Immature representation or immature deployment? Modeling ... · Observed Input Observed...
1
Observed comprehension behavior (Forsythe 2019) Observed Input H1 Noisy representation Intuition: Immature cue representations skew the information children can extract from their input. Likelihoods will be noisier than an optimally modeled learner. H2 Noisy deletion Intuition: Children may omit some cues when calculating pronoun meaning. Optimal model is mixed with sub-optimal cue deletion models. more noise Background and Research Questions Fig. 3: Best-fitting noise parameters for noisy representation model Fig. 2: Proportion subject antecedent responses for object-favoring ‒ subject-favoring cues Summary and Discussion Immature representation or immature deployment? Modeling child pronoun resolution Hannah Forsythe – [email protected] Lisa Pearl – [email protected] University of California, Irvine Computation of Language Lab Spanish subject pronouns are probabilistically associated with certain antecedents, depending on their (1a) FORM and the semantics of accompanying (1b) CONNECTIVES. They can also be categorically disambiguated by verbal number MORPHOLOGY (2). 54,757 utterances of naturalistic child-directed speech from Schmitt-Miller corpus (Forsythe et al. under review) hand-coded for reference to different antecedent types in the presence of each cue. (1) La maestra saluda a la niña … The teacher waves to the girl … a. … y { ø / ella } sale. and {pro/she} leaves. b. … { y después/porque } sale. {and then/because} leaves. (2) La maestra saluda a las niñas y … The teacher waves to the girls and …{ sale / salen }. leave-3S/3P %subject ant. (1,093/2,367) 0.46 (64/291) 0.22 (29/54) 0.54 (52/149) 0.35 % singular ant. (5,655/5,662) >0.99 (9/1,336) <0.01 (3) La maestra saluda a las niñas … The teacher waves to the girls … Subject antecedent Object antecedent … y después ø salen. … and then pro leave-3P Fig. 1: Example item, forced-choice picture selection (fully crossed) ø ella 0 1 porq desp PL SG 0 1 Q1 Can children use these cues to interpret subject pronouns? Q2 Are their non-adult-like interpretations due to immature representation or deployment of these cues? A1 Children’s use of Spanish verbal MORPHOLOGY is not fully adult-like, consistent with cross-linguistic findings (Pérez-Leroux 2005, Johnson et al. 2005, a.o.; but see Legendre et al. 2014). A2 This behavior is more likely caused by immature deployment of otherwise adult-like representations, consistent with other findings on child pronoun comprehension (ex. Principle B Conroy et al. 2009, Spenader et al. 2009). no noise noisy representation noisy deletion MSE ≤3 0.22 0.11 0.09 4 0.15 0.11 0.09 ≥5 0.15 0.11 0.08 adult 0.04 0.03 0.02 log likelihood ≤3 -1774 -913 -835 4 -1749 -956 -871 ≥5 -1264 -762 -668 adult -1201 -1113 -850 Table 2: Overall model fit to observed behavior Table 1: co-occurrence of antecedent types with cues in children’s input Fig. 4: Best-fitting noise parameters for noisy deletion model σ CON β CON σ MOR σ FOR β MOR β FOR P(α | fFOR, fCON, fMOR) P(fMOR | α) P(fFOR | α) P(fCON | α) P(fMOR | α) P(fFOR | α) P(fCON | α) P(α | f FOR , f CON , f MOR )∝ exp( log(σ FOR P(f FOR |α)) ) · exp( log(σ CON P(f CON |α)) ) · exp( log(σ MOR P(f MOR |α)) ) · P(α) (β FOR )(β CON )(β MOR ) P(α | f FOR , f CON , f MOR )+ (β FOR )(β CON )(1-β MOR ) P(α | f FOR , f CON )+ (β FOR )(1-β CON )(1-β MOR ) P(α | f FOR )+ …+ (1-β FOR )(1-β CON )(1-β MOR ) P(α) Acknowledgements: Many thanks to the teachers, children, and administrators at SEDI, Mexico City, MX, in particular Patricia de la Fuente and Beti López Juárez. Thanks to Lisa Pearl, Richard Futrell, Greg Scontras, Galia Bar Sever, Alandi Bates, and other members of the CoLaLab for helpful feedback. Support from NSF grant #SPRF-1810159 to Hannah Forsythe. ✓ ✓
Transcript of Immature representation or immature deployment? Modeling ... · Observed Input Observed...
Observedcomprehensionbehavior(Forsythe2019)ObservedInput
H1 NoisyrepresentationIntuition: Immature cue representations skew theinformation children can extract from their input.Likelihoods will be noisier than an optimallymodeled learner.
H2NoisydeletionIntuition: Children may omit some cues whencalculating pronoun meaning. Optimal model ismixed with sub-optimal cue deletionmodels.
morenoise
BackgroundandResearchQuestions
Fig.3:Best-fittingnoiseparametersfornoisyrepresentationmodel
Fig.2: Proportionsubjectantecedentresponsesforobject-favoring‒subject-favoringcues
SummaryandDiscussion
Immaturerepresentationorimmaturedeployment?Modelingchildpronounresolution
HannahForsythe– [email protected] LisaPearl– [email protected],Irvine
Computation of Language Lab
Spanish subject pronouns areprobabilistically associated withcertain antecedents, dependingon their (1a) FORM and thesemantics of accompanying (1b)CONNECTIVES. They can also becategorically disambiguated byverbal number MORPHOLOGY (2).
54,757 utterances of naturalistic child-directed speech from Schmitt-Millercorpus (Forsythe et al. under review) hand-coded for reference to differentantecedent types in the presence of each cue.
(1) Lamaestra saluda alaniña…Theteacherwavestothegirl…a.… y{ ø /ella }sale.
and{pro/she}leaves.b.… { ydespués/porque } sale.
{andthen/because} leaves.
(2)Lamaestra saluda alasniñas y…Theteacherwavestothegirlsand…{sale /salen }.leave-3S/3P
%subjectant.
(1,093/2,367) 0.46(64/291) 0.22
(29/54) 0.54(52/149) 0.35
%singularant.(5,655/5,662) >0.99
(9/1,336) <0.01
(3)Lamaestra saluda alasniñas…Theteacherwavestothegirls…
Subject antecedent Object antecedent
… ydespués ø salen.…andthenproleave-3P
Fig.1: Exampleitem,forced-choicepictureselection(fullycrossed)
øella
0 1
subj
"obj"
porq
ella
desp
ø
0 1
subj
obj
PLSG
0 1
Q1 Canchildrenusethesecuestointerpretsubjectpronouns?
Q2 Aretheirnon-adult-likeinterpretationsduetoimmaturerepresentation ordeploymentofthesecues?
A1 Children’s use of Spanish verbalMORPHOLOGY is not fully adult-like,consistent with cross-linguisticfindings (Pérez-Leroux 2005, Johnson etal. 2005, a.o.; but see Legendre et al. 2014).
A2 This behavior is more likelycaused by immature deploymentof otherwise adult-likerepresentations, consistent withother findings on child pronouncomprehension (ex. Principle BConroy et al. 2009, Spenader et al. 2009).
nonoise
noisyrepresentation
noisydeletion
MSE≤3 0.22 0.11 0.094 0.15 0.11 0.09≥5 0.15 0.11 0.08adult 0.04 0.03 0.02
loglikelihood≤3 -1774 -913 -8354 -1749 -956 -871≥5 -1264 -762 -668adult -1201 -1113 -850
Table2:Overallmodelfittoobservedbehavior
Table1:co-occurrenceofantecedenttypeswithcuesinchildren’sinput
Fig.4:Best-fittingnoiseparametersfornoisydeletionmodel
σCON βCON
σMOR
σFOR
βMOR
βFOR
P(α|fFOR,fCON,fMOR)
P(fMOR |α)
P(fFOR |α)
P(fCON |α)
P(fMOR |α)
P(fFOR |α)
P(fCON |α)
P(α|fFOR,fCON,fMOR) ∝ exp(log(σFOR P(fFOR |α)))·exp(log(σCON P(fCON |α)))·exp(log(σMOR P(fMOR |α)))
·P(α)
(βFOR )(βCON)(βMOR)P(α|fFOR,fCON,fMOR)+(βFOR )(βCON)(1-βMOR)P(α|fFOR,fCON)+(βFOR )(1-βCON)(1-βMOR)P(α|fFOR)+
…+(1-βFOR )(1-βCON)(1-βMOR)P(α)
Acknowledgements: Many thanks to theteachers, children, and administrators at SEDI,Mexico City, MX, in particular Patricia de laFuente and Beti López Juárez. Thanks to LisaPearl, Richard Futrell, Greg Scontras, Galia BarSever, Alandi Bates, and other members of theCoLaLab for helpful feedback. Support from NSFgrant #SPRF-1810159 to Hannah Forsythe.
✓
✓
