A sampling technique for worldwide comparisons of language contact scenarios

1.1 Biases and controls in typological sampling

Put simply, sampling allows typologists to delimit the population of languages between which linguistic comparisons are drawn. This section presents a brief overview of a few fundamental principles for sampling in language typology, and how they relate to the overall goal of researching linguistic structure. For more detailed discussions of the vast literature on sampling methods, we recommend Bakker (2011), Miestamo et al. (2016), as well as Guzmán Naranjo and Becker (2022).

In principle, generalizations about the global distribution of language structures should rely on a carefully designed sampling procedure. In this respect, two of the most important desiderata in any sampling procedure are the criteria used to achieve representativeness and independence. Representativeness criteria address how a language sample reflects linguistic diversity. The criteria for independence tackle the sources of similarities between languages, be it inheritance or diffusion.

Representativeness and independence criteria may interfere with one another in various ways. For instance, representativeness may increase with the size of the sample, since a larger sample is more likely to represent the distribution of cross-linguistic patterns in the total population of human languages. At the same time, increasing sample size may compromise independence since it can inflate the likelihood of genealogical and/or areal relations between the sampled languages. One way in which researchers can address issues of representativeness and independence is to consider biases in language selection.

Widely addressed biases in typological sampling are genealogical and areal biases (for an overview of bias types in typology, see Bakker 2011).

Linguistic features such as gender systems may be particularly frequent in certain language families or areas of the world. A typological dataset that contains an overrepresentation of languages from families where given structures are common will then show a bias towards the occurrence of those very structures. For instance, sex-based gender systems are common among Indo-European as well as Semitic languages. A typological sample for a study of gender systems that is heavily skewed towards these families would inflate the frequency of sex-based systems at the expense of non-sex-based systems, which are common among Nakh-Daghestanian or Atlantic-Congo languages, for instance. Similarly, a study that investigates the distribution of gender systems, but whose sample draws heavily from Africa, would potentially over-represent the non-sex-based gender systems commonly found on that continent. In order to address these biases, typological samples tend to use stratification techniques with both genealogical and areal controls.

Stratification is a control procedure for sample overrepresentation based on genealogical and areal information. Language samples can be stratified through established genealogical groupings, and/or areal classifications. Classifications chosen for genealogical and areal controls may differ across typological samples, and may also reflect individual researchers’ goals. Given that genealogical and areal classifications function as external measures of control in typological sampling, once classifications are chosen, it is important that those criteria inform the selection of the entire dataset for a given study.

The sampling technique presented here is designed for systematic comparisons of contact scenarios on a global scale. Given that contact may be more likely to occur between geographically adjacent language communities (but see Campbell 2006: 1 for a discussion), areal stratification is a crucial step in our sampling procedure, detailed in Section 1.2.1. In order to help disentangle contact effects from genealogical inheritance, our technique focuses on sampling contact scenarios between maximally genealogically unrelated languages, that is, between languages belonging to distinct language families. Through this approach, similarities uncovered between languages in contact gain weight, given that the genealogical distance between these languages decreases the likelihood that they would show the same structures due to inheritance from a common ancestor. It follows from this consideration that genealogical stratification is another key aspect of our technique (see Section 1.2.2).

1.1.1 Areality control in typological sampling

Areality is defined here as the effect of geographic proximity on the distribution of linguistic features. Areal biases affect the independence and representativeness of language samples because unrelated languages in contact may come to share a great deal of structural similarities as a result of diffusion. For instance, contact with Indo-European languages (different Germanic languages and Russian, rus) has led to changes in the syllable structure of Estonian (ekk, Uralic) (Napoleāo de Souza and Sinnemäki accepted 2022). Lack of control for areality may thus be particularly detrimental to probability samples, which aim to uncover statistical tendencies in the distribution of linguistic features. At the same time, as Bakker (2011: 95) also underscores, factoring in information about areal diffusion into typological investigations can be helpful when assessing the stability of linguistic features and their likelihood to be diffused through contact, which is one of the goals of current research agendas in distributional typology (Bickel 2010). In this section, we describe the most widely used methods to control for areality in present-day typology.

The commonest approach to areal control in typological sampling is to use continent-wide classifications as a basis for language selection. For instance, Dryer (1989, 1992 divides the world into six macro-areas,^[1] roughly corresponding to major continental zones: Africa, Eurasia, Southeast Asia and Oceania, Australia and New Guinea, North America, and South America. These six macro-areas are assumed to be maximally independent from each other both in terms of language history and of contact dynamics, while also reflecting comparable levels of genealogical and typological diversity. A similar approach is taken by Hammarström and Donohue (2014), whose six-way areal classification is nonetheless based exclusively on the distribution of landmasses versus water bodies: Eurasia, Africa, Australia, Multinesia/Papunesia, North America, and South America. Hammarström and Donohue’s classification is the one currently used by Glottolog (Hammarström et al. 2020) and the latest online versions of the WALS database (Dryer and Haspelmath 2013).

A third areal classification in typology, which is also based on continent-wide areas, is the one adopted for the purposes of the Autotyp database (Bickel et al. 2022). The Autotyp areal classification consists of two levels. The first level identifies 10 wide continental areas that are roughly comparable to Dryer’s (1989, 1992 and Hammarström and Donohue’s (2014) macro-areas, but are established based on “assumptions about contact events in history, informed by current knowledge of the historical, genetic, anthropological, and archeological record” (Bickel et al. 2022). These continent-wide zones are: Africa, Western and Southern Eurasia, Northern and Central Asia, South and Southeast Asia, Australia, New Guinea and Oceania, Western North America, Eastern North America, Central America, and South America. The second level focuses on areal divisions on a smaller scale, identifying 24 different linguistic areas. These are: Alaska-Oregon, Basin and Plains, California, East North America, Mesoamerica, Andean, Northeast South America, Southeast South America, Europe, North Africa, African Savannah, Greater Abyssinia, South Africa, Greater Mesopotamia, Indic, Inner Asia, Southeast Asia, North Coast Asia, Oceania, North Coast New Guinea, New Guinea Highlands, South Coast New Guinea, North Australia, and South Australia. The 24 Autotyp Areas are shown in the Map in Figure 1. The boundaries between these smaller scale areas were established based on non-linguistic factors, such as (bio-)geography and socio-economic history (Johanna Nichols, pers. comm.). The Autotyp Areas are thus an attempt at tackling the representativeness and independence of typological samples through a procedure of areal stratification that is based on natural and cultural-economic regions of the world. The 24 Autotyp Areas have featured in typological studies investigating how the interaction of historical contingencies and cognitive abilities affect the distribution of language structures (aka ‘distributional typology’, e.g., Sinnemäki and Di Garbo 2018; Widmer et al. 2021).

Figure 1:

The 24 Autotyp Areas (Bickel et al. 2017, used under CC-BY 4.0 license).

Our method uses the 24 Autotyp Areas as a grid for identifying candidate contact scenarios within diverse geographical areas of the world while also providing us with a tool for areal control at a smaller scale than the continent-wide level.^[2] This procedure is discussed in detail in Section 2.

In addition to the use of pre-established geographical classifications such as those surveyed above, spatial statistics has become increasingly common in quantitative typology as a way to assess the degree of contact between speech communities. For instance, Murawaki and Yamauchi (2018) use autoregressive models to investigate the relationship between inheritance and diffusion in a selection of typological features. In a similar vein, Jaeger et al. (2011) and Guzmán Naranjo and Becker (2022) address contact by using both macro-areas and distances between languages.^[3] While these approaches can be useful for automatic control, and to estimate the degree of contact between languages, the current technique builds samples directly based on established contact relations between language communities. We argue that this type of phenomenon-based sample is especially useful in order to test hypotheses about contact-induced change and the socio-historical correlates of typological distributions.

1.2 Establishing and comparing contact scenarios: insights from areal typology, areal linguistics, and contact linguistics

Areal typology and areal linguistics investigate the processes through which languages in contact may become similar because of interactions between neighboring communities. However, what characterizes studies in areal typology vis-à-vis areal linguistics is the focus on areal effects from a cross-linguistic perspective (cf. Dahl 2001: 1456). Areal typology has gained ground over the past 20 years due to a growing interest in disentangling functional-communicative pressures from socio-historical dynamics in typological distributions (Bickel 2010; Koptjevskaja-Tamm 2011).

Areal studies typically conceive of contact influence as resulting from bundles of languages in interaction. Linguistic outcomes of contact within these bundles are generally distributed along multiple trajectories, which can be depicted in the form of isoglosses. Yet another approach characterizes contact studies, which typically investigate language contact from the perspective of individual languages, and are mostly based on case studies. In this context, contact influence is described as resulting from interactions between pairs of languages in contact.

Our procedure attempts to reconcile the research traditions of areal and contact studies. It relies on geographical classifications that take in to account linguistic geography and areality. At the same time, it constructs sampling units in a pairwise fashion, based on a common practice in language contact research. As outlined in Section 2 and subsequently demonstrated in Section 3, this approach results in a sampling tool that is easily applicable to a variety of contact scenarios beyond those that we have sampled for the purpose of this demonstration.

2.1 Geographical area

This first criterion allows us to draw a sample that represents contact scenarios from all parts of the world. It thus complies with general representativeness desiderata in typological sampling (Section 1.2). Given that one of the goals of our methodology is to generate a sample for testing contact effects on linguistic variables, we opted for using a geographical classification that is at least in part delimited on criteria other than linguistic ones (Section 1.2.1). Thus, the 24 areas established for the Autotyp database (Bickel et al. 2022) form the basis for areal stratification.

The Autotyp Areas differ from other geographical classifications in that they derive from a combination of archeological, anthropological, historical, and genetic data rather than from linguistic features alone (Nichols et al. 2013: 6). The 24 Areas in the Autotyp classification (Figure 1) also present the advantage of zooming in onto continent-wide areas, which is not the case in any of the other widely used classifications in typology, such as the one adopted by Glottolog. This classification thus provides us with a grid to pursue representativeness within continent-wide areas. Table 1 lists the areas we used in the current technique in comparison to (a) continental landmasses, and (b) the Glottolog areas.

Initially, we aimed at selecting two sets from each of the 24 Autotyp Areas (i.e., 2 sets × 3 languages × 24 areas = 144 languages). In order to arrive at an even 150 languages,^[7] we decided to include two extra sets, one from Northeast South America, and the other from the North Coast of New Guinea. We selected South America and New Guinea as the sources for the two additional sets because of the high degree of linguistic diversity found in those areas as a whole (e.g., Dahl 2008; Hammarström 2016). For reasons explained in Section 2.2, only one set could be sampled for Southern Australia, which yielded a final sample of 49 sets and 147 languages in total.

2.2 Independently reported contact scenarios

This second parameter guides the choice of Focus and Neighbor Languages to analyze within each Autotyp Area. To fulfill this criterion, we relied on descriptions of contact scenarios in the language contact and areal linguistics literature, primarily through macro-area surveys of contact situations. For instance, we drew our Northeast South America sets based on the surveys of contact in the Amazon Region by Aikhenvald (2002, 2012. From those surveys, we then selected the ‘Vaupés Area’ as a starting point from which to draw specific contact scenarios and individual languages.

Descriptions of individual cases of contact varied in detail across the different areas. For instance, some language contact situations have a longer history of study, such as the Pacific Northwest in North America (Thompson 1979), whereas others stem from more recent proposals, for instance those involved in the Mamoré-Guaporé Area in South America (Crevels and van der Voort 2008). Whenever possible, we opted for contact scenarios with multiple instances of documentation in the literature. Table 2 illustrates some of the materials we consulted to arrive at our individual sets.

The task of matching contact scenarios with the Autotyp Areas was straightforward in most cases. Whenever we were uncertain where to place a case within an Autotyp Area, we relied on the geographical location of the Focus and Neighbor languages as stated in Glottolog. In a few cases, we also consulted specialists in the specific contact situations. A few Autotyp Areas presented challenges of their own. These are the three Autotyp Areas in the island of New Guinea, the North Coast Asia Area, Eastern North America, and Southern Australia.

The New Guinea Areas, and North Coast Asia are all relatively small, so that finding a Benchmark proved especially challenging in some cases. Additionally, New Guinea shows less in-depth documentation given the linguistic diversity in the region compared to other parts of the world. Eastern North America showed high degrees of language obsolescence, which had some impact on sampling composition (see Section 5). Finally, the spread of the Pama-Nyungan family in Southern Australia posed an insurmountable problem to the application of the selection criteria. The ubiquity of Pama-Nyungan led to a very reduced choice of contact scenarios between genealogically unrelated languages in that part of the world. As a result, only one set was sampled for the Southern Australia Autotyp Area.

We wish to underscore that aside from selecting sources that provided a detailed description of the contact scenario from a linguistic standpoint, we made no further evaluations as to the nature of the contact relations described in the literature. As such, we remain neutral to any disagreements or disputes over the plausibility of individual proposals, especially regarding larger linguistic areas. Finally, neither the type of contact profile instantiated by the candidate contact scenarios (e.g., regarding ‘intensity’ or ‘symmetry’), nor the linguistic outcomes of contact reported in the literature played any role in the selection procedure.

2.3 Genealogical distance

In order to minimize possible confounds due to common ancestry, we established that each set in our sample would include only Focus and Neighbor Languages that belonged to distinct language families. This third criterion helps us ensure that linguistic effects on Focus Languages stem from the contact situation rather than from shared inheritance (but see how the current technique would fare in studies of more closely related languages in Section 4). Genealogical distance thus represents a way to address the independence desideratum in typological sampling (Section 1.2).

As previously mentioned, we used the classifications in Glottolog as a parameter for genealogical distance between Focus and Neighbor Language within a given set. For the purposes of this paper, a language family corresponds to the highest level of genealogical classification in Glottolog (i.e., “top-level family”), meaning that a grouping such as ‘Finnic’ constitutes a subfamily in our classification, rather than a family (in this case, Uralic). All Glottolog classifications used in this paper refer to the database as of April 2020 (i.e., version 4.2.1; Hammarström et al. 2020).

For instance, applying this criterion to the contact scenarios in the Vaupés Area gave us the option of selecting the Naduhup language Yuhup (yab) as the Focus language, and a Tucanoan language such as Macuna (myy) as the Neighbor Language (Aikhenvald 2002). Determining which language should be the Focus, and which should be the Neighbor Language often hinged on the level of detail in the description of contact, and on the availability of reference materials (see Section 2.4). Additionally, in a few cases discussed below, the internal structuring of a given language family influenced sample composition, especially regarding the choice of the Benchmark Language. For each Focus Language, a Benchmark Language was chosen from among the Focus Language’s closest relatives^[8] as stated in Glottolog 4.2.1 (Hammarström et al. 2020).

The choice of Benchmark poses a challenge in itself, since two closely related languages may indeed be involved in the same contact scenario.^[9] Such a possibility led us to establish further criteria for the selection of Benchmark Languages. First, the Benchmark Language must not be part of the Focus/Neighbor contact zone. Ensuring that the Benchmark language is not subject to the same contact influences as the Focus language strengthens its status as a source of external control. When more than one candidate was available, we chose the one that was geographically closest to the Focus Language. For the current sample, we relied on the sources described in Section 2.1.2 to rule out the Benchmark Language’s participation in the contact scenarios we sampled. More specifically, we double-checked reference materials for mentions of contact between Benchmark, Focus and Neighbor Languages, and replaced any Benchmarks that participated in contact scenarios with either FL or BL.^[10] Contact between the Benchmark and other languages, possible as they may be, played no role in the selection process.

The Vaupés contact zone once again serves as an illustration. Following the steps outlined above, we find that Yuhup’s closest relative is Hup (jup), which also belongs to the Hup-Yuhup sub-branch of the Naduhup family in the Glottolog classification. Since Hup also participates in contact scenarios involving languages in the Tukanoan family, possibly with the Neighbor Language Macuna itself, selecting Hup violates one of the criteria for selection as Benchmark. The next closest relative to Yuhup would then be Dâw (kwa). However, upon double-checking our references, we found that Dâw too played a role in this particular contact scenario. As such, we moved on to the next relative, finally selecting Nadëb as the Benchmark Language in the Hup-Macuna set. Figure 2 illustrates the genealogical relations between the languages of the Naduhup family according to Glottolog. The location of Hup, Yuhup, and Nadëb with respect to the other language families of the Vaupés is shown in Figure 3.

Figure 2:

The Naduhup family tree based on Glottolog (Hammarström et al. 2020).

Figure 3:

The Naduhup languages with respect to the other languages of the Vaupes region. Map taken from Epps (2013: 333), available Open Access. Scale: 1 cm–100 km (our estimate).

2.4 Availability of reference materials for linguistic data collection

This criterion relates to the overall goals of our larger project, namely to test the effects of contact on a range of linguistic variables. Our research aims to compare given variables in the Focus Language to those of the Neighbor Language, using the Benchmark Language as a parameter for comparison. Thus, one relevant caveat that this method introduces is that each sampling unit requires three different descriptions of appropriate quality for the languages in each set. Reference materials such as grammars, journal articles, and dictionaries were deemed appropriate for the goals of our study. While we may use additional sources such as databases and data from conference presentations, we established that the three languages in each of the sets must have primary linguistic sources for them to be included in the sample.

Selecting reference materials for Benchmark Languages was a relatively uncomplicated task given that there were often different options for each set. On the other hand, the lack of appropriate descriptions for Neighbor Languages posed a much greater challenge. In many cases, the Focus Language was in contact with a single language belonging to a different family, so that the lack of materials for that particular Neighbor Language led to the exclusion of the entire set. In cases for which there were two or more candidate Neighbor Languages, we selected the one we deemed had the most suitable description. As discussed in Sections 4 and 5, in some cases, the choice of Neighbor Language was narrowed down through discussions with experts of the sampled contact scenarios.

4.1 Applicability of the sampling procedure

The dataset generated through our sampling procedure fully complies with the four criteria introduced in Section 2, namely: geographical area, independently reported contact scenarios, genealogical distance, and availability of reference materials. By using the Autotyp Areas and Glottolog classification as bases for areal and genealogical control, our dataset meets key desiderata in typological sampling, in that it is both geographically and genealogically stratified. Moreover, by selecting two-three contact scenarios per Autotyp Area, the method also achieves representativeness in that it includes a diverse range of sociolinguistic contexts from various parts of the world (Section 1).

The sampling procedure utilizes insights from areal linguistics, areal typology, historical linguistics, and language contact research, being further guided by the contact literature. This method makes one important contribution to sociolinguistic typology (Croft 2021; Trudgill 2011). It is the first technique of its kind to allow for worldwide comparisons of contact scenarios, one of the key goals of the sociolinguistic typology enterprise. While there have been other attempts to study contact from a cross-linguistic perspective, and from the perspective of quantitative typology, (e.g., Chang and Michael 2014; Murawaki and Yamauchi 2018), the novelty of our approach resides in providing a sampling method that builds directly on established contact scenarios, which in turn allows for conducting large-scale comparisons of language contact phenomena.

4.2 The procedure regarding the study of contact effects

The sample presented here is currently being used to gather linguistic data in the context of our larger project (see Section 1). We are coding for 200 different variables in the realms of phonology, morphosyntax, and lexicon in each of the languages in the sample. More specifically, the coding encompasses syllable structure, lexical prosody, nominal number, locus of marking of nominal possession, and adnominal demonstrative systems. The project uses a fine-grained approach to variable design that assesses the intricacies of each variable, including their behavior under different morphosyntactic and/or semantic conditioning factors (see Bickel 2010 for similar approaches in language typology). Our coding scheme is based on documented cross-linguistic variation in each of these domains (e.g., Maddieson 2013 for syllable structure; Kibort and Corbett 2008 for nominal number; Nichols 1992 for locus of marking; and Maurer 2013 for demonstratives). Additionally, the data collection follows a bottom-up approach that examines which aspects of a given domain of language structure align between languages in contact. Our coding aims to provide answers to the following questions:

1. How similar are the Focus and Neighbor Languages with respect to the parameters describing the properties of each linguistic variable?

2. How do similarities between Focus and Neighbor Languages compare to corresponding structural properties of the Benchmark Language within each set?

Preliminary results obtained using the current sample indicate that our method captures structural similarities between Focus and Neighbor Languages that strongly suggest explanations derived from language contact. That is, we find systematic correspondences between a range of phonological, morphosyntactic, and lexical features of the contact pairs which differ from those attested in the respective Benchmark Languages. Data from six of our sets illustrate those findings in the domain of nominal number. The examples sets were chosen to represent each continental area, that is, Africa, Australia, Eurasia, Papunesia, North America, and South America. The languages in each set, their genealogical affiliations, and a selection of possible contact effects followed by the total number of FL–NL feature matches for each set are given in Table 5. Short descriptions follow.

One striking pattern emerges from the data presented in Table 5. Focus and Neighbor languages across these six sets are similar to each other by either possessing or lacking one number category (notably the dual or collective plurals), whereas the opposite pattern is found in the Benchmark for each set. While number systems across these sets differ in their morphosyntactic and semantic behaviors, the patterns observed suggest that pairs of Focus and Neighbor languages show similarities by either possessing or lacking typologically less frequent values compared to the respective Benchmarks.

The scenario in which Focus and Neighbor languages share number values missing in the Benchmark is illustrated by Sets 09, 15, 21, and 49. In Set 09, from Northern Australia, both Mawng and Kunbarlang have dual inflections on independent personal pronouns (Kapitonov 2019; Singer 2006), but these are absent in the Benchmark Language Iwaidja. In Set 21, the Focus Language Papanana marks dual number in the collective forms of the articles (Smith-Dennis 2021: 191), similarly to the Neighbor Language Rotokas, in which dual marking is generally pervasive (Robinson 2011), whereas the Benchmark Marshallese lacks dual entirely. Similarly, both Yurok and Karok in Set 15 have verb forms described as “collective plurals” which are used to index that the subject denotes a group of entities (Bright 1957: 88; Garrett 2014: 41–46). These forms are not reported for the Benchmark Naskapi. Finally, in Set 49, both the Focus Language Western Toba and the Neighbor Language Wichí Noctén are reported to have a series of nominal affixes that encode collective and distributive meanings. This type of affixes is not attested in the Benchmark Language Kadiwéu (cf. Carpio 2012: 71–73 for Western Toba; Terraza 2009: 88–93 for Wichí Noctén).

The opposite pattern, where certain number values are attested in the Benchmark, but absent in the Focus and Neighbor languages, is found in Set 02 and Set 14. In Set 02, dual forms are attested in the Benchmark Lele (Frajzyngier 2001), but absent in the Focus Language Bade (Schuh 2005) and in the Neighbor Language Manga Kanuri (Hutchinson 1981). This situation is akin to Set 14, where dual inflections are frequent in the Benchmark Gulf Arabic (Holes 1990) while nearly absent in the Focus Language Maltese (Borg 1994); the Neighbor Language for this set, Sicilian, lacks dual entirely (Bigalke 1997).

Having first described these findings, we now turn to a word of caution. While the fact that a feature is present in two unrelated languages in contact but absent in the Benchmark may be taken as indication of contact-induced change, the absence of a feature in both languages arguably constitutes less clear evidence of contact. In other words, instances when two unrelated languages in contact both lack a given feature which is present in the Benchmark Language (which would correspond to a ‘0–0–1’ in our coding method) would be less indicative of contact-induced change than when both Focus and Neighbor Languages contain a feature absent in the Benchmark (1–1–0 in our coding method). Such a conclusion is warranted especially in cases of typologically frequent properties.^[11] Put simply, one should be cautious to use ‘shared absences’ as the sole indication of contact. Expanding the basis of comparison to more than one Benchmark would seem especially important for better contextualizing shared absence as indication of contact.^[12] Nevertheless, all things being equal, both ‘shared presence’ and ‘shared absence’ contribute relevant datapoints to the overall assessment of a given contact profile.

It should be noted that the method being proposed here does not preclude more in-depth, comparative analyses of contact effects in individual languages. Rather, our method provides a sample which researchers can use as a starting point for further investigations of contact. It mitigates the issue of similarity due to inheritance by proposing the analysis of maximally unrelated languages. It offers an additional tool for control by proposing the inclusion of a Benchmark Language against which variables should be compared.

The current sampling method thus provides typologists with a testing ground for investigating contact effects on a large scale. Preliminary observations from six sets from across the globe suggest that data collection using this sampling technique allows for the detection of structural similarities between given pairs of Focus and Neighbor Languages. The addition of Benchmark Languages provides a systematic way to further ensure that any similarities uncovered may be due to contact. The next section discusses limitations in the procedure.

4.3 Limitations

As hinted above, the current procedure does not preclude detailed areal and historical-comparative analyses of specific cases of contact, all of which remain paramount. Additionally, while successful in generating a sample of languages suitable for the study of contact from a typological perspective, the current method presents limitations, three of which we discuss here. Two limitations concern Benchmark Languages, namely how many Benchmark to consider per set, and the differing degrees of relatedness between Benchmarks and Focus Languages across the sets. The third limitation addressed in this section is the project-specific constraint that requires that we find living experts on the Focus Languages and/or contact scenarios with whom to collaborate.

4.4 Current implementation of the procedure

As mentioned above, the sampling technique presented in this paper is part of the research design of a larger project investigating how various sociolinguistic contexts impact around 200 variables in the realms of morphosyntax, phonology, and the lexicon. The dataset presented in Section 3 and Appendix A (in the Supplementary Materials) is a demonstration of the sampling technique as well as the actual source of linguistic data for the project. This section briefly illustrates how the sampling procedure serves the goals of the project at large.

The sampling technique serves three interrelated goals within the larger project. First, the dataset resulting from the sampling procedure enables us to conduct comparisons of contact scenarios from a sociolinguistic perspective. Secondly, the sample serves as the source for data on linguistic variables in the languages of each set (illustrated in Section 4). Thirdly, the sample allows us to study how the sociolinguistic profiles observed in different contact scenarios relate to the patterns uncovered through the comparison of language structures within sets. Achieving these three goals would thus generate a systematic account of linguistic contact, encompassing both its sociolinguistic underpinnings and its potential impact on linguistic structure.

In this design, a sociolinguistic questionnaire probes into contact dynamics between Focus and Neighbor language communities in six different domains: family and kin, ceremonial exchange, daily interactions, labor, knowledge, and trade. Investigating contact in these domains serves as a description of the nature of contact between the Focus and Neighbor communities. The questions in the questionnaire are informed by the literature on contact, psycholinguistics, historical linguistics, second-language acquisition, among others (see Di Garbo et al. 2021). Questions address topics such as the duration and density of contact between Focus and Neighbor languages, social network structures, language attitudes, patterns of language use and transmission, and so forth.

Table 7 provides an overview of the sample composition by six macro-areas as of November 2021, focusing on the match between the contact scenarios identified in the literature (‘Target’) and those for which we were able to identify respondents to our questionnaire (‘Obtained’).

As illustrated in Table 7, we attained adequate matches between the number of contact pairs identified in the literature and expert responses for Africa and Eurasia, with a coverage of 100 and 85 percent, respectively. High coverage (between 78 and 58 percent) was also reached for Papunesia, Australia and South America. This result is particularly promising given that these areas tend to be underrepresented in typological research, often due to lack of accessible resources. However, for North America, only 50 percent of the initial candidates for contact pairs yielded questionnaire responses from experts. This observation contrasts with the fact that the area is usually well-covered in typological studies focusing only on linguistic features.

The low targets reached for an area such as North America may reflect the insufficient documentation of sociolinguistic phenomena in Native languages, as discussed in Section 3. The lack of responses to our questionnaire also reveals the degree of language obsolescence in indigenous communities. We note that we are retaining the full-fledged sample for collecting linguistic data regarding our 200 variables (Section 4). However, it follows that the sociolinguistic data come only from those sets for which we established collaborations with experts.

4.5 Further applications

The current sampling procedure is applicable to different research goals. While the focus of the current dataset is on contact dynamics between genealogically unrelated languages, the language selection criteria introduced here could also serve investigations of contact between related languages (see Section 4.3 for an illustration using the Balkans). For related languages, generating a sample would only entail reconfiguring the internal composition of the three-language sampling units.

What’s more, the current sample partly depended on the availability of living experts with whom we could collaborate in the sociolinguistic component within the larger project. In Section 4.3 we discuss how this project-specific strategy may have led to an emphasis on more recent contact scenarios. Future studies based solely on reference materials may thus be less constrained in their choice of contact scenarios, possibly permitting the investigation of contact situations for which no sociolinguistic information is available.

Finally, as reiterated throughout this article, this sampling technique is phenomenon-based, stemming from the manual assessment of the contact literature. A possible development of the current method would be to implement an automatized procedure of language selection based on similar design principles. This would enable the researcher to generate possibly larger datasets than the one presented here.