1. Children must be biased learners.(given the indeterminacy problem…?)How do young children learnto name objects?They must know something abouthow nouns map onto categories.Names for thingsshape bias material bias“dax”“blicket”Novel Noun Generalization Task(e.g., Landau, Smith & Jones, 88; Soja, Carey & Spelke, 92; Samuelson & Smith, 99)2Children seem to honor a distinction betweensolid objects and non-solid substancesThe Phenomenon• Biases that constrain word learning• Abstract, apply to things never seen before• Divide the world into two ontologicalcategoriesEmerge with development3The nativist idea:milksandchaircastleeggguidesAn opposing idea from theconnectionist perspective:milksandchaircastle egg4An opposing idea:chaircastleeggmilksandeggmilksandchaircastleegggeneralizeballballballStep 1.cupcupChild learns thenames of individualobjects (slowly, effortfully).Same name goeswith same shape inthe input.Attention to shape isrewarded.5Balls are round.Step 2.Cups arecup-shaped1st ordergeneralization:Child forms individuallexical categoriesorganized by objectshape.Step 3.“ ‘s ”are _____-shaped.Balls are round.Cups are cup-shaped.That is…a “shape bias”.Child develops ageneral tendency toattend to objectshape whenextending an objectname to newinstances.2nd ordergeneralization:6Step 4.-shaped thing is “giraffe”.-shaped thing is “crayon”.-shaped thing is“tractor”Child shows an increasein rate of new objectname acquisitions.In consequence…• Are the words children know reallyorganized this way?– Yes!• Is that organization enough to createbiases?– Yes!– I’ll show you in a computational model.• Well, ok, but can your model make anynovel predictions?– Yes!7Are the words children know reallyorganized this way? (Samuelson & Smith, 1999)• examined the similarity structure of 300early nouns• for each noun:– Are the items in this category solid ornon-solid?– Are the items in this category alike in shape?– Are the items in this category alike in material?– Are the items in this category alike in color?Samuelson & Smith, 1999shapematerialshapematerial8• Are the words children know reallyorganized this way?– Yes!• Is that organization enough to createbiases?shapematerialshapematerialExperiment 1 - NetworksWord LayerHidden LayerObject LayerShapeMaterial Solidity9TrainingSHAPE MATERIAL SOLIDITY“ ball”“ ball”“ ball”Testingshape biasmaterial biasSim( , ) > Sim ( , )Sim( , ) < Sim ( , )10Resultssolid non-solid67%61%• Are the words children know reallyorganized this way?– Yes!• Is that organization enough to createbiases?– Yes!• How about novel predictions?11A new prediction: the ontology biasnames do not refer to categories that violateontological boundariesshapematerialshapematerialeggExperiment 2 - NetworksDoes the network believe that names will not violatethe boundary between solid and non-solid?solidsolidnon-solidsolidnon-solidnon-solidsolid non-solid12solidsolidnon-solidnon-solidnon-solidsolidExperiment 1 - Ontology maintainingExperiment 2 - Ontology violatingnon-solid solid67%61%Experiment 2 - NetworksResults55%63%Experiment 3• given a solid object, will children refuseto generalize its name to an object of thesame shape if it is not solid?• given a non-solid substance, willchildren refuse to generalize its name to amaterial match that is solid?13Stimuli - in the networksSHAPE MATERIALSOLIDITYSOLIDITYStimuli - for children- ontology violating shape matches14Stimuli - ontology violating material matchesSubjects: 24 children, 30-36 months oldSolid vs. Non-solid exemplarbetween subjectsOntology Maintaining vs. Ontology Violatingwithin subjectEach child had 12 trialsExperiment 3 - ChildrenDesign15Experiment 3 - ChildrenOntology Maintaining71%47%Solid TargetNon-solid TargetResultsOntology Violatingsolidnon-solidSolid TargetNon-solid TargetExperiment 3 - ChildrenResults55%76%16Conclusions• Early words present an organized structure• Enough to create a shape bias for solids anda material bias for nonsolids• Enough to create an “ontology:”a belief in solid objects and nonsolid substancesas different kinds• In two kinds of statistical learners: networks and childrenOther predictions of the model• Early material bias for some kinds of shapes• Similarities and differences in noungeneralizations by Japanese- vs. English-speakers.• Similarities and differences in noungeneralizations by Spanish- vs.