The aim of the current study was to explore whether CLI occurred if the two languages had inverse properties. According to Hulk and Müller (2001) and Müller and Hulk (2001), CLI is likely to occur when two conditions are met: (i) there is surface structure overlap between the two languages; and (ii) the relevant property lies at the interface of two linguistic domains. Surface overlap refers to a situation where Language A has two grammatical structures (e.g., α and β), but Language B has only one of them (β); bilingual children may overproduce β in Language A compared to monolingual peers. This overproduction is driven by the reinforcing input for β from both languages. A concrete and relevant example comes from Westergaard and Anderssen (2015) in their investigation of possessive structures in Norwegian-English bilingual children: Norwegian, as thoroughly described is Section 2.1.2, has both prenominal and postnominal possessives, whereas English only has the prenominal (i.e., my car). The prenominal possessive is thus the overlapping structure for the acquirers of these two languages. In relation to (ii), a structure is considered to be at a particular interface when both options are grammatical in a given language, but their use is semantically or contextually conditioned, either signaling a meaning difference (syntax-semantics interface) or being context-dependent (syntax-pragmatics interface). Possessive variants are placed in the latter group: the role of context in guiding the choice between grammatical alternatives, as pre-and postnominal possessives, is contextually driven (see overview in Sections 2.1.1 and 2.1.2).

Since then, numerous studies have been conducted on CLI in simultaneous bilingual children reconfirming both (i) and (ii) (Paradis and Navarro, 2003; Serratrice et al., 2004; Sorace and Serratrice, 2009), but none so far on a double and inverse structural overlap. The ever-expanding body of research has shown how CLI occurs at the syntax-semantics interface: studies on copula omission in English-Spanish bilinguals (Fernández Fuertes and Liceras, 2010; Liceras et al., 2012) and generic vs. specific expression in plural NPs in English-Italian and Spanish-Italian bilinguals (Serratrice et al., 2009); within syntax, Anderssen and Bentzen (2013) argue that the structural similarity between English and Norwegian facilitates the acquisition of the Norwegian determiner system in a Norwegian-English bilingual child. Additionally, in a study on Greek-English bilinguals, both phenomena from narrow syntax and the syntax semantics interface were found to be equally affected by CLI (Argyri and Sorace, 2007). These results suggest the possibility that external factors such as language dominance may affect CLI (Bernardini, 2003). However, language dominance is a complex phenomemon, and there is no operationalized definition for it, which causes researchers to rely on either proficiency or the majority language status to determine dominance (Hamann et al., 2019). Several factors contribute to the outcome of language dominance, and Hamann et al. (2019) list the L1 maintenance, the direction of CLI, and the general effects of bilingualism on cognitive development as factors contributing to the effects of dominance on the competence and processing of bilinguals. Language dominance will be revisited later on.

Thus, the conditions initially deemed necessary for CLI to occur have steadily been expanding: structural overlap is predictive of CLI, but CLI is not limited to it (Serratrice et al., 2009). External factors must also be considered to predict the direction of CLI (Kupisch and Bernardini, 2007; Rodina et al., 2020; Serratrice et al., 2009). For example, Serratrice et al. (2009, p. 252) found that the place of residence of the participants was relevant: English-Italian bilinguals residing in Italy consistently outperformed the bilingual group that resided in the UK on the interpretation and use of plural noun phrases with definite articles in Italian.

Claim (ii) from Hulk and Müller (2001) and Müller and Hulk (2001) fits squarely with the interface hypothesis (IH) (Sorace, 2011; Sorace and Serratrice, 2009). The IH identifies both the syntax-pragmatics and the syntax-semantics interface as vulnerable domains for bilinguals; yet the two interfaces pose different challenges for speakers. For the former, the IH argues that there might be processing difficulties regardless of whether there is a structural overlap between the languages, whereas the overlap plays a more relevant role for the syntax-semantics interface. Sorace and Serratrice (2009) also specify how both interfaces are sensitive to reduced input, causing the linguistic setting in which the child lives to influence the amount of CLI. This feeds into the research on the status of the HL addressed in what follows.

When studying bilinguals, one of the languages will inevitably be a HL (unless research is conducted in a fully bilingual society). This societal circumstance should not be ignored when predicting potential outcomes. I take the definition of HL provided by Rothman (2009, p. 156):

HL is simply another instance of simultaneous bilingual acquisition (Montrul, 2016). In the field of HLs it is widely accepted that acquisition outcomes are determined by extra-linguistic factors (Rodina et al., 2020). Specifically for this study, the two languages differ in their status: Italian is the HL, whereas Norwegian is the majority language. The status of HL is inversely linked to dominance, as typically heritage speakers (HSs) are less dominant in the HL compared to the majority language (Polinsky and Kagan, 2007) once they enter the kindergarten/school system. Language dominance is thus the relative strength, proficiency, and active use of one language over another within a bilingual individual, shaped dynamically by input and social factors. At the same time, place of residence and majority language exposure exert an influence, as HL speakers fully embedded in society will have a high and diverse level of exposure to the majority language, with the result usually being that it will be the dominant language. Bilgory-Fazakas and Armon-Lotem (2025) claim that dominance is dynamic, influenced by family and community factors, and balanced exposure predicts better HL maintenance.

Polinsky and Scontras (2020) synthetize the potential deviation of heritage grammars from the same grammars developed in a majority language setting as avoidance of ambiguity, shrinkage of structure, and resistance to irregularity. The avoidance of ambiguity means the heritage grammar reduces the number of possible mappings in relation to the baseline; a practical example comes from heritage Japanese, where multiple interpretations of topics are reduced to merely the contrastive reading. In addition to this, Polinsky and Scontras (2020) claim that HSs reduce the overall hierarchical structure in their language. Examples of this come from comprehension research of aspect in Russian and agreement in Spanish. The last point, resistance to irregularity, manifests itself as an overregularization of input (similar to L1 acquisition); this is observed primarily in morphology, where regular forms replace irregular forms, i.e., in the regularization of English past tense forms by HSs. Since this study does not delve into irregular structures, but rather focuses on the syntax-pragmatic interface and the use of appropriate structures, I will naturally focus more on the first two aspects of the potential deviations of heritage grammar. They link the simplification of the heritage grammar to high cognitive demands, specifying how linguistic phenomena at the interface of two language domains require more processing resources compared to phenomena within a single domain. Thus, interface phenomena, such as the one currently being explored, are more vulnerable to change in the heritage grammar (p. 14). A consequence of this is the shrinking of structure, which Montrul (2010) refers to as simplification, while Polinsky and Scontras (2020) state that

I will consider this when making predictions in Section 3.

Many researchers have found that the grammatical system of the HL may be simplified compared to the input baseline (Montrul, 2010). This simplification usually takes the form of retention of the default and the loss of marked forms and may be guided by internal language factors, i.e., if the majority language has a simplified grammar with respect to the HL. For example, heritage Spanish or Russian may converge toward majority English, the latter having less morphology. Numerous studies have found a simplification of gender in Spanish and of the case system in Russian; for an overview see Montrul (2016). However, HL simplification may be guided by external factors: Rodina et al. (2020) found that factors such as the size of the HL community and the current exposure to HL instruction positively correlated with retaining the target gender system in HL Russian. They compared the acquisition of grammatical gender in HL Russian in different majority contexts, i.e., with majority languages with fewer gender values and/or less transparent gender systems than the target Russian. They found no differences based on the majority languages, but a clear effect of external variables.

In the current study, both the HL grammar and the grammar of the majority language are equally complex, with both having prenominal and postnominal variants whose use is pragmatically motivated. Nevertheless, a shrinkage of structure (i.e., simplification) is nevertheless possible as a baseline process of HL because interface phenomena are particularly vulnerable to change since they demand greater processing resources.

Bernardini (2003) analyzed Italian-Swedish bilingual children and found that CLI effects depended on dominance: Swedish-dominant children largely omitted postnominal possessives in Italian, whereas Italian-dominant children resembled Italian monolinguals. This could be an effect of CLI, as the prenominal possessive is where Swedish and Italian overlap. As acquisition progressed, Italian-dominant children paired with Italian monolinguals who had a similar distribution of possessive structures, whereas Swedish-dominant children did not produce postnominal possessives. The productions of the Swedish-dominant child were comparable to those of Swedish L2 learners of Italian in the same study. This means that dominance plays a role in CLI, but its effect remains dependent on the properties of individual languages. The CLI from Swedish to Italian consists of a lack of production of postnominal possessives in Italian, a structure not attested in the children's dominant language.

As mentioned previously, Westergaard and Anderssen (2015) found that Norwegian-English bilingual children favored prenominal possessives, consistent with surface overlap effects.

From these studies we can conclude that Italian and Norwegian possessive structures are indeed prone to CLI, so investigating how these structures might influence each other in Italian-Norwegian bilinguals can reveal more about the nature of CLI.

If we consider HL status in the aforementioned studies, we see that Italian was the HL in Bernardini (2003), but Norwegian was not the HL in Westergaard and Anderssen (2015). Thus, we may see the data as heritage Italian being simplified, i.e., only the unmarked form persists, whereas the majority Norwegian is affected by CLI from English but crucially retains both variants.

The participants were 24 bilingual children living in Norway who had either one or two Italian parents. The age range was from 4.1–10.0, and there were 14 boys and 10 girls. Sixteen children had two Italian parents, and eight had one Italian- and one Norwegian-speaking parent. The participants were recruited through personal communication and through social media on various parent groups. The parents were asked to sign an informed consent form, complete the Bilingual Lexical Experience Calculator (BiLEC) (Unsworth, 2013), and to assign an ID to the children based on the parameters provided by the researcher. The ID had to stay consistent in all tasks.¹ Prior to commencing the tasks, the children were informed about the task (game) and were asked if they wanted to take part. If a child did not respond during the testing phase, they were not forced to continue, and testing was interrupted. The children received a gift card with a value of Norwegian krone (NOK) 100 (approximately EUR 10) for their participation.

As controls, I chose to include bilingual rather than monolingual children. The reasoning behind this is that a debate has been taking place in the field of bilingualism and HL regarding the comparability of bilinguals to monolinguals using the same yard stick (Kupisch and Rothman, 2018; Rothman et al., 2022). Similarly, Vulchanova et al. (2022) had questioned the definition of native speaker. In relation to the issue of our participant choice, they defined HSs as simultaneous bilinguals, making them native speakers of both the heritage and majority language. They also argued that this, taken together with the findings that the L1 changes in lieu of successful L2 acquisition, undermined the notion of the native speaker as a point of reference (p. 4). These recent discussions have resulted in the view that comparing monolinguals to bilinguals does not meet the standards of an empirical control. An increasing number of studies are investigating the possibility that within-group comparisons of bilinguals/HSs can reveal more about the factors guiding their development (Bayram et al., 2019; van Osch, 2019), and some studies even exclude monolingual controls altogether (Meir, 2018).

Thus, I opted not to use monolingual controls. The controls were also simultaneous bilinguals, with a combination of English and one of the target languages. The 27 bilingual controls were divided into two groups: Italian-English controls (n = 12, ages 4.0–7.5) and Norwegian-English controls (n = 15, ages 4.4–9.8). The procedure of the task remained unvaried, but these children completed only half of the task; they were not tested in English. They received a gift card with a value of NOK 100 or GBP 8.

The Norwegian-English controls were children from Norwegian parents residing in the UK (n = 5), and children from one or two British parents residing in Norway (n = 10). The former group thus consisted of participants for which Norwegian was the HL, while for the latter it was the majority language, comparable to our target participants. While this division of the controls is not ideal, having the two subgroups will allow us to check for differences in the production of possessives based on the societal status of Norwegian, i.e., since English has the prenominal possessive variant only, the direction of CLI is theoretically grounded to cause a heightened production of the prenominal in Norwegian. By having the two languages have a different status, it will help us gain insight into how the societal status of the language mediates possessive production in Norwegian. All the Italian-English controls resided in the UK and, thus, maintained the HL status of Italian across the target and control participant groups.

The role of the controls was to check for CLI potential, in a more theoretically controlled setting: English only has prenominal possessives, and thus the direction of CLI is easily predicted based on previous studies. Nevertheless, the absence of monolingual controls is a potential limitation, since it is not possible to know exactly how monolingual children would respond to this task. At the same time, the bilingual controls provide a theoretically grounded benchmark for assessing CLI in our participants.

The elicitation task was designed to elicit the full array of possessive constructions in pragmatic conditions that should yield different possessive variants in each language. The task consisted of two PowerPoint presentations, one for each language, consisting of animations in which characters depicting first, second, and third person were shown doing actions and interacting with objects. Each target scene had an introductory slide, on which two characters were present along with the items that belonged to them (Figure 1). The scene then proceeded with only one of the characters being in a room, and after a few slides they interacted with their own object (Figure 2), something that happened to said object—maybe it was broken or dirty or hard to reach (Figure 3)—which then caused the character to use the object of the other character present in the introductory scene (Figure 4). The character for the participant had two versions, boy and girl, so that it was the same gender as the child. The experimenter character was different for the Italian and Norwegian versions of the task, as the experimenter administering the task was a different person for the two languages. All the images used in this task for assembling the scenes were downloaded from Vecteezy² with a pro license.

The Italian version of the task had nine target scenes (three for each character), and each scene depicted first a neutral condition (character interacting with their own object) and then a contrast condition (character interacting with someone else's object). The Norwegian version had 12 target scenes because the third-person possessive is marked for gender of the possessor. The presentation for each language had an equal number of fillers, which made the Norwegian version longer. The full list of items used in the task is displayed in the Supplementary Appendix.

The filler scenes also had an introductory slide depicting two characters and some objects. The scene proceeded with the two characters using the presented objects together; possession was not prompted or elicited.

The experiments were conducted on Zoom due to Covid-19 restrictions. The experimenter shared the screen with the participant, showing the prepared elicitation task. The interaction was recorded through Zoom, set up in such a way as to record only the shared screen and the audio, but not the faces of the participant and experimenter. The data were collected by the author (Italian task) and research assistants (Norwegian task), so that child participants would not see the experimenter speaking the other language, thereby ensuring that each task was conducted entirely in the target language.

The elicitation task started by showing the child a telescope in the Italian version and a mirror in the Norwegian version of the task. The experimenter explained how this object is magic and will take us to the land of drawings, where we (the experimenter and the participant) were also drawings. The characters for first (participant), second (experimenter), and third (friend) were then presented to the child. This way the child could relate to one of the characters and relate the experimenter to another character, which allowed us to elicit first- and second-person possessive pronouns. This worked extremely well, and the children had no problem identifying with the assigned character. Before the beginning of the task, there were two practice scenes in which each character was shown an object that belonged to them. One of the characters then played with their own object and the child was asked to describe the scene, then the same character played with someone else's object, and the child was asked to describe what was happening. The practice scene allowed us to explain to the child what description we were looking for by prompting them to include not only what the object was, but also who it belonged to. It was crucial not to prime the child with either structure, and the experimenter never provided either target structure when making suggestions to the child. In the introduction slide of the target scenes the experimenter introduced the scene with

It was of crucial importance for the experimenter not to use a possessive DP, which is why the possessive pronoun was used predicatively, as can be seen in the foregoing example. In the predicative position, the possessum is elided, and thus there is no surface order to prime the elicited structures. It took around 15 min to complete the task.

The main statistical analysis will be based on the syntactic variants of prenominal and postnominal possessives. As specified earlier, 241 target structures from the Italian task and 346 from the Norwegian task were obtained. For the statistical modeling, I coded the unmarked variant as 0 (prenominal for Italian, postnominal for Norwegian) and the marked variant as 1 (postnominal for Italian, prenominal for Norwegian). This made it possible to obtain comparable results in terms of unmarked and marked variants when including both language tasks in a single model. Figure 5 shows the distribution of the two structures, divided by condition and by language.

The most striking difference between the two languages was how these bilinguals used almost exclusively one of the structures in Italian, the unmarked prenominal one, whereas they used both structures in Norwegian.

I first plotted a base model that simply compared the productions in the two languages. This was a generalized linear mixed-effects model (GLMM) (Bates et al., 2015) with a binomial logit link to examine the effects of condition (neutral vs. contrastive) and language (Norwegian vs. Italian) on participants' production of possessive structures. The binary dependent variable was coded as 1 for marked structures and 0 for unmarked structures. The model included random intercepts for participant (ID) and test item, accounting for repeated measures within both factors. The neutral condition and the Norwegian task were set as the reference levels. The outputs of all the models are displayed in the Supplementary Appendix.

The model revealed a significant main effect of condition: participants were more likely to produce marked structures in the contrastive condition than in the baseline condition (p < 0.001). The comparison between Norwegian and Italian shows that marked variants are less likely to be produced in Italian than in Norwegian (p < 0.01). Critically, the model revealed a significant interaction between condition and language (p < 0.001). The model with Norwegian as the reference only shows that children are less sensitive to discourse distinctions in Italian; it does not tell us whether they are sensitive at all. To assess this, I ran a separate model on the Italian data to determine whether children's production patterns reflect discourse sensitivity in that language. Here I set condition as the independent variable and participant as a random effect.⁴

The model displayed a significant negative intercept, suggesting a preference for unmarked variants. Comparison of the neutral and contrast conditions revealed a marginally significant (p = 0.0538) increase in the likelihood of use of the marked variant in contrast conditions, which is an appropriate pragmatic use. Compared to the results of the Norwegian task in the base model, the distinction between conditions in the Italian task is less pronounced, in line with the visual representation of the data in Figure 5.

Additional factors, such as gender, age, education level of each parent, and the language each parent spoke, were added as additional independent variables to the models plotted for Norwegian and Italian (separately, in order to reduce the number of variables). The factors that yielded significant effects will be outlined in the following sections.

One factor impacting the results was whether the parents of the tested bilinguals were both Italian (It-It) or if one parent was Italian and the other Norwegian (Nor-It). The generalized linear mixed regression models were run separately for each language task, with response as the dependent variable and condition and parents as the independent variables; as for the random effects, the Norwegian model included both participant and test item, whereas in the Italian model, item had to be excluded as a random effect due to a singularity issue.

In the Norwegian task (Figure 6), there was no difference in the comparison of parents in the neutral condition. The interaction between condition and parentage was significant (p < 0.05), indicating a lower usage of the marked variant in the marked, contrastive context when both parents spoke Italian in the home.

Visually (Figure 6), we see that the children with two Italian parents used more prenominal variants in neutral contexts compared to the group with one Norwegian and one Italian parent. This is not target-like as the prenominal variant is marked in Norwegian, and its use is not warranted in neutral contexts. The observed overuse of the marked variant is not statistically significant, but it could be an indication of CLI.

For the Italian task, apart from the strong preference for the use of the unmarked variant, no significant distinctions were found in the fixed-effects comparisons or in interactions. This suggests that the use of the unmarked variant is very pronounced in the Italian task, and this remains unvarying regardless of whether there is one or two parents who speak Italian.

The models, one for each language, were plotted by adding age as an independent variable to our base model. Age was plotted as a continuous variable, and the model was set to show an interaction between age and condition. There was no significance in the fact that in the Italian model, age was not a relevant factor for the contextual use of the possessive variants in the Italian task. Thus, there was no observable development for this linguistic phenomenon in our dataset. This is displayed in Figure 7. To provide a clearer overview of the data, age is treated as a categorical variable in the figure (unlike in the statistical analysis, in which age is a continuous variable). At age 4, there is no observed use of the postnominal form. This differs from what previous studies found for Italian monolinguals, where there was an overuse of the postnominal form from a young age (Antelmi, 1997; Cardinaletti and Giusti, 2011; see Velnić, 2024, for an overview). One possible explanation is that the bilingual children in our study may have initially followed a similar developmental trajectory, producing postnominal forms early on, but that this stage was not captured due to the age range in our dataset. Here, the bilingual children start using the postnominal variant at age 5, but after that there is no development in the proportion of usage in the two contexts. I will address this in the discussion in relation to Italian being the HL.

The Norwegian model showed a marginally significant development with age as the older children were more target-like in the neutral condition (Figure 8). There was also an interaction between age and condition as more marked variants were used in the contrast condition with increasing age. We see in Figure 8 that children use both variants when the testing begins and that they become more target-like with age. Younger children overused the prenominal variant more in the neutral conditions than the postnominal variant in contrast conditions. This is not surprising as studies involving monolingual corpora have found an overuse of the marked variant in unmarked contexts (Anderssen and Westergaard, 2010; Velnić, 2024).

The core aim of this study was to explore how Italian-Norwegian bilinguals used the syntactic variants to express possession, which is why the main focus was on the use of the prenominal and postnominal variants. However, in both languages, emphatic stress on the unmarked variant can be used to express a contrastive meaning. In this section, I will focus on the use of emphasis. Possessives carrying stress were marked as emphatic. Table 5 summarizes the use of emphasis in the task. In both languages only the unmarked variant receives emphatic stress, with the marked variant indicating contrast is never emphasized.

The use of emphasis is mostly confined to the contrastive condition, which indicates that the children understand the distinction between the two pragmatic conditions in both of their languages.

These data points were added to the dataset on which the main statistical analysis was conducted, and the models were set as previously. The emphatic examples were coded with 1, matching the coding for the marked structures. As before, I plotted a model comparing the output per condition and language and an additional model for the data from the Italian task only.

I do not report the model on the task language comparison, as the added data points do not change the original observation. In the model on the Italian task only, there was a strong significant difference between the two conditions, as they use a marked variant significantly more in a contrast condition (p < 0.001).

This section compares the bilingual controls to our target Norwegian-Italian bilinguals. The controls were Italian-English and Norwegian-English bilinguals, and the CLI effect should be different in those languages as English only has the prenominal possessive (e.g., my car, ^*car my), which is the unmarked form in Italian but the marked form in Norwegian. Thus, I expect a high production of prenominals in the English-Italian bilinguals, though this effect is expected to be diminished for the English-Norwegian bilinguals. The responses from the controls are displayed in Table 6.

From the data presented in Table 6, English seems to cause CLI in Norwegian as the prenominal (marked) variant is used more frequently in neutral contexts compared to our target group. Nevertheless, the role of our control groups is to help us assess the role of our two target languages on each other, and this is done by comparing the results of the Italian task and Norwegian task in our target and control groups.

As before, the unmarked variants were coded as 0 and the marked variants as 1. Statistical comparisons were in line with those conducted in Section 5.3 on the main results. The markedness of the structure was set as the dependent variable, while condition and group (target vs. control) were plotted as independent variables. Participant and test item were set as random intercepts, though test item had to be removed from the model on the Italian task due to a singularity issue.

Comparison of the target children with the controls in the Italian task revealed no differences between target bilinguals and our controls. The use of the marked variant was already low in our target participants, and it continued to be low in the controls.

Comparison of targets and controls in the Norwegian task showed a clear difference between the groups. The controls used the prenominal significantly more in the neutral condition (p < 0.05), but no interaction was found for group and condition, meaning these two groups used the two variants in roughly the same way in the two conditions. This result is in line with previous studies conducted on English-Norwegian bilinguals (Westergaard and Anderssen, 2015).

While the status of Italian as the HL remained unvaried in the target and control groups, owing to recruitment issues, the status of Norwegian in the controls was mixed. The participants in the English-Norwegian control group were recruited in both Norway (from one or two English-speaking parents) and the UK (from Norwegian parents), which altered the status of Norwegian from majority to HL. This made it possible to observe different uses of the variants in the controls based on the language status of Norwegian. Figure 9 shows that the target group and the controls with Norwegian as the majority language look comparably the same, but there are statistical differences with the group residing in the UK. It was not possible to plot a single model with residency added as a factor, because target participants (Norwegian-Italian bilinguals) residing in the UK were not tested. For this reason, I ran two separate models, one comparing the targets to the controls in Norway and the other comparing the targets to the controls with Norwegian as the HL. The first model found no statistical difference at the group level, indicating that English does not have a diverse effect on Norwegian, as long as Norwegian is the majority language. The second model yielded a significant difference between the groups, with the UK controls accepting the marked form significantly more in neutral conditions (p < 0.05). I attribute this to an effect of CLI from English. The interaction of group and condition was not significant, meaning the markedness of the context did not affect the two groups differently.