargetdis-tribution,whichisunrealisticforthecomplexdomainsNLPandCVmodelsusuallydealwith[

Bronsteinetal.,2021

].Wearguethatsymboliccomponentscanbetheregulariza-tionforforcingmorecomplexbehaviorintoneuralmod-els.Bycombiningdata-driveninsightswithexplainable,logic-basedrepresentations,wecanachievehighergeneral-izationcapabilities[

Besoldetal.,2021

].Otherdata-drivenparadigmshavealsobeenconsidered[

Lodietal.,2010

;

Domeniconietal.,2015

].

Asthekeywords“reasoning”and“explainability”aregain-ingmoreattentioninthecommunity,wehighlightakeynotationalmisunderstandingbetweenrule-guidedandpost-

ProceedingsoftheThirty-FourthInternationalJointConferenceonArtificialIntelligence(IJCAI-25)

SurveyTrack

10419

inferenceforensicstrategies.Whilebothmethodsaimtopro-videarationalebehindpredictions,theyfundamentallydifferintheirapproach.Rule-guidedmethodsconstraintheout-puttomeetspecificcriteria,whilepost-inferenceapproachesretrospectivelyreconstructplausibleinterpretablereasoningpathwaysthatledtotheresults.Wefurtherclaimthatthesetermsshouldnotbeusedtoadvocatehuman-likefacultiesbutasproxiesforgeneralizationbehaviorwithinahierar-chicalbackbonearchitecturegovernedbypredefinedrulesorcomponent-interactionschemas.

OurContributionThelackofacleardefinitionofNeSyintheAIfield,theinconsistencyinevaluationbenchmarks,andthemultitudeofuncoordinatedstudydirectionswithnocommoncomparisongroundsprimarilydrivethissurvey.Inthiswork,weconductatask-directedliteratureanalysisfo-cusingonhowNeSymethodsapplyandscaleindiverseap-plicationcontexts.Weassessthemethodologicalsoundnessofsuchapproachesandcomparethemtoblack-boxsystemsinreal-worldscenariostoidentifythecurrentlimitationsofthesestrategies.Departingfromrecentsurveysthatattempttoclassifyresearchworksalongtaxonomicalaxesornet-workarchitecturalvariations,weproposeanewmethodolog-icalapproachthatdisentanglestheseconceptualandpracti-calinconsistenciesbyprovidingacomprehensiveandcriticaloverviewofNeSysystems.Weaimtodeepentheunderstand-ingofthesehybridsystems’advantagesandinherentlimita-tions,layingthegroundworkforfutureresearchandbridgingthegapbetweenexplainabilityandperformanceinAI.

2Taxonomy

Thissectiondiscussesthestructureandsystematicrepro-duciblemethodologiesfollowedtodefineourtask-directedtaxonomyforNeSymethods.Wefurtherexaminedatasets,evaluationbenchmarks,andtheirlimitationsindelineatingfaircomparisonsacrossdifferentapproaches.Aspreviouslydiscussed,weacknowledgethenuancedusagesandincon-sistencyinthedeclinationofthetermNeSyinNLPandCVliterature,whereitisoftenmisappliedtomethodsclaimingreasoningandinterpretabilitywithoutstructuredsymbolicframeworks.Inthispaper,NeSyexclusivelyreferstoap-proachesintegratingneuralnetworkswithsymboliccompo-nentssuchassolvers,logicalrules,orstate-actionschemas.

InclusionCriteriaThissurveyaimstoinvestigateNeSysystemsfromthelastAIrevolutionin2017.Toiden-tifyrelevantresearchsystematically,weusedtheDBLPSPARQLendpointtocollectresearchpaperspublishedbe-tween2017and2024fromgeneral-purposeleadingvenuesinthosefieldswheresuchtechnologicalleaphastakenholdmore—NaturalLanguageProcessing(NLP)andComputerVision(CV).AcustomizedquerywasexecutedtoretrievepublicationsrelatedtoNeSymethodsusingthekeywords:“neuro-symbolic”,“nesy”,“rule-based”,“probabilistic-logic”,“probabilistic-reasoning”,“logic-based”,“soft-logic”,“fuzzy-logic”,“concept-learning”,“inductive-logicprogram-ming”.Figure

1

illustratesthenumberanddistributionoftheresearchworksmeetingourrequirements.Wemeticulouslyanalyzedtheresulting172papersandcategorizedthemto

identifyworksthatdeviatedfromthedesignatedNeSyfor-mulation,excludingnon-interpretablemethods,purelylogi-calapproaches,andshortpapers.Thiscollectionwasthor-oughlystudiedtoderiveapoolofresearchworksfocusingonNeSyapproachesthatintegratesymbolicreasoningwithneuralnetworks.WefurtherexploredrelevantrelatedworkstogainaprecisepictureoftheNeSyresearchlandscape.

DatasetsandBenchmarksWeanalyzedreal-worlddatasetsandbenchmarkstoensureanunbiasedcomparisonacrossmethodswhilefocusingonpracticalapplications.Thisstudyhighlightedreproducibilitychallengesintasksinvolvingsamplingoperations,suchasnegative-examplemining.WefoundsubstantialinconsistenciesontheWN18RR

1

benchmarkfordifferentresearchworksusingthesameblack-boxmodelbaselines.WefurtherspotlightanevaluationtrendintheNesyliteraturewheresyntheticdatasetsandtoytasksarespecificallytailoredforthefeaturesofthenewlyproposedmethod.Forthisreason,weconsiderthemunsuitablefordrawingconclusionsandcomparisons.InFigure

2

,welabelmostoftheidentifiedNeSytaskswiththeirrespectiveevaluationbenchmarkiftheyrespectourfairnesscriteria.Missingvaluesappearinimagegeneration,wheredifferentinstructionpromptssignificantlyimpactresults;inReinforcementLearning(RL),whereeachmethodisevaluatedoncustomtasksorgames;andincausaleffectestimation,duetooutdatedbenchmarkswithlimitedentries.

ATask-OrientedFormulationWeorganizeouranalysisbyidentifyingthreedistinctmethodologicalframeworksthattraversetheneural-symbolicresearchlandscape:(1)RuleMining,(2)RuleEnforcement,and(3)ProgramSynthesis.Withineachframework,wecategorizetechniquesbasedontheirtheoreticalgroundingandthetaskstheyaddress,focusingonhowconstraints,symbolicstructures,andinterpretableprogrammaticconstructsarerespectivelyextracted,integrated,orsynthesized.OurtaxonomyinFigure

2

isdesignedtobereadintwocomplementaryways.Atop-downperspectivehighlightshowtheseframeworksinterconnect,emphasizingsharedlogicalformalismsandlearningparadigms.Abottom-upapproachbeginswithaspecifictask,movingupwardtoidentifythemostrelevantmethodologicalfamiliesandtechniquesforthatobjective.Thisdualperspectivepreservesthefield’stheoreticalcoher-encewhilepromotingatask-directedformulationofNeSysolutions.Enablingbidirectionalexplorationbetweenrelatedmethodologiesallowsresearcherstoidentifyanalogousstrategiestorefineorexpandexistingapproachesbasedontheirobjectives.InFigure

3

weshowexamplesoftheformallanguagesemployedbyeachtheoreticalframeworkinthetaxonomy.

Inthefollowingsections,weanalyzeeachmethodologyindepth,illustratingitscoreprinciplesandhighlightingoppor-tunitiesforadvancementsinneural-symbolicintegration.

1

/TimDettmers/ConvE

ProceedingsoftheThirty-FourthInternationalJointConferenceonArtificialIntelligence(IJCAI-25)

SurveyTrack

10420

DeterministicFiniteAutomata

Synthesis

Program

Rule

Enforcement

Gaming

RuleMining

DeterministicFiniteAutomata

Context-FreeGrammars

CausalEffectEstimation

LearningShaping

NaturalLogic

SemanticParsing

Reinforcement

ProbabilisticLogic

+

ReferIt3D

3DQuestionAnswering

TemporalCommonsense

Reasoning

McTACO,TIMEDIAL

FirstOrderLogic

Claim

Verification

Inference

NaturalLanguage

MultiWoZ2.1

Generation

Dialog

+

Reinforcement

LearningShielding

HornClauses

RefCOCO

FEVER/FEVEROUS

e-SNLI

Image

Manipulation

Relation

Extraction

Link

Prediction

Classification

TemporalProcessModelling

FOLIO

VerificationWeibo

Image-Text+

PredicatedDiffusion

Numerical

VisualQuestion

+

LogicalReasoning

ReasoningAnswering

CUB

DWIE

ALGEBRA

ADE20K

MIMIC-IV

WN18RR

NeuralNetworkVerification

GTSRB

Figure2:Task-DirectedNeSyTaxonomy.WeorganizethemostrelevanttaskfamiliesbyclassifyingeachaccordingtoapplicableNeSy

techniquesandgroupingthemintothreemacro-categories:(1)RuleMining,(2)RuleEnforcement,and(3)ProgramSynthesis.Wealsolabelthemostcommonlyuseddatasetstohighlighttheirroleinreal-worldbenchmarkingandmodelevaluation.Tasksareorganizedfrombottomtotopandfromlefttorightwithineachcategory,mirroringthesurvey’sconceptualprogression.

3Rule-Miningtechniques

Neuro-Symbolictechniquesemployrule-miningasacoremethodologyformodelconstruction,focusingonextract-inginterpretablerulesfromdiverseinputdata.Thecom-putationalcomplexityofruleextractiondirectlycorrelateswiththechosenformalspecificationlanguage.Hornclauseapproachesexhibitthehighestcomplexity,requiringmatrixrepresentationsforgroundatomtruthvalues.NaturalLogicframeworksshowmoderatecomplexity,combininglinguis-ticpreprocessingwithquestion-answeringforentityandop-eratorextraction.DeterministicFiniteAutomata(DFAs)presentthelowestcomplexity,leveragingoperationaltraceswithSAT-solvingforrulelearning.

3.1HornClauses

HornClauses(HCs),afundamentalcomponentoflogicpro-gramming,provideastructuredframeworkforrule-basedreasoning.Definedasdisjunctionsofliteralswithatmostonepositiveliteralastheclausehead,HCsenablelogicalin-ferenceacrossvariousdomains.

TemporalpointprocessmodelingreliesonHC-basedrea-soningtocaptureeventdependencies,whichisparticularlycrucialinfieldssuchasmedicineandautonomousdriving,wheresuddenchangesinconditionscanhavesevereconse-quences.[

Yangetal.,2024

]introducedaNeSyruleinductionmethodthatleveragesasequentialcoveringalgorithmandacustomattentionmechanismtoextractHCs.Whiledemon-stratingscalability,trustworthiness,andstrongperformanceoverexistingmodels,itsabilitytocaptureincreasinglycom-plexrulesremainsanopenresearchquestion.

TheInductiveLogicProgramming(ILP)paradigmauto-matesthediscoveryofHC-basedrules,supportinggeneral-izedreasoningandinferenceoverstructureddata.Inthecon-textofextractingrulesfromstructureddataliketablesandgraphs,ILPseekstoidentifymissingconnectionsbetweenentitiesbyuncoveringpatternsinexistingrelationships.DFORL[

Gaoetal.,2024

],arecentlightweightmethodforefficientruleextraction,introducesadepth-limitedbreadth-firstsearchforneighborhoodextraction.Thisproposition-alizationtechniqueconvertsrelationalfactsintovectorrep-resentationssuitableforneuralnetwork-basedlearningandappliessyntacticconstraintstoreducetherulesearchspace.

Byintegratingauxiliarymatricesandcurriculumlearning,DFORLuncovershiddenpredicatesandenhancestheeffi-ciencyandprogressionofthetrainingprocess.Althoughthisapproachdemonstratespotentialinreal-worldapplica-tionssuchasdrugdesign[

Kingetal.,1995

],itsperformanceontheWN18RRlinkpredictionbenchmarkremainsincon-sistent.Thisvariabilityprimarilystemsfromthedataset’sdesign,whichomitsinverserelationsinthetestset,com-plicatingdirectcomparisonswithbaselinemethodsandsug-gestingthatmethodslikeDFORLcouldbenefitfromincor-poratingnegationinrulebodiestoenhancethemodel’sabil-itytohandlehigher-aritypredicates.NCRL,acomplemen-taryapproachby[

Chengetal.,2023

],focusesonrulecom-positionality,samplingalternativepathsbetweenconnectednodestoconstructHornclauses.Thismethodemploysanit-erativealgorithmcombininganRNN-basedselectionprocesswithanattentionmechanism,aimingtomaximizethelike-lihoodthatheadpredicatescanbereconstructedfromsam-pledpathpredicates.Byleveragingpredicateembeddings,NCRLoutperforms[

Gaoetal.,2024

]onWN18RR,mainlyduetoitshigherabilitytoinfersemanticrelationshipswheninverserelationsaremissing.Itsstrongperformanceinlow-dataregimesandrapidconvergenceonGPUmakesNCRLacompellingalternativeforthesametask.

Appliedtotabulardata,InductiveLogicProgrammingfoundationsareleveragedinFold-SE[

WangandGupta,

2024

]toextractinterpretablelogicalrulesforsupervisedtabularclassification.Designedtohandlebothcategoricalandnumericaldatawithminimalpreprocessing,itscalesef-ficientlytolargedatasetswhilepreservingrulecoherence.Followingasequentialcoveringstrategy,Fold-SEiterativelyrefinesrulesbyoptimizingamodifiedGiniImpuritymet-rictoenhanceclassificationperformance.BuiltuponFold-SE,NeSyFOLD[

Padalkaretal.,2024

]adoptsILPonex-tractedhighlevelfeaturesforanimageclassificationtask.NeSyFOLDderivesrulesfrombinaryactivationmaskspro-ducedbyConvolutionalNeuralNetworks(CNNs),assign-ingsemanticlabelstokernelsposthoc.ItsrelianceonCNNarchitectures—extendedandoutperformedbymoderntransformer-basedmodels—constrainsitseffectiveness,andtheexclusionofexceptionpredicatesinsemanticlabelingfur-therrestrictsitsrepresentationalcapacity.

ProceedingsoftheThirty-FourthInternationalJointConferenceonArtificialIntelligence(IJCAI-25)

SurveyTrack

10421

Inthetextualdomain,Document-levelRelationExtrac-tion(DocRE)isasupervisedlearningtaskthatidentifiesre-lationsr(h,t)betweenentities,whererrepresentsthere-lation,histheheadentity,andtisthetailentity.BothNeSymodelsandblack-boxapproacheshavebeenevalu-atedonDWIE,awidelyrecognizedbenchmarkforDocRE.JMLR,anovelframeworkintroducedin[

Qietal.,2024

],integratesDocREwithruleextractionthroughresidualcon-nections.Thisapproachcomputesrulesupportasaweightedsumofatomsupports,facilitatingtheimplementationofsoft-proofmechanismsthatcombinerelationsintoHCs.[

Jainet

al.,2024

]adoptedanalternativeformulation,treatingDocREasaknowledgebaselinkpredictiontaskandutilizingtheblack-boxmodelDocRE-CLiPtoinferrelationalstructures.FedNSL,introducedin[

Xingetal.,2024

],extendsNeSymethodstoafederatedlearningsettingforDocRE,achiev-ingperformancebetweenthatofDocRE-CLiPandJMLR.Whileitsevaluationofreal-worlddatasetsremainslimitedtoDWIE,theframeworkdemonstratesthepotentialforprivacy-preservingalgorithms.Byleveragingvariationalexpectationmaximization,itefficientlyconstrainstherulesearchspace,reducingtrainingroundswhilemaintainingcompetitiveper-formance.Acomparativeevaluationofthesemethods—asshowninTable

1

—indicatesthatJMLRsignificantlyout-performsDocRE-CLiPonDWIEwhileachievingsimilarre-sultsonotherbenchmarks,highlightingtheadvantagesofrule-basedreasoningincapturingcomplexmulti-sentencere-lationships.However,JMLRdoesnotexplicitlyincorporateexternalknowledge,afeaturemorenaturallyintegratedintoknowledge-base-drivenmethods.

3.2NaturalLogic

NaturalLanguageInference(NLI)determinesentailmentre-lationsbetweenapremiseandahypothesisusingformalop-eratorssuchasequivalence(=)andforwardentailment(c).[

Fengetal.,2022

]proposedatransformer-basedNeSymodelthatintegratedGPT-2withreinforcementlearningforop-eratorcompositionandintrospectiverevisionusingWord-Net.ThisapproachoutperformsbaselineslikeBERTonmostbenchmarksbutremainssensitivetolinguisticnoise,suchasadverbsandprepositionalphrasemodificationsinbench-markdatasets.QA-NatVer[

Alyetal.,2023

]extendedthisframeworktoclaimverification,constructingproofsthroughsentencealignment,operatorassignment,andacustomDFA.AsshowninTable

1

,QA-NatVerhighlightsthetrade-offbe-tweenNeSymodels’explainabilityandperformance,withanotableaccuracydropduetoitsrelianceonasmalltrainingset.Itsblack-boxcounterpart,SFAVEL[

Bazagaetal.,2024

],overcomesthislimitationusingself-supervisionanddistilla-tiontoremovetheneedforlabeleddata.TabVer[

Alyand

Vlachos,2024

]furtherextendedNLItotabularclaimver-ificationbyincorporatingnumericalreasoning,recognizingequivalences,alongsidepragmaticreasoningthroughdeter-ministicrules.WhileachievingstrongresultsonFEVER-OUSdatasets,itstruggleswithexactnumericalmatching,un-derscoringongoingchallengesinadaptingNeSymethodstolessrigidreasoningtasks.

3.3DeterministicFiniteAutomaton

DeepQ-learninghasproveneffectiveinreinforcementlearning(RL),particularlyintasksthatrequiresequen-tialdecision-makingandstructuredexploration.Traditionalblack-boxmethodsstrugglewithenvironmentsliketheAtariclassicgame“Montezuma’sRevenge”,wheresuccessde-pendsoncomplexinteractionsbetweenobjectsandpreciseroomnavigation.ADeterministicFiniteAutomatonprovidesastructuredwaytomodelsequentialdependenciesbyrep-resentingstatetransitionsthroughafinitesetofrules.Thisformalismcanbeleveragedtoguideexplorationandim-provedecision-makingincomplexRLenvironments.Lever-agingthisformalism,[

Hasanbeigetal.,2024

]proposesadeepQ-learningframeworkaugmentedwithDFAsynthesistoenhanceexplorationefficiencyandpolicyoptimization.Bystructuringlearnedbehaviorsasstatetransitions,thisap-proachenablesmoreeffectivereasoningoversequentialde-pendencies.Withinthisframework,theRLalgorithmgener-atesexplorationtracesthatcapturesequencesofstate-actionpairsalongsiderewardestimates.ThesetracesinformaDFAsynthesismodule,whichformulateslogicalconstraintsandemploysaSATsolvertoconstructaminimal-stateDFAen-codingtheagent’slearnedbehavior.Byleveragingthisstruc-turedrepresentation,thedeeplearningmodulerefinespolicytransitions,optimizingstate-actionmappingsanditerativelyimprovingdecision-makingwithintheRLprocess.Experi-mentsdemonstratethattheframeworkachievesfasterconver-genceintasksrequiringsequentialplanningandobjectinter-actions,outperformingconventionalmodelsthatfailtocon-verge.Integratingexpert-designedDFAsfurtherenhancesef-ficiencybyeliminatingtheinitialexplorationphaseandexpe-ditingtraining.Futureresearchmayextendthisapproachbe-yondgamingandinvestigateusingmoreexpressiveautomata,suchasPushdownAutomata,tomodelhierarchicalmemorystructures[

Sipser,1997

].

4Rule-EnforcementTechniques

Thissectionpresentspromisingmethodologiesforenforcingconstraintsintheformofrulesoversystemsforvariousappli-cationpurposes.Theseconstraintscanbeenforcedthroughtailorednetworksorbyregularization.

4.1First-Order-Logic

IntegratingFirst-OrderLogic(FOL)rulesintoconstrainedimagegenerationhasbeenexploredtoenhancecontrolla-bilityandtrustworthiness.[

SueyoshiandMatsubara,2024

]introducedapioneeringapproachthatextractsFOLcon-straintsfromtextusingdependencyparsingandtranslatesthemintoequationsgoverningtheintensityofattentionmaps.Ratherthanmininglogicalconstraintsfromlargedatasets,thismethodemploysaregularizationstrategy—awell-establishedpracticeinneuro-symbolicreasoning—asdemonstratedbylaterworksinthissurvey[

Caietal.,2022

;

Pryoretal.,2023

].Inthisapproach,thelogicalstructureisnotinferredfromdatabutexplicitlyimposedthroughaspe-cializedlossfunctionthatguidesthegenerationprocess.No-tably,thismethodnativelysupportslogicalquantifiers,which

ProceedingsoftheThirty-FourthInternationalJointConferenceonArtificialIntelligence(IJCAI-25)

SurveyTrack

10422

NaturalLanguageInference

Thekid

doesnotlovetabletennis

≡≡c

Thekid

sports

doesnotlike

SemanticParsing

SubjectsAllrectangleshavefour-sides

PredicatesAll ObjectsAll

four-sidedthingsareshapesrectanglesareshapes

HornClauses/FirstOrderLogic

Allrectangleshavefoursides→FourSided(x)←Rectangle(x)

/

Allrectangleshavefoursides→∀xFourSided(x)←Rectangle(x)

ProbabilisticLogic

I'malmostsurethatallrectangleshavefoursides

0.9:FourSided(x)←Rectangle(x)

DeterministicFiniteAutomaton

run

start

idle

ContextFreeGrammars

Wellformedparentheses:{λ,(),(()),...,(n)n}R1:S→(S)Terminalsymbols:{"(",")",λ}

R2:S→λNonterminalsymbols:{S}

(λistheemptystring)

Figure3:Overviewoftheintermediateformallanguagesemployedbyeachtaxonomymethod.

allowforformulatingmorepreciseandinterpretablecon-straints.QualitativeanalysessuggestthatincorporatingFOLconstraintsinimagegenerationsignificantlyenhancescon-tentfidelity,makingthisapproachhighlyrelevantforapplica-tionslikechatbotsandautomatedcontentcreation.However,theabsenceofacountingmechanismpreventsconstrainten-forcementonrepeatedentities(e.g.,“ablackdogandawhitedog”),whileabiastowardprototypicalexampleslimitsflexi-bility.Visualconceptlearningandsemanticparsingcouldbeincorporatedtoaddressthesechallenges[

Maoetal.,2019

].

Reasoningoverimagesisessentialfordetectingmulti-modalmisinformation,wheredeceptionstemsfromthein-terplayoftextandvisuals.Whiletraditionalblack-boxap-proacheshaveachievedhighaccuracyinthisdomain,thelackofinterpretabilityremainsamajorlimitation.[

Dongetal.,

2024

]introducedaNeSypipelinethatenhancestransparencybyintegratinglogicalreasoningintotheclassificationpro-cess.Usingtwoencodermodels,themodelextractspatternsfrombothmodalitiesandemploysateacher-studentnetworktoestimateprobabilitiesforthreelatentvariables:imagema-nipulation,cross-modalinconsistency,andimagerepurpos-ing.AuniqueFOLruleclassifiesasampleasfakeifanyofthethreevariablesholdtrue.Table

1

showsthismethodtrail-ingitsblack-boxcounterpartbyafewpoints,like

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论