3.1 Hypotheses regarding transfer in third language acquisition

Hypotheses regarding language interaction or transfer in L3 vary greatly depending on the language, the linguistic domain and the theoretical approach within which the empirical research is conducted. While some studies suggest that the L1 is a default source of transfer to the L3 (Lozano 2003; Na Ranong & Leung 2009; Hermas 2010; 2015), others indicate that L2 has a privileged status (Bardel & Falk 2007; Bardel & Sanchez 2017). Both these models, while having distinct hypotheses, are based on the biological status of L1 vs. L2. Specifically, the L1 Status Model considers that the native language has a privileged status as a source of transfer, while the L2 Status Model is grounded in the difference between the memory sources: L2, as L3, is sustained by declarative memory and learning is similarly based on explicit knowledge. The profile of our participants presents a challenge for these models: while Romanian is the native language learned first from a chronological point of view, it is also the heritage language, used mostly at home, typically with a restricted number of family members, and with scarce school support. Moreover, English, even though chronologically learned second, as the majority language, it is the language used in the widest settings, with diverse input sources. Our participants often identify English as their stronger language and our production measures detailed in section 4.3 confirm English to be the dominant language. As such, English could be considered as another L1.

One approach that considers both L1 and L2 roles in transfer is the Cumulative Enhancement Model which predicts that both L1 and L2 can transfer to L3, but only when this transfer is facilitative (Flynn et al. 2004; Berkes & Flynn 2012). Two current models advance the idea of transfer from either background language: the Typological Primacy Model (Rothman 2013; 2015) and the Linguistic Proximity Model (Westergaard et al. 2017). According to the Typological Primacy Model, elaborated for the initial stages of L3 acquisition, typological similarity is based on overall perceived proximity between the languages’ lexicons, phonology, morphology and (surface) syntactic structures. Typological similarity can be the source of both facilitative and non-facilitative influence. In the Linguistic Proximity Model (Westergaard et al. 2017), the predictive factor for transfer (both facilitative and non-facilitative) is structural similarity (or proximity) between properties and aspects in languages, regardless of whether the languages are overall typologically close or not. We are going to elaborate below on this latter model since it pertains to more advanced stages of acquisition, while the Typological Primacy Model aims mainly at explaining transfer in the initial stages of L3 acquisition.

The two main features of the Linguistic Proximity Model are the following: cross-linguistic influence can be facilitative and non-facilitative and it is based on certain similarities between the L3 and either of the previously learned languages, irrespective of their order of acquisition. Structural similarity refers to the similarity between abstract linguistic properties as opposed to surface word order and it can obtain between languages which are either typologically similar or typologically distant. Facilitative influence is the result of strict structural similarity between computations. For example, Westergaard et al. (2017) show that in the acquisition of L3 English, the previously acquired languages, Norwegian and Russian, are different with respect to the V2 property in declarative sentences. Russian, like English, does not have the V2 property, while Norwegian does. The V2 order is assumed to result from verb movement across the adverb to the second position. Comparing L2 English (L1 Russian), L2 English (L1 Norwegian) and L3 English (2L1 Russian and Norwegian), the results show similar and more target-like performance for L2 and L3 participants who had Russian as one of their previously acquired languages. This result is interpreted as facilitative influence based on the strict structural similarity between English and Russian with respect to the V2 property. However, as far as general typological similarity is concerned, these two languages, English and Russian, are further apart than English and Norwegian.

Studies on non-facilitative influence in bilingualism in general and in L3 acquisition in particular are numerous, and attest that this type of transfer is a robust phenomenon (Rothman et al. 2019; Puig-Mayenco et al. 2020). According to Westergaard et al. (2017), non-facilitative influence happens “when learners misanalyse L3 input (and/or have not had sufficient L3 input), and mistakenly assume that a property is shared between the L3 and either or both of the previously acquired languages” (Westergaard et al. 2017: 671). How does the mistaken assumption of shared properties between two languages relate to the structural similarity approach to transfer? Westergaard et al. (2017) do not explicitly discuss this, but they give the example of non-facilitative influence from Norwegian (a V2 language) to English (a non-V2 language), in the form of transfer of verb movement from Norwegian (a V2 language) to L2 English. While contemporary English is considered a non-V2 language, “earlier versions of English had certain V2 properties” and this order “survives as a rule inverting subjects and auxiliaries in questions” (Roberts 1996; Westergaard 2007). However, in declarative clauses, English does not have V2 and therefore the order with respect to adverbs is adverb-verb. Influence of Norwegian to English is predicted to be both facilitative (facilitating subject-auxiliary inversion) and non-facilitative (resulting in the ungrammatical verb raising past the adverb). This amounts to saying that learners are predicted to overgeneralize the V2 property of Norwegian even in declarative clauses, a context where this is ungrammatical in English, and the influence is predicted to happen from a more comprehensive system to a more restricted one. As we will see in Sections 3.2.2 and 3.2.3, French as L3 presents important structural overlap (similarity) with both Romanian and English in the domain of verb movement and discourse fronting, the properties that are directly relevant for adverb placement (linearization).

3.2 The syntax of adverbs across languages

This section presents the parametric settings for verb movement and discourse fronting in the adult grammars of the standard varieties of the three languages. The participants of this study were most likely exposed to such varieties in the school environment for English and French. Romanian displays no such syntactic variations between standard and regional varieties.

3.3 Research questions

This paper focuses on the L3 acquisition of modal and aspectual adverbs by school-age bilingual children (Romanian and English) who learn French as an L3 in school immersion programs. The research questions are as follows:

• (i)   How is adverb placement learned in a French L3 setting?

• (ii)  What is the contribution of the background languages, Romanian and English, to the acquisition of French L3? Specifically, do we see non-convergent word order involving adverbs and, if yes, does transfer play a role in these? If so, what is the source of transfer?

The assessment of adverb placement takes into account parametric settings, namely verb movement and the possibilities for dislocations under discourse triggers (when the adverbs do not surface in their first merge position). The objectives are to determine if there is transfer and if there is, the source, the direction and the type of transfer, as well as to identify the theoretical L3 model that may most adequately account for the data.

As the three languages abide by Cinque’s hierarchy in (1), we expect the hierarchy to be well acquired by the children in all three languages. With respect to French L3, if non-convergent linearization of adverbs occurs, it should be the result of differences in parametric settings concerning verb movement and/or discourse fronting. These parameters display opposite settings in Romanian and English, the background languages. In the domain of verb movement, influence from Romanian would mostly show as facilitative influence on finite verb movement in French (correct SVAdv order), and possibly non-facilitative influence on non-finite verb movement (incorrect SVAdv order). Influence from English will predict finite verbs in-situ (incorrect SAdvV order). In the domain of discourse fronting, influence from either Romanian or English could result in adverbs being placed in various positions within the sentence. The influence of one or the other language has to be determined on a case-by-case basis. However, note that Romanian is the language with the most comprehensive setting, with routine dislocations, and, according to the definition of transfer under structural similarity, the source of non-facilitative influence in French L3.

5.1 Distribution of adverbs

First, Table 4 presents all the adverb occurrences by child and by language and their totals. This table counted all the adverbial occurrences (“Total Adverbs”), as well as those that figure in Cinque’s hierarchy (“Total Target”).

The results show that the highest overall adverb production is in Romanian, while the lowest is in French. Adverb occurrences are also more uniform across children in Romanian (except for the child L) than for English and French, where adverb production is variable, with some children producing very few adverbs (as low as one adverb), and others producing many adverbial occurrences. The child L produces a very low number of adverbs across all languages.

On the other hand, Table 5 quantifies only the adverbs of Cinque’s hierarchy, as distributed per child: the modal and aspectual adverbs that merge in the vP (i.e., from Voice to Asp_{completive II}) and the TP (i.e., from T_past to Asp_{completive I}) fields. In this group, the lowest production is also in French. Overall, the distribution of these adverbs is variable across children in every language, with French more uniformly on a low frequency.

Regarding the types of adverbs, English and Romanian texts display a similar number while in French, the children produce half the number of types. Tables 13 and 14 in Appendix 2 show the types produced in each language along with the adverbial classification. A noticeable difference concerns the class distribution in the three languages; most English adverbs belong to the Mood class (e.g., actually, probably, really), whereas most Romanian and French adverbs belong to Mode and Aspect classes (TP/vP field). For example, in the English narratives, there are 17 occurrences for really and ten for probably, but only four for again and two for still. There is no evidence for a transfer of these Mood adverbs into Romanian or French, where the use of the English equivalents is virtually absent.

5.2 Convergent adverb placement in Romanian and English

For all three languages, there are no deviations in the semantic properties of individual adverbs (e.g., no over or under extensions). However, when it comes to the syntax of adverbs, Romanian and English narratives display convergent use, whereas French narratives do not always do so. Examples of target-like adverb placement are provided in (7) for Romanian and in (8) for English (i.e., adverbs that belong to Cinque’s hierarchy in the TP and vP areas).

In (7) and (8) the same children produce sentences with adverbs in Romanian and English. These sentences show differences in the acquisition of the English verb morphology, for all three children, displaying inflectional errors for present and past and present progressive, which contrasts with adult-like production of indicative and subjunctive verb forms in Romanian. However, adverb use and placement are correct in each language.

5.3 Non-convergent adverb placement in L3 French

5.4 Summary of results

Our results point out that French L3 provides examples of non-convergent adverb use (unlike Romanian and English), with the following characteristics:

• a)  Non convergent word orders, where the adverb surfaces in clause-final position, or in preverbal position, or in clause-initial position;

• b)  Verb movement is well established but there is a tendency of non-convergent adverb placement in non-finite contexts;

• c)  Children are capable of placing adverbs, at the same time, in a target-like and non-convergent manners, and for some children, adverb production, although limited, is entirely target-like;

• d)  Non-target word orders tend to cluster in younger participants who also have fewer years of experience in French;

• e)  Non-target word orders seem linked to a specific lexical item, i.e. the adverb encore.

Although the findings presented are limited from a quantitative point of view, they align well with previous results. Word order perturbations related to specific adverbial lexical items have been noted in studies of production of adverbs in work such as Herschensohn (1998) and Hancock and Sanell (2009). In our corpus, the adverb encore is the one most used in a non-convergent manner.

Variability in adverb placement – participants who use convergent and non-convergent word orders with the same adverb - has been also attested in previous studies, and taken to indicate an intermediate stage of acquisition: the acquisition of the relevant parameter (in this case, verb raising in French) while the background language parameter (a non-raising language such as English) is still in use (Herschensohn 1998; Leung 2006). In our corpus, there seems to be a tendency for non-convergent use to appear in the non-finite contexts. While we are not aware of similar results for L2 or L3 French, the same result was found in L2 Spanish by Ayoun (2005); the participants performed at chance level in the non-finite condition with adverb placement. In contrast, the verb with respect to negation was relatively well placed, however, negation and adverb did not occur in the same sentence. The author speculates that verb raising past the adverbs in non-finite contexts might not have been well acquired given the rarity of these structures in the input.

Preference for adverb placement in clause-final position has been noted in other production corpora, and while sometimes this was attributed to influence from the previous language, at other times it was hard to find an explanation in terms of cross-linguistic influence. For instance, Hancock and Sanell (2009) speculated that preference for the clause-final position in French L2 found especially with the adverb aussi could be the result of superficial influence from L1 Swedish. Similarly, Herschensohn (1998) noted the clause-final position of the adverb mal in French L2 (following the use of well in English L1) but also of the adverb souvent (the corresponding adverb often cannot appear in final position in English), concluding that not all non-target adverbial positions are the result of cross-linguistic influence. The preverbal placement (SAdvV/T) in our corpus is also quite prevalent. The position of adverbs in front of inflected verbs, in language combinations where one of the prior languages is English, has been taken to indicate lack of verb raising (Herschensohn 1998; 2009; Ayoun 2005; Leung 2006; Stadt et al. 2016).

While alignment of our findings with those from previous research validates our small data points, we diverge from the interpretation of these findings in previous literature. More specifically, we show, in the following section, that word order perturbations are not the result of verb movement, or lack thereof, but can be better explained in terms of various dislocations.

6.1 Discourse fronting in French L3

This section discusses the syntactic patterns that underlie the erroneous linearizations listed in the previous section. The gist is that the main trigger for these errors is the way adverbs undergo left or right dislocation, while the setting for the verb movement parameter is orthogonal to the issue.

In our study verb movement does not seem to be a trigger for linearization errors. Although the non-convergence percentages of Table 8 indicate a higher incidence of errors in infinitive clauses and in the presence of complex tenses, when we look at the respective data, we notice that verb movement follows the rules, whereas the placement of adverbs does not. What is the parametric setting responsible for the deviations?

The erroneous clause-final linearization of the adverb (see all examples in Table 12 in Appendix 1) may theoretically arise from two structures: (i) verb movement past the adverb in situ; or (ii) right dislocation of the adverb, meaning that the adverb remains in situ, whereas the vP fronts (as Remnant Movement). The correct placement of pas in relation to the verb in these contexts (especially with infinitives, as in (22a) — also see previous sections) eliminates pattern (i). We are thus left with pattern (ii). Evidence for pattern (ii) comes from sentences where either the repetitive or the continuative encore is preceded by direct objects and/or prepositional phrases, as in (22a, b).

In Cinque’s hierarchy, both Asp_repetitiveI and Asp_continuative are in the TP field (see 2b). On the other hand, direct objects and adjunct prepositional phrases are base generated within vP. For these constituents to precede the adverb in situ, as in (22), Remnant Movement (vP movement after being vacated by one or more items) must be at work.¹¹ There is no structure lower than vP, so we cannot assume that the child’s grammar maps the repetitive or continuative in a slot that is hierarchically lower than V or sister to V – the adverb must be higher, in situ.

Right dislocation along this pattern is productive in all three languages under consideration, although the adverbs that occur in such contexts may be different. For example, English routinely allows it with again but not with still. French also allows for the right dislocation of some adverbs (e.g. brièvement ‘shortly’ base generated in Asp_durative in (2b))¹² but not of encore. Romanian shows no restrictions for the right dislocation of any class of adverbs, except for those that have a proclitic status. For example, the Romanian clause in (23), produced in our corpus, shows încă in final position, and the interpretation can be either continuative or repetitive, according to the context – in this particular context it is continuative.

Since the syntactic rule is present in all three languages, no parametric transfer applies. The problem concerns the specific lexical items that undergo right dislocation, and in this respect the children work with the most comprehensive pattern, which happens to apply in Romanian; it is an example of the generalized system of right dislocation in Romanian applied in French.

Clause-initial linearization of adverbs may also be analyzed as transfer of the more generalized parametric pattern. All three languages allow for adverb fronting; however, they differ with respect to the class of adverbs compatible with this operation. Thus, in the examples xv. and xvi. of Table 12, in the Appendix 1, the errors consist in the choice and pairing of tokens, not in the illegality of fronting per se: i.e., continuative encore is not a frontable item in French L1, whereas maintenant is, but not above the transition formula aussi (too high in the CP). It also happens that Romanian is the least discriminatory in this respect, and allows for the counterparts of xv. and xvi., see example (24), with the Romanian equivalent of continuative encore:

A more challenging linearization noted in section 5.3.3 has the adverb in between the subject and the verb, as shown again in (25).

In Cinque’s hierarchy (2b), the continuative encore is base generated in the lower part of the TP field, whereas maintenant is in a (non-argumental) Spec of TP. Thus, the word order in (25) may arise from two underlying syntactic patterns: (i) the verb remains in vP, while the subject is in the (argumental) Spec,TP, so naturally, the adverb in the TP field will be in-between; or (ii) the verb is in T and the adverb is fronted to the CP field. Notably, other adverbs also occur in this position (juste, peut-être).

Pattern (i) cannot be considered because all these contexts necessarily contain finite verbs (or else no subjects could be present), and simple finite verb movement to T is well established in these children’s grammar (see previous sections). For example, the verb is higher than pas in (25a) while the adverb is still preverbal. Crucially, the adverb precedes the auxiliary in (25b), where the verb is in complex tense and should have displayed the adverb in situ between the auxiliary and the verb (as it would in the English counterpart). Erroneous movement of the past participle cannot be at work because that would have left the adverb in post-verbal position. Since these observations eliminate pattern (i), the alternative is pattern (ii): the adverb is fronted from its in situ position to the CP field, under discourse triggers.

Adverb fronting is possible in all three languages, but it would leave the subject lower in either French L1 (e.g., maintenant il va…) or English (e.g., now he goes…). However, in (25), the subject is fronted higher than, and concurrently, with, the adverb. This is typical of Romanian grammar, where fronting does not involve quantificational chains (see example (4)), so various constituents can concurrently undergo fronting. Romanian examples of similar linearization are found in the corpus; see (26).

French also allows for non-quantificational chains upon subject fronting (through CLLD), but does not extend it to concurrent free fronting of other constituents (e.g., *Jean la maison il ne la veut pas/*La maison maintenant il ne la veut plus).

Hence, although CLLD applies in both Romanian and French, the pairing of CLLD with the non-quantificational fronting of any type of constituents, including AdvP, applies only in Romanian, and it is visibly extended to French L3. Further indications of transfer from Romanian come from the subject treatment: in (25a), CLLD fails to apply to the fronted subject (i.e., there is no resumptive clitic), which is the way subjects are generally dislocated in Romanian (no overt resumptive clitic in (26)); in (25b), the subject pronoun undergoes dislocation, although it is a proclitic, which is the way subject pronouns are fronted in Romanian, since they are not clitics. This indicates morphological transfer (lack of subject clitics, non-clitic treatment of il) in addition to the partial transfer of the fronting/discourse parameter from Romanian to French L3.

Finally, our results also show linearization errors where the adverb surfaces between two subject forms: a full-fledged DP and the corresponding subject clitic, as in (27).

This is a variant of the partial fronting/discourse parametric transfer discussed above, with the exclusion of the additional morphological transfer of subject treatment: CLLD applies as in French L1, but it is concurrent with adverb fronting, which is not a property of French L1. This order occurs routinely in Romanian (with dislocations for both the subject and the adverb and optional resumption through a non-clitic pronoun), but yields ungrammaticality in English.

Table 11 sums up all the points of possible transfer discussed in sections 5 and 6.

Comparing French L3 and French L1, there are five non-convergent areas. All five areas are convergent in Romanian L1, while only three are partially (i.e. +/-) convergent in English. Hence, we conclude that the influence of Romanian grammar is stronger on French L3 than the English grammar. To these observations, we also add the remarks on the lexical information, that is, children could produce quantitatively more adverbs in Romanian, and the type of adverbs used in the French L3 texts were also the counterparts of those in the Romanian texts rather than in the English texts (e.g., no use of most English Mood adverbs).

6.2 Transfer to French L3

In light of the analysis in the previous section, the proposal is that there is evidence of transfer from Romanian L1 to French L3, and the transfer involves the discourse parameter. More precisely, both Romanian and French display a positive setting for the discourse parameter with non-quantificational chains, but the extent of its applicability differs: the setting is more comprehensive in Romanian, for both left and right dislocations (no restrictions on the type and number of fronted constituents, no lexical restrictions) but more limited in French (e.g., only arguments front, one at a time, some adverbs do not undergo fronting). Hence, the cases of Romanian transfer to French fall under the proposed concept of structural similarity, whereby Romanian, which has more generalized mechanisms in the dislocation system, influences French which presents more restricted options in the same area. This results in non-facilitative influence insofar as French L3 learners do not show the lexical and constituent restrictions on dislocation as per the target grammar. The phenomena observed can be further characterized as falling under the notion of syntactic complexity or interface syntax (Jakubowicz 2000 and Sorace 2005 respectively) – features that belong to the interface between syntax and other domains such as the lexicon and discourse (discourse, in this case) are more difficult to acquire and are more prone to cross-linguistic influence from a previously acquired language. Moreover, a syntactic operation is less complex if it applies obligatorily in every sentence and more complex if there are restrictions on its application. The influence applies from the background language into the language being learned when the background language presents the more economical system (less complexity).

We note that the data presented show limited scope of transfer, as children are also capable of using the correct word order, and adverbial non-convergent uses are scarce. This is something that we observe in previous studies as well, as learners are usually capable of using the target construction to a certain extent. Until fully acquired (if ever), the grammar of adverb placement in French displays what Sorace (2005: 7) calls “residual optionality”. We provided arguments that this optionality is due to transfer and that it happens at the level of the syntactic representation. However, we also note that the data do not exclude the possibility that language processing is the underlying cause (and no data can exclude this unless an experiment tapping processing is used).

In addition, Romanian is the learners’ L1 and it is a Romance/Latin language like French. As such, the proposed transfer from Romanian to French also satisfies the predictions of the Typological proximity model (Rothman 2013; 2015) as well as those made by the L1 factor model (Hermas 2010; 2014) if indeed we consider Romanian as the sole L1 (with English the L2). Therefore, the three accounts are confounded in this case but the data do fall under the more precise analysis of structural similarity, irrespective of the possible and plausible role of the typological proximity and/or the L1 factor. In this respect, to the best of our knowledge, the present study is the first to indicate that generalized discourse triggered fronting (resulting in movement of adjuncts as well as arguments) can also be transferred in the domain of the acquisition of adverb placement. Previous work on non-convergent linearization of adverbs concentrated by and large on transfer from the domain of verb movement.

Studying the acquisition of adverb placement presents some methodological issues that should be observed in future research. The present study enhances previous findings in the domain of adverb placement by indicating that, when a previous language has mechanisms underlying what is called “free word order”, all these mechanisms can transfer to a newly acquired language in the particular domain of adverb placement. This points to the need for studies that test domains larger than the vP area where the stimuli only contain the verb, its arguments, and one contrast in adverb placement, as in the highly preferred judgment/preference tasks used in this domain (Léa regarde souvent la télévision. / Léa souvent regarde la télévision. ‘Léa watches often TV.’/‘Lea often watches TV.’; cf. Hermas 2010: 361). While this simple contrast is necessary in order to minimize parsing complexity in preference/judgment tasks (McDaniel & Smith Cairns 1998), and crucially important in determining the verb movement setting, considering tasks with more clausal material (for instance, including negation and adverbs at the same time) or choosing contrasts that reveal other types of movement is also necessary.

Moreover, previous studies on (ad)verb placement have mentioned that accuracy results can vary according to the type of task (see Prévost 2009: 110 and references therein). For example, Leung (2002) reported higher accuracy on the production task (written sentence completion) than on the preference task (PT). She interpreted this difference as a task effect of the PT, which included an option “both right”, or as a possible weakness of the PT (similar to judgment tasks) in tapping into learner’s underlying competence. While we cannot directly compare our results to those obtained through other type of tasks, they do align with the generally good performance found by the previous production studies. It is also possible that children’s good performance on adverb placement can be attributed, in part, to the nature of the production task, where the participant is free to choose not to produce a particular syntactic construction or lexical item (as opposed to an elicited production task, for instance). Lastly, production data have been found to be more compatible with studies reporting L2 or hybrid (L1 and L2) influence and might, for example, “capture processing-based influence at a more superficial level” (Puig-Mayenco et al. 2020: 49) due to considerations such as attentional resources. It is therefore advisable to treat our results (as well as any results from production studies) with caution until a comprehension study confirms them.

In addition to pinpointing the source language of the transfer, we are interested in the identification of structural requirements for transfer at more advanced stages of acquisition. In a more general context, our study supports the approach of UG-constrained sequential acquisition and we note the importance of discussing not only what seems difficult in acquisition but also the areas that seem to conform to the predicted patterns.