The Yale Grammatical Diversity Project: Morphosyntactic variation in North American English

1.1 Our approach to the study of morphosyntactic variation

Generative linguists are primarily interested in the mental grammar of individual speakers, modeled as a system of rules that can form some linguistic units (syllables, words, sentences, etc.) but not others. While it is obvious that there are systematic grammatical differences across speakers of different languages or dialects, differences also exist among people within their speech communities (see Trousdale and Adger [2007] and Cornips [2015] for some interesting perspectives on this issue). If every individual has a mental grammar, we might expect that what it means to “speak the same language” is to have mental grammars that are similar but not necessarily identical (and, of course, to share a significant proportion of the lexicon).

Generative microcomparative syntax, then, is the study of the differences between similar mental grammars, with the goal of furthering our broader understanding of the human language faculty. As Kayne (2005: 283) points out, “microcomparative syntax work provides us with a new kind of microscope with which to look into the workings of syntax.” Several projects conducted within this paradigm are listed at www.dialectsyntax.org, including ASIS (Poletto and Benincà 2007), SAND (Barbiers et al.2005, Barbiers et al.2008), and ScanDiaSyn (http://uit.no/scandiasyn/?Language=en), in Italy, the Netherlands/Belgium, and Scandinavia, respectively.

For our project, we recruit (and in some cases develop) methodologies especially suited to the kinds of questions we are interested in asking. We collect data in the form of acceptability judgments in order to determine which kinds of sentences can be generated by individual speakers’ mental grammars and which cannot. In addition to eliciting judgments from speakers of particular varieties, we also conduct large-scale surveys, map the results of those surveys geographically, conduct statistical tests taking geography and other social variables into account, and look for theoretically significant linguistic correlations. In all cases, we do this with the primary goal of understanding variation between speakers at the individual level.^[1] While our goals and methodologies are informed by our theoretical perspective, we expect that our work and results will be of interest to linguists working in other frameworks and even to the public more generally. Therefore, in addition to our technical theoretical work, we are committed to highlighting our descriptive findings and providing freely accessible resources intended for a broader audience.

1.2 Situating the YGDP within the study of variation in English

From an empirical perspective, we are interested in morphosyntactic variation across the varieties of English spoken across North America. Interspeaker variation within North American English has long been an object of linguistic investigation (see Schneider [2008] and Wolfram and Schilling [2016] for useful overviews). However, most studies of morphosyntactic variation, as noted by Kortmann (2003), have focused on particular phenomena such as negative concord, positive anymore, subject–verb agreement, and multiple modals (Reed and Montgomery 2016), or on particular varieties such as Alabama English (Feagin 1979), Appalachian English (Wolfram and Christian 1976; Montgomery and Hall 2004b; Hazen et al. 2013), African American English (Labov et al. 1968; Baugh 1983; Rickford 1999; Green 2002; Lanehart 2015), and Canadian English (Tagliamonte 2006).

When it comes to large-scale investigations (considering multiple phenomena across multiple varieties), researchers of North American English have conducted surveys, compiled corpora, or collected data from a variety of resources and expert contributors. The most prominent nationwide surveys of individual speakers of American English have focused on phonology (Labov et al. 2006), the lexicon (Carver 1987; Hall 2013), or both (Kurath et al. 1939–1943 and subsequent work on the Linguistic Atlas of the United States and Canada;^[2]Vaux and Golder 2003). Grieve (2009, 2016) has created a large corpus of letters to the editor from regional newspapers all over the country, which he uses to conduct statistical and geographical analyses of lexical variation.^[3]

Perhaps most comparable to our project in empirical focus and scope is a large body of work (Kortmann and Schneider 2004; Kortmann 2005; Kortmann et al. 2005; Szmrecsanyi and Kortmann 2009; Hernández et al. 2011; Hickey 2012; Kortmann and Lunkenheimer2012; Kortmann and Lunkenheimer2013; Siemund 2013; Gerwin 2014) that considers questions of morphosyntactic variation in English within a research paradigm that integrates functional typology with dialectology. This approach is sometimes called variationist typology or sociolinguistic typology (Siemund 2013, 283). These researchers provide systematic overviews of the features attested (along with the degree of attestation) in particular varieties of English and useful summaries of existing literature on these topics. In many cases, they utilize quantitative techniques to reveal generalizations that illuminate how different varieties of English reflect crosslinguistic tendencies. Moreover, Kortmann and Lunkenheimer (2013) (eWAVE) provides an interactive interface to explore, compare, and geographically visualize the presence of certain features across varieties of English worldwide.

Our project differs from this work in several respects. First, our empirical scope is distinct: we focus exclusively on North American English, whereas most of the work in this line of research is on English as spoken in the British Isles and around the world.^[4] Second, variationist typologists approach dialect variation from a top-down perspective: they begin with a list of varieties of English and a corresponding list of experts on these varieties, who are responsible for determining the prevalence of a given syntactic property or phenomenon in each variety.^[5] Our approach, on the other hand, does not take specific varieties as a given; we are trying to discover what they are bottom-up. In order to do so, we quantify over individual speakers’ acceptability judgments on specific sentences rather than over varieties or over categorical ratings of the prevalence of a grammatical property as assigned by researchers at the level of the variety. Third, while many of our goals overlap with those of the variationist typologists (e.g., mapping the geographic distribution of morphosyntactic phenomena that vary across English speakers/varieties), one key goal of the typological investigation is to evaluate whether varieties with distinct historical trajectories or contact scenarios (e.g., first-language varieties versus second-language varieties versus English-based pidgins and creoles) have fundamentally different structural properties. By contrast, we are interested in discovering how observations gleaned from the study of English morphosyntactic variation can answer formal questions framed within generative syntactic theory. In sum, both lines of research seek to explore morphosyntactic variation in English, but our scope, methods, and goals are complementary.

1.3 Roadmap

Our next section outlines the overall goals of our project in more detail, and then we move on (Section 3) to discuss methodological issues, including the use of geographical maps of our findings. Finally, we describe in broad strokes some of the results we have obtained so far (Section 4) and some of the ways we have been able to share our findings with others, inside and outside academia (Section 5).

3.1 Examining existing sources

In our efforts to document grammatical variation within North American English, we have examined a broad swath of relevant literature on English morphosyntactic variation (http://ygdp.yale.edu/references) in order to compile a list of phenomena, build a bibliography, and collect examples. We have also explored less formal forums such as blogs and social media sites in search of the same kinds of information – we have found that these can be a particularly rich source of data and commentary on understudied and/or new phenomena.

We store bibliographic metadata and PDFs of our sources in a group Zotero database, which allows us to search across all our sources, tag and filter sources by the phenomena they discuss, and export references in whatever format is required for a particular publication venue. The bibliography is publicly accessible at https://www.zotero.org/groups/yale_grammatical_diversity_project/items.

3.2 Plotting attested examples

We use Google Fusion Tables to catalog the attested examples we have found in the literature. Each example is associated with metadata including the source where we found it, the phenomenon (or phenomena) it illustrates, the nature and date of attestation, the acceptability of the sentence for the speaker, and the speaker’s speech variety, ethnicity, age, socioeconomic class, locale, and region. We assign geographic coordinates to each example based on the descriptions given in the original source, which allows us to display these examples in interactive Google Maps embedded on our website. Each example is displayed as a pin on the map, and hovering over a pin will reveal the metadata available for that example.

3.3 Administering surveys online

One key methodology we have used to gather new data is to administer acceptability judgment surveys online.

3.4 Conducting geospatial analysis

Our survey methodology lends itself nicely to studying geographical variation in acceptability judgments, since we get participants from all across the country. This benefit does come with an analytical problem: acceptability judgments are gradient, and in most cases, their spatial distribution is gradient too. So we face two kinds of questions. First, how do we represent the spatial distribution of judgments graphically, so that one can examine a map and glean patterns from it? Second, when are the patterns we see statistically reliable?

For the first question, there are many possibilities. After piloting some of them, we determined that a twofold approach is the most useful. First, we plot the individual participants’ primary childhood residence as points on a map. Often, we present judgments of 1–2 in one color, and 4–5 as a second color, omitting 3s. This is only for visual presentation; 3s are included in all calculations, including the interpolation and hot spot analyses discussed below.

Second, we use interpolation to visualize patterns of values. Interpolation fills in, for each part of the map, what its expected value is, based on a computation over the values of the points closest to it. We use the inverse-distance weighted algorithm (see Wood [2016]). We visualize the interpolation with different shades for 1–2, 2–3, 3–4, and 4–5. The result smooths over the points to reveal broader patterns in a visually clear way.

Together, collapsing the values of the point data (1 and 2 as the same color and 4 and 5 as the same color) and projecting interpolation under them generally provide a good visual overview of a dataset. But it is not always clear whether a geographic pattern is statistically reliable. One useful tool for this is the Gi* statistic (Grieve et al. 2011; Tamminga 2013), referred to as a “hot spot” analysis in ArcGIS, the software we use for mapping and geospatial analysis. The hot spots test is conducted for each data point. If there are any hot spots, we then draw borders around them to indicate an overall “region” of contiguous hot spots.^[7]

An example is shown in Figure 1, which shows the results of the sentence Here’s you some money. The shaded interpolation reveals that the sentence is judged better in the South than in other areas. The preponderance of green dots there (representing judgments of 4 or 5), versus black dots (representing judgments of 1 or 2), reinforces this, and tells us what the interpolation is based on. Some areas, for example, have more data than others, and this is valuable information. The red border indicates a hot spot region, and blue borders indicate cold spot regions. This tells us that the pattern we see is statistically significant.

Figure 1:

Here’s you some money.

Another technique we find useful is to overlay previously postulated dialect boundaries over our data. We can append each data point with its dialect region, and include that information in a regression analysis along with age, sex, race, etc. We also average over such regions and map them out. For example, the map in Figure 2 shows the average judgments for do-support with the have yet to construction (see Section 4). The dialect regions come from the Atlas of North American English (Labov et al. 2006). The darker the shade of blue, the higher the judgment.^[8] This map shows that such sentences are degraded for Southern speakers, but we find many acceptances in several areas in the North. Figure 3 shows the values and 95% confidence intervals for these regions; a one-way ANOVA reveals that the differences among means is statistically significant (F[12, 494] = 3.054, p = 0.0004); Tukey-corrected multiple comparisons reveal significant pairwise differences between the South and both Inland North and Western Pennsylvania, and between Western Pennsylvania and New York.

Figure 2:

Do-support with the have yet to construction.

Figure 3:

Do-support with the have yet to construction.

4.1 Interspeaker variation “in every room”

In many cases we find that some sentence or construction has a particular geographical distribution (see Figure 1). We will discuss examples of this kind of result below. In many other cases, however, we find rampant interspeaker variation without any geographic or demographic correlate. Some examples of this type are shown in (4).

In (4a), two auxiliaries appear to the left of the subject of a yes–no question (rather than the standard single auxiliary). Sentence (4b) exemplifies subject control in the presence of an indirect object, an important construction type in the control literature (see Landau [2013: 149 and references therein]). Copy raising, recently discussed by Landau (2011) and Asudeh and Toivonen (2012), is exemplified in (4c) – in this case involving an embedded object pronoun (him) that matches the matrix subject (John). All three of these sentence types vary across speakers, but none of the variation, as far as we can tell, has geographic correlates. See Wood et al. (2015a, 307) for a map of (4a) and Wood (2016) for a map of (4b). Figure 4 presents a map of a variant of (4c).

Figure 4:

John seems like Mary offended him.

These kinds of results are certainly open to interpretation; the point here is just that a variety of sentences that are reported in the syntax literature to vary across speakers exhibit “variation in every room” – that is, interspeaker variation that might be expected in any given room of native speakers. Knowing what kinds of syntactic phenomena vary across individuals in the same speech community raises a host of interesting questions about language acquisition and online sentence processing. Moreover, generative syntacticians aim to understand interspeaker variation in general, whether it is tied to geography or not (Kayne 2013: 133).

4.2 Implicational relationships

One kind of result that has been theoretically illuminating, independent of geographic concerns, involves implicational relationships of the kind familiar from linguistic typology (Greenberg [1963, 1966] and later work; see Szmrecsanyi and Kortmann [2009] and Siemund [2013] for further discussion of these relationships across English dialects). For example, Tyler and Wood (forthcoming) study survey results focusing on the have yet to (HYT) construction, illustrated with sentences like (5). One of the questions they pursue is whether have is an auxiliary, leading us to expect a yes–no question as in (6a), or a main verb, leading us to expect a yes–no question as in (6b).

There is quite a bit of variation in the judgments of sentences like those in (6). The quantitative results lead Tyler and Wood (forthcoming) to two conclusions. First, there are enough speakers who accept (6b) to take it to be a genuine syntactic option for many speakers. Second, speakers who accept (6b) are overwhelmingly likely to accept (6a) as well, but not the other way around: many speakers accept (6a) but reject (6b). Tyler and Wood (forthcoming) develop a syntactic analysis of the HYT construction intended to derive this asymmetry. Results like this strike us as one key area where quantitative studies can help raise (and answer) interesting theoretical questions.

4.3 Geographic diffusion

Because we ask for the ages of our survey participants, we are able to analyze changes in a grammatical phenomenon’s geographic distribution over time. An example of this comes from the “personal dative” construction (Christian 1991; Webelhuth and Dannenberg 2006; Horn2008, Horn2013; Gerwin 2014: Ch. 7; Hutchinson and Armstrong 2014). As illustrated in (7), the dative (her) is obligatorily pronominal and coreferent with the subject (she):

Personal datives were thought to be a phenomenon characteristic of the South (Webelhuth and Dannenberg 2006), though some (e.g., Christian 1991: 18) have conjectured that it may be found in other vernacular varieties as well. Our research has shown that it may be spreading geographically in “apparent time” (Bailey et al. 1991). Among speakers over 40, the construction is primarily accepted in the South. Among speakers between 18 and 30, however, acceptance is much more widespread. (Speakers between 31 and 40 fall somewhere in the middle.) This result, which may reflect what Horn (2008: 176) calls the “Braxton effect,”^[9] illustrates the kind of results we might expect when we intersect geographic region with other linguistically relevant social categories.

4.4 Known geographic distributions

A final kind of result involves confirming or elaborating on previously hypothesized geographic distinctions. We will discuss one example of each, both taken from Wood (2016), to which the reader is referred for more details.

For an example of confirmation, the so don’t I construction previously mentioned in (1e) has long been thought to be restricted to eastern New England (Labov 1972: 815; Hall 2013). This is borne out in our survey data. While nearly half of the participants in eastern New England accept sentences like (1e), exceedingly few outside of eastern New England do.

For an example of elaboration, the be done my homework construction previously mentioned in (1e) has been thought to be characteristic of Canadian, Vermont, and Philadelphia English (Labov 2001; Yerastov2008, Yerastov2010, Yerastov2015; Fruehwald and Myler 2015).^[10] This is borne out in our survey data, but we also find a more complex picture: it is favored not only in Philadelphia, but also in surrounding areas such as Delaware, southern New Jersey, and Maryland. It is also highly favored in much of New England, including not just Vermont, but New Hampshire, eastern Massachusetts, and Maine.

4.5 Summary

In sum, our primary research objectives stem from theoretical questions, and our survey methodology is designed to collect data that will bear on syntactic theory. But since our surveys also collect geographic and demographic information, we can analyze this information to determine which demographic factors, if any, constrain the phenomena we study. Knowing this can certainly help us to advance knowledge relevant to syntacticians, at the very least to help us find speakers of a construction we are interested in. Beyond that, we also aim, with this aspect of our project, to build on a rich tradition of dialect research and contribute to a broader understanding of how syntactic variation distributes across speakers of North American English.