Bilinguals have a single computational system but two compartmentalized phonological grammars: Evidence from code-switching

Classic linguistic models, such as Chomsky’s minimalist schematization of the human language faculty, were typically based on a ‘monolingual ideal’. More recently, models have been extended to bilingual cognition. For instance, MacSwan (2000) posited that bilingual individuals possess a single syntactic computational system and, crucially, two phonological systems. The current paper examines this possible architecture of the bilingual language faculty by utilizing code-switching data. Specifically, the natural speech of Maria, a habitual Spanish-English code-switcher from the Bangor Miami Corpus, was examined. For the interface of phonology, an analysis was completed on the frequency of syllabic structures used by Maria. Phonotactics were examined as the (unilingual) phonological systems of Spanish and English impose differential restrictions on the legality of complex onsets and codas. The results indicated that Maria’s language of use impacted the phonotactics of her speech, but that the context of use (unilingual or code-switched) did not. This suggests that Maria was alternating between encapsulated phonological systems when she was code-switching. For the interface of morphosyntax, syntactic dependencies within Maria’s code-switched speech and past literature were examined. The evidence illustrates that syntactic dependencies are indeed established within code-switched sentences, indicating that such constructions are derived from a single syntactic subset. Thus, the quantitative and qualitative results from this paper wholly support MacSwan’s original conjectures regarding the bilingual language faculty: bilingual cognition appears to be composed of a single computational system which builds multi-language syntactic structures, and more than one phonological system.


Introduction
Since the onset of language research, one particular model has dominated the field: that of the monolingual speaker.Historically, monolinguals were viewed as the 'baseline' language user, while bilinguals were defined in relation to their simpler counterparts, notably as being "two monolinguals in one person" (see Grosjean 1989 for a discussion).While it is reasonable for science to first focus on less complex cases (i.e., monolinguals), global numbers indicate that monolinguals may not be a baseline after all.In fact, it is estimated that over half of the world's population speaks more than one language (Grosjean & Miller 1994), making bilingualism and multilingualism the norm rather than the exception.
With the development of the field, more and more interdisciplinary experts now agree that the bilingual mind may be distinct from that of the monolingual, simply by virtue of holding two languages within one single cognitive system (e.g., MacSwan 2000).For instance, a wealth of data indicates that many bilinguals possess cooperative lexicons (e.g., Beauvillain & Grainger 1987;Hartsuiker, Pickering & Veltkamp 2004;Sabourin, Brien & Burkholder 2014), seemingly integrated within at least one linguistic interface.Despite this upsurge in empirical interest, the underlying composition of the bilingual language system is still obscure.Fortunately, one practice may illuminate its particulars: code-switching.
you have told that story 'You have told that story.'Such consistent acceptability judgements offer a glimpse into the underlying bilingual system.They reveal the ways in which languages can and cannot be mixed, and therefore, shed light on the possible cognitive associations between the code-switched languages at each linguistic interface (e.g., phonology, morphology, syntax).In short, the examination of code-switched utterances allows linguists to model the bilingual language faculty while considering the bilingual as a baseline, rather than the monolingual.
As such, the current paper utilizes code-switching as a mechanism to assess previously purported models of bilingual cognition.In particular, the underlying structure of the bilingual language faculty is probed by two main questions: How many phonological systems are employed for code-switched utterances?How many syntactic subsets characterize code-switched sentences?
Evidence both empirical (i.e., corpus-based data analyses) and theoretical in nature confirms that code-switching is derived from a unique syntactic grammar but dual phonologies, as past models have suggested (MacSwan 2000;MacSwan & Colina 2014).
The current paper is structured as follows: the next section expands on background concepts relative to code-switching and the bilingual language faculty under a minimalist perspective.
Section 3 is devoted to the interface of phonology.Here, I analyze syllabic structure frequencies by using a Spanish-English code-switching corpus (the Bangor Miami Corpus; Deuchar, Davies, Herring, Parafita Couto & Carter 2014) in order to tap into the underlying phonological codeswitching systems.In the fourth section, the interface of syntax is addressed; I expand on the theoretical support for the singularity of the syntactic code-switching grammar and once again utilize the code-switching corpus as empirical support.The final section concludes.
(2) Spanish/English a. Ayer fui al supermercado.I bought some apples.yesterday went to the supermarket I bought some apples 'Yesterday I went to the supermarket.I bought some apples.' b.El niño está reading the book.
the boy is reading the book 'The boy is reading the book.' As researchers recognized the systematicity of code-switches, a multitude of (primarily syntactic) language mixing constraints were introduced.For instance, the Equivalence Constraint (3), brought forth by Poplack (1980), claimed that code-switching boundaries are prohibited where such a switch may give rise to an illicit word-order in one of the two languages.Following the Equivalence Constraint, a series of language mixing constraints were proposed: the Closed class constraint (Joshi 1982), the Government constraint (Di Sciullo, Muysken & Singh 1986), the Matrix language frame (Myers-Scotton 1997), the Functional head constraint (Belazi et al. 1994), among others.While these theories often possess reasonable predictive power, most have been critiqued by proponents of the constraint-free approach (e.g., MacSwan 2000;2014;MacSwan & Colina 2014).This is because the Equivalence Constraint and many of its successors directly refer to the act of code-switching and therefore impose upon it an otherwise vacuous 'third grammar'.By contrast, constraint-free formalism maintains that code-switching does not (and should not) possess constraints that are independent from monolingual speech.As it turns out, Chomsky's minimalism is well equipped to carry out a constraint-free analysis of code-switching (MacSwan 2000;2014).
(3) Equivalence Constraint: Code-switches will tend to occur at points in discourse where juxtaposition of Ll and L2 elements does not violate a syntactic rule of either language, i.e. at points around which the surface structures of the two languages map onto each other.(Poplack 1980: 586) In minimalism, patterns observed from linguistic data are preferentially elucidated by the most simple and elegant assumptions.Consider Chomsky's generally accepted hypothesis that the (monolingual) human language faculty is made up of A) a lexicon and B) a computational system (C HL ) (see Chomsky 1995).In brief, C HL is responsible for Selecting lexical arrays from the lexicon, combining them into a syntactic structure with the operation Merge, and then Spelling-Out the derivation to logical form (LF: the interpretative level of language) and phonological form (PF: the sensorimotor aspect of language).
Adopting the constraint-free approach for code-switching data, MacSwan (2000) extended This system very simply accounts for code-switching while avoiding switching-specific grammars: in the bilingual language faculty, C HL can select lexical items from multiple lexicons and combine them into a single syntactic construction.On the other hand, a bilingual individual can easily speak in a single language by selecting items from a single lexicon instead of two.
Thus, by adopting a minimalist framework, MacSwan convincingly illustrated that an adequate code-switching grammar should not explicitly or conceptually refer to the act of using multiple languages (an insight previously broached by Mahootian 1993;Belazi et al. 1994).Instead, his model is anchored in the following assumptions (MacSwan 2000: 43): (4) i.
[…] All grammatical relations and operations which are relevant to monolingual language are relevant to bilingual language, and only these.ii.Nothing constrains code switching apart from the requirements of the mixed grammars.
In short, assuming that two languages, L x and L y , are involved in a mixed sentence, and that these languages each possess their own grammar, G x and G y , any code-switched sentence is determined by G x ∪ G y .We do not need to know the arbitrary identity of any L, and we do not need to generate a switching-specific grammar.
Once the syntactic structure (the overt component) is derived, it must surface as a phonological/phonetic form.The monolingual model simply sends the sensorimotor elements of the derivation to the unique phonological system in the configuration.The bilingual language faculty, on the other hand, possesses more than one lexical set, and thus, more than one phonological system.In a monolingual utterance, bilinguals can easily respect the phonology of any of their languages: items from each lexicon are merely rendered to their corresponding phonological system.Contrastively, during code-switching, all lexical items must be parsed by their appropriate phonological systems after Spell-Out. 1 Hence, to give rise to a felicitous code-1 Strikingly, this schematization also appears to stand for bimodal bilinguals.Instead of simply spelling-out a derivation to a given phonological system, bimodal bilinguals may spell-out the structure to the phonology of distinct articulators: speech or sign (see Lillo-Martin, Müeller de Quadros & Chen Pichler 2016).To recapitulate, MacSwan's model of the bilingual language faculty implies that codeswitching is possible at the syntactic level, but that code-switching at PF is unavailable.This means that each segment involved in the switch respects the phonology of the lexical set it originates from, and furthermore, that there is no singular alternate phonological grammar that can account for the mixed utterance in its entirety.Presently, it may be asked exactly why the syntactic and phonological interfaces, by nature, differ within the architecture of the bilingual system.That is, what motivated MacSwan to posit that syntactic grammars can be combined, while the phonological counterparts could not?Essentially, this conjecture is based on the computational features of each interface-regardless of whether the system is monolingual or bilingual.
Within the syntactic interface, operations can take place at any given time.For instance, inserting a post-nominal adjective into a derivation before introducing a noun will still give rise to a correct surface form, as the lexical item is labeled with [adj. +post-nom.].Based on this imperviousness to underlying order relations, there is no concern in stipulating that items from code-switched sentences are syntactically derived from an assembled set (i.e., a single lexical array constructed by C HL ).
On the other hand, phonological rules or constraints have been noted as being order-dependent at least since Bromberger and Halle (1989), who penned the Principle in (5).Since phonological grammars are sensitive to ordering, it would not be parsimonious to assume that code-switching is characterized by the united phonologies of the languages involved in the switch.Consider, for instance, a set of phonological rules {R What would take place if Spell-Out introduced a code-switched sentence composed of L x and L y into a combined PS?At the very least, the result of such a computation is challenging to obtain (though see Chapter 6 of López 2020 for a potential analysis).
(5) Phonological rules are ordered with respect to one another.A phonological rule R does not apply necessarily to the underlying representation; rather, R applies to the derived representation that results from the application of each applicable rule preceding R in the order of the rules (Bromberger & Halle 1989: 58-59).
Thus, the separation of the phonological systems within the architecture of the bilingual mind directly falls from the "imperfection" of phonology, notably its poor computational properties (Chomsky 2002: 118).It is a consequence of the order-specific nature of the phonological interface; if phonological systems were united in the model, their respective orders could be compromised.As strikingly noted by MacSwan & Colina (2014: 188), this entails that "language switching is phonological switching", and nothing more.Such an observation falls into line with conceptualizations of language which claim that variation is confined to the externalization of language (Chomsky, Gallego & Ott 2019), while the internal language remains largely invariant.
The ii.Each set of internally ranked constraints is a constraint dominance hierarchy, and a language-particular phonology is a set of constraint dominance hierarchies.iii.Bilinguals have a separately encapsulated phonological system for each language in their repertoire in order to avoid ranking paradoxes, which result from the availability of distinct constraint dominance hierarchies with conflicting priorities.iv.Every syntactic head (word) must be phonologically parsed at Spell-Out.Therefore, the boundary between heads represents the minimal opportunity for codeswitching.
Before continuing, it is important to distinguish code-switching from nonce-borrowing (and attested loanwords).According to original theory, code-switching, whether inter-or intrasentential, involves unique processes (e.g., Grosjean 1995).Unlike switched segments, borrowed words are said to be morphologically, and often phonologically integrated into the target language (Poplack & Meechan 1998;though see Poplack, et al. 2020 for recent data).Since borrowings (and not code-switches) may be parsed by the phonological system of the recipient language (e.g., 'rid-er' is pronounced as French-like /ʁɑ͡ ɪde/, even though it originates from the English riding /ɹɑ͡ ɪdɪŋ/), these types of items do not give us much insight into the architecture of the bilingual's phonological systems.From a computational perspective, sentences featuring borrowed words or loan words are not at all different from single-language speech; these items will not be considered in the empirical analyses of this paper.
In the subsequent section, past phonetic studies examining code-switching will be reviewed; these indirectly support MacSwan's model of the bilingual language faculty.

Experimental phonetic/phonological literature
A moderately large literature has investigated whether bilinguals abruptly switch between the phonetics of their languages when they produce mixed utterances, or alternatively, whether the phonetic systems somehow converge towards a new schema that is dissimilar from that of both languages in unilingual contexts.The results from these studies remain conflicting.
Early findings examining the voice-onset time (VOT) of French-English code-switchers indicate that bilinguals shift from one phonetic system to the next, with absolutely no phonetic remnant of the opposite language after a switching boundary (Grosjean & Miller 1994).These results have seldom been replicated (though see Brown 2015; Olson 2013 (bilingual contexts) for comparable conclusions).For the most part, studies have observed some sort of effect between the languages involved in the code-switch.
On one hand, some research indicates that code-switched speech may actually be hyperarticulated.This result has primarily been found for suprasegmental features, such as pitch (Olson 2012; see also Olson & Ortega-Llebaria 2010) and stress, as indicated by vowel length (Olson, 2012;Muldner et al. 2019).In other words, the suprasegmental features of the languages involved in the switch may diverge from each other, possibly in order to ensure maximal distance between their acoustic features.Hyper-articulation of code-switched items may also occur since switches are said to be "less predictable" (Olson 2012: 451) than single-language speech (see Bell, Brenier, Gregory, Girand & Jurafsky 2009 for supporting results).In any case, hyperarticulation is not problematic to MacSwan's model of the bilingual language faculty; it still presupposes that bilinguals have encapsulated phonological systems and that code-switches are not phonologically integrated.
On the other hand, several studies indicate that phonetic convergence does indeed take place between languages involved in a code-switch; this is by far the most common finding ( Pérez & Wayland 2018).For instance, when participants produced switches between Spanish and English, the VOT values of stops in these languages may shift towards each other.Interestingly, this shift is often asymmetric: it is primarily observed within the language that contains a broader VOT spectrum.Thus, English VOT may shift towards shorter Spanish-like values, since these segments can afford to shift while respecting their own phonological thresholds (Bullock et al. 2006;Balukas & Koops 2015;Piccinini & Arviniti 2015).
Despite these complex and often antithetical findings, one result remains unchanged: the presence of phonetic convergence does not result in a merge of the systems.In fact, the phonetic shift never extends as far as the respective phonological categorization boundaries of each language.For example, Balukas and Koops (2015) observed a durational shift for English voiceless stops in code-switched compared to monolingual speech, notably towards shorter Spanish-like VOT values.Nevertheless, the VOTs of the code-switched English stops still remained within the typical English range (see also Bullock et al. 2004;2006;Antoniou et al. 2011;Piccinini & Arvaniti 2015, etc.).This means that the often-observed phonetic convergence is, as its name implies, simply phonetic in nature.It does not negate the duality of the phonological systems according MacSwan's model of the bilingual language faculty.
In brief, the acoustic literature illustrates that phonetic convergence (among other possibilities) is available for code-switches, but that this shift is virtually never phonological.These results rather support the notion that code-switching does not impact the phonologies of the languages involved in the mixed utterances.Up to this point, however, mostly indirect evidence concerning this matter has been examined.In the subsequent section, the duality of the phonological systems, as postulated by MacSwan's model of the bilingual language faculty, is assessed.Unlike previous research, this is done by examining a facet that is quite distinct from the phonetic system: that of syllabic structure.

Background
The goal of the current section is to investigate the underlying phonological system of codeswitched utterances, in order to verify MacSwan's model of the bilingual language faculty.One way of tapping into these rather abstract notions is by comparing code-switched languages which differ within a specific aspect of their respective phonological systems.For instance, MacSwan and Colina (2014) utilized the Spanish-specific phonological rule resulting in the realization of intervocalic stops as approximants.Unilingual English displays no pattern similar to the Spanish approximant rule.These authors observed that in Spanish-English code-switches favouring the application of the Spanish rule (e.g., /migowst/ 'mi ghost' → */miɰowst/), stops were only realized as approximants in about 10% of the recordings (compared to 93% in unilingual Spanish).However, phonological processes were successfully triggered for segments belonging to the same language in which the process was active (e.g., Spanish voicing rule: /misgowsts/ 'mis ghosts' → /mizgowst/).MacSwan and Colina thus cleverly demonstrated that code-switching boundaries involve an abrupt shift in the bilingual's phonological systems.
This means that examining such patterns effectively allow us to hone into the phonological component of the language faculty, as modelled by Chomsky and MacSwan.
The possible empirical outcomes are presented in (7).Outcome A) supports phonological The general phonotactic differences between Spanish and English are exploited in order to test the outcomes in (7).Spanish and English have distinct constraints regarding the legality of certain syllabic structures.English has a relatively lax syllable structure; it can accommodate up to three consonants in the onset position, and even more in the coda (e.g.strengths /strɛŋ(k)θs/  CCCVC(C)CC, glimpsed /ɡlɪmpst/  CCVCCCC).Contrastively, consonant clusters within a given syllable are greatly disfavoured in Spanish (e.g., Barlow 2003).As can be seen in Table 1, this is reflected in the relative frequency of syllable types in unilingual English (Delattre & Olsen 1969) and Spanish (Villanueva-Reyes 2013).
As it turns out, the grammatical distinction between English and Spanish is easily modeled with the Optimality Theory framework (OT).In OT (see Smolensky & Prince 1993), the ranking of universal constraints forms a language-specific grammar; the numerous permutations for hierarchical constraint rankings give rise to cross-linguistic differences in phonological systems.
When comparing languages with different constraints rankings (and thus different superficial phonological attributes), one can use OT to determine which system is currently active in codeswitched discourse.
In the present study, the relevant constraint which primarily accounts for the phonological differences between Spanish and English is *Complex (i.e., "Assign a violation for every syllable with a consonant cluster").That is, *Complex possesses a distinct ranking in Spanish and English, being of a relative higher rank in the former than the latter (8).This simple distinction accounts for the resulting phonotactics of both languages, as can be seen in Tables 2 and 3 (loosely adapted from Misker & Anderson 2003).
Simply put, this pattern is obtained since English grammar 'prefers' outputs that preserve all of their original segments (i.e., faithful candidates), rather than outputs without consonant clusters.
Crucially, this preference is determined by the outranking of Max-IO over *Complex (8a).
2 Delattre and Olsen obtained these syllabic frequencies by examining 2000 syllables in texts from contemporaneous authors (e.g., short stories, plays). 3Villanueva-Reyes (2013) originally separated syllables according to whether they contained a simplex vowel or a diphthong.Since vowel quality is of no interest to the current paper, diphtongues and simple vowels were collapsed.
It is not specified whether these relative syllabic frequencies were obtained from speech or text samples.With only a handful of constraints, we can observe quite simply from where arise the differential phonotactic patterns in English and in Spanish.While this analysis may be slightly idealized, it demonstrates the key idea upon which the current section hinges: cross-linguistic differences in the ranking of *Complex give rise to distinct relative syllabic frequencies in Spanish and English.As such, this paper capitalizes on these distinct rankings of the *Complex constraint while examining natural speech from a Spanish-English corpus.If bilinguals do indeed possess encapsulated phonological systems (in line with MacSwan's model and the constraint-free approach), the hierarchy of constraints in code-switched speech should respect the hierarchy of constraints of each respective language involved in the code-switch.

The corpus
The publicly available Bangor-Miami corpus was consulted (see Deuchar et al. 2014; available at http://bangortalk.org.ukthrough the Free Software Foundation's General Public License).This corpus contains transcribed audio files of several hours of recorded speech from Spanish-English bilinguals in the region of Miami, Florida.The current paper focused on the recordings (n = 15) from a 45-year-old female speaker with the pseudonym 'Maria', who was known to the researchers as an avid code-switcher and a balanced bilingual.As such, she was given a microphone and invited to record code-switched conversations in her daily life (Deuchar et al. 2014).The corpus contains a total running time of approximately 14 hours of Maria's speech.

Coding procedures and variables of interest
Based off the audio files and their transcriptions, sentences produced by Maria were coded according to two crucial variables: Context and Language.The variable of Context refers to the type of utterance: code-switch or unilingual.Code-switched segments correspond to streams of speech containing both Spanish and English, as shown in the examples in (9).Unilingual segments correspond to streams of speech occurring in either English (10a) or Spanish (10b).
In order to avoid possible inter-sentential switching points, segments were only extracted and coded as unilingual if they were preceded by at least one other single-language sentence in the same language.Occurrences of single other-language items such as (11a-b) were excluded, since they may be borrowed words; attested loanwords were also omitted (see Section 2.1).
The variable of Language simply corresponds to whether each segment was produced in Spanish or in English, regardless of Context.In code-switched sentences, the portions of speech are also coded as belonging to Spanish or English (e.g., from (9b): tienen ese salon grande, pero en, etc. are coded as Spanish and county, so it's very easy to hang out, etc. as English).The language tag (Spanish or English) is determined by the lexicon of origin of each word (i.e., L x or L y in Figure 1).For instance, in (9a), menos is a lexical item in Spanish, but not in English.In the case of cross-linguistic homonyms (e.g., tan in means ‛so' in Spanish, and ‛shade of brown' in English), the syntactic context provided a disambiguation for the language tag (e.g., for (9b) to be well-formed, tan must possess the Spanish meaning and is thus tagged as belonging to Spanish).Notice that the phonotactics of the words are not consulted during the coding of the language tags.
Thus, the language tag of the cognates could be established according to lexical and morphological features of the word.For example, resistir in (9c) contains Spanish verbal inflection, and is thus tagged as Spanish; similarly, compañía possesses the Spanish feminine declension class -a, and is thus tagged as Spanish.With respect to orthographically identical cognates, single phonemes were used to disambiguate the language tag.For instance, natural may be tagged as Spanish if the rhotic is produced as /ɾ/ rather than /ɹ/.Consequently, even in the case of cognates, no reference to the phonotactics of the word is required in order to attribute a language tag to Maria's segments of speech.Note that proper nouns and other pronounced speech sounds were not analyzed ('Henry', 'Naomi', 'uh', 'oh', 'woo', etc.), as the language tag is often ambiguous.

(9)
Code-switched sentences (Spanish/English) a.We may have to put the house shields you know home house shields on those para por lo menos quitar un poquitico porque he wants to at least show the mira cuando no tenia un shield […].(Maria,7,61) 'We may have to put the house shields you know home house shields on those in order to at least remove a little bit because he wants to at least show the… look when he didn't have a shield…' b.In XXX county tienen ese salon grande so it's very easy to hang out pero en XXX county no es tan comfy como eso.(Maria,24,588) 'In XXX county they have that big lounge so it's very easy to hang out but in XXX county it's not so comfy as that.' c.No lo pude resistir cause they were too cute.(Maria,40,153) 'No I can't resist because they were too cute.' (10) Unilingual sentences (Spanish/English) a.I saw that we have some proposal due on the tenth.Are you going to want me to review it?(Maria, 2, 1-5).
b. Pedro me dijo.Vamos a dar tragos a todo el mundo.(Maria,16,(48)(49).'Pedro told me.We're offering drinks to everybody.' (11) Excluded other single-language items (Spanish/English) a. Mi hermano llego el sabado antes de Christmas.(Maria,4,164) 'My brother came the Saturday just before Christmas.'b.They are not going to remove it but they wanna be able to say bueno you know it's no different than whatever your house your street light or something.(Maria,7,59) Once each segment was coded for Context and Language, its corresponding syllables were manually extracted and, based on impressionistic judgements, categorized as one of the following types: CV, CVC, V, VC, CCV, VCC, CCV, CCVC, CVCC, CCVCC, CVCCC, CCVCCC, VCCC, CCCVC and CCCV.This process was accomplished with reference to both the transcription and the corresponding auditory files for Maria's conversations.In the case of any uncertainty (i.e., degraded sound quality, differences between the transcription and the audio recording), Spanish informants were consulted (see acknowledgements) to identify the syllables.Since vowel quality was of little interest to the current project, syllables with simplex vowels and diphthongs were collapsed into a single syllabic type.
Note that only intra-sentential (and not inter-sentential) code-switches were extracted (see 9), since this is the environment in which phonological convergence is most likely to materialize.
In other words, if convergence were to take place in code-switched speech, it is most probable during instances of lengthy switches, when Maria is freely using both of her languages to express her thoughts.If convergence does not appear even in these types of contexts, it serves as robust evidence that Maria is actively alternating between compartmentalized phonological systems.

Potential empirical outcomes
The syllabic structure extracted from Maria's code-switched speech was compared to that of her unilingual speech.More specifically, the proportions for each syllable type within the Spanish portions of Maria's code-switched speech were compared to their relative proportions in the Spanish portions of Maria's unilingual conversations.Similarly, the proportions for each syllable type within the English portions of Maria's code-switched speech were compared to their relative frequency in the English portions of Maria's single-language conversations.
Since the constraints relative to syllable structure (e.g., *Complex) vary between Spanish and English, Maria's use of different syllables structures during code-switched speech will give insight into the underlying phonological system(s) that is-or are-employed when she mixes her languages.As such, the possible empirical outcomes from ( 7 As per MacSwan's minimalist model of the bilingual language faculty, the analysis is expected to yield outcome A; the relative frequency of syllabic structure used by Maria is not hypothesized to differ significantly from her unilingual speech in each respective language.

Analyses
In order to verify this hypothesis empirically, two separate analyses were realized: generalized linear models and repeated measure ANOVAs.As the frequencies of the syllable types are 4 Miami Latino English (colloquially referred to as the "Miami-Dade accent") appears to present its own dialectal attributes.In particular, linguistic studies have observed differences in the prosody (e.g., Alvord 2006;Carter, López Valdez & Sims 2020), the realization of the lateral /l/ (e.g., Rogers & Alvord 2019) and the realization of low vowels (e.g., Carter, López Valdez & Sims 2020) in Miami English.To my knowledge, research up to this point has not yet examined the phenomenon of cluster reduction in this dialect.
interdependent for a given analysis, and are skewed for values below 0.5, all proportion values were transformed into logits.This was the dependent variable in both analyses.
Firstly, three generalized linear models were constructed for the data (see Supplementary material for R script).The first ("Baseline") contained the following independent (categorical) predictors: Onset (zero Cs, one C, two Cs, three Cs) and Coda (zero Cs, one C, two Cs, three Cs).
In the second ("LanguageModerator") and third ("ContextModerator") models, the additional categorical predictor of, respectively, Language (English or Spanish) and Context (unilingual or code-switched), were introduced.The interaction between this additional predictor and each other initial predictor was also included in the last two models (e.g., Context*Onset, Context*Coda, Context*Onset*Coda).The goal of this analysis was to examine the goodness of fit of each model, in order to assess whether the introduction of Language or Context enable the model to make better predictions about the frequency of each syllable type.The goodness of fit of each model was interpreted via the Akaike Information Criterion (AIC; Akaike 1987).The AIC gives information about the quality of a given model: the lesser the AIC, the better the fit of the model.AIC is useful, as it penalizes the addition of any new predictor to the model.This prevents more complex models from being a better fit simply by virtue of being more complex.
A difference in AIC larger than +/-2 between models is interpreted as a significant difference between the models.If Maria is using compartmentalized phonological systems when she is code-switching, the inclusion of Language should provide the truest model for the data.
The second analysis was used to confirm the results from the AIC analysis.Separate 2 x 2 repeated-measures ANOVAs were conducted for each syllable type (i.e., CV, CVC, V, VC… etc.)

The data
From the 14 hours of Maria's recorded conversations, 37 609 syllables were coded, with 88% of syllables belonging to unilingual speech (within unilingual speech: English = 69%; Spanish = 31%) and 12% belonging to intra-sentential code-switching (within code-switches: English = 45%; Spanish = 55%).The number and proportions belonging to each syllable type according to the experimental factors (Context and Language) are detailed in Table 4 and depicted visually in Figure 2. Note that Maria's syllable proportions largely match those that have been reported previously for unilingual English and Spanish speech (see Table 1).

Generalized linear models
The results from all three models are presented in Table 5.As can be seen, the inclusion of the Language predictor (LanguageModerator) produced the best fitted model; its AIC is the smallest in magnitude, and the difference in AIC between LanguageModerator and both other models is larger than 2. As such, even when additional predictors were penalized, Language (Spanish vs. English) improved the predictability of the data.ContextModerator was not better fitted than the baseline model (difference in AIC = 0.5), indicating that Context (monolingual vs codeswitched) does not improve the goodness of fit of the model to the data.

Repeated-measure ANOVAs
The results from the ANOVAs are presented in Table 6.As can be seen, the factor of Language produced significant (or near-significant) differences between syllable proportions for the

Discussion
In line with the hypotheses, statistical models indicated that the language of production of each syllable was a significant predictor in determining syllabic frequency in intra-sentential codeswitches.Context was neither a significant predictor for syllabic use in Spanish nor in English.As such, even when Maria was code-switching, the Spanish portions of her mixed utterances were Spanish-like, and the English portions of her mixed utterances were English-like (see assumption A in ( 12)).By contrast, context was neither a significant predictor for syllabic use in Spanish nor in English-Maria was not using an excess of complex onsets and codas in the Spanish portions of her code-switched speech, and she was not limiting these same syllabic structures in the English portions.These main findings demonstrate that Maria was actively and abruptly switching between the encapsulated phonological systems of her language.
Recall that MacSwan suggests that code-switching at the phonological form of language is unavailable, since PF is composed of hierarchical constraints: if the phonological forms were united, in a code-switching context, the order of the constraints would not be upheld.In consequence, when a derivation arrives at Spell-Out, two separate phonologies (PS x and PS y ) are compartmentalized before PF.This means that lexical items originating from either lexicon can only be conditioned by L x or L y .The results from the current study thoroughly support MacSwan's model: Maria's language-specific orderings of *Complex were preserved during code-switching (i.e., the syllabic structure in her code-switched speech matched the syllabic structure of respective monolingual speech).Consequently, the current data also fail to support models which maintain that bilinguals possess fully integrated grammars (see López 2020); Maria displayed no convergence between her phonological systems.The interface of phonology is clearly dichotomized in the bilingual language faculty.
How can independent phonological systems be analyzed within OT? Preliminary considerations tend to support the notion of co-phonologies.The formalism of co-phonology has mostly been are treated with different grammars (e.g., one for suffixes, one for roots).Co-phonologies have previously been implemented for words from different lexical classes (Inkela & Zoll 2007).From this line of thinking, it is not a huge leap to suppose that co-phonologies can be applied to different lexical sets, as in code-switching.

Syll. Type Effect
A co-phonology seems to correspond to what was observed in Maria's code-switched speech: Maria abruptly alternated from one phonological system to the next when she mixed Spanish and English.In a code-switching co-phonology, each language involved in the utterance would possess its own compartmentalized phonological system.That is, phonological systems would map to a single language involved in the switch.Spoken works would thus be parsed by the appropriate system, depending on the language from which they originate.

Section summary
Up to this point, the minimalist model of the bilingual language faculty (MacSwan 2000) has been supported.Utilizing the phonotactic differences between Spanish and English (i.e., regarding the *Complex constraint), the speech of a Spanish-English code-switcher was analyzed in order to determine whether she abruptly alternated between her phonological systems during codeswitching.The analysis revealed differences in syllabic proportion between languages (Spanish vs. English), but not between contexts (unilingual vs. code-switched).These results reinforce the idea, in line with MacSwan's model of bilingual language faculty, that A) distinct phonological systems are accessed during code-switching and B) that these systems correspond to those of each monolingual language involved.
We can now answer the first half of the primary research question: the bilingual language faculty is composed of two compartmentalized phonological systems.Just like MacSwan originally conjectured, "language switching is phonological switching" (MacSwan & Colina 2014: 188).In the next section, I will answer the second part of the research question.That is, I will review theoretical evidence that suggests that code-switching is composed of a single syntactic grammar.This evidence will once again be supported by in-the-field instances of Maria's codeswitched speech.
4 Syntax: Single and ready to mingle

Background
While the independence of the phonological systems is supported, this does not reflect the entire conceptualization of the bilingual language faculty.Recall that MacSwan's model also ascertains that bilinguals (just like Chomsky's monolingual) possess a computational system (C HL ) which combines items into a single syntactic derivation.That is, once items are chosen from the lexicon(s), the array is Spelled-Out in one fell swoop (or, potentially, in a series of phases; see López et al. 2017); structures are built when items are derivationally combined via the operation of Merge (MacSwan 2000).
Up to this point, I have taken the singularity of C HL for granted.Nonetheless, another option is possible when it comes to the bilingual language faculty: individuals who speak more than one language may possess globally compartmentalized systems for each language, as in Figure 3 (right panel).Luis López (2017 and onward) coins this framework as separationism or the commonsense framework, since it is reflective of the way humans intuitively think about language (e.g., "I speak three languages, but my mom only speaks two.").As MacSwan's model also involves some degree of architectural separation (i.e., at the level of phonology), I will refer to the common-sense framework as global separationism.Global separationism entails that bilinguals' morphosyntactic grammars 5 are just as independent as their phonological systems.
Crucially, global separationism and MacSwan's model make different assumptions with regard to multi-language structures.If global separationism is correct, the bilingual language faculty 5 When alluding to the morphosyntactic system or grammar, I refer to this component in a purely minimalist sense.
That is, I examine morphemes that are essentially realizations of syntactic features (e.g., case, person, gender, number, etc.).This should not be confused with morphological traits privy to integration, which has often been linked to a change in the phonological system (e.g., MacSwan 2000).would only be able to Merge items from a single lexicon, since the C HL is language-dependent, or language-encapsulated.Contrastively, MacSwan's minimalist model of the bilingual language faculty allows for the combination of items originating from multiple languages.
You may think that global separationism is immediately falsified, since bilinguals evidently do combine items from more than one lexicon-that is the very core of intra-sentential codeswitching, after all.However, a sceptic might argue that the mere existence of code-switching does not negate global separationism, since bilinguals may simply be building partial syntactic structures within their compartmentalized computational systems, before pronouncing these structures in juxtaposition.In other words, some may argue that code-switching has the outward appearance of a combined syntactic system, but that it is actually composed of independent structures being outputted side-by-side.This would entail that code-switching is a surface phenomenon that exists in performance but not in competence.
How can we verify whether this is the case?It seems that the solution lies in the examination of syntactic dependencies (i.e., c-command, grammatical case, complement relations), which presupposes the singularity of C HL .The theoretical outcomes are the following.The diverse syntactic code-switching literature, as well as Maria's speech from the Bangor-Miami corpus, provide us with ample opportunities to verify these assumptions.Much of the argumentations provided in this section are inspired by Luis López' generativist body of work, most notably his recent book (2020).Recall that, unlike MacSwan, López believes that the bilingual language faculty is combined at every interface, including PF.While the syllabic data from Section 3 markedly refutes the integration of a bilingual's phonological systems, this does not mean that López' entire model is unsupported.In fact, MacSwan and López appear to agree in their assumption that bilinguals derive code-switches from a unique syntactic subset.Maria's naturalistic data provides more evidence for this conjecture.
Given that much of the evidence provided in the subsequent sections was also presented by López (2020), novel data (i.e., Maria's data from the Bangor-Miami corpus; syntactic judgements obtained by the author) are identified as such.

Negative polarity items
According to Hoeksema (2000: 115), negative polarity items (NPIs) are linguistic units that "give rise to minimal pairs of affirmative and negative sentences".Crucially, in order to produce grammatical sentences, NPIs must be c-commanded (see Chomsky 2014) by negative elements, like not (Giannakidou 2011).Thus, the difference between the sentences in ( 16) is that any is licensed (or triggered) in (16a) but not in (16b).
( In each of these cases, a syntactic dependency is established between an NPI and a trigger that originate from distinct lexical sets.That is, a negative particle from L x licences an NPI from L y .Thus, the singularity of the syntactic component is supported: the examples above clearly demonstrate that code-switching is not simply the juxtaposition of two partial syntactic structures constructed in language-dependent C HL .In actuality, the lexical items from the code-switched language pairs are united into a combined array, from which the unique C HL can build a single syntactic construction.

Grammatical case
Grammatical case is associated to the syntactic or grammatical function of an item within a given sentence.Many languages possess complex case systems, wherein grammatical case is indicated by an inflectional morpheme; the sentence becomes unacceptable in the absence of case marking or when the wrong case is assigned.For instance, in (20b) bedny-'poor' is assigned the genitive case instead of the dative case (as in 20a), and the resulting sentence is ill-formed.If case assignment occurs between a verb from L x and an item from L y , we can say that there is a syntactic dependency between these items, and that they co-exist within the same C HL .Past evidence suggests that these syntactic dependencies do indeed take place in code-switching (21-24).For instance, in the sluicing constructions in (21), the verb 'to threaten' assigns the dative case in German (21a), but the accusative case in Spanish (21b).However, in the code-switch, the German remnant 'who' must possess the accusative case (21c a. Anau star-y ploshod-tï ne-ler-di zönder-di.this old-nom/acc square-acc thing-acc renovate 'There the old square and so were were already renovated.'Altogether, these data illustrate that grammatical case assignment can take place between items from different languages.As case assignment entails syntactic dependency (assumption A), it is clear that the intra-sentential code-switches shown above were derived from a singular syntactic construction.The evidence above therefore supports the singularity of the syntactic component, in line with MacSwan's formulation of the bilingual language faculty.

Mixed selection
López (e.g.2020) originally introduced the 'Mixed Selection' problem by examining codeswitching between light verbs and infinitives.To generalize this concept, I assume that Mixed Selection has occurred when an item from L x selects one from L y as its complement, or vice versa.

Determiner phrases
The general consensus in syntax is that constituents such as that flower, la flor are headed by the determiner (see Bernstein 1991).In Spanish and English, the determiner-head requires a nominal complement.For instance, the sentence in (25) becomes unacceptable if the noun phrase within the scope of the DP is empty.
(25) a. Ben went to the store to buy apples./ *Ben went to the to buy apples.b.Ben fue a la tienda a comprar manzanas./ *Ben fue a la a comprar manzanas.
Thus, if a determiner from L x selects a nominal complement from L y , we can assume that Mixed Selection has occurred.Under global separationism, the derivation containing the DP should crash within the language specific C HL , since the complement position would be empty.However, if we assume that code-switching is built from a single syntactic construction, Mixed Selection involving determiners and nouns is entirely unproblematic.

Verbal complements
Some transitive and ditransitive verbs require complements.For instance, the sentence in (27), which contains the transitive verb 'to buy'/'comprar', is rendered unacceptable if the object DP is absent.In this case, one would say that the verbal head 'selects' (or Merges with) its complement, 'a pair of socks'/'un par de calcetines'.naturally produced speech (28).It is also abundant in the code-switching literature (Pfaff 1979;Ewing 1984;Azuma 1993;Mahootian & Santorini 1996;Gonzalez-Vilbazo & López 2011).
The presence of such constructions provides solid evidence of a united syntactic component, in line with MacSwan's model of the bilingual language faculty.If the C HL were languagedependent, the partial structure involving the verb would crash before Spell-Out in the absence of its complement.

Section summary
In the current section, the second half of MacSwan's model of the bilingual language faculty was supported: code-switching is in fact governed by a singular morphosyntactic grammar.The evidence from the literature and Maria's spontaneous speech shows that syntactic dependencies do indeed take place between different-language items.This is not predicted by global separationism, which presupposes globally encapsulated systems for each language known by a bilingual or multilingual.
We can now answer the second half of the primary research question.Contra the "commonsense" hypothesis, the bilingual language faculty is composed of a unique computational system, wherein syntactic constructions are built.

Concluding remarks
The current paper has examined the architecture of the bilingual language faculty by reviewing code-switching data.As MacSwan originally conjectured, bilingual cognition is composed of a single computational system, which can build multi-language syntactic constructions, and two phonological systems.MacSwan's model was supported using both empirical (corpusbased) data and theoretical frameworks.This means that code-switching, essentially, is 'just phonological switching".Interestingly, neuroimaging studies have shown that language switching does indeed engage brain regions associated to phonological processing (Price, Green & von Studnitz 1999).
This paper is not only relevant for formal bilingualism research; it may also allow researchers to contextualize past experimental studies.For instance, past research has demonstrated that cognates (i.e., orthographically and/or phonologically similar words cross-linguistically, such as table in French and English) may act as catalysts for code-switching (e.g., Brown 2015).That is, the alternation between languages (code-switching) is most likely to occur after a cognate has been produced.This finding also appears to be predicted by the architecture of the bilingual language faculty.If the assembling process is cognitively costly, then this effort may be mitigated by the presence of a cognate, since such words possess similar entries in each of the lexicons.
Bilinguals may thus utilize this temporary dip in the required cognitive effort to juxtapose items from their languages.Future research should examine this interesting pattern more closely.
Before ending this paper, one last peripheral point can be made about MacSwan's model of the bilingual language faculty.Recall that, in the pursuit of computational efficiency, MacSwan theorizes that bilinguals possess separate lexicons for each language that they speak (for alternative perspectives see work by Artemis Alexiadou, Kay González-Vilbazo, Luis López and Sara Stefanich).This is not the prevalent view, particularly in psycholinguistics, since a wealth of data indicates that bilinguals possess cooperative or (partially) overlapping lexicons (e.g., Beauvillain & Grainger 1987;Hartsuiker et al. 2004;Sabourin et al. 2014).While this issue is beyond the scope of the present paper, it does not appear that the architecture of the lexicons is particularly pivotal to MacSwan's model.Whether or not the lexicons are overlapped, the reality remains that bilinguals possess one C HL and two phonological systems.It is not immediately apparent how the presence of an integrated lexicon would challenge these assumptions.
Altogether, in bolstering MacSwan's model of the bilingual language faculty, this paper supports the constraint-free approach for code-switching, as it achieved by the minimalist program.That is, code-switching is uniquely governed by the languages that compose it, and not a third grammar.Future studies should examine switching using this minimalist tradition.
Chomsky's minimalistic view of the human language (see Figure1, left panel) to the bilingual language faculty.The bilingual version of the model (see Figure1, right panel) includes two differences.First, there are two lexicons available to the bilingual speaker.Second, bilinguals possess discrete phonological systems, which are compartmentalized before PF.Just like the monolingual language faculty, however, bilinguals are thought to construct sentences within a single syntactic component.
Spanish syllabification of absorb-to ('absorbed').Maria's recordings were chosen over those of other speakers in the corpora, since her conversations took place in-the-field.In fact, Maria was conversing with friends, family and colleagues(Deuchar et al. 2014), and her recordings are thus very likely to reflect her natural speech.Recent reviews indeed call for both a) the use of habitual code-switching participants; and b) more ecologically valid data sets or collection methods(González-Vilbazo et al. 2013;   Beatty-Martínez, Valdés Kroff & Dussias 2018; Blanco-Elorrieta & Pylkkänen 2018; van Hell,   Fernandez, Kootstra, Litcofsky & Ting 2018).By using Maria's spontaneous recordings, these methodological considerations are addressed.
following syllable types: CVC, V, VC, CCV, CCCV, VCC, CVCC, CCVCC, CVCCC.The factor of Context was only significant for one syllable type (devoid of any consonant clusters): VC.Based on these results, it is clear that Language (i.e., belonging to an English vs. a Spanish portion of speech) is a better predictor of syllable identification than Context (i.e., belonging to a unilingual vs. a code-switched portion of speech).

Figure 2 :
Figure 2: Proportion of syllables according to Context and Language.

Figure 3 :
Figure 3: Left: Simplified version of MacSwan's (2000) model of the bilingual language faculty.Right: The bilingual language faculty from a globally separatist perspective (adapted from López 2020: 17).
If syntactic dependencies are found between items of L y and L x , code-switching can be represented by G x^y .B. If items from L y and L x do not interact syntactically, code-switching cannot be represented by G x^y .
Bogatye nijogda ne zavidujut bednym.the-rich.nomnever neg envy the-poor-dat 'The rich never envy the poor.' b. *Bogatye nikogda ne zavidujut bednyx the-rich-nom never neg envy the-poor-gen bought a pair of socks./ Mona compró un par de calcetines.b. *Mona bought./ *Mona compró.If an obligatorily transitive verb from L x selects a complement from L y , this means that Mixed Selection has occurred.There are countless examples of this type of Mixed Selection in Maria's as it is posited in current model of the bilingual language faculty (MacSwan 2000); (7)Potential empirical outcomes: Let a code-switch be composed of x and y A. If x respects the phonotactics of L x , and y respects the phonotactics of L y , the phonologies of L x and L y are impermeable.B.Alternatively, if x does not respect the phonotactics of L x , or y does not respect the phonotactics of L y the phonologies of L x and L y are permeable.
Contrastively, in Spanish (Table3), an input such as /absorb-to/ 'absorbed' is syllabified into three units, none of which include a consonant cluster: VC, CVC and CV syllables.Syllabification occurs in this manner since consonant clusters are less desirable in Spanish than deleted segments, as determined by the ranking of * Complex over Max-IO (8b).(8) Crucial constraint rankings of interest: a. English: Max-IO > *Complex b.Spanish: *Complex > Max-IO

Table 4 :
Overall summary of coded data, collapsed across audio files.

Table 6 :
Inkelas 1998;Inkelas & Zoll 2007) p-values are flagged with asterisk(s) and bolded; p-values trending towards significance are also bolded.In all cases, df = (1,14).appliedtomorphologically conditioned phonological systems.In a co-phonology (e.g.,Inkelas 1998;Inkelas & Zoll 2007), different morphological components (within a single language) ) a. I do not have any more candy.b. *I have any more candy.Past data (and novel judgements) 6 indicate that sentences containing opposite-language NPIs and negation triggers are acceptable, as indicated in(17)(18)(19).In each case, the NPI is bolded and the license is underlined.Under a globally separatist point of view, such sentences should be ill-formed, since syntactic relations like c-command can only take place within the languagecompartmentalized C HL .On the other hand, MacSwan's model predicts that such constructions should be entirely acceptable, since lexical arrays are built into a single syntactic structure and can thus Command one another.
Juan threatened someone.accbut I know not who.acc/ dat 'Juan threatened someone, but I don't know who.