International Child Phonology Conference 16 June 2011 New sound patterns are learned first in frequently heard words Mits Ota University of Edinburgh Sam Green University College London 1 Phonology vs. words Targetlike production /tw/ /pl/ /kr/ Age 2 Phonology vs. words Targetlike production [pl]um /pl/ [pl]ease [pl]ay [pl]ate Age 3 Factors behind lexical variability Age of word acquisition (AoWA) (Garlock et al., 2001) Neighborhood density (Garlock et al., 2001; Munson & Swenson, 2005) Lexical frequency (Gierut, 2001; Stoel-Gammon, 2010) In production In the input 4 Previous work: Results Input frequency effects on production accuracy? Population Yes No TD; 5-7 years Leonard & Ritterman (1971) Moore, Burke, & Adams, (1976); Garlock, Walley, & Metsala (2001) PD; 3-7 years Gierut, Morrisette, & Champion (1999); Gierut & Storkel (2002); Morrisette & Gierut (2002) TD; 1-3 years Ota (2006) 5 Sosa (2008) Previous work: Sources of mixed results Population Sound patterns analyzed Frequency estimates Confounds Input frequency ~ production frequency (≈ sampling), AoWA, neighborhood density, word length Method Elicited production of selected words Longitudinal spontaneous production data 6 Current study Population : 1-4 year old TD children Sound patterns analyzed : Word-initial consonant clusters Frequency estimates : Based on mother’s speech Confounds : Controlled in regression analysis Input frequency ~ production frequency (≈ sampling), AoWA, neighborhood density, word length Method Elicited production of selected words Longitudinal spontaneous production data Survival analysis 7 Sampling problem in corpus analysis 20 25 30 35 40 45 plus Pluto Accuracy ?→ 1.0 0.8 0.6 0.4 0.2 0.0 please pleasure Production accuracy of /pl/- in a spontaneous speech corpus plenty plow plum plate play playdoh plan plane planet ←? ←? plant 1.0 0.8 0.6 0.4 0.2 0.0 plastic ?→ 1.0 0.8 0.6 0.4 0.2 0.0 place 20 25 30 35 40 45 plain 20 25 30 35 40 45 Age (months) 8 20 25 30 35 40 45 1.0 0.8 0.6 0.4 0.2 0.0 Survival analysis start Proportion of people alive 4 3 1 2 Patient ID 5 6 end 0 20 40 60 80 Time since start of study = Death (observed) = Death (unobserved) = Last confirmation 9 100 Time A survival curve for heart transplant patients (critical event = death) Cox regression Proportion of people alive non-smokers smokers Time 10 Covariates Smoking** Gender Age*** No. of heart attacks** Survival analysis of cluster acquisition start Proportion not acquired 4 3 1 2 WordID Patient 5 6 end 0 20 40 60 80 100 Time since start of study = Accuracy ≥ 80% (observed) = Accuracy ≥ 80% (unobserved) = Last obs of accuracy < 80% 11 Time A ‘survival curve’ for words with an initial cluster (critical event = acquisition) Cox regression on cluster acquisition Proportion not acquired high frequency Low frequency Time 12 Covariates Input frequency? AoWA? Word length? Neighborhood density? Cluster type? Specific research questions All other factors being equal, does input frequency increase the proportion of words in which an initial consonant cluster is acquired (= 80%+ accuracy) before 4? Does input frequency equally affect different types of wordinitial clusters (e.g., /pl/, /sw/, /skr/)? Production of early-acquired prosodic forms in Japanese is less affected by input frequency (Ota, 2006) 13 Data Providence Corpus (Demuth, Culbertson & Alter, 2006) Longitudinal Phonetically transcribed Includes maternal speech Child 14 Age No of sessions Tokens of words with cluster Types of words with cluster Lily 1;3-4;0 80 5,209 481 Naima 0;11-4;0 88 7,140 521 Violet 1;2-4;0 54 1,536 271 Age of cluster acquisition Age of cluster acquisition for each word = First 3-month bin with production accuracy above 80% Nontargetlike production Deletion: [pe] ‘play’ Epenthesis: [pəle] ‘play’ Substitution: [pwe] ‘play’ Targetlike production Everything else 15 Predictor variables (covariates) Input frequency: Summed token count of each lexical item in mother’s speech (log) Production frequency: Mean monthly token count of each lexical item in child’s speech (log) Age of word acquisition: Month of first production attempt Neighborhood density: Number of neighbors (log) Word size: Number of phonemes 16 Predictor variables (covariates) Cluster size: CC (e.g., /st/, /pl/) vs. CCC (e.g., /str/, /spl/) Cluster type: C(C)w: /tw/ (twinkle), /skw/ (squash) C(C)j: /mj/ (music), /skj/ (skewer) C(C)r: /kr/ (cry), /spr/ (spring) C(C)l: /kl/ (clean), /spl/ (splash) SN: /sn/ (snow), /sm/ (small) SP: /st/ (star), /sk/ (skip) 17 Frequency effects on survival curves 0.8 0.0 0.2 0.4 0.6 0.8 0.6 0.4 0.2 0.0 Proportion of words with cluster not acquired 1.0 Below-median frequency 1.0 Above-median frequency 15 20 25 30 35 Months 18 40 45 15 20 25 30 35 Months 40 45 Correlations between covariates 35 45 -3 -1 0 1 2 4 6 8 10 12 0.0 1.0 2.0 3.0 0 1 2 3 4 5 6 15 25 35 45 25 First AoWA 1 15 25 35 45 15 Production frequency 12 -3 -1 LogMeanAttempts 3.0 4 6 8 NPhon Word size 0.0 6 4 2 Input frequency LogMotFreq 0 = Lily = Naima = Violet 1.5 N’hood Log.Nb density 0 1 2 3 4 5 6 19 Cox regression: Words learned before 2 C(C)r and C(C)l slower AoWA Cluster type -18%*** Input frequency 20 +16%* List 1 (AoWA ≤ 2;0) Input frequency x Cluster type Frequency effect weaker in C(C)r and C(C)j Cox regression: Words learned between 2 and 3 C(C)l, C(C)j and SN faster; C(C)r slower Cluster type Frequency effect weaker in C(C)l, C(C)j, SN and C(C)r Input frequency x Cluster type AoWA Word length -7%** Input frequency 21 +11%*** -8%** List 2 2;0 < AoWA ≤ 3;0 -15%*** N’hood density Finding 1 Production of initial clusters is mastered first in frequently heard words. Why input frequency matters? Exposure updates lexical representation Misrepresentation as a source of production inaccuracy (Macken, 1992) Improvement of representation and exposure (Schwartz & Terrell, 1983; Swingley, 2007) Perception-production mismatch places pressure to overcome output restrictions (Coetzee & Pater, 2009) 22 Finding 2 Lexical input frequency effects are weaker in fastacquired clusters – except C(C)r. The problem with C(C)r is /r/, not the cluster. Lily (3;1.0) Naima (3;10.10) Violet (3;7.22) [wɛd] ‘red’ [wuf] ‘roof’ [waɪʔ] ‘right’ The rest is consistent with Ota’s (2006) observation: ‘Easy’ sound pattern = less frequency effects Generalization of learning from words to sound patterns? 23 Finding 3 For words learned between 2 and 3, acquisition of initial clusters is also faster in short words and in sparse neighborhoods. Implications for the emergence of lexical neighborhoods in children (Charles-Luce & Luce, 1990, Coady & Aslin, 2003, Dollaghan, 1994) Counter evidence to the Lexical Restructuring Model (Metsala & Walley, 1998)? 24 Implications Frequency: Sounds vs. words Frequency: Input vs. output Development: Phonology vs. words 25