New patterns of pathological language mixing in bilingual aphasia

Abstract

Background:

A central challenge in bilingual aphasia research is the absence of clear markers that distinguish pathological language mixing (PLM) from mixing used as a compensatory mechanism, leaving ongoing debates unresolved. Also, prior work suggests that PLM occurs primarily during retrieval of open-class words and increases with utterance length during structured tasks, yet these findings, and the hesitation behaviors accompanying PLM, remain largely untested in spontaneous speech.

Method:

This study examines PLM in Patient P3, a 59-year-old female bilingual speaker of Spanish and English with severe Broca’s aphasia who showed extensive mixing and reported significant communicative disruption as a result. She was administered the Western Aphasia Battery-Revised in both languages and transcribed samples from her response to narrative prompts were coded for word class and language mixing, and hesitation durations were acoustically analyzed.

Results:

P3 mixed closed-class words more frequently than open-class words. PLM frequency was unrelated to utterance length, and longer hesitations preceded non-target language intrusions, indicating increased retrieval cost. P3 showed elevated Aphasia Quotients in both languages when cross-language responses were computed, suggesting symmetric (bidirectional) mixing rather than assymetric (unidirectional) compensatory intrusion into the weaker language.

Conclusion:

Together, these results propose symmetric language mixing and higher hesitation duration preceding non-target intrusions as potential markers of control deficits in bilingual aphasia.

Introduction

Language mixing refers to the insertion or alternation of lexical items and larger constituents from one language into the other, either within a clause (intra-sentential mixing) or across clause or utterance boundaries i.e., inter-sentential mixing (Mooijman et al., 2025). This is usually intentional, pragmatically motivated code-switching characteristic of neurotypical bilinguals (Lehtonen et al., 2023). This linguistic behavior may be pathologized in bilingual aphasia due to stroke which is marked by language control and/or retrieval deficits (e.g., Riccardi, 2012). Pathological language mixing (PLM) has been described as reflecting language control deficits marked by uncontrolled or inappropriate alternation between languages following brain damage (Abutalebi et al., 2000). Empirical work has documented PLM in bilingual aphasia across multiple linguistic levels, including single-word intrusions, multi-word constituents, and clause-level switches, often concentrated in the post-stroke weaker language (Bihovsky et al., 2023; Goral et al., 2019). However, not all mixing observed in bilingual aphasia reflects a loss of control. Recent studies demonstrate that some mixing is systematic and compensatory rather than pathological. For example, Balasubramanian et al. (2025) showed that a Hindi – English bilingual with Broca’s aphasia produced selective English noun insertions within a rigid Hindi morphosyntactic frame, reflecting underlying lexical – semantic vulnerabilities rather than generalized switching failure. This debate persists, in part, because the field lacks well-defined markers that differentiate pathological language mixing from compensatory mixing in bilingual aphasia.

PLM in bilingual aphasia has been shown to be more frequent in the less proficient language and during constrained tasks, especially when retrieving semantically rich content words than function words (e.g., Goral et al., 2019; Lerman et al., 2019; Mooijman et al., 2025). However, most evidence comes from structured tasks like picture naming or description, while spontaneous speech remains underexplored, particularly across grammatical categories such as open- and closed-class words. The impact of utterance length on PLM is unclear, with mixed findings regarding its relationship to language control demands (Martin & Schnur, 2019; Paplikar et al., 2014). Hesitation duration especially those occurring before switches, may signal internal monitoring or control difficulties (e.g., Hlavac, 2011; Mooijman et al., 2025). This study therefore examines whether PLM differentially affects open- and closed-class words, whether mixing increases with utterance length, and whether hesitation rates differ between utterances with and without PLM. Our research is guided by three primary questions:

1. Does language mixing affect specific word classes (i.e., open-class vs. closed-class words) during elicited spontaneous speech?

2. Does language mixing increase with increased utterance length?

3. Does the duration of hesitation before retrieving a word differ depending on whether the word is produced in the target language or the non-target language during elicited spontaneous speech in PLM?

Clinically, PLM can complicate differential diagnosis and severity ratings, as frequent cross-language intrusions may be misinterpreted as confusion or “non-compliance” with language-of-testing instructions rather than as manifestations of impaired language control or strategic compensation (Ansaldo et al., 2010). Understanding whether language mixing reflects control deficits, lexical retrieval support, or adaptive translanguaging has direct implications for whether clinicians aim to suppress PLM or intentionally harness it to support functional communication (Lerman et al., 2019; Mooijman et al., 2025). Single-case studies such as the present one therefore contribute to refining bilingual assessment protocols and to designing interventions that are sensitive to individual patterns of PLM.

Method

We employed a single-subject experimental design (Imaezue et al., 2025), collecting two baseline measures of elicited spontaneous speech at Timepoints 1 (T1) and 2 (T2), one week apart. Baseline performance on the English and Spanish versions of the Western Aphasia Battery-Revised (WAB-R: Kertesz, 2007) was also assessed at T1 and T2 respectively. A mobile device with a pre-installed assessment application was mailed to the participant’s home and the entire study was conducted remotely via Microsoft Teams.

Participant

P3 is a 59-year-old right-handed woman who was a premorbid bilingual dominant speaker of Spanish (native) and English (acquired). Spanish was acquired in early childhood at home, and English was acquired at age 19 through work and school opportunities. Following immigration to the United States at age 37, P3 subsequently attained fluency at age 38. In 2019, P3 suffered a left middle cerebral artery aneurysm with subarachnoid hemorrhage resulting in chronic aphasia that impaired her ability to communicate in both languages, as confirmed by clinical report. Following her stroke, P3 demonstrates PLM characterized by the unintentional combination of Spanish and English in spontaneous speech. An excerpt of P3’s speech is provided here:

I like go out rain [EN] pero [ES] I like to stay out. Yes I like it. But [EN] en [ES] Peru [PN], go there [EN] no hay [ES] rain [EN] poquito eso no es pero aqui si [ES] I like rain. [EN]

Note: EN = English; ES = Spanish; PN = Proper noun. Language tags indicate language membership of each lexical item to highlight points of mixing. (Imaezue et al., 2025)

P3 resided in a predominantly English-speaking environment and complained that her language mixing significantly disrupted her everyday communication. Thus, her goal was to minimize unintentional language mixing and increase her use of English words when speaking. P3 provided signed written consent forms approved by the Institutional Review Board of the University of South Florida (USF IRB ID: STUDY006201) before participating in the study and did not receive financial compensation for her participation.

Language and cognitive assessments

We collected background information on P3 through self-reported questionnaires, including demographics and medical history (see Imaezue et al., 2025). The Language Experience and Proficiency Questionnaire (LEAP-Q; Marian et al., 2007) and the Bilingual Switching Questionnaire (BSWQ; Rodriguez-Fornells et al., 2012) were administered to ascertain P3’s language history, proficiency and her perception of language mixing. LEAP-Q revealed high pre-stroke proficiency in both languages and significant post-stroke decline while BSWQ indicated frequent unintentional language switching post-stroke (see Imaezue et al., 2025). Baseline aphasia severity and type were determined using the Western Aphasia Battery-Revised (WAB-R: Kertesz, 2007) administered in English and Spanish with P3 diagnosed with severe Broca’s aphasia in both languages. Cognitive abilities were assessed using the English version of the Cognitive Linguistic Quick Test-Plus (CLQT+; Helm-Estabrooks, 2017). CLQT+ results indicated overall mild cognitive impairment, with executive functions and attention within normal limits but significant memory deficits (see Imaezue et al., 2025).

Elicited spontaneous speech task

The speech samples analyzed in this study were gathered using a custom mobile application. First, P3 was instructed to ensure the mobile device was connected to the internet and to place the headphones on. Then, she navigated to the app where 20 narrative prompts were presented visually and auditorily (e.g., “What do you like to do on the weekends?”). P3 then initiated a one-minute spontaneous response by pressing the app’s “Record” button and she was instructed to respond in English alone (the target language). We used this approach because P3 was primarily interested in attempting to respond in English alone since she wanted to reduce her PLM, particularly Spanish words intrusions. Consequently, the intrusion of Spanish words in her utterance was considered as PLM for this task. Each prompt-response pair was designed to elicit naturalistic language production. The app automatically captured each response with accompanying metadata and uploaded the files to secure cloud storage, enabling subsequent acoustic and linguistic analyses. P3 received initial training on how to use the app and reported no technical difficulties throughout data collection. No concurrent speech-language therapy was provided during the data collection period. No corrective feedback was given with respect to speech content or language control.

Linguistic and speech analysis

Authors CD and CS – both proficient in English and Spanish – orthographically transcribed the samples and coded all words by language and word class. Language mixing was defined as intra- or inter-sentential insertion of one language into another, excluding proper nouns like names and geographical locations (e.g., “Peru”). The exclusion of proper nouns follows established code-switching methodology because they usually remain unchanged regardless of linguistic context. For instance, Poplack (1980) excluded proper names and place names from her code-switching corpus, contending that these forms operate as fixed, socially entrenched lexical units rather than as active switch points. In line with this interpretation, subsequent accounts treat proper nouns as linguistic islands or essentially unanalyzed insertions, thereby distinguishing them from true code-switches (Bullock, 2009). Authors DA and CD quantified the number of English and Spanish words, calculated correct information units (CIUs), and determined the proportion of open- and closed-class words in each language. For each narrative sample, language proportions were calculated separately for open-class and closed-class words categories. All English and Spanish open-class tokens (nouns, verbs, adjectives, adverbs) and closed-class tokens (articles, prepositions, conjunctions, auxiliaries, pronouns) were identified. Proportions were then computed using category-specific denominators rather than the total word count. Specifically, the English open-class proportion was defined as English open-class tokens ÷ (English + Spanish open-class tokens), and the English closed-class proportion as English closed-class tokens ÷ (English + Spanish closed-class tokens). Spanish proportions were calculated analogously. Cognates were also identified, and utterance length was computed by the number of words per response across all samples. Hesitation duration analysis was conducted using Audacity (v3.7.3). Spectrograms were used to generate fine-grained visualizations of hesitations, defined as the onset and offset of silent pauses preceding word retrieval in either language, measured in milliseconds (e.g., Antoniou et al., 2011). Hesitations were categorized as either non-language-mixing or language-mixing-related. Using Audacity’s selection tools, the authors marked onset and offset timepoints to establish each pause duration.

Results

Non-parametric Wilcoxon signed-rank tests were used to examine differences in language mixing patterns across word class and cognate status. A Spearman’s rank-order correlation was used to evaluate the relationship between utterance length and the frequency of language mixing.

Summary of WAB-R performance

P3’s performance across the WAB-R subtests varied as a function of single or combined language responses across English and Spanish versions of WAB-R. Table 1 shows that the highest scores, particularly Aphasia Quotients (AQs), occurred when responses in both languages were combined compared to when only responses in one language were used to establish language performance and aphasia severity.

Table 1.

Participant performance on WAB-R English and Spanish versions.

Language Condition	AQ	Spontaneous Speech	Auditory Comprehension	Repetition	Naming
L1	33.7	50%	60.35%	10.7%	1.8%
L2	32.3	20%	50.85%	10.8%	40.5%
L1/L2	49.2	50%	60.35%	10.7%	60.9%
L2/L1	37.7	25%	50.85%	10.8%	40.5%

Note: L1 = Spanish, L2 = English, AQ = Aphasia Quotient.

Language mixing and word class

Language mixing (Spanish words intrusion) occurred more during retrieval of closed-class during elicited spontaneous speech than open-class words (Table 2). A Wilcoxon signed-rank test confirmed a statistically significant difference in the proportion of closed-class words versus open-class words across P3’s responses. On average, closed-class Spanish words were used more frequently than open-class words.

Table 2.

Descriptive and inferential statistics on P3’s Spanish-English language mixing patterns.

Measure	Statistic	Open-Class Words	Closed-Class Words	Test Statistic	p-value
Proportion of word class usage in Spanish	Mean (SD)	35.78% (19.27)	55.52% (22.34)	Z = 4.713	<.001
Correlation between utterance length and language mixing frequency	Spearman’s p			0.05	0.272
Pause duration and language mixing frequency during language mixing	Mean (SD)	869 ms (1215)
Pause duration and language mixing frequency during non-language mixing	Mean (SD)	500 ms (680)

Note: ms = milliseconds.

Language mixing and utterance length

A Spearman’s rank-order correlation was conducted to examine the relationship between utterance length and frequency of language mixing. The analysis yielded a weak, non-significant correlation, indicating that variation in utterance length did not reliably predict the frequency of language mixing.

Language mixing and hesitation duration

To examine whether hesitation duration differed as a function of language mixing, we descriptively compared durations preceding utterances that contained language mixing to those that did not. As shown in Figure 1(C), hesitation duration prior to language-mixed utterances were longer than those preceding non-mixed utterances.

Figure 1.

Language mixing effects on lexical class proportions, hesitation durations, and hesitation behaviors. Note: A. the mean proportion of open- and closed-class Spanish words produced during language mixing. B. Average pause durations (in milliseconds) for utterances with and without language mixing. Error bars represent standard deviations. C. Spectrogram and corresponding waveform depicting hesitation behavior. LM = language mixing

Discussion

This study offers novel insights into PLM in bilingual aphasia by analyzing word class distribution, utterance length, and hesitation duration during elicited spontaneous speech in a bilingual individual with severe Broca’s aphasia. Notably, P3 demonstrated greater mixing of closed-class than open-class word- a pattern not typically reported in the PLM literature. Prior studies have more commonly documented cross-language intrusions in open-class words, often attributed to increased lexical retrieval demands, whereas closed-class items are generally considered language-specific and relatively resistant to switching (e.g., Goral et al., 2019; Lerman et al., 2019). Against this background, P3’s pattern raises questions about the locus of disruption.

Within the Matrix Language Frame (MLF) model, mixed utterances are structured by a Matrix Language, typically L1, that provides the morphosyntactic frame for bilingual processing (Myers-Scotton, 1993; Myers-Scotton & Jake, 2000). Because closed-class items are typically supplied by the Matrix Language and are relatively resistant to switching, disproportionate intrusion of such items from the non-target language may signal difficulty maintaining the intended grammatical frame. We therefore use the term morphosyntactic control vulnerability to describe reduced stability in sustaining the target language as the Matrix Language during production. Importantly, most prior reports of PLM have emphasized open-class intrusions, making P3’s pattern novel rather than expected.

At the same time, interpretation requires caution. P3 was a premorbid unbalanced bilingual yet demonstrated symmetric aphasia severity across languages post-stroke. Under lexical accounts, cross-language intrusions in the weaker language often arise when a speaker fails to retrieve a content word and instead accesses its translation equivalent in the stronger language (Goral et al., 2006; Lerman et al., 2019). However, the elevated rate of closed-class intrusions observed here suggests that difficulties may extend beyond lexical retrieval to instability in maintaining the grammatical frame itself. Because premorbid proficiency asymmetry and executive control demands may interact, we cannot fully disentangle morphosyntactic vulnerability from proficiency-driven influences. Accordingly, we interpret the closed-class effect as hypothesis-generating rather than definitive evidence of a distinct control deficit. These findings contribute to emerging evidence that PLM patterns are heterogeneous and not uniformly characterized by open-class mixing.

Also contrary to predictions, utterance length did not correlate with mixing frequency. Language intrusions occurred across various utterance lengths, suggesting that control failures may arise from local lexical access issues rather than increased planning load (cf. Green & Abutalebi, 2013). Furthermore, longer hesitation duration preceded retrieval of non-target language utterances, pointing to increased monitoring demands or response conflict (Mooijman et al., 2025). Although increased monitoring or response conflict offers one plausible interpretation, longer hesitations may also indicate lexical access difficulty in bilingual aphasia, particularly in individuals with non-fluent aphasia with lesions in the frontotemporal networks of the brain (Zhang et al., 2025). Under lexical retrieval accounts, prolonged pauses can arise when speakers struggle to access a target lexical item and subsequently activate a translation equivalent in the non-target language (Goral et al., 2006; Lerman et al., 2019).

These findings support hesitation duration as a sensitive marker of control disruption in bilingual aphasia. Prior work (e.g., Karpathiou et al., 2018) has demonstrated through bidirectional assessment in a bilingual individual without pathological language mixing (PLM) that hesitation may indicate underlying cognitive and linguistic processing demands. However, unlike those studies, the present investigation did not implement a fully bidirectional assessment of hesitation behavior across both languages. The elicitation task was intentionally structured to promote English-only production in accordance with P3’s clinical objective of reducing Spanish intrusions. Consequently, hesitation measures primarily captured pauses preceding intrusions into the non-target language rather than symmetric hesitation patterns across both languages. Future research should determine whether these hesitation effects reflect switching costs into the weaker language (e.g., Gollan & Ferreira, 2009) or instead represent a broader temporal signature of PLM that operates independently of switching direction.

Finally, P3’s AQ across languages also intersect with a growing body of work examining dual-language use in bilingual aphasia. Studies have shown that engaging both languages during assessment or intervention may increase overall verbal output, facilitate cross-language transfer, and in some cases support communicative recovery (Ansaldo et al., 2014; Ansaldo & Marcotte, 2007; Goral et al., 2025; Marcotte, 2008; Paplikar et al., 2014; Paplikar et al., 2019). These findings suggest that dual-language contexts can provide additional lexical access routes and communicative flexibility. At the same time, research on executive control in bilingual aphasia indicates that switching demands may place strain on regulatory mechanisms depending on individual profiles (Mooijman et al., 2022). In the present case, although P3 demonstrated higher combined English – Spanish performance under a dual-language framework, she reported that PLM substantially disrupted her everyday communication. Consequently, P3’s stated clinical goal was to reduce Spanish intrusions and improve English use within her home environment (Imaezue et al., 2025). From an evidence-based practice perspective, clinical decision-making requires integration of research evidence, clinical expertise, and client values (American Speech-Language-Hearing Association, 2005). Thus, increased aggregate output alone does not necessarily reflect communicative success if it conflicts with the individual’s functional priorities. Moreover, P3’s symmetric (bidirectional) mixing, with intrusions occurring from each language into the other, may reflect impaired language control (Imaezue et al., 2025) rather than adaptive translanguaging, as suggested in related work (Lerman et al., 2025). Given these findings, future research should therefore examine not only aggregate language output but also communicative effectiveness, patient and caregiver perspectives, and goal attainment across single-language and dual-language contexts.

Conclusion

This single-case study shows that PLM in bilingual aphasia can diverge from prevailing patterns, as evidenced by disproportionate closed-class intrusions, prolonged hesitations preceding non-target language utterances, and symmetric mixing across the two languages. Importantly, these behavioral features were consistent with the patient’s report that mixing substantially disrupted everyday communication, underscoring their clinical relevance. Together, the findings indicate that PLM is not uniformly driven by open-class lexical retrieval failures and suggest that both temporal hesitation and symmetric mixing may serve as candidate markers of impaired language control. Although necessarily preliminary, these observations motivate future studies in larger and cross-linguistic samples to test their generality and to clarify the mechanisms underlying PLM.