The emergence of linguistic knowledge in infant speech ... 4_Cristia... · The emergence of linguistic knowledge in infant speech perception ... Kaplan et al. (2002); Uther et al.

1

The emergence of linguistic knowledge in infant speech perception

When, why and how to build a holistic view

Alex CristiaLSCP, CNRS

Vowels (between 2 and 14)Consonants (between 6 and 122)

Figures from the World Atlas of Language Structure

Few (2-4)Many (7-14)

Few (6-14)Many (34-122)

The input

Infant & adult-directed speech (IDS-ADS)

These two registers differ in most cultures (Lieven, 1994)

Infants attend preferentially to IDS (meta-analysis in Dunst et al., 2011)

Acoustic characteristics that distinguish IDS from ADS similar across languages/cultures → functional pressures (Ferguson, 1964)

Why is IDS different from ADS?

Attributed to (at least) 2 possible goals1. Engage attention, modulate arousal, and express

affection → IDS ≈ happy speechFernald (1990); Kaplan et al. (2002); Uther et al. (2007)

… Higher absolute pitch & expanded pitch range (Scherer, 2003;(Ohala, 1983)

… Smiling → shorter vocal tract and tenser vocal configuration – affects some vowels more than others (Tartter, 1980)



affection → IDS ≈ happy speech

2. Provide “good” linguistic material to facilitate language acquisition → IDS ≈ hyperspeech

Fernald (2000); Kuhl et al. (1997); Sundberg (2001)

Clear ADS: Language-specific expansion along identifying acoustic dimensions (e.g., Andruski, 2006, Kang & Guion, 2009)

Note that prosody also affected: slower speech rates & sometimes higher pitch and larger pitch ranges (Bradlow et al., 2003; Picheny et al., 1985; Smiljianic & Bradlow, 2005))



affection → IDS ≈ happy speech

2. Provide “good” linguistic material to facilitate language acquisition → IDS ≈ hyperspeech

Caregivers enhance phonologically/lexically distinctive dimensions

→ Phonemic distinctions will be stretched→ Allophonic distinctions will not

• Moms asked to talk about a set of objectsFirst to infant alone; then to experimenter &

confederate– 28 American English moms of 4-month-olds– 18 American English moms of 11-month-olds

• Objects carefully chosen to elicit tense/lax, nasal/oral, and point vowels in controlled lexical & phonotactic environments– Similar vowels (8 words w/ each vowel) with

matched preceding and following Cs and similar word frequency• Tense/lax: ε/e (pesto/basil), ɪ,i (picnic/peekaboo) • Oral/nasal: æ~/æ (dancer/tassle), ɛ/̃ε/ (tender/teddy)• Cs: (b,p,d,t,k,g)V(s,ʃ,b,p,d,t,k,)

Methods

IDS does not show category enhancement

Cristia & Seidl (2013) JCLScripts etc. athttps://sites.google.com/site/acrsta/ Home/nsf_allophones_corpora

Nu

mbe

r of results

Variation evident in a systematic review

Cristia (2013), Language and Linguistics CompassAnnotated tables downloadable from sites.google.com/site/acrsta/IDS_meta-analysis

Not an isolated findingFactors

Benders (2012)McMurray et al. (in press)

Lam & Kitamura (2012)Burnham et al. (in press)

Impact of the input

Symbolic: Presence of categories

/s/ /ʃ/

Hypothetical acoustic dimension

Distributional: modulated by distributions

of acoustic cues

/s/ /ç/ /ʂ/

/s/

Hawaiian

English

Polish

Frequency of occurrence

Cf. Ladefoged & Cho (2005)

Mom speech

Sally!/s/ pole ~8kHz

Prosody

Sounds

42 moms+ babies

4-6m: 18

12-14m: 24

…

/ʃɔ .. ʃa/

/ʃɔ .. ʃa/

/ʃɔ .. ʃa/

Habituation

Average of preceding 3 trials falls below 40% of longest 3 trials

/sɔ .. sa/

/ʃɔ .. ʃa/

/ʃɔ .. ʃa/

/sɔ .. sa/

Test: fixed duration (15s)Baby /s-ʃ/

discrimination

/s/

Mom speech Baby /s-ʃ/ discrimination

Cristià 2011 JASA

4-6mo 12-14mo

Cristià et al. 2011 JPhon

Beyond correlations: AGL

Some learning (→ meta-analysis of [un]published results with Sophie ter Schure <[email protected]>)

Frequency

/ç/ /ʂ/

Frequency

Exposure type 1 Exposure type 2

2

1 Individual variation in natural language acquisition

Laboratory learning (strict control input)

/ç/ /ʂ/

Impact of the input in the context of 'real' language acquisition

First steps

Phonology-first

Lexicon-first

Interactive view

Rääsanen (2012)

Johnson (2004); Beckman & Edwards (2010)

Using individual variance longitudinally: Initial benchmarks

1. Is the individual variation measured meaningful?

2. Are we tapping linguistic skill

(or just general lab performance)?

A longitudinal approach: Initial benchmarks


Can it predict vocabulary size?

Searches in scholar.google.com, Pubmed, Science Direct, and Proquest

Identified 20 articles and theses 43 effect sizes [r] (+ 8 without comparable ES) Median r within infant group Grouped into three types: speech sounds, words, prosody

Example: Processing of sounds



24

Median effect size and weight per infant group

25

Median effect size and weight per infant group

26

Median effect size over all studies

r=.31 [.22, .4]

Speech perception at 4-12 months explains 5-15% variance

in vocabulary size at 11-48 months

Speech sounds r=.35 [.22, .47]Words r=.28 [.14,.4]

Prosody r=.42 [.18, .61]

Cristia, Seidl, Junge, Soderstrom, & Hagoort (in press), Developmental ScienceDatabase available at sites.google.com/site/invarinf (Individual Variation in Infancy)

A longitudinal approach: Initial benchmarks


2. Are we tapping linguistic skill

(or just lab performance)?

For now, 2 ways to answer

28

1. Reuse existing data

Habituation r=.45 [.2, .65]Dishabituation r=.42 [.29, .54]

Rapid auditory processing r=.54 [.25, .74]

Overall non-linguistic infant predictors of vocabulary r=.43 [.35, .5]

Speech-speech: r=.43 [.28, .57]

Correlations within infancy Speech + non-linguistic: r=.25 [.08, .40]

29

2. Direct experimental comparison

Predictions: General lab performance Cognitive demands of the task Emergent language skills

30



31



32



33



Trochees Vowels VRM

Vowels 0.30*

VRM 0.05 -0.07

A/B 0.1 -0.2 0.09

Seidl, Cristia, Wang, & French (variably rejected)

Kuhl, 1994; Kuhl et al. 2008 Curtin & Werker, 2005 : Curtin et al., 2011

Extension to current theories

PHONEME[k]

WORDFORM

generalizations

The synthetic learner (BOOTPHON) DUPOUX

The synthetic learner+

+ Deployment of attention+ Memory systems+ ...

Increasing sensitivity: better measures

High replicability (+portability)

High reliability

'Intensive'- Compatible with a multivariate approach

Input, output, perception

+ state-of-the-art algorithms, e.g.SOUNDS: Supervised phone recognizer WORDS: Semi-supervised term discovery

Other problems of individual variance

Limited pool of variance:

Twin studies → heritability 25-67% (Stromswold, 2001); single factor emerges (e.g., Bornstein et al., 2012)

Conceptual ambiguity:

Precise theories (quantitative predictions)

Causality can never be demonstrated

AGL, models, animal parallels...

Summary

Multivariate approaches

Not only in vogue, but even necessary (often)

Infant studies' low SNR

Interactions may fall below sensitivity

Principled way to structure search

Models could play a key role

Face conceptual inseparability

Thank you.

The emergence of linguistic knowledge in infant speech ... 4_Cristia... · The emergence of linguistic knowledge in infant speech perception ... Kaplan et al. (2002); Uther et al.

Documents