مقاله

Language Learning ISSN 0023-8333
EMPIRICAL STUDY
Error-Correction Mechanisms in Language

Learning: Modeling Individuals
Adnane Ez-zizi ,a,b Dagmar Divjak ,a and Petar Milin a
a
University of Birmingham b University of Suffolk
Abstract: Since its first adoption as a computational model for language learning, ev-
idence has accumulated that Rescorla–Wagner error-correction learning (Rescorla &
Wagner, 1972) captures several aspects of language processing. Whereas previous stud-
ies have provided general support for the Rescorla–Wagner rule by using it to explain
the behavior of participants across a range of tasks, we focus on testing predictions gen-
erated by the model in a controlled natural language learning task and model the data
at the level of the individual learner. By adjusting the parameters of the model to fit the
trial-by-trial behavioral choices of participants, rather than fitting a one-for-all model
using a single set of default parameters, we show that the model accurately captures par-
CRediT author statement – Adnane Ez-zizi: conceptualization; methodology; investigation; data

curation; formal analysis; software; visualization; writing – original draft preparation; writing – re-
view & editing. Dagmar Divjak: conceptualization; methodology; funding acquisition; writing –
review & editing. Petar Milin: conceptualization; methodology; funding acquisition; validation;
writing – review & editing.
A one-page Accessible Summary of this article in non-technical language is freely available in
the Supporting Information online and at https://oasis-database.org
We are grateful to Emma Marsden and four anonymous reviewers for their insightful comments
and suggestions on this article. We would also like to thank Maciej Borowski for his help with
the choice of the stimuli and for discussion of relevant aspects of the Polish language. We are
also grateful to Christian Adam for his help with running pilots of the experiment and for sharing
his feedback on those pilots. This research was funded by Leverhulme Trust Leadership Grant
RL-016-001 to Dagmar Divjak, which funded all authors. The first author was affiliated with the
University of Birmingham while work on the paper was carried out.
Correspondence concerning this article should be addressed to Petar Milin, Ashley Build-
ing, University of Birmingham, Edgbaston, Birmingham, B15 2TT, United Kingdom. Email:
p.milin@bham.ac.uk
The handling editors for this manuscript were Emma Marsden and Pavel Trofimovich.
This is an open access article under the terms of the Creative Commons Attribution License,
which permits use, distribution and reproduction in any medium, provided the original work is
properly cited.
Language Learning 74:1, March 2024, pp. 41–77 41

© 2023 The Authors. Language Learning published by Wiley Periodicals LLC on behalf of Language
Learning Research Club, University of Michigan.
DOI: 10.1111/lang.12569
14679922, 2024, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/lang.12569 by Readcube (Labtiva Inc.), Wiley Online Library on [21/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Ez-zizi, Divjak and Milin Error-Correction Mechanisms in Language Learning
ticipants’ choices, time latencies, and levels of response agreement. We also show that
gender and working memory capacity affect the extent to which the Rescorla–Wagner
model captures language learning.
Keywords language learning; error-correction learning; Rescorla–Wagner model;

morphology; agreement
Introduction
We humans share with other species many core learning mechanisms that
allow us to adapt to our environment (Rescorla, 1988). These mechanisms
include, among others, classical conditioning (i.e., Pavlovian conditioning;
Pavlov, 1927), instrumental conditioning (also operant conditioning; Skinner,
1938), and forms of social learning, such as vicarious learning (Bandura,
1962). Arguably the most uniquely defining human learning ability is lan-
guage learning, which also includes efficient transgenerational transmission
and is foundational for social inclusion and cohesion. However, whereas core
learning mechanisms are relatively well understood, language learning remains
much of a mystery (Ambridge & Lieven, 2011). An early attempt by Skinner
(1957) to account for language learning using the same principles as those gov-
erning lower-level cognitive tasks was quashed by Chomsky (1959). For much
of the remainder of the 20th century, language was seen as a by-and-large innate
system, governed by rules and handled by a uniquely human and specialized
cognitive structure. This structure was initially conceptualized as the language
acquisition device, and later extended to become universal grammar.
This dominant view was challenged from two sides simultaneously. The
emergence of usage-based linguistics in the 1980s (Langacker, 1987) promoted
a view of language as a dynamic and probabilistic system, resulting from
general cognitive capacities acting on language input (Dabrowska
˛ & Divjak,
2015). This view meshed well with connectionist frameworks, which showed
that rulelike behavior can emerge from exposure to usage alone and that lan-
guage knowledge is sensitive to properties of the input (Plaut & Gonnerman,
2000; Seidenberg & McClelland, 1989). Connectionism, arguably, paved the
way for changes in theorizing too, toward a view of language as being learned
like any other skill, and the early 2000s witnessed the start of a reintegration
of the basic principles of learning into the body of work on language (e.g., see
Bybee & McClelland, 2005; for more up-to-date works, see Ellis et al., 2016,
which addresses both first and second language learning, as well as Chuang
et al., 2021, which addresses lexical acquisition in second and third languages).
Language was now seen as being amenable to the same general-purpose

cognitive capacities and learning mechanisms that humans and animals use
to navigate and adapt to their environment (cf. Ellis, 2006a; Ellis & Sagarra,
2010, 2011; Sturdy & Nicoladis, 2017).
Among these learning models, Rescorla and Wagner’s (1972) model of
classical conditioning stands out for its simplicity and its ability to explain a
range of empirical learning phenomena (Siegel & Allan, 1996). This model
has been shown to be biologically plausibile (Chen et al., 2008) and to have an
evolutionary advantage over other more powerful learning mechanisms, in the
sense that it has a higher likelihood of being naturally selected and persisting
in the evolutionary process, compared to other plausible learning mechanisms
(for more details, see Trimmer et al., 2012).
Background Literature
The Rescorla–Wagner Model
As a model of classical conditioning, the Rescorla–Wagner (R–W) model is
concerned with situations where an entity (a human, an animal, or a machine)
has to learn the predictive relationship between objects and/or events (i.e., cues
and outcomes) in an environment, and where cues compete for their predictive
value for an outcome while iteratively (re)calibrating the learning (or associa-
tion) weights. More specifically, an association weight reflects the tendency of
an outcome to occur in the presence of a certain cue. A higher positive associ-
ation weight value for a particular outcome corresponds to a higher likelihood
of occurrence of that outcome in the presence of the cue; conversely, a highly
negative value corresponds to a greater likelihood of nonoccurrence of that
outcome (the cue is said to be inhibitory in this case). Values close to zero
mean low chances of observing (if the weight is positive) or inhibiting (if the
weight is negative) the outcome.
The R–W model assumes that the organism computes a simple error-
correcting learning rule used to update the association weights in each new
learning event (e.g., each trial in a behavioral experiment). The general idea
behind the correction rule is that the association between a cue and outcome is
(a) strengthened if both cue and outcome are present in the learning event, (b)
weakened if the cue is present but the outcome is not, and (c) kept the same
if the cue itself is absent. The updating of the association weights is driven by
the discrepancy between the expected and the obtained outcome, such that the
magnitude of the update—how much the association weights are adjusted—is
determined by two parameters called learning rates, and the direction of the
update—whether it increases the weight or decreases it—depends on the sign
of the difference between the expected and the observed outcome. In this way,
43 Language Learning 74:1, March 2024, pp. 41–77

most broadly, for the R–W model, learning is about the outcomes, and this sets
it apart from related models where learning is about the input cues (e.g., Pearce
& Hall, 1980).
Another feature of the R–W model is that, although the outcomes are up-
dated independently from each other, input cues compete for the predictivity
of outcomes. In other words, the adjustment of the weights depends not only
on the single cue being updated but on all the cues present in the learning
event through their sum of association weights. This cue competition principle
is what allowed the R–W model to explain many of the puzzling phenomena
of classical conditioning, some of which were also shown to be valuable for
understanding the mechanics of language learning (see the next section for a
discussion).1 One of the best-known examples of such learning phenomena is
the blocking effect (Kamin, 1969). This effect occurs when a cue is trained in
compound with a second cue to predict an outcome but when the second cue
is already a good predictor of the outcome. In such cases, the first cue cannot
form a strong association with the outcome (i.e., the first cue is blocked by
the second cue). More generally, the cue competition principle often results
in the observation that the best cues for the outcome prevent other cues from
developing a strong association with that same outcome.
The Rescorla–Wagner Model and Language Learning

Since its first mention within a linguistic context by Ellis (2006a), evidence
has accumulated showing that the R–W model can capture several aspects of
language learning (e.g., Baayen et al., 2011; Ellis, 2006b; Milin, Divjak, &
Baayen, 2017; Milin, Feldman, et al., 2017). So far, the available empirical ev-
idence stems from studies that train a R–W model that uses default parameter
values (here we allude to the two learning rate parameters used to update the
association weights after each new event), typically on either a small sample
from experiments on artificial languages or a large corpus of texts.2 Posttrain-
ing learning measures are then extracted from the simulated model and are
compared against observed response measurements from an experimental task.
A first issue is that predictions for (and from) such models are typically
generated independently from the experiment (with exceptions such as the
studies of Ramscar & Yarlett, 2007, and Divjak et al., 2021, where the model
generated the hypotheses to be tested experimentally). The parameters are typ-
ically set to their default values, missing the opportunity to take into account
the variability that can arise from simulating the model with different parame-
ter values (though see Olejarczuk et al., 2018, who used fixed parameter values
but fitted a separate model to each participant’s data using the same sequence

of examples encountered by the participant). Incorporating the variability aris-

ing from the model parameters when fitting learning models to language data
has the potential to improve the explainability of the individual differences
observed in the experiment, especially since language usage and representa-
tion is an area that shows huge individual variation (Dabrowska,
˛ 2018).
Training the model on a large-scale corpus comes at an even greater cost.
We leave aside here the issue of (lack of) similarity between the contents of
a corpus and the input that language users receive (which plagues converging
evidence studies generally; for a summary discussion, see Klavan & Divjak,
2016, and for collections of worked examples, see Divjak & Gries, 2012, and
Gries & Divjak, 2012). Here we focus on another issue: Training on a corpus
mutes the two main sources of variability of the model—namely, those related
to the choice of model parameters and the order of training examples—which
are mostly active during the early stages of learning (Shanks, 1995; also see
Milin et al., 2020, for a more general discussion of the trial order effect in
error-correction learning).3 These early biases, as Ellis (2006a) called them,
constitute a real test for the R–W model, before it can be deployed as a model
of language learning at a large scale. Modeling the parameters’ variability and
training the R–W model on the same examples encountered by the participants
represent novel opportunities for understanding language learning not yet fully
explored in previous studies.
The Present Study

The aim of the present study is to model how individual language learners
engage with the task at hand on a trial-by-trial basis, which constitutes a step-
changing challenge for the application to language learning of discrimination
or error-correction learning in general and the R–W model in particular.
Whereas previous studies have provided general support for the R–W rule
by using this model to explain the behavior of participants across a range of
tasks (Divjak, 2019; Milin & Blevins, 2020; Milin, Feldman, et al., 2017;
Pirrelli et al., 2020), we focus on testing predictions generated by the model
in a controlled natural language learning task and model the data at the level
of the individual language learner. In doing so, we treat each participant as a
separate learning entity governed by different capacities, which are, crucially,
formalized through the learning parameters of the chosen model.
Given that several studies have reported that classical conditioning per-
formance can be affected by cognitive and personal characteristics such
as working memory (Baetu et al., 2018; Sasaki, 2009), gender (Lonsdorf
et al., 2015; Merz et al., 2018), and age (e.g., Mutter et al., 2012), we also

investigate whether such characteristics could affect the adoption of a R–W-

like mechanism of language learning.
To achieve these goals and to address these questions, we designed a
simplified natural language learning task: simplified in order to exploit the ad-
vantage of tight empirical control, but only partly so in order to maintain a com-
mitment to ecological validity by offering a more naturalistic language input
experience. The task represents, to a reasonable extent, how people would learn
Polish subject–verb agreement mappings through natural exposure to exam-
ples.
We trained native English speakers on a set of carefully crafted examples,
which had both auditory and visual dimensions, and which incorporated some
of the complexities inherent to subject–verb agreement in Polish. Next, for
each participant, individually, we selected the best-fitting model (i.e., the pa-
rameters that led to the closest match between the responses of the participant
and the model), using the same training examples encountered by the partici-
pant. We then assessed the R–W model for its capacity to recover participants’
language choices as well as their time latencies, and compared it to other plau-
sible, yet rule-based response strategies. Finally, we tested whether cognitive
and personal characteristics such as working memory capacity, age, and gender
affect the extent to which the R–W model captures language learning.
Method
Participants
Sixty-six participants (Mdnage = 20 years; range = 18–65; 41 females) took
part in the experiment in exchange for a £7 Amazon voucher. Participants were
university students and staff. All of them were native English speakers without
knowledge of Polish or any other Slavic languages, had normal or corrected-
to-normal hearing and vision, and did not declare any learning disabilities.
Participants had different educational backgrounds, and many of them could
speak other languages in addition to English (the distributions of education
and language backgrounds are presented in Appendix S1 in the Supporting
Information online).
Materials and Procedure

All our materials, including data and code, are openly available on Github
(https://github.com/ooominds/Error-correction-mechanisms-in-language-
learning) and the University of Birmingham’s open-access repository, UBIRA
(https://doi.org/10.25500/edata.bham.00000911). Participants completed three
tasks and a short questionnaire in the following order: (a) a language learning

task (main task), (b) an explicit knowledge and demographic questionnaire,

(c) an implicit learning task, and (d) a working memory (WM) task. (A
detailed description of each task is provided in the next section.) The lan-
guage learning and implicit learning tasks were implemented and presented
to participants using OpenSesame (Mathôt et al., 2012; Mathôt & March,
2022). The demographic questionnaire was presented using Google forms,
and the WM task was administered using Tatool (von Bastian et al., 2013).
The experiment was run either individually or, whenever possible, in pairs,
in a quiet room, on Intel Core i7-8700 computers running Windows 10 and
equipped with Iiyama G-Master 24.5-in. monitors running at 59 Hz with a
screen resolution of 1,920 × 1,080 pixels. Participants heard the auditory
stimuli via Bose Quietcomfort 35 II noise-canceling headphones and regis-
tered their responses using a keyboard. The experiment took about 50 min to
complete.
Language Learning Task

Our simplified natural language learning task was inspired by the challenge of
learning subject–verb agreement in the plural past tense in Polish. In the past
tense, verbs are marked for the grammatical gender of the subject according to
the following rules:
1. If one of the referents is masculine personal (e.g., “man”), then the gen-
der of the subject as a whole is –li, which is sometimes referred to as the
masculine personal ending.
2. If the referents are feminine animate (e.g., “duck”), feminine personal
(e.g., “girl”), or neuter (e.g., “child”), then the gender of the subject
is –ły, which is sometimes referred to as the nonmasculine personal
ending.
3. Prescriptive grammars and native speakers of Polish disagree as to what
form should be assigned to a subject that includes multiple masculine
animate referents that are not persons (e.g., “the dog and the cat went
for a walk”) or mixes masculine animate and feminine personal ref-
erents (e.g., “the girl and the dog went for a walk”). Grammar text-
books prescribe the use of –ły whereas native speakers appear to fa-
vor the use of –li, according to Kiełkiewicz-Janowiak and Pawelczyk
(2014).
For the purpose of designing our simplified natural language learning task,
we implemented the first two rules but assumed that a subject consisting of
masculine animate referents should be used with the masculine personal form,

Figure 1 An example of stimuli presented in one trial to depict a scene.
as suggested by Kiełkiewicz-Janowiak and Pawelczyk (2014). We made such

a concession to improve the testability of our task from a learning perspective,
as explained in the task design below.
Stimuli
Each event in our learning task consisted of a scene that represented a joint
action performed by a group of human and/or animal characters, and for each
learning event, participants saw a picture that depicted the scene (Figure 1),
along with an audio recording of a Polish clause describing it. A new trial
started with a fixation dot that was shown at the center of the screen for about
500 ms, followed by the simultaneous display of the picture of the scene. Par-
ticipants heard the audio recording of the clause describing the scene 250 ms
after the onset of the picture of the scene while the picture remained on display.
A new trial was then presented after about 1 s.
We used the verb chodzić (“walk”), with its two possible plural past tense
forms chodziły (nonmasculine plural form) and chodzili (masculine plural
form), as the common action in all learning events. An example of a clause
heard by participants is Chłopiec i kaczka chodzili (“The boy and the duck
were walking”). The first three columns in Table 1 provide a list of all charac-
ters used in the experiment, along with their linguistic categories in terms of
gender and animacy; the last two columns concern the design of the task and
will become relevant in the next section.
The images representing the different human and animal characters were
extracted from Adobe Stock (https://stock.adobe.com) and then edited using
Adobe Photoshop CC 2018. The audio recordings of both the character labels
and the two verb forms were prepared using the speech synthesizer software
Speech2Go (Harpo Software, 2018).

Table 1 Information about the characters used in the language learning task
Character used Polish word for Linguistic category Cue Cue

in the task the character (gender & animacy) category
Duck Kaczka Feminine animate FA1 uFA

Lamb Owieczka Feminine animate FA2
Goat Koza Feminine animate FA3 bFA
Monkey Małpa Feminine animate FA4 ibFA
Girl Dziewczynka Feminine personal FP1 uFP
Woman Kobieta Feminine personal FP2
Granny Babcia Feminine personal FP3 bFP
Dog Pies Masculine animate MA1 uMA
Horse Koń Masculine animate MA2
Rabbit Królik Masculine animate MA3
Boy Chłopiec Masculine personal MP1 uMP
Man Meżczyzna
˛ Masculine animate MP2
Note. F = feminine, M = masculine, A = animate, P = personal. The lowercase let-
ters used with the cue categories reflect predictions from the Rescorla–Wagner theory
for each cue, as explained in the design section: b = blocked (i.e., the cue is blocked
according to the model); u = unblocked; ib = inhibitory blocked. For example, bFP
means that the cue is a feminine personal one and is predicted to be blocked.
Design
First, participants were taught the Polish labels of the different animal and
human characters used in the learning task. Specifically, participants were
presented with the images of all the characters along with their correspond-
ing labels, first individually and then in combination, as they appear later in
the learning task (e.g., a dog; a boy, a dog, and a monkey). There were eight
such character combinations, and participants were required to remember at
least seven of them (i.e., to reach a retention accuracy of 87.5%) before they
could proceed to the main task (see Appendix S2 in the Supporting Informa-
tion online for more details). Participants were allowed up to 10 attempts to
reach the required accuracy level.
The main task consisted of a training and a test phase. The design of the
training part of the task is summarized in Table 2. The task contained 12 cues
and two outcomes. The “+” sign indicates that the cues were presented in
compound, and the arrow symbol “→” indicates that the outcome on the right-
hand side followed the cues. Thus, for example, “FP1 + FP2 + FP3 → np”
represents a clause such as Dziewczyna, kobieta i babcia chodziły (“The girl,
the woman, and the grandma were walking”), where the subject of the clause is

Table 2 The learning events used for training
Block Learning events
Block 1 MP1 (uMP) + FA1 (uFA) → mp

MP2 (uMP) + FA2 (uFA) → mp
FA1 (uFA) + FA2 (uFA) → np
FP1 (uFP) + FP2 (uFP) → np
Block 2 FA4 (ibFA) + MP1 (uMP) + MP2 (uMP) → mp
MA1 (uMA) + MA2 (uMA) + MA3 (uMA) → mp
FA1 (uFA) + FA2 (uFA) + FA3 (bFA) → np
FP1 (uFP) + FP2 (uFP) + FP3 (bFP) → np
Note. The category of each cue is provided in parentheses after the cue. F = feminine,
M = masculine, A = animate, P = personal, mp = masculine plural verb form, np =
nonmasculine plural verb form, b = blocked cue, u = unblocked cue, ib = inhibitory
blocked cue.
made up of three female characters and the verb is in the nonmasculine plural
(np) past form, as opposed to the masculine plural (mp) past form. There were
two training blocks, each containing four events that were repeated 15 times
each. The order of the events was fully randomized within each block. The
events in the first block were composed of cue pairs, whereas those in the
second block were composed of cue triples.
We structured our task in this way to create blocking-like effects as usually
seen in Pavlovian learning experiments. For example, the addition of cues FA3
and FP3 to the compounds “FA1 + FA2” and “FP1 + FP2,” respectively, in
the second block should reduce the association strength that can be gained by
FA3 and FP3 for outcome np. Likewise, training MP1 and MP2 with outcome
mp in the first block should block FA4 from acquiring a positive association
with mp. Besides predicting that FA4 could get blocked, we also predicted that
it could become inhibitory for mp, that is, gain a negative association weight
with mp, as will be seen when we present the model fit simulation results.4 We
thus refer to FA3 and FP3 as blocked cues, and to FA4 as an inhibitory blocked
cue.
We categorized the cues into seven different categories based on their lin-
guistic properties and the blockinglike effects they predict (see the rightmost
column in Table 1). Specifically, the seven categories were based on whether
the cue is masculine or feminine, whether it is personal or animate, whether it
is predicted to be blocked or unblocked, and whether it is predicted to be an
inhibitory blocked cue. The similarity between the cues within each of these
categories is reinforced by the fact that they share the same association weights

with each outcome, according to the R–W theory, as will be shown in the Re-
sults section on learned noun–verb form association weights.
After training, the participant moved to the testing phase. The test consisted
of two components. By using a randomly generated cue from each category,
we tested learning once on all possible pairs mixing either cues from the same
cue category (e.g., FP1 + FP2 from the uFP group) or cues from different cate-
gories (e.g., MA1 + FP3 from the uMA and bFP groups). We also included the
four combinations consisting of cue triples presented in the training phase as
a sanity check for participants’ recall (these combinations were excluded from
our main analyses). Overall, in the test phase, each learner encountered in total
29 cue combinations, which were randomly selected from a total of 70 possible
cue combinations. (The exact format and instructions used while administer-
ing the task are provided in Appendix S2 in the Supporting Information online,
and the list of all test cue combinations is provided in Appendix S3.)
Finally, let us return to the question of why we adopted Kiełkiewicz-
Janowiak and Pawelczyk’s (2014) rule, whereby any subject combination that
contains a masculine referent takes the masculine personal plural form. First,
having the combination “MA1 + MA2 + MA3” associated with “mp” rather
than “np” made it possible to have a balanced number of mp and np events both
within the full task and within each block. This reduced the likelihood of any
bias towards np emerging purely due to the design. Second, this allowed us to
have more challenging combinations that better probe participants’ learning,
notably combinations intermixing feminine and masculine cues.
Analysis
From the learning task, data from three participants were discarded because
they persistently chose the same response across the test phase (27 or more
out of 29 responses; i.e., rate > 93%).5 To analyze participants’ choices and
response times, we used generalized mixed-effects modeling. The data con-
tained repeated measurements from the same participants and items on multi-
ple trials, hence we added random effects for both participants and items (i.e.,
cue combinations in the test phase). We selected the random effects structure
of the models by using a top-down strategy starting with all random intercepts
and slopes and then removing higher-order random effects step by step based
on Akaike information criterion scores. We ran the mixed-effects models in
R (R Core Team, 2019) using the lme4 package; the p values were obtained
using the lmerTest package based on Satterthwaite’s approximations, and the
model summary tables were generated using the sjPlot package. To determine
statistical significance, we used an alpha level of .05. In the analysis of

response times, we used the Box–Cox method as implemented in the car

package to transform the distribution to normality and facilitate statistical
modeling.
Explicit Knowledge and Demographic Questionnaire

After completing the language learning task, participants filled out a question-
naire that asked them whether they used any explicit rules to decide when to
use each of the two verb forms, and if they did, what these rules were. The ques-
tionnaire also collected information about participants’ gender, their age, the
languages they spoke (other than English), and their highest education level. A
full list of the questions used in the questionnaire is provided in Appendix S4
in the Supporting Information online. We focus specifically on the role of age
and gender in explaining any individual differences observed when fitting the
R–W model to the data. This is because age and gender have been shown to
affect both associative learning and second language acquisition. For example,
Mutter et al. (2012) showed that cue–outcome associations are less likely to be
acquired by older adults than by young adults. It is also well established that
older adults are less effective at learning a second language than young adults
(for a review, see Muñoz & Singleton, 2011) and experience more difficulties
with language production (Burke & Shafto, 2004). Several studies have also
reported that females show higher conditioning levels in associative learning
tasks (Lonsdorf et al., 2015; Merz et al., 2018) and acquire language more
effectively (Adani & Cepanec, 2019; van der Slik et al., 2015) than males.
Alongside the main learning task, we included a standard task of implicit
learning ability and one of working memory (WM). We selected these two
tasks because they capture salient properties of the learning setup: (a) the fact
that no explicit instructions were given; and (b) the fact that the linguistic
phenomenon can be considered discontinuous in that properties of (constel-
lations of) the agent, which is mentioned first, determine which past ending
will be used on the verb, which is mentioned second, so that some mainte-
nance of agent-related information in memory is required. Since our measure
of implicit learning ability is nonstandard and did not play a significant role
in our models, we report on this task in Appendix S5 in the online Supporting
Information.
Working Memory Task

Stimuli
To measure participants’ WM capacity, we used a slightly modified version
of the operation span test (Turner & Engle, 1989) used by Medimorec et al.

(2021). In each trial, participants were asked to retain a list of digits (between
1 and 9) presented one at a time. Each digit presentation lasted for 1 s and was
followed by a simple mathematical operation that could be either correct or
incorrect (50% of the mathematical operations were correct). Participants had
to verify the veracity of the mathematical operation before the next digit could
be displayed. At the end of each trial, they had to type in the digits in the same
order in which they had been presented to them. The length of the digit lists
increased gradually from two to eight, with each length repeated three times.
The task, thus, consisted of 21 trials.
Analysis
We calculated each participant’s WM span by first summing the number of
correct items they recalled in the correct order and then z-transforming the ob-
tained score. We excluded one participant whose WM score was discontinuous
from the rest of the sample (their WM score was −4.3 standard deviations from
the mean, whereas the second furthest WM score was −1.8 standard deviations
from the mean).
Computational Modeling
The Rescorla–Wagner Equations
The R–W model (Rescorla & Wagner, 1972) describes computationally how
the associations between cues and outcomes are established. In the context of
our experiment, a cue is the Polish label and image of one of the human or an-
imal characters appearing in the scene on a given trial, and an outcome is the
verb form describing their common action. For example, the clause Chłopiec,
meżczyzna
˛ i małpa chodzili (“The boy, the man, and the monkey were walk-
ing”) has as cues chłopiec, meżczyzna,
˛ and małpa, and as outcome chodzili. In
our case, the association weight (or strength) measures the tendency of a verb
form to occur in the presence of a certain noun.
After encountering a clause, the learner updates the association weight be-
tween a cue ci and an outcome o, depending on whether the cue and outcome
appear in the sentence, using a delta-type correction rule:
wt (ci , o) = wt−1 (ci , o) + αβδt−1
where:
⎧
⎪
⎪ 0, if ci absent
⎨λ − w c , o , if c present and o present
⎪
t−1 j i
δt =
⎪
⎪
c j present

⎪
⎩0 − wt−1 c j , o , if ci present and o absent
c j present

The subscript t refers to the present trial, thus wt (ci , o) is the association
strength between ci and o at trial t. α and β denote the learning rates for the
cue ci and outcome o respectively. λ refers to the maximum associability to an
outcome and is almost always set to 1.
Based on the equation, three cases determine how an association weight is
adjusted:
1. If the cue is absent, we make no adjustment to the weight.
2. If both the cue and outcome are present, then this provides positive evi-
dence that should strengthen the association weight, and the sum of the
weights of the cues present in the current event is adjusted towards the
maximum associability value.
3. If the cue is present but the outcome is not observed, then this provides
negative evidence that should weaken the association weight, and the
sum of weights is adjusted towards 0.
For the implementation of the model, we used the package that was developed
as part of the study by Milin et al. (2020).
Predicting Choices From the Model

To generate a verb form choice (or in the model’s terminology, an outcome)
from the model given a certain set of cues, we first calculate the activation of
each form by summing the association weights between the form and each of
the relevant cues. The predicted response from the model is then the form hav-
ing the highest activation. For example, if at a certain trial in the test phase, a
scene contained a girl and a monkey, then the activations of the masculine plu-
ral (mp; chodzili) and nonmasculine plural (np; chodziły) forms are calculated
as follows:
activ (mp) = w (girl, mp) + w (monkey, mp)
activ (np) = w (girl, np) + w (monkey, np)
where for the formulae, we used the final weights obtained at the end of the
training phase and hence omitted the trial subscripts (no learning happens in
the test phase). If activ(np) > activ(mp), the model would predict the np form,
and otherwise it would predict the mp form.
Model Fitting Procedure

In our simulations, we assumed that λ = 1 and β = 1 and considered the learn-
ing rate α as a free parameter to be estimated for each participant (henceforth,

whenever we refer to learning rate, we will always refer to the α parameter).

Specifically, we ran 50 computer simulations per participant using grid-search
for α ranging from .01 to .50. In each simulation, we programmed a virtual
agent to behave according to the R–W model and presented it with the same
training trials as the participant whose learning history we aimed to model.
From the trained model, we then generated form choices for the same trials
that the participant encountered in the test phase. We finally selected the learn-
ing rate (and hence the model) that maximized the match rate between the
participant’s observed responses and the model’s predicted responses (i.e., the
proportion of test items for which the model produced the same response as
the learner). Due to the nonidentifiability of the best-fit model, where in some
cases more than one learning rate value maximized the match rate, we selected
the median learning rate as the best parameter.
Model Evaluation
To help explain participants’ behavioral data, we derived an activation-based
measure from the fitted R–W model, which we call activation support for an
outcome. The measure aims to explain participants’ form choices and response
times, and is defined as the difference between the activation of the outcome of
interest and the activation of the remaining outcome. For example, the activa-
tion support for the nonmasculine plural form (np) is given by the following:
activation support (np) = activ (np) − activ (mp)
We hypothesized that the higher the activation support for a verb form (i.e.,
the stronger the evidence from the model supporting the verb form relative to
the other possible form), the higher the likelihood of that form being selected
by participants. We also expected that the magnitude of the activation support
would negatively correlate with participants’ response times. In other words,
the higher the magnitude of this measure, the quicker the participant’s response
would be. This should translate into a quadratic relationship between activation
support and response times, with the slowest responses expected when activa-
tion support values are near zero, and the fastest responses expected for high
positive or negative values.
Results
This section evaluates the extent to which the R–W model explains our par-
ticipants’ behavior by fitting a separate model to each participant’s data, and
tests whether the model fit quality is affected by individual differences such as
WM span, age, and gender. We first present some descriptive results on the

association weights of the fitted models, which summarize the linguistic

knowledge participants acquired in the language learning task. Next, we com-
pare the quality of fit of the model with that of other plausible, yet rule-
based response strategies. We then successively present analyses that assess
the model’s capacity to recover participants’ language choices, time latencies,
and levels of response agreement. The effect of cognitive and personal charac-
teristics on the extent to which the R–W model captures language learning is
analyzed at the end of the Results section.
Learned Noun–Verb Form Association Weights

Following the fit procedure described earlier in the section on computational
modeling, we selected the model that best captured the choices each partic-
ipant made over trials, by finding the “right” learning rate parameter (see
Appendix S6 in the online Supporting Information). Each participant was
characterized mostly by two regimes of model fit accuracy: one for learning
rates ranging roughly between .05 and .11, and one for learning rates be-
tween .12 and .50 (with some exceptions, as for Participants 12, 19, 27, and
35, for whom there were three regimes of accuracy), with neither of the two
regimes consistently leading to better model fit accuracy. Taken together, al-
though the explained variability in choices contributed by the learning rate
parameter was limited, making by-participants adjustments for that parame-
ter was still beneficial and insightful: We observed that there was not a single
learning rate value that led to the highest model fit accuracy for all partici-
pants. In other words, there appear to be considerable individual differences
in the rate of learning. Figure 2 depicts the distributions of the acquired as-
sociation weights of all possible noun–verb form pairs from the best-fitting
models.
Overall, the distributions of association weights were similar within each
cue category (e.g., MA1, MA2, and MA3 within the uMA category), rein-
forcing our grouping of the cues based on the grammatical gender and an-
imacy of the nouns they represent. Secondly, and unsurprisingly, the (un-
blocked) masculine cues gained a positive association weight with the mas-
culine plural form (i.e., these cues are more likely to result in a choice for
the masculine plural form), whereas the unblocked feminine cues gained a
positive association weight with the nonmasculine form (i.e., these cues are
more likely to result in a choice for the nonmasculine plural form). The
magnitudes of the weights also differed between participants for most of the
cues, thus creating a potential tool for capturing individual differences in our
data.

Figure 2 Histograms of cue–outcome association weights from the best-fitted

Rescorla–Wagner models. F = feminine, M = masculine, A = animate, P = personal,
mp = masculine plural verb form, np = nonmasculine plural verb form.
As predicted by (standard) blocking, the association weights between the

feminine blocked cues (i.e., FA3 and FP3) and the nonmasculine form were
more centered around zero than their unblocked counterparts (see the panes
for FA3 and FP3 in Figure 2). The blocking, however, is not yet fully re-
flected in the acquired weights, since for many participants, the association
weights between the blocked feminine cues and the nonmasculine form were
different from zero. An inhibitory blockinglike effect (i.e., a negative weight
between FA4 and the masculine plural form) appeared for about a third of
the participants. For the remaining participants, FA4 was more like a standard
blocked cue as its association weight with the masculine form was around zero.
Taken together, blockinglike effects showed tendencies in the predicted direc-
tions. We assumed that their relatively mild magnitude was because our exper-
iment captured early phases of learning, where expositions of the stimuli were

repeated only 15 times. This hypothesis was confirmed by rerunning the sim-
ulations presented in Figure 2, now with 1,000 repetitions per event, as shown
in Appendix S7 in the Supporting Information online; blocking and inhibitory
blocking effects occurred for all participants regardless of their learning rate or
event ordering. These results confirm what we pointed out before: Biases and
differences in learning are more likely to manifest early on in learning (Ellis,
2006a).
Participant–Model Match Rates

Next, we investigated to what extent these differences in learning can be cap-
tured by the R–W model if we take into account the order of events encountered
by each participant as well as differences in their learning rates. The model’s fit
accuracy (i.e., the proportion of matches between the responses from a given
participant and its best fitting R–W model) ranged from .24 to 1.00 (M = .68,
SD = .17): 17 out of 63 participants had a fit accuracy ≥ .80, and only nine
participants had a proportion of matches lower than .50. Evaluating the model
fit using leave-one-out cross validation6 shows that model fit accuracy was
equally high on unseen data, with an average accuracy of .68 (SD = .17) and
17 out of 63 participants reaching a fit accuracy ≥ .80. The fit accuracy rates
were highest for events containing a masculine personal cue (M = .74) or an
unblocked feminine personal cue (M = .68), and they were lowest for events
containing the inhibitory blocked cue (M = .61) or an animate cue (all means
≈ .65).
These results suggest a reasonably good fit of the R–W model to partici-
pants’ data, given that we considered a simple strategy for generating response
predictions based on the model activations—that is, for each event, we selected
the verb form that had the highest activation regardless of the difference in the
activation magnitudes of the two possible verb forms. We will later analyze
the sensitivity of the fitted models’ activations to the observed form choice
proportions and response times.
Comparison Between the Rescorla–Wagner Model and Other Decision

Strategies
The results presented above show that the R–W model captures our partic-
ipants’ behavior reasonably well, but how does the model compare to other
strategies that participants might have employed during the experiment? To
answer this question, we considered four decision strategies. The first two are
the prescriptive and normative strategies we presented earlier. The prescrip-
tive strategy is the one described, or prescribed, by Polish grammar books,

Figure 3 Proportion of participants that each model fitted the best. R–W = Rescorla-
Wagner.
and whereby a participant always chooses the nonmasculine verb form except
when a masculine personal cue is present (we also refer to this strategy as
the “feminine-biased” strategy). The normative strategy is the one generally
adopted by native speakers of Polish, whereby the masculine verb form is al-
ways selected except when all cues are feminine (referred to as the “masculine-
biased” strategy). We also included two basic strategies, whereby a participant
either always chooses the masculine verb form (referred to as the “masculine-
only” strategy) or always chooses the nonmasculine verb form (referred to as
the “feminine-only” strategy). The latter two strategies were included to cap-
ture participants’ behavior at the extremes.
Figure 3 displays the proportion of participants best fitted by each of
the five resulting models (R–W and our four decision strategies); we consid-
ered the model(s) with the highest participant–model match rate among the
five models as the best-fit model(s). The R–W model was the model that
best explained participants’ responses (31 out of 63 participants), followed
closely by the normative strategy (26 participants). The other three strategies
explained participants’ choices substantially less well than those two strate-
gies (< 12 participants). The fact that the R–W model and the normative
strategy were close in capturing participants’ behavior is not very surprising
since the verb forms used in the training events were selected based on the

Table 3 Fixed effects structures of the (generalized) linear mixed-effects models ex-
plaining participants’ nonmasculine plural form choices (left) and response times
(right) based on activation support from the fitted Rescorla–Wagner models
Nonmasculine plural choice Response times (transformed)
Predictors OR 95% CI p b 95% CI p
(Intercept) 0.75 [0.60, 0.95] .016 0.08 [−0.05, 0.21] .244

Activation 6.78 [3.82, 12.03] < .001 −0.02 [−0.14, 0.09] .685
support(np)
Activation −0.20 [−0.35, −0.04] .012
support(np)
squared
Note. α = .05.
normative rules and the predictions of the R–W model were largely in ac-
cordance with the normative strategy (Figure 3). It is interesting, though, that
the R–W model managed to learn this strategy implicitly without any prior
experience based on a simple general learning rule. The average percent-
age of response matches between the R–W model and the normative strat-
egy per participant was above 90%, and the average percentage of response
matches between the R–W model and the prescriptive strategy was above
85%.
Relationship Between the Model’s Activation-Based Measures and

Participants’ Choices and Response Times
In a further assessment of the quality of fit of the R–W model, we carried out
a generalized linear mixed-effects modeling analysis looking at the relation-
ship between participants’ choices and the activation-based measure derived
from the fitted R–W models—that is, activation support for the np form (see
the section on computational modeling for more details). We also analyzed
the relationship between participants’ response times and activation support
by fitting a polynomial linear mixed-effects model with both the linear and
quadratic terms of activation support, since we expected activation support to
have a quadratic effect on response times (Table 3). More detailed summaries
of the models with the random effects structures are provided in Appendix S8
in the Supporting Information online.
As expected, activation support for the nonmasculine plural form was
significantly positively associated with the likelihood of nonmasculine form

Figure 4 Relationship between the Rescorla–Wagner activation support for the non-
masculine plural form and the proportion of nonmasculine plural choices made by par-
ticipants (left), and relationship between the Rescorla–Wagner activation support for
the nonmasculine plural form and participants’ response times (right).
choices, OR = 6.78, p < .001, 95% CI [3.82, 12.03]. Figure 4 (left pane)
also shows that this relationship is asymmetrical around 0, reflecting a strong
bias towards the masculine verb form that, even with (activation-based) ev-
idence supporting the nonmasculine form, can still lead to a preference for
the masculine form. Also, and in line with our hypotheses, the second-order
polynomial term of activation support was a significant predictor of response
time, as there was a quadratic relationship between the activation support and
response time, with the slowest responses recorded for the least supported
events, b = −0.20, p = .012, 95% CI [−0.35, −0.04]; see also Figure 4, right
pane.
These results suggest that the fitted models performed well in predicting
participants’ form choices, and that the information encoded in the associa-
tion weights—the basic currency of a R–W model—is a good predictor of
both the likelihood of choosing a particular verb form and the speed with
which the response is made. Participants’ level of agreement regarding the
choice of a certain verb form thus differed depending on the activation sup-
port of that particular form, with a high level of agreement expected and at-
tested for high positive or low negative activation support values and with a

high level of disagreement expected and attested for activation support around
zero.
Level of Agreement Between Participants Through the Lens of the Model

We further analyzed participants’ behavior by exploring two additional ques-
tions. First, what level of agreement was there among language learners,
given a particular type of event (e.g., events made up of cues from the same
grammatical gender versus events intermixing cues of different grammatical
gender)? Second, and crucially, can the differences in levels of agreement be
adequately explained using the R–W model?
To answer these questions, we analyzed the effect of the presence of each
event category on the proportion of participants who chose one of the two verb
forms (for a full list of event categories used in the task, see Appendix S3
in the Supporting Information online). To obtain the model estimates of the
proportions, we used the generalized linear mixed-effects model that we built
in the previous analysis (Table 3, left) to model the relationship between ac-
tivation support and the choice of the nonmasculine plural form. Specifically,
for each event category, we averaged the model’s predicted proportions of the
nonmasculine forms based on its activation support values across participants
(given that each event category was encountered once by each participant).
The results are summarized in Figure 5. The events in the figure are sorted in
ascending order by observed choice proportions in the experiment, so the left-
most and rightmost sides represent regions of high level of agreement between
participants, whereas events situated in the middle part triggered a high level
of disagreement between participants.
Participants had a clear preference for the masculine plural form when
the event contained a masculine personal cue (uMP) or when the event in-
cluded only animate masculine cues (i.e., “uMA1 + uMA2”). Likewise, par-
ticipants clearly preferred the nonmasculine plural form when the events
solely contained feminine cues, whether personal or animate (e.g., “uFA +
uFP”). High levels of disagreement were mainly observed for events inter-
mixing the inhibitory blocked cue and a feminine cue or a masculine ani-
mate and a feminine personal cue (with proportions ranging between .44 and
.51).
Comparing the observed and predicted proportions, the model managed to
capture the difference in the levels of agreement between participants across
the different categories of events surprisingly well. The largest discrepancies
between the observed and predicted proportions appear to have occurred for
events involving the blocked and inhibitory blocked cues, suggesting that it

Figure 5 Observed and predicted proportion of participants choosing the nonmascu-

line plural form given a certain cue combination. F = feminine, M = masculine, A =
animate, P = personal, b = blocked cue, u = unblocked cue, ib = inhibitory blocked
cue.
was more challenging for the model to capture their effect on participants’
choices.
Relationship Between Model-Fit Quality and Individual Difference Measures

The model fit results demonstrated that the R–W model successfully accounts
for the behavior of a large proportion of our participants, including their re-
sponse times. However, the quality of fit varied across participants, with data
from a few participants fitted very poorly by the model. In an attempt to as-
sess whether the other measures collected during the experiment (demographic
characteristics, WM span, and implicit learning) can explain the observed in-
dividual differences in the model fit quality, we ran a multiple linear regres-
sion to predict participant–model match rates (logit-transformed) based on the
different individual difference measures collected. Specifically, the predictors
included WM span (z-transformed), gender, and age; the time-slopes extracted
for each participant from the implicit learning task did not make a significant

Table 4 Summary of the linear regression model assessing the effect of working mem-
ory (WM) span and gender on the proportion of participant–model matches
Participant–model match rate (transformed)
Predictors b 95% CI p
(Intercept) 0.59 [0.44, 0.75] < .001

WM z score 0.17 [0.05, 0.30] .008
Gender = male −0.28 [−0.54, −0.02] .033
Observations 61
R2 /R2 adjusted .148/.118
Note. α = .05.
contribution and are reported on in Appendix S5 in the Supporting Information

online.
From the full linear regression model that included all four individual mea-
sure variables as well as the interaction between gender and WM span, we de-
rived a final model containing only the significant variables by using backward
variable selection based on the likelihood ratio test. Data from two participants
were omitted from this analysis because one did not report their gender and
one was identified as an extreme outlier, as explained in the section on the
WM task. In total, data from 61 participants were fed to the linear regression
model. The best model after variable selection included both WM span and
gender, but not their interaction (Table 4).
The model fit accuracy increased significantly with increasing WM span,
b = 0.17, p = .008, 95% CI [0.05, 0.30], as illustrated in Figure 6 (left pane),
and was significantly higher for female participants (M = .69, SD = .17) in
comparison with male participants (M = .63, SD = .16), b = −0.28, p =
.033, 95% CI [−0.54, −0.02], as shown in Figure 6 (right pane). To check that
removing influential residuals did not affect our findings, we applied model
criticism as described in Baayen and Milin (2010); excluding the single ex-
treme residual with a z score greater than 2.5 in our model resulted in even
stronger effects (p = .004 for WM span and p = .007 for gender). Also,
WM span did not significantly correlate with gender (p = .058), which sug-
gests that their relationship is unlikely to have influenced the effects of gen-
der and WM span on model fit accuracy (see Appendix S9 in the Support-
ing Information online for more detail). Our findings thus show that having a
larger WM capacity or being female increased the likelihood of a participant

Figure 6 Effect of a participant’s (z-transformed) working memory (WM) span (left)

and gender (right) on the rate of matches between their responses and the predicted
responses from their best-fitting model. R–W = Rescorla–Wagner.
choosing verb forms in accordance with the R–W model in our language learn-
ing task.
Discussion
Summary of Findings
Our findings show that a R–W mechanism captures well how participants learn
subject–verb agreement in a morphologically complex language and, by ex-
tension, how they might learn language through mere exposure to it. With an
average fit accuracy of 68%, based on a simple activation-based decision strat-
egy, the model explained the verb form choices made by a large proportion
of participants rather well.7 More interestingly, an activation-based measure
extracted from the best-fitting models correlated strongly with both the like-
lihood of a particular verb form choice and the time required to make that
choice.
The model also provided insights as to why participants might display high
or low agreement levels when choosing a verb form, depending on the nature
of the subject of the clause. According to the model, this is due to the associ-
ation strengths that the participants acquire, which are used to calculate the

activation support for each of the possible verb forms. These association
strengths are mostly affected by (a) the learner’s learning rate for the cues (the
learning rate determines the magnitude of the correction of the weights, based
on the estimated error in each trial) and (b) the distribution of the learning
events they encountered during the learning stage (this would include the fre-
quency of each learning event and the order of the learning events, among
other things). Thus, one prediction from our study is that changing the order or
the relative frequencies of the learning events during the training might lead to
different choice patterns from those we observed here.
We also found a significant relationship between both gender and WM ca-
pacity and the participant–model match rates, which sheds light on what might
have driven the observed differences in the quality of model fit. The fact that
in our experiment a larger proportion of women than men acted in accordance
with a R–W mechanism is in line with findings from several previous stud-
ies that highlighted the association between gender and classical condition-
ing for both humans (Lonsdorf et al., 2015; Merz et al., 2018) and animals
(e.g., Velasco et al., 2019). This suggests a significant difference in learning
between men and women, with women being better modeled by the R–W error-
correction learning rule. Women are generally known to have a small language
advantage over men (see Kimura, 1999, for an extensive assessment), specifi-
cally in areas pertaining to lexical retrieval (Balling & Baayen, 2008, 2012). It
has been suggested that this might be due to women having a superior declara-
tive memory, which they could use to generalize over stored neighboring forms
(Hartshorne & Ullman, 2006).
The finding that the likelihood of a language learner behaving accord-
ing to the R–W mechanism increases with WM capacity provides evidence
that WM can play a role in classical conditioning by affecting the adoption
of a classical conditioning mechanism such as the R–W rule. Sasaki (2009)
and Baetu et al. (2018) previously provided evidence of disruption of classi-
cal conditioning performance when WM is loaded using dual-task paradigms.
The present finding adds to the mounting evidence that, against the predom-
inant belief, WM may be implicated in low-level cognitive processes such as
instrumental learning, more commonly referred to as reinforcement learning
within the neuroscience and machine learning communities (Collins & Frank,
2012; Ez-zizi, 2016) and in some forms of implicit learning (Medimorec et al.,
2021).
Blocking and inhibitory blocking-like effects did not emerge from the R–W
model for all participants. As shown, this was mainly due to the short duration
of the training phase. Increasing the number of training trials not only resulted

in the reemergence of blocking effects for all participants’ best-fitting parame-

ters, but also removed the variability in the association weights, thus predicting
that all participants should end up behaving in the same way in the long run
(see Appendix S7 in the Supporting Information online). This is not surprising
in light of Danks’ (2003) work, which shows that, in many cases, the R–W sys-
tem will converge to the same equilibrium regardless of the parameters used. In
other words, the destination of learners is often the same (this is also desirable
since we often want all learners to learn to make the same associations), but
the paths to those destinations can differ substantially depending on the nature
of the learning problem (e.g., amount of data available, relative frequencies
of the events, number of cues and outcomes). This has many implications for
language learning studies employing the R–W model, as discussed in the next
section.
Implications for Language Learning Studies That Use the

Rescorla–Wagner Model
Several language research groups have embraced the R–W model as a valuable
approach for modeling language learning phenomena due to its simplicity, its
cognitive plausibility, and its successes in explaining a wide range of language
learning phenomena. The present study provides further support for simulating
or modeling language learning using this model, but also draws attention to
the important issue of individual differences, which has so far been overlooked
in studies that combine computational modeling using the R–W model and
experimental data.
Given the amount of individual difference we observed in our data, it would
be prudent to move away from the currently predominant approach where the
R–W model is run once with its default parameter values and used to explain
data from all participants. Although the effect of the learning rate parame-
ter on model fit accuracy was not substantial for the chosen task, in practice,
R–W performance will always be affected by the choice of learning rate, irre-
spective of the particular modeling challenge (Milin, Divjak, & Baayen, 2017,
pp. 1739–1741).
Another common practice that might need to be reconsidered is training the
model on one large dataset or a small subset of it. Consequently, the features
that set the model apart from purely statistical classification models—namely,
the possibility of choosing parameters that capture how fast an individual can
learn, and the ability to account for input order effects—remain unused. It is
not surprising, then, that several studies have reported only minor or no differ-
ences between the R–W model and other statistical or learning models such as

logistic regression, memory-based learning, and decision trees (Baayen, 2011;

Baayen et al., 2013). With such an approach, the main advantage of the R–
W model over purely statistical techniques is its ability to perform incremen-
tal learning, as data become available. However, the same advantage could be
achieved from any neural network model with no hidden layer as such a model
can also produce weights between cues and outcomes, albeit with a different
and, arguably, less plausible learning rule (here we mainly allude to the back-
propagation learning rule, which is currently the predominant approach when
training neural network models). With such a model, the same model fitting
approach we used here can still be applied but with a different set of param-
eters to tune, such as the type of activation function, number of neurons, and
learning rate.
Recent work in usage-based frameworks has highlighted the vast individual
differences characterizing language knowledge in first language populations.
Individual differences in grammar have been found to be comparable in size to
those in lexical knowledge and are related to both the quality of the input and
the learner’s cognitive abilities (Dabrowska,
˛ 2018). Our findings demonstrate
that individual model fitting should be the default option when comparing the
R–W model or other computational models to participants’ data. Specifically,
model parameters and data inputs should be adjusted separately for each par-
ticipant to allow for a better account of individual responses and obtain a more
veridical picture of where knowledge of, for example, a rule resides, whether
in the aggregated mind of the linguist or in the individual minds of the users
(Divjak, 2018), and how it is distributed across the population. Our data also
support the usage-based stance that our linguistic knowledge is shaped by our
personal and cognitive characteristics, as attested by the significant role of WM
and gender in the quality of model fit; such factors should be considered by de-
fault when modeling language.
Limitations and Future Directions

Our study is the first to fit the R–W model to the behavior of individual learners
in an actual language learning task. We used the R–W model in the form avail-
able now, but our findings, despite being very promising, show that it might
be interesting to extend the model to handle WM capacity limitations. Fur-
ther investigations in such a direction could be inspired by work done in re-
inforcement learning—a closely related field to classical conditioning—where
learning has also long been assumed to occur in areas, often associated with
low-level cognitive functions, such as the dorsal and ventral striatum (Balleine
& O’Doherty, 2010), but is now recognized to also involve high-level cognitive

control via WM. This has led to the development of new learning frameworks
where WM is explicitly modeled as a key component that supports learning by
retaining information from previous trials (e.g., see Collins & Frank, 2012; Ez-
zizi et al., 2015). This could be the approach to take for R–W and other clas-
sical conditioning models, especially because in large simulation-based lan-
guage studies, learning events typically contain a large number of cues (e.g.,
all trigraphs or words in one sentence), which cannot be processed at once by
a human learner—as is required in the updates of the R–W model—due to
known WM capacity limitations (see Glautier, 2013, and Baayen et al., 2016,
for early attempts in this direction).
Another direction for future extension of our work is to collect partici-
pants’ responses over time while they are trained on the cue–outcome associ-
ations rather than having a separate postlearning test phase. This would have
the potential to improve the model fit further and to provide a broader picture
of the behavior of participants while they are learning the task. In addition, this
could allow the extraction of a learning measure based on time slope for the
language learning task, such as we did for the implicit learning task, and thus
would increase the likelihood of finding a link between implicit learning and
the quality of fit of the R–W model (see also our discussion in Appendix S5
in the Supporting Information online). A link between the two measures can
also be probed by fitting the R–W model to the response times collected in the
implicit learning task, as was done in Notaro et al. (2018), rather than using
time slopes only or a mixture of the two.
Finally, the particular structure of our language learning task favored the
normative (masculine-biased) strategy, but an interesting question that remains
unanswered is whether we can use the R–W model to predict the emergence of
different strategies as we vary the structure of the language input and control
for individual differences among language learners. The approach of using the
R–W model to explain or predict the level of agreement among language users
can be extended beyond Polish subject–verb agreement in the plural past tense
to cover other facets of language where a lack of consensus in language use
has been observed (e.g., see Geeraert et al., 2020; Milin, Divjak, & Baayen,
2017).
Conclusion
The R–W model is a very simple learning model, yet it has multiple sources of
variation that can be used to explain participants’ behavior in language learn-
ing experiments. These include the model’s learning rate, the order of pre-
sentation of learning examples, and the relative frequencies of cue–outcome

cooccurrences. In the present study, we systematically incorporated these

sources of variation when fitting the model to participants’ data, thus enabling
the model to successfully capture the choices and response latencies of most
participants in a language learning task on Polish subject–verb agreement. In
addition, cognitive and demographic characteristics such as WM and gender
determined the extent to which language learning was driven by R–W-like
learning principles.
Final revised version accepted 26 January 2023
Open Research Badges
This article has earned Open Data and Open Materials badges for making
publicly available the digitally-shareable data and the components of the re-
search methods needed to reproduce the reported procedure and results. All
data and materials that the authors have used and have the right to share
are available at https://github.com/ooominds/Error-correction-mechanisms-
in-language-learning and https://doi.org/10.25500/edata.bham.00000911. All
proprietary materials have been precisely identified in the manuscript.
Notes
1 The idea of cue competition is also at the core of the competition model of Bates
and MacWhinney (1987) for language acquisition. Their model, however, uses
mainly symbolic/linguistic cues such as word order or morphological features of
words and is based on a connectionist approach requiring a much more complex
architecture than the R–W model.
2 The contents of any corpus are, at best, a very rough approximation of the input that
language users receive. Conversely, artificial languages are illustrative and
informative for understanding natural languages but hardly a realistic reflection of
the complexity found in any given natural language.
3 The early implementations of the R–W rule as the naïve discrimination learning
model relied on a noniterative version of the algorithm, as provided by Danks
(2003), which eliminates the possibility of any order effects emerging.
4 It is important to note that here we were not interested in testing the blocking effects
per se as is typically done in behavioral experiments of classical conditioning. In
those experiments, only the events relevant to blocking are included (blocking is
tested separately from the other effects), and blocking is tested on a second cue
rather than a third cue as in our case (e.g., Kamin, 1969). Also, the learner is
usually trained for long enough to ensure that the “blocking” cue becomes a good
predictor of the outcome of interest. Such a clean experimental setup would not

fairly represent the “disarray” so pervasive in natural languages. As our study is

about language learning, we opted to mimic a realistic learning situation as closely
as possible.
5 An additional participant experienced equipment malfunction in the middle of the
test phase and had to retake the test. We retained this participant’s data since they
did not go through any extra training in the second run of the test phase and thus
started the new run with the same knowledge as in the first run (recall that the
outcome feedback is only provided in the training phase). The familiarization effect
should not play a role here since all participants underwent a few practice trials in
all phases before starting the actual experiment. We also reran all analyses without
this participant’s data and confirmed that removing them did not alter any of the
results presented in the paper.
6 The quality of model fit on unseen data was evaluated using leave-one-out
cross-validation. Specifically, for each participant, we held out one event of those
they encountered in the test phase and fitted a R–W model to the remaining events.
The model was then evaluated only on the reserved event by assessing whether the
response from the participant matched that of its best-fitting R–W model. We
repeated this for all test events and then computed the average (leave-one-out) fit
accuracy, that is, the proportion of matches between the responses from a given
participant on each of the reserved events and their associated predictions from the
participant’s best-fitting R–W model.
7 In fact, this accuracy level can be considered excellent since we assumed a very
simple nonprobabilistic action selection process where a verb form is chosen if it
has the highest activation. This does not take into account the variability that might
arise from exploration, lapse of attention, or inherent brain noise.
References
Adani, S., & Cepanec, M. (2019). Sex differences in early communication
development: Behavioral and neurobiological indicators of more vulnerable
communication system development in boys. Croatian Medical Journal, 60(2),
141–149. https://doi.org/10.3325/cmj.2019.60.141
Ambridge, B., & Lieven, E. V. M. (2011). Child language acquisition: Contrasting
theoretical approaches. Cambridge University Press.
Baayen, R. H. (2011). Corpus linguistics and naive discriminative learning. Revista
Brasileira de Linguística Aplicada, 11(2), 295–328.
https://doi.org/10.1590/S1984-63982011000200003
Baayen, R. H., Endresen, A., Janda, L. A., Makarova, A., & Nesset, T. (2013). Making
choices in Russian: Pros and cons of statistical methods for rival forms. Russian
Linguistics, 37(3), 253–291. https://doi.org/10.1007/s11185-013-9118-6
Baayen, R. H., & Milin, P. (2010). Analyzing reaction times. International Journal of
Psychological Research, 3(2), 12–28. https://doi.org/10.21500/20112084.807

Baayen, R. H., Milin, P., Filipović Ðurđević, D., Hendrix, P., & Marelli, M. (2011). An
amorphous model for morphological processing in visual comprehension based on
naive discriminative learning. Psychological Review, 118(3), 438–481. https://doi.
org/10.1037/a0023851
Baayen, R. H., Shaoul, C., Willits, J., & Ramscar, M. (2016). Comprehension without
segmentation: A proof of concept with naive discriminative learning. Language,
Cognition and Neuroscience, 31(1), 106–128. https://doi.org/10.1080/23273798.
2015.1065336
Baetu, I., Burns, N., & Child, B. (2018). Individual differences in working memory
capacity predict performance on an associative learning task [Paper presentation].
Australian Psychologist, Sydney, Australia. https://doi.org/10.1111/ap.12372
Balleine, B. W., & O’Doherty, J. P. (2010). Human and rodent homologies in action
control: Corticostriatal determinants of goal-directed and habitual action.
Neuropsychopharmacology, 35(1), 48–69. https://doi.org/10.1038/npp.2009.131
Balling, L. W., & Baayen, R. H. (2008). Morphological effects in auditory word
recognition: Evidence from Danish. Language and Cognitive Processes, 23(7–8),
1159–1190. https://doi.org/10.1080/01690960802201010
Balling, L. W., & Baayen, R. H. (2012). Probability and surprisal in auditory
comprehension of morphologically complex words. Cognition, 125(1), 80–106.
https://doi.org/10.1016/j.cognition.2012.06.003
Bandura, A. (1962). Social learning through imitation. In M. R. Jones (Ed.), Nebraska
Symposium on Motivation (pp. 211–274). University of Nebraska Press.
Bates, E., & MacWhinney, B. (1987). Competition, variation, and language learning.
In B. MacWhinney (Ed.), Mechanisms of language aquisition (pp. 157–193).
Lawrence Erlbaum.
Burke, D. M., & Shafto, M. A. (2004). Aging and language production. Current
Directions in Psychological Science, 13(1), 21–24. https://doi.org/10.1111/j.0963-
7214.2004.01301006.x
Bybee, J., & McClelland, J. L. (2005). Alternatives to the combinatorial paradigm of
linguistic theory based on domain general principles of human cognition. The
Linguistic Review, 22(2–4), 381–410. https://doi.org/10.1515/tlir.2005.22.2-4.381
Chen, Z., Haykin, S., Eggermont, J. J., & Becker, S. (2008). Correlative learning: A
basis for brain and adaptive systems. Wiley.
Chomsky, N. (1959). A review of BF Skinner’s Verbal behavior. Language, 35(1),
26–58. https://doi.org/10.4159/harvard.9780674594623.c6
Chuang, Y.-Y., Bell, M. J., Banke, I., & Baayen, R. H. (2021). Bilingual and
multilingual mental lexicon: A modeling study with linear discriminative learning.
Language Learning, 71(S1), 219–292. https://doi.org/10.1111/lang.12435
Collins, A. G. E., & Frank, M. J. (2012). How much of reinforcement learning is
working memory, not reinforcement learning? A behavioral, computational, and
neurogenetic analysis. The European Journal of Neuroscience, 35(7), 1024–1035.
https://doi.org/10.1111/j.1460–9568.2011.07980.x

Dabrowska,
˛ E. (2018). Experience, aptitude and individual differences in native
language ultimate attainment. Cognition, 178, 222–235. https://doi.org/10.1016/j.
cognition.2018.05.018
Dabrowska,
˛ E., & Divjak, D. (2015). Introduction. In E. Dabrowska
˛ & D. Divjak
(Eds.), Handbook of cognitive linguistics (pp. 1–9). Walter de Gruyter.
Danks, D. (2003). Equilibria of the Rescorla–Wagner model. Journal of Mathematical
Psychology, 47(2), 109–121. https://doi.org/10.1016/S0022-2496(02)00016-0
Divjak, D. (2018). Binding scale dynamics. In D. Van Olmen, T. Mortelmans, & F.
Brisard (Eds.), Aspects of linguistic variation (pp. 9–42). De Gruyter Mouton.
Divjak, D. (2019). Frequency in language: Memory, attention, and learning.
Cambridge University Press.
Divjak, D., & Gries, S. T. (Eds.). (2012). Frequency effects in language representation.
De Gruyter.
Divjak, D., Milin, P., Ez-zizi, A., Józefowski, J., & Adam, C. (2021). What is learned
from exposure: An error-driven approach to productivity in language. Language,
Cognition and Neuroscience, 36(1), 60–83. https://doi.org/10.1080/23273798.2020.
1815813
Ellis, N. C. (2006a). Language acquisition as rational contingency learning. Applied
Linguistics, 27(1), 1–24. https://doi.org/10.1093/applin/ami038
Ellis, N. C. (2006b). Selective attention and transfer phenomena in L2 acquisition:
Contingency, cue competition, salience, interference, overshadowing, blocking, and
perceptual learning. Applied Linguistics, 27(2), 164–194. https://doi.org/10.1093/
applin/aml015
Ellis, N. C., Römer, U., & O’Donnell, M. B. (2016). Usage-based approaches to
language acquisition and processing: Cognitive and corpus investigations of
construction grammar. Wiley-Blackwell.
Ellis, N. C., & Sagarra, N. (2010). The bounds of adult language acquisition: Blocking
and learned attention. Studies in Second Language Acquisition, 33(4), 553–580.
Ellis, N. C., & Sagarra, N. (2011). Learned attention in adult language acquisition: A
replication and generalization study and meta-analysis. Studies in Second Language
Acquisition, 33(4), 589–624. https://doi.org/10.1017/S0272263111000325
Ez-zizi, A. (2016). Reinforcement learning in partially observable tasks: State
uncertainty and memory dependence [Doctoral thesis, University of Bristol].
EThOS. https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.707711
Ez-zizi, A., Farrell, S., & Leslie, D. (2015). Bayesian reinforcement learning in
Markovian and non-Markovian tasks. IEEE Symposium Series on Computational
Intelligence (pp. 579–586). https://doi.org/10.1109/SSCI.2015.91
Geeraert, K., Newman, J., & Baayen, R. H. (2020). Variation within idiomatic
variation: Exploring the differences between speakers and idioms. East European
Journal of Psycholinguistics, 7(2), 9–27. https://doi.org/10.29038/eejpl.2020.7.2.
gee

Glautier, S. (2013). Revisiting the learning curve (once again). Frontiers in

Psychology, 4, Article 982. https://doi.org/10.3389/fpsyg.2013.00982
Gries, S. T., & Divjak, D. (Eds.). (2012). Frequency effects in language learning and
processing. De Gruyter.
Harpo Software. (2018). Speech2Go, Ivona, Nuance—Voices to Go—Text-to-Speech.
http://speech2go.net
Hartshorne, J. K., & Ullman, M. T. (2006). Why girls say ‘holded’ more than boys.
Developmental Science, 9(1), 21–32. https://doi.org/10.1111/j.1467-7687.2005.
00459.x
Kamin, L. J. (1969). Predictability, surprise, attention and conditioning. In B. A.
Campbell & R. M. Church (Eds.), Punishment and aversive behaviour (pp.
279–296). Appleton-Century-Crofts.
Kiełkiewicz-Janowiak, A., & Pawelczyk, J. (2014). Language and gender research in
Poland. In S. Ehrlich, M. Meyerhoff, & J. Holmes (Eds.), The handbook of
language, gender, and sexuality (2nd ed., pp. 353–377). Wiley-Blackwell.
Kimura, D. (1999). Sex and cognition. MIT press.
Klavan, J., & Divjak, D. (2016). The cognitive plausibility of statistical classification
models: Comparing textual and behavioral evidence. Folia Linguistica, 50(2),
355–384. https://doi.org/10.1515/flin-2016-0014
Langacker, R. W. (1987). Foundations of cognitive grammar: Vol. 1. Theoretical
prerequisites. Stanford University Press.
Lonsdorf, T. B., Haaker, J., Schümann, D., Sommer, T., Bayer, J., Brassen, S., Bunzeck,
N., Gamer, M., & Kalisch, R. (2015). Sex differences in conditioned stimulus
discrimination during context-dependent fear learning and its retrieval in humans:
The role of biological sex, contraceptives and menstrual cycle phases. Journal of
Psychiatry and Neuroscience, 40(6), 368–375. https://doi.org/10.1503/140336
Mathôt, S., & March, J. (2022). Conducting linguistic experiments online with
OpenSesame and OSWeb. Language Learning, 72(4), 1017–1048. https://doi.
org/10.1111/lang.12509
Mathôt, S., Schreij, D., & Theeuwes, J. (2012). OpenSesame: an open-source,
graphical experiment builder for the social sciences. Behavior Research Methods,
44(2), 314–324. https://doi.org/10.3758/s13428-011-0168-7
Medimorec, S., Milin, P., & Divjak, D. (2021). Working memory affects anticipatory
behavior during implicit pattern learning. Psychological Research, 85, 291–301
(2021). https://doi.org/10.1007/s00426-019-01251-w
Merz, C. J., Kinner, V. L., & Wolf, O. T. (2018). Let’s talk about sex… differences in
human fear conditioning. Current Opinion in Behavioral Sciences, 23, 7–12.
https://doi.org/10.1016/j.cobeha.2018.01.021
Milin, P., & Blevins, J. P. (2020). Paradigms in morphology. In Oxford Research
Encyclopedia of Linguistics. https://doi.org/10.1093/acrefore/9780199384655.
013.551

Milin, P., Divjak, D., & Baayen, R. H. (2017). A learning perspective on individual
differences in skilled reading: Exploring and exploiting orthographic and semantic
discrimination cues. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 43(11), 1730–1751. https://doi.org/10.1037/xlm0000410
Milin, P., Feldman, L. B., Ramscar, M., Hendrix, P., & Baayen, R. H. (2017).
Discrimination in lexical decision. PloS One, 12(2), Article e0171935. https://doi.
org/10.1371/journal.pone.0171935
Milin, P., Madabushi, H. T., Croucher, M., & Divjak, D. (2020). Keeping it simple:
Implementation and performance of the proto-principle of adaptation and learning
in the language sciences. arXiv:2003.03813. https://doi.org/10.48550/arXiv.2003.
03813
Muñoz, C., & Singleton, D. (2011). A critical review of age-related research on L2
ultimate attainment. Language Teaching, 44(1), 1–35. https://doi.org/10.1017/S026
1444810000327
Mutter, S. A., Atchley, A. R., & Plumlee, L. M. (2012). Aging and retrospective
revaluation of causal learning. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 38(1), 102–117. https://doi.org/10.1037/a0024851
Notaro, G., van Zoest, W., Melcher, D., & Hasson, U. (2018). Prediction and
information integration determine subtle anticipatory fixation biases. bioRxiv.
https://doi.org/10.1101/252809
Olejarczuk, P., Kapatsinski, V., & Baayen, R. H. (2018). Distributional learning is
error-driven: The role of surprise in the acquisition of phonetic categories.
Linguistics Vanguard, 4(s2), Article 20170020. https://doi.org/10.1515/lingvan-
2017-0020
Pavlov, I. P. (1927). Conditioned reflexes: An investigation of the physiological activity
of the cerebral cortex. Oxford University Press.
Pearce, J. M., & Hall, G. (1980). A model for Pavlovian learning: Variations in the
effectiveness of conditioned but not of unconditioned stimuli. Psychological
Review, 87(6), 532–552.
Pirrelli, V., Marzi, C., Ferro, M., Cardillo, F. A., Baayen, H. R., & Milin, P. (2020).
Psycho-computational modelling of the mental lexicon. In V. Pirrelli, I. Plag, & W.
U. Dressler (Eds.), Word knowledge and word usage (pp. 23–82). De Gruyter
Mouton.
Plaut, D. C., & Gonnerman, L. M. (2000). Are non-semantic morphological effects
incompatible with a distributed connectionist approach to lexical processing?
Language and Cognitive Processes, 15(4–5), 445–485. https://doi.org/10.1080/01
690960050119661
R Core Team. (2019). R: A language and environment for statistical computing
(Version 3.6.2) [Computer software]. R Foundation for Statistical Computing.
http://www.R-project.org

Ramscar, M., & Yarlett, D. (2007). Linguistic self-correction in the absence of

feedback: A new approach to the logical problem of language acquisition. Cognitive
Science, 31(6), 927–960. https://doi.org/10.1080/03640210701703576
Rescorla, R. A. (1988). Pavlovian conditioning: It’s not what you think it is. American
Psychologist, 43(3), 151–160. https://doi.org/10.1037/0003-066X.43.3.151
Rescorla, R. A., & Wagner, A. R. (1972). A theory of Pavlovian conditioning:
Variations in the effectiveness of reinforcement and nonreinforcement. In A. H.
Black & W. F. Prokasy (Eds.), Classical conditioning II: Current research and
theory (Vol. 2, pp. 64–99). Appleton-Century-Crofts.
Sasaki, T. (2009). The role of the central executive in associative learning.
Psychologia, 52(1), 80–90. https://doi.org/10.2117/psysoc.2009.80
Seidenberg, M. S., & McClelland, J. L. (1989). A distributed, developmental model of
word recognition and naming. Psychological Review, 96(4), 523–568. https://doi.
org/10.1037/0033-295x.96.4.523
Shanks, D. R. (1995). The psychology of associative learning. Cambridge University
Press.
Siegel, S., & Allan, L. G. (1996). The widespread influence of the Rescorla–Wagner
model. Psychonomic Bulletin & Review, 3(3), 314–321. https://doi.org/10.3758/
BF03210755
Skinner, B. F. (1938). The behavior of organisms: An experimental analysis.
Appleton-Century.
Skinner, B. F. (1957). Verbal behavior. Appleton-Century-Crofts.
Sturdy, C. B., & Nicoladis, E. (2017). How much of language acquisition does operant
conditioning explain? Frontiers in Psychology, 8, Article 1918. https://doi.org/
10.3389/fpsyg.2017.01918
Trimmer, P. C., McNamara, J. M., Houston, A. I., & Marshall, J. A. (2012). Does
natural selection favour the Rescorla–Wagner rule? Journal of Theoretical Biology,
302, 39–52. https://doi.org/10.1016/j.jtbi.2012.02.014
Turner, M. L., & Engle, R. W. (1989). Is working memory capacity task dependent?
Journal of Memory and Language, 28(2), 127–154. https://doi.org/10.1016/0749-
596X(89)90040-5
van der Slik, F. W. P., van Hout, R. W. N. M., & Schepens, J. J. (2015). The gender gap
in second language acquisition: Gender differences in the acquisition of Dutch
among immigrants from 88 countries with 49 mother tongues. PloS One, 10(11),
Article e0142056. https://doi.org/10.1371/journal.pone.0142056
Velasco, E. R., Florido, A., Milad, M. R., & Andero, R. (2019). Sex differences in fear
extinction. Neuroscience & Biobehavioral Reviews, 103, 81–108.
https://doi.org/10.1016/j.neubiorev.2019.05.020
von Bastian, C. C., Locher, A., & Ruflin, M. (2013). Tatool: A Java-based open-source
programming framework for psychological studies. Behavior Research Methods,
45(1), 108–115. https://doi.org/10.3758/s13428-012-0224-y

Supporting Information
Additional Supporting Information may be found in the online version of this
article at the publisher’s website:
Accessible Summary
Appendix S1. Distributions of Participants’ Educational and Language Back-
grounds.
Appendix S2. Experimental Procedure for the Language Learning Task.
Appendix S3. Cue Combinations Used in the Test Phase of the Language
Learning Task.
Appendix S4. Explicit Knowledge and Demographic Questionnaire.
Appendix S5. Implicit Learning Task.
Appendix S6. Effect of Learning Rate Parameter on Model Fit Accuracy.
Appendix S7. Long-Run Simulations to Assess Blocking Effects.
Appendix S8. Generalized Linear Mixed-Effects Models Explaining Partici-
pants’ Choices and Response Times.
Appendix S9. Correlations Between the Different Individual Difference Mea-
sures.

مقاله

Uploaded by

Copyright:

Available Formats

مقاله

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

مقاله

Uploaded by

Copyright:

Available Formats

Language Learning ISSN 0023-8333

Error-Correction Mechanisms in Language

CRediT author statement – Adnane Ez-zizi: conceptualization; methodology; investigation; data

Language Learning 74:1, March 2024, pp. 41–77 41

Keywords language learning; error-correction learning; Rescorla–Wagner model;

Language Learning 74:1, March 2024, pp. 41–77 42

43 Language Learning 74:1, March 2024, pp. 41–77

The Rescorla–Wagner Model and Language Learning

Language Learning 74:1, March 2024, pp. 41–77 44

of examples encountered by the participant). Incorporating the variability aris-

The Present Study

45 Language Learning 74:1, March 2024, pp. 41–77

investigate whether such characteristics could affect the adoption of a R–W-

Materials and Procedure

Language Learning 74:1, March 2024, pp. 41–77 46

task (main task), (b) an explicit knowledge and demographic questionnaire,

Language Learning Task

47 Language Learning 74:1, March 2024, pp. 41–77

Figure 1 An example of stimuli presented in one trial to depict a scene.

as suggested by Kiełkiewicz-Janowiak and Pawelczyk (2014). We made such

Language Learning 74:1, March 2024, pp. 41–77 48

Character used Polish word for Linguistic category Cue Cue

Duck Kaczka Feminine animate FA1 uFA

49 Language Learning 74:1, March 2024, pp. 41–77

Table 2 The learning events used for training

Block Learning events

Block 1 MP1 (uMP) + FA1 (uFA) → mp

Language Learning 74:1, March 2024, pp. 41–77 50

51 Language Learning 74:1, March 2024, pp. 41–77

response times, we used the Box–Cox method as implemented in the car

Explicit Knowledge and Demographic Questionnaire

Working Memory Task

Language Learning 74:1, March 2024, pp. 41–77 52

53 Language Learning 74:1, March 2024, pp. 41–77

Predicting Choices From the Model

Model Fitting Procedure

Language Learning 74:1, March 2024, pp. 41–77 54

whenever we refer to learning rate, we will always refer to the α parameter).

55 Language Learning 74:1, March 2024, pp. 41–77

association weights of the fitted models, which summarize the linguistic

Learned Noun–Verb Form Association Weights

Language Learning 74:1, March 2024, pp. 41–77 56

Figure 2 Histograms of cue–outcome association weights from the best-fitted

As predicted by (standard) blocking, the association weights between the

57 Language Learning 74:1, March 2024, pp. 41–77

Participant–Model Match Rates

Comparison Between the Rescorla–Wagner Model and Other Decision

Language Learning 74:1, March 2024, pp. 41–77 58

59 Language Learning 74:1, March 2024, pp. 41–77

Nonmasculine plural choice Response times (transformed)

Predictors OR 95% CI p b 95% CI p

(Intercept) 0.75 [0.60, 0.95] .016 0.08 [−0.05, 0.21] .244

Relationship Between the Model’s Activation-Based Measures and

Language Learning 74:1, March 2024, pp. 41–77 60

61 Language Learning 74:1, March 2024, pp. 41–77

Level of Agreement Between Participants Through the Lens of the Model

Language Learning 74:1, March 2024, pp. 41–77 62

Figure 5 Observed and predicted proportion of participants choosing the nonmascu-

Relationship Between Model-Fit Quality and Individual Difference Measures

63 Language Learning 74:1, March 2024, pp. 41–77