2.1 Previous study of Finnish dialects

Finnish dialects have traditionally been divided into eastern and western varieties (Rapola Reference Rapola1969, Wiik Reference Wiik2004) and further subdivided into eight dialects (Itkonen Reference Itkonen1964, Reference Itkonen1989). The traditional eight dialect categories include Southwest, Southwest transitional, Häme, South Ostrobothnia, Central/North Ostrobothnia, Far North, Savo, and Southeast. See Figure 1 for their geographical distribution.

Figure 1. Finland’s dialect areas: (1) Southwest, (2) Southwest transitional, (3) Häme, (4) South Ostrobothnia, (5) Central/North Ostrobothnia, (6) Far North, (7) Savo, (8) Southeast. The dark-grey areas are predominantly Swedish-speaking. Map by Tommi Kurki.

These dialect groups have been distinguished based on the distribution of a range of linguistic features. One of the more well-known surveys of Finnish dialects was conducted by Lauri Kettunen, later compiled and spatially mapped into his Dialect Atlas of Finland (Suomen murteet: Kettunen Reference Kettunen1940a,b).Footnote ¹ Many of the more dramatic dialectal differences leveled off by the nineteenth century and further leveling has occurred since. Several studies have applied quantitative methods to examine Finnish dialects and their collective variations (Wiik Reference Wiik2004, Leino, Hyvönen & Salmenkivi Reference Leino, Hyvönen and Salmenkivi2006, Hyvönen, Leino & Salmenkivi Reference Hyvönen, Leino and Salmenkivi2007, Leino & Hyvönen Reference Leino and Hyvönen2008). Most recently, modern population genetic approaches to dialectal variation have confirmed the computational soundness of the traditional eight-dialect categorization (e.g. Syrjänen et al. Reference Syrjänen, Terhi Honkola, Leino and Vesakoski2016).

We build on this quantitative work confirming the eight-way distinction and use it to categorize and map the speech of individual speakers. Previous approaches use aggregated survey data from Kettunen’s Dialect Atlas (Reference Kettunen1940a) as input and basis for the dialectal distinctions, rather than the speech of individuals (e.g. Honkola Reference Honkola2016, Syrjänen Reference Syrjänen2021). The Atlas only reports the most frequently used variant of a feature per municipality, rather than capturing the full range used in a municipality and even within a particular dialect. While additional discussion of feature variation is given in Kettunen’s explanation for the Dialect Atlas (Kettunen Reference Kettunen1940b), this additional variation is for the most part not included in recent quantitative approaches to dialects. This means that studies using the Atlas as a foundation are restricted from examining social variables of speakers, as responses were not recorded for individuals (see Honkola Reference Honkola2016:27). We build on these previous approaches and present a model that examines the speech of individuals and defines dialect categorization in distributional terms. This is more in line with some other quantitative approaches to the study of Finnish in contact settings by Lainio (Reference Lainio1989), whose model similarly included multiple possible realizations of dialect features per dialect for categorization. This allows for the natural complexity of the speakers’ language to be captured and more overlap between dialect features to be taken into account.

By using a distribution-based model and training it using actual speech data from the LAX corpus, natural inter-speaker and intra-speaker variation is inherently included in our model, and is thus more suitable for the study of Finnish in the North American context. The model that serves as a point of reference for the regional variation of spoken Finnish is built on actual speech samples from multiple individuals, and so this approach further allows for social variables to be considered for individual speakers. This is an important factor for the study of American Finnish, a contact language that speakers mostly acquired as a heritage language.

2.2 Heritage languages and speakers

A heritage language broadly defined is a language that an individual culturally identifies with because of a family or community tie (Polinsky & Kagan Reference Polinsky and Kagan2007:369). A heritage speaker of a language is narrowly defined as an individual raised in a home where a language other than the dominant community language was spoken, later resulting in some degree of bilingualism in the heritage and societal majority language – whether balanced or (usually) unbalanced (Scontras, Fuchs & Polinsky Reference Scontras, Fuchs and Polinsky2015:3). Many heritage speakers do (and did) not receive formal, written instruction in the heritage language (especially in the North American context), and thus the setting of acquisition, frequency of usage, and exposure to different registers and dialects has an impact on which language features they use most frequently in their own speech. Polinsky & Kagan (Reference Polinsky and Kagan2007) suggest that heritage speakers can be identified on a continuum of proficiency in order to best capture the nuances and differences between speakers’ abilities and contexts of acquisition (Reference Polinsky and Kagan2007:370–372). The nature of the input that heritage speakers receive is inherently different from that of speakers acquiring the language in a majority language context (Pascual y Cabo & Rothman Reference Rothman2012:451). This raises the complex question of what are appropriate baselines for comparison in the study of heritage languages. Polinsky (Reference Polinsky2018) discusses nuances of the issue suggesting that depending on the community and research questions, age-matched homeland speakers from the time of emigration or balanced bilinguals within the speech community may be appropriate alternatives, though there may also not be a relevant baseline available for comparison, as is often the case for endangered languages with few remaining speakers (Reference Polinsky2018:10–17; 329–333).

2.3 American Finnish as a contact variety in northern Wisconsin

American Finnish developed as a contact language in the generations after Finnish-speaking individuals migrated to North America in the early twentieth century. It contained influence from contact with English, but also features common in the Finnish dialects spoken in the places from which individuals first migrated. The majority of migrants came from Ostrobothnia and other western municipalities, meaning many also spoke Swedish or otherwise had traces of Swedish contact in their speech. Children of Finnish migrants born in North America grew up hearing Finnish spoken by their parents and within their smaller communities, but not used in official contexts or taught in public schools – common for minority languages past and present. While there are many similarities across speakers of Finnish in North America, this is not a formalized speech variety. Despite some general tendencies, there are many local and individual variations, especially in lexicon and pronunciation (Jönsson-Korhola & Lindgren Reference Jönsson-Korhola and Lindgren2003:409).

Researchers who have interviewed and spoken with many American Finnish speakers comment that the Finnish spoken by Finnish-Americans tends to have the dialect of a speaker’s parents or grandparents as foundation, but also clear influence from both English and the Finnish spoken by other speakers in their community (Jönsson-Korhola Reference Jönsson-Korhola1993:110; Jönsson-Korhola & Lindgren Reference Jönsson-Korhola and Lindgren2003:408–9; Männikkö Reference Männikkö2004:4). For this reason, while there are similarities across many speakers of American Finnish, there are also differences that tend to be individualistic. Many variations are due to different amounts of regular exposure to Finnish, as well as the locations from which a speaker’s parents or grandparents first migrated (and the related dialect that they spoke). American Finnish has been documented and studied by researchers focusing on contact effects (Larmouth Reference Larmouth1974, Hirvonen Reference Hirvonen and Niemi1992, Reference Hirvonen, Preston, Joseph and Preston2005, Virtaranta et al. Reference Virtaranta, Jönsson-Korhola, Martin and Kainulainen1993), as well as code-switching and borrowing (Poplack Reference Poplack1980, Lauttamus Reference Lauttamus, Ojanen and Palander1991, Halmari Reference Halmari1997, Männikkö Reference Männikkö2004). There have been few studies on American Finnish using quantitative approaches (aside from Poplack, Wheeler & Westwood Reference Poplack, Wheeler and Westwood1989 and Hirvonen & Lauttamus Reference Hirvonen and Lauttamus2000), mostly because the bulk of the fieldwork was conducted before such models and approaches were being widely used.

Observations about the grammatical structure of American Finnish tend to follow patterns seen in heritage languages across different types of contact settings. While some English influence is present, speakers tend to have more standard-like usage of cases (especially directional cases) than those of typical learners of Finnish as an additional language (Martin Reference Martin1993:97). Larmouth reports that directional cases (illative, elative, inessive, ablative, adessive, and allative) were preserved on nouns and pronouns in the speech of second and third generation Finnish Americans in Minnesota, while use of the partitive and accusative cases varied (Larmouth Reference Larmouth1974:365–366). This early observation of the endurance of case marking is in line with recent work that contradicts stereotypes of heritage speakers as losing case marking. Łyskawa & Nagy (Reference Łyskawa and Nagy2020) observed >90% normative case usage across heritage speakers born in Toronto of Russian, Polish, and Ukrainian parents in data collected in a naturalistic setting (Reference Łyskawa and Nagy2020:148). Other work suggests processes of restructuring of the dative and accusative cases in multiple speaker communities of heritage German, rather than explicit loss (Yager et al. Reference Yager, Nora Hellmold, Putnam, Rossi, Stafford and Salmons2015; Bousquette Reference Bousquette2020:491–492). The endurance of case marking and other features suggests that while American Finnish may differ noticeably from Finnish spoken in Finland, the language is nevertheless similar enough that case marking, consonant gradation, and other classic features of Finnish dialects can reasonably be expected in the speech of fluent or comfortable speakers, and to a greater degree than with typical non-native speakers of Finnish.

American Finnish is one of many language contact varieties spoken in Wisconsin during the nineteenth to twentieth centuries, and previous work on other immigrant languages in Wisconsin has indicated identifiable homeland dialect features, as well as variation therein. Previous research on Wisconsin German speech communities established in the early 1800s has shown that individual German dialects are not especially identifiable in individuals’ speech; rather, there is a mixing of distinct regional and standard features in a way that would not be found in a European setting, and have become a regional marker for the area (Litty, Evans & Salmons Reference Litty, Evans and Salmons2015:184). This was in contrast to previous work on German, which assumed retention and preservation of dialects from the time of migration (e.g. Eichhoff Reference Eichhoff, Trommler and McVeigh1985:234). This mixing is likely due to natural language change over time, whereby languages and dialects spoken in contact situations lose much of their original variation as a ‘leveling’, or reduction of marked variants, occurs, yet some features are maintained if they are in some sense ‘simpler’ (Trudgill Reference Trudgill1986:98). It is possible that the setting of acquisition of German as a heritage language in later generations played a role in the mixing of standard and regional features; some speakers learned to read and write standard German yet spoke Low German or other variants within the home (Bousquette Reference Bousquette2020:486–489).

Haugen (Reference Haugen1969) examined differences between home and community dialect features in Norwegian spoken by immigrants and their descendants in the Midwest from the 1930s onward (Reference Haugen1969:337–360; 26). Haugen notes that some Norwegian Americans had difficulty understanding differing Norwegian dialects spoken by neighbors and spouses, contributing to a switch to English for some (Reference Haugen1969:349). While many dialect features were preserved in the speech of later generation American Norwegians, there was still a leveling of some marked variants, evidence of contact with English, and on the whole, dialect preservation could only be assessed ‘for certain features … and for certain individuals, rarely if ever for whole communities’ (Reference Haugen1969:360). However, often there was still a community influence on language norms to some degree. Natvig (Reference Natvig and Brown2022) outlines how local business and church services were conducted in the community’s Norwegian heritage language in Ulen, Minnesota, until language shift to English had occurred on a community-wide level. Examining linguistic features in heritage communities at different stages of language shift (with corpora from different time periods), in conjunction with analysis of changing social patterns and rates of language shift, is another potential application for this model.

The situation of American Finnish differs from Wisconsin German and Norwegian in that there were fewer generations over which the language was maintained. The settlement of Finnish communities happened later, and most Finnish-speaking communities in northern Wisconsin were smaller than the state’s German-speaking communities and had a comparatively shorter period of bilingualism (Johnson Reference Johnson and Brown2022:25–26). For heritage speakers of the second and third generations, their own family or co-workers often provided the majority of their Finnish input. This suggests there was not a sufficiently long period of community-wide bilingualism for a leveled, regional dialect of the minority language to develop, though there may have been some community-level norms and influences. Importantly, this study is examining the dialectal background of individual American Finnish speakers rather than suggesting a cohesive American Finnish dialect amongst these speakers.

3.1 Data 1: Heritage languages and bilingualism in Wisconsin

Primary data for this study come from five interviews conducted in 2016 and 2017 with speakers of American Finnish in northern Wisconsin. The speakers are from three rural communities established around the same time and with similar population sizes over time, economic bases in farming, and timelines of language shift. Interviews consisted of personal history questions related to language usage, picture identification tasks, the ‘Frog Story’ narrative task, sentence translations, and some free conversation. Interviews each lasted about 50–90 minutes. Interviews were transcribed into ELAN and variants coded for the dialect features under study, first using an automated R script and then hand-checked. These particular speakers were chosen because of their minimal Finnish language instruction (as children or adults) and the degree of comfort they expressed in conversing in Finnish.

Table 1 presents each speaker’s background and their reported connection to Finland. We consider the first generation as born in Finland and relocated after the age of 10. Those of the second generation have at least one parent who migrated from Finland during adulthood. Those in the third generation have at least one grandparent who migrated from Finland as an adult. Section 3.1.1 offers more detailed discussion of each speaker’s background.

Table 1. Background of American Finnish speakers

3.2 Data 2: Finnish Dialect Syntax Archive

Models in this study use data from the Finnish Dialect Syntax Archive (LAX) as input. These data consist of 171 transcribed interviews annotated manually in terms of morphology and syntax, with informants from across Finland (1,037,999 tokens). Most of the data was recorded in the 1960s and the majority of informants were born between 1870 and 1890. This time period is roughly when the majority of the Wisconsin Finnish speakers’ parents or grandparents were born in Finland. Many of them migrated to America as adults or in later childhood. This means that the Finnish which Finnish American informants heard and were most exposed to is from roughly the same time period as the speech recorded in the LAX corpus. This particular collection is thus a reasonable comparison, and more similar to the speech that American Finnish speakers would have had as input, compared with more recently collected interviews. The analysis makes use of the version provided by the Language Bank of Finland (Kielipankki) via the Korp user interface.Footnote ²

3.3 Dialect features examined in model

While some Finnish language features are considered neutral and widespread, others are linked to particular region(s) (Mantila Reference Mantila, Makkonen and Mantila1997, Reference Mantila2004; Nuolijärvi & Sorjonen Reference Nuolijärvi and Sorjonen2010). An example of a neutral, widespread feature is final vowel deletion and deletion of i from unstressed diphthongs in words such as puna(i)nen ‘red’ (Mielikäinen Reference Mielikäinen1986:234). Some features that were originally regional then spread to become used widely in spoken language, such as using the passive verb form with the first person plural pronoun (Mielikäinen Reference Mielikäinen1986:235). These features can be considered part of a general spoken language (in Finnish: yleispuhekieli). Importantly, much of the movement that supported development of this general spoken language occurred from the 1950s to the early 1970s. This was a period of significant emigration and rapid rural depopulation as over 600,000 people left farming for newly created industry jobs in cities. Thus the language situation was relatively stable throughout the 1950s, but by the 1960s and 1970s there was rapid movement from rural to urban centers. This is notably after the time period in which individuals were interviewed for the LAX corpus. Most of those interviews were conducted in the 1960s and targeted older, non-mobile, rural individuals. This means that the speech documented was prior to rapid change, and captured examples of what had been relatively stable regional dialects.

To examine the dialect background of American Finnish speakers, we categorized the speech of speakers into the traditional eight dialectal categories based on the presence or absence of particular dialect features. We considered Finnish dialects in distributional terms, meaning that several of the classic features were assessed along with the degree to which they each contributed to the speaker’s dialect make-up. In practice, this means that a speaker’s dialect identification using the model is based on the consistency and frequency with which they use particular dialect variants over other varieties. We chose specific dialect features for the model based on their usage in previous studies (Rapola Reference Rapola1969, Itkonen Reference Itkonen1989, Lainio Reference Lainio1989, Wiik Reference Wiik2004, Nuolijärvi & Sorjonen Reference Nuolijärvi and Sorjonen2010), but also considered pragmatic factors which impacted the feasibility of coding – both within the data and in operationalizing the features so that they could be extracted from the LAX corpus in a fashion that makes comparison between the datasets possible.

Dialect features under study included those that are considered classic staples in the differentiation of Finnish dialects, spanning morphological, lexical, and phonological features. We included what are considered classic distinguishing features (such as /ts/ and /d/ variables, first person pronouns, and intrusive schwa vowel) and added additional features that helped improve the accuracy of the model. The model does not include some features that rely heavily on transcriber judgment or otherwise would have been difficult to operationalize between datasets (such as other lexical items, as well as diphthong opening and reduction). For each feature, every possible realization in a speaker’s transcript was identified and coded.

The rest of this section outlines the features used in the model and their realizations across different dialects. Importantly, we use classifications common for the dialect regions in the 1960s and prior to the post-urbanization movement which contributed to rapid language change and dialect leveling. Section 4 addresses the usage of the features by individual speakers, especially those which contributed most to the model’s overall classifications.

3.3.1 Morphological

Inessive case form

While the standard form -ssA is attested across all varieties of spoken Finnish today, it was unattested in some regions in the 1960s. In Finnish, it is one of six locative cases and used to indicate place, as in example (1), or attachment, as in example (2), as well as some expressions of time, seen in example (3):

Table 2 indicates the distribution of usage of each variant, showing how the same variant may be found across multiple different dialects. The R scripts, anonymized datasets, and distribution tables for all dialect features included in the study are available in an online repository.Footnote ³

Table 2. Distribution of variants of inessive case forms

The variants of this case form with one -s as well as with final vowel apocope are considered identifying features of western dialects. The deletion of the final vowel is considered a feature of the Southwest and South Ostrobothnian dialect groups. The variant -hnA refers to an ending found only in the South Ostrobothnian dialect on words with a possessive suffix such as talohnani for talossani, ‘in my house’. This particular usage was not attested in our data from American Finnish speakers.

Verb plural forms

The first person plural verb suffix in Standard Finnish is -mme. While this is often used in spoken language, the final vowel is sometimes a/ä (an alternation henceforth indicated as A) in some western dialects. Furthermore the passive is often used in spoken Finnish more broadly, resulting in sentences like those in examples (4a) and (4b).Footnote ⁴ In sum, the possible verb plural forms are -mme, -mmA, -VVn.

3.4 Methodology: Supervised machine learning using Random Forests

The methodological procedure comprises two sets of experiments, which, in turn, both have a validation experiment followed by an exploratory experiment focusing on the dialectal background of the heritage language speakers of Finnish. In the validation experiments we train two models. The first model predicts the dialectal background of individual speakers of Finnish based on their use of the above-discussed dialect features in their speech. The model outputs, for each of the eight dialect regions considered, the probability of the speaker hailing from that region. The second model, in turn, applies a double regression approach, as it predicts separately the longitudinal and latitudinal coordinates of each speaker’s dialectal home location, based on the use of the same above-discussed dialect features. Hence the output comprises the coordinates for the speaker. Both models are trained and tested using data from the Finnish Dialect Syntax Archive (LAX). Then, provided that the thus obtained models fare relatively well in their respective tasks of the validation experiment, we use these models in the exploratory experiments to study the dialectal profile of five individual American Finnish speakers.

Both models are based on Random Forests (henceforth RFs) as a supervised machine-learning algorithm (for the algorithm, see Breiman Reference Breiman2001). RFs are essentially large ensembles of inference trees, and they can be used for either classification, where the goal is to sort the data at hand into classes of a categorical response variable, or in regression, so as to predict the value of a numerical response variable (for various linguistically oriented examples on the use of RFs, see e.g. Tagliamonte & Baayen Reference Tagliamonte and Harald Baayen2012, Ivaska Reference Ivaska2015, Deshors & Gries Reference Deshors and Gries2016, Ivaska & Bernardini Reference Ivaska and Bernardini2020, Ivaska & Ivaska Reference Ivaska and Ivaska2022). In RFs, the values of the predicting variables (here, the distributions of the studied dialect features) are used to split the data (here, individual speakers of Finnish) according to the given response variable (here, either the dialect region or the coordinates of their geographical location). For example, if speakers hailing from the Savo region consistently use the pronoun mie for the first person relatively more often than speakers from other regions, the model learns to use this information and uses it when splitting the data into different branches of the inference tree. Such trees are constructed by randomly choosing a subset of the studied features to find the best predicting variables. This random sampling is repeated many times to find those predicting variables that are most consistent in the task at hand. This way, the models rely more on the consistently well-predicting features and place less emphasis on the less consistent features. In RFs, these weights can be explored in terms of the variable importance measure, a metric that helps in understanding the degree to which different included variables (here, dialect features) contribute to the outcome of the obtained models. In the present study, we use the ranger implementation of RFs (Wright & Ziegler Reference Wright and Ziegler2017). For the maps, we use the rnaturalearth and rnaturalearthdata packages (South Reference South2017a,b), and for measuring the distances between the true and predicted locations, we use Haversine great circle distances as implemented in the geosphere package (Hijmans Reference Hijmans2021).