ArticlePDF Available

When exceptions matter: bilinguals regulate their dominant language to exploit structural constraints in sentence production

Article

When exceptions matter: bilinguals regulate their dominant language to exploit structural constraints in sentence production

Abstract

What we say generally follows distributional regularities, such as learning to avoid "the asleep dog" because we hear "the dog that's asleep" in its place. However, not everyone follows such regularities. We report data on English monolinguals and Spanish-English bilinguals to examine how working memory mediates variation in a-adjective usage (asleep, afraid), which, unlike typical adjectives (sleepy, frightened), tend to resist attributive use. We replicate previous work documenting this tendency in a sentence production task. Critically, for all speakers, the tendency to use a-adjectives attributively or non-attributively was modulated by individual differences in working memory. But for bilinguals, a-adjective use was additionally modulated by an interaction between working memory and category fluency in the dominant language (English), revealing an interactive role of domain-general and language-related mechanisms that enable regulation of competing (i.e. attributive and non-attributive) alternatives. These results show how bilingualism reveals fundamental variation in language use, memory, and attention.
Full Terms & Conditions of access and use can be found at
https://www.tandfonline.com/action/journalInformation?journalCode=plcp21
Language, Cognition and Neuroscience
ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/plcp21
When exceptions matter: bilinguals regulate
their dominant language to exploit structural
constraints in sentence production
Christian A. Navarro-Torres, Paola E. Dussias & Judith F. Kroll
To cite this article: Christian A. Navarro-Torres, Paola E. Dussias & Judith F. Kroll
(2022): When exceptions matter: bilinguals regulate their dominant language to exploit
structural constraints in sentence production, Language, Cognition and Neuroscience, DOI:
10.1080/23273798.2022.2105915
To link to this article: https://doi.org/10.1080/23273798.2022.2105915
View supplementary material
Published online: 03 Aug 2022.
Submit your article to this journal
View related articles
View Crossmark data
REGULAR ARTICLE
When exceptions matter: bilinguals regulate their dominant language to exploit
structural constraints in sentence production
Christian A. Navarro-Torres
a
, Paola E. Dussias
b,c
and Judith F. Kroll
d
a
Department of Psychology, Princeton University, Princeton, NJ, USA;
b
Department of Spanish, Italian, and Portuguese, The Pennsylvania State
University, State College, PA, USA;
c
Center for Language Science, The Pennsylvania State University, State College, PA, USA;
d
School of
Education, University of California, Irvine, Irvine, CA, USA
ABSTRACT
What we say generally follows distributional regularities, such as learning to avoid the asleep dog
because we hear the dog thatsasleepin its place. However, not everyone follows such regularities.
We report data on English monolinguals and Spanish-English bilinguals to examine how working
memory mediates variation in a-adjective usage (asleep,afraid), which, unlike typical adjectives
(sleepy,frightened), tend to resist attributive use. We replicate previous work documenting this
tendency in a sentence production task. Critically, for all speakers, the tendency to use a-
adjectives attributively or non-attributively was modulated by individual dierences in working
memory. But for bilinguals, a-adjective use was additionally modulated by an interaction between
working memory and category uency in the dominant language (English), revealing an
interactive role of domain-general and language-related mechanisms that enable regulation of
competing (i.e. attributive and non-attributive) alternatives. These results show how bilingualism
reveals fundamental variation in language use, memory, and attention.
ARTICLE HISTORY
Received 8 September 2021
Accepted 8 July 2022
KEYWORDS
Sentence production;
bilingualism; individual
dierences; working
memory; category uency
Introduction
A hallmark feature of human speech is the ability to
produce novel utterances to describe events and
internal thoughts in a dynamic manner, while following
the grammatical regularities of a language. The form of
the utterances we speak is generally constrained by
semantic, phonological, and grammatical features of
the known language. However, spoken utterances can
at times be well formed and yet strongly dispreferred
for seemingly arbitrary reasons. For example, native
speakers of English might tend to avoid using the
asleep dog in favour of the dog thats asleep, despite
the fact that adjectives are typically used attributively
in English (Goldberg & Boyd, 2015; Yang, 2015). This pre-
ference showcases that multiple morphosyntactic forms
can serve similar functions, though they are not fully
interchangeable (see Labov, 1969; Poplack & Torres
Cacoullos, 2015; Sanko,1988 for discussion on form-
function asymmetries). However, despite having well-
established form-function asymmetries, individuals
may vary in the choices they make among competing
alternatives during conversation, as showcased in the
following examples taken from the Corpus of Contem-
porary American English (COCA):
She was the most alive person I knew. And I the most
alone. (Davies, 2008)
An awake patient is more complicated that an asleep
patient. (Davies, 2008)
Such variation may not be random but may instead
reect divergent and systematic tendencies of how
speakers engage linguistic and cognitive processes
(Fricke et al., 2019; Green et al., 2006). However, little is
known about the underlying processes that regulate
such choices in production.
Related to this question is the contention of whether
individualslinguistic choices are determined by factors
unique to the language system or whether they
emerge from domain-general principles (Bybee, 2012;
Hauser et al., 2002; Koranda et al., 2020; Vicari & Aden-
zato, 2014). The evidence has been mixed, with some
research suggesting that language draws from
domain-general control processes (e.g. Hsu et al., 2021;
Nair et al., 2021; Nozari & Novick, 2017; Swets et al.,
2007). Others have found little-to-no overlap (e.g.
Acheson & Hagoort, 2014; Engelhardt et al., 2017; Fedor-
enko et al., 2011; Van Dyke et al., 2014) and argue for a
strong division between the two (Fedorenko, 2014).
Part of the issue may be the framing of the question,
© 2022 Informa UK Limited, trading as Taylor & Francis Group
CONTACT Christian A. Navarro-Torres cn7491@princeton.edu Department of Psychology, Princeton University, Peretsman Scully Hall, Room 126,
Princeton, NJ 08540, USA
Supplemental data for this article can be accessed online at https://doi.org/10.1080/23273798.2022.2105915.
LANGUAGE, COGNITION AND NEUROSCIENCE
https://doi.org/10.1080/23273798.2022.2105915
which creates a false dichotomy between language and
other cognitive phenomena. As a result, very few studies
have empirically examined a more likely alternative: that
patterns of language use, and their processing, emerge
from an interactive coordination between distinct but
interrelated, cognitive processes, some of which may
be more domain-specic, but others of which will be
domain-general (Hsu et al., 2017; Ihlen & Vereijken,
2010; Van den Heuvel & Sporns, 2013). Such an alterna-
tive would have notable implications for research prac-
tice concerning replicability and generalisability of
eects (Wallot & Kelty-Stephen, 2018).
In the present study, we propose a regulatory model
in which speakers exploit and regulate form-function
asymmetries by interactively coordinating language-
related and domain-general processes. We refer to
such coordination as the construction-based regulation
hypothesis.We argue that speakers need construction-
based regulation given that constraints on memory
and attention can yield systematic variation in the
alternatives that become most accessible when planning
speech. We exploit this hypothesis in the context of a-
adjectives and propose that individuals regulate two
opposing, but interdependent, functional states: an attri-
butive state in which a-adjectives and typical adjectives
are used similarly on the basis of their function (i.e.
saying the asleep doggiven that a-adjectives modify
nouns in the same way typical adjectives do), and a
non-attributive state in which a-adjectives and typical
adjectives are treated dierently (i.e. knowing that a-
adjectives do not typically appear in attributive pos-
ition). We propose that eectively regulating these two
states enables statistical preemption, the process via
which speakers learn to avoid using a form by witnes-
sing a functionally related alternative in its place (Boyd
& Goldberg, 2011; Goldberg, 2016; Goldberg & Boyd,
2015). Key to understanding statistical preemption, in
our view, is working memory and the regulation of
attention, as established by functional models of
working memory (Cowan, 2005; Oberauer, 2002; Ship-
stead et al., 2016).
To test the construction-based regulation hypothesis,
we use a sentence production task and examine vari-
ation in the use of a-adjectives, which, unlike typical
adjectives, tend to resist attributive use. We rst
present group-level analyses replicating previously
reported eects on a-adjectives (Boyd & Goldberg,
2011) in a group of monolingual speakers of English
and a second group of English-dominant bilinguals
who speak Spanish as the home language. More impor-
tantly, we then examine individual dierences in verbal
uency and working memory to show that group-level
analyses on their own are insucient to identify how
language use draws from cognitive resources. Speci-
cally, we nd distinct patterns of association in each
group of speakers and show that only the variation
found in the bilingual group fully reveals evidence for
construction-based regulation in sentence production.
We argue that this is because bilinguals must actively
engage a language regulatory mechanism to manage
accessibility of their two languages when speaking and
that this cross-language regulatory mechanism impacts
how form-function asymmetries are exploited.
In the remainder of the introduction, we review the
evidence for categorisation and statistical preemption
following a usage-based linguistic framework to
account for form-function asymmetries in a-adjective
use. We then examine working memory and how it
may mediate accessibility in a-adjective usage. Finally,
we examine the evidence for a bilingual language regu-
latory mechanism, its importance for revealing inter-
actions between language-related and domain-general
processes, and its impact on accessibility.
Categorisation and statistical preemption
Usage-based approaches argue that language is an
emergent property of cognition and experience. In this
sense, grammatical constructions should not be taken
as rule-based features. Instead, they are dened as
form-meaning mappings with dierent degrees of
abstraction. They are employed through language use
and become conventionalised in the speech community
(Bybee, 2010,2013; Goldberg, 2013; Ibbotson, 2013).
Constructions can, therefore, range from morphemes
and single words to larger grammatical structures, and
phrasal units such as idioms. Constructions are pre-
sumed to arise from experience-dependent processes
that are domain-general, including frequency, relevance,
and priming (Goldberg & Ferreira, 2022), as well as
analogy, rich memory storage, and chunking (Bybee,
2010). These processes can shape accessibility of linguis-
tic representations, as well as the representations them-
selves. Critically, given that grammatical constructions
involve the same form-meaning associations as words
and morphemes, they are subject to the same cognitive
processes as lexical expressions.
Here we focus on the role of categorisation and stat-
istical preemption as experience-dependent processes.
Categorisation denotes the ability to cluster exemplars
(e.g. words or constructions) through comparison of pre-
viously established categories based on relevant linguis-
tic features (Bybee, 2010; Goldberg, 2016). To illustrate,
L2 learners of English may initially infer that the verb
explain can be used in both dative (e.g. Explain this to
me) and double-object (e.g. Explain me this)
2C. A. NAVARRO-TORRES ET AL.
constructions because many other transitive verbs do so
(e.g. Give/Send/Bring me the lettervs. Give/Send/
Bring the letter to me). Similarly, they may initially
infer that a-adjectives ought to be used as other adjec-
tives (i.e. saying the asleep dogbecause adjectives
are generally used attributively).
Statistical preemption, on the other hand, refers to an
inferential process in which continually experiencing a
novel linguistic form in particular contexts precludes
the use of an established form when both alternatives
share equivalent (or similar) meaning and function. To
further illustrate using the explain example, although
L2 learners could initially expect to encounter the verb
in both the dative and double-object constructions,
they may gradually learn to disprefer the use of the
latter as they only encounter instances of the former.
In a similar manner, prenominal instances should
become dispreferred with a-adjectives as L2 learners
hear functionally related alternatives (i.e. the dog
thats asleep) in the same contexts.
What is key from these examples is that categoris-
ation enables individuals to abstract common features
from dierent functionally related forms (e.g. all ditransi-
tive verbs can alternate between dative and double-
object constructions, or all modiers can be used attri-
butively), and that individuals can learn to constrain
that generalisation by identifying the contexts in
which alternate competing forms do and do not occur
(e.g. some verbs, such as explain, only appear in one con-
struction, even though an L2 learner may initially think it
could appear in either), provided they receive enough
input to do so.
The case of a-adjectives
The description of categorisation and statistical preemp-
tion is key for characterising patterns of language use
that cannot be easily explained by simple rules, such
as in the case of a-adjectives, the focus of the present
study. A-adjectives begin with a syllabic schwa (/ə/; rep-
resented orthographically by unstressed a), are mor-
phologically segmentable (a+ stem), are no longer
than two syllables (e.g. a/sleep, a/live, a/oat), and
contain meaningful stems related to near-synonym
adjectives (e.g. sleepy, lively/living, oating). Speakers
may store information about these morphophonological
features to generate an a-adjective network as a subca-
tegory of a more general adjectival network.
Evidence for statistical preemption with a-adjectives
comes from a production study by Boyd and Goldberg
(2011). English monolingual speakers were asked to
describe sequences of pictures containing adjectival
(real and novel) labels that would prompt the use of
attributive (i.e. prenominal) and non-attributive (i.e. rela-
tive clause) constructions. The results showed that, when
the target labels contained real a-adjectives, speakers
were more likely to produce non-attributive responses
(i.e. the dog thats asleep) than with typical adjective
labels, where they mostly produced attributive
responses (i.e. the sleepy dog). Additionally, when
the target labels were novel words that resembled real
a-adjectives (i.e. novel a-adjectives; e.g. ablim), a similar
pattern emerged (i.e. the dog thats ablim), although
to a lesser degree. These results were taken to support
the presence of categorisation: speakers partially gener-
alised the a-adjective canonical pattern to novel
members whose morphophonological features
resembled those seen in real a-adjectives.
In a follow-up experiment, the authors asked a new
set of participants to perform the same task, only this
time, participants experienced an exposure block that
contained a preemptive context: witnessing a few
instances of novel a-adjectives in relative clause position.
Although performance with real labels was similar to the
original experiment, the pattern of results for novel a-
adjectives changed, such that participants strongly disfa-
voured the use of attributive constructions with novel
a-adjective labels.
The ndings reported by Boyd and Goldberg (2011)
suggest that a combination of categorisation and stat-
istical preemption shape form-function asymmetries.
However, given that canonical a-adjective usage
follows tendencies rather than categorical constraints
(i.e. in Boyd and Goldberg (2011), real a-adjective pre-
nominal rates varied between .14 and .31 across exper-
iments), it is not entirely clear how statistical
preemption accounts for variation regarding these pre-
ferences. Speakers often vary in their ability to form
abstractions and categories (Dąbrowska, 2012) and in
how they plan speech (Mortensen et al., 2008; Swets
et al., 2014,2021). Here we consider the role of
working memory as an important source of variability
in a-adjective usage.
The role of working memory in language use
An integral part of language production is working
memory (Martin & Slevc, 2014), which we dene as the
active manipulation and maintenance of goal-relevant
information while avoiding interference from long-
term memory and/or previously relevant information.
Since the statistical preemption hypothesis states that
accessibility in production is shaped by competition in
context (Goldberg, 2019), working memory is likely criti-
cal to regulate that competition when planning speech.
Though there is substantial psycholinguistic research
LANGUAGE, COGNITION AND NEUROSCIENCE 3
examining working memory in language comprehen-
sion (e.g. Cunnings, 2017; Hopp, 2014; Kim et al., 2018;
King & Just, 1991; Lewis & Vasishth, 2005; Sprouse
et al., 2012; Van Dyke et al., 2014), very little research
has examined the role of working memory in production
(see Martin & Slevc, 2014 for a review). Further, although
memory is considered an important factor in usage-
based linguistic theories, it is usually delimited to chunk-
ing, the ability to integrate and store meaningful units in
a hierarchical fashion (Bybee, 2010; Green, 2017). Con-
versely, less attention has been given to more active
aspects of memory (i.e. working memory) and how
they mediate language use.
Broadly speaking, there are two classes of working
memory models: structured-oriented models (e.g. Bad-
deley, 2000; Baddeley & Hitch, 1974) and processing or
functional-oriented models (e.g. Cowan, 2005; Oberauer,
2002; Shipstead et al., 2016). Here we focus on the latter
class of models, particularly on the Concentric Model of
Working Memory proposed by Oberauer (2002,2019). In
the Concentric Model, there are three levels, or represen-
tational states, on which working memory operates:
(i) an activated part of long-term memory, (ii) a region
of direct access that exibly retains a subset of the acti-
vated content to be used, and (iii) a focal point of atten-
tion embedded in the region of direct access that holds
and selects one representation at a time. The region of
direct access allows for the temporary creation of new
associations between the focus of attention and acti-
vated representations from long-term memory. But
increasing the number of bindings in the region of
direct access will increase the probability of interference
in the focal point, resulting in a limited capacity storage
system that is subject to individual dierences.
We hypothesise that working memory is central to
the process of statistical preemption and that the Con-
centric Model is key in characterising that relation. Con-
sider an inexperienced L2 speaker of English who wishes
to use an a-adjective as a modier. Initially, the speaker
may be tempted to use it attributively (the focal point of
attention) by accessing the adjectival network from
long-term memory (the region of direct access). But
the speaker encounters instances of non-attributive con-
structions in contexts where attributive usage is
expected. This will generate new temporary bindings
in the region of direct access between previously estab-
lished (i.e. attributive) and newly acquired (i.e. relative
clause) forms for a-adjectives. Repeated instances of
relative clause constructions should preempt the attri-
butive construction, but in the process, proactive inter-
ference from long-term memory (i.e. an attributive
state from the adjectival network) may hinder preemp-
tion through the region of direct access. As such,
speakers with reduced working memory may use a-
adjectives attributively despite knowingthat a-adjec-
tives disprefer attributive placement. Further, retrieval
failures due to proactive interference would make
these individuals less able to extend the constrain to
new lexical members of the a-adjective category.
This scenario highly resembles the demands imposed
by working memory tasks such as the Operation-span
(O-span) task (Turner & Engle, 1989), in which lists of
items of various lengths need to be memorised for
later recall while solving mathematical equations. Main-
tenance of an item list in the focal point of attention can
become disrupted due to proactive interference from
previously relevant procedures (i.e. solving the
equations) and information (i.e. previously recalled
lists). Thus, examining individual dierences in working
memory may be useful to understand the nature of vari-
ation in a-adjective usage.
Bilingual language regulation
Arguably, the goal of speakers is to learn which forms
serve which functions in the language they speak (Gold-
berg, 2019). If this is the case, then individuals who grow
up speaking more than one language face a unique chal-
lenge: bilinguals need to learn an extended range of
form-function asymmetries, some of which will converge
across languages, but others of which will diverge
(Poepsel & Weiss, 2016). Learning to dierentiate and
constrain such distributions becomes critical to eec-
tively communicate in a given language. This raises the
question of how bilinguals are able to attain and main-
tain high linguistic skill in ways comparable to
monolinguals.
Unlike monolinguals, bilinguals are often required to
negotiate cross-language activation when speaking
(e.g. Martin & Nozari, 2021; for reviews, see Costa, 2005;
Hanulová et al., 2011; Hartsuiker & Bernolet, 2017; Kroll
et al., 2006,2012). Cross-language activation makes bilin-
guals susceptible to cross-language interference when
planning speech (Abutalebi & Green, 2007). Converging
evidence indicates that bilinguals manage such
demands in part by drawing from domain-general
control processes (e.g. Abutalebi et al., 2008; Baum &
Titone, 2014; Beatty-Martínez et al., 2020; Linck et al.,
2008; Pivneva et al., 2012). Thus, an important source of
variability in bilingual speakers may lie in their ability to
manage the accessibility of their two languages. For
instance, access to the dominant L1 is typically disrupted
after speaking in the weaker L2 (e.g. Guo et al., 2011;
Misra et al., 2012; Rossi et al., 2018; Van Assche et al.,
2013) or during prolonged immersion in an L2 environ-
ment (Baus et al., 2013; Linck et al., 2009).
4C. A. NAVARRO-TORRES ET AL.
These patterns of performance indicate that when
bilinguals speak in the weaker language, they actively
down-regulate the dominant L1 (Green, 1998; Kroll
et al., 2008), making it momentarily less accessible.
Shifts in L1 accessibility can often occur when bilinguals
engage in diverse interactional contexts where some
individuals only use one language or where there is
restricted use of the L1 (Baus et al., 2013; Beatty-Martínez
et al., 2020; Beatty-Martínez & Titone, 2021; Linck et al.,
2009). Thus, for bilinguals to eectively communicate
in one language at any given point, regulation of the
dominant language seems crucial. Studies examining
the neural basis of these regulatory eects indicate
that domain-general cognitive control networks are
involved (Branzi et al., 2016; Misra et al., 2012; Rossi
et al., 2018; Zhang et al., 2021). We refer to the ability
to coordinate an adequate balance between down-regu-
latory (i.e. suppression of highly accessible information)
and up-regulatory processes (i.e. recovery from that sup-
pression) as bilingual language regulation.
1
Recent research suggests that regulation of the domi-
nant language is key for attaining high linguistic skill in
L2 learners (Bice & Kroll, 2015; Pulido & Dussias, 2020)
and in procient bilinguals (Bogulski et al., 2019; Zirn-
stein et al., 2018). Of particular relevance here is the
study by Zirnstein et al. (2018), who examined electro-
physiological responses to semantically expected and
unexpected words during sentence reading in English
(e.g. After their meal, they forgot to leave a ten/tip for
the waitress) in monolinguals and in Mandarin-domi-
nant bilinguals who spoke English as the L2. Group-
level results showed that only monolinguals reliably
generated prediction costs when encountering unex-
pected words. However, individual dierence analyses
revealed that bilinguals were in fact able to generate
L2 predictions and recover from them, but that these
eects depended on an interaction between cognitive
control and category uency in the dominant L1 (Man-
darin). Specically, more ecient cognitive control was
associated with reduced prediction costs in the L2, but
this was particularly true in bilinguals who had high L1
category uency (i.e. those who had greater up-regulat-
ory abilities in the dominant language).
There are three key points regarding the interaction
between the cognitive and uency measures reported
by Zirnstein et al. First, the interaction indicates how
dierent patterns of association between language
and cognitive resources can emerge in dierent
groups of speakers. Second, regulation emerges from
an interaction of domain-general and language-related
processes. Finally, the eect of L1 uency suggests
that up-regulation of the dominant language is key to
understand how bilinguals engage domain-general
control processes during language processing. Notably,
however, the past studies examining bilingual language
regulation have focused on lexical processes (i.e. Bice &
Kroll, 2015; Bogulski et al., 2019; Zirnstein et al., 2018; c.f.
Pulido & Dussias, 2020). Here we examine bilingual
language regulation, as measured by the interaction
between working memory and category uency in the
dominant language, and its consequences for syntactic
variation in the use of a-adjectives.
The present study
In the present study, we investigate the case of a-adjec-
tives in monolingual speakers of English and in Spanish-
English bilinguals who speak English as the dominant
language. We adopt the paradigm reported by Boyd
and Goldberg (2011) using a sentence-production task
to examine speakersstructural preferences (i.e. prenom-
inal vs. relative clause placement) with a-adjectives rela-
tive to typical adjectives. Our rst goal is to further
provide evidence for categorisation and statistical pre-
emption by replicating the Boyd and Goldberg results
in monolinguals while also extending them to bilinguals.
The sentence-production task was performed after
one of two exposure conditions: non-preemptive
exposure or preemptive exposure. The non-preemptive
exposure version examined speakersbaseline knowl-
edge of real a-adjectives and their baseline ability to
extend the syntactic constraint to novel a-adjectives. In
this version, a strong tendency to avoid using real a-
adjectives attributively should emerge, indicating that
speakers have pre-existing knowledge of an a-adjective
category. Novel words that resemble a-adjectives (e.g.
ablim) may follow a similar pattern but to a lesser
extent, reecting speakersbaseline generalisation ten-
dencies. The preemptive exposure version was designed
to test whether exposure to the a-adjective pattern with
a subset of novel words (e.g. the dog thats ablim) prior
to the production task strengthens canonical use of a-
adjectives. If speakers use preemptive contexts to gener-
ate strong inferences about what not to say, then speak-
ers should show a stronger tendency to avoid attributive
responses with novel a-adjectives following preemptive
exposure.
The second, and more critical, goal of the present
study is to identify the cognitive mechanisms associated
with variation in the use of a-adjectives. As we previously
established, the exposure eects reported by Boyd and
Goldberg (2011) can explain how speakers learn to
avoid using a-adjectives attributively, but important as
these eects are, they cannot readily account for why
some speakers systematically produce attributive
forms. The reasons for such variation may depend on a
LANGUAGE, COGNITION AND NEUROSCIENCE 5
variety of factors, some of which may come from the
person, but others of which may be imposed by the
context.
We hypothesise that statistical preemption involves
the regulation of two interrelated functional states (see
Figure 1 for an illustration) via working memory. Some
speakers may consistently produce non-canonical (i.e.
attributive) forms with a-adjectives. This tendency may
reect what function a-adjectives serve relative to
other typical adjectives (i.e. using a-adjectives attribu-
tively because they are modiers, which typically
follow attributive placement). Other speakers may
choose to produce canonical (i.e. non-attributive)
forms. This tendency may reect how a-adjectives are
typically used (i.e. using a-adjectives non-attributively
despite being modiers because they rarely, if ever,
appear in the attributive position).
Both category uency and Operation-span (O-span)
tasks are ideal individual dierence measures to
examine this hypothesis. The O-span is a complex
working memory task thought to recruit procedures
associated with the control of attention (Engle & Kane,
2004; Linck et al., 2013; Shipstead et al., 2015,2016)
and has shown to account for variability in language
processing (e.g. Bice & Kroll, 2021; Navarro-Torres et al.,
2019; Tanner & Van Hell, 2014; see Linck et al., 2013 for
a meta-analytic review). Verbal uency tasks, on the
other hand, are unique in that they may reect a combi-
nation of linguistic and domain-general control pro-
cesses by imposing lexical-semantic constraints on
how speakers self-generate words. For instance, per-
formance on category uency reects how speakers cat-
egorise lexical-semantic information in the L1 and L2
(Borodkin et al., 2016; Kavé & Goral, 2017), everyday
language-use routines (Shao et al., 2014), and vocabu-
lary size (Sauzéon et al., 2011). Notably, uency tasks
are also related to cognitive control (e.g. Carpenter
et al., 2020) and task-regulatory procedures in older
adults (e.g. Federmeier et al., 2010; Shao et al., 2014),
neuropsychological patients (e.g. Martin et al., 1994;
Pettit et al., 2013; Zhao et al., 2013), and bilinguals (e.g.
Grogan et al., 2009; Luo et al., 2010; Taler et al., 2013;
Zirnstein et al., 2018).
Following the construction-based regulation hypoth-
esis, we predicted that both category uency and O-
span would account for the two opposing tendencies
in a-adjective usage. Specically, whereas increased
uency may reect greater usage similarity between a-
adjectives and typical adjectives (i.e. the tendency to
produce attributive forms), greater O-span ability may
reect greater usage dierentiation between the two
(i.e. the tendency to produce non-attributive forms).
Figure 1. Depiction of two functional states as captured by the construction-based regulation hypothesis in the context of a-adjective
usage. An attributive state (top) implies that a-adjectives are categorised with typical adjectives. Under an attributive state, speakers
are more likely to use a-adjectives attributively. Category uency is hypothesised to index this tendency. A non-attributive state
(bottom) implies that a-adjectives and typical adjectives are used dierently. Under a non-attributive state, speakers are more
likely to use relative clauses. Working memory is hypothesised to index this tendency. Individuals can shift between these states
through a regulatory process, as depicted by the crossed arrows.
6C. A. NAVARRO-TORRES ET AL.
Critically, in line with the interaction reported by Zirn-
stein et al. (2018), and following the idea that cognitive
systems are interactive (Ihlen & Vereijken, 2010; Wallot &
Kelty-Stephen, 2018), we expected both uency and O-
span to interactively (rather than independently) con-
tribute to the use of a-adjectives.
This hypothesis contrasts with more modular
accounts that emphasise independent roles of language
ability and/or domain-general processes. For instance,
some research indicates that complex span tasks,
which are thought to reect domain-general processes,
better account for individual dierences in L1 and L2
abilities than more domain-specic tasks such as
simple span tasks (e.g. Daneman & Merikle, 1996; Linck
et al., 2013). Conversely, other research suggests that
working memory eects can be better explained in
terms of language-specic factors, such as verbal and
reading abilities, input quality and quantity, and phono-
logical encoding (Acheson et al., 2011; Daneman &
Carpenter, 1980; Farmer, Misyak, & Christiansen,
2012; Farmer, Fine, Misyak, & Christiansen, 2017; MacDo-
nald & Christiansen, 2002; Misyak et al., 2010a,2010b).
The key prediction following a modular view is that we
would expect O-span and uency to account for a-adjec-
tive use either uniquely (e.g. a signicant eect of
O-span but not of uency) or jointly (e.g. both uency
and O-span signicantly accounting for variance)
without any interactions between the two.
Testing the construction-based regulation hypothesis
and its alternatives may be dicult, if not impossible,
by only assessing young adult monolinguals who are at
their cognitive and verbal peak. This is because young
adult monolinguals may not actively engage language
regulation to the same extent as bilinguals, who are
required to engage regulatory processes when choosing
to speak in one language. Therefore, we hypothesise
that, through such cross-language regulatory mechan-
isms, bilinguals would be more likely to reveal the con-
struction-based regulatory processes (i.e. the interaction
between working memory and category uency) associ-
ated with categorisation and statistical preemption.
The bilingual speakers in the present study oer an
interesting scenario to examine this issue. These bilin-
guals, typically referred to as heritage speakers (Polinsky
& Scontras, 2020; Staord & Azevedo, 2015), initially
acquired Spanish as their L1 at home but then became
educated and socialised in the L2 (English). As a result,
their use of Spanish is often limited to the home environ-
ment, whereas English has become the dominant
language across most social contexts (Hurtado & Vega,
2004;Kohnert et al., 1999).
One possibility is that both monolinguals and heri-
tage speakers show similar a-adjective usage tendencies
(both at the group and individual level), given bilinguals
higher level of prociency and dominance in English.
However, research on production shows that some bilin-
guals, particularly unbalanced bilinguals, tend to be
slower to produce words and sentences and are more
prone to speech errors in both languages relative to
monolinguals (e.g. Gollan & Goldrick, 2012; Sadat et al.,
2016; Runnqvist et al., 2013). Such production costs
may be because bilinguals have reduced experience in
each language relative to monolinguals but may also
reect cross-language activation of related alternatives
(Kroll & Gollan, 2014; Sadat et al., 2016; Sullivan et al.,
2018). Thus, for English-dominant bilinguals such as heri-
tage speakers, learning the distributions of typical and a-
adjectives may still pose unique challenges, as they need
to learn the adjectival regularities of each language
while at the same time actively managing cross-linguis-
tic discrepancies in adjectival use (Sánchez & Camacho,
2021).
2
Notably, heritage speakers vary in the extent to
which they are able to reach high levels of attainment
in English while maintaining uency in the home
language (Winsler, 2022). Here we argue that regulation
of the dominant language (English) may be important to
mediate cross-linguistic variation in adjectival use while
remaining sensitive to the a-adjective distributions
unique to English.
Method
Participants
Monolingual participants included a total of 58 individ-
uals (29 for the non-preemptive version and 29 for the
preemptive version). Monolingual participants were
uent in English only, had not enrolled in any second-
language courses during college, and had no more
than minimal exposure to an L2 prior to college. Mono-
linguals were recruited at the Pennsylvania State Univer-
sity or at the University of California, Riverside, and
received either course credit or $10/hour. Bilingual par-
ticipants included a total of 56 Spanish-English speakers
(30 for the Non-preemptive version and 26 for the Pre-
emptive version) raised in the US. Bilinguals were
recruited at the University of California, Riverside, and
received either course credit or $10/hour. All partici-
pants gave informed consent, and the procedures had
the approval of the Institutional Review Board at the Uni-
versity of California, Riverside.
Tables 1 and 2provide the characteristics of the
monolingual and bilingual samples, respectively. Partici-
pants completed a category uency task and the O-span
to assess verbal abilities and working memory (see
Materials section). Participants also completed a
LANGUAGE, COGNITION AND NEUROSCIENCE 7
language history questionnaire to assess language
history, as well as self-rated levels of language ability.
Monolinguals across the two exposure conditions did
not signicantly dier in age, English uency, O-span
recall, and O-span processing speed (all ps > .140).
Both monolingual subgroups also self-rated themselves
similarly in English prociency, p> .190.
For bilinguals, participants across both exposure con-
ditions did not signicantly dier in age, the age at
which the two languages were acquired, current fre-
quency of exposure and use in each language, as well
as exposure prior to age seven (all ps > .540). Critically,
they also displayed similar category uency in each
language, similar O-span recall, as well as similar O-
span processing speed (all ps > .200). We note that bilin-
gual participants in the preemptive group self-rated
their English abilities lower than the non-preemptive
group, t(25) = 3.61, p= .001. However, given that self-
rated prociency measures have been found to be less
reliable than objective prociency measures
(Tomoschuk et al., 2019), and given the subgroup simi-
larities in all other measures (language exposure,
language use, and category uency), we attribute the
discrepancies in self-rated prociency primarily to meta-
linguistic and/or sociocultural awareness, rather than
prociency or language ability per se.
Finally, both bilingual subgroups reported acquiring
Spanish before English, Non-preemptive: t(27) = 3.15,
p= .004; Preemptive: t(25) = 6.87, p< .001, and having
greater exposure to Spanish relative to English before
age seven, Non-preemptive: t(27) = 3.67, p= .001; Pre-
emptive: t(25) = 3.82, p= .001. Despite this, both sub-
groups produced more category exemplars in English
than in Spanish, Non-preemptive: t(27) = 5.30, p
< .001; Preemptive: t(25) = 6.60, p< .001, and reported
greater current exposure, Non-preemptive: t(27) =
3.23, p= .003; Preemptive: t(25) =3.28, p= .003, and
use, Non-preemptive: t(27) = 4.62, p< .001; Preemp-
tive: t(25) = 5.30, p< .001, to English than to Spanish.
These patterns are consistent with a general characteris-
ation of heritage speakers in Southern California, who
typically acquire Spanish rst but eventually become lin-
guistically dominant in the societal language English.
Materials
Sentence Production Task. The materials used for the sen-
tence production task were adapted from Boyd and
Goldberg (2011). During the production task (Figure 2
(B)), pairs of animal images were presented onscreen,
each containing a target or foil label underneath.
Target labels included picture-label items that per-
formed a visual action on the screen (e.g. a fox moving
Table 2. Participant characteristics from the bilingual samples.
Group 1 (n= 30)
Non-preemptive
exposure
Group 2 (n= 26)
Preemptive exposure
M(SD)
Valid
NM(SD)
Valid
N
Age 19.41 (1.50) 29 19.62 (2.68) 26
Age of
acquisition
Spanish 1.81 (2.42) 27 1.38 (1.20) 26
English 3.96 (1.89) 27 3.73 (1.76) 26
Self-rated
prociency
Spanish 8.73 (1.23) 29 8.81 (1.27) 26
English 8.92 (0.87) 29 8.04 (1.37)*** 26
Self-rated
exposure
Spanish 8.43 (2.04) 28 8.88 (1.66) 26
English 9.64 (1.10) 28 9.92 (0.27) 26
Self-rated use
Spanish 7.54 (2.52) 28 8.42 (1.58) 26
English 9.64 (1.03) 28 9.77 (0.71) 26
% exposure
before age
seven
Spanish 70 (30) 28 71 (28) 26
English 30 (30) 28 29 (28) 26
Verbal uency
Spanish 33.36 (6.01) 28 33.00 (6.31) 26
English 43.27 (7.06) 30 41.27 (6.26) 26
O-span
Recall score 38.59 (8.04) 29 40.27 (7.87) 26
Response time
(ms)
2310.29 (362.02) 29 2285.06 (208.79) 26
Note: Self-ratings (i.e. prociency, exposure, and use) were made on a 10-
point scale ranging from 1 (no prociency or exposure/use) to 10 (high
prociency or exposure/use). Self-rated prociency reects average self-
ratings across four dimensions (speaking, listening, reading, writing).
Fluency score was measured as the average number of exemplars pro-
duced across semantic categories. O-span recall score was measured as
the number of correctly recalled to-be-remembered words out of 60 poss-
ible. O-span response time indicates the amount of time it took to correctly
solve an equation. Some data were excluded due to experimental or
equipment errors.
***p<.001.
Table 1. Participant characteristics from the monolingual
samples.
Group 1 (n = 29)
Non-preemptive exposure
Group 2 (n = 29)
Preemptive exposure
M(SD)
Valid
NM(SD)
Valid
N
Age (in years) 21.34 (3.18) 29 20.11 (2.11) 28
Self-rated
prociency
9.59 (0.91) 29 9.62 (0.60) 28
Verbal uency 52.69 (8.98) 29 54.50 (10.71) 28
O-span
Recall score 46.07 (7.17) 28 49.28 (6.95) 29
Response time
(ms)
2184.61 (364.83) 28 2156.73 (348.32) 29
Note: Self-ratings were made on a 10-point scale ranging from 1 (not pro-
cient) to 10 (highly procient). Self-rated prociency reects average self-
ratings across four dimensions (speaking, listening, reading, writing).
Fluency score was measured as the average number of exemplars pro-
duced across semantic categories. O-span recall score was measured as
the number of correctly recalled to-be-remembered words out of 60 poss-
ible. O-span response time indicates the average amount of time it took to
correctly solve a math equation. Some data were excluded due to exper-
imental or equipment errors.
8C. A. NAVARRO-TORRES ET AL.
towards an apple). After each trial, a question appeared
onscreen (e.g. which fox moved to the apple?), which
probed the participant to answer out loud using the
target label with a sentence containing either an attribu-
tive construction or a relative clause construction.
We specically followed the design of Experiments 1
and 2 in Boyd and Goldberg (2011), where prior to the
production task, participants witnessed an exposure
block in which the experimenter demonstrated the pro-
cedure of the task using novel adjectives as illustrations.
Participants were assigned to one of two exposure ver-
sions (Figure 2(A)). In the non-preemptive version, the
experimenter demonstrated the task procedure with
four novel labels with typical adjective characteristics
(i.e. gorpy,oggy,tooky,zinky) that were never used as
target labels in the production task. Participants saw a
total of 12 examples (each label repeated 3 times) in
the exposure block. To reduce any biases for producing
attributive vs. non-attributive responses, half of the
example labels appeared in attributive position (e.g.
the gorpy fox) while the remaining half appeared in
relative clause position (e.g. the fox thats gorpy). In
the preemptive version, participants also saw a total of
12 examples but were presented with 2 novel a-adjec-
tives in a relative clause structure (e.g. the fox thats
ablim). These novel a-adjectives were repeated three
times, both of which also appeared in the production
task. The remaining six example trials included novel
typical adjectives in attributive position (e.g. the
gorpy fox).
There was a total of 56 trials in each version of the
production task, 28 of which contained a target label
corresponding to 1 of 4 experimental conditions
(Table 3). Target labels were manipulated for adjective
novelty (i.e. whether the target label was a real or
novel word) and adjective type (i.e. whether the target
label was an a-adjective or a typical adjective). Thus,
each of experimental condition had seven target labels
total. The remaining 28 trials consisted of llers
(Table 4). Half of the ller trials contained high-frequency
adjectives that typically follow attributive placement.
The remaining ller trials contained third-person singu-
lar verbs in the present tense that required the use of
relative clause constructions. The inclusion of both
types of llers would help minimise any potential
biases that would globally favour the use of one
Figure 2. Experimental procedure illustrating the dierence between preemptive and non-preemptive exposure (A) and a trial
sequence within the production block (B).
LANGUAGE, COGNITION AND NEUROSCIENCE 9
structure over the other, while also guarding against
possible structural priming eects within the task.
Participants were randomly assigned to view one of two
counterbalanced versions of the task. Trials were pseudo-
randomized across each counterbalanced version. The
rst trial was always a ller trial. Afterwards, critical and
ller trials always alternated between one another. There
were no more than two llersofthesametype(adjective
labels vs. verb labels) that occurred consecutively. There
were no more than two critical trials from the same con-
dition that occurred in a row. Each trial was counterba-
lanced in such a way that half of the target labels
appeared on the left, and half on the right.
There are three crucial ways in which our version of
the sentence production task materials diered from
the materials used in Boyd and Goldberg (2011). First,
we added three additional items per condition, includ-
ing three additional real a-adjectives taken from Gold-
berg and Boyd (2015), to increase power and
generalisability. Second, we modied the list of novel
labels such that novel typical adjectives had the same
stem as those used with novel a-adjectives. This was
done in part to make comparisons between the novel-
word conditions analogous to the comparisons
between the real-word conditions (e.g. blimsy and
ablim vs. sleepy and asleep). Importantly, if participants
consistently avoid attributive responses of novel a-
adjectives, this would indicate that speakers are genu-
inely extending a pattern based on how words are
used in context and not just based on isolated lexical-
semantic features. Lastly, unlike the Boyd & Gold-
berg (2011) study, we increased task diculty by impos-
ing a time limit in each trial (see Figure 2(B)) and by
varying the types of questions that would be used to
elicit a response (see Procedure below).
Category Fluency Task. Monolingual participants com-
pleted a category uency task to measure language pro-
duction abilities in English. In this task, they were asked
to generate as many exemplars as possible that belong
to a semantic category within a 30-second time limit.
Four categories were presented out of eight possible
categories (animals, body parts, clothing, colours,
fruits, furniture, musical instruments, and vegetables).
Two of the four categories were animate, while the
remaining two were inanimate. The order of presen-
tation for all categories was counterbalanced in two ver-
sions of the task. Participants were asked to avoid
producing repetitions of words and names of people
or places. Responses were recorded on a digital recorder.
949-999-2941.
In the bilingual version of category uency, four
semantic categories were presented for Spanish and
an additional four categories for English. Each language
was assessed in separate blocks. Two of the categories in
each language block were animate and two were inani-
mate. Categories were counterbalanced and evenly dis-
tributed between language blocks. Language block
order was also counterbalanced across participants.
Operation Span Task. The operation span (O-span)
task was used to assess working memory ability
(Turner & Engle, 1989). Participants were instructed
that they would solve math equations while simul-
taneously memorising English words. Each trial began
with a xation sign (+), presented in the centre of the
screen for 1000 ms. Next, a simple algebraic equation
(e.g. (4 * 2)2 = 2) appeared, on which participants
judged the accuracy of the solution to the equation. Par-
ticipants were told to respond as quickly and accurately
as possible to the equation by pressing yes when it was
correct or no when it was incorrect. The equation
remained on the screen until their response or up to
3750 ms. After the equation disappeared, an English
word appeared for 1250 ms and was followed by
another xation that indicated the beginning of a new
equation. This sequence of events was repeated until
the complete set of equations and words was presented.
After the complete set of equations and words was pre-
sented, the word RECALLappeared on the screen to
prompt the participant to use the keyboard to type in
all the words they had seen during the set. Participants
were instructed that they did not have to type the
words in the exact order in which they were presented,
but that they could not type the last word of the set as
their rst response. Set sizes ranged from two to six
equations and words and set size incrementally
increased as the task progressed.
Table 4. Stimuli list for ller trials.
High frequency adjectives 3rd person singular verbs
angry (happy)
bad (good)
fast (slow)
nice (smart)
old (young)
rich (poor)
strong (weak)
happy (angry)
good (bad)
slow (fast)
smart (nice)
young (old)
poor (rich)
weak (strong)
bites (votes)
cheats (lies)
dances (swims)
gambles (smokes)
golfs (snowboards)
ghts (sings)
teaches (complains)
votes (bites)
lies (cheats)
swims (dances)
smokes (gambles)
snowboards (golfs)
sings (ghts)
complains (teaches)
Note: Distractor labels are in parenthesis.
Table 3. Experimental design and stimuli list.
Real Novel
A-adjective Typical adjective A-adjective Typical adjective
afraid (brave)
aoat (sinking)
alive (dead)
alone (friendly)
asleep (vigilant)
awake (tired)
aware (ignorant)
frightened (brave)
oating (sinking)
living (dead)
lonely (friendly)
sleepy (vigilant)
wakeful (tired)
conscious (ignorant)
ablim (zecky)
abod (impy)
adax (zedgy)
ait (tammy)
afraz (zibby)
agask (zampy)
awass (glimful)
blimsy (oud)
bodsy (frodly)
daxy (zoopy)
itzy (zappy)
frazzy (zunderful)
gasky (zinky)
wassy (gralliant)
Note: Distractor labels are in parenthesis.
10 C. A. NAVARRO-TORRES ET AL.
Procedure
Participants rst completed the sentence production
task, followed by category uency, O-span, and the
language history questionnaire. At the end of the
session, participants were debriefed, rst by asking
whether they had guessed the purpose of the exper-
iment or had explicit knowledge of the a-adjective con-
straint, followed by the experimenter explaining the
purpose of the study in more detail. Participants were
tested in a sound-attenuated room while seated in
front of a computer that was connected to a button
box and a digital recorder.
At the beginning of the main task, participants were
briefed on the experimental procedure and were told
to use labels exactly as they appeared onscreen. They
were also encouraged to use full sentences and avoid
saying sentence fragments only. Prior to the exposure
block, these points were emphasised in six practice
trials (using real labels) in which participants received
direct feedback from the experimenter. The rst three
practice trials were performed by the experimenter,
while participants were encouraged to perform the last
three trials. After successfully reviewing all six examples,
they were then told that the real task would contain
words that they had never seen before and were encour-
aged to use them as if they were real words.
The experimenter then proceeded to the exposure
block. At the beginning of each trial, the rst pair of
animal images appeared on the screen for two
seconds (Figure 2(B)). Here participants were told that
they would have to say both labels (target and distrac-
tor) out loud before the images disappeared. The exper-
imenter responded as a would-be participant to ensure
that participants understood this component of the
task. The next sequence showed the target image per-
forming an action within one second. These actions
varied across trials. That is, sometimes the target
image moved towards an object (or simply towards a
location without reference); other times, the image
shrunk or enlarged in size; and at other times, it disap-
peared. This was done to make the task less predictable,
thus forcing participants to remain attentive throughout
the task. Finally, a question appeared in the middle of
the screen for ve seconds. During the exposure block,
answers were provided underneath the question and
the experimenter read them out loud. After the question
disappeared, a xation sign appeared at the centre of
the computer screen indenitely. They were told that
this was the only time where they would have to press
a key to proceed to the next trial. After completing the
exposure block, the experimenter left the room to let
the participant begin the production task. Participants
had to formulate their own descriptions on each trial
and were encouraged to provide an answer as quickly
and as best as possible. Two microphones were used
to record both syntactic choices and response times.
Coding and analysis
Sentence responses were coded for syntactic choice;
that is, whether the target label was used attributively
(i.e. prenominally) or non-attributively (i.e. in a relative
clause). Responses were considered incorrect if the par-
ticipant modied and/or used the wrong label construc-
tion (e.g. the fox thats cool and asleep moved to the
apple), if the formulation was incomplete (e.g. the
one on the left …”), or if there was no response. False
starts, reformulations, and hesitations were included in
this part of the analysis if the response was correct. For
the monolingual sample, there were 1623 critical trial
responses in total, and 27 were excluded due to incor-
rect responses. For the bilingual sample, there was a
total of 1568 critical trial responses, 45 of which were
excluded due to incorrect responses.
3
Statistical analyses were performed with generalised
mixed-eects models using the lme4 software package
(Bates, Mächler, et al., 2015) in the R programming
environment (R Development Core Team, 2019). We
rst conducted group-level analyses by creating core
models for each group. These core models included con-
trast coded xed eects of adjective novelty (real = 0.5,
novel = 0.5), adjective type (a-adjective = 0.5, typical
adjective = 0.5), exposure type (non-preemptive = 0.5,
preemptive = 0.5), and their interactions. For these
analyses, we attempted to t random eects using a
maximal procedure (Barr et al., 2013), including
random intercepts for items and participants, by-partici-
pant random slopes for adjective novelty and adjective
type, and their interaction, as well as by-item random
slopes for exposure type. However, the logistic model
returned a warning message for singularity. Following
the recommendation of Bates, Kliegl, et al. (2015), we
conducted a principal component analysis to simplify
the random eects structure. Based on the analysis, we
removed the correlation between the by-item intercept
and the random slope for exposure type in the monolin-
gual analysis. In the bilingual analysis, we removed the
random slope for exposure type altogether. This resulted
in the following R codes for the monolingual and bilin-
gual models, respectively:
glmer(Syntactic Choice Adjective Novelty
Adjective TypeExposure +(Adjective Novelty
Adjective Type|Participant)+(Exposure||Item))
LANGUAGE, COGNITION AND NEUROSCIENCE 11
glmer(Syntactic Choice Adjective Novelty
Adjective TypeExposure +(Adjective Novelty
Adjective Type|Participant)+(1|Item))
Next, we proceeded to create follow-up generalised
mixed models to examine whether working memory
and uency accounted for variability in syntactic
choices across conditions. In these models, we included
O-span recall and English uency scores as xed eects,
both of which were allowed to interact with adjective
novelty and adjective class, as well as with each other.
Given that the purpose of adding the individual dier-
ence measures was to explain participant-level variabil-
ity across conditions, by-participant random slopes for
these models were removed (see Gullifer et al., 2018;
Gullifer & Titone, 2020 for a similar procedure). Since
our goal was to capture variation that is common to
both versions of the experiment, we included exposure
type as a covariate. This resulted in the following R code:
glmer(Syntactic Choice Exposure +Adjective Novelty
Adjective TypeOspanFluency +(1|Participant)
+(Exposure|Item))
All continuous factors were centred and standardised
to aid model interpretation. Signicant interactions and
follow-up comparisons in the core models were exam-
ined by retting the model with a dummy coded categ-
orical factor to examine simple eects at each level of
the categorical factor. In the case of signicant inter-
actions involving continuous factors, simple eects were
also conducted by rescaling a given continuous factor 1
SD above/below the mean (Aiken & West, 1991). For
example, a signicant interaction between O-span score
and adjective type might indicate that the eect of the
latter depends on working memory ability. To conrm
this statistically, we can ret the model to determine
the eect of adjective type for individuals with high or
low working memory (i.e. 1 SD above or below the
mean). Note that retting the model does not aect the
goodness of t or the type-1 error rate. Rather, it simply
uses a dierent reference point to provide a dierent
interpretation of the coecients while keeping the var-
iance constant (Gelman et al., 2012;Gelman&Hill,2007).
Results
Group-level usage patterns
Monolinguals
Figure 3 (left panel) shows the raw proportion of attribu-
tive responses for monolinguals across exposure type
and across conditions (see also Table B1 in the Appendix
for descriptive means). The mixed logistic regression
analysis (Table 5) conrmed a signicant eect of
exposure, indicating that the likelihood of overall attri-
butive responses was reduced for monolinguals
exposed to a preemptive context relative to those
exposed to a non-preemptive context. Signicant main
eects of adjective type and adjective novelty indicated
that the likelihood of an attributive response was lower
for real vs. novel words and for a-adjectives vs. typical
adjectives. A signicant two-way interaction between
adjective novelty and adjective type indicated that,
across the two exposure types, the tendency to avoid
attributive responses was greater for real a-adjectives,
β= 3.50, SE = 0.44, z= 7.93, p< .001, than for novel a-
adjectives, β= 2.12, SE = 0.40, z= 5.29, p< .001.
Critically, and as previously shown, the rate of attribu-
tive responses with novel a-adjectives depended on the
type of exposure. A signicant three-way interaction
conrmed that the two-way (adjective novelty × adjec-
tive type) interaction was signicant for monolinguals
with non-preemptive exposure, β=2.36, SE = 0.64, z
=3.68, p< .001, but not for monolinguals with preemp-
tive exposure, β=0.40, SE = 0.63, z=0.63, p= .528.
Simple eects analyses further conrmed that attribu-
tive rates for novel a-adjectives were lower for monolin-
guals with preemptive exposure relative to
monolinguals with non-preemptive exposure, β= 1.66,
SE = 0.38, z= 4.43, p< .001. No exposure eects were
observed with real a-adjectives, β= 0.18, SE =0.51, z=
0.36, p= .720, real typical adjectives, β= 0.69, SE = 0.45,
z= 1.55, p= .122, or novel typical adjectives, β= 0.20,
SE = 0.46, z= 0.44, p= .657, suggesting that the eect
of preemptive exposure was specic to novel
a-adjectives.
Bilinguals
As shown in Figure 3 (lower panels) and Table 6 (see also
Table B2 in the Appendix), the bilingual data also yielded
main eects of exposure, adjective type, and adjective
novelty, indicating an overall reduction in attributive
response rates for bilinguals with preemptive exposure
vs. non-preemptive exposure, for a-adjectives vs. non
a-adjectives, and for real adjectives vs. novel adjectives.
The interaction between adjective novelty and adjective
type was also signicant, conrming that the eect of
adjective type was greater for real adjectives, β= 4.00,
SE = 0.42, z= 9.60, p< .001, than for novel adjectives, β
= 1.93, SE = 0.43, z= 4.48, p< .001. However, unlike
monolinguals, the three-way interaction was not signi-
cant. Instead, there was a signicant two-way interaction
between adjective type and exposure. The statistical
breakdown of this interaction indicated that the
12 C. A. NAVARRO-TORRES ET AL.
exposure eect was signicant for both real, β= 1.84, SE
= 0.56, z= 3.30, p= .001, and novel, β= 1.98, SE = 0.47, z
= 4.26, p< .001, a-adjectives. Importantly, the
production choices for typical adjectives did not
change as a function of exposure, real: β=0.37,SE = 0.46,
z= 0.80, p= .425; novel: β=1.05, SE = 0.70, z= 1.51,
Table 5. Estimated coecients from the mixed logistic model
on syntactic choice for monolinguals.
Fixed eects Estimate SE
z-
values p
(Intercept) 0.79 0.20 4.05 <.001
Adj Novelty 0.89 0.29 3.12 .002
Adj Type 2.81 0.32 8.82 <.001
Exposure 0.69 0.32 2.16 .031
Adj Novelty × Adj Type 1.38 0.55 2.50 .013
Adj Novelty × Exposure 0.50 0.35 1.42 .156
Adj Type × Exposure 0.48 0.44 1.09 .277
Adj Novelty × Adj Type ×
Exposure
1.97 0.62 3.16 .002
Random eects Variance SD Correlations
Intercept | participant 1.1303 1.06
Adj Novelty | participant 0.5477 0.74 0.23
Adj Type | participant 1.4883 1.22 0.11 0.46
Adj Novelty × Adj Type |
participant
0.7242 0.85 0.31 0.29 0.87
Intercept | item 0.3100 0.56
Note: SE = standard error of the estimate.
Table 6. Estimated coecients from the mixed logistic model
on syntactic choice for bilinguals.
Fixed eects Estimate SE
z-
values p
(Intercept) 0.83 0.23 3.60 <.001
Adj Novelty 1.43 0.35 4.08 <.001
Adj Type 3.01 0.32 9.42 <.001
Exposure 1.09 0.40 2.71 .007
Adj Novelty × Adj Type 2.09 0.56 3.76 <.001
Adj Novelty × Exposure 0.14 0.54 0.26 .792
Adj Type × Exposure 1.26 0.48 2.61 .009
Adj Novelty × Adj Type ×
Exposure
0.23 0.73 0.32 .749
Random eects Variance SD Correlations
Intercept | participant 1.7396 1.38
Adj Novelty | participant 2.2390 1.50 0.33
Adj Type | participant 1.2154 1.10 0.21 0.72
Adj Novelty × Adj Type |
participant
0.1477 0.38 0.87 0.17 0.19
Intercept | item 0.2008 0.45
Exposure | item 1.7396 0.16 1.00
Note: SE = standard error of the estimate.
Figure 3. Mean proportion of attributive (prenominal) vs. non-attributive (relative clause) responses by exposure type, adjective type,
adjective novelty, and group. Higher bars indicate increased attributive use. Circles plot values for individual participants. Error bars
represent 95% condence intervals. Exp1 = Non-preemptive exposure; Exp2 = Preemptive exposure; ***p< .001; **p< .01; ns denotes
adierence that was not statistically signicant.
LANGUAGE, COGNITION AND NEUROSCIENCE 13
p= .132, conrming that the decrease in attributive rates
was specictoa-adjectives.
Interim summary and discussion
The group-level analyses indicated that speakers, both
monolinguals and bilinguals alike, tended to avoid
using a-adjectives attributively relative to their typical
adjectival counterparts. Without exposure to a preemp-
tive context, speakers partially extended the a-adjective
constraint to novel words such as ablim,adax, and
awass. Attributive rates for novel a-adjectives decreased
with exposure to a preemptive context (i.e. using relative
clauses with a subset of novel a-adjectives in a context
where attributive use is expected). Notably, our version
of the task had several modications relative to Boyd
and Goldberg (2011) to assess the generalisability of
the eects. These modications included a larger
stimuli set, novel labels with identical stems across con-
ditions (e.g. ablim/blimsy, awass/wassy), and increased
task diculty by imposing time constraints and unpre-
dictable variations in the kinds of actions performed by
target items. Together, these ndings conrm the
results reported by Boyd and Goldberg (2011) and are
consistent with the idea that speakers learn the distri-
bution of a-adjectives via statistical preemption.
4
Even though monolinguals and bilinguals patterned
similarly, group dierences emerged regarding how
they were impacted by exposure. While monolinguals
with preemptive exposure only showed a reduction in
rates with novel a-adjectives, bilinguals with preemptive
exposure showed a reduction with both real and novel
a-adjectives. Since novel words were devoid of any con-
ventionalised meaning, this suggests that bilinguals
were more open to generating extensions of a form
regardless of its semantic associations. This contrasts
with monolingual speakers, who often extend forms to
novel contexts based on semantic similarity (Floyd &
Goldberg, 2021; Harmon & Kapatsinski, 2017). Recent
studies also provide evidence for a dissociation
between monolinguals and L2 speakers in judging
unconventional sentences via statistical preemption
(Robenalt & Goldberg, 2016; Tachihara & Goldberg,
2020). For instance, L2 speakers may judge sentences
such as Amber explained Zach the answeras being
more acceptable than monolinguals, presumably
because they are less likely to take its competing alterna-
tive Amber explained the answer to Zachinto account.
Because such a discrepancy was not predicted in our
data, we can only provide two speculative interpret-
ations based on previous work.
One interpretation is that bilinguals display an
increased non-semantic extension of a form because
their form-meaning mappings are less well-dened or
noisier. Such an interpretation is consistent with a
weaker-links, or frequency-lag, account (Gollan et al.,
2011) in which linguistic accessibility in bilinguals is
diminished because of reduced cumulative experience
in each language relative to monolinguals. Another
interpretation (i.e. a competition-for-selection account)
is that bilinguals may show increased non-semantic
extension not because of reduced experience but
because of a need to regulate an increasing number of
competing forms with similar meaning. These many-
to-one mappings may create more opportunities for lin-
guistic interference during speech planning (e.g. Abuta-
lebi & Green, 2007; Sadat et al., 2016; Sullivan et al.,
2018), and may also impact how bilinguals extract distri-
butional regularities from the input (Poepsel & Weiss,
2016). Consistent with this idea is the nding that bilin-
gual infants are less likely to follow a mutual-exclusivity
bias, whereby objects are assumed to have a single
designated label (e.g. Byers-Heinlein & Werker, 2009;
Houston-Price et al., 2010). The implication of the com-
petition-for-selection account is that bilinguals may
possess a language system that is more open to new
kinds of form extension regardless of its semantic associ-
ations. We will return to this issue in the General Discus-
sion following the individual dierence analyses that we
report below.
Importantly, while some individuals systematically
avoided using a-adjectives attributively, others consist-
ently preferred using them attributively (see Figure 3).
This was true for both monolinguals and bilinguals.
Although the statistical preemption hypothesis can
account for how speakers learn to avoid attributive
forms, it cannot readily explain why some individuals
in the experiment persisted in producing them. Rather
than treating these sources of variation as noise, we con-
sider the possibility that they reect systematic discre-
pancies in how speakers regulate competing
alternatives during utterance planning.
In an attempt to reconcile these patterns of language
use, we presented the construction-based regulation
hypothesis, in which speakers regulate attributive and
non-attributive states by interactively coordinating
language-related and domain-general processes that
are tied to working memory and attention. Under this
account, category uency and O-span are expected to
interactively index attributive and non-attributive
usage tendencies, respectively, and that bilinguals
would be more likely to reveal these interactive eects
given their need to regulate the dominant language
(Branzi et al., 2016; Boguslki et al., 2019; Pulido &
Dussias, 2020; Rossi et al., 2018; Zirnstein et al., 2018).
We also presented the alternative hypothesis that
14 C. A. NAVARRO-TORRES ET AL.
distinct cognitive processes may account for variation in
a-adjective use in an additive manner. Under this
account, uency and O-span would independently
predict task performance.
Individual dierences in category uency and O-
span
Monolinguals
For monolinguals, O-span signicantly accounted for
production tendencies. As Figure 4 (left panel) shows,
this eect depended on whether the produced sentence
contained an a-adjective or a typical adjective. This was
conrmed via a signicant two-way interaction between
O-span and adjective type (Table 7). Follow-up analyses
indicated that higher O-span scores were
signicantly associated with a reduction in attributive
responses for (both real and novel) a-adjectives, β=
0.40, SE = 0.19, z=2.11, p= .035. For typical adjec-
tives, the association was not signicant, β= 0.12, SE =
0.21, z= 0.60, p= .550. As Table 7 indicates, Fluency did
not predict performance on any of the experimental
conditions, nor did it interact with O-span.
5
Bilinguals
Model estimates for bilinguals are shown in Table 8. The
results conrmed the predicted interaction between
English uency and O-span, but the pattern of inter-
action varied across experimental conditions. Therefore,
we report the results in an incremental fashion, focusing
on the lower-order terms rst, followed by the higher-
order interactions.
First, like monolinguals, O-span signicantly inter-
acted with adjective type (Figure 4, right panel), indicat-
ing that higher O-span was signicantly associated with
a decrease in attributive rates for a-adjectives, β=0.53,
SE = 0.20, z=2.73, p= .006, but not for typical adjec-
tives, β= 0.01, SE = 0.20, z= 0.07, p= .947. However,
unlike monolinguals, bilingualsEnglish uency
additionally predicted production tendencies. And
unlike O-span, the eect of uency was positive and pre-
dictive of overall performance (Figure 5, right panel). In
other words, increased English uency was associated
with increased attributive rates across conditions. This
suggests that uency and O-span capture two distinct
(i.e. attributive vs. non-attributive) production
tendencies.
Second, there was a signicant interaction between
O-span and English uency.
6
Simple slopes analyses
revealed a signicant positive eect of uency at low
O-span levels, β= 0.88, SE = 0.26, z= 3.37, p= .001, but
no eect of uency at high O-span levels, β=0.14,
SE = 0.30, z=0.47, p= .635. Conversely,