1

Vocabulary Development inFirst and Second Language Acquisition

Ping LiCognitive Science Lab

Department of PsychologyUniversity of RichmondRichmond, VA 23173

pli@richmond.eduhttp://cogsci.richmond.edu/

Overview of research

Does it differ in which language we speak? Crosslinguistic studies of Chinese and English

Does it differ in how many languages we speak? Bilingual language representation

Does it differ across a developmental trajectory? Child language acquisition

How does

process and organize

??

Approaches: behavioral, computational, & neural

Outline The lexicon in acquisition: Major issues

Vocabulary spurt Bilingual lexical representation Age of acquisition/critical periods

Connectionism and development

Self-organizing neural network models DevLex, DevLex-II, SOMBiP

Conclusions

Lexical Development in Monolinguals

Lexical development in children: it’s fast 14,000 words by age 6

Changes across time: not spread out evenly 0-18 months: 50 words 18-30 months: 500 words

Vocabulary spurt: rapid learning following slow growth Word-learning intrinsic factors

Phonological memory, word retrieval, lexical organization Cognitive and social abilities

Naming insight, communicative awakening

Lexical Development in Bilinguals

Simultaneous bilingual acquisition No spurt found for both L1 and L2 Individual differences: size, rate, and frequency of L2 input

Sequential bilingual acquisition Shared versus distinct lexical memory

(one store vs. two stores)

Learning (SLA) and representation (Bilingualism)

A computational perspective

Why computational modeling? Flexibility in parameter variation and

hypothesis testinge.g., Timing, size, and rate of input

Computational mechanisms required forlexical acquisition

e.g., association, organization, and competition

2

Localization vs.Organization

Neurons in the brainconnect with one

another to form networks

The brain learns by modifyingcertain connections in

response to inputs

Neural Networks

units

weights

learning rule

activations

connections

Neurons in the brain connect with one

another to form networks

The brain learns by modifyingcertain connections in

response to inputs

Neural Networks

Neurons in the brain connect with one

another to form networks

The brain learns by modifyingcertain connections in

response to inputs

Information is distributed through large groups of connected units

Knowledge is represented bypatterns of activation

Learning is accomplished by adjustment of the weights that connect the units

Emergent Properties

Connectionist language learning models

Limitation of previous computational models

Artificial input patterns Small size of lexicon Supervised learning (e.g., BP network) Monolingual learning & representation

Connectionist language learning models

Our Model

Realistic input (based on parental speech) Early child lexicon (500 words based on CDI) Unsupervised learning (Self-Organizing Maps) Mono-&-bilingual learning and representation

This has been made possible by the availability of large-scale speech corporaonline (e.g., CHILDES, CDI, HAL) and the computational tools therein

3

Computational principles of our model

Self-Organizing Map (SOM; Kohonen, 1982, 2001) Unsupervised Learning Topography-preserving Gradual formation of structures with soft boundaries

Forming lexical categories

Computational principles of our model

Self-Organizing Map (SOM; Kohonen, 1982, 2001) Unsupervised Learning Topography-preserving Gradual formation of structures with soft boundaries

Hebbian Learning Different maps can be connected via Hebbian learning,

according to which associative strengths of the correspondingnodes increase through co-activation

Li, P., Farkas, I., & MacWhinney, B. (2004). Early Lexical Development in a Self- Organizing Neural Network. Neural Networks, 17, 1345-1362.

Li, P., Zhao, X., & MacWhinney, B. (2007). Dynamic self-organization and early lexicaldevelopment in children. Cognitive Science, 31.

DevLex: a Developmental Lexicon model (Li et al. 2004)

The DevLex Bilingual Model (SOMBiP; Li & Farkas, 2004)

ENGLISH SEMANTICS

CHINESE SEMANTICS

CHINESE PHONOLOGY

ENGLISH PHONOLOGY

ASSOCIATIVE CONNECTIONS (Hebbian

learning)

Self-organization

Self-organization

Wo

r d F

or m

P

hon olo gi c a

l W

or d

Mea n

i ng

C

o- oc c ur re nc e- bas e d re pr es ent a

ti on ( deri ved fr om

separ ate com

pon ent

expos ed to bil ingua

l corp us)

Phonological Map

Semantic Map

Input phonology map(SOM)

Self-Organization

Word Form(Phonetic features)

Output sequence map(SARDNET)

Self-Organization

Word Sequence(Phonetic features)

Hebbian Learning(Comprehension)

Hebbian Learning(Production)

Semantic map(SOM)

Word Meaning Representation(WCD)

Self-Organization

The DevLex-II model (Li et al. 2007)

4

Simulating lexical disorders(Miikkulainen, 1997; Silberman et al., 2007) Input Representations

Phonological representations PatPho - a phonological pattern generator

(Li & MacWhinney, 2002)

Phonemic representations Articulatory features of phonemes

(Ladefoged, 1982)

Semantic representations WCD - a word co-occurrence detector

(Farkas & Li, 2002; Li et al., 2004)

Emergence of Lexical Categories in DevLex Emergence of structured semantic representations

50 150

250 500

Rapid learning in early vocabulary

5

Rapid learning and lexical organization

Vocabulary spurt occurs when Structured representations in word meaning

and phonological shape are established Associative mapping is consolidated “Setting up the basic framework”

Word confusion rates across time

Effects of word frequency and length Effects of phonological working memory

Effects of associative capacity Effects of simulated lesion

6

Summary of DevLex Results for L1

Developmental changes in word learning are modulated bylexical organization; Lexical organization is reflected in thedevelopment of structured representation

Individual differences with respect to the shape and function ofearly vocabulary are explained by the interaction betweenseveral parameters, including phonological short-term memoryand associative capacity

Learning itself can determine the shape of change, anddiscontinuous patterns of development can emerge from thesame underlying mechanisms

Emergence of lexical representations based onbilingual input

ENGLISH SEMANTICS

CHINESE SEMANTICS

CHINESE PHONOLOGY

ENGLISH PHONOLOGY

ASSOCIATIVE CONNECTIONS (Hebbian

learning)

Self-organization

Self-organization

Wo r d Fo r m

P

h on olog ica l W

o r d Me an in g

Co- oc cu rr en ce -b as ed

re pr es en t at ion ( der i ved fr om

s epar ate componen t

expo sed t o bil ingu al c or pus)

Phonological Map

Semantic Map

Emergence of lexical representations(Simultaneous Learning)

Effects of age of learning

Early L2 learning Late L2 learning

Effects of age of learning

32:10.3:1L2:L1

2%3244English

64%956ChineseLateL2 Learning

1.87:10.75:1L2:L1

11%2397English

20.6%1803ChineseEarlyL2 learning

1:10.94:1L2:L1

12.8%2162English

12.8%2038ChineseSimultaneouslearning

Confusion rateLexical space

Table 1: Size of lexical space and rate of confusion for Chinese (L2)vs. English (L1) on the semantic map

Summary of DevLex Results for L2

When the learning of L2 is early relative to that of L1,functionally distinct representations of the two lexiconsmay be established.

When the learning of L2 is delayed relative to that ofL1, the structural consolidation of L1 will significantly(sometimes dramatically) impact on the representationof L2 (e.g., resulting in parasitic L2).

Plasticity and competition: The network’ ability toreorganize its structure decreases as the structure ofL1 has consolidated

7

Li, et al. 2004

Neural correlates ofnouns and verbs inChinese

Neural correlates of nouns and verbs in early bilingualsChan, et al. (2006)

ERP signatures of subject-verb agreement in L2 learningChen, et al. (2006) Conclusions

DevLex is a developmentally plausible and computationallyrealistic model designed to account for the dynamic self-organization in lexical learning and representation

The model captures important mechanisms underlyingdevelopmental phenomena (e.g., early plasticity, competition,and experience-dependent structural changes; Bates, 1999).

Our model projects lexical development at a dynamicalsystems level, in which phonological and semanticrepresentations of words continuously interact and evolve.

New research in language acquisition point to directionsbeyond the nature-nurture debate (e.g., infant studies; cf. Kuhl;Saffran), and our research exemplifies such directions.

ReferencesLi, P. (2006). Modeling language acquisition and processing: Connectionist networks. In

Li, P., Tan, L.H., Bates, E., & Tzeng, O. (eds.), Handbook of East Asian Psycholinguistics(Vol. 1: Chinese). Cambridge, UK: Cambridge University Press.

Li, P., Zhao, X., & MacWhinney, B. (2007). Dynamic self-organization and children’s wordlearning. Cognitive Science, 31.

Li, P., Farkas, I., & MacWhinney, B. (2004). Early Lexical Development in a Self-OrganizingNeural Network. Neural Networks, 17, 1345-1362.

Li, P. (2003). Language acquisition in a self-organizing neural network. In P. Quinlan (ed.)Connectionist Models of Development: Insights from real and artificial neural networks.Taylor & Francis: Psychology Press.

Li, P., & Farkas, I. (2002). A self-organizing connectionist model of bilingual processing. InR. Heredia & J. Altarriba (eds.), Bilingual sentence processing. North-Holland: ElsevierScience Publisher.

Hernandez, A., Li, P., & MacWhinney, B. (2005). The emergence of competing modules inbilingualism. Trends in Cognitive Sciences, 9, 220-225.

Hernandez, A., & Li, P. (2007). Age of acquisition: Its neural and computational mechanisms.Psychological Bulletin.

For reprints or preprints, visit http://cogsci.richmond.edu/