Learning Words in Two Languages: Manipulating Exemplar Variability for Within- and Cross-Language Generalization

Abstract

Purpose:

This article aims to describe how exemplar variability can manipulate the word learning environment to maximize within- and cross-language generalization in Spanish–English bilinguals. Furthermore, we examined sources of individual variability that predicted word learning.

Method:

Nineteen Spanish–English bilingual children participated in a word learning task presenting words in both languages. Children learned words either in a high variability condition (in which multiple exemplars are introduced with the target word) or in a no variability condition (in which the same referent is used with the target word). Word learning was tracked over the course of the training, and retention was examined once the training was discontinued. Children's generalization of referents within and across languages was also examined.

Results:

The exemplar variability effect was observed in within-language generalization trials, whereas cross-language generalization was less robust. Nevertheless, cross-language associations emerged in examining the role of language proficiency, such that semantic skills in English predicted word retention across languages. Similarly, children's propensity to code-switch during language production was positively correlated with retention of words learned in the high variability condition.

Conclusions:

The findings show that Spanish–English bilingual children may make use of exemplar variability to support word learning in different ways compared with monolinguals. The exemplar variability effect interacts with children's acquired language skills and word learning abilities at the start of the intervention. This study provides preliminary evidence from which future research can develop word learning interventions that are responsive to the needs of multilinguals.

Supplemental Material:

https://doi.org/10.23641/asha.19241856

Vocabulary growth and word learning are a common area of service delivery for many speech-language pathologists. Indeed, a wide range of etiologies lead to delays in vocabulary development; in fact, the majority of children served in early intervention are referred due to communication concerns (Morgan et al., 2016). Increasing vocabulary size is a common intervention goal because it is well established that early vocabulary abilities and other oral language skills set the foundation for later literacy (Duff et al., 2015; Lee, 2011; Muter et al., 2004). Literacy skills, in turn, predict academic achievement and other significant long-term outcomes (e.g., Johnson et al., 2010).

Word learning is a complex and dynamic process, and a variety of intervention approaches exist given that challenges in word learning are well documented in children with early language delays and disorders (McGregor et al., 2013). Yet, there is a dearth of research examining treatment approaches for dual language learners (DLLs) specifically. Approximately one third of children in the public school system in the United States are DLLs (e.g., they have at least one parent that speaks a language other than English), and the majority (59%) of DLLs are hearing Spanish at home (Park et al., 2017). Dual language learners are typically defined as children learning two languages at the same time. Although many extant approaches created within Spanish and English monolingual populations may extend to Spanish–English DLLs, it is important that these approaches be adapted to maximize their efficacy (Cycyk et al., 2021).

Emerging research suggests that the dual language experience affords a unique word-learning context that has both similarities and differences compared with a single language context (Fennell & Lew-Williams, 2018; Kalashnikova et al., 2018). Of interest to the present discussion is DLLs' ability to generalize categories across two languages. Bilinguals can take advantage of two languages in the context of word learning: Findings show that, for bilinguals, learning a word in one language makes it easier to learn it in the second (Bilson et al., 2015; De Anda & Friend, 2020; Goodrich et al., 2016). The ability of bilinguals to use such cross-language generalization to support vocabulary growth has been documented in multiple studies and across a wide age span. This relative strength in facilitating word learning may be a skill to be leveraged in interventions targeting word learning in dual language contexts for children with early language delays and disorders. Given the importance of developing word learning strategies that support intervention in this population, this study aims to extend and adapt a monolingual word-learning intervention approach to the unique cross-language context of DLLs. The overall aim is to provide an example of an adaptation of a monolingual word learning intervention to DLLs while also yielding preliminary data on its efficacy in this population.

Exemplar Variability

Learning a word requires more than mapping a word to a referent. True learning is when this mapping allows the learner to generalize to a semantically related word or referent in decontextualized contexts (Borovsky et al., 2016; Goldenberg & Sandhofer, 2013). This learning supports the formation of categories and rich lexical–semantic networks, a process that begins at least as early as the first year (e.g., Friedrich & Friederici, 2005). One way to maximize this word learning during intervention is to increase the distributional cues of exemplars by introducing children to a variety of referents when learning a new word. That is, to teach the category of a chair, it is best to see variable examples of different sizes, colors, and textures than it is to see any single one of those examples. Indeed, this type of exemplar variability in the input has been shown to be an effective intervention for supporting word learning and generalization in English monolinguals (Aguilar et al., 2018). Aguilar et al. taught a group of preschool children (n = 18) with specific language impairment eight new words in two conditions across three intervention sessions. In one condition, children learned words with three different exemplars (i.e., a high variability context), whereas the second condition presented words with only one exemplar (i.e., a no variability context). In order to measure children's learning over the course of the training task, a vocabulary test presented children with objects that they had previously seen during training as well as new within-category objects they had not seen previously to test generalization. Those children in the high-variability condition showed the strongest retention of words 3 weeks after the last training session, whereas there were no exemplar variability effects over the course of the training sessions. Taken together, these findings suggest that high exemplar variability specifically supports retention of words within English monolingual preschoolers. This finding has recently been replicated in slightly older French monolinguals with developmental language disorder (Krzemien et al., 2021) and is consistent with a body of work showing that variability in linguistic input facilitates learning (Gomez, 2002; Plante et al., 2014; Richtsmeier et al., 2009; von Koss Torkildsen et al., 2013).

Predictions for the Bilingual Context

It is well established that bilinguals are not two monolinguals in one, but instead form a highly interactive language system that exhibits competing activation within and across languages (Grosjean, 1989; Marian & Spivey, 2003). A growing body of research shows that this extends to word learning specifically, in that bilinguals are more likely to learn a word in a language if they already know it in the other (e.g., Bilson et al., 2015; De Anda & Friend, 2020; Goodrich et al., 2016; Sheng et al., 2016). That is, bilinguals seem to leverage translation equivalents in word learning. Such findings provide support for shared lexical–semantic processes across languages, consistent with prominent theoretical models of bilingual language representation (e.g., Grainger et al., 2010; Kroll & Stewart, 1994).

How does the exemplar variability effect extend to bilinguals, who also experience variability in word forms in their exposure to translation equivalents across their two languages? Translation equivalents are words across languages with the same referent, like chair and silla. The original findings by Aguilar et al. (2018) showed that monolingual children retained words best when they learned them with different exemplars. We expect that these findings will replicate within each of bilinguals' two languages given that such an effect is generally thought to be a reflection of basic cognitive learning processes (i.e., statistical learning; e.g., Erickson & Thiessen, 2015). That is, learning a word in Spanish with multiple exemplars will help support the extension of the category to the English translation equivalent, and vice versa.

Although we expect to observe exemplar variability effects in a DLL context, this study asks whether exemplar variability further facilitates this cross-language transfer, a unique phenomenon of multilingual development. Recall that exemplar variability helped monolingual children retain words even when they generalized them to new within-category referents they had never seen before (Aguilar et al., 2018). In the context of cross-language generalization, it may, therefore, also be possible that high exemplar variability will further facilitate children's propensity to generalize across languages to learn translation equivalents. In particular, the mechanism by which exemplar variability theoretically supports word learning is by strengthening the underlying semantic representation of a word and its category. In the context of translation equivalents, a stronger underlying representation may influence the speed and strength with which such a category is extended across languages. If the semantic category for a word is already strong in one language, language transfer may occur when bilinguals encounter a referent in the second language as they are more likely to identify it as part of the same word category. That is, we expect that exemplar variability may also help bilingual children extend words learned in one language context to a second language within translation equivalents specifically.

Language Proficiency

In addition to describing the overall effects of exemplar variability in dual language contexts, this study aims to understand individual differences in DLL children's word learning abilities based on language proficiency within and across languages. Prior research shows that children's acquired language skill (i.e., their language proficiency) predicts their ability to learn from distributional cues. Those monolingual children with small vocabularies are more likely to learn phonotactic patterns that violate the rules of their native language unlike their peers with larger vocabularies (Graf Estes et al., 2011, 2016). Similarly, vocabulary size predicts children's ability to extend new words (Thom & Sandhofer, 2009). This pattern of results suggests that the degree to which children make use of variability in word learning is related to their past experience with language. In a recent study, monolingual children with developmental language disorder (DLD) showed a correlation between their receptive vocabulary size and their ability to extend a novel noun category (Krzemien et al., 2021).

As with the literature on exemplar variability, findings regarding language proficiency in monolingual populations only partially inform our understanding of bilinguals. In general, it is reasonable to hypothesize that those bilingual children with the strongest language proficiency will also demonstrate the strongest word learning outcomes. Yet, bilingual children have lexical–semantic skills within and across their languages, making it important to understand how proficiency in one language can influence skills in the other. Furthermore, the relative proficiencies across languages can lead to dominance patterns, as language proficiency can be stronger in one language compared with another within bilinguals, and this can vary across language domain (i.e., semantics, grammar, phonology, etc.; Bedore et al., 2012) and shift over development (Oppenheim et al., 2020). Language dominance effects in lexical–semantic development have been documented in early bilingual first language acquisition within young Spanish–English bilinguals as early as the second year of life (Conboy & Mills, 2006). For example, within Spanish–English bilingual toddlers, strong vocabulary skills in the dominant language predicted speed of lexical retrieval in the nondominant language (De Anda et al., 2018). This study asks whether proficiency in Spanish can support further word learning in Spanish, but also whether proficiency in Spanish supports word learning in English as well, and vice versa. Understanding the ways in which bilingual children leverage both languages is crucial for developing word learning interventions that maximize efficacy and efficiency.

In addition to describing the relative importance of Spanish and English language proficiency, this study investigates the role of overall language proficiency by examining lexical–semantic skill across languages in predicting word learning. Similarly, we examine the role of code-switching in predicting the degree to which children can generalize word knowledge within and across their two languages. This growing body of work suggests that although code-switching can lead to processing costs during language production in adults (Bobb & Wodniecka, 2013; Fricke et al., 2016; Meuter & Allport, 1999), code-switching may, nevertheless, be preferred by bilingual children to help fill lexical gaps (Kuzyk et al., 2020; Smolak et al., 2020). Code-switching in bilingual children has also been linked to components of executive function (Crivello et al., 2016; Kuzyk et al., 2020). It has been speculated that code-switching in children's production reflects children's linguistic competence; in one study of 5- to 6-year-old Mandarin English learners observed in child care centers, children's number of code-switched utterances were positively associated with language proficiency (Yow et al., 2018).

Prior research on code-switching is consistent with work on translanguaging that has shown that bilingual children leveraging two language systems in educational contexts have stronger learning outcomes (e.g., Creese & Blackledge, 2010; Hornberger & Link, 2012). It has been posited that the pervasiveness of code-switching and the fact that it is associated with language proficiency in young children suggests that code-switching reflects children's emerging linguistic competence (Yow et al., 2018). Given this emerging area of research, we sought to explore the associations between code-switching and children's word learning skills. This study examines the degree to which children could generalize word categories within and across their languages and whether this is facilitated by exemplar variability. Thus, we hypothesized that children with high degrees of code-switching within an utterance (i.e., intrasentential code-switching) would demonstrate stronger word learning outcomes than peers with less propensity for code-switching.

This Study

As we have reviewed, word learning is an important skill in the development of early oral language abilities that can be further facilitated by manipulating exemplar variability. This article aims to describe the extent to which such manipulation of the word learning environment through exemplar variability facilitates within- and cross-language generalization of translation equivalents in Spanish–English bilinguals. This study provides preliminary evidence from which future research can develop word learning interventions that are responsive to the needs of multilinguals. Specifically, the first research question of interest asks: (1) Does high exemplar variability support word learning and generalization within and across languages in Spanish–English bilingual children? As described previously, we expected that exemplar variability would significantly predict word learning outcomes within Spanish and English and support cross-linguistic generalization.

The second set of research questions examines the association between children's language proficiency and the retention and generalization of new words. In particular, we examined language proficiency within and across languages. Specifically, we ask: (2a) How does semantic skill in Spanish versus English predict word learning outcomes within and across languages? Given the well-established findings in the bilingual literature showing cross-language interaction, we expected that semantic skills in one language would predict learning within and across languages, with the dominant language explaining the most unique variance. Furthermore, to examine skills across languages, we investigate (2b) whether lexical–semantic skill across languages predict retention of learned words, and whether grammatical skills also contribute unique variance to word learning above overall lexical–semantic proficiency. Lastly, we evaluated (2c) whether children's propensity to code-switch during language production predicted their ability to generalize words within and across languages. To the extent that code-switching reflects underlying language proficiency, we expected that those children that alternated between languages within a single utterance would also be more likely to generalize words across languages. Together, these questions examine language proficiency with language-specific measures (Question 2a), overall language measures (Question 2b), and measures that capture language mixing (such as code-switching in Question 2c).

Method

Participants

A group of 19 children (six male, 13 female) participated in this experiment (see Table 1 for summary of demographic variables). The participants ranged in age from 4;0 (years;months) to 6;9 (M _years = 5;5, SD _years = .78). All of the children were reported by their parents to be exposed to both English and Spanish. All but two children had greater exposure to Spanish than English as measured on the Language Exposure Assessment Tool (De Anda et al., 2016; M _Spanish = 75.06%, SD = 19.38; M _English = 25%, SD = 19.55). Three caregivers reported that they had or have had concerns about their child's language and communication at some point, but no parent reported a language delay or disorder diagnosis. The majority of mothers had earned at or below a high school, General Educational Development, or equivalent level (67%) with a smaller group also reporting vocational training or college attendance (28%). Participants were recruited through flyer postings at local Latinx-serving grocery stores and organizations, as well as through attendance at Latinx events in the local community.

Table 1.

Demographic characteristics of participants.

Variable	M (SD)
Cumulative language exposure
%Spanish	75.06% (19.38%)
%English	24.94% (19.55%)
	n (%)
Dominant language of exposure
Spanish	17 (89)
English	2 (11)
Language concern history
Yes	3 (16)
No	16 (84)
Maternal ethnicity
Mexican	15 (79)
Guatemalan	3 (16)
Prefer not to answer	1 (5)
Maternal education
Some school	7 (37)
High school, general educational development, or equivalent	5 (26)
Technical or vocational training	1 (5)
Some college	1 (5)
Bachelor's degree	3 (16)
Prefer not to answer	2 (11)

Word Learning Training

The word learning task employed in this study was a modification and adaptation of the published procedure in English monolinguals (Aguilar et al., 2018; see Supplemental Material S1 for the task script). Specifically, the training took place in the context of a gender-neutral treasure hunt activity to maximize child engagement. Children were taught eight words in Spanish and eight words in English, for a total of 16 words (see Table 2). The order of presentation of the target words was counterbalanced. Each target word was presented to the child 6 times per visit, for a total of 18 presentations of each word over the course of the study. Three corresponding object exemplars were provided for each word. In each language, half of the words have object exemplars that are identical (the no variability condition), and for the other half, the object exemplars are different (the high variability condition).

Table 2.

List of trained words across language and exemplar variability conditions.

High exemplar variability		No exemplar variability
English	Spanish	English	Spanish
Clamp	Gafa	Bobbin	Canilla
Curler	Rulo	Bolt	Tornillo
Fabric	Tela	Hinge	Bisagra
Tassel	Borla	Scourer	Estropajo

Prior to the child's arrival, the research assistant set up the testing room such that one of the three object exemplars (e.g., “the pirate treasure”) was hidden from the child's view under colorful foam squares. After building rapport with the child, the bilingual research assistant escorted both the child and caregiver to the testing room. The child was provided with verbal instructions before the “treasure hunt” began. The word learning training task was done first in the child's dominant language of exposure and then in their nondominant language. The research assistant reorganized the testing materials between language conditions, and verbally cued the transition between languages (e.g., “Now we are going to do the same activity in English”).

Recall that at the beginning of the task, one of the three object exemplars are hidden around the test room, whereas the remaining two exemplars are in the research assistant's possession and out of view from the child. The task begins when the research assistant shows the two object exemplars in their possession for the first target word (as mentioned previously, objects may be the same or different exemplar depending on the variability condition). The target word is then subsequently presented 6 times across a series of prompts fixed in the following order: (a) directions, (b) statement, (c) imperative command, (d) question, (e) prompt for production, and (f) clean-up elicitation. Specifically, the target word is first presented within (a) a set of directions (e.g., “We are going to use the ___for the treasure hunt.”) as the research assistant shows the child the two exemplars. Next the assistant presents the target word within (b) a statement (e.g., “This is a ___,” or “Here is the ___,” or “This is my favorite___.”). Having introduced the two objects, the researcher then directs the child to find the third object exemplar hidden around the room using (c) an imperative command (e.g., “Find the ___,” or “Get the___.”). Once the third object has been retrieved by the child conducting the “treasure hunt,” the research assistant then repeats the target word within (d) a question, (e.g., “Which ___do you want?” or “Which ___ should we use?” or “Which ___ is your favorite?”). For the fifth presentation, the research assistant produces the target word when providing (e) feedback to the child after prompting them to say the target word (e.g., “What is this called?” Feedback: “Yes/No, this is a ___.”). The sixth and final presentation of the word was (f) an elicitation to help clean up the three objects before moving on to the next target word (e.g., “Put the ___ in the treasure chest.”). Both the research assistant and child were able to interact with the three corresponding exemplars until the presentation of the next target word and corresponding object exemplars.

Target Words and Their Referents

As mentioned, children were taught eight words in Spanish and eight words in English, for a total of 16 low-frequency words taught at each of the training visits. At the start of the study, the words were presented to parents who all confirmed that all of the words were previously unknown to their children. The eight target words in Spanish and eight in English were noncognate translation equivalents to ensure minimal phonological overlap. The words were the same words used in Aguilar et al. (2018) with Spanish translations also added. All words refer to physical objects (i.e., nouns). The participants were exposed to the target words in two language conditions (i.e., English and Spanish) and two exemplar variability conditions (i.e., high exemplar variability or no exemplar variability). In the high variability condition, children saw three unique objects for each target word. However, in the no variability condition, children saw three identical objects for each target word.

For the high variability condition, the three unique referents varied by texture, color, and/or size (see Figure 1 for an example) consistent with the Aguilar et al. (2018) stimuli. Children were exposed to three different object exemplars when trained in the high variability condition. The object exemplars presented in the high variability condition for English were unique objects, different from the object exemplars presented for Spanish. That is, as shown in Figure 1, although the trained words in Spanish and English were translation equivalents and part of the same object category, different exemplars were presented in each language. This partially reflects the reality of the bilingual experience as variable referents are encountered in different language contexts due to cultural variability (e.g., the prototypical exemplar of “bread” is different in English vs. Spanish speaking contexts). Using different exemplars for each word also represents a conservative estimation of children's cross-language abilities since we are interested in measuring how children extend and generalize categories when referents are variable (rather than identical) across languages. Words in the no-variability condition were represented by three identical object exemplars. As with the high variability condition, the identical object exemplars differed for English and Spanish. A chi-square goodness of fit test showed that the number of syllables for each of the 16 training words did not differ significantly (p > .05) as a function of language (English vs. Spanish) and condition (high variability vs. no variability).

Figure 1.

Example stimuli for word learning training task across exemplar variability conditions and language.

Word Learning Test

Children completed a test to examine comprehension and generalization of word learning from the training. The test is relatively conservative in capturing word learning outcomes in that it assesses children's comprehension of the trained words and their generalization in a decontextualized context. The word learning test was presented to the child as a slideshow on a touch screen tablet, which was controlled by the research assistant. Each target word item appeared across three different trial types: training, within-generalization, and cross-generalization (see Figure 2 for an example). The target object appeared among a field of four images. The foil images represented different versions of objects that the children had been exposed to during the world learning training task. These objects were selected for their contrasting characteristics to the objects used in the training task, differing significantly in color, texture, and size. At no point did the research assistant provide feedback on the child's accuracy.

Figure 2.

Example stimuli for word learning test task across exemplar variability conditions and language.

On training test trials, the field of objects included exact object exemplars used in training. These trials examined whether children learned the word-to-referent associations established during the training task. The within-language generalization test trials presented objects the child had never seen before but were still part of the object category. This trial type measured the child's ability to generalize their knowledge of a previously learned word to a new category referent. The cross-language generalization test trials consisted of pictures of objects that had been presented in the nontarget language. This measured the child's ability to generalize a semantic category and the corresponding referents to a second language (essentially the translation equivalent). As with the training task, all trials were blocked by language, such that testing began in the child's dominant language of exposure, followed by testing comprehension of words in the nondominant language.

Language Proficiency Measures

Bilingual English Spanish Assessment (BESA)

The English and Spanish semantic subtests of the BESA (Peña et al., 2018) were used to capture semantic proficiency in each language. These subtests were administered according to the standardized procedure in the BESA manual. The subtests can be used independently to assess children's language skills. The semantic subtest utilizes receptive and expressive items to examine children's lexical knowledge. The BESA was normed on Spanish–English bilingual children in the United States. Children provided semantic associations and definitions through six tasks: analogies, characteristic properties, categorization, functions, linguistic concepts, and similarities and differences. The six tasks capture breadth and depth of the semantic system based on extant literature describing bilingual children's organization and access of lexical–semantic information. In terms of reliability, the English and Spanish semantic subtests demonstrate strong internal consistency (coefficient alpha = .86 in English and .88 in Spanish), temporal stability (test–retest correlation = .65 in English and .71 in Spanish), and validity (e.g., logical validity, empirical validity, and construct validity). The semantic subtest scores in English versus Spanish were used to examine the role of semantic skill in each language predicting word learning and generalization within and across languages.

Spontaneous Language Sample

Children's expressive language proficiency and use of Spanish and English was captured during a 10-min interaction with their caregivers. Dyads had access to (a) toy cookware and plastic food items, (b) farm animals and farmhouse or building blocks, and (c) the same book in Spanish and English. Similar to the Three Bags Task commonly used to study early parent–child interactions (e.g., Tamis-LeMonda et al., 2004), caregivers were encouraged to use all three sets of toys in ways that felt natural to them using their language(s) of preference.

Children's language productions and vocalizations during the interaction were transcribed and coded by trained bilingual Spanish–English research assistants using the conventions from the Systematic Analysis of Language Transcripts (SALT; Miller & Iglesias, 2012). In order to ensure accuracy of transcription and coding, a second transcript was reviewed alongside the video of the recording by a second research assistant. A third review was completed by SALT staff to ensure coding accuracy.

The following key variables were extracted from transcripts using SALT software to describe children's language production: number of different words (NDW), number of total words (NTW), mean length of utterance in words (MLUw), and subordination index (SI). MLUw was preferred over MLU in morphemes (MLUm) because it is widely accepted in languages other than English given variation in inflectional morphology across language (Gutiérrez-Clellen et al., 2000). Given that caregivers and children were not given specific language instructions, the extracted variables represented children's overall language skill across Spanish and English in a relatively naturalistic context.

In addition, each complete and intelligible utterance was coded to indicate whether the utterance was produced entirely in Spanish, entirely in English, or whether a code-switch was produced within the utterance (i.e., intrasentential code-switching). Using these codes, two key variables were calculated of relevance to the present research questions. First, to estimate children's propensity for within-utterance code-switching, we divided the number of code-switched utterances by the sum of code-switched and single-language utterances produced (i.e., total complete and intelligible utterances). Furthermore, for the purposes of calculating reliability, a random set of 20% of transcripts were reviewed in full by a second group of independent research assistants who were not involved in their original transcription (interrater reliability = 87%). The reliability transcripts were also reviewed by SALT for coding accuracy.

Procedure

Each participant visited the lab a total of 4 times over approximately a 4-week period (see Figure 3). During the first visit, children's baseline semantic skills were assessed using a parent–child language sample and the BESA followed by the presentation of the word learning training task. At the second and third visits, children once again completed the word learning training followed by the test task. All three visits were completed within a 2-week period. The final test session was conducted at the fourth visit, approximately 2 weeks after the final training session, to examine maintenance of the trained words.

Figure 3.

Timeline of data collection procedures.

Analysis Plan

To answer the first research question regarding the exemplar variability effect, a generalized linear mixed-effect model with accuracy (0, 1) as the binomial dependent variable (link: logit) was used with variability condition (high vs. low) as a fixed effect (Brown, 2021). Performance on the within- and cross-language generalization was examined specifically as it is considered a stronger measure of children's learning than performance on the training trials that do not require generalization to new words or exemplars as they present the same objects and labels used in the training. The central hypothesis was that children's performance on the word learning test task would differ based on exemplar variability (high vs. low) as well as generalization type (within-language vs. cross-language generalization), time (Visits 1, 2, 3, and 4), and language (Spanish vs. English). Thus, because we expected that the exemplar variability effect would vary as a function of visit (Tests 1, 2, 3, 4), language (English vs. Spanish), and trial type (within vs. cross language), these too were entered as fixed effects along with random effects of participants and word. Likelihood ratio tests were used to indicate whether such a model would provide a significantly better fit for the data than a base model including only the random effect structure. Summary model output in R was used to compare each level in the fixed effect factor against the reference category (visit: Test 1; variability condition: high; language: English; trial type: cross-language trials). In addition, a Type II Wald test of linear hypotheses was used to identify main effects and interactions using the chi-square test statistic for comparison with maximum likelihood ratio tests.

To answer the second research question regarding the association between language proficiency and the retention and generalization of new words, analyses included children's performance at Test 4 only in order to examine individual differences in retention of words following training. A generalized mixed-effect model with a binomial predictor (accuracy) was once more utilized. Where comparisons were appropriate, each predictor of interest was entered into the model and Akaike information criterion was compared with determine whether model fit improved. All analyses were completed using RStudio Team (2020) using the lme4 (Bates et al., 2015) and lmertest packages (Kuznetsova et al., 2017).

Results

Table 3 provides a summary of descriptive statistics across factors. See Supplemental Material S2 for a graphical representation of the data in Table 3. Note that the table includes performance on all three trial types. Recall that training trials presented children with the same referents/objects they had seen during the word learning training, whereas the generalization trials examined children's new word-referent mappings they had not been exposed to previously and, therefore, provide a stronger test of children's learning.

Table 3.

Descriptives for the test task across trial types, languages, and word learning conditions (high vs. no variability).

Trial type	Spanish		English
	High variability	No variability	High variability	No variability
	M (SD)
Trained objects
Test 1	1.53 (.84)	2.53 (1.22)	1.26 (.87)	1.95 (1.43)
Test 2	2.78 (1.06)	2.37 (1.12)	2.11 (1.41)	2.00 (1.15)
Test 3	2.42 (.84)	2.79 (1.27)	2.47 (1.17)	2.47 (1.17)
Retention	3.00 (.84)	2.44 (1.20)	2.67 (1.24)	2.56 (1.10)
Within-language generalization
Test 1	2.26 (1.15)	1.42 (1.12)	2.16 (1.54)	2.26 (.99)
Test 2	2.22 (.88)	1.89 (.96)	2.58 (1.22)	2.26 (1.15)
Test 3	2.58 (.90)	2.16 (1.21)	2.26 (1.28)	2.53 (1.17)
Retention	2.44 (.92)	2.22 (1.11)	2.06 (1.47)	2.72 (1.02)
Cross-language generalization
Test 1	2.11 (.66)	1.58 (.90)	2.47 (1.39)	1.74 (.81)
Test 2	2.56 (1.04)	2.44 (1.58)	2.11 (1.37)	2.05 (1.13)
Test 3	2.47 (1.07)	2.58 (.77)	2.53 (1.12)	1.84 (.90)
Retention	2.61 (1.09)	1.94 (1.16)	2.22 (1.40)	2.44 (1.25)

Exemplar Variability and Generalization

The first research question of interest concerned the role of exemplar variability in supporting generalization of trained words within and across languages for Spanish–English learners. A binomial generalized mixed-effect model was conducted with accuracy (0, 1) on the test task as the dependent variable. Time (Tests 1, 2, 3, 4), language (English vs. Spanish), variability condition (high vs. low), and trial type (cross- vs. within-language) were entered as fixed effects. Random effects of participants and training word were also included. A likelihood ratio test indicated that the model including the fixed factors provided a significantly better fit for the data than a base model including only random effects, X ²(31) = 59.55, p = .002. See Supplemental Material S3 for estimates and confidence intervals.

Examination of the summary output for the full binomial generalized mixed-effect model indicated a significant interaction between condition, trial type, and time. Specifically, an interaction between low variability condition, within-language generalization trial type, and the Test 3 timepoint (z = 2.109, p = .035) was observed. This suggests that accurate word identification is significantly increased by 1.534 log odds (odds ratio: 4.637) at the third test visit for the low variability condition in within-language generalization trials compared with the reference category at the first timepoint (Test 1 cross language trials in the high variability condition). Confidence in the effect appears weak as the 95% interval is relatively wide (95% CI [0.11–2.96]). No other effects were significant.

In addition, examination of correlations among random effects showed a negative by-participant correlation between the slope and intercept (r = −.30). The correlations showed that participants with high accuracy word identification scores at Test 1 showed slopes that were smaller (closer to zero) between the high and low variability condition compared with their peers with lower accuracy scores. This suggests that children with higher accuracy in the word identification task at the beginning of the study were less affected by the variability condition manipulation compared with their peers with lower word identification accuracy at Test 1. Similarly, the negative association between slopes and intercepts for the language condition (r = −.72) showed that children with high word accuracy scores had more comparable scores across English and Spanish compared with children with lower word accuracy scores.

Lastly, a Type II Wald test of linear hypotheses was used to identify main effects and interactions using the chi-square test statistic to compare findings against the likelihood ratio test results. Results for the fixed effects in the full model showed a significant effect of visit, X ²(3) = 19.295, p = .0002; a Visit × Language interaction, X ²(3) = 8.006, p = .046; and a significant Language × Trial Type interaction, X ²(1) = 9.224, p = .0024.

Lexical and Semantic Proficiency

The second research question of interest examined individual differences in children's retention of trained words as a function of lexical–semantic proficiency. As noted previously, these analyses included children's performance at Test 4 only in order to examine individual differences in retention of words following training. The first analysis under this research question examined whether children's within-language semantic skills in Spanish versus English as measured by the BESA predicted retention. The subsequent language sample analyses (including NDW, NTW, MLUw, and SI) examined children's overall language skill (including code-switching) in a naturalistic context across Spanish and English in predicting learning outcomes. Table 4 provides descriptives across measures of language proficiency.

Table 4.

Descriptives for key language and lexical–semantic measures.

Measure	M (SD)	Range
BESA semantics subtest
Spanish standard score	99.76 (11.71)	80–123
English standard score	93.13 (10.15)	75–115
Spontaneous language production
NDW	82.47 (31.06)	20–129
NTW	187.16 (93.50)	28–359
MLUw	2.41 (.70)	1.22–3.95
SI	0.96 (.11)	0.6–1.09
% of total utterances with code-switching	2.16 (3.18)	0–12

Note. BESA = Bilingual English and Spanish Assessment; NDW = number of different words; NTW = number of total words; MLUw = mean length of utterance in words; SI = subordination index.

To begin, we examined whether Spanish BESA scores would predict performance and, if so, whether English BESA scores would predict additional variance. In this way, we examined the unique contribution of the nondominant language in the sample of Spanish-dominant children. A base model was first constructed that included accuracy (0, 1) as the binomial dependent variable and language (English vs. Spanish), condition (high vs. low variability), and trial type (cross- vs. within-language) along with random effects of participant and word. For the purposes of model comparison, a second model was constructed including Spanish semantic scores on the BESA as a continuous predictor. A likelihood ratio test revealed that the second model including the Spanish BESA score did not explain significant additional variance compared with the base model. A third model included English semantic subtest scores in place of the Spanish scores. This time, the inclusion of the English scores did improve the model fit compared with the base model, suggesting that the English scores explained additional variance and provided a significantly better fit for the data than a base model, X ²(1) = 4.322, p = .038. Examination of model output for the model showed a significant main effect of English scores (z = 2.23, p = .026) and no interactions.

Next, we evaluated the role of language proficiency as measured during the observed parent–child language sample. Whereas the BESA analyses compared English and Spanish to examine the role of within-language skills in promoting word learning, the language sample analyses examine overall language skill. Thus, we ask how semantic and morphosyntax measures across Spanish and English predict word learning, and whether morphosyntactic skill predicts unique variance above semantic skill alone as observed during the language sample. We began by including lexical–semantic skills, which are the most proximal to the word learning outcomes we were attempting to predict. Using a series of nested models, we examined how each additional predictor improved model fit indices using likelihood ratio tests. The base model was the same as in the BESA analyses, such that it included accuracy (0, 1) as the binomial dependent variable and language (English vs. Spanish), condition (high vs. low variability), and trial type (cross- vs. within-language) along with random effects of participant and word. We first included NDW as a predictor that significantly improved model fit, X ²(1) = 10.92, p < .001. Including NTW (in addition to NDW) in a second step did not improve model fit. Similarly, including MLU and SI in a third and fourth step also did not improve model fit (all ps > .07).

Lastly, to examine the role of language alternation, we included code-switching as a continuous variable in the same base model used in the BESA and language sample analyses. The model predicted additional variance above the base model, X ²(2) = 10.20, p = .006. Examination of the model output showed a significant interaction between code-switching and variability condition (z = −2.796, p = .005). Follow-up correlations showed that the degree of children's within-utterance code-switching and word retention accuracy scores were significantly and positively associated, but only for words learned in the high variability condition (r = .37, p = .001). Words learned in the low variability condition did not show a significant correlation.

Discussion

There remain gaps in our understanding of effective approaches to expanding word learning in young DLLs. This study represents a step toward extending existing research on monolingual children to the multilingual context. In particular, this study examined a group of Spanish–English bilingual children learning words across their languages under conditions designed to maximize retention based on recent findings in monolingual learners. The research questions of interest examined the role of exemplar variability and language proficiency in promoting word learning and generalization within and across Spanish and English.

Exemplar Variability Effects in Within- Versus Cross-Language Generalization

The first aim of this study was to establish the effect of exemplar variability for words learned in Spanish and English. Results showed that exemplar variability effects were minimal and nearly absent. Specifically, we observed only an effect of low exemplar variability for within-language generalization, and this effect was found only for Test 3 when compared with word learning performance at Test 1 and with a relatively wide confidence interval. That is, children were more likely to extend a trained word to a new referent in that same language if it was trained with the same exemplars, but this effect was not robust and observed only by the third training session and did not continue into the retention timepoint (Test 4). At the group level, the effect of low variability generally tended to improve word learning (all else being equal) as shown in the model output for the full binomial generalized mixed-effect model. Yet a main effect of exemplar variability was not observed in a separate analysis. Overall, the effect of exemplar variability was not observed at the group level as expected in the current study compared with extant monolingual findings.

There are some preliminary hypotheses that can explain the results. Regarding cross-language generalization, children as a group seemed to have more difficulty extending a referent they had already learned in one language to the second language compared with their ability to extend a category to a new referent within the same language. Results also suggested a language effect such that word learning outcomes differed in Spanish versus English in the current group of Spanish-dominant bilinguals. This may be consistent with a growing body of literature documenting the nuanced word learning strategies used by bilinguals as a function of language dominance. For example, it has been posited that word learning occurs under dynamic conditions, such that the strategies that support fast referent selection may be relatively comparable in monolinguals and bilinguals (Kalashnikova et al., 2018; McMurray et al., 2012; Weatherhead et al., 2021). However, the strategies that scaffold slow associative learning to support retention may require additional time to develop in bilinguals given that they are building two language systems. The results here suggest that further refinement occurs in bilingual children's slow associative word learning skills.

The differences in bilingual word learning could reflect differences in task demands (e.g., Munakata, 1998), such that there may be less cognitive load in processing words in the dominant language compared with additionally extending them to a weaker language. Such differences in within- versus cross-language associations are well documented in the bilingual literature (e.g., Bitetti et al., 2020; Prevoo et al., 2016). Indeed, in this study, cross-language generalization required children to extend an object category while also extending the word form to the other language. In addition, generalization to the second language could have been attenuated due to the different exemplars presented across languages. Although this is common for bilingual language learners, it is possible that the specific degree overlap in referents across languages influences cross-language generalization. However, and as noted previously, confidence in the results of the interaction of the exemplar variability effect with word learning and generalization is limited in this study. As such, these conclusions require further research.

Alternatively, it is possible that the exemplar variability and dominance effects reflect underlying language proficiency differences that influence the cognitive load and task demands. This is consistent with several prior studies demonstrating positive associations between vocabulary breadth, depth, and word retention (Nicoladis & Laurent, 2020; Repnik et al., 2021). Indeed, the random effects correlations lend support to the idea that individual differences may explain the pattern of results in the present group of Spanish–English bilinguals. Results showed that those children with strong accuracy scores at the first test were less influenced by the variability condition compared with those children with lower scores. Similarly, strong accuracy scores at the first test were associated with smaller language dominance effects (i.e., difference between English and Spanish word learning scores). This suggests that there are individual differences evinced as early as the first test that predict children's degree of word learning as a function of exemplar variability and language over time. We return to this point below in discussing the language proficiency findings of this study.

Comparisons to Prior Monolingual Findings

Although results showed that individual differences may play a role in the exemplar variability effects, findings also showed that the overall exemplar variability effect was not significant at the group level. This means that there were less robust exemplar variability effects for the present group of Spanish–English bilinguals, in contrast to the published findings in English monolinguals. Recall that findings in English monolinguals showed an exemplar variability effect at the critical retention timepoint (e.g., Time 4; Aguilar et al., 2018). What could explain the constraints on the exemplar variability effect? The present findings alone do not rule out the presence of the exemplar variability effect for retention in bilinguals, and we should be cautious in interpreting comparisons between monolinguals and bilinguals. Indeed, the task demands for monolinguals were different from those presented to bilinguals in this study given the unique research questions of interest. For example, in this study, bilinguals were learning the same eight English words as the monolinguals and in comparable conditions, but bilinguals were also tasked with learning the Spanish translation equivalents (using different exemplars across languages). This means bilinguals in this study were learning twice the words that monolinguals learned therefore increasing the difficulty of the task and potentially leading to the lack of an exemplar variability effect for retention. Such an increase in task demands may place an added burden on working memory that interacts with bilingual children's language experience in promoting language learning (e.g., Kaushanskaya & Crespo, 2019), in addition to the cognitive demands in extending categories across languages as referenced previously.

It is also possible that the underlying word learning strategies differed across monolingual and bilingual children. As reviewed in the introduction, prominent models of word learning posit that heuristics arise out of language experience, which includes the experience of learning two or more languages (e.g., Byers-Heinlein & Werker, 2009). Indeed, unlike monolinguals, bilingual children have two languages that interact. In this study, children's semantics in English predicted word learning performance across languages, whereas Spanish semantic skill did not. Thus, although cross-language generalization was not as robust as within-language generalization, significant cross-language associations still emerged. This means that semantic skills in the less-dominant English language predicted retention of words in both English and Spanish. These results highlight two key findings. First, bilingual children leverage both languages, and this is consistent with a large body of work showing bilinguals demonstrate parallel processing of their two language systems (Goodrich & Lonigan, 2018; Marchman et al., 2020); bilinguals' two language systems are interactive, and both languages are accessed over the course of word processing. Therefore, it is also possible that in addition to increased task demands, bilingual children may make differential use of their two languages in promoting word learning compared with their single language peers who only have a one language as described previously. Second, the fact that English, but not Spanish, predicted word learning suggests that relative language dominance influences the pattern of cross-language associations. This is consistent with several studies that show dominance effects in word processing and learning as early as 22 months of age (Conboy & Mills, 2006). In one study, children's vocabulary size in the less dominant language predicted speed of lexical retrieval in the dominant language (De Anda et al., 2018). Thus, although the exemplar variability effect was not observed at the group level in this study, bilingual children nevertheless made use of their interactive language system to support further word learning.

Despite the lack of a robust exemplar variability effect, bilingual children in this study learned words at the same rate as monolinguals. Collapsing performance across Spanish and English yielded accuracy scores comparable to the published monolingual findings, such that by the time of retention testing children had learned five words on average. Consistent with prior work, these findings suggest that bilinguals may be deploying different word learning strategies while still meeting similar developmental milestones (e.g., Nicoladis & Genesee, 1997).

The Role of Language Proficiency

In terms of individual differences in children's word learning, the present findings underline the importance of promoting language, broadly construed, irrespective of whether it is Spanish or English. Specifically, children's word retention was predicted by their lexical–semantic productivity (i.e., NDW) but less well by their morphosyntactic skill. Similarly, children's propensity for intrasentential code-switching was positively associated with retention of words learned in the high variability condition. The overall model also showed random effect correlations suggesting individual differences in the exemplar variability effect such that the difference between high and low variability conditions was greatest for children with relatively weak word learning skills at the beginning of the intervention. Taken together, the individual differences in lexical–semantic, morphosyntactic, and language-alternation skills and their association with word retention corroborate leading theories of word learning that suggest that language proficiency influences word learning strategies and retention (Graf Estes et al., 2011, 2016; Thom & Sandhofer, 2009; Krzemien et al., 2021).

The language alternation findings also provided further support for theories that suggest that code-switching in children's language production reflects increasing linguistic competence (e.g., Yow et al., 2018). Indeed, code-switching seems to require additional cognitive and linguistic resources for both comprehension and production (e.g., Kaushanskaya & Crespo, 2019). Extending this to this study and the examination of exemplar variability, it may be that children's ability to make use of word-referent mappings in the high variability condition as in this study may reflect precocious cognitive and linguistic processing skills, which, in turn, is associated with children's propensity to code-switch. Thus, while exemplar variability does not seem to predict word learning retention at the group level, individual differences in children's ability to alternate languages during language production were positively associated with their use of exemplar variability to learn new words. Furthermore, while our initial hypothesis was that code-switching would support word learning across conditions, the findings imply that this hypothesis is only partially correct and in need of further refinement given that the effect was localized to the high variability word learning condition specifically. This suggests that children's ability to make use of multiple exemplars for retention is associated with their propensity to alternate languages during production.

Clinical Implications for Word Learning Interventions

Though the findings of this study cannot conclusively report an effect of exemplar variability, we know that high exemplar variability is as good if not better than no variability in DLLs depending on the learning context. Clinicians should, therefore, introduce exemplar variability whenever possible when supporting both monolingual and bilingual children. In addition, clinicians should keep in mind that supporting language skills across both languages supports strong cross-language outcomes, and that extant vocabulary size as well as morphosyntactic skills across languages each contribute to children's ability to learn words. Furthermore, code-switching in children's language production should not be mitigated or seen as a deficit; on the contrary, a growing body of evidence points to children's language alternation reflecting growing cognitive and linguistic skills. The results of this study also highlight that word learning may be more complex in DLLs compared with monolinguals. Whereas we experimentally manipulated variability in referents, children from bilingual backgrounds also experience natural variability in word forms (e.g., labels in Spanish vs. English) and can demonstrate distributed skills across languages that together support further processing even in single-language contexts. Thus, all word learning interventions must take this complexity in the language learning contexts into consideration when planning for treatment.

Limitations and Future Directions

This study represents an early step in providing recommendations for word learning interventions that specifically support DLLs. Future research should test varying intervention dosage conditions and testing intervals to first determine whether the exemplar variability effect can be maintained well into the retention period. It is possible that the exemplar variability effect would be observed at retention with higher treatment sessions and by testing retention at a longer interval. Additionally, future research should test DLLs with language delays and disorders specifically and examine word learning outcomes across a continuum. As a reminder, children were asked to select a decontextualized referent from a field of four, which is harder than having children identify referents from a smaller set or in a more naturalistic context for example. It may be that the exemplar variability effect is evinced in measures that are sensitive to emerging knowledge rather than more conservative measures of robust skill. Whereas we intentionally replicated the degree of variability used in prior research, the degree of exemplar overlap within and across languages should also be manipulated in future research. It is likely that the degree of within- and cross-language generalization is influenced by the similarity between the exemplars. Lastly, to further refine our understanding of the individual differences reported here, future research must examine which children benefit from exemplar variability and when such a strategy should be introduced in development. Thus, while the present results offer promise for developing interventions that specifically meet the needs of DLLs, there remains a need for researchers and clinicians to continue refining approaches for this specific population of children.

Funding Statement

Research reported in this publication was supported in part by the National Institute on Deafness and Other Communication Disorders under Award Number K23DC018033 to the first author.