1. Introduction
The processes involved in third language (L3) acquisition are yet to be fully understood. Of particular interest is whether L3 learners draw on one or both of their previously acquired languages (PALs) in acquiring an L3. A key point of interest in this line of enquiry is the initial stages of L3 acquisition, where learning is yet to take place. Understanding L3 development beyond early acquisition is a more recent endeavour, and a burgeoning research area (González Alonso et al., Reference González Alonso, Bernabeu, Silva, DeLuca, Poch, Ivanova and Rothman2025). Indeed, pinpointing the roles of structural and lexical similarity in CLI in the L3 in intermediate-to-advanced speakers is not so heavily researched. In accordance with the Linguistic Proximity Model (LPM) (Westergaard et al., Reference Westergaard, Mitrofanova, Mykhaylyk and Rodina2017), we take the position that L3 structures are initially weak representations, strengthened with further input and use, and thus becoming increasingly target-like. This brings up two questions: (1) we would expect that speakers with higher proficiency in the L3 would have more target-like use of the language, but would CLI still be present at this stage? Does structural similarity play a role in this? And (2), what is the TARGET that target-like use refers to? And what kind of input has an impact here? Here, we define input rather broadly, in alignment with the L2/L3 literature, in terms of the information available to the learner in the linguistic environment (Gagliardi, Reference Gagliardi2013). We recognise here that L1 grammar may not always be the same as prescriptive rules taught in the classroom, due to factors including ongoing language change in the L1, dialectal differences and nuanced grammatical distinctions not typically formally taught to L2/L3 learners. In this study, we thus aim to examine whether this target may represent a move towards the learned rules of a language (e.g., input from the classroom and prescriptive resources), or a move towards L1-like use (e.g., based on input through interaction with others in the community/media).
This study examines the acquisition of Norwegian as a third or additional language by L1 Polish–L2 English speakers. We designed an Acceptability Judgement Task (AJT) that was distributed online. Our participants were L1 Polish–L2 English–L3 Norwegian, L1 English–L2 Norwegian and L1 Norwegian–L2 English speakers, chosen to adhere to a subtractive language design (Westergaard et al., Reference Westergaard, Mitrofanova, Mykhaylyk and Rodina2017) so that we can isolate the potential (facilitative and non-facilitative) effect of Polish. The properties under investigation were indefinite articles (present in Norwegian and English, not in Polish) and possessive reflexives (present in Norwegian and Polish, not in English) in Norwegian (see Table 1).
Table 1. Conditions (+ indicates acceptable, – indicates unacceptable, ✘ indicates absent)
a This is listed in Żychliński et al. (Reference Żychliński, Skałba, Wrembel and Kaźmierski2023) as being unacceptable in Norwegian; however, recent literature suggests that in fact the reflexive is increasingly being supplanted by the possessive (see Section 2.3).
All three languages have non-reflexive possessives, but the extent of their reference depends on whether the meaning is divided across the types of possessives (reflexive and non-reflexive). We refer to the non-reflexive possessive simply as third-person possessives. If participants draw from both PALs at higher proficiency levels, showing co-activation, we expect the Polish group to perform better than the English group on possessive reflexives and worse on articles. We included a group of native Norwegian speakers (Norwegian–English group) to assess whether improved performance on grammaticality judgements in the learner groups necessarily reflect L1-like choices, or rather, adherence to prescriptive grammatical rules learned in classroom and textbook scenarios. Both properties represent phenomena where prescriptive rules and L1 use of the language may not align.
We first provide a broader summary of the relevant L3 models, followed by a section on comparison groups and language learners and information on the morphosyntactic properties investigated. We then describe our specific research questions and predictions, followed by our methodology, including participant information and the procedure for the experimental tasks. This is followed by the results pertaining to both research questions, and in the discussion, we interpret these in light of the current literature. Finally, we provide the study limitations and conclusions regarding L3 acquisition and learning in higher proficiency speakers and what it may mean to move towards target-like language use.
2. Background
2.1. L3 acquisition models and later-stage influence
Much of the L3 literature investigating the roles of lexical and structural similarity to date assess the initial stages of language acquisition. The various models available which theorise how this acquisition functions have a particular focus on this stage, as it precludes (as far as possible) the variables of learning (the L3) and input. However, this article aims to investigate the roles of these factors across proficiency levels in the L3. We now move to a brief discussion of the models with respect to their predictions relating to L3 acquisition.
The Typological Primacy Model (TPM) focuses primarily on and has been primarily applied to the earliest stages of L3 acquisition. The model predicts that transfer into the L3 will occur ‘wholesale’ from one of the previously acquired languages, based on a ‘Big Decision’ (Schwartz & Sprouse, Reference Schwartz and Sprouse2021) which is made upon the parser recognising similarity at one of a hierarchy of levels (the lexicon, phonology, functional morphology and syntax) (Rothman, Reference Rothman2011, Reference Rothman, Baauw, Drijkoningen, Meroni and Pinto2013, Reference Rothman2015). However, it is not clear how much time it takes for the parser to recognise this similarity and make the transfer. Recent ERP studies have been interpreted as indicating that perhaps more exposure than previously thought is required for the parser to make this decision (González Alonso et al., Reference González Alonso, Alemán Bañón, DeLuca, Miller, Pereira Soares, Puig-Mayenco, Slaats and Rothman2020; Pereira Soares et al., Reference Pereira Soares, Kupisch and Rothman2022). Once the parser has made this decision, it is possible for transfer to also come from the other language, labelled within this paradigm as Cross-Linguistic Effects (CLE), to be distinguished from Cross-Linguistic Influence occurring due to the Big Decision. Thus, it is possible within this model to see later-stage influence from both languages, through CLI and CLE.
According to the LPM and the Scalpel Model, both previously acquired languages are co-activated in the acquisition process and L3 acquisition occurs property by property (Slabakova, Reference Slabakova2017; Westergaard et al., Reference Westergaard, Mitrofanova, Mykhaylyk and Rodina2017). Thus, within the LPM and the Scalpel model, it is possible for learners’ L3 to show CLI from both lexical and structural similarity, from both previously acquired languages. The LPM and Scalpel models make predictions about CLI after the initial stages (though many studies investigating these models also focus on early L3 acquisition, e.g., Jensen & Westergaard, Reference Jensen and Westergaard2023; Mitrofanova et al., Reference Mitrofanova, Leivada and Westergaard2023). In these models, it is possible to see later-stage influence from both languages, but via a different conceptualisation of the processes occurring to that of the TPM. In these models, it is presumed that both previously acquired languages are always co-activated, in line with the co-activation literature (Blanco-Elorrieta & Caramazza, Reference Blanco-Elorrieta and Caramazza2021). Thus, they are both always available for CLI. The languages compete in the mind, with one, the more strongly activated/less inhibited structure, ‘winning out’. This does not mean it will win out every single time – but rather that, for example, someone who speaks Language A and Language B, where a structure is shared by Language A and target Language C, will be more accurate on the target Language C structure than someone who only speaks Language B. In these models, acquisition occurs property by property, and both structural and lexical similarity can play a role, from both previously acquired languages. Factors like complexity and saliency of different properties can also play a role. In these models, L3 structures are conceptualised as initially weak representations, which are strengthened with further input and use of the language (Slabakova, Reference Slabakova2017; Westergaard et al., Reference Westergaard, Mitrofanova, Rodina, Slabakova, Cabrelli, Chaouch-Orozco, Alonso, Soares, Puig-Mayenco and Rothman2023). Thus, within these models, later-stage CLI can occur but would become less likely with further proficiency and use.
The Cumulative Enhancement Model also makes predictions beyond the initial stages – in alignment with the LPM and Scalpel Model, any previously acquired properties are readily available to the L3 learner, but importantly, prior language experience may enhance subsequent language acquisition or remain neutral (Flynn et al., Reference Flynn, Foley and Vinnitskaya2004). Other L3A models which are not similarity driven, but rather default driven, predict that either the L1 (Jin, Reference Jin and Leung2009; Leung, Reference Leung and Liceras2003) or the L2 (L2 status factor model; Bardel & Falk, Reference Bardel and Falk2007, Reference Bardel, Falk, Amaro, Flynn and Rothman2012) is the source of influence on the L3. However, recent research indicates that factors including working memory, metalinguistic knowledge and language experience can have an effect on whether CLI is to come from the L1 and/or the L2 (Bardel & Sánchez, Reference Bardel, Sánchez, Angelovska and Hahn2017; Sanchez & Bardel, Reference Sanchez, Bardel, Angelovska and Hahn2017). In this study, we do not test these default-driven models, as we have a subtractive-language rather than a mirrored-groups design.
Returning to the conceptualisation of L3 structures as being strengthened with further input and use of the language, here we ponder the possibility of CLI in ultimate attainment – or differences in what ultimate attainment looks like for speakers with different previously acquired languages. This study defines ultimate attainment in alignment with Hermas (Reference Hermas2014), as a relatively stable inter-language system resulting from a prolonged and sustained exposure to the target input, and which does not exclude possible further L3 development. Previous work on L3A in terms of ultimate attainment focuses on whether it can be ‘nativelike’ – that is, whether there are differences between learners and L1 controls (Hermas, Reference Hermas2014; Leung, Reference Leung2002; Lozano, Reference Lozano2003; Rothman, Reference Rothman2011). Here, we focus on whether there are differences between learner groups in a subtractive language design – and thus, whether CLI remains throughout and much later in the acquisition process.
2.2. Comparison groups and language learners
Making comparisons with L1 ‘monolingual’ control groups as a baseline, as has been done extensively throughout L2 acquisition literature, can be problematic for various reasons, not least because L1 groups are different to Ln learners in various ways (De Houwer, Reference De Houwer2023; Rothman et al., Reference Rothman, Bayram, DeLuca, Dunabeitia, Gharibi, Hao, Kolb, Kubota, Kupisch, Laméris, Luque, van Osch, Soares, Prystauka, Tat, Tomic, Voits and Wulff2022). One difference is the multicompetence of L3 learners, who have at least two constantly co-activated languages in their multilingual minds (Cook, Reference Cook and Cook2003, Reference Cook, Cook and Wei2016). Such learners thus operate with a cognitive mindset different from monolingual speakers. Additionally, current research is moving away from an idealised characterisation of the native speaker due to the recognition of widely documented individual variation in the speakers’ linguistic competence (Vulchanova et al., Reference Vulchanova, Vulchanov, Sorace, Suarez-Gomez and Guijarro-Fuentes2022). Experience-based variation exists within both monolingual and multilingual groups, which should be taken into consideration.
Furthermore, on a global scale, multilingualism is the norm and monolingual normativity bias is thus an issue. On a practical note, it can also be difficult to find true functional monolinguals for data collection. The issue of monolingual normativity bias is especially problematic where L1 ‘monolinguals’ are compared with heritage speakers (HS), as discussed in Rothman et al. (Reference Rothman, Bayram, DeLuca, Dunabeitia, Gharibi, Hao, Kolb, Kubota, Kupisch, Laméris, Luque, van Osch, Soares, Prystauka, Tat, Tomic, Voits and Wulff2022).
Having recognised this (and in alignment with Rothman et al.’s (Reference Rothman, Bayram, DeLuca, Dunabeitia, Gharibi, Hao, Kolb, Kubota, Kupisch, Laméris, Luque, van Osch, Soares, Prystauka, Tat, Tomic, Voits and Wulff2022) article), we posit that there are some situations where the inclusion of L1 SPEAKERS in an Ln study may be useful, specifically from a sociolinguistic perspective. We propose that this is particularly true for Ln learners living in the country whose majority language they are acquiring, as opposed to Ln learners residing in their home country and acquiring a language in an instructional setting. The latter group is not extensively exposed to input from interaction with L1 speakers and thus comparing them to native speakers may be less relevant. But our participants are residing in the country where their Ln is spoken (Norway), and thus the inclusion of an L1 group acts as a realistic input our learner groups are exposed to through interaction. We use the terminology L1 speakers here, rather than monolinguals, for two reasons. First, in taking this sociolinguistic perspective where it is the NATURE OF THE INPUT provided to the Ln learners that is of interest, it is not of concern whether the L1 speakers are monolingual – only that they are input providers for Ln learners. Second, as is generally the case in this particular environment, the Norwegian participants are at least bilingual, with all of them knowing English (mean years of study = 17.4, SD = 3).
In further defining input to develop the research question, we refer the reader to the diagram proposed by Zhao (Reference Zhao2021), wherein input leads to intake (storage in temporary memory, incorporation of data available in the input), affecting the interlanguage system (becoming part of long-term memory), leading to output. Interaction in the form of conversation can also act as input and it can provide positive feedback from the interlocutor, which can further modify output. Moving back to the conceptualisation of L3 structures as initially weak representations, strengthened with further INPUT and use, we thus question what a move towards this TARGET-LIKE USE refers to – and therefore, what input seems to have a greater effect on the language use of more proficient L3 speakers. Is this the prescriptive ‘rules’ of a language (e.g., as learned in a classroom), or L1-like use (from input via interaction in the community)? To understand this, one needs to understand what L1-like use looks like – how (some of) the people that provide the L3 language input and interaction use language. One can assess how frequently speakers communicate with L1 speakers, or how frequently they use the L3, but without knowing what the L1 speakers are doing, it can be difficult to answer this question.
Part of this relates to the fact that there is variation in L1 speakers and that how they use the L1 changes over time as well. For example, in Norwegian, the L3 in question for this study, there is an ongoing dialectal change in the Tromsø dialect of Northern Norway, where speakers are moving towards the use of ikke as a replacement for ikkje (NEG) (Fiva, Reference Fiva2016; Nesse & Sollid, Reference Nesse and Sollid2010; Sollid, Reference Sollid2019). Despite this, L3 learners of Norwegian who use the Tromsø dialect acquire and use the ikkje form (Castle et al., Reference Castle, Jensberg, Velnić, Malarski, Jensen and Wrembel2024). Another reason for examining L1-like use to understand L3 acquisition is that form-to-meaning mappings are often more complex than what morphosyntactic rules may suggest. For example, Norwegian has articles – but if one considers the interface of syntax/semantics, and thus the genericity paradigm, one can see that Norwegians do not use articles in quite the same way that English speakers do, there are different form-to-meaning mappings (Velnić et al., Reference Velnić, Slabakova and Dahl2025). This is not necessarily explicitly taught in the classroom, as both teachers and textbooks generally tend to offer very little instruction regarding the teaching of articles for generic reference (Snape & Yusa, Reference Snape, Yusa, Whong, Marsden and Gil2013). Such constructions can therefore be more difficult for learners to acquire. Thus, we need a clearer view of the current state of the target language in the contexts that we are testing and a simple way of obtaining this is by testing native speakers on the same items as our target learners. For the L3 speaker, for example, input comes from both the classroom and prescriptive resources (e.g., use articles in Norwegian) and from interaction and use in the community (e.g., in some contexts, nouns do not require articles in Norwegian). To investigate which type of input is having a greater effect, it is useful to examine properties where output based on prescriptive resources versus interaction and use in the community may differ.
2.3. The properties investigated
The properties under investigation in this study are indefinite articles and possessive pronouns (specifically, third-person possessive and reflexive possessive pronouns). English and Norwegian use articles to indicate definiteness, and both have pre-nominal indefinite articles. Polish, on the other hand, does not have articles.
Norwegian has post-nominal definite articles, though we only test indefinite articles in this study. Thus, for this property, POL ≠ ENG = NOR. It is important to note here, though, that English and Norwegian have slightly different norms for when use of an article would be appropriate. For example, Norwegian has a much wider use of the bare noun, while bare nouns are only acceptable in English when referring to a mass noun (i.e., gold is valuable, *dog is a mammal), Norwegian employs bare count nouns much more freely signalling a TYPE rather than a TOKEN (Borthen, Reference Borthen2003). Nevertheless, the DPs that we test in the article condition were all in object position with a specific-indefinite referent, and there should be no discrepancy between the two languages there.
Possessive pronouns, on the other hand, represent a property where ENG ≠ POL = NOR. Polish and Norwegian require a different form where possessive pronouns are reflexive, as opposed to object oriented.Footnote 1 English is ambiguous, using the same form for both of these meanings (4). Examples (5) and (6) show use of the reflexive possessive pronoun in Norwegian and Polish, respectively, while examples (7) and (8) show use of the object-oriented possessive in the two languages.
It is important to note that there is gradient acceptability for the use of possessive pronouns where their use would be normatively incorrect in Polish (i.e., according to the rules of the language, a reflexive possessive should be used) (Żychliński et al., Reference Żychliński, Skałba, Wrembel and Kaźmierski2023). Żychliński et al. (Reference Żychliński, Skałba, Wrembel and Kaźmierski2023), in their study on the effect of gradient acceptability in L1 Polish on L3 acquisition of Norwegian, found that AJT ratings on sentences with subject-oriented pronouns were low (compared to definite and indefinite articles, which they also tested). They propose that this may be due to the negative influence of gradient acceptability in L1 Polish, as well as differences to L2 English. They cite Hestvik (Reference Hestvik1992) in noting that acceptability of (non-reflexive, anti-subject oriented) possessive pronouns in subject-oriented scenarios is not gradient, and that such readings are impossible. However, variability in L1 Norwegian has recently been observed, suggesting an ongoing change in the Norwegian possessive system (documented for Swedish in Lundquist, Reference Lundquist2013; Sundgren, Reference Sundgren, Gustafsson, Holm, Lundin, Rahm and Tronnier2016; Julien, Reference Julien2020), with third-person possessives increasingly supplanting reflexive forms (Gieselmann, Reference Gieselmann2024, Westergaard, p.c.). Thus, it seems that both Polish and Norwegian have gradient acceptability here, while English has only one form available. This will be an important distinction for examining the results. Table 1 shows the acceptability of each condition in each of the languages.
3. The current study
3.1. Research questions
The current study aims to investigate CLI in L3 acquisition across proficiency levels towards the higher end of the proficiency spectrum, while considering what target a move to target-like language use represents. Thus, our research questions are as follows:
-
1. Does CLI based on structural and lexical similarity occur from both previously acquired languages in L3 speakers with higher proficiency levels?
-
2. With increasing L3 proficiency, do learners approximate the target language in a manner that mirrors actual L1 speaker performance or do they align more closely with prescriptive grammatical rules of the L3?
3.2. Predictions
3.2.1. Research question 1
For research question 1, in terms of whether CLI based on structural and lexical similarity can occur from both previously acquired languages, we base our predictions on the LPM and the Scalpel model. These models propose that, as both/all previously acquired languages are co-activated in the brain, structural and lexical similarity can occur from both (Slabakova, Reference Slabakova2017; Westergaard et al., Reference Westergaard, Mitrofanova, Mykhaylyk and Rodina2017). We use a subtracted language groups design to test this, with a Polish–English–Norwegian group and an English–Norwegian group (see Section 4.3). Thus, we expect that the Polish–English–Norwegian group will perform more accurately on the possessive pronoun conditions (facilitative CLI from Polish), and less accurately on the article conditions (non-facilitative CLI from Polish), and vice versa for the English–Norwegian group. Note also that English and Norwegian have greater lexical similarity overall, possibly increasing the general activation of English in both groups. In terms of whether CLI can occur at higher proficiency levels – the LPM and Scalpel model allow for this, as it is noted that, with input and learning, L3 use becomes MORE target-like (but this does not necessarily entail that the speakers will always be at ceiling in their production once the property has been acquired), and learners become better at inhibiting influence from previously acquired languages (Jensen et al., Reference Jensen, Mitrofanova, Anderssen, Rodina, Slabakova and Westergaard2021). Indeed, Jensen et al. (Reference Jensen, Mitrofanova, Anderssen, Rodina, Slabakova and Westergaard2021) find that CLI obtains at developmental stages of L3 acquisition, though the developmental slopes for the different properties were not equal.
Note also here that the other L3 acquisition models also predict that both languages are likely to influence L3 DEVELOPMENT, but that the influence of both languages may not be equally strong at all stages. González Alonso et al. (Reference González Alonso, Bernabeu, Silva, DeLuca, Poch, Ivanova and Rothman2025) note this as an area where the field is (more or less) in agreement. This study design, however, was informed by the predictions of the LPM.
3.2.2. Research question 2
For research question 2, to test whether participants move towards target-like performance with proficiency, we include proficiency as a variable. Based on the literature regarding ultimate attainment, we predict that it is possible for L3 (and L2) acquirers’ grammars to converge with ‘native’ knowledge in terms of parameter resetting, though this may not always occur, and depends on the parameters under study (Hermas, Reference Hermas2014). There are some features in the L2 which can be persistently difficult to acquire despite favourable circumstances for acquisition (Schmid et al., Reference Schmid, Gilbers and Nota2014). Thus, we predict that both Norwegian learner groups will move towards target-like use with higher proficiency levels, though by property this may be modulated by factors such as complexity and saliency, for example, the well-documented difficulty associated with the acquisition of articles by native speakers of languages without articles (Hermas, Reference Hermas2018; Jaensch, Reference Jaensch, Slabakova, Rothman, Kempchinsky and Gavruseva2008). Indeed, this may be an indication of the Cumulative Input Threshold Hypothesis at work, wherein input over the lifetime is inversely correlated with the rate of developmental recovery from non-facilitative transfer (Cabrelli & Iverson, Reference Cabrelli and Iverson2024).Footnote 2 In line with this theory, the Polish group’s threshold for acquisition of the article may be quite high and would thus require a considerable amount of L3 exposure to adopt the L3 ‘rules’.
To test whether this TARGET is L1-like use, or more in alignment with prescriptive grammatical rules, we include a Norwegian–English speaker group. We selected two properties wherein the grammatical RULE is predicted to not necessarily match with the behaviour of L1 speakers (see Section 2.3). We predict that the behaviour of the L1 Norwegian participants may not be strictly in line with prescriptive morphosyntactic rules around articles and possessive pronouns as learned in the classroom, due to differing form-meaning mappings and gradient acceptability (see Section 2.3). Native speakers are included in the study in order to compare our L2/L3 groups against real intuitions and not prescriptive rules. Hence, our analysis of the native performance (i.e., what the ‘target’ represents) is also exploratory, but the comparison between the native controls and the learner groups is theoretically grounded. If learners behave in alignment with L1 speakers on the ‘matching’ properties (e.g., if there are no differences between the Norwegian and English groups for the grammatical article condition), with a movement towards more L1-like behaviour with increased proficiency, we may conclude that their behaviour is L1 like. However, it may also be the case that L3 learners’ behaviour adheres more to prescriptive rules. Such a conclusion could be made if the learner groups are stricter in their grammatical judgements than the L1 Norwegian group (in alignment with prescriptive rules) on the ‘matching’ properties and/or move closer to a strict alignment with prescriptive rules that differ from the performance of the L1 Norwegian group with greater proficiency.
4. Methodology
4.1. Participants
Participants were L1 Polish–L2 English–L3 Norwegian, L1 English–L2 Norwegian and L1 Norwegian–L2 English speakers living in Norway. They were recruited via flyers shared on university campuses and via online expatriate groups. A total of 60 participants were recruited and four were excluded. Exclusions were made on the basis of knowledge of either the subtracted language (Polish) in another group (e.g., an L1 English speaker with knowledge of Polish), or participants not from the L1 Norwegian group reporting knowledge of Norwegian and other Northern Germanic languagesFootnote 3 from childhood. There was thus a total of 56 participants included in our analyses (41 women, 2 non-binary) (Table 2). The lowest participant score was 68.8% on the proficiency test.
Table 2. Participant meta-data on age, proficiency score and length of residency
Participant proficiency scores were high among all three groups, with the L1 Norwegians unsurprisingly having the highest mean score with a lower standard deviation. In terms of input and contexts of L3 acquisition, for both the L1 Polish and L1 English groups, a majority of participants had engaged in a Norwegian language course (74% and 61% respectively) (Table 3). The L1 English group spent on average 11 years learning Norwegian, while the L1 Polish group spent 5 years learning Norwegian, but 16 learning English. A greater proportion of participants from the L1 English group used Norwegian than from the L1 Polish group in every category except ‘with colleagues’ and it was only speakers from the L1 English group that used Norwegian in all five possible input environments (with partner, with family, with friends, with colleagues, with others). Indeed, 47.8% of the L1 English participants used Norwegian in a family environment, while no L1 Polish participants did so.
Table 3. Participant meta-data on contexts of L3 input and use for learner groups
4.2. Procedure
Participants used the QR code on the flyer to sign up via Nettskjema.Footnote 4 The signup form operated as both a consent form and a screening tool.Footnote 5 Participants were sent instructions on how to engage in the experiment in their L1 (as indicated on the signup form). All instructions for the entire experiment were in Norwegian, except for the language background questionnaire, which was in the respective L1s (Polish, English or Norwegian).
Participants were instructed to first engage in the main experiment. This was an Acceptability Judgement Task with a five-point Likert scale followed by a proficiency task hosted on JATOS (Lange et al., Reference Lange, Kühn and Filevich2015). Participants were then instructed to fill out a background questionnaire. Participants did the study entirely online (e.g., not in the laboratory). Our results are, however, comparable to those of a similar study with a different focus regarding gradience in L1 Polish and L3 Norwegian where data were collected in-person (Żychliński et al., Reference Żychliński, Skałba, Wrembel and Kaźmierski2023). The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
4.3. Experiment design
We used a subtractive language design (Westergaard et al., Reference Westergaard, Mitrofanova, Mykhaylyk and Rodina2017) in order to isolate the roles of individual previously acquired languages. Thus, by comparing the Polish–English–Norwegian group to the English–Norwegian group, we may be able to isolate the effect of Polish on L3 Norwegian. Differences between the L3 group and the L2 group(s) may be attributed to the influence of the subtracted language (Westergaard et al., Reference Westergaard, Mitrofanova, Rodina, Slabakova, Cabrelli, Chaouch-Orozco, Alonso, Soares, Puig-Mayenco and Rothman2023). For example, if Polish–English–Norwegian speakers and English–Norwegian speakers are tested on a property in Norwegian which also exists in Polish but not in English, if Polish speakers perform better in terms of accuracy in the property than the English–Norwegian speaking group, this may indicate CLI from Polish. The English–Norwegian group does not have access to facilitative CLI for this property. A full subtractive language group design would also include a Polish–Norwegian group that would warrant the isolation of the effect of English. Unfortunately, it was not possible to find such a group. However, this situation is not uncommon and use of just one subtracted language group (e.g., English–Norwegian) is prevalent in the literature using this type of design (Castle et al., Reference Castle, Jensen, Mitrofanova and Westergaard2025; Mitrofanova et al., Reference Mitrofanova, Leivada and Westergaard2023).
In the current design, we specifically tested for differences between the English and Polish groups in conditions wherein (1) Norwegian matched with English but not Polish (articles) and (2) Norwegian matched with Polish but not English (possessive and reflexive pronouns) (Table 1).
We included both a condition wherein accepting a possessive pronoun would be ‘correct’ in Norwegian and one where accepting a reflexive pronoun would be ‘correct’ in Norwegian (as opposed to only possessives or only reflexives ever being correct, which may lead to participants learning the pattern in the test and acting accordingly).
Each test sentence was preceded by a contextualising sentence. In the article condition, for example, the contextualising sentence(s) Elisabeth er glad i hester. I sommer jobbet hun for å spare penger ‘Elisabeth loves horses. Over the summer she worked to save money’ were used before the participant would judge a sentence like the one below.
Providing a contextualising sentence was especially important for the reflexive and possessive conditions, because the semantic content of the sentence indicates what the ‘correct’ answer should be in the test sentence. For example, for the contextualising sentence Thomas og Per jobbet sent, og etterpå dro Per uten å ta med seg jakken ‘Thomas and Per worked late, and afterwards Per left without taking his jacket’, the sentence below, using a possessive pronoun, would be ‘correct’.
Each test sentence was rated on a Likert scale from 1 to 5. These numbers were colour coded to minimise perceptual differences between what each number indicates for different participants (Figure 1). Six practice sentences with filler items were introduced first, so that participants knew how the experiment worked. There were 38 test sentences, with six conditions, equal grammatical and ungrammatical test sentences and two lists (19 test sentences per list). There were 36 filler sentences. The filler items investigated genericity and definiteness in plural nouns and are the test items for a separate study.
Figure 1. AJT task screen shown to participants.
The proficiency task was an adapted proficiency test from Language Trainers (2015), which allowed us to test proficiency up to a C2 level of Norwegian. All groups were required to do this task.
The background questionnaire consisted of general meta-data questions, as well as additional questions regarding bilingualism. A full list of the questions in all three languages is available on OSF (see Data availability statement).
4.4. Statistical analysis and coding
Data from the JATOS server (experiment data) and Nettskjema (meta-data and background questionnaire) were downloaded. The data from the JATOS server were run through OpenSesame to create a spreadsheet for each participant. All CSV files were then cleaned and merged in R. All response times of under 1000 ms were removed to mitigate potential noise in the data caused by participants clicking through and not engaging properly with the task.
We fitted an ordinal model to predict the response value (Likert scale: 1–5) against the L1 (Polish, English or Norwegian) and the condition (article grammatical, article ungrammatical, poss_3rd grammatical, poss_3rd ungrammatical, poss_r grammatical, poss_r ungrammatical) using the ordinal package (Christensen, Reference Christensen2023). Participant and test item were included as random effects (clmm(resp_value ~ L1×cond + (1|id)). Item was not included as the boundary fit was singular. A likelihood ratio test indicated no improvement in model fit in including this variable (Pr(>Chisq = .98). The joint_tests function from the emmeans package was used to examine the significance of L1, condition and their interaction (Lenth et al., Reference Lenth, Bolker, Buerkner, Giné-Vázquez, Herve, Jung, Love, Miguez, Riebl and Singmann2023). To assess whether the L1 Polish and L1 English groups differed on each individual condition (RQ1), pairwise analyses were conducted also using the emmeans package.
To assess how sentence ratings changed with increased proficiency in L3 Norwegian, we created an ordinal effects model to predict response value against L1, condition and Norwegian proficiency (proficiency questionnaire score), with the L1 Norwegian data removed. We also included participant and test item in the random effects structure here (mdl < − clmm(resp_value ~ L1×cond×proficiency + (1|id)). As earlier, item was not included as the boundary fit was singular, and a likelihood ratio test indicated no improvement in model fit in including this variable (Pr(>Chisq = .974). We then examined variable significance (joint_tests) followed by conducting pairwise analyses using the emtrends function in the emmeans package.
5. Results
We first present the results of average Likert scale responses in grammatical and ungrammatical conditions, for each property, for the three L1 groups. Table 4 indicates the mean ratings for each condition.
Table 4. Mean ratings per condition
These results are visualised in the box plot below (Figure 2).
Figure 2. AJT results per condition and L1.
In the article condition, there is a greater difference between the grammatical and ungrammatical conditions for the Norwegian and English groups than there is for the Polish group, a pattern led by the higher acceptance of the ungrammatical condition by the Polish group. The average acceptance of the grammatical article by the Norwegian group is quite low (mean = 3.33), and there is a lower ceiling (Figure 2).
For the possessive_3p condition, the Polish group appears to be more precise than the English group. All groups make a clear distinction between grammatical and ungrammatical items, but for the English group there seems to be the most overlap. The Norwegian natives are precise in judging ungrammatical items, the Polish group have more precision with grammatical items.
For possessive_refl, the Norwegian natives do not have a clear ‘grammatical stair’. They have a very varied acceptability of the ‘ungrammatical’ condition – wherein the third-person possessive is used in the place of a grammatically required reflexive possessive. The English group has a wide range of accepting the grammatical items. The Polish group is less nuanced in rejecting the ungrammatical items.
To address research question 1 and predict the Likert scale response value based on L1 group and condition (e.g., article grammatical, article ungrammatical, poss_3rd grammatical, etc.), we fitted an ordinal model. The model’s total explanatory power is substantial (conditional R 2 = 0.41) and the part related to the fixed effects alone (marginal R2) is of 0.30. A joint test showed a non-significant effect of L1 (F(2, Inf) = 1.2, p = .30), a significant effect of condition (F(5, Inf) = 55.8, p < .0001) and a significant interaction between L1 and condition (F(10, Inf) = 6.46, p < .0001).
The predicted values can be viewed in Figure 3. The English group (estimate = −1.12, p = .01) and Norwegian group (estimate = −1.65, p = .0007) are better in rejecting ungrammatical article sentences (sentences without articles) than the Polish group. The Polish group (estimate = −1.37, p = .007) and the English group (estimate = 1.12, p = .02) are more accepting of grammatical article sentences than the Norwegian group. The Polish group is more accepting of the grammatical third-person possessive than the English group (estimate = −1.47, p = .007) and the Norwegian group (estimate = −1.47, p = .014). The English (estimate = −1.11, p = .043) and Polish (estimate = 1.18, p = .035) groups both rate ungrammatical possessive reflexives lower than the Norwegian group. The full output of the model can be viewed in Appendix 1.
Figure 3. Predicted response values per group and condition.
To address research question 2, we fitted an ordinal model to predict the Likert scale response value with L1 group, condition and proficiency level. The model’s total explanatory power is substantial (conditional R 2 = 0.44) and the part related to the fixed effects alone (marginal R 2) is of 0.32. Joint tests revealed that the effects of condition (F(5, Inf) = 42.02, p < .0001) and the interactions between condition and L1 (F(5, Inf) = 4.59, p = .0003) and condition and proficiency (F(5, Inf) = 5.98, p < .0001) were significant.
With increasing proficiency, performance on the AJT improves in rejecting ungrammatical articles (estimate = −7.66, p = .023) in the English group. Performance on the AJT also improves in rejecting ungrammatical possessives (estimate = −10.52, p = .009) in the Polish group. The full output of the model can be viewed in Appendix 2.
The predicted patterns for the two significant conditions can be viewed in Figures 4 and 5. Detecting ungrammaticality seems to be easier than attuning to grammatical items and this may be due to the yes-bias that the lower proficiency participants seem to have.
Figure 4. Predicted response values per group, condition and proficiency (accuracy) in ungrammatical possessive condition.
Figure 5. Predicted response values per group, condition and proficiency (accuracy) in ungrammatical article condition.
6. Discussion
In this section, we discuss our findings in relation to the research questions. We provide the research questions for ease of reading, followed by a brief mention of our predictions and then the results in relation to this and the literature.
6.1. Research question 1
Does CLI based on structural and lexical similarity occur from both previously acquired languages in L3 speakers with higher proficiency levels?
For this question, in alignment with the LPM, we specifically predicted that the Polish group would perform more accurately (accepting grammatical and rejecting non-grammatical) than the English group in the possessive and possessive reflexive conditions (facilitative CLI from Polish) and less accurately than the English group in the article condition (non-facilitative CLI from Polish). However, we also expected that this pattern may not be so strong for reflexive possessives due to gradient acceptability in Polish and L3 Norwegian.
Indeed, results indicate that the English group (p = .01) and the Norwegian group (p = .0007) are better in rejecting ungrammatical article sentences (sentences without articles) than the Polish group. Additionally, the Polish group tends to rate the grammatical article sentences highly. We argue that this difficulty with articles and tendency to rate all article-containing sentences highly could be indicative of CLI from the ‘subtracted language’ Polish, as Polish does not have articles. Those with L1 Polish have more difficulty with and uncertainty around articles than L1 English and L1 Norwegian groups. Despite broader overall lexical similarity with L2 English (and structural similarity in the grammatical condition), structural influence from Polish appears to occur at least in the ungrammatical article condition. These issues may also persist in Polish learners as the use of Norwegian articles can be quite complex. The vast body of literature on L2 acquisition is in agreement that learners of article-less languages struggle with the acquisition of this linguistic feature (Ionin et al., Reference Ionin, Ko and Wexler2004; Park, Reference Park2013; Snape et al., Reference Snape, Umeda, Wiltshier, Yusa and Stringer2016; Trenkic, Reference Trenkic2007). Additionally, Norwegian is more permissive with use of the bare singular when compared to languages such as English where the use of bare count nouns is undoubtedly ungrammatical (*Mari caught red butterfly). Nevertheless, we must address that the L1 Norwegian speakers did not always reject test items with bare singular nouns, which we have labelled as ungrammatical in the current task. There were no obvious outlier items when we looked at the average per item, with a few items reaching a mean judgement of 2.09 out of our five-point scale. Some participants were more permissive than others, averaging at around 3 (n = 2). However, when we looked into the individual judgements, we found an item that was judged as good or very good. The item in question was Mari fanget rød sommerfugl (‘Mari caught red butterfly’). Provided the preceding context (Mari likes butterflies. Yesterday she tried to catch one), it is possible that the Norwegian participants that rated this item interpreted red butterfly as a type-denoting item, a reading that defines the NP as a species rather than a butterfly that is red. A parallel example could be black bird in English: it can refer to a bird that is black or to the species Turdus merula. While English would not morphologically distinguish between the reference of these two (both need an article), Norwegian is different as it is much more permissive with its use of the bare forms (Borthen, Reference Borthen2003; Velnić et al., Reference Velnić, Slabakova, Dahl and Listhaug2024). However, an investigation into the Norwegian article system and generic and type reference is outside the scope of this article. That said, it is clear that the Polish speakers face multifaceted challenges when navigating the Norwegian article system: not only do they have to acquire an article system, but they must also learn when to refrain from using articles, and the optional cases and subtle differences in meaning make this task all the more challenging.
In terms of participant performance on the possessive conditions, there were no statistically significant differences between the English and Polish groups for the reflexive possessive. The lack of difference between Polish group and English group, despite Polish having reflexive possessives, may possibly relate to gradient acceptability of reflexive possessives in Polish. Indeed, Żychliński et al. (Reference Żychliński, Skałba, Wrembel and Kaźmierski2023) also did not find significant differences between their Polish–English–Norwegian group and English–Norwegian groups for subject-oriented pronouns and related this to gradient acceptability in Polish. It may also be related to a general relative difficulty of acquisition of this property (Helland, Reference Helland2017; Slabakova, Reference Slabakova2017) and the neutralisation of positive (Polish) and negative (English) CLI (Westergaard, Reference Westergaard2021). However, it must be noted that there may also be an L3 input-based effect – as there is increasing gradient acceptability in L1 Norwegian itself (discussed in 6.2 next).
There were, however, differences between the groups in the third person possessive condition. The Polish group is better in accepting the grammatical possessive condition than the English group (p = .0067) and the Norwegian group (p = .014). Both the Polish and English groups have the third-person possessives available in their previously acquired languages. The Polish group also has both the possessive and possessive reflexive available from their L1, Polish. For the English group, this difference between their performance in this condition and that of the Polish group may be related to uncertainty regarding whether the possessive or the reflexive possessive is appropriate in different contexts, due to English not having the reflexive possessive property. However, this does not explain why the Polish group also rates this condition more highly than the Norwegian group. The difference appears to be led by the Polish group’s very high rating of this condition (see Figure 3) – they are particularly accepting of the grammatical possessive. This may be linked to the following of grammatical rules regarding this morphosyntactic condition, as opposed to other aspects of the sentence affecting well-formedness (i.e., for the Norwegian native speakers, similarly to the article condition discussed earlier).
6.2. Research question 2
With increasing L3 proficiency, do learners approximate the target language in a manner that mirrors actual L1 speaker performance, or do they align more closely with prescriptive grammatical rules of the L3?
With increasing proficiency, performance on the AJT improves in rejecting ungrammatical articles for the English group. For the Polish group, performance on the AJT improves with increasing proficiency for rejecting ungrammatical possessives (sentences using the reflexive where context calls for the possessive). Adherence to grammatical rules appears to improve with greater proficiency in some properties, while other properties remain more difficult to acquire, likely due to differences in complexity and saliency and thus different learnability levels of different properties (Jensen et al., Reference Jensen, Mitrofanova, Anderssen, Rodina, Slabakova and Westergaard2021; Slabakova, Reference Slabakova2017) and possibly the influence of cumulative input from the L1 (Cabrelli & Iverson, Reference Cabrelli and Iverson2024). The conditions in which the Polish and English groups become more target like with greater proficiency are those where both learner groups already match the L1 Norwegian group in the Model 1 comparisons. There are no statistically significant differences in performance between the L1 English and L1 Norwegian groups for ungrammatical articles (a condition where Norwegians are quite strict in their rejection, in line with prescriptive rules) and between the L1 English or L1 Polish groups and the L1 Norwegian group for ungrammatical possessives (a condition where all three groups can be expected to behave similarly).
However, to fully examine whether adherence to grammatical ‘rules’ equates to L1-like choices, the differences between the L1 Norwegian group and the intermediate-to-advanced Ln groups across proficiency levels must also be examined. In terms of articles, the Polish group (p = .007) and the English group (p = .026) are more accepting of grammatical article sentences (sentences with an article) than the Norwegian group. For the Polish group, this may be linked to uncertainty around article use and a tendency for acceptance where errors cannot be identified. However, the same cannot be said for the English group, who have articles in the only previously acquired language from which CLI could occur. Indeed, overall, Norwegians generally have a lower ceiling for grammatical articles (~4, see Figure 2). We relate this difference between groups to differences in form-to-meaning mapping between Norwegian and English. Norwegian allows for bare count nouns, whilst English does not. This adherence to stricter grammatical rules in the L1 English group shows evidence of CLI from English as well as a possible preference for prescriptive rules.
Across proficiency levels, the English (p = .043) and Polish (p = .035) groups both rate ungrammatical possessive reflexives lower than the Norwegian group. The Norwegian group are quite accepting of ungrammatical possessive reflexives (i.e., using the object-oriented as opposed to the subject-oriented pronoun, using hans/hennes instead of sin). As predicted in Section 3.2, the variability in ungrammatical possessive reflexives for L1 Norwegians may be linked to recent observations suggesting an ongoing change in the Norwegian possessive system (documented for Swedish in Lundquist, Reference Lundquist2013; Sundgren, Reference Sundgren, Gustafsson, Holm, Lundin, Rahm and Tronnier2016; Julien, Reference Julien2020), with third-person possessives increasingly supplanting reflexive forms (Gieselmann, Reference Gieselmann2024; Westergaard, p.c.). Indeed, Lødrup (Reference Lødrup, Hartmann, Mursell and WurmbrandPreprint) finds that the non-reflexive plural possessive is replacing the reflexive in Norwegian, posited to be strengthened by translation from English. Further research is needed here.
Thus, a rating across proficiency levels for ungrammatical possessive reflexives which is already below that of the Norwegian group (and does not change with increased proficiency) may indicate that it is perhaps rather grammatical rules which participants are in line with (i.e., from classroom learning and prescriptive resources), as opposed to (adequate) input from interaction with L1 Norwegian speakers. In other words, the Polish and English groups both behave more in line with the prescriptive morphosyntactic rules of Norwegian – to use the possessive reflexive for subject-oriented possessives – than with the behaviour of L1 Norwegian speakers, for whom this feature is being supplanted by third-person possessives. This is also the case for the English group in the grammatical article condition. Such results occur for both groups despite differing levels of use of Norwegian in daily life and the fact that the L1 English group have, on average, lived in Norway for longer than the L1 Polish group (see Tables 2 and 3). We thus tentatively suggest that input from prescriptive rules (e.g., in the classroom or learning resources) may have a greater effect on the interlanguage system and thus output and language use than does input from interaction with the community, at least in this speaker group.
We additionally ponder three, possibly compounding, explanations regarding this result. First, it is possible that the ‘adequate’ input level from community interaction required for movement towards more L1-like use may be higher for some properties than others, due to their complexity, saliency and consequently difficulty in learning (which may differ for learners with different previously acquired languages). Second, participants may have a standard language ideology (Milroy, Reference Milroy, Bex and Watts1999, Reference Milroy2001; Milroy & Milroy, Reference Milroy and Milroy1999). This is especially prevalent in L2 classroom learners, whose preferences may be tied to pedagogical norms regarding the notion of a standard language (Milojičić, Reference Milojičić2023; Ott, Reference Ott2025). They may thus view prescriptive norms as the ‘correct’, ideal form and prioritise them over naturalistic variants (even if aware of and able to use the latter). Relatedly, the demand characteristics of the experimental context may have had an influence here (Leustek, Reference Leustek and Allen2017). Participants knew they were taking part in a study on language learning and they thus may have defaulted to rule-based performance rather than naturalistic usage, assuming that the task required ‘correct’ answers.
7. Limitations
Ideally, for a full subtracted-language groups design, we would also include an L1 Polish–L2 Norwegian group, to allow us to pinpoint possible effects of L2 English in the L1 Polish–L2 English–L3 Norwegian group. Those with no English may behave differently regarding articles and perhaps have less difficulty with possessive reflexives. However, this group is difficult to find due to the ubiquity of English. It would also have been useful to include an article condition where L1 Norwegians would prefer no article (e.g., where a noun functions as a descriptor of a person’s profession or identity, jeg er lege, lit. ‘I am doctor’). This would form an equivalent to our testing of sentences where the reflexive possessive would be both correct and incorrect.
Finally, it would be useful to assess learners at different stages of acquisition – essentially, a longitudinal study assessing the same learners over time. There is a need for longer term longitudinal studies to better understand the L3 developmental process and use over the lifetime.
8. Conclusion
The current study aims to understand the roles of structural and lexical similarity in CLI in L3 acquisition at higher proficiency levels. In line with the LPM, TPM, and Scalpel model, we find that structural (dis)similarity plays a role throughout the developmental process. This is in addition to other factors including gradience of acceptability and differences in salience and complexity of each property.
We also show that L1 behaviour does not necessarily represent the set of prescriptive morphosyntactic ‘rules’ of a language that Ln speakers learn. This behaviour is useful to know about, as it informs us as to how speakers that Ln speakers are in contact with (and receive input from) use the language, as opposed to relying on standard prescriptive grammatical rules for comparison. This provides insights as to what Ln speakers do in terms of moving towards learned rules (e.g., from input in the classroom) versus building L1-like representations (from input through interaction with community members). Lners align more with grammatical rules in conditions where the L1 Norwegians behave similarly. However, in comparing with L1 speaker judgements, we see that intermediate-to-advanced stage Lners do not necessarily align with changes occurring for native speakers in the Ln. Future studies may aim to further disentangle these factors by specifically assessing those taught only in the classroom (e.g., L1 Polish Ln Norwegian learners in Poland) as opposed to naturalistic learners in Norway.
Data availability statement
The data that support the findings of this study are openly available in OSF at https://osf.io/f5gt6/?view_only=d1c1dfbb64364412900a8ff3a291325a
Acknowledgements
The research leading to these results has received funding from the Norwegian Financial Mechanism 2014–2021 project number 2019/34/H/HS2/00495. Open Access publication of this article was funded through a Transformative Agreement with UiT The Arctic University of Norway. We thank Dr Kamil Kaźmierski for his assistance with cleaning the dataset.
Competing interests
The author(s) declare none.
Appendices
Appendix 1: Post hoc pairwise results, Model 1
Model 1 is an ordinal model used to predict the response value against the L1 and the condition, with participant as a random effect.
cond = grammatical_article:
cond = grammatical_poss_3rd:
cond = grammatical_poss_r:
cond = ungrammatical_article:
cond = ungrammatical_poss_3rd:
cond = ungrammatical_poss_r:
p-value adjustment: Tukey method for comparing a family of 3 estimates.
Appendix 2: Post hoc pairwise results, Model 2
Model 2 is an ordinal model predicting the response value against L1, condition and Norwegian proficiency, with participant as a random effect.
$emtrends.
L1 = Engelsk:
L1 = Polsk:
$contrasts
L1 = Engelsk:
L1 = Polsk:
p-value adjustment: Tukey method for comparing a family of 6 estimates.
Open access
