Role-reference associations and the explanation of argument coding splits

3.1 Two types of usual associations

To make Universal 2 (in (6) above) more specific, the most straightforward possibility is to consider the arguments individually, regardless of what happens elsewhere in the clause, and to state their usual associations (or association tendencies) with the referential prominence scales:

For example, clauses such as ‘The dog found a bone’ and ‘She gave the boy a key’, with definite A and R, and indefinite P and T, are more usual than ‘A rock hit the hiker’ or ‘She gave a boy the key’.

On the other hand, we can look at both arguments within a two-argument scenario simultaneously and compare their referential prominence values. In monotransitive constructions, the scenario consists of the A and the P, and in ditransitive constructions, the scenario consists of the R and the T.^[4]

Because of these tendencies, we can distinguish three kinds of scenarios:

The downstream scenarios (e.g., ‘I caught a rabbit’) are the most usual ones, and the upstream scenarios are the least usual (e.g., ‘A dog bit you’), with balanced scenarios in between. The “upstream/downstream” metaphor is introduced here because the most common scenarios are expressed in the “easiest” way (like swimming downstream), while the least common scenarios can be thought of as more “difficult” (like swimming upstream) and thus need more “coding energy”.

3.2 Two types of splits: Single-argument splits and scenario splits

Corresponding to single-argument associations and scenario associations, we have two types of argument coding splits:^[5] single-argument splits (Sections 4 and 5) and scenario splits (Sections 6 and 7).

A single-argument split is defined in (12), and the corresponding universal is formulated in (13) (which is restricted to flagging and does not make claims about indexing, cf. Note 6).

For example, prominent P-arguments have special longer flags in languages with differential object marking (e.g., Spanish, where specific human Ps get the accusative preposition a), and non-prominent R-arguments have special flags in languages with split R flagging (e.g., in English, where full nominal Rs get the dative preposition to, cf. Section 5.1.2).

This universal makes use of the notion of an asymmetric split, i.e., a coding split in which one of the coding types involves longer coding (also called special coding). In most cases, this means overt coding contrasting with zero coding, but we will also see examples of shorter and longer overt coding.^[7]

The universal in (13) is a more general formulation that subsumes the famous split P flagging universal in (14).^[8] An illustration was given in (2) above, and more will be said in Section 4.1.

The more general formulation in (13) also includes split A flagging (“differential subject marking”), as illustrated above in (1) and treated further in Section 4.2, as well as split R and split T flagging (treated in Section 5).

The other type of coding split is called scenario split because it is based on the scenario associations. A scenario split is defined in a general way in (15), and the most general scenario universal is given in (16).

For example, in (3) from Teop, the downstream scenario (locuphoric > aliophoric, in (3b)) does not have the special coding with the Object marker ben-, and in (4) from English, the upstream scenario (N > pers, in 4c) requires the Dative preposition to (Recall from 11b that an upstream scenario is one where R is less prominent than T, and from (8a) that full nominals are less prominent than person forms). The balanced scenarios (N > N, pers > pers) do not require this special coding, and neither do the downstream scenarios (an example of a downstream scenario in English is She gave him_Rthe money_T, pers > N).^[10]

The next four sections (Sections 4–7) are devoted to further illustration (including some discussion) of the two types of argument coding splits, both in monotransitive constructions and in ditransitive constructions (see Table 1 above). Within each type of coding split and for both construction types, I will illustrate how various prominence scales condition coding splits. Not all prominence dimensions are found to be relevant (equally frequently) with all the coding split types and both construction types, for unclear reasons, but most of the expected phenomena are attested in the literature. Crucially, constructions that are not expected and ruled out by the role-reference association universal are not attested (or found only marginally). In Section 11, I will discuss the explanation of these strikingly general patterns, and I will propose a correspondingly general explanation.

In addition to single-argument splits and scenario splits, I will also discuss obligatory and optional verbal voice constructions (Sections 9 and 10). I will end up formulating 12 additional universals (summarized in Figure 1 in Section 11 below), all of which are ultimately special cases of Universal 1.

Figure 1:

Taxonomy of universal claims in this paper.

4.1 Split P flagging (or differential object marking)

As mentioned earlier, split P flagging, traditionally known as “differential object marking”, has been widely discussed in the earlier literature, so there is no need to dwell on it much in this paper (see Seržant and Witzlack-Makarevich 2018 for a recent overview). It is particularly often conditioned by animacy and definiteness, but sometimes by givenness and nominality, and occasionally by person (focus-conditioned split P flagging seems to be very rare, and I currently lack examples).

The universal nature of the phenomenon, formulated as Universal 4 in (14) above, has generally been taken for granted. Aissen (1999: 673) counts it “among the most robust generalizations in syntactic markedness”, and much of the literature focuses on the precise factors in particular languages as well as on possible explanations, not on providing evidence for or against the universal. Bossong (1985: 3–8) makes it very clear that he regards the phenomenon as an implicational universal (first described systematically by Thompson 1912), and even Filimonova (2005), who documents a few counterexamples, does not challenge its status as a universal tendency. Bickel et al. (2015) seem to suggest that the dependence of split P and split A flagging on referential prominence is not a universal tendency, but Schmidtke-Bode and Levshina (2018) show that Bickel et al.’s data can be interpreted rather differently, as providing support for Universal 4.^[11]

The following subsections illustrate five subtypes of split P flagging, with conditioning by different prominence scales.

4.2 Split A flagging

Since Silverstein (1976), split A flagging (or split ergativity, i.e., ergative flagging restricted to lower-prominence A) has often been seen as the mirror image of split P flagging. Kiparsky (2008: Section 2.3) argues that the pattern represents a “true universal”, i.e., not merely a typological generalization. By now it is well-known that split A flagging (also called “differential subject marking”,^[13] cf. de Hoop and de Swart 2009) is rarely conditioned by animacy or definiteness, and much more often by person or nominality. Because in quantitative terms, the observed coding phenomena are not an exact mirror image of what we find in P flagging, some authors have questioned the proposal that they should be understood in an analogous way (Fauconnier and Verstraete 2014). However, accusative and ergative systems are not fully symmetrical in other ways either, but even so it has often proved useful to consider them as mirror images, and here I hypothesize that a very similar generalization applies to ergative systems. Universal 6 in (21) is completely parallel to Universal 4 in (14) above.

In other words, if a language has ergative marking but not on all kinds of arguments, then it tends to have it for the less prominent arguments, in particular for full-nominal arguments, for aliophoric (third-person) arguments, and for focused arguments (see Dixon 1994: Section 4.2).^[14]

5.1 Split R flagging

Split R flagging is much more common than split T flagging, just as generally R flagging is more common than T flagging in ditransitives, so we begin with this.

5.2 Split T flagging

Split T (theme) flagging is quite rare, just as T flagging in general is uncommon in ditransitive constructions (English has a T preposition only with a few verbs, e.g., provide [someone]_r [with something]_t). It seems that it is mostly found in languages of West Africa.

6.1 Person-conditioned special P flagging

In Kolyma Yukaghir, a special accusative flag is required on the P when the A is aliophoric (cf. Keine 2010: Section 6.3), i.e., when the construction is not person-downstream.

A completely parallel situation is found in Teop, an Oceanic language (illustrated above in (3)). Yurok (an Algic language of California) is similar as well, requiring an accusative suffix if the A is aliophoric and the P is locuphoric (Robins 1958: 21; see Keine 2010: Section 6.2; Georgi 2012).^[18]

6.2 Person-conditioned special A flagging

In Sahaptin, a special ergative flag is required on the A when the P is locuphoric (cf. Keine 2010: Section 6.1), i.e., when the construction is person-upstream.

6.3 Definiteness-conditioned special A flagging

Eastern Khanty has an ergative flag on the A-argument when the P-argument is specific (see Baker 2015: 128), i.e., when the construction is not definiteness-downstream.

Baker (2015: 128) also cites Ika (a Chibchan language of Colombia), and Kalin (2017: Section 3.4) cites Niuean, both of which have been described as having very similar systems.

6.4 Animacy-conditioned special P flagging

A similar phenomenon from a less exotic language is Spanish differential object marking that depends on the (in-)animacy of the A: As discussed by García García (2007: 64), Spanish uses the preposition a on the P also in cases such as (38), where the A is inanimate (i.e., when the construction is not animacy-downstream).

7.1 Special R coding conditioned by person of T

This kind of coding split covers most of what has become widely known in the literature as “person-case constraint” (PCC). For example, in Bulgarian, when T is third person (aliophoric), the R can be a clitic pronoun, but the special dative preposition na is required on the R when the T is a locuphoric clitic pronoun (see 39b–c).

As was pointed out in Haspelmath (2004), the term “person-case constraint” is a misnomer, because very similar phenomena are found in languages that have no case at all, like Shambala.

For constructions like these, I had earlier discussed the universal in (41), based on the “person-case constraint” of Bonet (1994) and subsequent work.

In the present context, this can be reformulated as follows:^[19]

A scenario with locuphoric R and aliophoric T is a downstream scenario, so we expect it to be expressed by short forms. Universals 9a and 9b are thus merely special cases of Universal 5 in (16) (the fact that the person-role universal was not fully general was already noted in Haspelmath 2004: Sections 6.2–6.4, and the present paper expands on those observations).

7.2 Special R coding conditioned by nominality of T

In many varieties of English (especially American, it seems), the R cannot be coded in the simplest way when the T is a person form rather than a full nominal. In these varieties, (44a) is unacceptable (*Pat showed him it), not only (44b), which seems to be unacceptable in all varieties of English (see (4c) above).

In these varieties of English, special coding of R (with the preposition to) is thus conditioned by a high position of T on the nominality scale: If T is a personal pronoun, R must be coded in a special way.

7.3 Special T coding conditioned by nominality of R

While the use of a special R marker (like na+ in Bulgarian and kwa+ in Shambala) to code the person-upstream scenarios 3 > 1 and 3 > 2 is perhaps the most widespread pattern, some languages use special focal forms for T arguments when the R is a person form.

For example, Modern Greek has a set of Genitive (i.e., dative) and Accusative proclitics used in downstream and aliophoric balanced scenarios, as seen in (45a). But these proclitics cannot be used in upstream scenarios, as seen in (45b).

The form eména in (45c) does not contain a special flag, but is simply the independent personal pronoun. However, it is longer and thus conforms to the general prediction of longer coding for less usual situations.

7.4 Special R coding conditioned by animacy of T

In Icelandic, the preposition fyrir is required on the R when the T is animate, according to Siewierska and van Lier (2013: 194).

Both constructions have accusative flagging for the T and dative flagging for the R, but the dative flag is the (shorter) Dative case form in the downstream scenario, while the dative flag is a (longer) preposition in the upstream scenario.

10.1 Classical passive, antipassive and dative alternations

Another widespread phenomenon in languages, closely related to argument coding splits, is argument coding alternations. An alternation is a situation where two different coding patterns can be used alongside each other, with roughly the same meaning. Well-known examples of coding alternations are passives and antipassives for monotransitives, and dative alternations for ditransitives.

There is a great variety of such argument coding alternations, but when we single out those that show asymmetric coding, we can generalize over them in a way that has apparently not been done before. By asymmetric coding, I mean a situation where either (i) one of the alternates has special verb coding, as is normally the case in passives (Haspelmath 1990), and also in the antipassive construction in (59), or (ii) the argument flagging in one of the alternates is clearly shorter. The latter is the case in the English Dative alternation, where the Double Object construction (in 60a) shows no preposition, while the Prepositional Dative construction has an extra dative preposition. Many alternations are asymmetric in both ways at the same time: Thus, the English passive alternation has special verb coding (the passive auxiliary be plus the Past Participle form of the verb), and in addition the argument flagging is longer (the preposition by on the agent argument).

Given this notion of an asymmetrically coded alternation, we can formulate the generalization in (61).

Here the most relevant subtype of referential prominence is topicality or givenness. For passives, which are by definition asymmetric, this means that they tend to be used when the A is not given/topical, and/or when the P is not new information. We can formulate this as a universal:

That this is indeed the case has been known for quite some time (e.g., Shibatani 1985; Siewierska 1984), although I am not aware of earlier formulations that are as general as Universal 13.

For dative alternations, which are generally (and almost by definition) asymmetrical, we can likewise say that the longer alternant occurs when unexpectedly the R is not given/topical, and/or when the T is not new information.

For English, this is well established (e.g., Collins 1995; Thompson 1990), and the situation in related languages is not very different (see van der Beek 2004 for Dutch, for example). However, dative alternations are not very common in the world’s languages (Siewierska 1998), and I am not aware of in-depth studies of dative alternations in non-European languages. Thus, the available evidence for Universal 14 is currently slim, but there is no counterevidence either, and both 13 and 14 are special cases of Universal 12, so I would like to claim that it is indeed a universal generalization.

Universal 12, in turn, is evidently a special case of Universal 1, the general role-reference association universal.

10.2 Splitting alternations

In addition to coding splits and coding alternations, we also find an intermediate phenomenon that provides further confirmation for the present approach: Some construction pairs alternate under some conditions, but are in complementary distribution in other conditions. I call these situations splitting alternations.

For example, in Lummi (a Salishan language; Jelinek and Demers 1983), the ordinary Active construction is used only when the scenario is nominality-downstream (as in 64a) or nominality-balanced (as in 64b). When the scenario is upstream, the Passive construction (with the verb suffix -ŋ, and the Oblique preposition on the A) is obligatory (see 64c).

For the scenarios that we have seen so far, the Active-Passive pair can be seen as a split pattern that shows both an argument coding split (because the A is sometimes Oblique-marked) and verb coding (because the non-usual situation requires the verb marker -ŋ). However, when both the A and the P are full nominals, either the Active or the Passive can be used:

A very similar situation is described for Northern Tiwa of Picurís by Nichols (2001).

In the domain of ditransitive constructions, we find a parallel case in Koyra Chiini (a Songhay language of Mali; Heath 1999). On the one hand, the Postpositional Dative construction (with the postposition + se) is the only possibility in not maximally usual situations, i.e., in nominality-upstream and nominality-balanced scenarios, as in (66a)–(66c). The shorter Double Object construction (66d) is possible only in nominality-downstream patterns.

However, in the nominality-downstream scenario, the Double Object construction is not obligatory, but either construction is possible:

Heath (1999) has not investigated the differences between these two constructions, and Jelinek and Demers (1983) do not tell us about the usage differences between the Lummi Active and Passive either, but I suspect that further investigation will show that they differ in usage not unlike the corresponding English constructions, as suggested by the universals in Section 10.1.

Since the English Dative Alternation is also an optional alternation in some cases (cf. (60)) but a grammatically required split in others (cf. (4)), it is also an example of a splitting alternation.

11.1 Summary of the generalizations

I have reviewed substantial evidence for the role-reference association universal that was introduced in the first section of this paper and is repeated here:

The subsequent Universals 3–14 are all special instances of this super-universal. Figure 1 diagrams their taxonomic relationships.

The next question is what explains this universal tendency. In this section, I mention three possible explanations: frequency-based coding efficiency (Section 11.2), disambiguation (Section 11.3), and innate biocognitive-representational constraints of a grammar blueprint (Section 11.4). I will argue for frequency-based coding efficiency and against the other two explanations.

11.2 Frequency-based coding efficiency

The first explanation, which I argue for here, invokes a functional-adaptive constraint: Languages tend to have efficient coding systems, with zero or short coding for more frequently occurring meanings and functions, and overt and long coding for more rarely occurring functions. Through piecemeal adaptation in language use, languages come to have (or restore) efficient patterns. This explanation accounts for a very large number of other coding asymmetries, as summarized in the universal in (68).

This accounts for coding asymmetries in pairs like singular/plural, present/future, affirmative/negative, allative/ablative, and many others. They were often called “markedness asymmetries” in the past (Croft 2003: Chapter 4; Greenberg 1966), but it is now clear that the formal patterns are due to frequency-based coding efficiency (Haspelmath 2008; 2021). The explanation is the same as for length differences in lexical forms (Zipf 1935): frequently expressed meanings are more predictable and can therefore be expressed by shorter forms.

For role-reference associations, this kind of explanation has long been advocated, for example by Caldwell (1856: 276), whose early remarks on special P flagging deserve to be quoted again here (cf. Filimonova 2005: 78):

[…] the principle that it is more natural for rational beings to act than to be acted upon; and hence when they do happen to be acted upon – when the nouns by which they are denoted are to be taken objectively – it becomes necessary, in order to avoid misapprehension, to suffix to them the objective case-sign.

It is of course unclear what is meant by “natural” and a vague appeal to “nature” is not explanatory by itself. However, if we replace “natural” by “frequent”, then the quotation is identical to what I claim here: Frequency asymmetries lead to differences in predictability, and grammatical coding is more efficient if more predictable meanings get less coding, and vice versa. A very similar formulation is found in Comrie (1989: 128), still using “natural” but evidently meaning “frequent”:

[…] it has been noted that in actual discourse there is a strong tendency for the information flow from A to P to correlate with an information flow from more to less animate and from more to less definite. In other words, the most natural kind of transitive construction is one where the A is high in animacy and definiteness, and the P is lower in animacy and definiteness; and any deviation from this pattern leads to a more marked construction. … the construction which is more marked in terms of the direction of information flow should also be more marked formally, i.e., we would expect languages to have some special device to indicate that the A is low in animacy or definiteness or that the P is high in animacy or definiteness[…] (Comrie 1989: 128).

Comrie also uses the “markedness” terminology, which does not lend itself to a clear causal explanation, but if “marked formally” is replaced by “coded overtly” and “marked in terms of information flow” is replaced by “non-usual (= rare) association of role and referential prominence”, then this is exactly the explanation that I am proposing. Perhaps the clearest statement in the earlier literature that appeals to frequency of use in explaining a subset of the phenomena discussed here is Dahl and Fraurud (1996) (as well as Dahl 2008).

Importantly, the functional view of role-reference association effects has no problem explaining some of the features of the crosslinguistic patterns that are otherwise hard to account for:

1. The patterns that we find are implicational universals, not unrestricted universals. In general, implicational patterns suggest functional motivations rather than innate representational constraints.

2. We often find optionality of coding around the cut-off points, as has been described, for example, in Spanish (e.g., von Heusinger and Kaiser 2005; for optional case-marking more generally, see Lestrade 2013).

3. Asymmetric coding splits (Sections 4–9) and asymmetric coding alternations (Section 10) can be seen to fall under the same larger generalization, even though their status as grammatical rules is very different. The idea that “soft constraints mirror hard constraints” (Bresnan et al. 2001) is much more readily explained in a functional view of the origins of these grammatical patterns than if one attributes crosslinguistic regularities to an innate grammar blueprint.

4. What matters for the crosslinguistic generalizations is the overt markers, not the abstract patterns. For example, while abstractly, many Australian languages exhibit tripartite case alignment patterns (Baker 2015: Section 1.2.2), these patterns show realizations by overt markers that are fully in agreement with the predictions of frequency-based coding efficiency. This is not explained by generative theories of abstract case (as is admitted by Baker 2015: 23, Note 12).

Finally, let me make a few remarks on the causes of the associations of role rank and referential prominence that we observe empirically. These associations are of course no surprise: Agent and recipient arguments have a strong tendency to be animate (and therefore definite), because humans are primarily interested in actions carried out by humans, as well as in transfer events with human recipients. This is what leads us to say that agents and recipients have a “high role rank”. I am somewhat hesitant to say that this is a real explanation, because one might justly ask what leads speakers to classify agents and recipients as coherent role classes to begin with (it could be that these are innate categories of a grammar blueprint). But regardless of the explanation of the associations, they are empirically testable, because linguists generally agree that agent roles and recipient roles can be identified in languages (and texts) independently of their referential prominence properties. For the efficiency explanation of form-frequency correspondences to go through, it is not necessary that we have an explanation of the usage frequencies (Haspelmath 2021).

11.3 Ambiguity avoidance

So far, I have not mentioned ambiguity avoidance as an explanatory factor, although intuitively, this makes a lot of sense: Languages serve to convey speakers’ thoughts to their interlocutors, so their expressions should not be ambiguous. And indeed, the classic cases of split P flagging (differential object marking) help the hearer to distinguish between the A and the P, in that they provide special marking precisely for those P-arguments that are more like typical A-arguments (in that they are definite and/or animate). The same reasoning applies to split A flagging, as well as to split R and T flagging, and also to the scenario splits.

So why am I not saying that the explanation of the grammatical patterns lies in ambiguity avoidance,^[23] and instead appeal to a more abstract principle of efficient coding and predictability? There are three reasons for this.

First of all, ambiguity avoidance can be seen as merely a special case of efficient coding. If grammatical markers are preferentially used when the grammatical meaning is least predictable, then it follows that they should tend to be used when there is a danger of ambiguity, because ambiguity is an extreme form of lack of predictability. For differential object marking (DOM), this point was made by Newmeyer (2005: Section 4.9.2) (and endorsed by Hawkins 2014: 194).

All that one needs to adopt is the well-established hypothesis that within a given domain, more frequent combinations of features require less coding that less frequent ones. There is no need to appeal to ambiguity reduction to explain the phenomenon of DOM (Newmeyer 2005: Section 4.9.2).

Second, and more importantly, if ambiguity avoidance were the primary explanation, we would expect that languages can use diverse coding means to ensure unambiguous interpretation, including anti-efficient coding. For example, one could imagine languages in which topical agents are always ergative-marked and indefinite patients are always accusative-marked, but focused agents and definite patients are zero-coded. Such a language would have very little potential ambiguity (because only clauses with focused agents and definite patients would lack flagging, and these clauses are very rare), but this type is completely unattested. By contrast, languages of the mirror-image type, which are actually widely attested, show at least as much potential ambiguity (when the patient is indefinite and has no accusative marking). This shows that efficiency of coding is crucial for understanding limits on the world-wide distribution of languages.

Third, there are cases where a longer expression is required (or is possible) for the less frequently used meanings even though there would seem to be nothing wrong with the more regular, shorter expression. For example, in ditransitive person-role combinations, languages sometimes ban perfectly regular and non-ambiguous forms, as we saw in (45) from in Modern Greek, repeated here.

(45b) is perfectly clear and unambiguous, but the language still requires the longer form Tu éðose eména ((45c) in Section 7.3 above). A similar argument on the basis of experimental results is made by Diessel (2019: Section 11.5.2).

Of course, all cases of asymmetric alternations are precisely of this kind: A sentence such as A girl saw the stork is perfectly unambiguous, but languages still often prefer more complex forms (with additional flagging and/or additional verb coding) such as The stork was seen by a girl. This is explained by the preference to have additional coding for less predictable forms, even if these forms are no less ambiguous than the simpler forms.

11.4 Biocognitive-representational constraints of a grammar blueprint

Over the years, there have been a large number of proposals in generative grammar for how to account for various aspects of the phenomena that are here subsumed under role-reference associations. Since split P flagging (differential object marking) is particularly prominent, this has been treated frequently, but it seems that no consensus has emerged (Kalin (2018) surveys quite a few different recent approaches). Likewise, for person-role interactions (cf. Section 7.1), there are many competing ideas (see Tucker 2013: 254–276 for an overview). There is also some generative literature on other scenario splits, but most of this work does not establish clear links with single-argument splits.^[24] What most of the generative papers share is that they talk about particular languages, but at the same time seem to propose highly general analyses that are motivated by many other considerations. This kind of work is thus very hard to compare to the present proposal, which focuses on claims about universal tendencies in the world’s languages.

There is really only one line of research in the generative tradition that is similar in generality to my proposal, and this is Aissen’s (1999, 2003) work, which has been widely cited. Aissen uses optimality-theoretic (OT) notation to express the universal tendency that role rank and referential prominence go hand in hand. There had been earlier hints in the literature about such role-reference connections, as in the following quotation from Farkas and Kazazis (1980).

[I]n the Rumanian clitic system, the case hierarchy [Ethical > Goal > Theme] and the personal hierarchy [1 > 2 > 3] are not supposed to conflict. Where there is no conflict […], the string is grammatical. Where there is strong conflict […], the sequence is unacceptable […] (Farkas and Kazazis 1980: 78; see Haspelmath 2004: Section 2.6).

But Aissen was the first to incorporate the prominence scales known from the functional-typological literature into generative grammar, in an attempt to provide a truly general explanation for all languages.

However, Aissen’s papers diverge from the standard generative approach in that they do not really make crucial reference to representational constraints of a grammar blueprint (“universal grammar”). On the one hand, Aissen’s work is generative in that it appeals to universal constraints in order to simultaneously describe individual languages and explain the gaps in attested languages. Her complex system of fixed constraint subhierarchies with constraint conjunction, and with intercalated constraints that penalize overt case-marking, derives the implicational universals nicely. But on the other hand, she does not really claim that these constraints are specific innate biocognitive mechanisms (as part of a grammar blueprint). Instead, she connects her approach with functionalist ideas about markedness and iconicity:

The effect of local conjunction here is to link markedness of content (expressed by the markedness subhierarchy) to markedness of expression (expressed by *Ø). That content and expression are linked in this way is a fundamental idea of markedness theory (Greenberg 1966; Jakobson 1939). In the domain of Differential Object Marking, this is expressed formally through the constraints [shown immediately above]. Thus they are iconicity constraints: they favor morphological marks for marked configurations (Aissen 2003: 448).

Another quotation makes it particularly clear that for Aissen, the optimality-theoretic machinery is primarily a notational device:

“OT provides a way … to reconcile the underlying impulse of generative grammar to model syntax in a precise and rigorous fashion with a conception of DOM which is based on prominence scales. The purpose … is to develop an approach … that is formal and at the same time expresses the functional-typological understanding of DOM” (Aissen 2003: 439).

Even though the orthodox position seems to be that all constraints (plus the OT architecture) are innate, Aissen is not alone in linking OT notation with functional considerations. Whatever one thinks of the machinery used in her papers, it seems clear that there is no argument against my functional-adaptive view of role-reference associations here.