Spoken syntax in a comparative perspective: The dative and genitive alternation in varieties of English

This paper introduces a new resource designed to facilitate the quantitative investigation of syntactic variation in spoken language from a comparative perspective. The datasets comprise homogeneously annotated collections of “interchangeable” (i.e. competing) genitive and dative variants in four varieties of English: American English, British English, Canadian English, and New Zealand English. To showcase the empirical potential of the data source, we present a suggestive analysis that investigates the extent to which the probabilistic grammar of genitive and dative variant choice differs across varieties. The statistical analysis reveals that while there are a number of subtle probabilistic contrasts between the regional varieties under study, there is overall a striking degree of cross-varietal homogeneity. We conclude by outlining directions for future research. This article is part of the Special Collection:  Probabilistic grammars: Syntactic variation in a comparative perspective


Introduction
In an effort to pool various pre-existing materials, work has been underway by the authors to create two comprehensive and homogeneously annotated datasets designed to facilitate the investigation of syntactic variation in spoken language from a comparative perspective. 1 Dataset 1 (henceforth: the "genitive dataset") covers the English genitive alternation, as in (1); dataset 2 (henceforth the "dative dataset") covers the English dative alternation with the verb give, as in (2). (1) The genitive alternation ( Szmrecsanyi, Benedikt, et al. 2017. Spoken syntax in a comparative perspective: The dative and genitive alternation in varieties of English. Glossa: a journal of general linguistics 2(1): 86. The datasets cover four spoken regional varieties of English (or groups of spoken varieties): dialects of British English (henceforth: BrE), US American English (AmE), New Zealand English (NZE), and Canadian English (CanE). In World Englishes parlance, all of these are "central" or "inner circle" varieties of English (Kachru 1992;Mair 2013), which as advanced native varieties have reached the ultimate stage ("differentiation") in Schneider's (2007) Dynamic Model of the evolution of postcolonial Englishes.
In this paper, we introduce the new datasets to the community, and summarize the sources, the definitions of the variable contexts, annotation procedures, and so on. We use these datasets to address the extent to which the probabilistic grammars of genitive and dative choice (by which we mean the set of constraints, their effect directions and effect strengths, and the constraint rankings) differ or resemble each other across varieties of English, and we discuss how these differences or resemblances speak to theoretical questions about the usage-and experience-basedness of (knowledge of) variation. We would like to stress at the outset that the paper is not intended as the "last word", but is instead designed to inspire readers to extend our analysis in various ways.
The remainder of the paper is structured as follows. In Section 2, we provide some background on the syntactic alternations we study. Section 3 introduces the datasets. In Section 4 we investigate the overall importance of constraints on variation in a crossvariety perspective. In Section 5 we conduct careful pairwise comparisons between the varieties under study for the sake of establishing probabilistic contrasts with regard to the effect of constraints. Section 6 discusses key findings, and sketches directions for future research.

Background: The genitive alternation and the dative alternation
Both the genitive and the dative alternations are positional alternations: by choosing competing variants, language users can switch the order of possessor/possessum and recipient/theme. As a result, the alternations are sensitive to a number of ordering principles such as the principle of end weight (Behaghel 1909;Eitelmann 2016), given-before-new (Clifton & Frazier 2004), and -more generally speaking -Easy First (MacDonald 2013).
Historically speaking, both alternations have been available in the grammar of English for a considerable while, though the frequencies of use of the variants have changed. For the genitive alternation, we know that the s-genitive was used in writing much more often than the of-genitive up until the twelfth century (Thomas 1931: 284). This pattern reversed in the Middle English period (Mustanoja 1960: 70), to an extent that the s-genitive appeared to be dying out (Jucker 1993: 121). However, the s-genitive underwent a functional expansion in the Early Modern English period "against all odds" (Rosenbach 2002: 184), allowing both variants to survive with reasonable frequency today. For the dative alternation, the prepositional dative variant was not widely available for most of the Old English period (Mitchell 1985;Traugott 1992), but constituent order in the ditransitive construction was variable (e.g. Kemenade 1987) and sensitive to essentially the same factors that constrain the dative alternation in Modern English (De Cuypere, 2010). The prepositional dative variant began to appear in Late Old English texts (Fischer 1992), but was initially subject to lexical restrictions (Allen 2006). It developed into "a fully productive alternative" (Fischer & van der Wurff 2006: 166) to the ditransitive variant during the Middle English period, though it is still less frequent than the ditransitive today.
As to probabilistic variation in the genitive alternation across speech communities, Hinrichs and Szmrecsanyi (2007) show that in written-edited-published standard English, non-animate possessors discourage s-genitive usage less strongly in AmE than in BrE, and that long possessums favor s-genitive usage in AmE but not in BrE. Hundt and Szmrecsanyi (2012) find that possessor animacy is overall a more important predictor in (earlier, written) NZE than in (earlier, written) BrE; Jankowski and Tagliamonte (2014) report that in vernacular CanE, the s-genitive is near-categorically used with animate possessors, while inanimate possessors exhibit more variation and apparent-time change. Heller et al. (2017) establish that the top constraints fueling probabilistic differences across a set of nine international varieties of English include possessor animacy, constituent length, and final sibilancy.
As for the dative alternation, probabilistic differences across varieties reported in the literature include the following. Bresnan and Hay (2008) find that non-animate recipients are more likely to be used in the prepositional dative construction in spoken NZE than in spoken AmE (Bresnan & Hay 2008: Figure 1); Wolk et al. (2013) report that end-weight of themes has a stronger effect in (written) AmE than in (written) BrE; Tagliamonte (2014) diagnoses greater use of the ditransitive dative variant in BrE as well as a greater degree of spread of the ditransitive dative variant into animate recipients in contrast to CanE where the prepositional dative variant is fairly stable. Finally, with a regression model comparing each of nine varieties (including BrE, CanE and NZE) to a collective "average" variety, Röthlisberger et al. (to appear) show, among other things, that non-pronominal recipients favor the prepositional dative significantly more so in CanE than in other varieties.

Dataset description
To construct the datasets under study in this paper, we aimed to combine smaller, preexisting materials (listed below) in a way that would maximize analyzability and comparability. To achieve this, we homogenized the definition of variable contexts as far as possible (excluding tokens that did not meet our guidelines for interchangeability) and standardized the annotation for the major constraints on variation (both across sub-datasets and across alternations). These efforts notwithstanding, two limitations should be spelled out at the outset. First, different sub-datasets span different ranges of real and apparent time, to the extent that this information is known. In the analysis to be presented in this paper, we take the liberty to ignore this variance, but encourage future research to include demographic information in statistical modeling, for the sake of investigating change in probabilistic grammars over time. Second, there are also some minor register differences across the sub-datasets, which we likewise ignore in the analysis to follow while acknowledging that they are potentially significant. Notice, for example, that the BrE genitive tokens were extracted from a corpus that samples oral history interviews; AmE genitives come from a corpus covering telephone conversations; and the CanE and NZE genitive tokens derive from sociolinguistic interviews. This and the more or less subtle topic differences these register differences entail is the likely reason, as we will argue in Section 5.1, why genitive rates (but not necessarily the underlying probabilistic grammars) vary to some extent across the sub-datasets.
With these caveats in mind, we stress that we investigated the datasets for any potential imbalances in data sampling, with a careful eye toward cross-sub-dataset asymmetries in the distributions of important predictors. In spite of the aforementioned limitations, we found a high degree of consistency among the four varieties with respect to, among other things, the distributions of possessor and possessum lengths and possessor animacy in the genitive dataset, as well as the distributions of the lengths and pronominality of recipients and themes in the dative dataset. To the extent that even small variations occurred, close inspection of the tokens themselves revealed no obvious patterns to indicate potential sampling issues. Nonetheless, we acknowledge that differences in sampling could skew our results, though the extent to which they may do so remains to be established. In what follows, we provide more information about the two datasets.

The genitive dataset
3.1.1 Sources AmE: Most tokens were extracted from the Switchboard corpus of American English (Godfrey, Holliman & McDaniel 1992), as described in Shih et al. (2015). The Switchboard corpus covers telephone conversations collected at the beginning of the 1990s. A secondary, smaller batch of AmE genitives which is in the dataset but not included in the analysis presented in this paper, derives from the Corpus of Spoken American English (CSAE) 2 and was first analyzed in Szmrecsanyi (2006: Chapter 5). (N = 1104 Switchboard; N = 319 CSAE) BrE: This sub-dataset derives from the Freiburg English Dialect Corpus (FRED) (Hernández 2006) and was first analyzed in Szmrecsanyi (2006: Chapter 5 Gordon, Maclagan & Hay 2007). (N = 1953) The more detailed documentation accompanying the datasets provides more information about the sources. Note that the original sub-datasets were adapted to match the joint dataset's specifications as regards the definition of the variable context and annotation for constraints.

Definition of the variable
Only potentially alternating genitive constructions were included. In particular, only determiner s-genitives (i.e. s-genitives which have specifying function; see the documentation accompanying the dataset, where the constructions included are described meticulously) 3 and of-genitive constructions beginning with a definite article were considered in the analysis. For the specific selection criteria of the individual sub-datasets we refer to the references given the documentation accompanying the joint dataset. As different studies (and datasets) tend to use their own selection criteria, these sub-datasets were post-edited to make them conform to uniform criteria; for the criteria we again refer to the dataset documentation. Such post-editing could of course only eliminate constructions from the sub-datasets but not retrieve data that was excluded in the original materials, so inevitably some differences between the sub-datasets remain. The main differences are documented in the dataset documentation. The NZE genitive tokens were extracted from the outset according to the guidelines mentioned above.

Predictors
In addition to language-external factors (variety, speaker ID, speaker sex [available for a subset of observations and not subject to analysis in the present study], speaker year of birth [available for a subset of observations and not subject to analysis in the present study]), genitive observations in the dataset were annotated for the following predictors, all of which are well-known determinants of genitive variation (see Rosenbach 2014 for discussion): • Response.variable: s-genitive versus of-genitive. The documentation accompanying the datasets provides more information about the technicalities. For the analyses to be presented in the following sections, semantic relation between constituents was modeled as a binary predictor, prototypical ((1) to (4) in the above scheme) versus non prototypical.

AmE:
The dative tokens were extracted from the Switchboard corpus of American English (Godfrey, Holliman & McDaniel 1992), as described in Bresnan et al. (2007). The Switchboard corpus covers telephone conversations collected at the beginning of the 1990s (N = 1190 -we reiterate that here and in the other sources attention is restricted to the dative verb give).
BrE and CanE: Datives were extracted from materials collected in the UK and Canada between 1997 and 2010 according to standard sociolinguistic procedures. In the UK the data come from York and small towns and villages all over the UK. In Canada, the data come from Toronto, as well as numerous small towns and villages in Ontario (see Tagliamonte 2014). We add that the Canadian data has an apparent time difference (BrE: N = 944; CanE: N = 1157).

NZE:
The observations come from the Origins of New Zealand English corpora (ONZE) (see Gordon, Maclagan & Hay 2007), and were collected separately but overlap substantially with those presented by Bresnan and Hay (2008). Note that the New Zealand dative sub-dataset includes historical data from early New Zealand English speakers, and so covers speakers born between 1851 and 1984 (N = 845).
The documentation accompanying the datasets provides more information about the sources. Note that the original sub-datasets were adapted to match the joint dataset's specifications as regards the definition of the variable and annotation for constraints.

Definition of the variable
Attention was restricted to the dative verb give. The definition of interchangeable ditransitive and prepositional dative variants broadly follows Bresnan et al. (2007), which essentially means that all instances of give with two argument NPs minus non-interchangeable constructions were considered. The documentation accompanying the dataset provides more information.

Predictors
In addition to language-external factors (variety, speaker ID, speaker sex [available for a subset of observations and not subject to analysis in the present study], speaker year of birth [available for a subset of observations and not subject to analysis in the present study]), dative observations in the dataset were annotated for the following predictors, which are well-known determinants of dative variation (see Bresnan & Ford 2010 for discussion): • Response.variable: Ditransitive dative versus prepositional dative. (In addition, the dataset also documents a small number (N = 8) of non-standard patterns of the type he had to give it her, which were excluded in the analysis to be presented in this paper). • Recipient/Theme.type: The annotation distinguishes between the following categories: (1) noun phrase; (2)  The documentation accompanying the datasets provides more information about the technicalities. For the analyses to be presented in the following sections, the predictors were simplified as follows: Recipient/Theme.type were reduced to a binary contrast: pronominal (

The bird's eye perspective: Establishing constraint importance across varieties
To set the stage, this section ranks individual constraints per variety and alternation according to explanatory importance, using Conditional Random Forest (CRF) analysis (see Tagliamonte & Baayen 2012 for an accessible introduction). A desirable property of CRF models is that they are robust to problems of predictor correlation, as their predictions are derived from ensembles of conditional inference trees grown on randomly sampled subsets of the data (Breiman 2001;Strobl, Malley & Tutz 2009). Each tree is built on a random subset of predictors whose explanatory power is assessed through random permutation of their values, which breaks the predictor's association with the outcome. A "permutation variable importance" measure for each predictor can then be calculated by measuring the decrease in accuracy of the model with the permuted predictor compared to the accuracy of the model with the non-permuted predictor, and used to rank predictors. We conducted an independent CRF run for each of the varieties, per alternation, utilizing the cforest() function in R's party package (Hothorn, Hornik & Zeileis 2006;Strobl et al. 2008).
The variable importance measures for the genitive alternation are shown in Figure 1. Possessor animacy is the most important predictor in all varieties by a wide margin; possessor length is the next most important predictor in all varieties except in CanE ( Figure 1c), a variety in which possessor animacy is supremely important but other constraints are largely not. According to CRF, other constraints have only a minor influence on genitive choice overall, and their relative ranking tends to vary across varieties. For example, it appears that final sibilancy is relatively important in NZE (Figure 1d), but much less so in AmE ( Figure 1a) and BrE ( Figure 1b). The corresponding variable importance rankings for the dative alternation are shown in Figure 2. Compared to the genitive alternations rankings, more constraints are crucially implicated in variation patterns; there also seems to be more variability between varieties. That said, recipient/theme type and theme definiteness are consistently among the top-ranked predictors, while recipient definiteness is rather unimportant, except to some degree in BrE. AmE ( Figure 2a) is different from the other varieties thanks to the important role that the length difference between recipient and theme plays. CanE ( Figure 2c) is special because recipient animacy is comparatively important.
In summary, we would like to highlight the comparative similarity of the genitive alternation rankings and the relative dissimilarity of the dative alternation rankings: in the genitive alternation, possessor animacy is the uncontested prime constraint (which is of course not surprising given the literature -see e.g. Rosenbach 2005) wherever we look, while the situation is more "pluralistic", as it were, in the dative alternation. However, some stability also exists across varieties for the dative alternation, in that recipient/theme type and theme definiteness are consistently among the top-ranked constraints.

The jeweler's eye perspective: Pairwise regression modeling
In this section we take a close look at differences between the varieties under study regarding the effect size and direction of the various constraints on genitive and dative choice. The analysis technique we use is binary logistic regression analysis with mixed effects (Gelman & Hill 2007;Zuur et al. 2009) as implemented in the lme4 package in R (R Core Team 2014; Bates et al. 2015). The investigation proceeded in a round-robin fashion, where the data from each variety were compared to the data from each other variety, resulting in a total of six regression runs per alternation comprising every possible pairwise combination. Compared to an alternative unitary model pooling all varieties for each alternation and incorporating interaction terms between variety and the language-internal constraints, our approach adds considerable resolution because it is more sensitive to pairwise probabilistic differences between varieties than the unitary-model approach, in which one would have to define a default variety or use other contrast designs.
Fixed effect predictors in regression analysis are those mentioned in the preceding sections. All models also initially contained an interaction of the variable "variety" (BrE vs. AmE vs. NZE vs. CanE) with each of the language-internal predictors. 4 No higher order interactions were included. The initial random effects structure, prior to pruning, consisted of random intercepts for speaker (idiolectal differences) as well as random intercepts for the head nouns of the possessor and possessum/recipient and theme NPs (note that following Wolk et al. 2013: 399, we collapsed all speaker and head noun levels that did not reach a threshold of at least five observations). The same initial model structure was fit to each of the six pairwise sub-datasets, and was then individually optimized following the top-down procedure outlined by Zuur et al. (2009: 120-122). Starting with the maximal model structure, the contribution of each random effect was first assessed via likelihood ratio tests. After settling on the optimal random structure, we removed non-significant interaction terms whose standard errors were larger than the absolute value of their coefficients. 5 Standard model diagnostics were then applied to each model, checking for overdispersion, leverage, residual structure and normality of random effects (Harrell 2001;Gelman & Hill 2007;Baayen 2008). Lastly, models were run through a 10-fold cross-validation procedure to check for the possibility of overfitting. Multicollinearity is not a serious issue, as the condition number κ in all pairwise regression models is below the threshold of 15 which indicates medium collinearity (Baayen 2008: 200).
In what follows we summarize pairwise comparisons that exhibit significant probabilistic contrasts. Detailed model outputs are available as supplementary materials (supplementary file 1).

Comparing BrE to CanE
We begin with a comparison of constraints on genitive choice among speakers of CanE and BrE. In these varieties, we find approximately equivalent proportions of genitive variants across the two varieties (Table 1). The CanE/BrE regression model (in which CanE serves as the default variety) fits the data very well (C = 0.98), and correctly predicts 93.2% of the observations, from a baseline of 63.1%. All language-internal constraints have main effects in the directions predicted by the literature (see e.g. Rosenbach 2014): animate possessors favor the s-genitive, as do definite possessors, and longer possessums; the pres-4 Preliminary data exploration of the genitives dataset suggested possible interactions of animacy and constituent length, so interactions of animacy with both possessor and possessum length were also included in the initial models. 5 This is a more conservative criterion for elimination than the usual p < .05, as it only removes predictors whose estimated effects are extremely small (or variance in the estimate is extremely large) and therefore unlikely to be affected by the presence/absence of other factors in the model. ence of a final sibilant disfavors the s-genitive, while a prototypical relation between possessor and possessum (e.g. kinship or ownership) favors it; and so on. This pattern is consistent across all the pairwise genitive models, so we will not discuss main effects hereafter.
As for probabilistic contrasts, the model reveals a probabilistic contrast between CanE and BrE with regard to the effect of possessor animacy: BrE speakers' genitive construction choices are less sensitive to possessor animacy (as measured by the probabilistic difference which an animate possessor makes, compared to an inanimate possessor) than those of CanE speakers.

Comparing BrE to NZE
We now turn to a comparison of BrE and NZE, where, unlike in the previous case, we do find substantial cross-varietal differences in the overall frequencies of genitive variants ( Table 2). According to our data, NZE speakers use the s-genitive notably less often than speakers of BrE (see this section's Interim summary for a discussion of this frequency differences).
The BrE/NZE model (default variety: NZE) fits the data very well (C = 0.986), correctly predicting 94.1% of the observations, from a baseline of 52.8%. Two language-internal constraints interact significantly with variety. First, as in the CanE/BrE comparison, we see a difference between BrE and NZE with regard to the effect of possessor animacy, which has a stronger effect in NZE than in BrE: what we are seeing is that BrE speakers' genitive construction choices are less sensitive to the difference between inanimate and animate possessors than those of NZE speakers.
Second, the effect of a final sibilant in the possessor is significantly stronger in NZE than in BrE, in that NZE speakers avoid the s-genitive in the presence of a final sibilant to a larger extent than do BrE speakers. The partial effects plot in Figure 3 visually depicts the contrast; note how the NZE line is steeper than the BrE line.

Comparing CanE to NZE
Let us set aside BrE for the time being, and turn to the third comparison for which we find cross-varietal differences: CanE versus NZE. In terms of overall frequency, here too we find varying proportions of genitive variants across the two sub-datasets (Table 3): CanE speakers use the s-genitive considerably more often than speakers of NZE do.
The fit of the regression model (default variety: CanE) is very good (C = 0.989; accuracy = 94.8%, baseline = 50.2%). There are two significant interactions between variety and language-internal constraints on genitive choice. First, similar to the BrE/NZE comparison the effect of a final sibilant in the possessor is significantly stronger in NZE than in CanE in that NZE speakers avoid the s-genitive in the presence of a final sibilant to a larger extent than CanE speakers do. Second, the effect of the semantic relations between the genitive constituents is significantly stronger in NZE than in CanE: when choosing genitive variants, NZE speakers are more sensitive to the difference between prototypical and non-prototypical semantic relations than CanE speakers are. The difference is visually depicted in the partial effects plot in Figure 4.

Remaining comparisons
Beyond the aforementioned interaction effects, we find no other significant differences between varieties in the three remaining comparisons (AmE/BrE, AmE/CanE, AmE/NZE). We thus take the liberty to skip a discussion of these models, and refer the reader to the supplementary materials instead.

Genitive variation: Interim summary
We begin by discussing why we see differences in genitive rates: recall that the s-genitive is strikingly more frequent in CanE (64.4%) and BrE (61.5%) than in AmE (40.5%) and NZE (35.1%). It seems that these differences primarily boil down to topic differences: as we have seen in Section 4, the most important constraint on genitive variation by far is possessor animacy. Animate possessors favor the s-genitive, and it so happens that in the sociolinguistic interviews and oral history materials that make up our CanE and BrE sub-datasets, people talk more about animate NPs than in the AmE and NZE sub-datasets: in the CanE sub-dataset, a full 66.8% of all possessors are animate, and in the BrE sub-dataset, the figure 65.2%. By contrast, the share of animate possessors is only 37.6% and 40.7%, respectively, in the NZE and AmE sub-datasets. These frequency differences are certainly interesting, but we stress that they do not necessarily speak to the issue of probabilistic variation patterns. Instead, we are dealing here with what Szmrecsanyi (2016) calls an "environmental" difference: the exact same probabilistic grammar will yield different variant rates if the input frequency of relevant stimuli (e.g. animate possessors) is sufficiently dissimilar. But such differences do not tell us much about the extent to which language users may or may not have different ways of choosing between variants. Turning to actual probabilistic differences, Table 4 summarizes the four significant probabilistic contrasts that have surfaced in the foregoing analysis. Animacy of the possessor is implicated in two of the five contrasts, and so we may conclude that possessor animacy is not only one of the most important constraints on genitive variation, but also one of the most variable (see also Rosenbach 2017). 6 Final sibilancy likewise takes part in two comparisons. More generally speaking, we see that the probabilistic genitive grammars under study exhibit subtle yet measurable differences not only with regard to the ranking of the constraints (as determined in CRF analysis -see Figure 1) but also in terms of effect size.
Looking at the varieties that are involved in probabilistic contrasts, observe that NZE is implicated in four of the five contrasts we diagnosed. One pattern, in this connection, is that final sibilancy tends to have a stronger effect in NZE than in some other varieties. The pattern that emerges with regard to BrE is that the influence of possessor animacy appears to be notably weaker in BrE than in other varieties.
What is the extent to which the contrasts uncovered in the pairwise analyses are consonant with the previous literature? To begin with, our finding that the effect of possessor animacy is stronger in NZE than in BrE ties in with Hundt and Szmrecsanyi (2012), however their significant effects were limited to collective and locative, rather than animate, 6 The interpretation of interaction terms involving possessor animacy is to some extent contingent on the fact that the default level for animacy was set to "inanimate" in regression analysis, hence the models quantify the difference that e.g. animate possessors make vis-à-vis inanimate possessors. That said, the distribution of genitives with non-animate possessors indicates differences that are not well captured in our regression modeling: in particular, the NZE data stand out in showing a great reluctance to use the s-genitive with non-animate possessors (collapsing collective, locative, temporal and other inanimate possessors) compared to the other varieties, as reflected in the relative frequencies of the s-genitive with non-animate possessors (NZE: 4.0%; CanE: 8.7%; BrE: 10.8%; AmE: 10.7%). A more in-depth study of this issue is reserved for another occasion.
possessors (again, vis-a-vis inanimate possessors). We will not speculate as to why our results are different from Hundt and Szmrecsanyi (2012) except to note that their study examined data of a different time period (late 19th and early 20th century) and medium (written) from ours. Next, we saw earlier that previous studies comparing AmE to BrE have reported a number of probabilistic contrasts, such as inanimate possessors discouraging s-genitive usage less strongly in AmE than in BrE, and/or animate possessor more strongly favoring the s-genitive in BrE than in AmE (Jahr Sohrheim 1980; Hinrichs & Szmrecsanyi 2007;Szmrecsanyi & Hinrichs 2008;Szmrecsanyi et al. 2014). The previous literature also suggests that in the written medium at least, long possessums favor s-genitive usage in AmE but not in BrE (Hinrichs & Szmrecsanyi 2007). However, our pairwise regression-based comparison between BrE and AmE fails to detect any significant probabilistic differences with regard to genitive choice. The most likely reason for this failure is that the present study examines spontaneous spoken data, while Hinrichs & Szmrecsanyi (2007) examined written-edited-published language from sections of the Brown corpora. Recent work has found sizeable differences in genitive usage across spoken and written modes and registers, even within the same variety (Grafmiller 2014), thus the results presented here should be interpreted within the larger context of what we know about genitive variation across different modes, registers and communities. 7 Finally, we note that none of the probabilistic contrasts that our analysis unearths emerge as significant in Heller et al. (2017), a recent study on probabilistic genitive variation that covers the same varieties that are represented in Table 4, albeit using different multivariate designs and on the basis of different, more acrolectal and less vernacular data sources (the International Corpus of English and the Corpus of Global Web-Based English). 8 7 As an aside, we note that in the BrE data a good number of interviewees are fishermen, who talk a lot about their fishing boats. Since Early Modern English, boats and ships have had a tendency to be encoded with the s-genitive (Breejen 1937: 55). Future study may want to investigate the extent to which this peculiarity of the BrE dataset may conceal actual probabilistic differences between BrE and AmE. 8 The only effect that seems to shine through, upon closer inspection, in the spoken-only subpart of the Heller et al. dataset is the final sibilancy difference between BrE and NZE, which happens to be the only contrast in Table 4 that is significant at the p < 0.01 level.

Comparing AmE to BrE
A frequency analysis reveals that the ditransitive dative variant is overall a bit more popular in BrE than in AmE (Table 5). The pairwise regression model fit on the combined sub-dataset (default variety: BrE) has a very good fit (C = 0.987) and correctly predicts 96.4% (baseline: 88.8%) of all dative outcomes. The constraints all have the effect directions predicted by the literature: longer recipients favor the prepositional dative variant, as do non-pronominal recipients, pronominal themes, indefinite recipients, and definite themes -and so on. This pattern is fairly consistent across all the pairwise dative models, so we will not discuss main effects in dative variation hereafter.
Regression analysis diagnoses one significant probabilistic contrast between BrE and AmE: in both varieties, non-pronominal recipients favor the prepositional dative variant (vis-à-vis pronominal recipients), but the effect is weaker in AmE than in BrE. The difference is shown in Figure 5 (notice the steeper line in the UK diagram).

Comparing AmE to CanE
Variant proportions are similar in CanE and AmE (Table 6). The pairwise regression model (default variety: CanE) has a very good fit (C = 0.980) and correctly predicts 96.1% (baseline: 85.3%) of all dative outcomes.
We find two significant interaction effects. First, in CanE communication uses of give disfavor the prepositional dative compared to transfer verbs whereas in AmE communication tokens actually favor the prepositional dative, vis-a-vis transfer tokens (this difference is visually depicted in Figure 6). Second, in both varieties longer recipients favor the  prepositional dative variant -as expected per the principle of end weight -but the effect is stronger in AmE than in CanE. In other words, the principle of end weight is a more powerful determinant of variant choice in AmE than in CanE. Figure 7 illustrates the difference.    Table 7 indicates that there are no major frequency differentials. The regression model (default variety: NZE) based on the combined NZE/AmE sub-dataset has a very good fit (C = 0.990) and correctly predicts 97.0% of all dative outcomes (baseline: 87.9%). The regression model identifies a significant interaction between variety and the semantics of give: in NZE, communication uses of give disfavor the prepositional dative compared to transfer verbs, but in AmE communication tokens actually marginally favor the prepositional dative, compared to transfer tokens. This is parallel to the pattern that we saw in the AmE/CanE comparison (see Figure 6).

Comparing BrE to CanE
Comparing variant rates in CanE and BrE, we find that the prepositional dative variant is somewhat more popular in CanE than in BrE (see Table 8). The regression model based on the CanE/BrE sub-dataset (default variety: NZE) has a good fit (C = 0.979) and correctly predicts 96.1% of all outcomes, against a baseline of 87.4%.
We find two significant interaction effects between variety and language-internal constraints. First, as recipient length increases relative to theme length, the prepositional dative variant is favored more strongly in BrE than in CanE. In other words, the principle of end weight is more powerful in BrE than in CanE. This is similar to the patterns we saw in the AmE/CanE comparison (see Figure 7). Second, vis-à-vis pronominal recipients nominal recipients favor the prepositional dative variant more robustly in BrE than in CanE. Although this effect is naturally correlated with that of recipient length (pronominal recipients always being very short), tests for data multicollinearity provide reasonable evidence that the two effects are independent.

Comparing BrE to NZE
In NZE, variant proportions are roughly similar to those in BrE (Table 9). The regression model (default variety: NZE) has a very good fit (C = 0.992) and correctly predicts 97.3% of all dative outcomes (baseline: 90.7%).   There is one significant probabilistic contrast: in NZE, communication uses of give disfavor the prepositional dative compared to transfer uses (the default category); there is also a tendency in NZE that abstract uses disfavor the prepositional dative. In BrE, by contrast, abstract uses favor the prepositional dative (again, compared to transfer uses), while communication uses tend to have no effect. Figure 8 visually depicts the interaction.

Comparing CanE to NZE
As can be seen from Table 10, the prepositional dative is moderately more frequent in CanE than in NZE. The regression model based on the combined sub-datasets (default variety: CanE) has a very good fit (C = 0.980) and correctly predicts 96.8% (baseline: 86.4%) of all dative outcomes.
The model diagnoses one significant interaction effect: the effect of length differences between recipient and theme is different in the two varieties. In both varieties, longer recipients favor the prepositional dative variant -as expected per the principle of end weight -but the effect is stronger in NZE than in CanE (similar to what we saw in the AmE/CanE comparison -see Figure 7). In other words, the principle of end weight is a more powerful determinant of variant choice in NZE than in CanE. Table 11 summarizes the eight significant probabilistic contrasts identified in pairwise regression modeling. Adopting a constraint-centered perspective, our findings concern-Figure 8: Partial effects plot of the interaction of variety with the semantics of give ("T" -transfer;

Dative variation: Interim summary
"A" -abstract; "C" -communication) in the AmE/CanE model. Log odds (Response.variable) are for the prepositional dative. ing the dative alternation may be summarized as follows. The semantics of the dative verb make a significant difference in three comparisons, and what we see is in fact different effect directions across varieties -for example, in AmE communication tokens of give tend to favor the prepositional dative, unlike in other varieties. It might be expected that one would see probabilistic differences like this particularly with constraints such as verb semantics, which are perhaps particularly amenable to cultural/regional (re-) interpretation. This set of contrasts clearly merits more scrutiny, and future research is encouraged to investigate the issue further. Length effects are likewise involved in three contrasts; here the very consistent pattern is that end weight effects are weaker in Canadian English than in all other varieties. Lastly, Recipient.type surfaces in two comparisons, and in each case the prepositional dative-favoring effect of non-pronominal recipients is stronger in BrE than in other varieties.
What is the extent to which these patterns are consistent with findings reported in the literature? Consider Bresnan and Hay (2008), who find -based on many of the same dative tokens that feed into the NZE/AmE sub-datasets under study here -that non-animate recipients are more likely to be used in the prepositional dative construction in spoken NZE than in spoken AmE (Bresnan & Hay 2008: Figure 1). Our analysis fails to replicate this effect, but we note that our data were coded differently, 9 and that we do not, in contrast to Bresnan and Hay, consider information status in the model. Moving on in the literature, we note that Wolk et al. (2013) find that theme length effects have a stronger effect in AmE than in BrE. We do not see this effect in our data, but we would like to caution that Wolk et al. studied written and historical data while we study contemporary spoken data. Tagliamonte (2014) reports two differences between CanE and BrE: first, a greater use of the ditransitive dative in BrE (a finding which we replicate), and second, a greater degree of spread of the ditransitive dative into animate recipients in BrE compared to CanE, where the prepositional dative was still robust up to the more recent generations. Our analysis was unable to confirm significant differences with regard to recipient animacy; instead we find two contrasts not reported by Tagliamonte (2014). It should be borne in mind however that the modeling techniques we used differ substantially from those employed in Tagliamonte (2014). How replicable are these findings? The BrE/CanE, BrE/NZE, and CanE/NZE contrasts detailed in Table 11 are not replicated in Röthlisberger et al. (to appear), a study on the probabilistic grammar of the dative alternation that covers, among others, the same varieties that are represented in Table 11 (however, the regression models are designed differently and the data sources subject to study -the International Corpus of English and the Corpus of Global Web-Based English -are more acrolectal and less vernacular than the data we study here).

Discussion and directions for future research
In this paper we introduced two datasets designed to facilitate the comparative investigation of spoken syntax. Based on this new resource, we investigated the extent to which patterns in the English genitive and dative alternations differ across four major regional varieties of English. Drawing inspiration from work in Probabilistic Grammar (Bresnan & Ford 2010), variationist (socio)linguistics (Labov 1982) and -more specifically -comparative sociolinguistics (Tagliamonte 2001), we took a closer look at the extent to which constraints on genitive and dative choice are fluid or stable in a cross-variety perspective.
Our key findings may be summarized as follows. We utilized conditional random forest analysis (CRF) to probe the overall importance of constraints, and found that the constraint rankings looked more similar across varieties in the genitive alternation (where possessor animacy is consistently the most important constraint by far) than for the dative alternation, where there is more cross-varietal fluidity. Subsequently, we used binary logistic regression analysis with mixed effects to investigate the effect size and significance of the constraints on dative and genitive variation. We specifically conducted careful pairwise regression comparisons, in which we were interested in potential interactions between variety and language-internal constraints. We found five such pairwise probabilistic contrasts in the genitive alternation, and eight in the dative alternation. Analogous to the CRF analysis, therefore, we observe that with regard to the sheer number of probabilistic contrasts the genitive alternation is more uniform in a cross-variety perspective than the dative alternation. In addition, we noted that some constraints on the dative alternation (semantics of the dative verb give) have different effect directions in different varieties, a sort of reversal that we do not see in the genitive alternation.
All things considered, we are struck by how similar and homogeneous the probabilistic grammars are across the spoken varieties under study in this paper, despite more or less subtle differences between the component sub-datasets on which our analysis is based. To highlight this homogeneity further and to provide some quantitative context for the discussion of the contrasting effect(s) later on, we take the liberty now to report two supplementary and deliberately very simple fixed-effects binary logistic regression models, one per alternation (see Tables 12 and 13). 10 Crucially, the models do not include any variety-related information, either as main effect or in interaction terms, and are thus 10 The predictors in these models are the constraints discussed in Section 3. The models were calculated using R's glm() function. The models were not optimized in any way.  entirely agnostic about variety differences. Nonetheless, most of the predictors are significant and have the theoretically expected effect directions (and the models have excellent fits). The point is that the parallel main effects of the constraints across the varieties can be interpreted as evidence for a solid probabilistic core grammar. In other words, while comparative analysis and the quest for contrasts took center stage in this study, probabilistic homogeneity seems to be the key trait of the datasets under study.
Against the backdrop of this overall homogeneity, we did identify several significant probabilistic contrasts (Tables 4 and 11), and the existence of such contrasts is expected if grammar indeed constitutes "the cognitive organization of one's experience with language" (Bybee 2006: 711), because such experiences vary across communities. For example, one of the constraints that is malleable across varieties is final sibilancy in the genitive alternation: the effect of final sibilancy is stronger in NZE than in BrE and CanE (see Table 4). It is interesting to note in this context that the avoidance of a clash in final sibilancy is a predictor of the genitive but not the dative alternation: the same speakers who use the of-genitive to avoid a possessor's final sibilant in the s-genitive, do not also avoid the ditransitive dative when the recipient has a final sibilant adjacent to an initial sibilant in the theme (Bresnan 2011). Avoidance of the final sibilant thus appears to be a (probabilistic) grammatical property of the genitive construction, which cannot be dismissed as a product of general articulatory constraints on adjacent sibilants. Hence our finding in the present study that some varieties are reliably differentiated by the likelihood of final-sibilance avoidance in the genitive is very strong evidence for probabilistic grammars of these varieties.
But in all, the number of probabilistic contrasts we observed is relatively small, taking into account the numerous comparisons between sub-datasets differing in geographical origin, speaker attributes, and collection methods. The degree of cross-varietal homogeneity as regards the strengths and directions of the various constraints is rather remarkable: in the big picture, it is the parallels across varieties that really stand out, and so we suggest the findings here are one further step toward identifying a model of syntactic variation that forms part of the "common core" (Quirk et al. 1985: 16) of the grammar of spoken English.
The homogeneity (again, despite some heterogeneity in the materials) we see in the models shows a common statistical pattern, termed "quantitative harmonic alignment" (Aissen 1999;Bresnan et al. 2007): One of the main findings of previous corpus work on the dative alternation is the existence of a statistical pattern in which, all else being equal, animate, definite, pronominal, discourse-accessible, and shorter arguments tend to precede inanimate, indefinite, nonpronominal, less discourse-accessible, or longer arguments in both of the dative constructions. (Bresnan & Ford 2010: 181) Likewise in the genitive alternation, animate, definite, pronominal, discourse-accessible, and shorter genitive constituents should precede inanimate, indefinite, nonpronominal, less discourse-accessible, or longer genitive constituents (Rosenbach 2014: 238-239). What we are seeing in our analysis is that the basic common phenomena of quantitative harmonic alignment are very robust indeed across the varieties, even as we have looked hard for probabilistic differences between varieties.
We were unable to find some contrasts in our speech-only materials that have been reported for written varieties (e.g. that long possessums favor s-genitive usage in AmE but not in BrE; see Hinrichs & Szmrecsanyi 2007). The reason for this, then, may well be that spoken vernaculars of the type covered in our datasets are more homogeneous than written varieties. This may seem surprising to those who would suppose that spoken language should be more heterogeneous in nature than written language, for which stylistic norms (which may be regional, however) seem to be in place. But it is probably not entirely unreasonable to assume that the production of spoken syntax -we emphasize that this study is not about lexis or accent differences -is subject to processing and production constraints and biases (e.g . Hawkins 1994;MacDonald 2013) in a way that the production of written syntax is not; hence the homogeneity. That being said, we believe that the cross-variety differences that we do find are consistent with the predictions of a framework assuming that subtle usage differences are learnt statistically. We know, for example, that "processing" concepts such as "working memory" in comprehension tasks are affected by training (i.e. experience) (see, e.g., Wells et al. 2009). Broadly viewed, then, statistical learning is a usage-based and/or experience-based framework, and the probabilistic contrasts that our modeling has revealed, although not exceedingly numerous, are consistent with predictions generated by this framework.
Contrary to our expectations, we also did not find many of the cross-variety contrasts reported in recent studies of spoken language (e.g. the animacy-related AmE/NZE contrast reported in Bresnan & Hay 2008). At the same time, we mentioned how recent follow-up research (Heller et al. 2017;Röthlisberger et al. To appear) fails to replicate contrasts that we have uncovered in this study. These replication failures are potentially troubling -it seems that comparative variation analysis of the type we have performed here is more sensitive to the particulars of the data source and coding decisions and to the design of the analysis than one would suspect or hope. In this regard, however, our empirical work makes a valuable contribution; our careful scrutiny of inconsistencies in coding decisions within and across varieties and constructions, by multiple authors, has yielded datasets that are both more principled and more comparable than those used in previous studies. Consequently, while the differences between previous and present results call into question the extent to which any set of results can be generalized, we remain confident that the picture gained from the present datasets is more accurate than the picture gained from previous datasets.
Nevertheless, the degree to which probabilistic contrasts are robust across data sources and research designs needs more attention. For one thing, we conducted multiple pairwise comparisons but -in a rather exploratory spirit -did not adjust the significance threshold accordingly. Thus (except for final sibilancy in the genitive alternation) many of our interaction effects do not reach Bonferroni-corrected thresholds for significance (p = .05/6 = .008), and so it is maybe not too surprising that many of the contrasts that we describe here fail to replicate in other studies. In general, then, it may be the case that the datasets we investigate are too small, and the predictor sets we are using too large, for regression analysis to reliably pick up and replicate probabilistic contrasts. Other analysis techniques, such as conditional inference trees (see also below), may be better suited to explore "small n and large p" datasets such as ours. Another issue that should be kept in mind in this connection is that we have taken the liberty in this study to not consider social factors (for which the datasets are partly annotated), but it is clear that in doing so we may have introduced confounds. Therefore, including sociolinguistic predictors in future analysis is high on the agenda.
Other directions for future research include the following. It would be highly desirable to provide a systematic comparison of different ways to model the data; these need not necessarily be regression-based, as in this paper, but may also include less widely used techniques such as memory-based learning (Daelemans & Bosch 2005) or naïve discriminative learning (Baayen 2011). As for designing models of the genitive alternation specifically, with possessor animacy being such a crucial (and in some cases near-categorical) constraint it may be worth considering taking possessor animacy out of the regression models at a later step in the analysis in order to zoom in on the attributes of the variation grammars that are actually divergent and highly variable (Tagliamonte 2014 is a recent study that uses this technique; the idea goes back to Labov 1969: 729, who argues that inclusion of (near-)categorical contexts will obscure the real patterns of variation). Alternatively, it may be worthwhile to turn to conditional inference trees (e.g. Hothorn, Hornik & Zeileis 2006), an analysis technique which is crucially based on this sort of data sub-setting. Lastly, future research should also study the possibility that the (subtle) differences across varieties are due to differential stages in the evolution of the syntactic variants themselves: the datasets on which the present study's analysis is based are an excellent resource to investigate how speakers of particular vernacular varieties are slower or quicker to adopt progressive syntactic variants in particular contexts (see Tagliamonte, Durham & Smith 2014).