ebook img

Concept Generation in Language Evolution PDF

0.22 MB·
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Concept Generation in Language Evolution

Concept Generation in Language Evolution MarthaLewis,JonathanLawry DepartmentofEngineeringMathematics,UniversityofBristol,BS81TR,UK [email protected],[email protected] Abstract upofqualitydimensionsandequippedwithadistancemetric, 6 1 forexampletheRGBcolourspace. Thisthesisinvestigatesthegenerationofnewcon- 0 Label semantics proposes that agents use a set of labels cepts from combinations of existing concepts as a 2 LA = {L ,...,L } to describe a conceptual space Ω with 1 n language evolves. We give a method for combin- distancemetricd(x,y). LabelsL areassociatedwithproto- n i ing concepts, and will be investigating the utility a typesPi ⊆ Ωanduncertainthresholdsεi,drawnfromprob- of composite concepts in language evolution and 5 J thencetheutilityofconceptgeneration. athbailtitayndeisletrmibeunttioxns∈δεiΩ. TishesuthffirecsiehnotlldyεcilocsaepttuorePsithteonboetiloan- 2 belledLi. TheappropriatenessofalabelLi todescribexis 1 Introduction quantifiedbyµLi(x),givenby ] I Humans are skilled at making sense of novel combinations A (cid:90) ∞ of concepts, so to create artifical languages for implemen- µ (x)=P(d(x,P )≤ε )= δ (ε )dε . tation in AI systems, we must model this ability. Stan- Li i i εi i i s d(x,Pi) c dard approaches to combining concepts, e.g. fuzzy set [ theory, have been shown to be inadequate [Osherson and LabelscanthenbedescribedasLi =<Pi,d(x,y),δεi>. Smith, 1981]. Composite labels frequently have ‘emergent 1 v attributes’ [Hampton, 1987] which cannot be explicated by 3 ANewModelofConceptComposition 2 decomposing the label into its constituent parts. We argue Experiments in [Hampton, 1987] propose that human con- 3 thatinthiscaseanewconceptisgenerated.Thisprojectaims 7 to determine conditions for such concept generation, using cept combination can (roughly) be modelled as a weighted 6 multi-agentmodelsoflanguageevolution. sum of attributes such as ‘has feathers’, ‘talks’ (for the con- 0 cept‘Bird’). Theseattributesdifferfromqualitydimensions . 1.1 ThesisOutline in conceptual spaces: they tend to be binary, complex, and 1 multidimensional. Wethereforevieweachattributeasalabel 0 The project divides into three parts. Firstly, we have devel- inaconceptualspaceΩ andcombinetheselabelsinabinary 6 oped a model of concept combination within the label se- i v:1 mmaondteilcsisfirnasmpeirweodrbkyasangdivreenfleinct[sLraewsuryltsanind[THaanmg,p2to0n0,91].98T7h]e, sspuacchela{b0e,l1s}α˜n=ill(cid:86)usnit=ra1te±dLiinmfiagpusreto1a, wbihnearreyaveccotonrju(cid:126)xnαcttiaokninogf value 1 for positive labels L and 0 for negated labels ¬L . i inwhichmembershipinacompositeconceptcanberendered i i X Wetreatmembershipinα˜inthebinaryspacewithinthelabel astheweightedsumofmembershipsinindividualconcepts. semanticsframework. Soα˜ isdescribedinthebinaryspace r Secondly, wemustshowthatcompositionalitycanevolve a byα˜ =<(cid:126)x ,d((cid:126)x,(cid:126)x(cid:48)),δ>asbefore. within a population of interacting agents. Preliminary work α inthisareaexaminestheabilityofapopulationofagentsto convergetoasharedsetofdimensionweights. {0,1}n Thirdly, we will investigate the generation of new unitary concepts from existing composite concepts, building further uponthemulti-agentmodel. 2 Background Ω1 Ω2 Ωn Thisworkisbasedonthelabelsemanticsframework[Lawry, 2004;LawryandTang,2009],togetherwithprototypetheory Figure1: Combininglabelsinabinaryspace [Rosch, 1975], where membership in a concept is based on proximitytoaprototype,andconceptualspaces[Ga¨rdenfors, Wedefineadistancemetricinthebinaryspace{0,1}nas: 2004]. Thelatterviewsconceptsasregionsofaspacemade Definition1 WeightedHammingDistance 1 1 SD SD For (cid:126)λ ∈ (R+)n, ∀(cid:126)x,(cid:126)x(cid:48) ∈ {0,1}n, where (·) is the scalar 0.8 λ 0.8 λ product, H ((cid:126)x,(cid:126)x(cid:48))=(cid:126)λ·|(cid:126)x−(cid:126)x(cid:48)| 0.6 0.6 (cid:126)λ Theorem2 Let α = (cid:86)n ±L and λ = (cid:80)n λ . Let 0.4 0.4 i=1 i T i=1 i ε∼U(0,λ ),d=H . Then: 0.2 0.2 T (cid:126)λ µα˜(Y(cid:126))=(cid:88)n λλi µ±Li(Yi) 0 0.6 w0.8 1 0 0.6 w0.8 1 T i=1 (a) x ∼ U[0,1], x ∼ (b) x ∼ U[0.25,0.75], 1 2 1 Compound concepts θ˜,ϕ˜ may be combined in a higher U[0,0.5]. λ converges to x2 ∼ U[0,0.5]. λ con- levelbinaryspace. Thenθ˜•ϕ˜canbeexpressedinthecontin- 0.5forallvaluesofw vergestovaryingvalues. uousspaceasaweightedsumofθ˜andϕ˜. Figure 2: Mean SD and λ at time t = 2000 for different Theorem3 Letθ˜•ϕ˜=<{(1,1)},H ,δ>.Thenµ (Y(cid:126))= valuesofw. Eachpointaverages25simulationsrunwith10 w(cid:126) θ˜•ϕ˜ agents. (cid:80)ni=1(w1λϕTwλTθλiθ+Twλ2ϕλTθTλϕi)µ±Li(Y(cid:126)). We have therefore shown that combining labels in a When w = 1 we can predict the value to which λ will weightedbinaryspaceleadsnaturallytothecreationofcom- converge. Consider the quantity A − λt which determines positeandcompoundconceptsasweightedsumsofindivid- whethertheupdateispositiveornegativeateachstep. ual labels, reflecting results in [Hampton, 1987]. We have Definition4 A positive region R+ ⊂ Ω is a set of points furthercharacterisednotionsofnecessaryandimpossibleat- R+ ={(cid:126)x∈Ω:A−λ ≥0} t tributesusingideasfrompossibilitytheory. Theorem5 Let p+ denote the probability of a point(cid:126)x ∈ Ω fallinginapositiveregionandletw = 1acrossthepopula- 4 ConvergenceofDimensionWeightsAcross tion. Thentheexpectedvalueofλconvergestop+. aPopulation 5 FurtherWork We investigate how a population of agents in a multi-agent simulationplayingaseriesoflanguagegamesmightconverge Wearecurrentlyworkingonanalyticalresultstopredictthe to a shared set of dimension weights. Agents with equal la- value of λ to which agents converge. Under certain circum- belsL =L =<1,d,U[0,1]>∈Ω =Ω =[0,1](disEu- stances, such as the case where w = 1, or with an altered 1 2 1 2 clidean distance), and randomly initiated weights λ ∈ [0,1] updatingmodel,analyticresultsarepossible. Wewillextend engage in a series of dialogues about elements in the con- thisworktolookattheutilityofusingconjunctiveassertions ceptual space, adjusting their weights after each dialogue is withinthesesimulations. completed. Ateachtimestep,speakeragentsmakeassertions Work in the third year will focus on examining how new α = ±L ∧±L aboutelements(cid:126)x ∈ Ω ×Ω whichmax- concepts might be generated from the combination of exist- i 1 2 1 2 imiseµ ((cid:126)x)=λµ (x )+(1−λ)µ (x ). ingones. Wewillbuildonthelanguageevolutionmodelcur- αi L1 1 L2 2 Thelisteneragentassessesα againstitsownlabelset. If rentlyindevelopment. i µ (x) ≤ w,thereliabilityofthespeakeragent,thelistener αi agentupdatesitslabelset. References Theupdateconsistsinincrementingthedimensionweight [Ga¨rdenfors,2004] P. Ga¨rdenfors. Conceptual spaces: The λtowardsavalueA,sothatλ = λ +h(A−λ )where t+1 t t geometryofthought. TheMITPress,2004. h=10−3and [Hampton,1987] J.A. Hampton. Inheritance of attributes A= w−µL2(x2) in natural concept conjunctions. Memory & Cognition, µ (x )−µ (x ) 15(1):55–71,1987. ±L1 1 ±L2 2 [LawryandTang,2009] J. Lawry and Y. Tang. Uncertainty Thisisthequantitythatsatisfiesµ (x)=w. IfA<0(or αi modelling for vague concepts: A prototype theory ap- A>1)wesetA=0(orA=1). proach. ArtificialIntelligence,173(18):1539–1558,2009. Theconvergenceacrossthepopulationismeasuredbythe standarddeviation(SD)oftheλacrossthepopulation. [Lawry,2004] J. Lawry. A framework for linguistic mod- Figure2showstheresultsoftwosetsofsimulationsacross elling. ArtificialIntelligence,155(1-2):1–39,2004. varyingvaluesofw.Thetwosetsofsimulationshavedistinct [OshersonandSmith,1981] D.N.OshersonandE.E.Smith. distributionsofelementsencounteredwithinthespace.When On the adequacy of prototype theory as a theory of con- wis0.5orbelow,theagentsdonotconvergetoshareddimen- cepts. Cognition,9(1):35–58,1981. sionweights(notshown).Whenw >0.5,agentsdoconverge [Rosch,1975] E. Rosch. Cognitive representations of se- to shared dimension weights: SD is low. The weights con- mantic categories. Journal of experimental psychology: vergedtodependbothonthereliability,w,ofeachagent,and General,104(3):192,1975. thedistributionofelementsintheconceptualspace.

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.