Nisi paria non pugnant: argument structure alternations with lexical reciprocal verbs in Latin

4.1 Data and annotation

For the purpose of this study, I have selected a sample of Latin verbs intended to cover the three general semantic classes of lexical reciprocals proposed by Knjazev (2007: 122): (i) verbs of spatial relationship of proximity/remoteness, e.g., congrego ‘assemble’, (ii) general relations of identity/difference, e.g., congruo ‘coincide’, (iii) and human relationship of rivalry/collaboration, e.g., pugno ‘fight’ (see also Fanelli 2009: 86–89). To select the verbs, I have looked for translational equivalents of the meaning ‘fight, argue’, ‘agree’, ‘be equal’, ‘discuss, converse’, ‘join, gather’, ‘separate’ in the Lewis and Short dictionary. I have only selected prefixed verbs in those cases in which the base verb has no symmetric meaning (this means that I have included iungo ‘join’ and contendo ‘contend’ but not coniungo ‘join’).^[5] Occurrences of these verbs have been retrieved from the LASLA corpus.^[6] Only predicative subject-oriented (see fn. 3) forms of the verbs have been included (this means that for verbs such as iungo ‘join’ only mediopassive intransitive forms with anticausative/passive function have been selected). Verbs with less than ten occurrences have been excluded (e.g., rixor ‘fight’ n = 4, socio ‘agree’ n = 5), whereas for verbs with more than 100 occurrences, only a random sample of 100 tokens has been selected. Following this procedure, I have collected 17 verbs for a total of 985 tokens.

The verbs that I have selected can be arranged into four classes based on two parameters. The first parameter, proposed already by Nedjalkov (2007: 13), concerns the distinction between spatial versus proper reciprocals. As the labels suggest, the former refer to symmetric spatial relationships, while the latter to more abstract situations (these roughly correspond to groups [ii] and [iii] in Knjazev’s classification). The second parameter is polarity. This parameter concerns the nature of the symmetric relationship between R1 and R2, and reflects the distinction in the proper reciprocal domain between verbs of rivalry versus collaboration, and, likewise, the spatial distinction between verbs of separation versus union. The list of the verbs selected for this study, together with their semantic classification and their token frequency, is given in Table 1.^[7]

All tokens have been manually annotated for a number of parameters. The list of parameters as well as their values is given in Table 2. In addition (see Section 6), only for discontinuous constructions have I annotated the following information concerning R1 and R2: morphological realization (noun, pronoun, zero), encoding of R2, animacy (animate/inanimate), and relative linear order (R1 > R2, R2 > R1).

4.2 Modeling the alternation: random forest model

In order to explore which of the parameters in Table 2 constitutes a better predictor for the choice of the argument structure construction, I have created a random forest model. Random forest models are particularly suitable when one deals, as in our case, with a small dataset with many variables, and allow one to rank variables, or predictors, based on their explanatory power (see discussion in Gries 2020, 2021]; Levshina 2015: chap. 14, 2020; Tagliamonte and Baayen 2012). The model has been fitted with the package randomForest(Liaw and Wiener 2002), setting Construction as our response variable (with three values simp = simple, abs = absolute, and disc = discontinuous) and the rest of the parameters in Table 2, with the exception of Lemma (see Section 5.5), as predictors. Overall, the model has an accuracy of 63 % (the best model is obtained by setting ntree = 3,000 and mtry = 5), which means that it predicts the correct construction with almost twice the accuracy of a completely random model (which would have an accuracy of 1/3 given that we are dealing with a trinomial response variable).

Figure 1 reports the variable importance score for each of the predictors taken into consideration. This score gives us an indication of “how much (if at all) […] a certain predictor contribute[s] to predicting the response” (Gries 2021: 465). As can be seen, three factors seem to be highly predictive: the impersonality of the verb, prefixation, and the occurrence of reciprocal marking. I discuss these and other factors in detail in the following sections. Other parameters seem to have a lesser explanatory power. Crucially, there does not seem to be a strong effect of literary genre in term of prose versus poetry, nor a particular effect of aspectual morphosyntax.^[8] Concerning the latter, this counters Mocciaro’s (2011: 334) observation that the absolute construction (at least in the variety of Italian spoken in Sicily) often triggers a habitual reading of the sentence, as in Latin one fails to observe a significant correlation between the absolute construction and the infectum (on the aspectual properties of Latin verb stems, see Haverling 2010).

Figure 1:

Random forest model: variable importance score.

5.1 Impersonal verb forms

Impersonal passives in Latin can be defined as those mediopassive forms of either transitive or intransitive verbs that lack a referential subject and take the default third person singular ending, of the type eo ‘go’ > itur ‘one goes’ (Pinkster 2015: 267–272 with references). As impersonal passives lack a syntactic subject, which is the position typically occupied by at least R1, a clarification is in order on the syntax of impersonal lexical reciprocals. A closer look at the occurrences reveals that lexical reciprocals in the impersonal passive forms showcase the same variety of constructions as their personal forms. The main difference lies in the coding of R1 and R2. In the simple impersonal construction, R1 and R2 are syntactically treated in the same way in the sense that both are omitted, so that only the verb remains, as in (7a). In the discontinuous impersonal construction, R2 is expressed via an oblique, as in (7b). Finally, the absolute impersonal construction features R1 expressed as the oblique Agent of the impersonal passive via an a(b) + ablative phrase, as in (7c) (Napoli 2010, 2013]).

As already remarked upon by Napoli (2013: 390) and Pinkster (2015: 267), the impersonal construction is particularly frequent with fighting verbs, especially among historians. My data fully confirms this claim: 110 out of 116 impersonal constructions occur with fighting verbs.

As Figure 2 shows, the impersonal construction is a good predictor for construction selection in the sense that it greatly favors the simple construction, while it disfavors both the absolute and the discontinuous ones.

Figure 2:

Partial dependence score for impersonal. Note: The plots for partial dependence score are to be read as follows: a positive score means that a given value of the predictor favors the choice of a specific construction, while a negative score means that it disfavors it. The length of the bar visually indicates the magnitude of the effect.

The correlation in Figure 2 can be explained by keeping in mind the nature of impersonal passives, which typically consist in “filling an argument position of a predicate […] without establishing a referential link to any entity from the universe of discourse” (Van der Auwera et al. 2012: 123). In fact, the Latin impersonal passive is typically used in contexts in which either the agent is generic/indefinite, or, if it is definite and overtly expressed, as in (7c), it is discursively backgrounded. As Pinkster puts it, “the impersonal passive is a perfect means to concentrate the action” (Pinkster 1992: 174), that is, to foreground the event denoted by the verb without the need to profile its participants (see Sansò 2006). It is therefore fitting that the simple construction, in which both R1 and R2 are likewise defocused, is mostly used in impersonal contexts. This is also in line with the finding that, whereas R1 and R2 can be occasionally expressed via oblique phrases as in (7b) and (7c), in my corpus there are no occurrences in which both are expressed simultaneously.

5.2 Reciprocal marking

The second-best predictor is the occurrence of additional analytic reciprocal marking in the sentence. Specifically, as the data in Figure 3 shows, reciprocal marking strongly predicts the simple construction and, conversely, it substantially disfavors the absolute construction.

Figure 3:

Partial dependence score for reciprocal marking.

As an example, consider the sentence in (8), in which the verb contendebant is used in a simple construction and co-occurs with the reciprocal phrase inter se.

The observed scores fit the universal proposed by Haspelmath (2007: 2093) whereby languages only allow discontinuous constructions when reciprocal marking occurs on the verb, whereas these are incompatible with analytic marking of the type instantiated by Latin inter se.^[9] Notably, one also detects a small positive correlation between the discontinuous construction and reciprocal marking. A closer look at that data reveals that one only finds three such occurrences, which do not offer serious counterevidence. In fact, these occurrences feature other markers of reciprocity, which are arguably less conventionalized than inter se, such as the adverb invicem ‘mutually’ in (9).

Overall, inter se, which necessarily expresses a symmetric relation among multiple participants, appears to be both incompatible with the asymmetric conceptualization imposed by the discontinuous construction, as well as with the lack of R2 in the absolute construction. The question remains why inter se is used with simple constructions at all. In fact, as the verbs already lexicalize a symmetric relation, the addition of inter se may be redundant. A possible answer is that inter se is used to disambiguate the reciprocal reading in those contexts in which the latter might be less clear (thus Revuelta 2017). In support of this idea, note that with the verbs congruo, consentio, and iungo, inter se is only used with inanimate reciprocants, which are possibly less likely to be understood as also playing the role of Agent and thereby require the use of inter se to strengthen the reciprocal interpretation of the verb, as in (10).

5.3 Verbal morphology: prefixation

The third predictor in order of importance concerns the occurrence of verbal prefixation. As discussed in Section 4.1, the sample includes a number of prefixed verbs: seven con-verbs (confligo, congredior, congrego, congruo, colloquor, contendo, consentio) and two dis-verbs (dimico, discepto). Before we move on to evaluate the role of prefixation, a note is in order on the nature of these verbs. For the purpose of this paper, I have considered verbs such as colloquor as lexical reciprocals, but this choice is admittedly problematic, because one could also argue that verbs such as colloquor ‘converse’ and congredior ‘meet’ are better understood as grammatical reciprocals formed via com-prefixation from non-reciprocal verb bases, in this case loquor ‘talk’ and gradior ‘go’, respectively. Nevertheless, these are the only verbs that in Latin lexicalize symmetric situations such as ‘discuss’ and ‘meet’. In addition, in other cases com-verbs cannot simply be regarded as grammatical reciprocal counterparts of a non-reciprocal base, because either the base is synchronically missing (e.g., congruo, which synchronically lacks a base verb, and confligo from fligo, which is barely attested in pre-classical authors), or the verb has developed a lexicalized meaning unpredictable from the base verb, e.g., tendo ‘stretch, bend’ > con-tendo ‘fight, dispute’. More generally, as preverbs tend to have a wide range of meanings (on the polysemy of com-specifically, see Brucale and Mocciaro forthcoming; Moussy 2005; Rosén 1992), it is disputable whether preverbation in ancient Indo-European languages serves as a dedicated means of creating grammatical reciprocals or rather as a way to derive new lexical ones (see discussion in Inglese and Zanchi 2020: 141–151). For these reasons, while acknowledging their problematic nature, I treat all com- (and dis-) verbs on par with the other lexical reciprocal verbs.

The partial dependence scores for prefixation are reported in Figure 4. As Figure 4 shows, the general trend is for prefixed verbs to positively predict the discontinuous construction, while disfavoring the absolute and partly also the simple one.

Figure 4:

Partial dependence score for Prefix.

The negative association between prefixation and the absolute construction can be explained in a similar way as the incompatibility of the absolute construction with analytic reciprocal marking: as preverbs, in particular com-, highlight co-participation, they are less compatible with the construal imposed by the absolute construction. An example of a com-verb used in an absolute construction is congrego ‘assemble’ in (11), in which R2 is omitted because it is generic, similarly to (7c) above.

The association between prefixed verbs and the discontinuous construction is less straightforwardly explained. Things become clearer if one looks at the distribution of individual prefixes. As the data in Figure 4 shows, it is com-verbs that are particularly more drawn to the discontinuous constructions, even when compared to the simple construction (with the latter, in fact, a small positive effect of dis-can be detected). Specifically, leaving dis-verbs aside, if one zooms in on the distribution of com-verbs in the discontinuous constructions with the coding of R2, as shown in Table 3, a striking pattern emerges, whereby com-verbs significantly correlate with the coding of R2 via the comitative phrases with cum (Fisher’s Exact Test for Count Data, p-value < 0.01). In other words, there appears to be a remarkable solidarity pattern whereby com-verbs tend to be used in a discontinuous construction with R2 expressed by the cognate preposition cum (I return more extensively on the realization of R2 in Section 6.1).^[10] This pattern is illustrated in (12), where both confligo and colloquor co-occur with cum prepositional phrases.

Notably, the same pattern is also found with other pairs of cognate preverb/preposition, as in the case of inter-/inter in (13), so that it remains unclear to what extent the pattern involving lexical reciprocal com-verbs and cum-adverbials instantiates a more general behavior of preverbs in combination with cognate prepositions.

Finally, concerning the interaction between prefixed reciprocal verbs and analytic reciprocal marking, Revuelta (2017) observes that inter se frequently occurs with dis-prefixed reciprocals as a disambiguator (especially in those contexts in which, given the polysemy of the prefix, other readings could be available). Unfortunately, data from my corpus does not lend empirical support to this claim, as inter se equally occurs with prefixed and unprefixed verbs, as shown in Table 4 (Fisher’s Exact Test for Count Data, p-value = 0.6519).

5.4 Adverbial phrases: Purpose and Topic

The fourth parameter for importance is the co-occurrence with adverbial phrases that indicate the semantic roles of Purpose and Topic. The former is broadly understood as the goal that participants want to achieve when undertaking a given action, whereas the latter typically consists in the object of discussion or contention with verbs of fighting, discussing, and agreeing. From a purely semantic perspective, Purpose and Topic are not always easily distinguishable (and they are also closely connected with the role of Cause). In order to operationalize this distinction in my dataset, I have systematically annotated a given adverbial phrase as expressing either Purpose or Topic based on its coding. As extensively discussed by Luraghi (2010), Purpose and Topic can variously be encoded in Latin. Based on the data from the corpus, realizations of Purpose include prepositional phrases introduced by pro, ob, ad (typically with the gerundive) or adverbial clauses introduced by ut plus subjunctive. An example is pro patria in (14):

Topic expressions include the bare ablative, prepositional phrases headed by de, in (with either ablative or accusative), as well as a variety of interrogative or complement clauses. An example is the phrase de potentatu ‘over power’ in (15).

Partial dependence scores, reported in Figure 5, support the need to keep Purpose and Topic adverbials distinct. To put it simply, Purpose adverbials positively predict the absolute construction and less so the discontinuous and the simple ones, which instead display a small positive dependence on Topic adverbials. This asymmetry can possibly be explained as follows. Topic and Cause conceptualize a shared interest between the two reciprocants and are therefore preferred when both reciprocants are expressed (in the simple and discontinuous construction). By contrast, Purpose highlights the higher involvement of either party (in the sense that, even though we might be fighting over the same object, our individual goals may differ): this implies Purpose is more compatible with constructions in which only one participant is expressed (absolute construction).

Figure 5:

Partial dependence score for Purpose.

5.5 Verb semantics, symmetry, and lexical restrictions

As anticipated in Section 3, lexical reciprocal verbs can be classified along two main semantic dimensions, that is, the proper versus spatial and the polarity (negative vs. positive) distinctions. When one looks at the distribution of the three constructions under investigation with respect to these semantic dimensions, an interesting picture emerges.

Overall, as shown in Figures 6 and 7, these semantic distinctions appear be useful for predicting the absolute and the simple constructions, while the discontinuous construction seems to be unaffected (see the high positive scores for both values of the two predictors). There are two interesting effects to single out. First, as shown in Figure 6, negative polarity verbs strongly favor the absolute construction, while positive polarity is only marginally connected with the simple one. By contrast, as per Figure 7, spatial verbs robustly reject the absolute construction and instead show positive scores with the simple one.

Figure 6:

Partial dependence score for Polarity.

Figure 7:

Partial dependence score for Spatial.

The observed dependence scores can be explained as follows. Let us begin with the association between negative polarity and the absolute construction. If we focus on proper verbs, it turns out that the absolute construction occurs 209 times with negative polarity verbs, as in example (4) above, and only 34 with positive ones. To explain this distribution, a more careful consideration of the semantics of these verbs is in order.

As mentioned in Section 2, lexical reciprocals are generally connected with the notion of symmetry. A finer-grained distinction can be drawn between those verbs that necessarily express symmetric and simultaneous (or unary) events, that is, meet-verbs, and those which do so to a lesser extent, that is, fight-verbs (Kemmer 1993: 102; for a detailed discussion, see Dimitriadis 2008a, 2008b; Gleitman et al. 1996; Siloni 2012; Winter 2018). Meet-verbs have also been defined as irreducibly symmetric predicates, because “[they] can only describe individual events that are themselves symmetric for the two participants involved […]. There can be no event of John meeting Bill without that same event also being an event of Bill meeting John” (Dimitriadis 2008a: 376, 378). Conversely, even though “you can’t fight with no one to fight with or against” (Plank 2006: 248), fight-verbs have been described as partially symmetric events, because “at least a subset of them licences an interpretation in which only one of the arguments is a volitional participant” (Rákosi 2008: 423). In other words, fighting-verbs are more easily construed as asymmetrical than meet-verbs. This distinction is borne out by the Latin data: the higher frequency of the absolute construction with fighting verbs (which constitute the core of proper negative polarity verbs) is due to their nature as partially symmetric events.^[11]

It should be noted that, in the classification proposed by Dimitriadis (2008a), Rákosi (2008), and Siloni (2012), irreducibly symmetric predicates should also be incompatible with absolute constructions, because they should not license the omission of generic R2 (though they may admit zero anaphora). I believe that this incompatibility should better be understood in terms of a strong tendency, and not as a strict rule. As a matter of fact, empirical data from the corpus shows that generic R2 may occasionally be omitted even with verbs that are necessarily symmetric. An example is the absolute use of congrego ‘assemble’ in (11), and of congredior ‘meet’ and iungo ‘join’ in (16a) and (16b):

Even though they are admittedly marginal, cases such as those in (16) demonstrate that, in specific contexts, individual verbs can be coerced into constructions that may impose their own semantic profile and override the verbs’ lexical semantics (this is an instance of the more general mechanism of coercion or accommodation, see Goldberg 1995: 159). Note however that, in (16a) and (16b), the verbs congredior and iungo have undergone a semantic shift: the former means ‘meet an enemy in battle’ and the latter ‘have intercourse with’ (the tendency of lexical reciprocals to develop idiomatic meanings is also discussed by Siloni 2012).

Moving to spatial verbs, these negatively predict the absolute construction and prefer the simple and the discontinuous ones (even though with the latter the effect is more difficult to interpret). This distribution may find an explanation in the more general cognitive organization of spatial relations. As extensively discussed in the field of Cognitive Grammar (Langacker 1987: chap. 6; Talmy 1978), human beings tend to conceptualize spatial relations asymmetrically, by locating an entity (the figure or trajectory) with respect to another one (the ground or landmark). This means that spatial relations tend to be established within at least two entities, thus explaining the negative score with the absolute construction, and one is often more prominent than the other. This tendency also affects spatial lexical reciprocals, such as congredior and separo in (17a) and (17b). Even though these verbs express relations that are in fact symmetric, speakers prefer to profile them from the vantage point of one of the participants involved because this is the way in which they typically construe spatial relations (the same applies also to some abstract relations such as those of similarity, see Gleitman et al. 1996).

So far, I have discussed the role played by verbal semantics as a determining factor in predicting constructional alternation. Nevertheless, as the random forest model in Figure 1 shows, these parameters do not overall constitute good predictors when the entirety of the data and the variables are taken into account. In the random forest model, I have excluded the verbal lemma as predictor, because the model is not particularly suitable to handle predictors with more than a few levels (Levshina 2020: 622). Nevertheless, the role of lexical preferences cannot be ignored. As shown in Figure 8, with the exception of proelior, all the verbs in my dataset can occur in all three constructions, but do so to varying extent, with some verbs preferably selecting one construction over the other. Surprisingly, even verbs with a comparable semantics (e.g., bello, certo, luctor, proelior) do not show a consistent distribution across the three constructions. This behavior is not isolated, as it mirrors findings by Inglese and Zanchi (2020: 141) on the distribution of the three constructions with fighting verbs in Homeric Greek. We can better understand the distribution in Figure 8 as an instance of the more general and pervasive phenomenon of lexical restriction. As Van Lier and Messerschmidt (2022) point out “many verbs show statistical preferences for certain constructions […] lexically-specified preferences can override pragmatic factors in determining a speaker’s choice for one or the other construction” (Van Lier and Messerschmidt 2022: 4, 6). From a usage-based perspective (e.g., Bybee 2007), the crucial point is that such preferences are partly driven by the higher frequency with which speakers employ a given verb in a given construction, which in turn may lead to a greater entrenchment of that construction with that specific verb and its subsequent generalization even to contexts in which it is less immediately motivated by semantic/pragmatic factors.

Figure 8:

Constructional alternation and lemma.

6.1 The construal of R2

As Haspelmath (2007) observes “in discontinuous reciprocals, one of the arguments is always an oblique (most commonly a comitative) argument” (Haspelmath 2007: 2092). The encoding of R2 as a comitative argument reflects a somewhat neutral construal of its role as co-participating in the reciprocal event. While Haspelmath’s claim is undeniably true, I believe that a closer look at how R2 is actually realized in empirical data can shed interesting light on the function of the discontinuous construction. Table 5 reports data on the coding of R2 in the corpus (leaving aside those cases in which R2 is omitted, as discussed for example [6] above). As can be seen, in the vast majority of cases (239/384), R2 is encoded via a prepositional phrase with cum ‘with’ plus ablative.^[12] This is far from surprising, as cum is the default way to express the Comitative (and related) roles in Latin (Brucale and Mocciaro forthcoming; Luraghi 2010: 79–84). The reason why prototypical Comitative arguments (as in walk with the girl) are extended to encode R2 with reciprocal verbs (as in fight with the enemy) is to be sought in the general diachronic processes that link sociative, comitative, and reciprocal markers (Heine and Miyashita 2008: 182; Nedjalkov 2007: 32–41).

Nevertheless, as Table 5 shows, there are several other, albeit marginal, ways in which R2 can be expressed. The reasons why R2 may be expressed by an oblique non-Comitative phrase should be discussed within the much wider domain of the use of preposition and cases to express semantic roles in Latin (see Luraghi 2010; Pinkster 2015: esp. chap. 12). A finer-grained treatment of this issue is clearly beyond the scope of this paper, so that I limit myself here to some preliminary observations. First, a non-Comitative realization of R2 may be the default one with some verbs for semantic reasons. For example, R2 with separo is always expressed by a(b) + abl, because R2 with verbs of separation is expressed as a Source, whereas dative R2s are typical with the verbs iungo and congruo, as with these verbs R2 is construed as a Direction (which also accounts for the isolated use of ad + acc with congruo in Table 5). Stylistic factors also play a role: for example, one finds fighting verbs with R2 in the dative only in poetic texts, owing to a more general tendency to avoid prepositions in poetry (Pinkster 2015: 1182–1183).

More interesting are those cases in which the same verb shows alternative realizations of R2: for example, as shown in Table 5, pugno may occur in at least four distinct constructions (a behavior generally shared by other fighting verbs). Specifically, besides cases of Comitative R2, one also finds contexts in which R2 is expressed with either contra/adversus + acc or in + acc. Both realizations reflect a diminished participation of R2, either by simply construing it as a Direction (in + acc), or by also highlighting its affectedness, by construing it as Adversative Patient (contra/adversus + acc). Such cases show how the non-Comitative construal of R2 may stretch the semantics of the discontinuous constructions so as to include contexts in which R2 is portrayed as non-agentive, thus essentially construing an inherently symmetric event similarly to an asymmetric transitive situation. This extension is maximally visible in those few cases in which R2 is encoded by an accusative NP, which is the typical case of the Patient of transitive verbs (notably, these cases only appear to involve pronouns). From the perspective of constructional alternation, one may speculate that the discontinuous construction, by virtue of expressing R1 and R2 independently, is particularly suitable in those cases in which speakers not only need to assign major prominence to R1 (see Section 6.2) but also wish to construe R2 not merely as a co-participant in a symmetric event.

6.2 The discourse status of R1 and R2

Besides allowing for different conceptualizations of R2, the discontinuous construction has also been claimed to be favored in contexts in which R1 has a more prominent discourse status, typically as discourse topic (thus Allan 2003; Mocciaro 2011). Unfortunately, assessing the topicality status of referents is not always an easy task, especially in a corpus language for which no speaker intuition is available. I have therefore decided to rely on three parameters that are more objectively observable in out textual sources and that, combined, may be used as reliable proxies for topic status. The first parameter is animacy, as human participants tend to be more topical than inanimate ones. The second parameter concerns the realization of R1 and R2, that is, whether they are expressed by a full noun phrase, a pronoun or are omitted (Latin being a pro-drop language). This parameter is relevant inasmuch as topical referents are already accessible in discourse and therefore typically require lighter (or zero) anaphoric material to be reprised (Givón 2001: 419–420). Finally, another parameter is word order: specifically, I consider the linearization of R1 and R2, based on the assumption that topical elements come first, especially in a language such as Latin in which word order is largely dependent on pragmatic factors (see, e.g., Pinkster 2021: chap. 24; Spevak 2010). Quantitative data on these three parameters is given in Tables 6–8 (note that, in order to minimize stylistic factors, data for word order comes from prose text exclusively).

Data from the corpus fully confirms the hypothesis that discontinuous constructions seem to be favored in contexts in which there is an asymmetry in topicality between R1 and R2, whereby the former displays a higher topicality status. This is confirmed by the fact that R1 is more frequently expressed by pronouns or omitted, is more frequently animate (the difference is significant, Fisher’s Exact Test for Count Data, p-value < 0.01), and, when both are expressed, R1 typically precedes R2. In particular, the discontinuous construction is needed in those cases in which R1 is encoded by a pronoun and R2 by a full noun phrase, as such asymmetries appear to be incompatible with the simple construction, in which the two coordinated R1 and R2 must share the same part of speech. An example of a discontinuous construction with pronominal R1 and nominal R2 is (18):