Code-Switching or the mixing of multiple languages in a single conversation (abbreviated as CS) is a very well-known phenomenon of language contact. Examples (1) and (2) show CS within a sentence between English and French from Cook (1991) and a tweet from Montreal. In both sentences, the switch occurs from French into English (henceforth CS-Engl.):

CS can also occur from English into French as shown in the following tweet in (3) from a bilingual photographer from Montreal (henceforth CS-French):

Bilingual or multilingual speakers can use two or more languages without necessarily using Code-Switching, as shown by examples (4) and (5) from the same bilingual user. This particular user is from Montreal and has posted 23 tweets in French and 22 in English.

The influence factors on Code-Switching (CS) as in (1–3) and on Language Choice of Bilinguals (LCB) as in (4) and (5) have been investigated from various perspectives, such as the structural (Poplack, 1980; Cook, 1991; Myers-Scotton, 2002), social (Schweda, 1980; Poplack, 1985; Fishman, 2000; Bullock and Toribio, 2009; Gardner-Chloros, 2009; Holmes and Wilson, 2013; Valenti, 2014), cognitive (Müller, 2017; Kremin et al., 2021, among others), and discourse perspectives (Gumperz and Dell, 1972; Konidaris, 2004; Auer, 2007; Auer and Eastman, 2010). One important factor influencing CS and LCB is the relative size of speech communities (Schweda, 1980; Poplack, 1985), alongside other factors such as age, addressee, topic of discussion, and language attitude (see Berisso Genemo, 2022 for various factors). CS from French into English, for instance, has been shown to be more prominent in the city of Ottawa than in the city of Hull (Poplack, 1985). This difference has been explained by the difference in the size of the anglophone population (Poplack, 1985). Ottawa has a larger anglophone community than Hull. An intuitive explanation for this correlation is that the larger anglophone community in a city like Ottawa results in a higher amount of language use in English. This can affect the speakers of French from the same city in their language use by being exposed to a large amount of English use. Previous research has shown that frequency strongly affects language production (Unsworth, 2016, among others).

Schweda (1980) has made an observation similar to that of Poplack (1985) with respect to LCB. Bilingual speakers adapt to the local environment by choosing the language that is more often used in a given city or area. Bilinguals thus tend to use English more often than French in a city where more English than French is used (Schweda, 1980). The intuitive explanation of this effect is that multilinguals adapt their language use to their local environment in a way similar to how they adapt their language use to their addressees (Konidaris, 2004).

However, there are studies on LCB and CS that seem to contradict this intuitive correlation between the size of a language community and language use such as CS and LCB. Lamarre et al. (2002) study shows that despite the Canadian census data from 2011 predicting more LCB (English) in the west of the island of Montreal than in the east (Statistics Canada, 2011) because it shows a relatively higher number of anglophone speakers in the west than in the east (see Timiou, 2014 for visualization of Statistics Canada, 2011), language use does not depend entirely on the geographic distribution of the population (Lamarre et al., 2002). Bilinguals from Montreal use both languages independent of their location, especially in informal contexts such as on the street and in coffee houses (Lamarre et al., 2002). This is unexpected given the results from previous studies based on different methodologies ranging from tweet analysis to picture analysis of street signs, which show a clear geographic separation of languages in Montreal, indicating more English in the west and more French in the east (Bouchard, 2000; Laur, 2003; Termote, 2003; Mocanu et al., 2013; Leimgruber and Fernández-Mallat, 2021). However, the latter studies did not investigate language use by bilinguals as the authors in Lamarre et al. (2002) study did, which might explain the difference in the results. One possible explanation for the conflicting results in previous studies such as that of Poplack (1985) and of Lamarre et al. (2002) is that they are based on few location points and/or few bilingual speakers, which is very likely related to the challenge of data collection. CS, for instance, is a spontaneous phenomenon and is more often used in informal contexts. CS is almost never used in legal documents or other highly formal contexts. In order to collect natural language data with CS, natural language needs to be collected in authentic and informal communication contexts. This requirement excludes the use of many linguistic corpora that contain news articles, linguistic questionnaires, or survey-based methods. The latter two methods are based on asking a selected group of people about their language behavior in a specific context, mainly “at home” (Bouchard, 2000; Laur, 2003; Statistics Canada, 2011). People cannot be asked under what circumstances they code-switch and how often they do it, as CS often occurs spontaneously and speakers are not always aware of when they are code-switching. For this reason, a different methodology and a different data source are needed to test correlations between the size of the population of a particular city or neighborhood and CS or LCB from the same location.

The present study aims at clarifying the influence of the population size on CS and LCB by using tweets from Twitter associated with location information and user IDs that are necessary for identifying bilinguals and for measuring the intensity of CS and LCB per geographic area. This information will provide answers concerning whether the intensity correlates with the size of the speech communities in a given location. The size of a speech community will be taken from population data published by Statistics Canada (2011).

The outline of the paper is as follows. Section 2 expands on the General Hypothesis with several more operational hypotheses in more detail by looking at predictions from Statistics Canada (2011). Section 3 presents the data source and the methodology to test the detailed hypotheses from Section 2. Section 4 shows the results, and section 5 discusses the results and future research plans.

According to Figure 1 from Statistics Canada (2011), there is a much higher percentage of anglophones on the island of Montreal and in the city of Gatineau (a city on the border to Ontario), than in the city of Quebec.

According to Figure 1, the anglophone population is larger on the island of Montreal than in Greater Montreal (GM), which includes the island and the surrounding area. According to Statistics Canada (2011), visualized by Timiou (2014) on a geographic map, the anglophone population on the island of Montreal is larger in the western part of the city than in the eastern part. The population differences (+/−Greater Montreal) and urban zones (+/−western part of the island) should have an effect on language use.

These five hypotheses will be tested in section 4 after the section Data and methodology.

One of the major contributions of this paper was to test the General Hypothesis, repeated here below, which has led to conflicting results in previous studies (Poplack, 1985 vs. Lamarre et al., 2002):

The results of CS-Engl. in cities seem to show a trend confirming the General Hypothesis when comparing entire cities (see Hypothesis 1). However, not all cities have yielded enough data to be able to confirm the General Hypothesis due to low numbers of tweets. This result can be improved in the future. One possible improvement is to extend my word-pair list with additional word pairs or to use a completely different approach for measuring CS that does not rely on manually crafted word-pair lists (Mendels et al., 2018, among others). In the future, it will be necessary to test the automatic language classification of tweets and/or of words to improve the identification of CS and LCB. In this paper, I relied on the language classification performed by Twitter. Twitter’s language classifier from 2015 labels English Tweets with 99% precision (see Twitter website²). However, Twitter also mentions that minor languages and tweets with mixed languages such as English and French have a lower degree of precision (Mendels et al., 2018).

One important issue related to the identification of CS concerns the definition of CS. In this paper, as well as in many other computational linguistic approaches to CS (Mendels et al., 2018 for an overview), CS is defined very broadly as mixing of languages and thus also includes any kind of mixing such as formulaic expressions or “conventionalized” CS and language translations used in the same tweet such as We are open! Nous sommes ouverts! In linguistics, however, formulaic expressions are distinguished from “proper,” “genuine” or “spontaneous” CS (Poplack, 1980, among others). I will address this issue in detail in future research.

Another important result of this study is the contribution to testing the General Hypothesis at the level of Greater Montreal (island of Montreal and periphery) and on the sub-city level (east and west of the island) using precise geolocation information. On this level of spatial granularity, the results are mixed. Looking at geographic patterns of CS and LCB in Figures 5A, 8A, there is a trend of spatial division as predicted by the General Hypothesis or more precisely by Hypotheses 2–5. However, by looking into the statistical numbers and testing frequency correlations per bin, the effect of spatial pattern visible in Figures 5A, 8A is rather weak. Most locations do not show a big difference in frequency distributions of LCB (English and French) and of CS (+/–CS), especially in locations with higher frequency numbers. In the future, I will use methods to test spatial correlation in the data to see whether locations with preferences for English are clustered together (see Anselin, 1995 for Local Moran’s I).

This study has shown that there is considerably more English tweets than French tweets and considerably more Code-Switching in French tweets than in English tweets in Greater Montreal (see Figure 3). This observation suggests that language use on Twitter is not entirely predicted by population numbers of speech communities as shown in Figure 1 from Canadian statistics. The higher numbers of tweet counts in English in Figure 3 strongly indicate that English is the more dominant language on Twitter, which can be related to various factors such as English being used as Lingua Franca in social media (Laitinen and Lundberg, 2020). The observation of spatial patterns of CS-Engl. in cities of Quebec and spatial patterns of LCB and CS in Greater Montreal indicates that despite the dominance of English in the counts of tweets, the geographic context also influences language use to some extent, more strongly at city than sub-city level. This implies that the mechanisms of the digital and non-digital language contact are not the same.

In the future, I will test the influence of social contexts as defined by buildings or location contexts on CS and LCB. For instance, the airport has been shown to play an important role for +CS-Engl. (Figure 8B), but the airport is not the most important context for English tweets posted by bilinguals (Figure 5B). Instead, bilinguals tweet slightly more in French than in English at the airport. However, in order to compare similar contexts such as all coffee bars or airports requires information about the social use of buildings and urban districts. Lamarre et al. (2002) notice a difference in language use in formal/informal contexts by observing bilinguals’ language use as they move through the city. One could test the consequence of the difference between formal and informal contexts and CS or LCB in the future on the basis of tweets’ content. Formal and informal contexts usually correlate with different topics of discussion. Formal contexts often contain information of interest to the general public such as information about vaccination or elections. Informal contexts more often contain personal information of interest to particular user groups or people such as information about personal events or things. This hypothesis will be tested in the future on methods tailored for a topic analysis of tweets.

Another topic that needs to be explored in the future is user variation with respect to CS and LCB. User IDs can be used to find all tweets produced by the same user and then to sort users according to the intensity of Code-Switching and/or their language preferences. This method allows us to find users who are resistant to CS or who use CS very frequently. The geolocation of tweets produced by a specific user group, such as a user group with a high preference for English or French or users without a language preference, allows us to classify these user groups with respect to the locations they visit. Figure 14 shows an analysis that explores user IDs to classify user groups according to their language preferences and to classify users by their visited location. The results show that users with the highest preference for English are dispersed throughout the city (marked by red points), whereas users with no language preference are concentrated in the city center (marked by green points). One tentative explanation for the concentration of users without a language preference is the assumption that the majority of users from the city center are entrepreneurs of some kind, who tweet in both languages to address as many clients as possible, including both French and English speakers. This hypothesis predicts that tweets from this user group should more often correspond to translations or paraphrases of the same content in different languages. This prediction can be tested by looking at temporal features of the tweets, assuming that they will very likely be posted at a similar time and by analyzing the content of the tweets. These methods will allow us to study ways of describing users and to contribute to sociolinguistic studies (Labov, 2006).

Another question that needs to be addressed in the future is how the factor “time” influences the distribution of language use. The present study measures the distribution of language use in a particular period, namely on the basis of tweets from 2017 to 2021. It is possible that users from Montreal pass their time in different locations depending on the period (year or day). However, if we compare the language distribution of bilinguals in Greater Montreal performed on the dataset used in this study with the language distribution of all users in Montreal performed by Mocanu et al. (2013) on an earlier dataset, we see a similar effect: English is used in the west and French is used in the east. This result suggests that the temporal difference in years does not matter much for language choice. In order to confirm this conclusion, though, other periods need to be considered.

Finally, more research needs to be done to investigate the extent to which bilingual users on Twitter represent bilingual speakers outside Twitter. It is also important to note that users allowing their geolocation to be seen publicly represent only a small proportion of all users on Twitter (Kruspe et al., 2021). It is thus quite possible that the location differences observed in this study are only representative for a particular user group on Twitter. In order to address this issue, various methods have been suggested in the research on Twitter (Kruspe et al., 2021 for an overview). The representativity of Twitter users allowing their geolocation to be visible will be investigated in the future.

To sum up, this study has shown that sociodemographic factors measured by a speech community’s size tend to influence CS and LCB on various spatial scales—the city and sub-city scales—but to a different extent. Geolocated tweets offer the possibility of studying the influence of location on language behavior. In order to maximize the benefit of social media data for sociolinguistic research, more techniques are needed to be able to analyze large quantities of data. This includes location classification by their social use and classification of tweets by user, topic, and language.

The methodology of this article is based on OK’s previous work in Kellert and Matlis (2022).

The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding author.

OK: data collection, formulation of the hypotheses, design of testing the hypotheses, discussion of results, and article writing and editing.

OK acknowledges the support given by the Open Access Publication Funds of the Göttingen University and by the German Research Foundation (DFG) (Grant number: 468416293).

The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.