Data-driven identification of situated meanings in corpus data using Latent Class Analysis

4.1 Data and coding procedures

I extracted all n = 1,233 cases of tener que + infinitive, n = 84 cases of deber + infinitive, and n = 52 cases of deber de + infinitive from the Peninsular Spanish section of the PRESEEA. Together, these tokens add to a final dataset of n = 1,369 cases.

PRESEEA is a dialectal corpus of semi-structured spoken sociolinguistic interviews in Spanish. The Peninsular Spanish section of the PRESEEA includes 88 interviews of about 916,000 words dated between 1988 and 2011, recorded in Alcalá de Henares, Granada, Madrid, Málaga, and Valencia. The PRESEEA is an adequate corpus for the investigation of the three research questions raised in the previous section because (a) it includes spoken, relatively informal, language and (b) it is a socially stratified corpus that represents variation in terms of age, gender, and SES (education and current job) of the speakers.

The data were coded manually in terms of (a) the difference between different modal meanings assumed in the literature and (b) ten contextual parameters that were assumed to be indicators of these meanings.

The manual coding of types of modal meanings was carried out using the standards described by Larsson et al. (2020). First, a coding scheme was established in which three types of modal meanings that tener que + infinitive and deber (de) + infinitive can express were distinguished. Deontic obligation readings imply that external circumstances force the referent of the subject to realize an action (9). Deontic necessity readings imply that the referent needs to realize an action to fulfil a desired goal (10). Finally, probability readings can be described in terms of epistemic necessity; i.e., the proposition is presented as hypothesis (11).

From the n = 1,369 cases of tener que + infinitive and deber (de) + infinitive, a random selection of n = 100 was coded by both the author and a native speaker of Spanish in terms of this coding scheme. Inspection of the results revealed an Inter-Rater-Agreement score of 0.67. The commonly used Inter-Rater-Agreement scale by Landis and Koch (1977) considers scores between 0.61 and 0.80 to suggest ‘substantial agreement’ (Larsson et al. 2020, 243).

The data were also coded in terms of ten contextual predictors, which were hypothesized to indicate both the type of expressed situated meanings and the alternation between tener que + infinitive and deber (de) + infinitive on the basis of the previous literature on this topic (in particular, Blas Arroyo 2011, Rosemeyer 2017, Thegel and Lindgren 2020). These predictors describe properties of the subject (Person, SubjRef), verb (Aspect, Mood, Tense, and PredType), syntactic properties (Neg and Subordination), and pragmatic properties (Punctuation, InTurn). Table 1 summarizes these coding procedures, as well as the frequency distributions.

Before continuing to discuss the analysis itself, a brief discussion of the selection of these parameters as possible indicators of situated meanings is in order. In its original application for the analysis of sociological questionnaires, the selection of the possible indicators was relatively straightforward because it was limited by the elements in the questionnaire. When applied to corpus data, however, there does not appear to be an a priori limit to the linguistic indicators used in LCA. As mentioned above, in this article, I focused on indicators that previous literature believes to be proxies of the distinction between epistemic and deontic modal meaning precisely because this allows me to compare the relevance of modality for the description of the situated meanings identified by the LCA. This is not to say that my approach is exhaustive; in fact, adding more indicators may result in the identification of more situated meanings. Because an approach that includes (many) more indicators leads to increasingly complex models, it will necessitate more data and processing power. Finally, the description of the resulting latent classes will become more complex.

4.2 Analytical approach

I identified contextual profiles of the usage of tener que + infinitive and deber (de) + infinitive based on the ten variables using LCA. The analysis was carried out using the poLCA package (Linzer and Lewis 2011) in R (R Development Core Team 2024). The model estimated the posterior probability of membership in each class for each token in the dataset. Classes were interpreted on the basis of the average probability of each contextual predictor in a given class and compared to the manual coding of modal meanings of the tokens. Data and R code are publicly available on the OSF site for the project (Rosemeyer 2024).

The discrete class assignment of the selected five-class model to the data was subsequently used in a series of regression models that calculated the correlations between class membership and (a) selection of tener que + infinitive or deber (de) + infinitive, as well as (b) socioeconomic indicators for the speakers. These results are presented in Section 5.

4.3 Descriptive results

The first step of the analysis consisted of a LCA to identify the situated meanings of tener que + infinitive and deber (de) + infinitive in the dataset. I estimated models with up to ten classes without additional covariates. As proposed by Linzer and Lewis (2011, 6–7), a five-class model was chosen as the preferred model on the basis of the lowest BIC value (Table A1 in the appendix for model fit statistics). The discrete class assignment was used for the descriptive results showing the composition of classes. Figure 1 visualizes the composition of classes in terms of the ten contextual variables.

Figure 1

Posterior probability of predictor levels by latent class representing situated meanings.

A first approach to the description of these latent classes is to analyze their relationship to the manual coding in terms of modality. Table 2 summarizes the distribution of modal meanings in terms of the latent classes. Inspection of relative frequencies suggests that the latent classes partially represent the difference between the modal readings. In particular, Class 1 appears to mostly contain tokens expressing deontic obligation. While results for the other classes are less clear, deontic necessity readings are particularly frequent in Class 3 and epistemic necessity readings are relatively frequent in Class 4. Classes 2 and 5 mostly contain deontic obligation and necessity readings, with no clear specialization. The results of the LCA thus point to the necessity of considering more specialized situated meanings that can only partially be explained in terms of the modal meanings of tener que + infinitive and deber (de) + infinitive. Note also that the relative class sizes of the latent classes differ greatly. Classes 2 and 3 have the highest population share, followed by Classes 4 and 5. Class 1 has the lowest population share.

Let us now turn to the description of the latent classes in terms of situated meanings, starting with the classes that contain a high proportion of cases coded as expressing deontic obligation. While this qualitative analysis is inspired by notions from Conversation Analysis and Interactional Linguistics, it is important to mention that I cannot analyze the examples from my corpus in the same detail. Rather, I will try and derive the situated meanings on the basis of the description of the commonalities of the contextual configurations identified by the LCA.

Class 1 represents a contextual configuration in which the verb typically has first-person morphology and the subject refers to an animate referent. The verb is almost always inflected for perfective past and indicative mood and expresses a telic action. Compared with the other classes, the utterance is more likely to be coded as a declarative main clause, although there is a slightly higher proportion of exclamatives than in all other classes except Class 2. Both the probability for the utterance to be negated is lower, and for the utterance to be turn-initial is lower in Class 1 than in all other classes.

As a typical case for a Class 1 meaning, consider example (12).^[1] In examples of this type, the speaker is describing a past action as motivated by exterior circumstances. The reasons for this delegation of responsibility can be due to the fact that admitting to the action frequently threatens the face of the speaker. Dropping out of school or sending your child to a residence, as in example (9), are actions that are viewed as undesirable. In other examples, the actions are simply unexpected in terms of the ongoing narration, likewise calling for attribution to external causes. While the meaning of tener que and deber (de) + infinitive in this class can easily be described as a modal meaning, namely deontic obligation, I define the situated meaning expressed by Class 1 tokens as ‘delegation of responsibility’. Note that in terms of the description by Thegel and Lindgren (2020), this type of situated meaning expresses subjective modality, as the attitude in question is not shared by A and B but by A alone.

Class 2, the latent class with the highest population share, represents a contextual configuration that resembles Class 1 in various respects. In particular, the verb typically has first-person morphology and the subject consequently refers to an animate referent, although the posterior probability for first-person plural morphology is much lower than in Class 1. As in Class 3, the verb is more likely to express telic and sometimes atelic actions and to be inflected for past tense than in the other latent classes. However, in Class 1, it is much more likely to be inflected for present tense than in Class 2. Moreover, the verb in Class 3 configuration is typically inflected for the imperfective aspect. As in Class 1, negation is very unlikely in Class 2 configurations, the utterance is typically annotated as a declarative, and the utterance is unlikely to be turn-initial.

Given that the contextual configuration of Class 2 and Class 1 readings is similar, the readings are also expected to be similar in various respects. However, a close look does show that there are differences in these situated meanings. The tener que + infinitive token in (13) exemplifies Class 2 readings.

The main difference to Class 1 readings appears to be that Class 2 readings typically describe prospective actions. The link between deontic readings and futurity is well-known in Spanish and other languages (Bybee et al. 1994, 184–5, Schäfer-Prieß 1999, Hernández Díaz 2017). Indeed, in an example such as (13), tener que + infinitive seems to be quasi-synonymous with a future construction such as me voy a ir un poco más tarde ‘I will leave a bit later’. When used with imperfective past tense morphology, the action is prospective with regard to the deictic center. In example (14), the action encoded with tener que + infinitive is prospective as regards the described situation in the past.

Both in examples (13) and (14), the action described by tener que + infinitive is used to motivate a previous move by the speaker. Thus, in (13), the fact that E has to leave a bit later is used to explain why B declined A’s invitation to give her a lift. As in (12), tener que thus redresses a face-threatening act. In (14), the action encoded with tener que gives the reason for the fact that B and her family only stayed a little while at the village. I define the situated meaning expressed by Class 2 tokens as ‘motivation of another move’.

Crucially for my argument, examples such as (13) and (14) are difficult to code in terms of the distinction between deontic obligation and necessity, as is reflected by the fact that Class 2 contains a significant share of cases coded as expressing either meaning (Table 2). Indeed, whether or not the subject referent needs to realize an action due to external or internal circumstances is irrelevant as regards the situated meaning of motivating a previous move.

The contextual configuration of Class 5 presents an even more balanced distribution in terms of the coding of modal meanings. In Class 5, the verb is typically inflected for third person plural and the subject refers to a collective referent. The verb is inflected for indicative present or imperfective past tense and expresses a telic action. The utterance is likely to be coded as a non-negated declarative main clause in non-initial turn position. Example (15) is prototypical for Class 5 configurations.

In (15), A uses the tener que + infinitive construction to make a suggestion. This suggestion is motivated in terms of a type of intersubjective necessity; the speaker believes that it is in the best interest of society that TV moderators appear serious and thus speak norm-oriented Spanish. Previous studies describe this type of meaning as ‘modal necessity’ (López Izquierdo 2008, Hernández Díaz 2017, 210–1) or ‘weak obligation’ (Bybee et al. 1994, 186–7). In line with Bybee et al.’s description, weak obligation readings represent situations in which the fact that an obligation is not fulfilled, the consequences are not as serious as in strong obligation cases. This is clearly the case in Class 5 readings. I define the situated meaning expressed by Class 5 tokens as a ‘suggestion of a future collective action’.

Example (16), belonging to Class 5, shows that the situated meaning of suggesting a future collective action is not necessarily bound to the use of third-person morphology. The speaker A suggests that society should abolish the lottery because it is in their best interest. The first-person morphology is used to express inclusive reference, i.e., refer to both the speaker and society as a whole. Examples (15) and (16) furthermore demonstrate that suggestions of a future collective action are clearly intersubjective readings; in both cases, the speaker assumes that “the attitude in question is shared between the speaker and a larger group of people” (Thegel and Lindgren 2020, 5).

Let us now turn to the description of Class 3, the second largest latent class, which moreover contains a relatively large population of tokens coded as expressing deontic necessity. In Class 3, the verb is typically inflected for the second-person singular and the subject expresses impersonal reference. The verb is inflected for indicative present tense and expresses a telic action. The utterance is likely to be coded as a declarative main clause in a non-initial turn position. Negation is more likely in Class 3 than in all of the other latent classes. As a typical example for this class, consider (17).

In Class 3, impersonal reference is expressed using the so-called tú impersonal, whose use is much more frequent in spoken Spanish than the use of the more formal uno ‘one’ (DeMello 2000, Guirado 2011, Kluge 2016, Posio 2017). As described by Kluge (2016, 502), the use of the second person invites speaker B to insert themselves in the position of the person fishing for trouts. This use of the second person clearly expresses an intersubjective meaning very much similar to Class 5 readings; the steps for catching a trout described in (17) are necessary and ‘reasonable’ (Thegel and Lindgren 2020, 13) not only for the speaker but also for the hearer and indeed, everyone.

The situated meaning instantiated by Class 3 configurations is also similar to the one in Class 5. In (17), speaker A makes a suggestion about a type of action that is required in all cases where one is trying to catch a trout. The genericity of the statement is indicated by the use of present tense morphology. Class 3 readings thus express weak obligation in the sense of Bybee et al. (1994, 186–7). In contrast to Class 5 readings, however, Class 3 readings are generic and do not express future reference. Furthermore, Class 5 readings assume that the action is useful and beneficial to the entire community, whereas Class 3 does not. I define the situated meaning expressed by Class 3 tokens as a ‘suggestion of an impersonal action’.

As was described above, Class 3 tokens are more likely to be negated than all other latent classes. Example (18) illustrates that negation does not seem to affect the situated meaning represented by Class 3 configurations; A simply expresses a suggestion about an action that one does not need to realize in the generic situation of traveling in the city. Examples of this type are described as realizing an intersubjective meaning by Thegel and Lindgren (2020, 14–5).

Finally, let us discuss the type of situated meaning expressed by Class 4 configurations, which differ from the other latent classes in various respects. Recall that Class 4 unites a majority of tokens coded as expressing epistemic necessity. In Class 4 contexts, the verb typically has third-person singular morphology and the subject refers to an inanimate referent. The verb is almost always inflected for the present tense and very frequently expresses a state. Although the indicative mood is extremely frequent, Class 4 is more likely to involve conditional morphology than the other latent classes. As in the other latent classes, the utterance is frequently coded as a declarative main clause; however, in Class 4 configurations, the probability of use of question marks is higher than in the other latent classes. Likewise, the utterance to be turn-initial is higher in Class 4 than in all other classes.

Example (19) is typical for this type of contextual configuration. The speaker uses the tener que construction in order to postulate a hypothesis inferred from the previous context. Thus, based on the fact that the village inhabitants asked the helpers to leave, she infers that they are likely to be very proud about their achievements. This situated meaning, which I will call ‘postulation of a hypothesis’, consequently has to be described as a subjective meaning, since it represents an inference drawn by the speaker, which is not necessarily shared by the other participants in the interaction.

As shown by example (19), the hypotheses postulated in Class 4 contexts frequently concern mental attitudes because these attitudes are inaccessible to observation and need to be derived via inference. As a result, they typically serve evaluative functions. This explains why tener que and deber constructions in Class 4 configurations are more likely to be used at the beginning of a turn than in other latent classes. Consider example (20).

A’s utterance tiene que ser horrible is a reaction to B’s utterance hay gente casi con treinta años que no se van nunca, ¿no?, which is designed to elicit negative evaluation by A. In terms of the terminology from Conversation Analysis, this relationship can be described in terms of preference structure (Stukenbrock 2013, 234). The utterance tiene que ser horrible is a preferred reaction to B’s utterance and is consequently used turn-initially.

Finally, the fact that Class 4 configurations are more likely to be annotated as an information request can be explained in terms of the situated meanings expressed in these contexts. Due to the epistemic uncertainty involved in the postulation of a hypothesis, the speaker frequently asks the hearer for confirmation of this hypothesis. This correlation, which was already observed in a previous study on future tense and epistemic modality in Spanish (Rosemeyer and Sansiñena 2022), is exemplified in example (21).

Table 3 summarizes the description of the five latent classes in terms of situated meanings and the parameter of subjectivity in my data. In the next sections of this article, I will show how the relevance of social parameters for this data-driven description of situated meanings and how using LCA significantly improves the description of the variation between tener que + infinitive and deber (de) + infinitive.

5.1 Analytical approach

The analysis of the variation proposed in this article entails a major change to the standard practice of describing variation adopted in variationist analysis and probabilistic grammar. The description of situated meanings as social practices by the speakers developed in Section 3 implies that the use of a specific linguistic form is motivated by the speaker’s intent to express a certain situated meaning. Consequently, it is conceptually inconsistent to treat the variation between use of tener que + infinitive and deber (de) + infinitive as the dependent variable in a multivariate analysis. Rather, the periphrases are recruited by the speakers as contextual predictors of the situated meanings, in the same way as the other predictors that were used in Section 4.3 to establish the five latent classes. I consequently used a multivariate statistical approach in which the contribution of the use of tener que + infinitive and deber (de) + infinitive to the five latent classes was measured.

Moreover, this approach entails a new perspective on how to model the role of socioeconomic indicators for the variation between tener que + infinitive and deber (de) + infinitive. The third research question, developed in Section 4, asked to which extent the variation between tener que + infinitive and deber (de) + infinitive is governed by social speaker characteristics. If the use of tener que + infinitive and deber (de) + infinitive is a contextual predictor of the situated meanings modeled as five latent classes, this research question needs to be answered in terms of the manner in which the use of these periphrases to express the situated meanings is moderated by socioeconomic speaker status. In other words, it is necessary to investigate SES in interaction with tener que + infinitive and deber (de) + infinitive as predictors of latent classes. This can be schematically illustrated as in (22), where ‘∼’ means ‘predicted by’ and ‘:’ designates an interaction effect between the two predictor variables.

Note that in (21), SES is modeled as a single predictor, although SES is measured using several variables such as income, education, etc. Just like the situated meanings, the classification of speakers into groups representing SES was operationalized using LCA. This is standard practice in social demography (consider, for instance, Lowthian et al. 2021, Hammami et al. 2022). The results from this LCA will be presented in the next section.

5.2 Latent speaker classes

SES of the participants was measured using five variables coded in the PRESEEA corpus for the speakers, which are described in Table 4, namely level of education, speaker age, sex, income, and role in the conversation. The SpeakerAge variable was coded using the distribution quartiles.

Note that this information is not given to all of the speakers in the corpus. For the analysis presented in the remained of this section, I consequently had to eliminate the data from n = 28 speakers, corresponding to n = 132 tokens of tener que + infinitive and deber (de) + infinitive in our dataset, from the data. The final dataset thus consisted of n = 1,237 tokens of tener que + infinitive and deber (de) + infinitive, produced by n = 98 speakers.

In the same manner as in Section 4, I identified latent speaker profiles based on five variables using LCA. The model estimated the posterior probability of membership in each class for each speaker in the speaker dataset. I again estimated models with up to ten classes without additional covariates. Based on the lowest BIC value, a two-class model was chosen as the preferred model (Table A2 in the appendix for model fit statistics). The discrete class assignment was used for the descriptive results showing the composition of classes. Figure 2 visualizes the composition of the two speaker classes in terms of the five contextual variables.

Figure 2

Posterior probability of predictor levels by Latent Speaker Class.

The two identified latent classes clearly distinguish speakers in terms of SES. Class 1 speakers generally have a lower degree of education and a lower income than Class 2 speakers. Class 1 speakers are overwhelmingly participants, whereas a majority of the speakers assigned to Class 2 are interviewers. This fact probably explains the age and sex effects: Class 2 speakers are more likely to be women in the age group between 42 and 59 than Class 1 speakers, whereas, in line with the aim of the corpus to be representative in terms of social factors, the group of Class 1 speakers is more balanced.

5.3 Multivariate results

In this section, I report results from two multinomial logistic regression analyses measuring the correlation between situated meanings expressed by the three periphrases under study (tener que + infinitive, deber + infinitive, deber de + infinitive) and two predictor variables: (a) the formal difference between the three periphrases and (b) the posterior probability for a speaker to be assigned to the second speaker class, which represents a higher SES. I was able to model SES as a numerical variable representing the posterior probability of assignment of the second speaker class due to the fact that only two latent classes were identified by the LCA in Section 5.2. This operationalization increases the statistical resolution for the SES variable. I also tested for an interaction effect between the two predictor variables. The coding of the predictor variables is described in Table 5.

I was interested in the extent to which a description of situated meanings as latent constructs improves existing descriptions of the tener que - deber (de) alternation. Consequently, I calculated two multinomial regression models that only differ in terms of the dependent variable. The dependent variable in Model 1 was the manual coding of the data in terms of the distinction between deontic obligation, deontic necessity, and epistemic necessity. The dependent variable in Model 2 was the latent classes representing situated meanings identified in Section 4.

The use of multinomial logistic regression analysis (Orme and Combs-Orme 2009, Ch. 3, Levshina 2015, 277–89) was necessary because the dependent variable was modeled in terms of a discrete assignment of either the coded modal meanings (Model 1) or one of the five latent classes representing situated meanings (Model 2) to each case in the data. The analysis was performed in R (R Development Core Team 2024), using the nnet package (Venables and Ripley 2002). No model selection process was conducted due to the fact that only two predictor variables were analyzed. For both models, all levels of the dependent variables were tested as reference levels. The full results for the two regression analyses are given in Tables A3 and A4.

Model 1, with the dependent variable Interpretation, found significant main effects for the predictor variables Periphrasis and SES. However, the model failed to find statistically significant differences between the use of tener que and deber de, as well as deber and deber de, in terms of the dependent variable Interpretation. In other words, the model only successfully explains the choice between tener que and deber.

As evident in Figure 3, the model predicts deontic obligation and epistemic necessity readings to be expressed using tener que. In contrast, deber is most likely to express deontic necessity readings. Note that tener que is actually more likely to express epistemic necessity than deber, an effect that reached statistical significance.

Figure 3

Marginal effects plot of the main effects of Model 1 (multinomial logistic regression with the dependent variable Interpretation).

Model 1 also predicts epistemic and deontic necessity readings to be more likely the higher the probability for the speaker to belong to Latent Speaker Class 2, representing a higher SES. This effect is significantly stronger for deontic than for epistemic necessity readings. Finally, the interaction effect between Periphrasis and SES did not reach statistical significance irrespective of the reference level of the dependent variable.

Model 2, which measured the correlation between the latent classes measuring situated meanings, found significant main effects for the predictor variables Periphrasis and SES, as well as the interaction term. Figure 4 visualizes these results as marginal effects.

Figure 4

Marginal effects plot of Model 2 (multinomial logistic regression with the dependent variable LC).

Regarding the main effects, Model 2 finds tener que to be particularly likely to express Class 2 (‘Motivation of another move’) and Class 3 readings (‘Suggestion of an impersonal action’). The use of deber and deber de is significantly less likely in these contexts. While tener que is not particularly likely to be used to express Class 1 readings (‘Delegation of responsibility’), it is significantly more likely to do so than deber and deber de. As to deber, the findings from Model 2 show this periphrasis to be particularly likely to express meanings classified as Class 4 (‘Postulation of a hypothesis’). Finally, deber de is most likely used to express Class 5 meanings (‘Suggestion of a future collective action’), although it is also significantly more likely than tener que to be used to express Class 4 meanings.

The main effect of SES is not easily apparent in Figure 4 because of the significant interaction with Periphrasis. Model 2 predicts that the higher the probability that the speaker belongs to Latent Speaker Class 2, which represents the highest SES, the more likely Classes 2, 3, and 4 readings are. As evident in Figure 4, this effect is moderated by the used type of periphrasis; it is significantly stronger for Class 2 readings if the chosen periphrasis is deber de + infinitive. In other words, the effect for Class 2 readings is mostly due to deber de + infinitive.

5.4 Discussion of results

In this section of the article, I discuss the findings from the multivariate analysis in terms of the three research questions developed in Section 4. RQ 1 asked to which extent the analysis of situated meaning as a latent construct improves descriptions of the tener que - deber (de) alternation. A direct comparison between Model 1, which uses manually coded modal meanings as the dependent variable, and Model 2, which uses latent situated meanings, in terms of the fit to the data is impossible due to the different dependent variables. However, there are at least two reasons to believe that Model 2 is superior to Model 1. First, Model 1 only successfully explains the choice between tener que and deber, whereas Model 2 finds significant effects for the choice between all three periphrases and consequently better explains the choice of these periphrases as indicators of situated meanings. Second, the significant interaction effect between Periphrasis and SES implies that Model 2 successfully models the interplay between social factors and periphrasis choice in the expression of situated meanings. The same interaction effect does not reach statistical significance for Model 1.

The comparison between Model 1 and Model 2 also contributes to answering RQ 2, which asked whether the variation between tener que, deber and deber de + infinitive is best explained in terms of modality or subjectivity, with mixed results. Given that model 2 found the use of deber de + infinitive to be most likely with intersubjective Class 5 readings, whereas the use of deber + infinitive is extremely likely with subjective Class 4 readings, it appears that intersubjectivity explains the use of deber de + infinitive to a greater degree than the use of deber + infinitive. While tener que appears to be less specialized in terms of its meaning potential, it is relatively more likely to be used to express the subjective Class 1 and 2 readings. Given that manual classification of the data in terms of modal readings turned out to not be able to explain the use of all three periphrases, the findings presented in this article can be taken to partially support Thegel and Lingren’s (2020) analysis of the variation between the periphrases in terms of subjectivity.

Finally, the analysis developed in this article allows us to study the extent to which variation in the three periphrases is governed by social speaker characteristics (RQ 3). Both Models 1 and 2 found that the opposition between two latent speaker classes, which was interpreted in terms of a difference in SES, significantly influences the types of meaning expressed by the periphrases under study. Higher SES speakers are significantly more likely to use the three periphrases to suggest impersonal actions (Class 3 readings) and postulate hypotheses (Class 4) than lower SES speakers. While the exact reasons for this finding need further study, a possible explanation is that lower SES is generally associated with a lower degree of agency, understood in terms of concepts such as independence, ambition, and dominance (Shanahan 2000, Evans 2002, Abele and Wojciszke 2013, 2014). Consequently, the fact that lower SES speakers are more likely to characterize situations as independent of their actions may be the result of their real or perceived lower sense of agency.

The analysis also showed this effect of SES on expressed meanings to be moderated by periphrasis type. In particular, the higher the SES, the more likely speakers are to use deber de + infinitive to motivate another move (Class 2 reading). Given that Class 2 readings were described as subjective, this result does not conform to the hypothesis, proposed in Section 4, that higher SES speakers conform to linguistic norms to a greater degree. Recall that the function of Class 2 readings was described in terms of the mitigation of a face-threatening act. Examples such as (23) and (24), uttered by a speaker classified as belonging to Latent Speaker Class 2, seem to suggest that higher SES speakers have conventionalized the use of deber de + infinitive as this politeness resource to a greater degree than lower SES speakers. Note that both examples involve a verbum dicendi, which is typical for Class 2 readings.