The problem of the definition of the variable context

If syntactic variation in language reflects the competition between alternative expressions of the same meaning, a variationist analysis of a syntactic phenomenon presupposes the identification of this meaning. This corresponds to the identification of the variable context, within which the competition between the alternative expressions takes place. According to the principle of accountability (Labov, Reference Labov, Lehmann and Malkiel1982:30), a quantitative analysis of the variation between different means of expression has to take into account all occurrences of this target variable. Thus, “the occurrence of variants can be calculated out of the total number of contexts in which it could have occurred, but did not” (Tagliamonte, Reference Tagliamonte2006:72).

This approach creates the obvious problem of whether two types of linguistic expressions can really have the same meaning. According to the criterion of weak complementarity, two variants in a variable context do not need to express exactly the same meaning, in other words, they do not have semantic equivalence. Rather, these variants merely need to express similar discourse functions, that is, they have discourse equivalence (Sankoff & Thibault, Reference Sankoff, Thibault, Johns and Strong1981:207).

Although this criterion offers a remedy to the theoretical problem of equivalence, it does not always eliminate the practical problems with the selection of the variable context. In particular, it is frequently difficult to determine which linguistic elements count as variants with discourse equivalence.

As an example, consider the expression of nonpredicative possession in Old Spanish. Most research has focused on the variation in the use of the definite article preceding the possessive marker (la mi casa ‘the my house’, henceforth ART + POSS), a structure that is absent from present-day standard Spanish, as opposed to possessive alone (mi casa ‘my house’, henceforth POSS). A first research tradition has focused on stylistic factors. It has been suggested that, because ART + POSS is a structure that emphasizes possession, it is used with stylistic functions such as expressivity, solemnity, poeticality, or reverence (Eberenz, Reference Eberenz2000:265–319; Lapesa, Reference Lapesa, Aguilar and Elizondo2000 [1970]). A second research tradition focuses on the influence of structural factors; for instance, Wanner (Reference Wanner, Kabatek, Pusch and Raible2005:39–40) pointed out that the first and second person, or singular possessors, as well as possessive structures embedded in prepositional phrases, favor the use of the ART + POSS construction. Finally, syntax-discourse factors have been taken into consideration: Company Company (Reference Company Company2006) claimed that ART + POSS is relatively more frequent when the possessor or the possessed referent has been mentioned in the previous discourse or when the possessed referent has a high degree of accessibility by virtue of a relationship of inherent possession.

Crucially, however, there are other constructions that also serve to express possession in Old Spanish, such as a genitive phrase with a personal pronoun (la casa de él ‘the house of him’, henceforth GEN), the strong possessive adjective construction, in which the possessive is postposed (la casa suya ‘the house his’), or even a simple determiner + noun construction (ART/BARE) as in levantó la mano ‘he raised the [=his] hand’. All of these structures can appear in contexts similar to those of the ART + POSS and POSS structures, and their appearance correlates with a complex set of structural and contextual factors such as lyrical register, ambiguity of reference, or cognitive prominence (Company Company, Reference Company Company1994; Eberenz, Reference Eberenz2000:299; Enrique-Arias, Reference Enrique-Arias2012b:827–828).

The wide range of constructions with a possible possessive function illustrates an obvious analytical problem with the identification of the variable context. Prima facie, it is impossible to know which of these constructions fulfill the criterion of weak complementarity. A comparison of the use of, for example, the ART + POSS construction with the use of the POSS alone presupposes that these two means of expressing possession in Old Spanish have a stronger functional similarity with each other than with the other constructional types. Such a restriction to a subset of constructional types, however, might lead to an overgeneralization of the effects that are encountered in the analysis. For instance, some studies suggest that the parameter “inherent possession” (i.e., the degree of conceptual proximity between the possessor referent and possessed referent) influences the alternation between ART + POSS and POSS and the alternation between these two constructions and ART/BARE. In other words, at least with regard to this criterion, one could argue that the two variants la su mano ‘the his hand’ (ART + POSS) and la mano ‘the hand’ (ART/BARE) are more similar to each other than, for example, la su mano and su mano ‘his hand’ (POSS). Crucially, this would mean that existing analyses that do not take into account all these constructions might miss a piece of the puzzle, as the commonality between the constructions could constitute a criterion of its own.

This analytical problem has consequences for diachronic analyses as well. Although the diachronic trend whereby the use of the referentially overspecified forms (ART + POSS and GEN) is eroded in Early Modern Spanish is clear, the uncertainty regarding the exact nature of the functional interdependencies between the constructional types in synchrony entails an uncertainty regarding the nature of the historical process. For instance, it is usually assumed that between Old and Early Modern Spanish, ART + POSS was replaced by POSS. However, if the degree of functional similarity of these two constructions relative to other possessive constructions is lower than expected, it is uncertain whether it indeed was the POSS construction that replaced ART + POSS.

The problem of the comparability of contexts

A second important problem for the study of syntactic change concerns the comparability of contexts. It is a universal challenge to historical linguistics that historical data are always fragmentary and unrepresentative, as the selection of texts that has survived until today is the product of chance. Longitudinal analyses of syntactic change, however, need language examples that differ with regard to the state of development of the language rather than their usage contexts. This methodological challenge has been formulated in terms of a comparability paradox in historical corpus design (Enrique-Arias, Reference Enrique-Arias2012a:97): a historical corpus has to be diverse because it must contain texts that represent different periods, genres, or dialects. At the same time, this corpus must be uniform in that the distribution of content type, genres, or dialects along the different chronological sections in the corpus must be as similar as possible so they can be compared. In quantitative studies of syntactic change, we therefore face sample-related problems concerning which texts to compare. Even if we restrict our sample to, let us say, narrative texts, we may well end up with works created under diverse textual conventions and in which the distribution of narration, description, and dialogue will differ considerably. For instance, Company Company's study (Reference Company Company2009) on the alternation of ART + POSS and POSS uses a corpus that includes epic poetry (the 12th-century Cantar de Mio Cid), historiographical texts (such as the 13th-century General estoria), a theater play (the 15th-century Celestina), a picaresque novel (the 16th-century Lazarillo de Tormes), and other diverse texts. In such data, there is no way to be certain to what extent the frequency changes attested in the analysis correspond to structural changes affecting the construction under study and not to divergences in the setup of the data for each time period.

Using parallel corpora in quantitative approaches to syntactic change

The data used in our analysis comes from Biblia Medieval, a parallel corpus of Old Spanish biblical translations (Enrique-Arias & Pueyo Mena, Reference Enrique-Arias and Pueyo Mena2008–2016). The corpus includes the original texts (the Hebrew Bible and the Latin Vulgate) and their translational equivalents in medieval Spanish. The texts are aligned so that it is possible to identify the pairs or sets of sentences, phrases, and words in the original text and their correspondences in the Spanish versions. The use of parallel or comparable corpora in contrastive studies (between languages, dialects, contact varieties, historical periods, etc.) today constitutes a well-established practice within both corpus linguistics (McEnery & Xiao, Reference McEnery, Xiao, Kawaguchi, Takagaki, Tomimori and Tsuruga2007) and sociolinguistics (Tagliamonte, Reference Tagliamonte2012:162).

One of the limitations of corpus linguistics methodology is that is gravitates toward the search of explicit markers considering a closed list of elements. If we want to study the expression of possession in Old Spanish by a search in a conventional corpus, the relevant usage contexts are identified by searching for the linguistic elements instantiating these contexts, such as the possessive adjectives mi ‘my’, tu ‘your’, su ‘his/her/your/its’, etc. In contrast, parallel corpus methodology is much more open as we can search for any element used to express a given function. For instance, in a parallel corpus such as Biblia Medieval it is possible to search for possessive pronouns in the Latin Vulgate (e.g., meus ‘my’, tuus ‘your’, eius ‘his/her/its’) and then observe how they are translated in the Spanish versions. As a concrete example, consider the translations of Maccabees 1 5:5 in (1).Footnote ¹

1. (1) et incendit turres eorum

   ‘and he burned their towers’ (MAC1 5:5)

   1. a. e  quem-ó =  les    las    torres (E6)

      and  burn-pst.pfv.3.sg = pro.dat  det.f.pl      towers

   2. b. e  Puso    fuego   a    las     torres (General

      and  put.pst.pfv.3.sg    fire    to   det.f.pl   towers Estoria)

   3. c. e  Encendió      las      torres   d’  = ellos (E4)

      and  ignite.pst.pfv.3.sg     det.f.pl    towers   of = them

Here the researcher can observe without limitations what linguistic structures are used to convey the functions expressed by turres eorum ‘their towers’ in this context: we find a dative pronoun and a noun phrase (NP) with a definite article in (1a), an NP with no explicit possessive marker in (1b), and a genitive phrase in (1c).

Another problem that parallel corpus methodology helps us to overcome is that of the comparability of contexts. In parallel texts, we have direct access to the evolution of linguistic structures, as translation equivalents are likely to be inserted in the same, or very similar, syntactic, semantic, and pragmatic contexts of occurrence. As a parallel corpus such as Biblia Medieval puts the discourse contextual factors largely in control, the behavior of the elements used to express possession can be observed and compared in a focused manner. Another interesting feature of the Bible is that it encompasses texts of varied textual typology: narrative, legislative, lyrical poetry, wisdom literature, epistles, and dialogues. As a result, the Biblia Medieval corpus is particularly appropriate to explore register variation, as it is possible to examine how the same translator selects language options that are appropriate for each of the genres represented in the Bible.

There are, nonetheless, some conceivable problems associated with the use of biblical translations in linguistic research. Because they are translated texts, they pose the risk of interference from the source language. As sacred texts, they may also exhibit stylistically marked language (i.e., deliberate archaisms). We must keep in mind, however, that most Spanish medieval texts in current corpora come from translations (from Latin, Arabic, Hebrew, French, etc.) or are subject to the strong Latinate influence that is characteristic of 15th-century Spanish writing; moreover, nontranslated secular texts such as legal documents, or literature, especially poetry, may be highly artificial as well. In sum, biblical texts are not necessarily worse sources of data relative to other medieval text types.Footnote ²

At any rate, the methodological reliability of using translated texts in linguistic research crucially depends on the nature of the phenomenon to be studied. In the case at hand, variation between POSS, ART + POSS, and ART/BARE in Spanish can hardly be affected by features of the original texts, as Latin has no articles and Hebrew does not employ them in possessive structures. As for GEN, although it is true that its appearance has been associated with the imitation of eius type structures in Latin models, this does not constitute an automatic calque: it has been shown that translators use this structure selectively to remedy the referential opacity of POSS in regard to gender and number of the possessor or to exploit its stylistic possibilities (Enrique-Arias, Reference Enrique-Arias2012b:828).

Finally, we must acknowledge the difficulties in providing clear-cut criteria for defining what possession is. Rather than a closed list of neatly defined structures, we find a continuum of constructions that goes from morphologically marked ones to others that are only discourse-inferable and could simply have a relational meaning. Herein, we will restrict our queries to the different translations corresponding to the paradigm of Latin and Hebrew possessive pronouns. This procedure will exclude, for instance, possessive constructions with a lexical NP possessor such as la casa de Juan (‘the house of John’).

Combining parallel corpus data with multinomial logistic regression analysis

Like other regression analyses, multinomial logistic regression (MLR) analysis allows us to investigate the relationship of a set of predictor variables to a dependent variable (Orme & Combs-Orme, Reference Orme and Combs-Orme2009:91–122). The main areas in which these analyses have been employed are psycholinguistic studies (e.g., Li, Schweickert & Gandour, Reference Li, Schweickert and Gandour2000; Ratcliff & McKoon, Reference Ratcliff and McKoon2001) and corpus semantics (e.g., Krawczak, Reference Krawczak, Novakova, Blumenthal and Siepmann2014). MLR analyses rely on the assumption of independence of irrelevant alternatives, which “means that your choice of Candidate A over Candidate B is not influenced by whether Candidate X joins the fray” (Orme & Combs-Orme, Reference Orme and Combs-Orme2009:118). The independence of irrelevant alternatives can be understood as referring to the problem of the variable context: we are only allowed to conduct a MLR analysis if we are certain that we have included all expressions that might serve to express the relevant discourse-pragmatic function and not just a subset. Because the variable context is defined on the basis of the heuristic function of translated texts, we consider the combination between parallel corpus data and multinomial logistic regression analysis to be a “match made in heaven.” Using MLRs on parallel corpus data should allow us to model the domain of expression of a certain semantic relation much more reliably than statistical models that have a dependent variable with a number of levels corresponding to just a limited subset of possible expressions of the semantic relation.

Data collection and annotation

Data annotation

We undertook a token-by-token annotation of the Spanish data for a dependent variable (the type of possessive construction employed by the translator) and a series of predictor variables elected on the basis of the results of the previous studies summarized in the second section.

A variety of possessive constructions were used to translate the Latin possessive phrases, exemplified in (3–8). The English translations are taken from the King James Bible.

1. (3) fermosas   son     tus   quixadas [POSS]

   beautiful   be.prs.3pl     poss.2pl   cheeks

   ‘Thy cheeks are comely’ (CA 1:10, E3, 15th c.)

2. (4) los dientes así como el rebaño de = las ovejas [ART]

   det.m.sg teeth so like det.m.sg flock of = det.f.pl sheep

   ‘Thy teeth are as a flock of sheep’ (CA 6:6, E5/7, 15th c.)

3. (5) met-ió  mano  a  su  çurrón     [BARE]

   put-pst.pfv.3sg  hand  in  poss.3sg  Bag

   ‘[David] put his hand in his bag’ (SAM1 17:49, GE, 13th c.)

4. (6) Descubr-ió  como  la  huerta  la  su [ART + POSS]

   take.away-pst.pfv.3sg  like   the  garden  det.f.sg  poss.3sg

   choça

   tabernacle

   ‘And he hath violently taken away his tabernacle, as if it were of a garden’ (LA 2:6, E5/7, 15th c.)

5. (7) e  respond-ió  el  padre  d  = ella   [DET + GEN]

   and  respond-pst.pfv3.sg  det.m.sg  father  of = her

   ‘And her father said’ (JU 15:2, Arr., 15th c.)

6. (8) non = le  dex-ó  su  padre  d  = ella  entrar  [SU + GEN]

   not = him  let-pst.pfv.3sg  poss.3sg  father  of = her  enter

   ‘But her father would not suffer him to go in’ (JU 15:1, E4, 15th c.)

Most frequent are POSS and ART + POSS (Table 1). Two variants occurred with a rather low frequency and were collapsed with formally or functionally affine variants to obtain larger numbers. First, we only encountered 24 cases of the SU + GEN construction (8), which were combined with DET + GEN (7) in one new type, GEN, sharing a formal feature that clearly sets them apart from the others constructional types—the addition of a prepositional phrase indicating the possessor after the possessed entity. Second, the 56 cases in which the possessive was translated with a bare NP, as in (5), were assumed to behave similarly to cases in which the possessive was translated with a NP including an article, such as (4), particularly because both constructional types appear to strongly depend on a relationship of inherent possession. Consequently, we unified the constructional types BARE and ART under the new category ART/BARE.Footnote ³ As a result, the dependent variable Type has the four levels POSS (the reference level, n = 2611), ART + POSS (n = 1747), ART/BARE (n = 307), and GEN (n = 135).

Table 1. Usage frequency of possessive constructions in the corpus

The predictor variables, involving the factors that have been identified in previous studies, are summarized in Table 2.

Table 2. Summary of predictor variables

We coded for a series of features of the possessor (PS). First, we coded for person morphology (first, second, or third person) and number morphology (singular, plural) of the possessor (variables PS_Person and PS_Number). Second, we coded for the animacy of the possessor referent (variable PS_Animate) (note that divine entities such as God, angels, or spirits were classified as animate). Third, we coded for the social rank of the possessor, distinguishing the levels as (a) other, (b) upper class, and (c) God (variable PS_Status). Fourth, we coded the degree of activation of the possessor referent (PS_activation), giving the value as “true” if the possessor had been mentioned in the same or the previous sentence and “false” if it had not (following the operationalization in Company Company, Reference Company Company2006:79).

We included three variables that characterize the possessed entity (PD). The variables PD_Animate and PD_Activation were coded exactly as their counterparts for the description of the possessor. In order to capture the relationship of inherent possession between possessor and possessed entity, we introduced the variable PD_Inherent. This variable received the value “true” if the relationship of possession between possessor and possessed can be characterized as such that it cannot be undone by external factors (Chappell & McGregor, Reference Chappell and McGregor1989:26–28). PD_Inherent received the value “true” in cases such as parts of an entity (i.e., su siniestra ‘his left hand’), kinship terms, exuviae (blood, sweat, tears), aspects of personality including emotions, forms of personal representation (i.e., su nombre ‘his name’), concepts involving images of the person (i.e., la sombra ‘your shadow’), and important cultural concepts and objects of value (i.e., su dios ‘his god’, nuestros muertos ‘our dead’).

Two predictor variables aimed at capturing the syntactic context of the possessive construction. The first variable, Syntactic function, describes the syntactic function of the possessive phrase, with the values “subject” (e.g., examples (7) or (8)), “object” (e.g., example (9)), “preposition” (when part of a prepositional phrase, as in example (10)), vocative (11), and “apposition” (12). Predicate objects were counted as objects. The second variable, Dative, received the value “true” if the sentence contained a dative expression, as in example (9).

1. (9) saca-ron  le  luego  los  ojos

   put.out-pst.pfv.3pl  pro.dat.sg  right away  det.m.pl  eyes

   ‘and they put out his eyes’ (JU 16:21, E6/8, 13th c.)

2. (10) con  beso  de  la  su  boca

   with  kiss  of  det.f.sg  poss.3sg  mouth

   ‘with the kisses of his mouth’ (CA 1:2, GE, 13th c.)

3. (11) levánt-a = te   tú   mi   amiga

   get.up-imp = pro.refl   you   poss.1sg   friend

   ‘Rise up, my love’ (CA 2:10, E5/7, 15th c.)

4. (12) el  rey  Nabucodonosor   tu  padre  lo  fiz-o

   det.m.sg  king  Nabucodonosor   poss.2sg  father  him  make-pst.pfv.3sg

   mayoral

   master

   ‘[whom] the king Nebuchadnezzar thy father … made master’ (DAN 5:11, E5/7, 15th c.)

The last two predictor variables describe the type of text passage, whether the possessive was used in direct speech (variable Direct speech), and whether the possessive was used in a narrative or lyrical book (variable Narrative).

Model selection process

After extracting and coding the data, we subjected the data to two MLR analyses using the function multinom() (Ripley & Venables, Reference Ripley and Venables2015) in R (R Development Core Team, 2015), one for the data from the 13th century and one for the data from the 15th century. The dependent variable was the variable Type. We set the reference level of Type to “POSS” (possessive adjective cases, as in su casa ‘his/her house’). We included each of the 12 predictor variables as predictors in the two models.Footnote ⁴

MLR analyses are not easy to interpret, as the coefficients only compare odds ratios. To ease interpretation, we used marginal effects, a statistical concept developed in econometrics (Cameron & Trivedi, Reference Cameron and Trivedi2010:491–492; Freese & Long, Reference Freese and Long2001:127–128). A marginal effect is the effect of a predictor variable on the dependent variable when the predictor variable is changed while the other predictor variables remain constant. To calculate the marginal effects, we transformed the model coefficients into the predicted probability of each of the constructional types in a certain usage context (represented by the predictor variables), fixing the covariates at a specific value. Because most of the covariates are dummy variables (and there is no mean between true and false), we did not follow the usual practice to fix the covariates at their mean. Rather, we selected the level for each variable that is most frequent in our data. The corresponding fixed values are PS_Person = 3rd, PS_Number = Singular, PS_Animate = Animate, PS_Status = Other, PS_Activation = Activated, PD_Animate = Inanimate, PD_Activation = Not activated, PD_Inherent = Noninherent, Syntactic function = Subject, Dative = No Dative, Direct speech = Direct, Narrative = Narrative. Crucially, fixing the covariates for illustration of results does not change the results of the regression analysis, which are available in the appendix. However, it is important to keep this procedure in mind when interpreting our results in the next section. For instance, the analysis shows that ART/BARE constructions are much more likely in third-person contexts, whereas ART + POSS constructions are less likely. Given that we fixed the variable for person morphology to third person, the figures describing the results for the other parameters will often show that ART/BARE has a higher probability of use than ART + POSS. This would be different if we had fixed the variable for person morphology for, say, first-person morphology. Consequently, it is important to restrict the interpretation of the figures to the relative changes in probability due to the influence of the variable that is being analyzed.