Left dislocations across discourse types in monolinguals and bilinguals’ English

2.1 Left dislocation

Throughout the literature, left dislocation^[1] (LD) is described as a structure in which a constituent precedes its core clause (1). Specifically, a noun phrase occurs outside of its canonical position, to the left of the core clause or proposition. The argument position within the core clause is typically filled with a pronoun that is coreferential with the dislocated noun phrase (Gregory and Michaelis 2001; Prince 1984, 1997; Shaer and Frey 2004; Westbury 2016). Because the argument position is filled by the pronoun, the noun phrase is extra-syntactic and independent of the host sentence (Gregory and Michaelis 2001; Shaer and Frey 2004).

Three main sub-categories of LDs are described in the literature, which are distinguished by the type of coreferential element in the clause. In hanging topic LDs, the coreferential element can be a variety of different structures, including a strong or weak pronoun, an agreement morpheme, or an epithet (2a) (Anagnostopoulou 1997; Shaer and Frey 2004; Westbury 2016). Clitic LDs have a clitic pronoun in the argument position of the clause (2b) (Anagnostopoulou 1997; Westbury 2016). Contrastive LDs are similar to clitic LDs, but they have a d-pronoun, or weak pronoun, in the argument position of the clause (2c) (also called German Weak Pronoun LD) (Anagnostopoulou 1997; Frey 2005; Grohmann 2000; Shaer and Frey 2004). Notably, topicalization is not considered an LD; topicalization has a left dislocated referential expression, but there is no resumptive, coreferential pronoun in the clause (2d) (Gregory and Michaelis 2001; Prince 1984, 1997; Szűcs 2014).

Geluykens (1992) argues that a semantic classification of LDs is more effective than a syntactic one; in this way the superficial syntactic differences between the types of coreferential element can be ignored while the semantic similarities of LDs still allow them to be grouped. This then incorporates LDs in which the pronoun and noun phrase are linked through a partial rather than total coreferential relationship (3 and 4; also called Chinese-Style Topic Constructions) (Geluykens 1992; Westbury 2016).^[2]

LDs – regardless of the syntactic differences – have several functions in discourse. First, LDs are widely considered a topic promoting device (5) (Geluykens 1992; Givón 1983; Gregory and Michaelis 2001; Szűcs 2014; Westbury 2016; though see Prince 1984, 1997). Second, LDs are used to introduce new, not currently salient, or unexpected referents by removing them from the dispreferred subject position (6) (Gregory and Michaelis 2001; Manetta 2007; Prince 1984, 1997; Szűcs 2014; Westbury 2016). Third, LDs are used to indicate that the dislocated element is part of a partially ordered set that is already evoked in the discourse (7) (Gregory and Michaelis 2001; Szűcs 2014). As part of a set relation, LDs link the current utterance to previous discourse (Shaer and Frey 2004). In addition, LDs can be used to clarify the referent and avoid ambiguity (8) (Ashby 1988; McLaughlin 2011). The clarification function has mostly been discussed for French data and as a function of both left and right dislocation together (the example shows a right dislocation).

Various other functions of LDs have also been discussed in the literature although they are not directly relevant to this paper. These include, for example, contrast, turn taking (signal beginning of speaker’s turn), filler (hesitation with no clear pragmatic motivation), epithet (information about the referent), summary (regroups series of noun phrases), acknowledgement (echoing of referent in previous speaker’s utterance), and indication of politeness (Ashby 1988; Lange 2012; McLaughlin 2011; Tizón-Couto 2015, 2016). Discourse type restricts the functions of LDs, as some are interactional and would not occur in monologue-type speech or writing.

The use of LDs can be influenced by properties of the discourse, the speaker, or the sentential subject. LDs are more common in informal spoken discourse than in any other formality or mode (Geluykens 1992; Keenan 1977). English-speaking adults produce LDs most under casual, spontaneous circumstances and rarely in planned or written discourse (Keenan 1977). More recent findings by Biber et al. (2021: 948) suggest that LDs are most common in conversation and in fictional dialogues, while they are very rare in written prose. Research on written registers aligns with this conclusion – Tizón-Couto (2015) found that in present-day English (1900–1949), LDs and other left-detached sequences are most common in drama, fiction, and sermons, that is, in written registers that aim to recreate orality. For the written texts stemming from 1840 to 1914, it has also been demonstrated that speech-purposed writing (i.e., the writing that is meant to be performed or read aloud – drama, comedy, and sermons) contained a higher frequency of LDs than speech-like writing (i.e., writing of communicative immediacy – diaries, official and private letters), with mixed writing (prose fiction and trial proceedings) falling in between (Tizón-Couto 2017; see Section 2.2 of his paper for more references on the use of LD based on orality and informality). However, it remains open if these tendencies are present in today’s English.

As for the speaker properties, two studies on the use of LDs in English in two communities in Canada found differing patterns of use of LDs based on age in the two communities. In the primarily bilingual community, there was an age-graded pattern of use of LDs, meaning that middle-aged speakers used LDs least while younger and older speakers used LDs more. In the primarily anglophone community they found that the use of LD was relatively low by all speakers, and in the process of decline (Tagliamonte and Jankowski 2019, 2023). On the other hand, two other studies found no evidence of an effect of age or gender on the use of LDs (Gregory and Michaelis 2001; Lange 2012). Speakers’ language background can also play a role in the use of LDs: bilingual speakers of Indian English have been shown to have higher proportions of LDs than speakers of other English varieties (Lange 2012; Winkle 2015).

The animacy, complexity, and information status of the sentential subject have also been discussed as predictors of LDs. Some studies argue that proper names and humans occur in LDs significantly more than groups of humans and non-human subjects (Tagliamonte and Jankowski 2019, 2023). However, Snider and Zaenen’s (2006) study on LDs in a switchboard telephone corpus found no effect of animacy on the use of LDs overall. Instead, information status was a significant predictor of LDs with mediated entities (i.e., new but inferable (Nissim et al. 2004)) being dislocated more than new and old entities. Notably, the analysis on information status used a smaller subset of data than the analysis on animacy, and when animacy was investigated in just this small subset, it was a significant predictor of LDs. Tizón-Couto (2016) also investigated animacy and information status in written LDs in English historically (1500–1914). He found that LDs predominantly contain animate referents and that the majority of dislocated elements are mediated. Lastly, several studies investigated the complexity of the dislocated noun phrase in regard to the processing load of the structures. Tizón-Couto (2016, 2017) found that the length of the dislocated noun phrase is partially determined by the processing complexity of the LD. Longer noun phrases typically occurred with punctuation marks that clarified the overall structure. Winkle (2015) investigated the frequency of simple and modified noun phrases in LDs in several varieties of English with the expectation that modified noun phrases would be more common if LDs simplify the processing of an utterance. She found that modified noun phrases occur in 9.4 %–31.9 % of LDs depending on the English variety, with L1 English varieties generally having a higher percentage of modified noun phrases (pp. 115, 235; see also Lange 2012 for similar findings).

2.2 The current study

The current study examines if and how the use of LDs in English differs across the following discourse types: informal spoken, formal spoken, informal written, and formal written. We additionally ask (a) whether, and how, the use of LDs is conditioned by speaker characteristics and (b) whether LDs with particular characteristics are more likely to be used in certain contexts. We approach this question in several ways.

3.1 Participants and data collection

Participants are five groups of English speakers living in the United States: English monolinguals and bilingual speakers of English and one of four heritage languages (German, Greek, Russian, and Turkish). The majority of HSs (91.9 %) were first exposed to English at the age of 5 or earlier, with 45.7 % having the first contact with English from birth. Monolinguals were defined as those for whom English was the only language spoken at home, with no other language exposure before age 6. HSs are those who speak English as the majority language of their community as well as a heritage language which they learned from birth from at least one parent who is an L1 speaker of the heritage language. The participants are divided into two age groups – adolescent (13–18 years) and adult (20–37 years, Table 1).

Data collection followed the Language Situations methodology (Wiese 2020; Wiese et al. 2025), which elicits controlled, comparable, and quasi-naturalistic productions across formalities and modes. Participants watched a short video depicting a minor car crash and then recounted what they saw as if they were a witness to the accident. This procedure was completed in a formal context and an informal context, as well as in spoken and written modes. For the formal context, the participant recounted the video as a voice message to a police hotline and as written police report. For the informal context, the participant recounted the video as a WhatsApp voice message and a WhatsApp text message to a friend. Thus, each participant produced four narratives.

Based on these narrative data, we assessed the English proficiency of participants using speech rate (syllables per second) and two measures of lexical diversity: moving-average type-token ratio (MATTR) and measure of textual lexical diversity (MTLD; see Azar et al. 2020; Nagy and Brook 2020 for the association between speech rate and proficiency in HSs; Kyle et al. 2024 for the connection of lexical diversity and proficiency in L2 speakers). We found no difference in proficiency of the groups based on participants’ speech rates. As recommended in the literature (Zenker and Kyle 2021), we calculated MATTR and MLTD only on the narratives with at least 50 tokens – 1065 out of 1148. Contrary to the speech rate findings, Turkish HSs differ from English monolinguals in both lexical diversity measures. The other HSs groups do not differ from the monolingual speakers. Table 2 includes the three proficiency measures for each speaker group (group values and SEs predicted by linear models); for the models and data see the OSF repository at https://osf.io/ygk6m/?view_only=5c080b841a2b495aac73771bbd74dc3f, and for the spoken files that were used for the speech rate calculation, see release 0.4.0 of the Research Unit on Emerging Grammars (RUEG) corpus (Wiese et al. 2021).

All narratives^[4] were divided into communication units (CU) – a main clause plus any dependent clauses^[5] (Schneider et al. 2005) using EXMARaLDA software (Schmidt and Wörner 2014). CUs served as the unit of comparison for our study, since each CU could contain a maximum of one LD. The mean length of the narratives was 12.2 CUs (with a median of 11 and a range of 2–57). An example of an informal, spoken narrative is shown in (9), including the division into CUs and further annotations. The annotated data are accessible through the OSF repository.

3.2 Annotation

We first identified all LDs in the corpus according to the definition in Section 2.1. We then annotated each LD for four features as detailed in the following paragraphs: referent, type of pronoun, type of noun phrase, and function.

3.3 Analysis

Our analyses fall into three broad categories – confirmatory (narrative and speaker characteristics), exploratory (interaction of speaker characteristics with formality),^[7] and descriptive (LD characteristics). The data, R code, and Excel workbook that can be used to reproduce all analyses can be accessed through the OSF repository.

3.3.1 Confirmatory analysis: narrative and speaker characteristics

We first analyzed the data by the narrative characteristics of formality and mode and by the speaker characteristics of age and gender to see if the effect of these variables, as reported in the previous literature, manifests itself in our data. We coded each CU as containing an LD or not and then predicted the presence of LDs in CUs using generalized binomial linear mixed effects models employing the lme4 package (version 1.1-30, Bates et al. 2015) in R (R Core Team 2022). This type of modeling assumes that each CU provides exactly one chance for an LD to appear (e.g., “the man, he dropped the ball”) since more than one LD is generally not possible in a CU (e.g., “*the man, the ball, he dropped it”).^[8] Table 3 (Section 4.1) shows the raw numbers of LDs and CUs that the models were based on, and Figure 1 represents the proportions of LDs used by each speaker in each narrative (formal spoken, formal written, informal spoken, informal written), with speakers split by age and gender.

Figure 1:

By-speaker proportions of LDs out of CUs by age group and gender, split by formality and mode. Colored dots represent speakers, one dot = one speaker.

We included four predictors as fixed effects: formality (formal/informal), mode (spoken/written), age group (adult/adolescent), and gender (female/male), since these variables were explored in the literature (Geluykens 1992; Gregory and Michaelis 2001; Lange 2012; Tagliamonte and Jankowski 2019, 2023). We also included an interaction of age and gender since previous sociolinguistic research indicated that differences between men and women in the use of certain phenomena can be modulated by their age group (e.g., Labov 2001: 303–132). We did not include an interaction of formality and mode because there were only three data points in the written mode.

We used treatment contrast coding for the predictors, with formal setting, spoken mode, adult age group, and female gender as reference levels. We performed model selection as described in Gries (2021):^[9] first, we maximally specified the random effect of participants by including a by-participant random intercept and mode and formality random slopes. Next, we performed a step-by-step reduction of random effects: we removed the random effect that explained the least variance and compared the model fit of the new reduced model with the model fit of the previous model using ANOVA at each step. When the random effects could not be simplified anymore (i.e., any further reduction resulted in a worse model fit), we moved on to the second step – the reduction of fixed effects using drop1() function. We removed fixed effects until the reduced model had a significantly worse fit than the previous model.

We performed several diagnostics for the final model: we calculated a marginal R squared value using the r.squaredGLMM() function from MuMIn package (Bartoń 2024) to assess how much variance in the outcome variable was described by the model. Additionally, we checked the models for multicollinearity (using the vif() function from the car package; Fox and Weisberg 2019), overfitting (with isSingular() function), and overdispersion (using a custom function overdisp.mer() from Gries 2021: 439).

4.1 Confirmatory analysis: narrative and speaker characteristics

In this analysis, we examined the frequency of LDs by formality, mode, gender, and age group. We also included an interaction of age and gender. Table 3 shows the overall proportions of LDs out of CUs for the four combinations of gender and age group, split by formality and mode. Figure 1 shows individual by-speaker proportions of LDs out of CUs for each combination of gender and age group, split by formality and mode.

In the inferential analysis, the gender and age group predictors were dropped during the model selection, meaning that they did not significantly contribute to predicting LD use. The final model contained mode and formality as fixed effects and a formality random slope uncorrelated with a by-speaker random intercept. In this final model, we observed a strong main effect of mode (est. = −3.4129, SE = 0.6941, z = −4.917, p < 0.001), with fewer LDs in the written mode compared to the spoken mode. In addition, there was a main effect of formality (est. = 0.9293, SE = 0.2846, z = 3.266, p = 0.001), with more LDs in the informal setting than in the formal one (Figure 2). The fixed effects accounted for 35.8 % of variance in the outcome variable; no issues with multicollinearity, overfitting, or overdispersion were detected.

Figure 2:

Predicted probabilities of LDs by formality and mode and individual proportions of LDs out of CUs. Colored dots represent speakers; black dots with whiskers represent predicted probabilities of LDs based on the linear mixed effects model. The y-axis is zoomed from the original size of 0–0.3 to the size of 0–0.15 to make the model predictions more visible. When zooming, 10 data points above the 15 % mark were removed: two in formal spoken, eight in informal spoken.

4.2 Exploratory analysis: interaction of speaker characteristics with formality

Our second analysis examined the interaction of formality with the speaker characteristics of gender, age group, and speaker group. Mode was included as a control variable. Table 4 shows the overall proportions of LDs out of CUs for the five speaker groups, split by formality and mode. Figure 3 shows the individual by-speaker proportions of LDs out of CUs by speaker group, split by formality and mode, along with the predicted probabilities of LDs based on the linear model.

Figure 3:

Predicted probabilities of LDs by speaker group, split by formality and mode and individual proportions of LDs out of CUs. Colored dots represent speakers; black dots with whiskers represent predicted probabilities of LDs based on the linear mixed effects model. The y-axis is zoomed from the original size of 0–0.3 to the size of 0–0.15 to make the model predictions more visible. When zooming, 10 data points above the 15 % mark were removed: two in English MSs, three in German HSs, two in Russian HSs, and three in Turkish HSs.

In the inferential analysis, the final model’s fixed effects included mode, formality, and speaker group, as well as an interaction of formality and speaker group. Random effects included only a by-speaker random intercept. The significant interaction of formality and speaker group indicated that the formality slope was different in Greek HSs and Turkish HSs, compared to English monolinguals (Greek HSs est. = 1.863, SE = 0.777, z = 2.396, p = 0.017; Turkish HSs est. = 2.622, SE = 0.778, z = 3.370, p < 0.001). The simple effect of formality within English monolinguals indicated no evidence of the difference between formal and informal setting in English monolinguals’ productions (est. = −0.66, SE = 0.493, z = −1.338, p = 0.18). The fixed effects accounted for 42.5 % of variance in the outcome variable; no issues with multicollinearity, overfitting, or overdispersion were detected.

To see if there is a difference between the formal and informal settings in the four HS groups, we performed pairwise comparisons using emmeans() function and custom contrasts from the emmeans package (version 1.8.9, Lenth 2023). The resulting comparisons (with Holm-adjusted p-values) indicated that Turkish HSs had a significantly lower proportion of LDs in the formal setting than the informal one (est. = −1.961, SE = 0.601, z = −3.261, p = 0.004), while the other HS groups did not make a significant distinction between the formalities.

Overall, the results of the two inferential analyses show that mode is an important predictor of LD use, with the spoken mode having more LDs than the written one, as predicted based on Geluykens (1992). Formality also plays a role – when taken as a non-interacting predictor, we observe that the informal setting has more LDs than the formal one, as predicted based on the claim by Geluykens (1992). However, our exploratory analysis indicated that the formality effect is modulated by the speaker group: it is actually Turkish HSs that make the expected distinction between the formality settings (significantly more LDs in the informal one). In contrast, English monolinguals, along with German, Greek, and Russian HSs, do not show evidence of the significant formality distinction – the speakers in our sample used similar proportions of LDs in formal and informal settings. Finally, we observed no evidence of gender and age group effects, similar to Gregory and Michaelis (2001).

4.3 Descriptive analysis: LD characteristics

Lastly, we analyzed the LD characteristics – noun phrase, pronoun, referent, and function – used by formality. The frequency of simple and modified noun phrases in LDs by formality is shown in Table 5; the frequency is represented as a proportion^[10] of LDs of each type out of the total LDs used in the informal and formal contexts, as well as overall. In general, speakers mostly use modified noun phrases, and this difference is more extreme in the formal than informal context. The frequency of LDs with modified noun phrases is consistent with the idea that LDs may simplify the processing of a complex utterance or sentence (Lange 2012; Tizón-Couto 2016, 2017; Winkle 2015). However, our results show a higher frequency of modified noun phrases than Lange (2012) and Winkle (2015)’s results.

The frequency of different pronoun types in LDs by formality is given in Table 6. Personal pronouns are used most frequently, followed by possessive pronouns.

10 of the 19 referents in the data were produced at least once using an LD, and we categorized these referents as animate (man, woman1, couple, family, woman2, driver1, and drivers) and inanimate (car1, car2, cars). Table 8 shows the proportion of LDs with animate and inanimate referents out of the total LDs for each formality. We additionally determined the frequency of these 10 animate and inanimate referents overall in the data (Table 7). Animate referents are used most frequently in LDs in both formalities. Animate referents are also used more frequently in the narratives overall; however, this difference is more exaggerated in the LDs.

This suggests that the use of a referent in an LD depends on a combination of referent animacy and overall frequency of the referent (see analysis by each referent in OSF repository), which is consistent with Tagliamonte and Jankowski’s (2019, 2023) findings that a major predictor of left dislocation in English and French is subject animacy (see also Snider and Zaenen 2006; Tizón-Couto 2016, 2017).

Lastly, the frequency of the functions of LDs by formality is provided in Table 8. LDs are used mostly to reintroduce referents and to introduce new referents in the narrative.

Re-introduction is the most frequent function of LDs in both formalities. While we did not explicitly investigate information status of the dislocated referents in general, this suggests that overall LDs express old information rather than new information. This contrasts with Winkle’s (2015: 110) findings that dislocated referents more frequently contain new information in L1 varieties of English, but more frequently contain old information in L2 varieties of English. In L1 varieties, 34.7 %–46.2 % of LDs have a new referent, and 37.1 %–42.1 % have an old referent. In L2 varieties, 23.1 %–30.2 % of LDs have a new referent, and 43.2 %–61.5 % have an old referent. (p. 234). Thus, our results align more closely with the results for L2 English varieties. To gain more insight into if this may be an effect of bilingualism, we compared the use of LDs by function and by speaker group, as shown in Figure 4. For this analysis, we combined the functions of clarification and set as “other,” because we did not take the referents’ information status into account for those functions.

Figure 4:

Percentage of LDs of each function by speaker group.

German HSs are the only speakers to use more LDs with a new introduction function than with a re-introduction function. English MSs and all other HS groups use more LDs with a re-introduction function. These findings are substantially the same when additionally considering formality (see OSF repository for full analysis).