Present perfect and preterit variation in the Spanish of Lima and Mexico city: findings from a corpus analysis

2.1 Grammaticalization

The semantic instability of the PP is considered revealing of a more general process of semantic change in modern languages, i.e., grammaticalization. When a lexical form or grammatical item enters into grammaticalization, it gradually develops (new) grammatical functions, expanding to new domains of use (Bybee et al. 1994: 4; Heine et al. 1991: 2; Hopper and Traugott 2003: 1; Lehmann 2015: 13). An example is the English periphrasis “be going,” which was initially used with a literal meaning of “moving through space toward a goal,” but gradually acquired the additional meaning of future temporality, a process known as generalization of meaning (Bybee et al. 1994: 5; Hopper and Traugott 2003: 103). The two meanings, the spatial and the temporal, coexisted for some time (layering), but eventually, the former meaning demised (semantic bleaching or erosion). At this point, “be going” became a fully grammaticalized form expressing future. Such a development from spatial to temporal semantic value has been attested in several languages (Bybee et al. 1994: 268). The evidence that similar source material (e.g., verbs of movement) tends to undergo similar processes of semantic change across languages led to the hypothesis of source determination, according to which grammaticalization processes are not random but rather determined by the meaning of the construction that enters into grammaticalization.

One major finding in grammaticalization is unidirectionality, according to which once a grammaticalization process starts, it cannot be reversed. For instance, verbs of “movement toward a goal” tend to grammaticalize to futures cross-linguistically, but do not go back to their previous meaning (Bybee et al. 1994: 12–13). Similarly, it has been observed that, cross-linguistically, the PP first evolves from resultative constructions and then tends to acquire other semantic functions, being the encroachment into past perfective functions (typically carried out by means of the preterit) the most common outcome (Bybee and Dahl 1989; Bybee et al. 1994: 81; Squartini and Bertinetto 2000). However, the opposite direction of evolution is not attested (Bybee et al. 1994: 12). The hypotheses of source determination and unidirectionality predict that there are cross-linguistic clines, or universal paths, of semantic change (Bybee et al. 1994: 14). In other words, there are common patterns of semantic evolution of the linguistic material that enters into grammaticalization even in languages that are not geographically or genetically related, which may be due to “common cognitive and communicative patterns underlying the use of language” (Bybee et al. 1994: 14).

The validity of such hypotheses has been object of debate (Fløgstad 2017; Giacalone Ramat 1998; González et al. 2019; Janda 2001). On the one hand, Hopper and Traugott (2003: 139) state that the evidence for unidirectionality is systematic cross-linguistically, whereas counterexamples are sporadic and language-specific. On the other hand, Drinka (2017: 65–72) claims that there is weak evidence of the perfect-to-perfective shift as a universal process. She suggests that this linguistic innovation arose in the 12th and 13th centuries in the Parisian French vernacular in which the “perfect first came to be used to connote the past” (Drinka 2003: 20, 2017: 257). Drinka claims that this innovation later expanded across Europe in the areas that lay under the Parisian cultural influence (2017: 257–261). Therefore, Drinka indicates that the spread of the PP with preterit value can be better explained by areal diffusion and linguistic contact rather than by universal paths of development.

Most recently, Haspelmath (2019: 10) has critiqued Bybee's (2006) “common pathways of change in the tense-aspect domain.” Haspelmath specifically criticized Bybee’s (2006: 179–180) claim that “the true universals of language are the mechanisms of change that propel the constant creation and recreation of grammar.” Haspelmath (2019: 10) states that Bybee’s argument does not differentiate enough between “recurrent pathways of change” and the constraints that lead to such pathways of change. Haspelmath (2019: 3) claims that “universal tendencies cannot be explained by common pathways of change, only by change constraints.” He refers to change constraints as “mutational constraints,” which he proposes are two general types: source constraints and directionality constraints (2019: 6). The irreversibility of grammaticalization is the most well-known directionality constraint (Haspelmath 1999: 1043, 2004: 17, 2019: 6). As Haspelmath indicates, sources constraints, such as internal and external (i.e., language contact), impact the pathway of any grammatical change. Reinhöhl and Himmelmann (2017: 408) further echo the importance of source constraints in grammaticalization as “one needs to analyze change in context” in which “grammaticalization constructions are shaped by forms and structures they are built up from as well as by constructions they interact with.” Thus, it is important to consider source and directionality constraints in understanding pathways of change.

2.2 Present perfect and preterit variation

According to prescriptive grammars, the preterit expresses perfectivity; it denotes a situation temporally bounded in the past and reported for its own sake, while the PP indicates the present relevance of a past situation (Bybee et al. 1994: 54; Comrie 1976; Harris 1982).^[2] Comrie (1976) reports four cross-linguistic prototypical uses of the PP: resultative, which allows to infer a current state resulting from a past action; continuative, which marks past events that continue up to the present; experiential, which describes an event that occurred at least once in a certain period of time; and, recent past, which reports an event that occurred in a temporal interval that includes the moment of speech. As for Romance languages, the PP originated from a resultative construction attested in Pre-Classical Latin, as in the example by Harris (1982: 47), “Multa bona bene parta habemus” “Many goods well obtained we have” (Plautus, Trin. 347). Over time, these constructions lost their original resultative function to become a marker of a past event considered relevant to the present moment (Bybee and Dahl 1989). However, the indeterminacy of the notion of current relevance brought variable uses of the PP across Romance varieties (Cartagena 1999: 2945–6; Harris 1982).

Harris (1982) offered an account of the distribution of the PP in Romance varieties as being organized in four stages of grammaticalization. At stage I, the PP expresses a present state resulting from past actions (resultative function), as seen in Pre-Classical Latin. At stage II, it presents an aspectual function of durativity or iteration, referring to states and events that started in the past and continue up to the speech time. This stage is representative of Latin American Spanish varieties, Galician, and Portuguese. At stage III, the PP expands to an array of contexts that denote perfectivity, namely, an action or situation presented as completed in the past. This is what Squartini and Bertinetto (2000: 404) call the aoristic drift of the PP. Typically, the PP starts expanding to recent past contexts because they are more likely to be perceived as currently relevant. Several Peninsular Spanish varieties are at this stage. At stage IV, the PP expresses past actions regardless of the distance with the moment of speech, whereas the preterit is restricted to formal and written registers. This is the situation of French, Romanian, and Northern Italian. Of note is that Harris considered the varieties at stages III and IV as being “more developed” (i.e., some varieties of Peninsular Spanish), whereas those at earlier stages (i.e., varieties of Latin American Spanish) will eventually reach the more developed stages. However, Squartini and Bertinetto (2000) interpret Harris’ stages as independent developmental paths of a grammaticalizing PP. Therefore, a variety that is at stage II will not necessarily move toward stages III and IV. Similarly, Schwenter (1994: 78) claims that PP forms can be “frozen in their development” and that synchronic investigation is needed “to determine their ongoing development, if any, and to advance prognoses of future grammaticalization.” In the following sections, we provide an overview of PP/preterit variation in four varieties of Spanish: Peninsular Spanish that exhibits the most PP realizations with a grammaticalization in progress; Mexican Spanish that exhibits the most preterit realizations without signs of grammaticalization; and Peruvian and Argentinian Spanish that exhibit a high degree of PP variation different from other Latin American Spanish varieties.

2.3 Present perfect and preterit variation in Spanish

2.4 Research questions

Building on previous studies, the current endeavor aims to further explore the social and linguistic factors that affect the PP/preterit distribution in Lima, Peru Spanish (henceforth, LPS), including Mexico City Spanish (henceforth, MCS) as a point of comparison. Mexico Spanish is considered prototypical of the Latin American norm of use of the PP, whereas Peruvian Spanish shows an increased use of PP as compared to other Latin American varieties, although not as frequent as Peninsular Spanish (Howe and Scott 2008). To further explore this issue, we analyzed the PRESEEA^[8] Corpus (2014–) sociolinguistic interviews conducted in Lima and Mexico City. The current study sought to answer the following research questions:

RQ1: What is the overall frequency of PP and preterit in LPS and MCS? While we expect the preterit to be the most common form in both varieties, we hypothesize that the PP will be significantly more frequent in LPS than in MCS based on previous studies (Howe and Scott 2008; Rodríguez Louro and Jara Yupanqui 2011; Schwenter and Torres Cacoullos 2008).

RQ2: What linguistic factors govern the use of the PP in LPS and MCS? We hypothesize that the temporal reference will be the most significant factor in governing the PP/preterit distribution, in which the irrelevant reference will favor PP and the prehodiernal will favor the preterit as found previously (Howe and Scott 2008; Schwenter and Torres Cacoullos 2008). Moreover, we expect plural direct object to favor the PP as they favor iterative meanings (Howe and Scott 2008; Schwenter and Torres Cacoullos 2008). The presence of ya “already” is expected to favor the preterit in both varieties (Howe 2013; Howe and Scott 2008; Schwenter and Torres Cacoullos 2008).

RQ3: What social (extra-linguistic) factors govern the use of the PP in LPS and MCS? Is there any evidence of a change in progress in either speech community? As only a few studies (Bateman 2022; Howe 2013; Jara Yupanqui 2006) examined how social factors constrain PP/preterit variation in Peruvian Spanish and found contrasting results, there was no formal hypothesis as there was no reason to assume a priori that social factors would govern PP variation.

3.1 Participants

A total of 2,534 tokens of verbs in PP and preterit were coded from the transcriptions of 36 sociolinguistic semistructured interviews in the PRESEEA Corpus. 18 participants were from Lima and 18 from México City (online Appendices A and B for participant demographics). The PRESEEA corpus has 18 speakers for each speech community that are socially stratified by gender (male, female), age (20–34, 35–54, 55+), and educational levels (basic, medium, high). Table 1 summarizes the characteristics of participants in each speech community.

The semi-structured sociolinguistic interviews were collected between 2005 and 2007 in Mexico City and between 2008 and 2009 in Lima. All interviews were recorded in quiet environment such as a classroom, the interviewee’s house, or an office, with the exception of one interview in Lima (conducted in a street) and another in Mexico City (in a tianguis “street market”). The interviews are structured in eight thematic modules (online Appendix C) about each speaker’s life to encourage naturalistic speech following PRESEEA methods (Moreno Fernández 1996). While the Lima interviews closely follow this structure, the Mexico City interviews are a bit more variable in content. The interviews conducted in Lima present a similar length (40–50 min), whereas the Mexican ones are longer overall (60 min) but also present more variability in length. Despite these methodological discrepancies, the interviews were transcribed following the same coding procedure, and the same amount of participant demographic information is provided; therefore, we consider the two data sets comparable.

3.2 Dependent variable, data collection, and data coding

The alternation between preterit and PP was coded as a binary dependent variable. Coding began at the start of each interview. Although we included all instances of verbs conjugated in the preterit or PP, we considered a number of contexts outside of the “envelope of variation” (Milroy and Gordon 2003), including progressive structure estar + gerund (1), morphologically ambiguous (preterit or present) -AR and -IR verbs in first person plural (2), false starts (3), following previous studies (Poplack and Tagliamonte 2001; Schwenter and Torres Cacoullos 2008), and direct repetitions of the interviewer’s question (4). These contexts were, therefore, not included in the analysis.

After excluding these tokens, there were 2,534 tokens. However, due to too few tokens per level, we eliminated the hodiernal temporal reference level for the temporal reference variable (16 Mexico City, 19 for Lima) as well as the adverbial of recency level of the adverb type variable (7 Mexico City, 35 Lima). With so few tokens in these levels, we could not reliability claim a main effect or lack thereof when they were placed into the regression models, as it could skew the data with a Type I or Type II error. Thus, the final analysis was conducted with 2,457 total tokens (1,230 México City; 1,227 Lima; M: 68/speaker).

3.3 Independent variables

The study included three social variables, four linguistic variables, and speaker as a random factor. Following PRESEEA norms (Moreno Fernández 1996; PRESEEA 2014-), participants were divided by age group ^[10] according to three generations: younger (20–34), middle-aged (35–54), and older (55+); gender: male and female; and educational attainment, which included basic (0–5 years of schooling), medium (10–12 years of schooling), and high (15+ years of schooling). This study treats education only as educational attainment without reference to social class as operationalizing social class is quite complex, as it includes different measures of economic power, social prestige, and status (Ash 2013), and varies per the context of each community.

As for linguistic variables, we selected four factors that have been shown to be significant predictors of PP/preterit variation (Howe and Scott 2008; Schwenter and Torres Cacoullos 2008): temporal reference, temporal adverbial, presence of ya, and direct object. The choice of these linguistic variables is motivated by the methodological need of avoiding the notion of current relevance (Torres Cacoullos 2011), as it is impossible to establish objectively which tokens are currently relevant above intuitive characterizations (Schwenter and Torres Cacoullos 2008). Similarly, coding the tokens according to the four prototypical functions of the PP (resultative, continuative, experiential, and recent past; see Comrie 1976) is problematic when dealing with spontaneous speech (Howe 2006, 2013; Schwenter and Torres Cacoullos 2008).^[11] Linguistic variables such as the temporal reference and the presence of certain temporal adverbials are more objective measures of the constraints that govern PP/preterit variation. Each independent variable is explained below.

First, temporal reference included four levels. (i) Hodiernal: the event occurred in the same day in which the speech is taking place^[12] (5). (ii) Prehodiernal: the event occurred in the day before the day of the speech, or in a more remote past (6). (iii) Indeterminate: the analyst is unable to determine when an action occurred (7). It is possible to ask “when?” to the speaker to resolve the ambiguity; however, the speaker does not need to specify the temporal reference for their discursive purposes (Torres Cacoullos 2011: 160). (iv) Irrelevant: it cannot be queried when something occurred, as the event lacks a specific temporal relevance, as in the case of permanent states that continue up to the present, repeated actions, yes/no questions, and negative polarity (when an event has not occurred) (8).

Second, temporal adverbial included five levels. (i) Specific time adverbials: they answer to the question when? (9). (ii) Adverbials of recency: specific time adverbials that express that the event is perceived as recent (i.e., este verano “this summer,” ahora “now”) (10).^[13] (iii) Atelic adverbials: frequency and continuous time adverbials, which answer to questions: for how long? or how many times? (i.e., siempre “always,” a veces “sometimes,” hasta “until”) (11). The atelic context for temporal adverbials was almost always coded as irrelevant for the independent variable of temporal reference. Thus, these two independent variables are not fully independent, but the current study follows previous studies (Howe and Scott 2008; Schwenter and Torres Cacoullos 2008) in incorporating both independent variables. (iv) Connective adverbials: typically used to join the parts of a narration (i.e., después “after,” entonces “then,” luego “then”) (12). (v) Absence of adverbial: there is no temporal adverbial next to the verb (7). Atelic adverbs, which express duration and frequency, are expected to favor the PP, whereas specific time and connective adverbials are expected to favor the preterit (Howe and Scott 2008).

Third, presence of the adverb ya “already,” included two levels. (i) Present: ya was present in the sentence (13); (ii) Absent: there was no ya (14).

Fourth, direct object, included three levels. (i) Singular: the direct object was singular (15). (ii) Plural: the direct object was plural (16). (iii) Absent: there was no direct object (17).

3.4 Statistical analysis

Following the variationist comparative method (Poplack and Tagliamonte 2001: Chap. 5; Tagliamonte 2012) and previous PP/preterit variation studies (Schwenter and Torres Cacoullos 2008), the current study conducted separate multivariate analyses for each speech community. This allows for one to understand the underlying linguistic and social factors governing the use of the PP versus the preterit in each community, comparing similarities and differences. For each data set, following Tagliamonte and Bayen (2012), prior to any regression models, a random forest was calculated using the cforest function from the party package (Hothorn et al. 2020) to determine variable importance. This allows the researcher to order in importance each independent variable in the regression modeling, as order impacts the output of the regression. Following the random forest, a mixed-effects logistic regression model was fitted for each speech community using the lmer function (Bates et al. 2015) and lmerTest (Kuznetsova et al. 2017) in R (R Core Team 2022). The marginal R-squared (R ² m) and conditional R-squared (R ² c) values are listed to provide a goodness-of-fit of the variation for each final model (Nakagawa and Schielzeth 2013). All independent variables were tested in the model construction. Nonsignificant interactions were discarded from each subsequent model. For categorical independent variables with more than two levels, post-hoc analyses were conducted with estimated marginal means (Lenth et al. 2018). Different models were compared using ANOVA testing, relying on the lowest AIC being the best explanation of the data. While several interactions were included in the model, in general these models did not converge. Thus, to explore the relationship between significant main effects, a conditional inference tree was conducted using the cforest function from the party package (Hothorn et al. 2020).

4.1 Overall descriptive statistics

Of the 1,230 tokens in MCS, 111 (9.0 %) were realized in the PP and 1,119 (91.0 %) in the preterit. Of the 1,227 tokens in LPS, 303 (24.7 %) were realized in the PP and 924 (75.3 %) in the preterit. Figure 1 visualizes the proportional differences in realizations between communities. A Chi-square test revealed a significant difference between the two speech communities in which speakers from Mexico City produce significantly less PP than speakers from Lima, χ ²(1) = 106.5, p < 0.0001.

Figure 1:

4.2 Mexico City

The random forest results indicate that for PP/preterit realizations, the most important predictor is temporal reference, followed by direct object, gender, speaker, education, age (continuous), adverb type, polarity, age group, and the presence of ya (Figure 2).

Figure 2:

The best-fit mixed-effects logistic regression is presented in Table 2, displaying the estimate, standard error (SE), t-value, percent of PP per level, number (n) of tokens per level, and p-value. As the model is in reference to PP, positive estimates indicate that the listed factor level favors PP more than the reference level and negative estimates indicate that the listed factor level favors preterit realizations more than the reference level. Reference levels are displayed in the table.

The model demonstrated a significant main effect for temporal reference in which the irrelevant temporal reference most favored PP, followed by the indeterminate temporal reference, and then finally the prehodiernal temporal reference (Figure 3A). A post-hoc analysis of the estimated marginal means revealed that the irrelevant temporal reference significantly favored PP realizations more than both indeterminate (p < 0.001) and prehodiernal temporal references (p < 0.001). Additionally, the indeterminate temporal reference also favored more PP realizations than the prehodiernal temporal reference (p < 0.01).

Figure 3:

The model also demonstrated a significant main effect for direct object in which plural objects most favored PP realizations, followed by singular objects, and then lastly, no objects following the verb (Figure 3B). A post-hoc analysis of the estimated marginal means revealed that the only statistically significant difference between the three levels was that the plural object significantly favored PP realizations compared to no object (p < 0.01), but no significant differences compared to singular objects (p = 0.13). Furthermore, there were no significant differences between singular objects and no objects (p = 0.07).

To further explore the relationship between the two significant main effects, a conditional inference tree was conducted (Figure 4). The conditional inference tree indicates that temporal reference is the most significant predictor in which the irrelevant temporal reference demonstrates the most PP. When considering the prehodiernal and indeterminate temporal references, only the indeterminate context shows variation as the prehodiernal temporal reference demonstrates nearly categorical preterit. For the indeterminate context, it interacts with the presence of a direct object. Specifically, the indeterminate temporal reference with a plural direct object favors more PP than those of a singular or no direct object.

Figure 4:

4.3 Lima

The random forest results indicate that for PP/preterit realizations, the most important predictor is temporal reference, followed by speaker, age group, age (continuous), education, direct object, adverb type, gender, ya, and polarity (Figure 5).

Figure 5:

Age group and continuous age were placed into different regression models and compared to determine which version on this independent variable best explained the variation. The best model for both age group and continuous age included main effects for temporal reference, adverb type, ya, and either age group or continuous age. These model were compared using ANOVA testing, but there was no significant difference between the models and virtually identical AIC values. The R ² c values were identical as well (0.31), with only the R ² m values being slightly higher for the model with age group (0.23) compared to that of continuous age (0.22). Given this difference, as well as the higher variable importance value given to age group in the random forest, it was decided to use age group in the final regression model. However, follow-up analyses below explore continuous age.

The best-fit mixed-effects logistic regression is presented in Table 3, displaying the estimate, standard error (SE), t-value, percent of PP per level, number (n) of tokens per level, and p-value. As the model is in reference to PP, positive estimates indicate that the listed factor level favors PP more than the reference level and negative estimates indicate that the listed factor level favors preterit realizations more than the reference level. Reference levels are displayed in the table.

The model demonstrated three significant main effects for the linguistic factors. First, the model exhibited a main effect for temporal reference in which the irrelevant temporal reference most favored PP, followed by the indeterminate temporal reference, and then finally the prehodiernal temporal reference (Figure 6A). A post-hoc analysis of the estimated marginal means revealed that the irrelevant temporal reference significantly favored PP more than both the indeterminate (p < 0.001) and the prehodiernal temporal reference (p < 0.001). Additionally, the indeterminate temporal reference also favored PP more than the prehodiernal temporal reference (p < 0.001). Second, the model demonstrated a main effect for adverb type in which the presence of an atelic adverb most favored the PP (Figure 6B). However, a post-hoc analysis of the estimated marginal means only revealed one significant difference between levels in which specific adverbs favored the PP more than the connective adverbs (p < 0.05). Third, the model demonstrated a main effect for ya in which the absence of ya favored the PP more than the presence of ya (p < 0.01) (Figure 6D).

Figure 6:

Different from the MCS, the model for LPS data demonstrated one significant main effect for a social factor: age group. Specifically, those in the youngest age group (18–34) produced less PP than those in the middle age group (35–54) and the oldest age group (55 and older) (Figure 6C). Those aged 18–34 produced statistically significant less PP that those aged 55 and older (p < 0.05) but not statistically less than those aged 35–54 (p = 0.71). There were no significant differences in PP realizations between those aged 35–54 and those aged 55 and older (p = 0.11).

To further explore the relationship between the four significant main effects, a conditional inference tree was conducted (Figure 7). The conditional inference tree indicates that temporal reference is the most significant predictor in which the irrelevant context demonstrates the most PP production. Then when considering only the prehodiernal and indeterminate temporal references, only the indeterminate context shows variation as the prehodiernal context nearly categorically exhibits the preterit. Thus, the variation is most seen in the indeterminate temporal reference as it interacts with age group, ya, and adverb type. The first branch here is with age group, in which the oldest age group overall shows higher rates of PP, but even more so with all adverb types that are not connective. Then among the younger age groups, it depends upon the presence of ya in which the presence of ya favors more preterit, while the absence of ya shows less preterit, but this interacts with adverb type in which non-ya contexts with atelic, connective, and specific adverbs favors more preterit than non-ya contexts with no adverbs. However, as one can observe, the interactions with adverb type appear to be weak significant effects, with values just under p < 0.05, and thus these should be taken with caution.

Figure 7:

As mentioned previously, the model above opted for age group based on its higher R ² m value in the regression model and higher variable importance in the random forest than continuous age. However, as regression models showed no statistical difference between categorical and continuous age, here we conduct follow-up analyses with continuous age. Below we first visualize all the data points with continuous age (Figure 8A). A Pearson correlation was conducted and found a significant weak positive correlation between age and PP realizations (n = 1,227, df = 1,225, r = 0.15, R ² = 0.02, p < 0.001). Then we visualize the data as the percent of PP usage per individual speaker; thus, one data point per person (Figure 8B). A Pearson correlation was conducted and found a nonsignificant moderate positive association between age and PP frequency (n = 18, df = 16, r = 0.46, R ² = 0.21, p = 0.054). As this result was nearly significant, a follow-up Pearson correlation was conducted without the outlier LIMA_M11 (26-year-old, female, basic education), who although was part of the youngest group demonstrated the highest PP usage of all speakers. The follow-up Pearson correlation indicated a significant moderate positive association between age and PP frequency (n = 17, df = 15, r = 0.625, R ² = 0.39, p < 0.01).

Figure 8:

A conditional inference tree was conducted to explore the relationship between continuous age and three significant linguistic factors (temporal reference, adverb type, ya) for PP/preterit variation (Figure 9). Similar to the previous conditional inference tree, the most important factor is temporal reference in which the irrelevant context shows the most PP. Then the prehodiernal temporal reference shows nearly categorical preterit. Similar to the other model, within the indeterminate temporal reference, variation here is predicted by age and ya. Specifically, the model branches based under and over 47 years of age, showing that those over 47 years of age produce the most PP even in the indeterminate temporal reference. Then those 47 years of age and younger use more preterit in the presence of ya as compared to the absence of ya within the indeterminate temporal reference. Adverb type does not appear to interact with continuous age as it did age group. Of note is the model’s branching at 47, as opposed to the categorical value of 55 or older, indicating why the regression models yielded similar results.

Figure 9: