From sociolinguistic perception to strategic action in the study of social meaning

2.1 Design and materials

We developed a textual role-playing video game called L’installation à Paris ‘Moving to Paris’ in the Ren’Py engine (Rothamel 2022). A small open world was created around the city of Paris, where the player has to complete a series of quests (i.e. tasks) in order to stabilize their precarious situation: in the introduction to the game, it is explained that the player has recently arrived in the city and is sleeping on a cousin’s sofa. The player must complete four tasks: getting an apartment, opening a bank account, getting the attention of a server in a French café, and visiting a museum. At first the tasks cannot be completed, and the player has to explore and meet two non-playable characters (NPCs). Then, the player must request help from one of these NPCs in order to complete each task. To avoid biasing participants with nonlinguistic clues (clothing style, etc.), no visual indication of what the NPCs look like is ever presented, and only through the text can their personality, upbringing, and social status be imagined. One NPC was Monsieur Martin, an old-fashioned Parisian bourgeois who makes statements expressing status values: valuing family and tradition (Thévenot and Boltanski 1991; Woolard 1985). The other was Anthony, a down-to-earth middle-aged Parisian who makes statements valuing solidarity. Solidarity can be instantiated in concrete social interactions in many ways. Here, we took inspiration from Woolard (1985), who argues that the solidarity dimension is key for understanding sociolinguistic dynamics in situations of political conflict surrounding territory. We therefore had Anthony demonstrate solidarity values by resenting the gentrification of northeastern Paris and the recent displacement of less wealthy people from his neighborhood. A complete flowchart summarizes the course of a game in Figure 1.

Figure 1:

Flowchart of the game.

The first manipulated experimental condition behind the game was the linguistic guise of each NPC, which was introduced during the presentation dialogue of each, in the location that unlocks after all tasks have failed once. In this dialogue, both NPC either used a “standard” or “prestige” variety of French (standard guise; Bourdieu and Boltanski 1975) or a “vernacular” or “covert prestige” variety (vernacular guise; Trudgill 1972). This dialogue was the only moment in the game where the NPC guises were accessible. Table 1 illustrates the linguistic points of variation between the two guises. All of these variables have been independently shown to be socially meaningful (Ashby 1977; Coveney 2005; Thiberge 2020). Figure 2 exemplifies how the two guises differed for different participants (here, during the introduction dialogue for Anthony).

Figure 2:

Screenshots of the same piece of dialogue from Anthony: ‘I don’t get this world anymore …’. Top, vernacular guise; bottom, standard guise.

The second manipulated condition was the social context in which tasks had to be completed: all four places where the player had some quest pending also had two versions where different social values were put forward in the description. For instance, one player saw a status-oriented version of the bank, where financial stability and personal responsibility were explicitly expressed by a clerk, while another player saw a solidarity-oriented version of this location, where the clerk expressed solidarity and the need to take into account the very different personal situations of the clients. The “context” condition was laid out during the first time the player explored each location and before they failed each task, with three sentences alternating across conditions that described the social values important for each location. A one-sentence reminder of the social attributes for each context was presented to the player when they came back to the location with one of the NPCs to complete the task. Each player saw two status-oriented and two solidarity-oriented contexts (see Figure 3 for the two different versions of the bank location).

Figure 3:

Two versions of the bank. Top, solidarity-oriented: ‘In a solidarity bank, atypical cases are treated the same as any other.’ Bottom, status-oriented: ‘In life the rules are clear so that everything can run smoothly.’

Due to the length of the game it is not possible to provide an exhaustive presentation of all the text shown to players during a game.^[1] Participants had access to only one version of the game in a Latin square design, which allows for a good counterbalancing of experimental conditions. The textual descriptions of all four places and the guises of both NPCs can be found in the supplementary materials (https://osf.io/am5ur/).

2.2 Variables and predictions

Our independent variables are the guises of the NPCs and the status/solidarity-oriented versions of each location. Our dependent variable is the choice made by participants when they were prompted to select one of the two helping NPCs in each situation (recoded 0 = Anthony, 1 = Martin).

We only analyze here the first choice made by participants in the first situation. This amounts to 96 data points, which makes the dataset quite small. Our statistical analyses will thus use Bayesian logistic regression models rather than frequentist regression models, because Bayesian modeling is better suited for smaller datasets (Sorensen et al. 2016). Analyzing only the first choice excludes potential training effects from one quest to the next and does not take into account whether the choice allowed the quest to be completed or not. By “training effects” we mean that some participants may have worked out that two choices had to be “Martin” and two choices “Anthony” to win the game. Given that we did not constrain where participants made their first choice, and that they could fail multiple times in a row if they chose the same NPC in the same location over and over, players’ progression throughout the complete course of a game appeared too difficult to model.

Our predictions are based on existing sociolinguistics research, particularly classic works such as Bourdieu and Boltanski (1975), Woolard (1985), Labov (1973), and Trudgill (1972), and more recent experiments like Campbell-Kibler (2007), Podesva et al. (2015), and Beltrama et al. (2023):

1. Prediction 1: Monsieur Martin expresses status values and will be preferred to Anthony in contexts valuing status. We assume that players will be (mostly) rational and choose the NPC that will give them the best chance to win the encounter. The solidarity-oriented NPC Anthony will be preferred in contexts valuing solidarity. However, we predict that the linguistic guises could perturb this pattern.

2. Prediction 2: When Monsieur Martin is in his vernacular guise, we hypothesize that players’ preference for Martin in status contexts will decrease, compared to when he appears in his standard guise. Likewise, players’ preferences for Martin in solidarity contexts will be even more pronounced when Martin appears in his vernacular linguistic guise compared to his standard guise.

3. Prediction 3: Conversely, we predict that Anthony will be even more preferred in solidarity contexts when he appears in a vernacular guise, compared to when he appears in a standard guise. Likewise, we predict he will be more preferred in his standard guise in status-oriented contexts than in his vernacular guise.

2.3 Participants and procedure

Ninety-six participants were recruited on Prolific (http://prolific.com). All were self-declared L1 adult speakers of French from France, with 48 over 30 years old and 48 under 30. They were asked to play the game, but nothing was made explicit about the experimental variables and conditions. Players were told they could explore the city through a clickable map of Paris and skip through text descriptions and dialogues via the space bar of their keyboard. They were told that they could go to several locations indicated by stars on the map, and that they would be prompted with text choices, without any indication as to whether these choices were good or bad, and only when a quest was finished did they see any form of progress. Players were informed that they had to complete all main quests for the game to finish.

Participants typically took 12–30 min to complete the game, with some outliers (10 min–1 h). Participants’ progression was tracked automatically in a new text file as soon as they started a game, with timestamps and important choices being recorded (i.e. did the participant select option 1 or option 2 at a particular choice point). Participation was anonymous, with a unique but unreadable ID attributed to every new game, in accordance with ethical guidelines.^[2] Participants received standard compensation (approximately £5 per 30 min).

Then we went through each text transcript of a game and coded all the meaningful choices that participants had made and all relevant data (sociolinguistic profile and experimental conditions).

2.4 Results

We analyzed our data with the R suite version 4.3.2 (R Core Team 2023) and RStudio (Posit Team 2023), within the Bayesian framework (logistic regression modeling; see supplementary materials for full specifications and the list of R packages we used).

2.4.1 General results

In line with prediction 1, we find that Martin is chosen more often in status-oriented contexts than in solidarity-oriented contexts (Figure 4). Anthony is chosen much more frequently in solidarity contexts than he is in status contexts, but Martin was still chosen (slightly) more than half of the time in solidarity contexts, which is surprising.

Figure 4:

Choices by context (overall).

However, further inspection of the different contexts shows that all status/solidarity contexts do not behave in the same way. For one thing, players’ first choices are not equally balanced across the bank (n = 34/96), café (n = 31/96), museum (n = 20/96), and apartment (n = 11/96), since we left players the freedom to explore our open world. Furthermore, the bank location, in both its status and solidarity versions, highly favored Martin (Figure 5). This is in contrast to the other locations, where the solidarity version favored Anthony (or there was no preference).

Figure 5:

Choices by location and context.

2.4.2 Results: Martin’s guises

To analyze how this general pattern was affected by the linguistic guises of both NPCs, we first present how choices for Martin were related to his own linguistic guise (Figure 6).

Figure 6:

Choices by context and guise (Martin).

With Bayesian modeling (model mb1.ctx.guiseM), we found several meaningful effects even with the relative scarcity of our data. First, there is robust confirmatory evidence for an overall effect of context: Martin was more chosen in status-oriented contexts (β̂ = 1.66, 95 % CrI [0.60, 2.86], P(β > 0) = 1). Second, there is robust evidence for an overall effect of guise, with Martin being more chosen in his standard guise (β̂ = 1.24, 95 % CrI [0.20, 2.38], P(β > 0) = 0.99). Third, we also find some evidence for an interaction between context and guise: the preference for Martin in status contexts over solidarity contexts is even greater when he appears in his standard guise than when he appears in his vernacular guise (β̂ = 1.22, 95 % CrI [−0.88, 3.60], P(β > 0) = 0.86). In this way, the effects of the social meanings of the linguistic variants can be observed on the strategic choices another person will make in interaction with them.

It should be noted that these results do not capture the interaction between both NPCs guises, which we discuss below.

2.4.3 Results: Anthony guises

Another way to look at the data is to look at how the choices for Martin were modulated by Anthony’s sociolinguistic variants (Figure 7).

Figure 7:

Choices by context and guise (Anthony).

In another model (mb1.ctx.guiseA, where the dependent variable was still the proportion of Martin choice), we find the same robust confirmatory evidence for an overall effect of context (Martin was more chosen in status-oriented contexts; β̂ = 1.41, 95 % CrI [0.44, 2.45], P(β > 0) = 1). We also find robust evidence for an interaction between context and guise in this model (β̂ = −1.46, 95 % CrI [−3.53, 0.50], P(β < 0) = 0.93). This means that Martin is more highly preferred in status contexts when Anthony is in his vernacular guise than when Anthony is in his standard guise. Put another way, Anthony is chosen more often in solidarity contexts when he is in his vernacular guise, and he is chosen more often in status-oriented contexts when he is in his standard guise. Again, these results do not yet include the possible interaction between NPCs guises.

2.5 Discussion

The results of the video game study show that linguistic variants can affect the strategic choices that participants make, opening up a new way to study their social meanings. Our predictions are mostly borne out. To begin with prediction 1, the nonlinguistic social properties conveyed by NPCs affected how they were chosen. Martin is chosen more frequently than Anthony in status-oriented contexts and, although Martin is still slightly preferred in solidarity-oriented contexts, Anthony is chosen much more frequently in solidarity-oriented contexts than in status-oriented contexts. The linguistic guises of both NPCs nuanced these general findings.

Prediction 2 was partially borne out: players had an even greater preference for Martin in status-oriented contexts when he appeared in his standard guise. However, he was still chosen more often in status contexts in his vernacular guise, suggesting that, while his language did modulate participants’ perceptions of him, using vernacular language over standard language was not sufficient to compensate for the status-oriented content of his speech.

Prediction 3 was borne out. Anthony is more preferred in solidarity-oriented contexts when he appears in his vernacular guise than when he appears in his standard guise. Likewise, he is more preferred in his standard guise in status-oriented contexts than in his vernacular guise.

From this, we conclude that studying social meaning through looking at strategic choice, as in our video game, can reveal a wide range of results that are consistent with sociolinguistic theories. We now compare these results with those of a more traditional matched-guise experiment.

3.1 Design, materials, and participants

The experiment took place on a university-hosted instance of the IbexFarm platform (Drummond 2016). Forty-eight different self-declared adult L1 speakers of French from France (24 aged 30 or less, 24 older than 30) from Prolific read the textual^[3] descriptions of the status/solidarity contexts and the dialogues featuring the standard/vernacular NPCs (Monsieur Martin and Anthony).

All these were condensed into one text for each context/NPC, and we presented them with a Latin square and randomized design. Participants saw only one version of each place and one guise of each NPC, with a balanced number of standard-guise versus vernacular-guise NPCs and of solidarity-oriented versus status-oriented locations (6 items by list, no fillers). Participants had to give their impression on how important they thought some properties were in the location or for the character: education, tradition, hierarchy, and speaking “good” French (all four being status features), and solidarity and social justice (both being solidarity features).^[4] Participants answered on six corresponding 11-point slider scales (a frequent format in France, from school evaluations to general surveys, from 0 “not at all important” to 10 “very important”). We obtained 48 × 6 × 6 = 1728 answers in total. Considering the small number of items and the high variability we could expect on such social evaluations, we also opted for analyses with Bayesian models. Since our dependent variables were ratings on interval scales, we opted for cumulative-link models, best suited for ordinal data.

3.2 Results

Detailed analyses of the MGT results are provided in the supplementary materials (models mb1a.sopre for the NPCs, and mb1b.sopre for the locations). Figure 8 provides an overview of the social properties attributed to NPCs in their various guises, on the two kinds of scales (solidarity scales on the left, status scales on the right, mean and distribution of the ratings on the y-axis). On the solidarity scales, only Martin in his vernacular guise seems to yield higher results than all three other guises. In other words, the contrast between Anthony’s standard and vernacular guises that we observed in the strategic action paradigm was not observed in the MGT. On the status scales, we find a similar pattern, with Anthony showing no distinction between his two guises, although always being rated lower than Martin. Here again, it seems the contrast between the two linguistic guises that we observed in the video game is neutralized in the sociolinguistic perception task. On the other hand, the contrast between Martin’s two guises is observed in the MGT: as the sociolinguistics literature and the video game results would predict, Martin is rated higher on the status scales in his standard guise, and higher on the solidarity scales in his vernacular guise. This is captured in the model by the robust evidence for a three-way interaction between the NPC × scales × context variables (β̂ = 1.54, 95 % CrI [−0.34, 3.49], P(β > 0) = 0.95).

Figure 8:

Matched-guise task results for non-playable characters (black dot: mean; error bars: standard error; white lines within the violins: quartiles).

We find a similar pattern for the contexts: as shown in Figure 9, the status and solidarity versions of each context do indeed rate higher on the status and solidarity scales, respectively. However, we again see some neutralization for the café location. Although there is a trend in the expected direction (status properties rated as less important in the solidarity-oriented café compared to the status-oriented version), we find robust evidence for a three-way interaction between the location × scales × context variables in a model with museum as the reference location (and solidarity scales and solidarity-oriented contexts as references for the two other variables; β̂ = −2.91, 95 % CrI [−4.76, −1.03], P(β < 0) = 1). This indicates that the shift across scales is less important for the café than it is for the museum, bank, and apartment locations. Again, the lack of sharp differentiation in the MGT is surprising, because the two versions of the café are clearly distinguished in strategic action (Figure 5).

Figure 9:

Matched-guise task results for locations (black dot: mean; error bars: standard error; white lines within the violins: quartiles).