Abstract
Aims and Objectives:
This paper investigates the word order and adjectival agreement patterns in French-Dutch codeswitched Determiner Phrases (DPs). It examines the predictions made by two theoretical points of view: the approach by MacSwan (2009) within the Minimalist Program (MP) and the Matrix Language Framework (MLF) (Myers-Scotton & Jake, 2009).
Methodology:
The predictions of these frameworks are compared to data gathered in a grammaticality judgment task. In total, 120 codeswitched sentences were presented aurally to participants, who were asked to rate the sentences on a three-point scale.
Originality:
While some previous work on word order within codeswitched DP’s exists, this paper is the first study investigating the adjectival agreement patterns in codeswitching.
Data and Analysis:
Statistical analysis of the data showed that the MacSwan approach is a better predictor for the grammaticality judgments, as sentences predicted to be grammatical by the MP were rated higher than sentences predicted to be ungrammatical by the same model. This difference was statistically significant. There was no significant difference in rating for the predictions of the MLF.
Conclusions:
The results of the judgment task in combination with the results of previous research on codeswitching highlight the importance of a combination of data from both naturalistic and experimental settings.
Implications:
The predictions of the Minimalist approach have the upper hand over the predictions of the MLF. However, it remains is important to integrate results from other experimental methodologies, such as naturalistic data and results from psycholinguistic and neurolinguistic studies.
Introduction
In this paper, the term codeswitching is used to refer to the use of multiple languages within one conversation. Although early formal approaches argued switching is only allowed where the structures of both languages are equivalent (Poplack, 1980), more recent approaches take this to be untrue. The study of conflict sites – areas where the properties of the two participating languages differ – has become an important topic in codeswitching research (see Berg Grimstad, Lohndal, & Åfarli, 2014; Chan, 2008; González-Vilbazo & López, 2012; Hulk & Van der Linden, 1996; Nishimura, 1997; among others).
Word order is a relatively well-studied conflict site. Word order in the determiner phrase (DP) has been investigated and discussed by several authors (Cantone & MacSwan, 2009; Jake, Myers-Scotton, & Gross, 2002; Parafita Couto, Deuchar, & Fusser, 2015a; Parafita Couto, Munarriz, Epelde, Deuchar, & Oyharçabal 2015b. However, no systematic research has focussed on the linguistic factors that determine adjectival agreement in codeswitched DPs. To my knowledge, this is the first study to do so.
In the case of French and Dutch, the DP is a locus of two conflict sites: adjectival word order and gender. The next paragraph briefly outlines the relevant differences. For a more elaborate description, I refer to Devos, De Muyninck, and Van Herrewege (1991).
While Dutch only has pre-nominal adjectives, French has predominantly post-nominal adjectives. French, like most Romance languages, has a system of two genders: masculine and feminine. Gender agreement in French shows up on adjectives, articles, possessive pronouns and past participles. This is summarised in (1). In this example and all the following ones, italics represent French language data, while small capitals represent
(1) French Masculine Feminine Plural Indefinite article un une des Definite article le la les Adjective -ø -e -(e)s
Although in writing the feminine adjectival agreement marker is an -e, it never manifests as an [e]/[ε]/[ə]. Usually it makes an underlying consonant overt, though there can also be no overt effect. This is shown in (2).
(2) brown annoying pretty Masculine brun [brʏ̃] ambetant [ãbεtã] joli [ʒoli] Feminine brune [brʏn] ambetante [ãbεtãt] jolie [ʒoli]
Brabant Dutch, the variety of Dutch described in this paper, has a three-way gender system. The three genders of Brabant Dutch are masculine, feminine and neuter (Dekeyser, 1980).
Though the singular has three gender distinctions, the plural has none. The table in (3) gives an overview of the agreement of the articles of Brabant Dutch. Note that the adjectival agreement is not only sensitive to gender, but also to definiteness (Dekeyser, 1980, p. 107). The bracketed (n) is sensitive to phonological restrictions, which are not relevant here.
(3) Brabant Dutch Masculine Feminine Neuter Plural Indefinite article Definite article Adjective Definite Indefinite
As is shown in the table above, there is syncretism in the adjectival forms for all forms except the masculine (in some phonological contexts) and the neuter singular indefinite.
Taking into account only word order and adjectival form, every French-Dutch codeswitched DP has four possible variations, which are shown in (4).
(4) a. une pomme b. une pomme ‘the green apple’
This paper investigates which option bilingual speakers prefer. There was no investigation of each of the factors (word order and agreement) independently as the experimental design did not allow presenting these factors separately.
Literature overview
This section highlights the most relevant studies concerning gender and DP word order in codeswitching.
Gender is a quite well studied phenomenon in the field of bilingualism research. An early study by Poplack, Sankoff, and Miller (1988) on English words in French found that gender was assigned first and foremost on phonological criteria. Analogue gender (the gender of the translational equivalent) and default gender (masculine) were both found to be an additional influence on gender assignment. Parafita Couto et al. (2015b) found that feminine is the default gender for Basque nouns in Spanish, while Jake et al. (2002) found masculine to be the default for English loanwords in Spanish. Liceras, Fernández Fuertes, Perales, Pérez-Tattam, and Spradlin (2008) show that the assignment of gender to English loans in Spanish differs for different types of bilinguals. Simultaneous bilinguals prefer a default strategy, while L2 learners tend to assign analogue gender. Finally, Cantone and Müller (2008) showed that children simultaneously acquiring German and Italian, both of which have gender, allow for gender-agreement between the article of one language and the noun in the other.
This transfer of a gender feature from one language to the other was also observed by Treffers-Daller in her 1993 dissertation on Dutch-French codeswitching in Brussels. She found that French loans in Brussels Dutch have a tendency to keep their original gender. This stands in contrast to Standard Dutch, where loans tend to be assigned the neuter gender.
Treffers-Daller’s (1993) main findings concerning adjectives were that it is more common for Dutch adjectives to be embedded in a French sentence than vice versa. Furthermore, when French adjectives are embedded in a Dutch sentence, they tend to remain uninflected.
Word order within the DP has been studied for different language pairs. Cantone and MacSwan (2009) found that the adjective was the determining factor for DP internal word order for German-Italian codeswitchers. Jake et al. (2002) on the other hand studied Spanish-English bilinguals and concluded that the Matrix Language determined the word order in the DP. This finding was corroborated by Parafita Couto et al. (2015a) for Welsh-English, though they found the influence of the Matrix Language to be very slight.
There is to the best of my knowledge no work on codeswitching specifically addressing adjectival agreement.
Theoretical approaches and their predictions
While the phenomenon of codeswitching is studied from many different formal approaches, currently two are prominent: the Generative approach (MacSwan, 1999) and the Matrix Language Framework (Myers-Scotton, 1993), henceforth MLF. These theories went into a debate in a special issue of Bilingualism: Language and Cognition (Jake et al., 2002; MacSwan, 2005a; 2005b; Myers-Scotton, 2002). This section will highlight the relevant and most important elements of both frameworks.
The Matrix Language Framework
The MLF, developed by Carol Myers-Scotton in the early 90s, is a way to account for both the processing and production of bilingual speech. Several important principles are at work in this model. The first one also applies to monolingual speech: the Uniform Structure Principle (Myers-Scotton, 2006, p. 243).
(5) The Uniform Structure Principle A given constituent type in any language has a uniform abstract structure and the requirements of well-formedness for this type must be observed whenever the constituent appears. In bilingual speech, the structures of the Matrix Language are always preferred, but some Embedded structures […] are allowed if Matrix Language clause structure is observed.
A second fundamental principle is the asymmetric relationship between the languages involved in codeswitching. Often called the asymmetry principle, it demands that only one of the languages involved in the sentence provides its morphosyntactic frame. This language is referred to as the Matrix Language (ML), while the other language(s) involved is called the Embedded Language(s) (EL). Usually, it is possible to determine the Matrix Language for a whole corpus at once. Exceptionally, the ML can change from clause to clause within a corpus (Myers-Scotton & Jake, 2009, p. 337f). Two further principles are central: the Morpheme Order Principle and the System Morpheme Principle (Myers-Scotton, 2006, p. 44).
(6) The Morpheme Order Principle (MOP) A given constituent type in any language has a uniform abstract structure and the requirements of well-formedness for this type must be observed whenever the constituent appears. In bilingual speech, the structures of the Matrix Language are always preferred, but some Embedded structures […] are allowed if Matrix Language clause structure is observed. (7) The System Morpheme Principle (SMP) In [… mixed] constituents, all system morphemes which have grammatical relations external to their head consituents (which participate in the sentences thematic role grid) will come from the Matrix Language.
Exceptions to the SMP are Embedded Language (EL) islands (phrases from other varieties participating in the clause), which are allowed if they meet EL well-formedness conditions, as well as those ML conditions applying to the clause as a whole, such as phrase placement (Myers-Scotton & Jake, 2009).
Note that the SMP does not apply to all system morphemes, but rather to a subset called late outsider system morphemes (LOSM). These morphemes depend on information that is outside the word on which the morpheme attaches. This information can be retrieved from another word in the utterance or from the discourse. While for many morphemes LOSM-status needs to be determined language by language, verbal agreement is always considered to be a LOSM. This is illustrated in (8) for Shaba Swahili-French codeswitching (Myers-Scotton & Jake, 2009, p. 347). The ML, Shaba Swahili, provides the verbal agreement in the EL verb appartenir.
(8) Donc so c9.dem riches c9.all C9-non.past-obj-belong us c2-child c9-his ‘So, all these riches, it belongs to us, his children.’
The MOP (6) and SMP (7) together make predictions regarding the adjectival word order and agreement in a mixed language DP. The MOP predicts that in mixed DP constituents, the word order is determined by the Matrix Language. Since the ML supplies the late system morphemes and verbal agreement is an unambiguous late system morpheme, we can conclude that DP word order follows the word order of the language of the inflected verb.
With regards to adjectival agreement, the System Morpheme Principle (7) plays an important role. Recall that late outsider system morphemes must come from the Matrix Language. We need to determine whether or not adjectival agreement morphemes are late system morphemes. The definition of these morphemes is below (Myers-Scotton & Jake, 2009, p. 75):
‘outsider’ late system morphemes depend for their form on information outside their maximal projection. That is, they are co-indexed with forms outside the head of their maximal projections.
This definition can be applied to adjectival agreement, which is illustrated in Figure 1. 1

Indexation of adjectival agreement in Dutch and French.
Adjectives in Dutch are indexed with the determiner and the noun, as they are sensitive to definiteness and gender. Consequently, they look to both determiner and noun for information about their form. In French the adjectives are only co-indexed with the noun, as they do not need definiteness information from the determiner for agreement.
However, in both languages, the adjective looks outside its maximal projection (the AP) for information about its form. Consequently, the MLF would predict that any adjectival agreement should come from the Matrix Language, both when the adjective is French as when it is Dutch. As for the agreement form, the MLF predicts any conflict will be resolved in favour of the Matrix Language.
The Minimalist Program
The second framework considered in this article is the approach taken by MacSwan within the Minimalist Program (MP). The driving force behind the Minimalist Program is the elimination of all mechanisms not strictly necessary from a conceptual point of view. For proponents of this Program, it is essential not to posit a grammar specific to codeswitching. Making predictions about codeswitching data can be done by looking closely to the requirements of the monolingual grammars. Approaches of this type are known as constraint-free approaches.
Most researchers that adhere to the Minimalist Program hold that any language consists of two components: the computational system and the lexicon. The former is universal and invariant, while the latter is language-specific and contains the parameters responsible for the wide variation attested within the world’s languages (Chomsky, 1991). All language-specific information is found in the lexicon, which allows for a different view of bilingualism, in which the grammars of the two languages are less compartmentalised (MacSwan, 1999).
When a linguistic structure, or derivation, is built, the computational system performs several operations on lexical items. First, the operation Select chooses the items from the lexicon that will participate in the derivation. This subset of items is used to build the derivation is called the Numeration (or lexical array). In codeswitched speech, Select draws the elements that go into the Numeration from two lexicons.
The operation Merge then takes items from this Numeration and puts them together to form new hierarchical syntactic objects. To these objects, the operation Move can apply to form new structures.
Movement is driven by the process of feature valuation, a process that will be exemplified below. Two types of features drive movement: strong features drive overt movement, which is realised phonetically, while weak features drive covert movement, which is not. Both types of movement, however, are needed for interpretation (Chomsky, 1995). Whether or not a feature is strong or weak is language-specific. A difference in word order is thus attributed to a difference in feature strength, causing visible movement in one language, and invisible movement in another.
The predictions with regard to word order in codeswitched DPs have been discussed in Cantone and MacSwan (2009). Recall that movement during the derivation is caused by (strong) features. In the DP it is the Agreement Projection (AgrP) that is equipped with a movement-triggering feature. If this feature is strong, it triggers movement of the noun to a higher position. This process is illustrated in Figure 2. The t in the diagram indicates a trace left by the movement of the noun.

Upward movement of the noun triggered by a strong feature.
The word-order difference between French and Dutch is due to the movement-triggering feature being strong in French, and weak in Dutch. In codeswitched DPs, the prediction is that the language of the adjective determines the word order.
For agreement, it is important to consider what combinations would cause a derivation to crash, reflecting ungrammaticality. The first issue is whether or not an adjective can agree with a noun in a different language. MacSwan (1999) argues that it can, because gender features are inherent in the noun, and are therefore available for agreement with the adjective as soon as the noun is selected from the lexicon. There is also further evidence from the child-language literature that this is not problematic (Cantone & Müller, 2008).
The next step is the agreement process. According to Chomsky (2000, 2001), the Agree mechanism works in the following way: an unvalued feature (probe) on a head looks within its c-command domain 2 for another instance of the feature (goal) with which it can agree. Once the probe finds a valued goal, this value is assigned to the probe, and the feature is deleted.
For this mechanism a problem may arise if the probe cannot accommodate the value of the goal. This will only happen when a French adjective tries to agree with a Dutch neuter noun. The adjectival gender feature needs to be valued, so it searches for a goal, which it finds in the Dutch noun. However, the value of the gender feature on the Dutch noun is neuter. The adjective cannot value its own gender feature, as it only has two options: masculine and feminine.
The predictions of these two frameworks are compared and summarised in Table 1.
Summary of the predictions.
Main divergences between the approaches 3
A significant difference in the point of views of the frameworks discussed here is the type of data that are preferred. Research within the MLF is focussed on naturalistic data found in corpora. The Minimalist approach maintains that grammaticality judgments and naturalistic data will complement each other (MacSwan, 1999). The use of grammaticality judgments is not only a significant area of disagreement for these two frameworks; it is far from uncontroversial in the wider field.
There is disagreement as to what exactly it is that grammaticality judgments measure. While the Generativist position maintains that these judgments provide a (perhaps unique) window on competence, others argues that these judgments are just a different type of performance, and are as such also influenced by confounding factors, such as metalinguistic attitudes (Schütze, 1996).
This type of data also has several advantages. They allow eliciting very rare constructions, possibly absent by chance from naturalistic corpora. They have another significant advantage in providing negative evidence (Schütze, 1996), which is absent from naturalistic corpora. For Generativists, negative evidence is crucial to construct a theory of codeswitching (MacSwan, 2014).
When consulting bilinguals for grammaticality judgments, there are some additional issues. For many codeswitching communities, mixing languages is a stigmatised phenomenon. Judgments on codeswitching may be heavily influenced by prescriptive norms. Furthermore, Sanoudaki and Thierry (2014) showed that Welsh-English bilinguals had their Welsh syntax activated, even when they processed monolingual English sentences in an all-English context. This unconscious activation may or may not influence the judgment of grammaticality of conflict sites in bilinguals.
While the MLF is open to an incorporation of the Minimalist approach into their framework (Jake et al., 2002), the Minimalist opposition remains adamant that the notion of a Matrix Language should not be used to describe codeswitching data. They object to the introduction of a language label on theoretical grounds. As a language is generated by a grammar, a language label cannot be a primitive in such a grammar (MacSwan, 2015).
Methodology
To test the prediction of the frameworks discussed above, a grammaticality judgment experiment was carried out. An elicitation task was also conducted. Although the participants involved in the task did codeswitch, they did not do so within the DP and the data of this task are omitted here.
Participants
The participants were screened to comply with four main criteria, 4 as it has been shown that bilinguals that fit these criteria give homogeneous grammaticality judgments where codeswitching data are concerned (López, 2014):
acquisition of both languages before age four;
continuation of use of both languages;
native proficiency in both languages;
both languages are spoken on a daily basis.
To determine whether the participants fit the criteria, they filled in a background and proficiency questionnaire. This questionnaire also contained questions on language use and attitudes toward codeswitching. Participants were recruited by e-mail, social media and posters posted in both Dutch and French universities and libraries in Brussels. All participants signed a consent form. Ten participants took the survey in the presence of the investigator; the remaining ones took the survey by themselves.
Stimuli
For this task, 160 sentences were developed. Of the 160 sentences, 120 were test sentences and 40 were distractors. All sentences contained codeswitching in order to maintain a ‘codeswitching mode’ over the whole session. Three types of sentences were developed.
(9) Type I 24 sentences with a DP consisting out of a determiner and a noun Type II 96 sentences with a DP consisting out of a determiner, adjective and noun Type III 40 distractors in codeswitching mode
A list of the critical Type II stimuli can be found in the appendix.
Type I: Sentences of this type were included to study the factors involved in gender assignment in codeswitches. They are omitted due to space limitations, as results were not statistically significant and not central to the research question of this paper.
Type II: Sentences of this type had the following structure:
(10) Subject Verb [DP Object ]
The target DP consisted of a noun, determiner and adjective. The article and noun of the target DP were in the same language with the adjective in the other. The article and noun were in the same language and gender congruent, in order to isolate the effect of adjectival agreement in codeswitched DPs on grammaticality judgment.
The experimental design had four variables with two values: Matrix Language, DP word order, adjectival agreement (masculine versus feminine for French, and neuter indefinite versus others 5 in Dutch) and language of the adjective. Four variables with two values yield 16 possibilities. The nouns the adjectives modify had six gender conditions (see an example of each category in (11)), bringing the total number of unique conditions to 96.
(11) Gender in Brabant Dutch Gender in French Masculine Feminine Neuter Masculine ‘bike’ hat’ heart’ Feminine ‘tooth’ ‘star’ ‘house’
As there were 96 conditions, for each of the six gender categories, 16 nouns were selected, eight Dutch and eight French nouns, which were each others’ translational equivalent. The frequencies were checked in the CELEX database (celex.mpi.nl) for Dutch nouns and the Lexique2 database (http://www.lexique.org/) for the French nouns. The mean frequencies of the categories were not statistically different.
(12) is an example with French as the ML, while (13) has Dutch as the ML. A complete list of test sentences can be found in the appendix.
(12) Il ferme [DP ‘He closes the open window.’ ‘You are buying the green apple.’
All included nouns were inanimate, to prevent biological sex influencing the gender assignment, as this has been shown to override other factors (Poplack, Pousada, & Sankoff, 1982). French nouns with word markers were avoided, since the phonological shape might influence the gender assignment. For example, the overwhelming majority of nouns in French ending in -ette are feminine. 6
Considering the difficulty in procuring participants for experiments of this type, it seemed unlikely that a large number of responses would be collected. To improve the soundness of the statistical analysis, it was decided that each condition should be presented two times to each participant.
Because of the length of the experiment, it was preferable to halve the number of conditions, rather than doubling the number of sentences. Since the Matrix Language should not influence the gender condition, half of the gender conditions were presented with Dutch as the Matrix Language, while the other half was presented with French as the Matrix Language. A noun is not expected to change its lexical gender according to the language of the inflected verb, after all.
Definiteness was not included as a condition in this experimental design, again taking the length of the experiment into consideration. The choice between the definite and indefinite articles was driven by practicalities. When the article and noun were in Dutch, the DP was definite, as the indefinite article posed a dilemma. In Brabant Dutch it inflects, while in Standard Dutch it does not. This is shown in (14).
(14) Indefinite (singular) article Masculine Feminine Neuter Brabant Dutch ne een ee(n) Standard Dutch een een een
This would force a choice between these two variants, which would introduce an additional confound. When the noun and article were French, the DP was indefinite. This was necessary since the Dutch adjectival agreement is only overt in an indefinite DP (as was shown in (3)).
This means that DPs with a French noun and determiner were always indefinite, while the DPs with a Dutch noun and determiner were always definite. As definiteness is not predicted to interfere with the other factors under investigation (besides agreement), this should not be problematic within the context of this study.
As some French adjectives occur mostly pre-nominally (Devos et al., 1991), the adjectives were selected so that they had their unmarked position after the noun. Additionally, some French adjectives have no overt gender agreement and these adjectives were avoided. For every test sentence, it was determined whether or not they were predicted to be grammatical by each of the frameworks. Illustrations are provided below.
(15) a. J’ ouvre I open the box( ‘I open the green box.’ b. Word order Agreement Overall prediction MLF ✓ ✓ ✓ MP ✓ ✓ ✓
As the word order matches the language of the inflected verb, and the language of the adjective (French), both frameworks predict the word order in (15). The language of the agreement morpheme (French) also matches the Matrix Language, as the MLF would predict. The agreement on the adjective matches the gender of the noun (feminine), so the Minimalist Program would predict this derivation not to crash.
Neither framework would predict the following sentence to be grammatical.
(16) a. Il ferme he closes the open- ‘He closes the open door.’ b. Word order Agreement Overall prediction MLF ✗ ✓ ✗ MP ✗ ✗ ✗
The Minimalist Program would predict a derivation crash, as the gender features on the noun and adjective don’t match. Additionally, the word order in the DP is not as predicted, seeing that the French adjective is expected post-nominally. The MLF would reject the word order, because the DP does not follow the Matrix Language order.
For the next sentence, the two frameworks differ in their predictions.
(17) a. Odette Odette found a. ‘Odette found a new love.’ b. Word order Agreement Overall prediction MLF ✓ ✓ ✓ MP ✓ ✗ ✗
The gender features on the noun and adjective don’t match, hence the Minimalist Program expects a derivation crash. The MLF on the other hand predicts this sentence to be grammatical, as both the adjective and the DP word order are provided by the Matrix Language. Sentences such as these are the most interesting, as the frameworks make different predictions.
Procedure
The grammaticality judgment task consisted out of an online questionnaire developed using the survey software Qualtrics. The survey consisted of the following parts, in the following order:
Start-up screen. Contained a welcome message displayed in both Dutch and French.
Background questionnaire. Available in French or Dutch. Participants were able to choose which language to take the questionnaire in.
Written instruction for the grammaticality judgment task, in mixed language text.
Audio fragments of the 160 sentences, randomised.
Proficiency test. Participants were asked to indicate the correct gender for 10 Dutch and 10 French nouns. To test their proficiency in the Brabant Dutch genders, they were asked to indicate the article which ‘sounded the best’ for five nouns. 7 They could choose between the masculine indefinite article and the feminine/neuter indefinite article.
Language attitude questionnaire. Available in French or Dutch. Participants were able to choose what language to take the questionnaire in.
For practical reasons of loading time the sentences were presented in blocks of 10, four blocks at a time. These blocks were presented randomly and the stimuli types (I, II and III) were evenly distributed over the 16 blocks. The sentences within the blocks were also randomised. Participants were asked to rate the sentences on a three point scale, represented by emoticons.
The modality in which codeswitched stimuli are presented is the topic of some controversy. Traditionally, aural stimuli are generally preferred over written stimuli, since codeswitching is considered a predominantly oral phenomenon (González-Vilbazo et al., 2012). However, recent research has shown that the modality in which stimuli were presented had no significant influence on the grammaticality judgments of native codeswitchers (Koronkiewicz & Ebert, 2014).
In the case of French and Dutch, the choice for aural stimuli was inevitable. Orthographically the agreement morphemes of the Dutch and French adjective are the same (-e), while they are not the same phonetically ([-ə] in Dutch vs a whole range of manifestations in French, as shown in (2)). Aural stimuli consequently provided a way to distinguish these morphemes, whereas written stimuli did not.
The sentences were read aloud by a regular codeswitcher and recorded. The informant was instructed to make the sentences sound as natural as possible and to be extra vigilant to pronounce the adjectival agreement endings correctly. When a sentence did not sound or feel right according to the informant, it was recorded again. These recordings were edited into short fragments and converted to mp3 format (for compatibility with Qualtrix) using the audio editing software Audacity.
Results
The survey had a total of 47 responses. These responses were filtered with the information given in the background questionnaire, resulting in a total of 15 suitable responses. This high number of disregarded responses (32) is due to participants not finishing the questionnaire (11) or learning either Dutch or French after the age of four (22). Due to technical issues, the online survey occasionally refused to play a fragment. This means that some sentences only have 14 judgments. The results of the survey were analysed with the statistical software SPSS.
Language background and attitudes
Results of the proficiency tests were very high for the genders of French and Standard Dutch. Only one participant made one mistake in the Standard Dutch genders. These high scores can be attributed to the early acquisition of both languages of suitable participants.
Attitudes towards codeswitching ranged from neutral to positive, with the exception of two participants who had a very negative attitude towards codeswitching. However, a Kruskal-Wallis test showed no significant correlation between either reported use and mean scores or between attitudes towards codeswitching and mean scores.
Grammaticality judgments
The three possible judgments were converted to a score of ‘0’ for unacceptable, a score of ‘1’ for neither acceptable, nor unacceptable and a score of ‘2’ for acceptable sentences. The overall mean score of the test sentences was quite low at 0.71. Mean scores per participant ranged from 0.13 to 1.19. The lowest mean score for a sentence was 0.07 while the highest was 1.64.
No significant difference was found in the mean rating of sentences with Dutch as the language of the inflected verb compared to sentences with French as the language of the inflected verb, nor was there any preference for a default adjectival form.
The results are summarised in Table 2. As this table shows, the difference in mean rating was minimal for MLF predictions, and did not reach statistical significance. The sentences predicted to be grammatical by the MP had a mean rating of 0.25 higher than the ones predicted to be ungrammatical. This was a statistically significant result.
Mean ratings comparing the MLF and MP predictions.
As an anonymous reviewer pointed out, there is a tendency for determiners to be provided by the Matrix Language. In French and Dutch, the determiner would also be considered a LOSM, and should consequently be provided by the ML. This was not taken into account in the experimental design, as this would predict only single noun switches. All the critical stimuli had the shape illustrated in (18) (abstracting away from the DP internal word order).
(18) Subject
The bold faced elements are considered to be LOSMs within the MLF (for both participating languages), and are to be provided by the ML according to the SMP.
Nevertheless, when one compares the mean ratings of sentences, only looking at which language provided the determiner, another correlation surfaces. This is shown in Table 3.
Mean ratings comparing sentences in which the determiner will or will not be from ML.
This result is highly statistically significant. This difference is, however, not attributable to the LOSM status of determiners. Both determiners and adjectives are classified as LOSMs. Due to the experimental design, whenever the determiner was provided by the ML, the adjective was provided by the EL. Their morphological status is the same, yet violation of the SMP by an adjective seems to count for less than violation of the SMP by a determiner.
(19) a. ‘Linnea wants the short skirt.’ b. Nous achetaient ‘We bought the white house.’
In (19), experimental stimuli of the two types presented in Table 3 are given. For clarity, the SMP violations are underlined. Stimuli of type (a) – when ML and language of the determiner coincide – were rated higher than stimuli of type (b) – when ML and language of the determiner are different – despite their both violating the SMP. While these results indicate that there is a psychological reality for the language of the ML to coincide with the language of the determiner, this experiment also seems to suggest that this may not be explained by their LOSM status.
Discussion and suggestion for further research
The results in Table 2 indicate a clear win for the predictions of the Minimalist Program. As this table shows, the sentences which the MP predicts to be grammatical have a statistically significant higher mean rating, suggesting that this research program is more empirically adequate than the MLF. However, the results shown in Table 3 provide additional support for the observation that determiner and finite verb tend to be provided by the same language in codeswitched speech, a phenomenon usually used to argue in favour of the MLF.
This study has also shown that data gathered in a grammaticality judgment task may get different results from naturalistic data. The main findings regarding adjectives by Treffers-Daller (1993) — summarised in (20) — are not supported by this study. 8
(20) a. It is more common for Dutch adjectives to be embedded in a French sentence than vice versa. b. When French adjectives are embedded in a Dutch sentence, they tend to remain uninflected.
There was no statistically significant difference in mean rating of sentences with a Dutch Matrix Language and a French adjective and French ML sentences with a Dutch adjective. Moreover, there was no significant preference for Dutch adjectives to remain uninflected when embedded in a French ML sentence. This is a significant finding, as it contradicts the notion that a ‘default’ version of an inflecting part of speech will be used in codeswitching.
The results from this study could be consolidated and refined in several ways. For example, informal post-interviews with a subset of the participants suggested that they may have rated a sentence ungrammatical for reasons other than the ones under investigation in this paper, such as prosody or semantics. This could be countered by asking participants to choose between two alternatives, as shown in (21). 9 If participants are forced to judge only one dimension, it will be clear why the participants chose one alternative over the other. This may result in clearer data.
(21) a. I buy the house white b. I buy the white house
More traditional fieldwork methods could also prove useful. One on one interviews with a couple of invested informants have proven a valuable resource for the compilation of syntactic/morphological/phonological atlases. If informants can be found with suitable attitudes, this type of research may also provide more nuanced results than a study of the type conducted for this paper. Especially the possibility for a dialogue about the issues under investigation will give a more representative image of the factors influencing the acceptability of codeswitched sentences.
Another alternative approach can eschew grammaticality judgment tasks altogether. Sometimes participants have difficulty providing judgments to sentences that are prescriptively bad. This issue can be circumvented by taking a more subconscious route. Psycho- and neurolinguistic approaches have gained popularity in codeswitching and bilingualism research (Fairchild & Van Hell, 2015; Sanoudaki & Thierry, 2014). These types of studies unearth subconscious reactions to stimuli and provide a different point of view to the issue.
Because this is the first study looking at adjective-noun word order in Dutch-French codeswitching, and the first to look at adjectival agreement in codeswitching in general, these results in favour of the Minimalist Program can only be seen as tentative. Additionally, as was mentioned in an earlier part of the paper, data from grammaticality judgments are controversial. It has been argued that only an integration of corpus and naturalistic data, grammaticality judgments and neurolinguistic evidence can give a complete picture of codeswitching data (Gullberg et al., 2009). Since the evidence in this study has been limited to grammaticality judgments, these results will only gain a more definitive nature if and when they are replicated using other methodologies.
Footnotes
Appendix A
Acknowledgements
I would like to thank Eric Schoorlemmer, Philippe De Brabanter, Nicholas Ruyttenbeek, the participants of Bilforum 2015 and ISB10, and two anonymous reviewers who provided comments, suggestions and food for thought, all of which immensely improved this paper. I’d also like to thank Maricarmen Parafita Couto for all of this and much more.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
