Bilingual Insights into the Initial Lexicon

The Role of Cognates in Word Acquisition

Gonzalo Garcia-Castro

PhD Defence / Departament de Medicina i Ciències de la Vida

2024-11-03

The initial lexicon

Average 20-year-old knows ~42,000 lemmas: mental lexicon

Lexical representations: Phonological, conceptual, grammatical information of known words

First lexical representations at 6-9 months

Inter-modal experimental paradigms (Bergelson and Swingley 2012; Tincoff and Jusczyk 1999)
Parental reports and surveys (e.g., CDI, Fenson et al. 1994)

Normative trajectories of lexical development

Vocabulary size norms for 51,800 monolingual children learning 35 distinct languages (Frank et al. 2017)

Bilinguals face additional challenges, but do not lag behind

Increased complexity in linguistic context

Reduced linguistic input in each language

Increased referential ambiguity

Two overlapping codes

Split into two languages

> 2 labels per referent

Bilinguals face additional challenges, but do not lag behind

Hoff et al. (2012): bilinguals acquire words at similar rates as monolinguals

Despite these challenges, bilinguals reach most of their language acquisition milestones at equivalent ages as monolinguals
This is specially true for lexical development
For instance, I highlight this study by Hoff et al. (2012), in which they collected English and Spanish CDI administrations for a cohort of English monolinguals and English/Spanish bilingual children living in Florida.
Both groups were matched across many demographic variables, like SES
When the authors compared monolingual and bilingual vocabulary sizes in each language separately, they observed that English monolinguals knew more English words on average than bilinguals.
But when English and Spanish vocabulary sizes were combine into a measure of total vocabulary size, monolinguals and bilinguals’ trajectories of vocabulary growth were equivalent.
This begs the question: How do bilinguals keep up with monolinguals despite facing a more challenging lingusitic environment?

Lexical similarity modulates vocabulary growth in bilinguals

Floccia et al. (2018): CDI responses of 372 bilinguals learning English + additional language

Lexical similarity: Average phonological similarity (Levenshtein) between pairs of translations

Higher lexical similarity, larger vocabulary size

Stronger effect in the additional language (e.g., Dutch, Mandarin)

Recent studies have suggested that bilinguals might exploit the lexical similarity between their languages to boost the vocabulary growth.
In an influential monograph published in 2019, Floccia et al. collected vocabulary data from a large sample of bilingual 2-year-olds in the UK.
These children were learning English and an additional language, out of a diverse sample of 13 other languages from different typological families, linguistic groups, or that followed different grammatical systems.
Most children were English-dominant, but were exposed to the additional language in verying degrees
The authors calculated a measure of lexical similarity for each pair of languages.
For each language pair, they calculated the average phonological similarity between the translation equivalents included in the English vocabulary checklists and in each of the other additional languages
This measure was included as a predictor in a model that estimated participants’ vocabulary size in each of their languages
This index of lexical similarity turned out to be positively related with children’s vocabulary size in their additional language
For instance, English-Dutch bilinguals (learning two lexically similar languages) knew more Dutch words than Enlgish-Mandaring (low similarity languages) knew in Mandarin
Overall, these results pointed to lexical similarity as a possible facilitator of word acquisition

Lexical similarity modulates vocabulary acquisition in bilinguals

A cognate facilitation in lexical acquisition?

Cognates: Phonologically-similar translation equivalents

Cognate	Non-cognate
[cat] /ˈgat-ˈga.to/	[dog] /ˈgos-ˈpe.ro/

Some evidence that cognates acquired earlier than non-cognates (Mitchell, Tsui, and Byers-Heinlein 2023; Bosch and Ramon-Casas 2014; Tan, Marchman, and Frank 2024; Siow et al. 2022)

What mechanisms support a cognate facilitation during word acquisition?

Lexical access is language non-selective in bilinguals

Current explanations of this effect stem from the language non-selective account of the bilingual lexicon
This account states that during language comprehension and production, lexical representations from both languages are activated, even during monolingual situations
For instance, when a Catalan-Spanish bilingual hears the Catalan word cotxe, not only the lexical representation of cotxze is activated, but also that of other phonologically related words in Catalan and Spanish
There is extensive evidence in adults that this activation, which also percolates to semantic representations, influences the dynamics of word comprehension and production.
There is some evidence of such parallel activation in infants.
It has been suggested that this parallel activation of representations in both languages may facilitate the acquisition of cognates in infants, given that they are more strongly co-activated, given their higher similarity, compared to non-cognates.
But this effect still lacks a mechanistic account

The present dissertation

Study 1

Provide a mechanistic account for the cognateness facilitation
Test predictions of the model

Under review in Child Development (R2),

Study 2

Test core assumption of the model: Language non-selectivity in the initial lexicon

In preparation

Study 1

Cognate beginnings to lexical acquisition: The AMBLA model

Accumulator Model of Bilingual Lexical Acquisition (AMBLA)

Accumulation of information about form-meaning mappings:

Learning instances: Exposure to a word-form that results in the accumulation of information about its meaning

Age of acquisition: The infant accumulates a threshold amount of learning instances for a word-form

\[ \begin{aligned} \definecolor{myred}{RGB}{ 168, 0, 53 } \definecolor{myblue}{RGB}{ 0, 64, 168 } \definecolor{mygreen}{RGB}{0, 168, 87} \definecolor{grey}{RGB}{128, 128, 128} \textbf{For participant } &i \textbf{ and word-form } j \text{ (translation of } j'): \\ {\color{mygreen}\text{Age of Acquisition}_{ij}} &= \{\text{Age}_i \mid {\color{myred}\text{Learning instances}_{ij}} = {\color{myblue}\text{Threshold}} \}\\ {\color{myred}\text{Learning instances}_{ij}} &= \text{Age}_i \cdot \text{Freq}_j \\ \end{aligned} \]

I present here the Accumulator Model of Bilingual Lexical Acquisition or AMBLA
This is an accumulator model, in that it assumes that infants learn words by gradually accumulating information about word-referent mappings.
Information is collected by the child in the form of learning instances
We define learning instances as exposure to tokens of the word that result in the successful accumulation of information about their meaning
We assume that an infant acquires a word when they have accumulated some threshold amount of learning instances with
The age at which this occurs is the acge of acquisition of the word
The rate at which an infant accumulates learning instances is a function of two variables: they age (older infants will have accumulated more learning instance than younger infants), and the lexical frequency of the words (infants will accumulate learning instances faster for more frequent words than for less frequent words)

AMBLA: Simulating monolingual word acquisition

Catalan monolingual child

/’gos/ (Catalan), 100%

Parameters:

\[ \begin{aligned} \text{Threshold} = 300 \\ \text{Freq}_j \sim \text{Poisson}(\lambda = 50) \end{aligned} \]

AMBLA: Simulating monolingual word acquisition

Catalan monolingual child

/’gos/ (Catalan), 100%

Parameters:

\[ \begin{aligned} \text{Threshold} = 300 \\ \text{Freq}_j \sim \text{Poisson}(\lambda = 50) \end{aligned} \]

AMBLA: Simulating bilingual word acquisition

Linguistic input divided into two languages: Catalan 60%, Spanish 40%

Exposure: Proportion of time exposed to the language of \(j\) word

Accumulation of learning instances, a function of Exposure and Frequency.

AMBLA: Simulating bilingual word acquisition

Catalan monolingual child

/’gos/ (Catalan), 100%

Catalan/Spanish bilingual child

/’gos/ (Catalan), 60%
/’pe.ro/ (Spanish), 40%

Parameters:

\[ \begin{aligned} \text{Threshold} = 300 \\ \text{Freq}_j \sim \text{Poisson}(\lambda = 50) \end{aligned} \]

AMBLA: Simulating bilingual word acquisition

Catalan monolingual child

/’gos/ (Catalan), 100%

Catalan/Spanish bilingual child

/’gos/ (Catalan), 60%
/’pe.ro/ (Spanish), 40%

Parameters:

\[ \begin{aligned} \text{Threshold} = 300 \\ \text{Freq}_j \sim \text{Poisson}(\lambda = 50) \end{aligned} \]

AMBLA: Simulating a cognate facilitation

Words may accumulate additional learning instances from the co-activation of their (phonologically similar) translation equivalent

Degree proportional to their phonological similarity (Cognateness)

\[ \begin{aligned} \textbf{For participant } &i \textbf{ and word-form } j \text{ (translation of } j'): \\ \text{Age of Acquisition}_{ij} &= \{\text{Age}_i \mid \text{Learning instances}_{ij} = \text{Threshold} \}\\ \text{Learning instances}_{ij} &= \text{Age}_i \cdot \text{Freq}_j \cdot \text{Exposure}_{ij} + \\ &({\color{myred}\text{Learning instances}_{ij'} \cdot {\text{Cognateness}}_{j}})\\ \textbf{where:} \\ {\color{myred}\text{Cognateness}_{j,j'}}&{\color{myred} = \text{Levenshtein}(j, j')} \end{aligned} \]

Following the language non-selective account of the bilingual lexicon, we assume that infants co-activate translation equivalents in both languages, even in monolingual situations
The strength of this co-activation is a function of the phonological similarity between both word-forms
During language exposure cognate words will be more strongly co-activated than non-cognate words
As a result of this co-activation, when an infant encounters a given word-form in their speech inpt, they may not only accumulate a learning instance for such word, but also for its translation equivalent
The degree to which the translation accumulates this additional learning instance is a function of its phonological similarity between both word-forms
This is expressed in the model by adding a new term to the learning instance accumulator
This term adds the number of learning isntances that the translation eequivalent of the word has acquired, weighted by the phonological simlarity between them
This phonological similarity is measured using Levenshtein distance between both phonological transcriptions
As a result, identical cognates will accumulate cross-lingusitic learning isntances with every exposure to either translation, while words with no phonological similarity will not benefit from any additional learning instances

AMBLA: Simulating a cognate facilitation

Catalan monolingual child

/’gat/ (Catalan), 100%

Catalan/Spanish bilingual child

/’gat/ (Catalan), 60%
/’ga.to/ (Spanish), 40%

Parameters:

\[ \begin{aligned} \text{Threshold} = 300 \\ \text{Freq}_j \sim \text{Poisson}(\lambda = 50) \\ \text{Cognateness}_{j,j'} = 0.75 \end{aligned} \]

AMBLA: Simulating a cognate facilitation

Catalan monolingual child:

/’gat/ (Catalan), 100%

Catalan/Spanish bilingual child:

/’gat/ (Catalan), 60%
/’ga.to/ (Spanish), 40%

Parameters:

\[ \begin{aligned} \text{Threshold} = 300 \\ \text{Freq}_j \sim \text{Poisson}(\lambda = 50) \\ \text{Cognateness}_{j,j'} = 0.75 \end{aligned} \]

Predictions

Cognates acquired earlier than non-cognates
Cognateness facilitation stronger in the lower-exposure language

Predictions

Cognates acquired earlier than non-cognates
Cognateness facilitation stronger in the lower-exposure language

Barcelona Vocabulary Questionnaire (BVQ)

Online, open source
\(\approx\) 1,600 words (800 Cat., 800 Spa.)
4 sublists, random allocation

Results: Comprehension

366 participants (12-32 mo), 436 administrations \(\times\) 604 noun words

Ordinal, multilevel (Bayesian) regression model

\(p(\text{Comprehension}, \text{Production}) \sim \text{Exposure}_{ij} \cdot \text{Cognateness}_j\)

We gathered 436 questionnaire administrations from 366 12-32 month-old infants learning Catalan and or Spanish in the Metropolitan Area of Barcelona
We modelled the probabiity that each child would understand or produce each of the words
We used on ordinal Bayesian multilevel model that included several predictors of interest
The critical predictors were the main effect of \(Cognateness\), and its interaction with \(Exposure\)
First I am going to present the results for comprehension
In this figure I am showing the posterior predicted probability of comprehension across ages for three average word-forms
I am showing predictions for an identical cognate with 100% phonological similarity with its translation
I am also showing the predictions for a non-cognate with 50% similarity, and 0% similarity
Predictions are generated for word acquisition in the lower-exposure language (to the left) and for the higher-exposure language (to the right)
In the lower-exposure language, we observed a substantial facilitation effect of cognateness: the probability of acquisition grows faster for cognates
In the higher-exposure language, this effect was not observed

Results: Production

366 participants (12-32 mo), 436 administrations \(\times\) 604 noun words

Ordinal, multilevel (Bayesian) regression model

\(p(\text{Comprehension}, \text{Production}) \sim \text{Exposure}_{ij} \cdot \text{Cognateness}_j\)

Discussion

Earlier acquisition for cognates vs. non-cognates

Cognate facilitation moderated by exposure

Only words from the lower exposure benefit from cognateness

Cognateness as a candidate mechanism underlying Floccia et al.’s results

Cross-language facilitation via co-activation of phonologically similar translation equivalents

Is language-non selectivity already present in the initial lexicon?

In summary, we found an earlier acquisition of cognates than non-cognates, in line with the predictions of AMBLA
This effect was modulated by quantitative language-exposure
Only the acquisition of words from the lower-exposure language benefitted from their cognate status
This cognate facilitation provides an candiate mechanism for the results in Floccia et al.
We presented a mechanistic account for this effect in the form of a model, the AMBLA model
In this model, we propose that the cognate facilitation effect is the result of a cross-language accumulation of learning instances, in line with the language non-selective account of the bilingual lexicon
One core assumption of this model is that bilingual infants activate phonological word-forms in both languages in parallel, even in monolingual situations
In Study 2, we tested the plausibility of this assumption

Study 2

Developmental trajectories of bilingual spoken word recognition

Language non-selectivity in the initial lexicon

Some evidence in infants and children (e.g., Von Holzen and Mani 2012; Singh 2014)

Methodological pitfalls: “Bilingual” task

One language is task relevant, the other is covertly activated

Implicit naming task

Mani and Plunkett (2010, 2011)

Implicit naming task

Mani and Plunkett (2010, 2011)

Chance-level target looking in related trials
Prime-Target phonological interference
Implicit naming of prime pictures

Implicit naming task

Study 2: Design

Cross-language priming effects are short-lived

Change in design:

Auditory label before target-distractor images

Increased temporal proximity of prime and target

Study 2: Design

Data collection timeline

Predictions

Exp. 1: Monolinguals

Replicate within-language phonological interference from Mani and Plunkett (proof of concept)

Exp. 2: Monolinguals and bilinguals

If language non-selectivity, stronger interference in cognate vs. non-cognate trials

Data collection timeline

Experiment 1: Results, Bayesian GAMMs

English monolinguals

79 participants, 89 sessions

No evidence of phonological priming

Related trials \(\approx\) Unrelated trials

Experiment 2: Results, Bayesian GAMMs

Catalan/Spanish monolinguals

77 participants, 107 sessions

No evidence of phonological priming

Related trials \(\approx\) Unrelated trials Cognate trials \(\approx\) Non-cognate trials

Experiment 2: Results, Bayesian GAMMs

Catalan/Spanish bilinguals

78 participants, 133 sessions

No evidence of phonological priming

Related trials \(\approx\) Unrelated trials Cognate trials \(\approx\) Non-cognate trials

Discussion

Successful spoken word recognition across ages and language profiles

No evidence of priming effects, within or across languages

Unsuccessful retrieval of prime phonological forms?

Inconclusive results, revise design

General discussion

Summary

Cognateness facilitates word acquisition in the lower-exposure language

Candidate mechanism behind bilingual vocabulary growth

AMBLA: Cross-language accumulation of learning instances

Language non-selectivity in the initial lexicon: Pending testing

Discussion

Cognateness facilitation effect (Siow et al. 2022; Mitchell, Tsui, and Byers-Heinlein 2023; Tan, Marchman, and Frank 2024)
Candidate explanation for Floccia et al. (2018) and Hoff et al. (2012)
Standard Model of Language Acquisition (Kachergis, Marchman, and Frank 2022)

Future steps

The impact of cognateness in spoken word recognition: Re-analysing data from Study 2
More language pairs (lower overall similarity)
Train AMBLA model on vocabulary data

Whats not in this dissertation

Backward Semantic Inhibition

The emergence of inhibitory links in the initial lexicon

Vocabulary growth through the lens of bilingualism

Data collection ongoing

Whats not in this dissertation

Translation Elicitation

Levenshtein distance as a valid measure of word-level effects of phonological similarity

Monolingual participants listening to a non-native language

jtracer package

Methodological contributions

Sample size (N > 400)
Bayesian modelling: Quantifying uncertainty, estabilising statistical inference
Barcelona Vocabulary Questionnaire (BVQ)

bvq package +

Thank you!

Appendix

Introduction: Bilingualism

Classification of participants into monolinguals an bilinguals

Introduction: Cognate contents in the aggregated vocabulary

Cognate contents in the aggregated vocabulary

Study 1: Posterior regression coefficients

Aggregated vocabularies might conceal facilitation effects

Study 1: MCMC convergence (\(\hat{R}\))

MCMC convergence for the model in Study 1

Study 2: Predictions

Successful spoken word recognition across groups
If language non-selectivity, stronger interference in cognate vs. non-cognate trials

Study 2: Vocabulary size

Study 2 participant receptive vocabulary sizes across ages and language profiles

Study 2: Model convergence (Exp. 1)

MCMC convergence for model in Study 1 (Exp. 1)

Study 2: Model convergence (Exp. 2)

MCMC convergence for model in Study 2 (Exp. 1)

References

Bergelson, Elika, and Daniel Swingley. 2012. “At 69 Months, Human Infants Know the Meanings of Many Common Nouns.” Proceedings of the National Academy of Sciences 109 (9): 3253–58. https://doi.org/10.1073/pnas.1113380109.

Bosch, Laura, and Marta Ramon-Casas. 2014. “First Translation Equivalents in Bilingual Toddlers’ Expressive Vocabulary: Does Form Similarity Matter?” International Journal of Behavioral Development 38 (4): 317–22. https://doi.org/10.1177/0165025414532559.

Fenson, Larry, Philip S Dale, J Steven Reznick, Elizabeth Bates, Donna J Thal, Stephen J Pethick, Michael Tomasello, Carolyn B Mervis, and Joan Stiles. 1994. “Variability in Early Communicative Development.” Monographs of the Society for Research in Child Development 59 (5): 1–185. https://doi.org/10.2307/1166093.

Frank, Michael C., Mika Braginsky, Daniel Yurovsky, and Virginia A. Marchman. 2017. “Wordbank: An Open Repository for Developmental Vocabulary Data.” Journal of Child Language 44 (3): 677–94. https://doi.org/10.1017/s0305000916000209.

Kachergis, George, Virginia A. Marchman, and Michael C. Frank. 2022. “Toward a ‘Standard Model’ of Early Language Learning.” Current Directions in Psychological Science 31 (1): 20–27. https://doi.org/10.1177/09637214211057836.

Mitchell, Lori, Rachel Ka-Ying Tsui, and Krista Byers-Heinlein. 2023. “Cognates Are Advantaged over Non-Cognates in Early Bilingual Expressive Vocabulary Development.” Journal of Child Language, 1–20.

Singh, Leher. 2014. “One World, Two Languages: Cross-Language Semantic Priming in Bilingual Toddlers.” Child Development 85 (2): 755–66. https://doi.org/10.1111/cdev.12133.

Siow, Serene, Nicola A Gillen, Irina Lepadatu, Daniela S Avila-Varela, Gonzalo Garcia-Castro, Nuria Sebastian-Galles, and Kim Plunkett. 2022. “The Effect of Cognates on Bilingual Infant Vocabulary Trajectories: A Study Using Bilingual CDIs of English and One Additional Language.” In Proceedings of the 46th Annual Boston University Conference on Language Development. Boston, MA.

Tan, Alvin WM, Virginia A Marchman, and Michael C Frank. 2024. “The Role of Translation Equivalents in Bilingual Word Learning.” Developmental Science, e13476.

Tincoff, Ruth, and Peter W Jusczyk. 1999. “Some Beginnings of Word Comprehension in 6-Month-Olds.” Psychological Science 10 (2): 172–75. https://doi.org/10.1111/1467-9280.00127.

Von Holzen, Katie, and Nivedita Mani. 2012. “Language Nonselective Lexical Access in Bilingual Toddlers.” Journal of Experimental Child Psychology 113 (4): 569–86. https://doi.org/10.1016/j.jecp.2012.08.001.