1 Introduction

Bilingual speakers often alternate between two or more languages during a single conversation. This phenomenon is known as code-switching. While researchers have observed that intersentential code-switching is pretty free, in that code-switching between any two independently well-formed sentences produces a grammatical utterance, intrasentential code-switching is more constrained in that code-switch location affects whether or not a code-switched sentence is judged to be licit or illicit (see Poplack 2015 for a review). This literature is based on the observation that not all intrasentential code-switches are licit, as seen in the contrast between (1a) and (1b) which differ only with respect to code-switch location, indicated by | :

Much of the work on intrasentential code-switching has focused primarily on investigating the syntactic conditions under which code-switched utterances are categorized as either grammatical or ungrammatical (see MacSwan 2013 for a review). While these approaches have been highly influential in arguing for a single computational mechanism responsible for building both unilingual and code-switched expressions, they often stop at this point. Focusing only on what code-switched sequences are licit or illicit fails to consider all of the possible sequences that are generatable (i.e structures that are available in at least one of the bilingual’s languages even if they are not the canonical or preferred order in the other language) to convey an intended meaning. This focus overlooks valuable insights we can gain from investigating bilinguals’ flexible processing of various word orders. Since we know that sentences can vary in how preferable they are across language users even within a bilingual group (e.g., Królikowska et al. 2019, as discussed in Beatty-Martínez & Dussias 2019), our larger goal in this study is to investigate how bilinguals process structures that may be dispreferred or less natural but not necessarily illicit. For example, Welsh-English bilinguals produced complex noun phrases that varied the relative ordering of adjectives and nouns between both of their languages (e.g., English Noun-Welsh Adjective, Welsh Adjective-English Noun), although they favored certain orders more than others (Parafita Couto & Gullberg 2019). It is this variability in structural preferences which makes code-switching bilinguals an ideal population for investigating whether, when confronted with less-preferred but generatable structures, bilinguals prioritize specific elements in the input they receive, or if they show no clear preference.

In this study, we aim to contribute to our understanding of bilingual language processing by testing bilinguals’ sensitivity to various word orders across both single- and mixed-language expressions. Under the assumption that bilinguals use a single parser to process both single- and mixed-language expressions, we investigate whether they prioritize certain elements when processing what we term congruent (i.e., structures where elements appear in their single-language canonical word order) and incongruent (i.e., structures where elements appear in non-canonical positions) expressions, or whether they show no clear preference. To do this, we conducted a picture identification task with three bilingual groups – Arabic/English bilinguals, Korean/English bilinguals, and Spanish/English bilinguals – where participants were presented with three-word expressions that varied in word order (Determiner-Modifier-Noun, Determiner-Noun-Modifier) and location of the code-switch (No Switch, after Word 1, after Word 2, Double Switch). We selected these three populations because, for each group, the preferred word order in one language differed from that of the other language. This task allowed for a comparison of processing efficiency across congruent and incongruent structures. Faster processing was observed when the word order aligned with the typical structure of the noun’s language. These findings suggest that language users prioritize the most conceptually informative lexical items (which are nouns in our task) during online comprehension processing. And with these results, we argue that the lexical item that functions as the lexical head of the expression—nouns for nominal expressions—serve as the anchor for determining the underlying syntactic structure.

2 Relevant Background

A structure is classified as congruent or incongruent based on the single-language preferences for canonical word order. Since single-language utterances are understood to follow specific word order requirements, we focus on a structure that has varying canonical word orders requirements across different languages, mainly three membered nominal phrases consisting of a singular definite determiner, a color-denoting modifier, and an inanimate object-denoting noun.

Consider the following examples from English, Arabic, Korean, and Spanish in (2):

All four expressions depict the same concept: some red-colored house. However, the canonical word order for these stand-alone expressions differ between Arabic, Korean, Spanish, and English. While in English, adjectives typically appear before nouns they are modifying (i.e., pre-nominal), in languages like Korean, Spanish, and Arabic, comparable modifiers follow the noun they are modifying (i.e., post-nominal). Therefore, we consider unilingual expressions like “the red house” and “la casa roja” as being congruent because they follow the canonical word orders of English and Spanish, respectively. By contrast, unilingual expressions like “the house red” and “la roja casa” would be considered incongruent because they do not follow the canonical word order. Though these would generally be treated as illicit,¹ what gets considered licit for mixed-language expressions has been a matter of considerable theoretical debate (see MacSwan 2013 for a review).

Since the early 1970s, linguists have attempted to account for the rule-governed nature of code-switching expressions. Some of these earliest approaches were either highly language specific, for example positing constraints for Korean/English code-switching that are different from constraints on Spanish/English code-switching (e.g., Timm 1975), or postulated code-switching specific constraints that made explicit reference to lexical items being associated with a particular language variety (e.g., Sankoff & Poplack 1981; Belazi et al. 1994). Later work, however, suggests that code-switching can, and should, be accounted for by the same structure-building principles that govern unilingual utterances (Pfaff 1979; Woolford 1983; Santorini & Mahootian 1995). This perspective has led to proposals that attempt to explain code-switching asymmetries within a constraint-free framework.

Competitions between language systems within the bilingual mind have been examined at all levels of language, from phonological (e.g., Gonzales & Lotto 2013) through semantic (e.g., Lee & Williams 2001). While some argue for language-specific mechanisms during phonological processing and lexical access, there has been convergence on the view that a single computational mechanism is responsible for constructing both single- and mixed-language syntactic structures, which lead to unified theories about grammatical vs. ungrammatical mixed-language expressions. However, none of the existing studies consider how their theoretical approach can be modeled into a bilingual parser that constructs mixed-language expressions, particularly in contexts where the switch happens between phrases that are structurally incongruent with respect to their preferred single-language word orders. This gap leaves unanswered questions about how bilinguals process these transitions, especially when confronted with shifts in surface word order that do not overlap across languages. We avoid this debate for now by seeking to first determine what is congruent for mixed-language expressions.

To identify the congruent structures for mixed-language expressions, we must determine which lexical item(s) contains a critical cue to guide the parser’s word order preferences. Language processing involves generating multiple structural representations in parallel, which are then evaluated and weighted against each other (Traxler 2014). Within this framework, the parser (i.e. the mechanism within the language processing system responsible for building and interpreting structural representations) plays a crucial role. As we recognize each word, we integrate that item into our current structural representations and then use the updated representations, along with contextual information, to anticipate the next word (Ferreira & Chantavarin 2018). Our ability to anticipate input, using grammatical knowledge and contextual information, facilitates online processing.² But because the parser generates many viable representations in parallel, it needs to weigh possible representations against each other by the probability that the structure is what the speaker intended. Although it may seem plausible to assume that bilinguals develop distinct parsers for each language, with a supervisory system managing when to use each parser, we reject this assumption based on recent bilingual processing studies which suggest that combinatorial operations are not language-specific (see Declerck et al. 2020; Phillips & Pylkkänen 2021). Assuming a single parser that generates and compares structures for both single- and mixed-language expressions requires a biasing mechanism for determining which structures, perceived as expressions, are preferred. Thus, having a bias for a particular lexical item is particularly relevant given how bilinguals can access words from both of their languages at any time (see Costa et al. 2006 for a review), inviting different language structures to compete.

3 Methods

Our goal is to investigate how bilinguals ease the processing of (in)congruent code-switched expressions by anticipating the most likely structure intended by a speaker. Assuming bilinguals rely on a single parser to process both single- and mixed-language expressions, we now explore whether or not bilinguals prioritize specific elements during online processing. We hypothesize that if bilinguals do bias a specific lexical category (or information contained by the items belonging to that category) when determining the most likely structure, this will be reflected in faster processing times for expressions where the biased word appears in its congruent word order position. Specifically, one of the following three outcomes could emerge: If bilinguals privilege determiners, we expect shorter response times when English determiners appear in Det-Mod-N sequences, and when Korean, Spanish, or Arabic determiners appear in Det-N-Mod sequences. If they instead prioritize modifiers, participants should respond more quickly when English adjectives appear in Det-Mod-N sequences and when Korean, Spanish, or Arabic adjectives appear in Det-N-Mod sequences. Similarly, if nouns are the privileged category, faster response times should be observed when English nouns appear in Det-Mod-N sequences and Korean, Spanish, or Arabic nouns appear in Det-N-Mod sequences. On the other hand, if bilinguals do not bias any particular element when processing (in)congruent structures, we would expect no significant difference in processing times across varying word order positions, as processing would be more reliant on some other mechanism (e.g., a shallow, conceptually-driven approach).

3.2 Stimuli & Experimental Design

A set of determiners, four colors, and four inanimate objects were selected to create the auditory stimuli for each bilingual group (see Table 2 in Supplemental Materials for full list of lexical items). Determiners that behave most like the English article “the” were selected for Arabic (“el”), Korean (“ku”), and Spanish (“el” for masculine nouns, “la” for feminine nouns). Color-denoting modifiers were selected because they (i) transparently modify the object-denoting nouns with which they compose and (ii) consistently appear post-nominally in Arabic, Korean, and Spanish. Object-denoting nouns were selected for being commonly occurring inanimate items with unambiguous, cross-linguistically consistent lexical equivalents in the languages studied. All instructions and auditory stimuli were elicited from highly proficient bilingual consultants. The visual stimuli were created using Microsoft Powerpoint to generate images of the selected objects and colors for each bilingual pair. (For more information about our stimuli and experimental design, see A1 in Supplemental Materials).

To measure the online processability of various single-language and code-switched sequences across three bilingual groups, we used a picture identification task that minimally differed between all groups (see A2 in Supplemental Materials for a description of the procedure). In each trial, participants heard one of two types of sequences: determiner-modifier-noun (Det-Mod-N) sequences, or determiner-noun-modifier (Det-N-Mod) sequences (see Tables 3–5 in Supplemental Materials for example stimulus sets for each bilingual group). Each auditory sequence was presented with a four-field visual world; and each field contained a unique inanimate object (see Figure 1). For Det-Mod-N sequences, participants saw four different objects in one of two colors. We refer to this as the object-identification task. For Det-N-Mod sequences, participants saw two different objects in one of four colors. We refer to this as the color-identification task. We recorded which colored object they selected (accuracy) and how long it took for them to click on the target image using their computer mouse (response time).

Figure 1: Task Designs by Stimulus Type – Participants completed two picture identification tasks that depended on the auditory stimulus they heard: the object identification task for Det-Mod-N sequences (left) and the color identification task for Det-N-Mod sequences (right). Examples of single-language stimuli are also presented to illustrate the in-/congruent contrast.

4 Results

Data from each bilingual group was preprocessed and analyzed separately (see A3 in Supplemental Materials for a full description). Across all three bilingual populations, participants were overall highly accurate (Korean-English bilinguals M = 97.40%: Spanish-English bilinguals: M = 96.51%; Arabic-English Bilinguals: M = 94.79%). Given that participants were able to successfully complete the two tasks for our two sequence types, we further analyzed their response times on accurate trials to determine whether word order, language of the determiner, language of the adjective, language of the noun, and/or any interactions thereof affected how quickly participants responded as a reflection of processing. Tables 6–8 in the Supplemental Materials contain the outputs from the statistical models.

4.1 Korean/English Bilingual Participants

First, we observe a main effect of the language of the noun: participants were faster for sequences with English nouns (t = 5.76, p < 0.001). Second, we observe a significant interaction between word order and the language of the noun (t = –2.64, p = 0.008). We performed a post-hoc pairwise comparisons test to test for significant differences between the relative ordering of Korean nouns and English nouns. These results reveal a significant difference between the relative ordering of Korean nouns (p = 0.0003) but not between the relative ordering of English nouns (p = 1.00). Participants were faster when Korean nouns were presented in their canonical position. This can be seen in Figure 2.

Figure 2: Korean/English bilingual participants were faster for English nouns across sequence and task types, as well as for when Korean nouns appeared in their canonical position (Det-N-Mod).

4.2 Spanish/English Bilingual Participants

First, we observe a main effect of the language of the noun (t = 3.99, p < 0.001): participants were faster for sequences with English nouns. Second, we observe a significant interaction between word order and the language of the noun (t = –2.20, p = 0.03). We conducted a post-hoc pairwise comparisons test on this interaction, and observed a significant difference in the relative ordering of English nouns (p = 0.0004) but not Spanish nouns (p = 0.26). Ultimately, our participants were faster when English nouns appeared in their canonical position. This can be seen in Figure 3.

Figure 3: Spanish/English bilingual participants were faster for English nouns across sequence and task types, as well as for when English nouns appeared in their canonical position (Det-Mod-N).

4.3 Arabic/English Bilingual Participants

First, we observe a main effect of the language of the noun (t = 3.78, p = 0.0002): participants were faster for sequences with English nouns. However, we did not observe a significant interaction between word order and the language of the noun (t = –0.48, p = 0.63). We performed a post-hoc pairwise comparisons test on the interaction between word order and language of the noun, given the main effect of the noun. Those results revealed a significant difference in response times for the relative positioning of English nouns (p = 0.02) but not Arabic nouns (p = 1.00). Taken together, participants were faster for English nouns, and this was likely driven by when English nouns appear in their canonical position. This can be seen in Figure 4.

Figure 4: Arabic/English bilingual participants were faster for English nouns across sequence and task types, as well as for when English nouns appeared in their canonical position (Det-Mod-N).

5 Discussion

This study set out to investigate how bilinguals process object-denoting expressions in both single-language and code-switched contexts, particularly when the two languages involved differ in canonical word order. We asked whether bilinguals rely on specific lexical categories when interpreting (in)congruent structures and examined whether this reliance is modulated by structural differences across language pairs.

We observed a significant interaction between word order and language of the noun for two of our three groups (Korean/English, Spanish/English). However, the results from conducting post-hoc pairwise comparisons for the interactions between word order and the language of the noun revealed that all three groups were faster when nouns appeared in their canonical position. Taken together, our results suggest that bilinguals are not responding solely based on task structure or surface position. If participants’ response times were simply a reflection of task demands, we would have expected a main effect of word order across all conditions. For instance, since the final word in each phrase was always the one that participants had to identify (nouns in the object-identification task, modifiers in the color-identification task), response times should have been uniformly faster or slower based on that position alone. This, however, was not the case. Instead, they appear to privilege a specific lexical category (namely, the noun) when determining the most likely structure. We interpret these results as evidence that nouns play a central role in guiding bilinguals’ word order preferences in object-denoting expressions.

Why would the noun take a privileged position when committing to a specific, object-denoting structure? In both the processing and syntactic literature on nominal phrases, we take for granted the intuition that nouns provide the core of the conceptual representation derived from a nominal expression, akin to how verbs provide the core of the conceptual representation derived from a verbal expression. Some proposals have tried to capture this intuition in the syntactic literature by claiming that it is lexical heads, not functional heads, that determine the grammatical combination of words contained within a given expression (Grimshaw 2000). We argue that the preferred order of a code-switched expression depends on the language of the lexical item that functions as the lexical head of the expression.

As we discuss the relevance of this finding, it is important to note that the nominal constructions examined here were presented in isolation and not embedded within a larger carrier phrase. This is relevant to note as prior research has shown that bilingual word order preferences and processing can be influenced by the broader morphosyntactic context surrounding mixed nominal phrases (Cantone & MacSwan 2009; Balam & Parafita Couto 2019; Parafita Couto & Gullberg 2019; Parafita Couto & Stadthagen-Gonzalez 2019; Vaughan-Evans et al. 2020; De Nicolás & López 2022; van Osch et al. 2023; Olson 2024). Consider the following example sentences from previous studies interested in the relative ordering between nouns and adjectives among Spanish/English (3a) and Papiamento/Dutch (3b) bilinguals:

Both studies found that bilinguals preferred adjective placement that matched the word order patterns of the verb’s language in code-switched sentences. Although our study specifically focuses on isolated nominal expressions, the observed effect of verbs in the prior work and our effect of nouns converge on the finding that bilinguals often privilege the lexical head when processing mixed language expressions. What would be interesting is if we observe a subject-object asymmetry in word order preferences in future work. The examples in (3) have code-switched nominals function as the internal argument of the verb (i.e. object), such that the verb would shape the preferred word order of the embedded code-switched nominal. If the lexical head of an expression serves as an anchor, a code-switched nominal functioning as the external argument (i.e. subject) would likely follow the preferred word order of the noun’s language (as opposed to the verb’s). This would further suggest that factors at the clause level, including the morphosyntactic environment, play a crucial role in how bilinguals parse and prefer certain word orders. Theoretical frameworks such as Myers-Scotton’s (1997) and Myers-Scotton & Jake’s (2000) work differentiating between a matrix and an embedded language in code-switching, as well as exoskeletal models of language mixing (e.g., Alexiadou & Lohndal 2018; Riksem, 2018; Grimstad et al. 2018), further emphasize the importance of these structural factors in influencing how bilinguals may construct code-switched structures.

While this general pattern held across language groups, examining each group’s results separately also raises several questions about the consequences of how specific language grammars interact. First, we found that Korean/English bilinguals were sensitive to Korean nouns appearing in Det-N-Mod sequences; whereas the Spanish/English bilinguals and Arabic/English bilinguals were sensitive to English nouns appearing in Det-Mod-N sequences. This could reflect distributional differences between groups: Korean modifiers can appear pre-nominally when they are morphologically marked, which we did not manipulate in this study; whereas Arabic and Spanish lack this option. Interestingly, post-hoc pairwise comparisons for the interactions between word order and the language of the modifier revealed that Spanish/English bilinguals and Arabic/English bilinguals were slower when Spanish modifiers and Arabic modifiers appeared in their canonical position. One possibility for this result is an effect of gender agreement on processing. Korean and English do not exhibit gender agreement between nouns and their modifiers, but Spanish and Arabic do. And while there is evidence that suggests Spanish/English bilinguals will default to using the Spanish masculine form with English nouns in code-switched nominal phrases (e.g., Valdés Kroff 2016), we propose that there could have been a processing delay among our Arabic/English bilingual group due to a lack of definiteness spreading (see A1 in Supplemental Materials for further discussion). While we stand by our methodological decisions, they may have partially contributed to the specific result differences observed between groups (see A4 in Supplemental Materials for further discussion). Finally, we found that all three groups were faster for English nouns. This could reflect the high preponderance of English noun insertions produced by bilinguals who acquire English as one of their languages (Parafita Couto & Gullberg 2019) or an effect of our participants being largely English-dominant (see A4 in Supplemental Materials for further discussion).

6 Conclusion

The current study explored how bilinguals are able to process both single-language and code-switched expressions, particularly when the two languages involved exhibit (in)congruent word orders. We presented Korean/English, Spanish/English, and Arabic/English bilingual adults three-word sequences and asked them to identify the intended target from a visual world containing four objects. Our findings show that all three groups exhibited faster processing when the word order of the phrase aligned with the word order preferences of the language of the noun. These results suggest that not only do bilinguals prioritize the most conceptually-informative element when processing single- and mixed-language expressions, but they are sensitive to the distributional preferences of that element when determining the preferred structural representation. Our findings contribute to the growing evidence that word order are not categorical but gradient and shaped by multiple interacting factors. The linearizations we examine in this study are all generatable, in that both the Det-N-Mod and Det-Mod-N structures are viable word orders for our bilingual participants. While bilingual speakers may show a preference for certain linearizations when mixing languages, there is still much we can learn from looking closer at the structures that they may not prefer but are still generatable. In our study, we saw that bilinguals’ preferences are not arbitrary. Instead, there was a clear tendency for certain structures to be preferred over others. Had we only focused on identifying the conditions under which bilinguals labeled a code-switched utterance (un)acceptable, we would have missed this generalization. Because of that, we argue that these results should encourage us as researchers to treat structures that are canonical in one language but non-canonical in another as a continuous variable, rather than categorical when investigating bilingual language use, in a way that is similar to recent arguments that have been made for word order variability in general (see Levshina, Namboodiripad, et al. 2023; Beatty Martínez et al. 2025; and references therein).

Supplemental materials

Supplemental Materials – Appendix can be found here: https://doi.org/10.16995/glossa.23254.s1

Acknowledgements

We would like to thank Alejandra Reinoso for her assistance in the development of the Spanish/English bilingual stimuli, and all of the bilingual participants who engaged in our study despite widespread uncertainty during the early days of the COVID-19 pandemic. We are especially grateful to Gary Thoms for his feedback on an earlier draft of this manuscript. We also thank the audiences at PLC 46 for their comments and discussion, and the anonymous reviewers for their helpful feedback. This work was supported through funds allocated to SFP from New York University.

Competing interests

The authors have no competing interests to declare.