2 Theoretical background

This article lies at the interface of two well-known research paradigms. On the one hand, it adheres to the probabilistic grammar framework in that it assumes that grammatical knowledge is partially probabilistic and that multiple constraints operate simultaneously, sometimes with opposite effects, on the alternation between competing variants (e.g. Bresnan Reference Bresnan, Featherston and Sternfeld2007; Bresnan et al. Reference Bresnan, Cueni, Nikitina, Harald Baayen, Bouma, Krämer and Zwarts2007; Bresnan & Hay Reference Bresnan and Hay2008; Bresnan & Ford Reference Bresnan and Ford2010). Research in that spirit has shown that speakers are able to predict, with high accuracy, the odds of finding a particular linguistic variant in a particular context. This, in turn, entails that speakers’ grammatical knowledge must necessarily include intuitions about the underlying probabilistic constraints governing linguistic behavior. Bresnan and colleagues further show that grammatical knowledge is gradient and subject to restructuration as a result of changes in speakers’ experience with language, which is at least partly dependent on their sociocultural environment. The present study combines this probabilistic viewpoint with an interest in the connection between the structural characteristics of varieties of English and their sociohistorical background, in the spirit of the World Englishes framework (e.g. Schneider Reference Schneider2007; Mesthrie & Bhatt Reference Mesthrie and Bhatt2008). Crucial to the structural characteristics of varieties of English is the concept of nativization or indigenization. Nativization or indigenization refers to the process whereby speakers of postcolonial varieties make English their own, expressing themselves by means of ‘locally characteristic linguistic patterns’ (Schneider Reference Schneider2007: 6). Indigenization processes have been claimed to exist mainly at the lexis–syntax interface (Schneider Reference Schneider2003: 249): rather than inventing novel syntactic patterns from scratch, these postcolonial varieties of English are characterized by innovative combinations of lexical items and existing syntactic constructions.

A growing body of literature has recently emerged from the incorporation of the principles of probabilistic grammar into the World Englishes paradigm with the aim of exploring and delimiting the extent to which the strength of probabilistic constraints fluctuates across varieties of English (e.g. Rosenbach Reference Rosenbach2002, Reference Rosenbach, Rohdenburg and Mondorf2003; Hinrichs & Szmrecsanyi Reference Hinrichs and Szmrecsanyi2007; Bresnan & Hay Reference Bresnan and Hay2008; Szmrecsanyi & Hinrichs Reference Szmrecsanyi, Hinrichs, Nevalainen, Taavisainen, Pahta and Korhonen2008; Bresnan & Ford Reference Bresnan and Ford2010; Bernaisch et al. Reference Bernaisch, Gries and Mukherjee2014; Szmrecsanyi et al. Reference Szmrecsanyi, Grafmiller, Heller and Röthlisberger2016; Heller et al. Reference Heller, Szmrecsanyi and Grafmiller2017; Röthlisberger et al. Reference Röthlisberger, Grafmiller and Szmrecsanyi2017; Szmrecsanyi et al. Reference Szmrecsanyi, Grafmiller, Bresnan, Rosenbach, Tagliamonte and Todd2017; among others). Common to most of these studies is the observation that varieties share a fairly robust probabilistic grammar in that the constraints affecting a particular syntactic phenomenon are largely stable across varieties and fuel the same kind of syntactic choices. However, gradient regional differences seem to exist with respect to the strength with which such constraints impact speakers’ constructional choices in each variety. For instance, American English (AmE) and BrE speakers differ in that speakers of AmE favor the s-genitive over the of-genitive more strongly with inanimate possessors and as the length of the possessum increases than BrE speakers (Hinrichs & Szmrecsanyi Reference Hinrichs and Szmrecsanyi2007; Szmrecsanyi & Hinrichs Reference Szmrecsanyi, Hinrichs, Nevalainen, Taavisainen, Pahta and Korhonen2008). Similarly, Bresnan & Hay (Reference Bresnan and Hay2008) report that the animacy of the recipient impacts the choice of dative variant more strongly in New Zealand English (NZE) than in AmE, with inanimate recipients being more likely in the ditransitive dative variant in the former than in the latter variety.

In order to refer to these gradient regional differences, Szmrecsanyi et al. (Reference Szmrecsanyi, Grafmiller, Heller and Röthlisberger2016: 133) extended the notion of indigenization from the World Englishes paradigm to the probabilistic domain and coined the term probabilistic indigenization, which they defined as ‘the process whereby stochastic patterns of internal linguistic variation are reshaped by shifting usage frequencies in speakers of post-colonial varieties’. Probabilistic indigenization thus refers to a linguistic process that leads to statistical differences across varieties in the effects of probabilistic constraints. Szmrecsanyi et al. (Reference Szmrecsanyi, Grafmiller, Heller and Röthlisberger2016) argue that divergences in the odds of finding a given syntactic variant in a given context across varieties, even if these patterns are not stable, are evidence of the existence of variety-specific grammars tied to unique sociolinguistic backgrounds. Comparing the effect of probabilistic constraints in three syntactic alternations – the genitive, dative and particle placement alternations – across four varieties (i.e. BrE, Canadian English (CanE), IndE and SgE), they show that the four varieties largely share a common probabilistic grammar, since the effect direction of constraints remains stable across varieties. Nonetheless, quantitative differences emerge with regard to the strength of these constraints. For instance, they observe that the influence of a directional prepositional phrase after the verb phrase on the particle placement alternation was weaker in IndE than in the other varieties they studied. Interestingly, the three alternations in Szmrecsanyi et al. (Reference Szmrecsanyi, Grafmiller, Heller and Röthlisberger2016) turned out not to be equally sensitive to probabilistic indigenization effects, with the particle placement alternation exhibiting stronger variety-specific patterns than the dative and genitive alternations. Szmrecsanyi et al. associate this difference in sensitivity with the lexical specificity of the alternation in question: they conclude their study by suggesting that probabilistic indigenization effects arise as a function of the lexical specificity of syntactic alternations, with those that are strongly connected to specific lexical items being the ones most likely to exhibit cross-varietal indigenization effects. Their argumentation finds support in previous work in World Englishes that has shown that cross-varietal differences mainly emerge at the lexis–syntax interface. Similar tendencies were found by Szmrecsanyi et al. (Reference Szmrecsanyi, Grafmiller, Bresnan, Rosenbach, Tagliamonte and Todd2017), a study of dative and genitive variation in spoken language in four native varieties of English, namely AmE, BrE, CanE and NZE. Their data also shows that syntactic alternations are not equally homogeneous across varieties, with the dative alternation displaying stronger variety effects than the genitive alternation.

In the present article, we extend our previous knowledge of the probabilistic grammars of varieties of English in two ways. First, we propose a corpus-based variationist method for calculating the extent to which syntactic alternations display probabilistic indigenization effects: what counts is not if and/or how often people use particular constructions, but how – that is, subject to which probabilistic constraints – they choose between ‘alternate ways of saying “the same” thing’ (Labov Reference Labov1972: 188). Our approach is inspired by the Variation-Based Distance and Similarity (VADIS) method proposed in Szmrecsanyi et al. (Reference Szmrecsanyi, Grafmiller and RosseelMS), which assesses the degree of alternation-internal homogeneity across varieties of English along three lines of evidence, as proposed in the comparative sociolinguistics literature (see Poplack & Tagliamonte Reference Poplack and Tagliamonte2001: 92; Tagliamonte Reference Tagliamonte, Chambers and Schilling2002: 731):

1. 1. Statistical significance: Do the same constraints have a statistically significant effect across varieties?

2. 2. Size and direction: Are probabilistic constraints similar with respect to the size and direction of their effects? Are there any constraints that have, e.g. a stronger effect, or an effect in the opposite direction in one variety as compared to rest?

3. 3. Constraint ranking: Is the overall ranking of constraints homogenous across varieties? In other words, do the constraints have the same relative importance in all the varieties considered?

Second, we quantitatively test the hypothesis in Szmrecsanyi et al. (Reference Szmrecsanyi, Grafmiller, Heller and Röthlisberger2016) according to which an alternation's degree of probabilistic indigenization is proportional to its lexical specificity, that is, the more lexically specific an alternation, the more indigenized it is.

3 Data and method

For the purposes of this article, we investigate variation in the choice of syntactic variants in three varieties of English, namely BrE, IndE and SgE.Footnote ² BrE is an Inner Circle L1 variety, while IndE and SgE are Outer Circle L2 varieties which are considered to have reached phase 4, endonormative stabilization, in Schneider's (Reference Schneider2007) Dynamic Model. Despite their similarities regarding typological classification, IndE and SgE differ from each other in that the English in India has been described as being in a ‘steady state’ in which both progressive and conservative forces are at play (Mukherjee Reference Mukherjee2007: 158). Moreover, whereas the number of L1 speakers of SgE has been on the rise since the 1980s – in 2010 more than 32 percent of Singaporeans claimed that English was their dominant language (Leimgruber Reference Leimgruber2013: 9) – L1 users represent only about 0.25 percent of the total number of IndE speakers (Sharma Reference Sharma, Kortmann and Lunkenheimer2012: 523). Therefore, the set of varieties of English studied, despite being restricted in number, represents very different variety types, evolutionary stages, and even more fine-grained distinctions as regards the varieties’ social ecologies.

The data for the present study were extracted from the British (ICE-GB), Indian (ICE-IND) and Singaporean (ICE-SIN) components of the International Corpus of English (ICE). Each national component in ICE contains 500 texts, which amount to a total of approximately one million words: 600,000 words of spoken material and 400,000 of written material. A useful feature of ICE is that all its components follow the same design and annotation scheme, which makes them particularly appropriate for establishing comparisons between varieties. Interchangeable instances of the genitive, dative, particle placement, and subject pronoun omission alternations were retrieved as follows:

1. – In the case of the genitive alternation, being a relatively frequent syntactic phenomenon, 10 percent of the texts in ICE was enough for statistical analysis. A sample of texts containing text one, eleven, twenty-one and so on was created and then used to extract tokens of the genitive alternation in an automatic fashion. Appositive genitives, classifying genitives, double genitives, idiomatic genitives, partitive genitives and genitives involving indefinite possessums were excluded to make sure that both variants could have been used interchangeably (see Rosenbach Reference Rosenbach2002, Reference Rosenbach2014; Wolk et al. Reference Wolk, Bresnan, Rosenbach and Szmrecsanyi2013). This process yielded 3,108 genitive tokens (for further details on the genitive database, see Heller et al. Reference Heller, Szmrecsanyi and Grafmiller2017; Heller Reference Heller2018).

2. – Instances of the ditransitive and prepositional dative variants were retrieved using a list of dative verbs, shown in (5), adapted from previous literature (Levin Reference Levin1993; Mukherjee & Hoffmann Reference Mukherjee and Hoffmann2006; Bresnan et al. Reference Bresnan, Cueni, Nikitina, Harald Baayen, Bouma, Krämer and Zwarts2007; De Cuypere & Verbeke Reference De Cuypere and Verbeke2013; Wolk et al. Reference Wolk, Bresnan, Rosenbach and Szmrecsanyi2013). To ensure interchangeability, we excluded tokens involving particle verbs, passivized verbs, elliptical structures, coordinated verbs, clausal or non-overt constituents, beneficiary constructions, constructions containing a spatial goal and idiomatic expressions. In addition, extremely long recipients (more than 18 words) and themes (more than 23 words) were eliminated from the prepositional and ditransitive dative variants respectively. This rendered a database of 3,012 tokens of the dative alternation (for further details on the dative database, see Röthlisberger et al. Reference Röthlisberger, Grafmiller and Szmrecsanyi2017; Röthlisberger Reference Röthlisberger2018).

3. (5) accord, advise, allocate, allot, allow, answer, appoint, ask, assign, assure, award, bequeath, bid, bring, call, carry, cause, cede, charge, concede, convey, cost, deal, deliver, demonstrate, deny, develop, drop, entrust, explain, extend, feed, flick, flip, forward, get, gift, give, grant, guarantee, hand, impart, inform, issue, keep, lease, leave, lend, loan, lose, mail, name, offer, owe, pass, pay, permit, play, pose, post, prescribe, present, promise, propose, provide, quote, read, recommend, refuse, render, sell, send, serve, set, show, sing, slip, submit, suggest, supply, take, teach, tell, throw, toss, vote, wish, write, yield

4. – All verb-particle combinations including a transitive particle verb and one of the ten following particles were extracted from the corpus: around, away, back, down, in, off, on, out, over and up. Cases involving passive sentences, sentences with extracted direct objects, modified particles, names, titles, or other fixed expressions, and prepositional verbs were subsequently filtered out. In addition, instances involving pronominal direct objects and direct objects longer than six words were excluded. This process returned 2,480 tokens of the particle placement alternation (for further details on the particle placement database, see Grafmiller & Szmrecsanyi Reference Grafmiller and Szmrecsanyi2018).

5. – Due to the difficulty of automatically identifying all the relevant instances of omitted subject pronouns in the corpus, these had to be manually extracted. A balanced sample of 40 texts per ICE component was created by randomly selecting 10 spoken informal, 10 spoken formal, 10 written informal and 10 written formal texts, from which 1,229 interchangeable instances of omitted pronominal subjects were retrieved.Footnote ³ A random sample of 1,229 overt subject pronouns was then automatically obtained from the same texts, totaling 2,458 instances of both omitted and overt subject pronouns. Interchangeable tokens excluded non-referential omitted/overt subject pronouns, cases in which both the subject pronoun and the auxiliary verb of the clause were dropped, imperative sentences, and overt pronouns in tag questions (for further details on the subject pronoun omission database, see Tamaredo Reference Tamaredo2018).

The datasets were then annotated for several language-external and language-internal constraints although, for the purposes of the present article, we restricted our analysis to the five most important language-internal predictors of each syntactic alternation. This was a measure to ensure model convergence but, as the findings of Röthlisberger et al. (Reference Röthlisberger, Grafmiller and Szmrecsanyi2017) suggest, the most prominent constraints of a syntactic alternation are also the ones most sensitive to probabilistic indigenization effects (see also Grafmiller Reference Grafmiller2014). Therefore, we can be relatively confident that our set of constraints capture most potential differences between the varieties at hand. The five most important predictors were selected on the basis of per-alternation random forest analyses (see below) fitted to the whole dataset of all three varieties. Language-external constraints, such as register or medium of production, were excluded from this study because they did not consistently show up among the five most important predictors in all alternations. External factors have been shown to vary considerably across varieties of English, as they basically boil down to cultural and social differences (e.g. Heller et al. Reference Heller, Szmrecsanyi and Grafmiller2017), so including them would have potentially added extra heterogeneity to some alternations but not to others. Tables 1 to 4 display the five probabilistic constraints chosen for each syntactic alternation.

Table 1.

The five most important predictors of the genitive alternation

Table 2.

The five most important predictors of the dative alternation

Table 3.

The five most important predictors of the particle placement alternation

Table 4.

The five most important predictors of the subject pronoun omission alternation

In addition, the data were annotated for the lemmas of the specific lexical items occurring in each of the instances: genitive tokens were annotated for possessor and possessum head nouns, datives for verb lemma, recipient and theme head nouns, tokens of the particle placement alternation for verb lemma, particles and verb-particle combinations, and subject pronouns for the following main verb lemmas.

The degree of alternation-internal homogeneity across the three varieties was estimated in three steps. First, we fitted a mixed-effects binary logistic regression model and a random forest per variety using the same model formula per alternation. Statistical analyses were carried out in R (R Core Team 2017) using the glmer() function in the lme4 package (Bates et al. Reference Bates, Maechler, Bolker and Walker2015) for mixed-effects models and the cforest() function in the party package (Hothorn et al. Reference Hothorn, Buehlmann, Dudoit, Molinaro and Van Der Laan2006; Strobl et al. Reference Strobl, Boulesteix, Zeileis and Hothorn2007, Reference Strobl, Boulesteix, Kneib, Augustin and Zeileis2008) for random forests. Both mixed-effects models and random forests seek to predict the choice between variants of a syntactic alternation (e.g. s-genitive vs of-genitive) given a set of predictors (here: restricted to five) and, as we detail below, they enable us to assess the strength, direction and relative predictive importance of our predictors. While mixed-effects models make their predictions on the basis of a mathematical equation, random forests establish the usefulness of a predictor through trial and error. Mixed-effects models are well-suited for analyzing corpus data of the kind used here because they allow us to take into account the non-independence of our observations via random effects adjustments for, e.g., lexical items or speakers sampled. Mixed models thus provide more reliable generalizations about broader patterns beyond the specific lexical items or speakers observed in our datasets. Random forests, on the other hand, enable us to explore more idiosyncratic patterns within our datasets, e.g. non-linear effects and interactions. Random forests, as implemented in the most common packages, are not well suited to deal with so-called random effects, since they cannot handle categorical predictors with very large numbers of levels.Footnote ⁴ However, an advantage of random forests is that they are quite robust to common issues in linguistic analyses, such as data sparseness or predictor non-linearities (see also Tagliamonte & Baayen Reference Tagliamonte and Harald Baayen2012: 158–61 for details). The method averages over a defined number of conditional inference trees using random subsampling and a permutation scheme (see Strobl et al. Reference Strobl, Boulesteix, Kneib, Augustin and Zeileis2008 for details).

The computed mixed-effects models included the five most important language-internal predictors per alternation as fixed effects and the lemmas of lexical items as random effects (see Appendix). Due to the low number of some lexical items (and ensuing issues with model convergence), we grouped infrequent items together for the modelling process. For each alternation and random predictor, we identified the frequency value which separated lexical items into two groups: one containing 10 percent of the items, which occurred more often than the selected threshold, and another group containing the remainder 90 percent, which occurred less often than the specified frequency value. Lexical items in the low frequency group were subsequently bundled together, with the exception of particles in the particle placement alternation and verbs in the dative alternation which were not grouped due to their high frequency. No interaction terms were added. We are well aware of the fact that previous research on the alternations that we study often show robust interaction effects, although mostly between language-internal and -external predictors and hardly ever between language-internal predictors. Hence, we decided not to test for interaction effects in order to keep our models simple and because large models with many interactions often lead to serious fitting/convergence problems. Random forest model formulas comprised only the five most relevant constraints per alternation.

In a second step, we calculated the similarity between varieties along the ‘three lines of evidence’ (Tagliamonte Reference Tagliamonte, Chambers and Schilling2002; see section 2) using the method proposed in Szmrecsanyi et al. (Reference Szmrecsanyi, Grafmiller and RosseelMS):

1. Statistical significance: the number of shared significant and non-significant constraints in per-variety mixed-effects models (at p < 0.05, following Szmrecsanyi et al. Reference Szmrecsanyi, Grafmiller and RosseelMS).

2. Relative strength: the inverse of the (Euclidean) distance between the coefficient estimates in per-variety mixed-effects models (calculated without the intercept).

3. Constraint ranking: Spearman's rank correlation coefficient between the predictors’ variable importance values obtained from per-variety random forests using the varimpAUC() function in the party package (Strobl et al. Reference Strobl, Boulesteix, Kneib, Augustin and Zeileis2008).

Note that when calculating Euclidean distances using coefficient estimates, a change in the reference and predicted level(s) of the constraints and dependent variable might lead to different results. To overcome this potential issue, we transformed all binary predictors to numbers (e.g. recipient animacy = {animate, inanimate} changed to = {0, 1}) and centered the values around the mean (following Gelman Reference Gelman2008). Furthermore, the reference level and predicted levels were set the same for each alternation. We chose Euclidean distance instead of Spearman's rank correlation to calculate similarity between varieties on the basis of coefficient estimates because the latter does not take into account patterns across the sizes of the predictors’ effects, only their relative (absolute) size. Coefficient estimates with the same values but opposite signs would thus be maximally similar using Spearman's rank correlation, while Euclidean distance would recognize the distance between them.

For each of the three lines, we obtained one similarity score for each alternation separately by variety (see tables 6 to 9). On this basis, we calculated a mean similarity score for each line and each alternation averaging across all varieties (see table 5). And lastly, we calculated the mean similarity between the varieties for each alternation as a measure of its overall stability by averaging across all three lines, which we interpret here as reflecting the alternations’ degree of probabilistic indigenization across BrE, IndE, and SgE: the lower the value, the more heterogeneous the alternation and thus the greater its degree of probabilistic indigenization.

Table 5.

Similarity scores across alternations

Table 6.

Per-variety similarity scores of the genitive alternation for each line

Table 7.

Per-variety similarity scores of the dative alternation for each line

Table 8.

Per-variety similarity scores of the particle placement alternation for each line

Table 9.

Per-variety similarity scores of the subject pronoun omission alternation for each line

Finally, the alternations’ lexical specificity was operationalized on the basis of the concordance index C, which represents how well the model discriminates between the levels of the response variable. In order to tease out the lexical effects from the random structure in the mixed-effects models, we additionally computed fixed-effects only models per variety and alternation by means of the glm() function in R (R Core Team 2017). After computing the C statistic for both the mixed-effects and fixed-effects model, we subtracted the C statistic of the fixed-effect models from the C index obtained from mixed-effects models. To calculate C, we made use of the somers2() function in the Hmisc package (Harrell Reference Harrell2014). The resulting value indicates the increase in discriminative power from a fixed-effects only model to a model comprising both fixed effects and lexical items as random effects, thus signaling the importance of lexically specific constituents. We also considered an alternative heuristic to quantify lexical specificity by making use of R ² values. R ² is a goodness-of-fit statistic which is usually equated to the proportion of variance accounted for by the model: an R ² value of 1 would correspond to 100 percent of the variance accounted for by the model. An alternation's degree of lexical specificity could hypothetically be operationalized as the increase in R ² values from a fixed-effects only model to a model with both fixed effects and lexical items as random effects. This would in theory reflect the importance of random effects in the model and, therefore, indicate how strongly associated each alternation is with specific lexical items. However, we refrained from using R ² as a measure of lexical specificity since its interpretation is not as clear as in linear regression models, where it accounts for the proportion of variance in the response variable that is explained by the predictors (see Levshina Reference Levshina2015: 259). Furthermore, R ² values are usually lower in logistic regression than in linear regression models, even when they are equivalent in terms of goodness of fit. This is why the concordance index C is commonly reported in logistic regression analysis instead of R ² (e.g. Hosmer & Lemeshow Reference Hosmer and Lemeshow2000: 162), and why we chose to rely on lexical specificity values calculated on the basis of the former statistic.