2.1 (Split) ergativity and agent prominence

The first deviation from the typologically prevalent pattern of coding agent prominence comes from ergative languages, of which Tima, discussed in 2.2, serves as an example. Ergative systems differ from nominative-accusative systems in which the sole argument of an intransitive sentence (S) and the agent-like argument (A) of a transitive sentence are assigned the same case (nominative), while the more patient-like argument (O) is marked differently (accusative). By contrast, in an ergative system S and O are coded by the same morphological case (absolutive), while A is assigned a distinct case (ergative).² This situation is schematized in Figure 1, illustrating accusative and ergative alignment systems.

Figure 1: Accusative and ergative alignment.

It should be noted that languages differ in how exactly they express ergative alignment (i.e., how exactly they apply morphological and tonal case marking, verb agreement patterns, etc.; see, e.g., Comrie 2013; Stoll & Bavin 2013). Furthermore, in most ergative languages, the ergative alignment is not exceptionless but rather confined to certain subsystems of grammar. That is, “[m]ost morphologically ergative languages […] show what is referred to as ‘split’ ergativity, where different referring expressions (e.g. nouns, pronouns) pattern according to an ergative or accusative system, depending on their person and animacy features […]” (Dunn & Meakins 2023: 217, referring to Australian languages with ergative alignment systems). Such splits may correlate with grammatical conditions, e.g., word order (as is the case in Tima, Niger-Congo) or TAM (as in Gujarati and Hindi, both Indo-Iranian), or by verb type (as with intransitive verbs in Basque, linguistic isolate). Other languages “exhibit ‘optional’ ergativity where the use of the ergative marker is variable, but its non-use does not affect the grammatical role of A, […]” (Dunn & Meakins 2023: 217).³ Cross-linguistically, ergative systems of alignment appear to be dispreferred (Nichols 1992; Bickel et al. 2015). For instance, Bickel and colleagues found that most languages in their sample tend to prefer a case system that treats S and A alike (accusative alignment) (Bickel et al. 2015: 2). However, even though rare, ergative alignment seems to entail a different notion of agent prominence compared to nominative alignment (e.g., Dowty 1991). This is because the association between agent and patient roles and grammatical relations (or morphosyntactic marking) has been argued to be reversed in languages with an ergative system (Dowty 1991: 582). That is, languages with an ergative system differ from languages with a nominative/accusative system in how they correlate proto-roles (agent vs. patient) and morphosyntactic marking (or even grammatical relations), resulting in an inverse relationship between the two. Unlike in languages with a nominative/accusative system, ergative systems mark the agents of transitive and intransitive sentences differently and even align the intransitive S with the O (patient) rather than with A. Conversely, the agent of a transitive event (A) receives a special marking which sets it apart from S.⁴ Given the opposite directions of correlating agents and morphosyntactic marking, this may result in different interpretations of the agent category in speakers of (split) ergative languages compared to languages with nominative/accusative systems. To illustrate this idea, take language comprehension as an example. Since the grammatical subject in many nominative/accusative languages receives the default case with the weakest marking, speakers commonly tend to interpret such a default or base form as an agent argument (e.g., Bickel et al. 2015). However, since, in languages with (split) ergative systems, agents receive a special (ergative) case marker but patients receive the default case with the weakest marking, speakers of these languages may be encouraged to interpret unmarked noun phrases as patients, rather than agents (see, e.g., Bickel et al. 2015, for this argument). In other words, the default interpretation of base forms (or unmarked forms) as agents may be attenuated in languages with a (split) ergative system, thereby potentially casting doubt on the notion of agent prominence as defined above.

Although some typological and neurophysiological studies have touched upon the role of ergativity in agent prominence (e.g., de Hoop & Narashimhan 2008; Bickel et al. 2015), it is still far from clear whether and how agent prominence may be realized in languages with such systems. To shed more light on this issue, we will zoom in on Tima, a language with a split ergative system (which we will describe in more detail in the following section). We employ picture-based elicitation to make the agent or patient visually more or less prominent (e.g., by manipulating the affectedness of the patient, by varying animacy, and by occluding the agent so that it is only partly visible). This fine-grained manipulation of agent and patient properties provides insights into how ergative marking is applied in response to agentivity features and likewise offers a test case for whether the notion of agent prominence can be upheld in a language with a (split) ergative system.

2.2 Split ergativity in Tima

Tima, investigated by a number of scholars from Germany and Sudan since 2006 (following a pilot study by Dimmendaal conducted between 2002 and 2004, published in 2009) is one of the three languages of the Katloid group.⁵ They belong with greatest probability to the Niger-Congo phylum (Dimmendaal 2018; 2023; for a more doubtful view see Güldemann 2018). Tima makes use of a split ergative system which provides speakers with the option to choose a transitive construction that does (example (1)) or does not (example (2)) place the agent first. In the former case, neither agent nor patient receives any marking, while in the ergative construction in (2), the agent is morphologically marked. This means, in more detail, that we find a homorganic nasal precliticized to the agent noun phrase. The ergative marker N= is homonymous with the nominal instrumental marker in Tima. We follow Casaretto et al. (2020: 119), who argue that “the ergative construction developed from an active construction that introduced a cause by means of the instrumental preposition”. (For more details see Section 4.3 in Casaretto et al. 2020.)

Example (3) has a preverbal intransitive subject which is not ergative-marked. Under restricted circumstances, the subject of an intransitive clause may also appear after the verb. In this case the intransitive subject is also ergative-marked, but the verb needs to be derived for instrumental marking as well (see example (11)).

Table 1 presents an overview of the marking of arguments in correlation with word order. As will be discussed below, additional word orders (not included in the table) are possible when focus marking is involved.

To sum up, ergative marking in Tima closely correlates with word order: whenever the subject is not in sentence-initial position, it is obligatorily marked. In this sense, one may consider Tima as exhibiting split ergativity. Speakers “have a choice whether to present a proposition in the form of, e.g., AVO […] (without case marking of the agent), […], or OVA_ERG (with case marking of the agent) […]. That is, from this perspective, these languages can alternatively be characterized as having optional ergativity […]” (Casaretto et al. 2020: 112). Irrespective of the chosen perspective, the situation poses a potential challenge for agent prominence. First, it involves a deviation from the basic word order AVO (with object-first OVA order) and, second, the agent bears case marking while the patient does not.

Having reviewed the marking options of nominal arguments, specifically of S/A and O, we will now take a look at the verb. In general, the verb carries information referring to the main arguments. This consists, to begin with, of a prefix which shows agreement in number and person for S/A. In the case of the third person singular vs. plural, this prefix is not differentiated, however. That is, if both A and O are third person, as in ‎(1) and ‎(2), the assignment of the prefix ám- to the argument in either preverbal or postverbal position can only be made by analogy with examples with 1^st or 2^nd person participants, as here it becomes clear which argument is cross-referenced on the verb. Consider ‎examples (4a) and (4b), the former illustrating an unmarked construction, the latter an ergative construction. We not only find the prefix cém- indexing 1sg agent in both sentences, but also the non-ergative pronominal enclitic =dá in (4a) and the ergative pronominal enclitic =ná in (4b) cross-referencing the agent as well. For third person subjects, no such enclitic exists. If appropriate, third person pronouns for 3sg and 3pl are used. These pronouns and proper names behave like nouns with regard to ergative marking. Pronominals referring to 1sg, 2sg, 1pl (incl/excl) and 2pl are instead encliticized to the verb. Due to syncretism, the difference between ergative and non-ergative encliticized forms is only recognizable with 1sg (see Dimmendaal 2010: 235), as in (4).

Under the circumstances described, the question arises as to whether the agent in Tima is still more prominent than other arguments, as claimed above. To be able to answer this question, we first searched comparable Tima data sets regarding the frequency of constructions and the conditions of their occurrence. The data taken into consideration comes partly from text recordings, partly from the usage of photo stimuli, and partly from utterances we elicited with the help of the questionnaire on information structure developed by Skopeteas et al. (2006), especially Task 3 (described on page 39ff. of the manual). Several variables, such as the givenness, animacy and identifiability of the agent, are considered to provide insights into the choice of construction.

The following observations let us speak in favour of agent prominence. First of all, transitive sentences in elicitation contexts almost exclusively have AVO word order. We considered 100 sentences elicited with picture stimuli (source: Skopeteas et al. 2006, Task 1), as well as photos taken in the Tima area, all displaying transitive events, e.g., the stimulus used for example (7), or photos like the one used as a stimulus for (5). Of these stimuli, 98 were described with clauses having AVO(Obl) word order, whereas only two have OVAerg word order (see example (13) as one of these two). (These results resemble the proportions between the many sentences elicited through translation, which we did not include in this context to avoid false numbers regarding the basic word order when counting translation equivalents.)

Furthermore, and most decisively, interpreting the first of two participants of a Tima sentence as the patient is not possible when both participants are unmarked. Not surprisingly, the first-mentioned participant is interpreted as A when O is lower on the animacy hierarchy than A, as in (5); ìɽɨ́nkɨ̀k ‘sorghum beer’ as the agent is, of course, also pragmatically excluded.

Consider then (6) with two human participants, a ‘woman’ and a ‘man’, either of whom could pragmatically be A or O. However, the only possible correct reading is that ‘the woman has hit the man’, thus following an agent-first interpretation.

As these examples illustrate, word order is decisive concerning the interpretation of participants as agent vs. patient. Thus, we are dealing here with a case of word order freezing which, as described by Himmelmann & Primus (2015: 52), “[…] refers to the fact that in the absence of other clues (such as case, agreement, animacy, and context/intonation), it is the word order which straightforwardly determines the interpretation of verbal arguments in an ambiguous sentence”. This concerns comprehension, as exemplified with (6), but also sentence production, as demonstrated by elicitation work with picture stimuli, exemplified in (7).

Crucially, when the A argument is prominent in a number of different dimensions, including having features of proto-agentivity (such as being highly agentive, animate), speakers select the non-ergative construction, thus displaying an agent-first bias. That is, if shown picture stimuli with an identifiable A acting on O, it is exclusively the AVO construction that is chosen. This holds true when both A and O are animate, as in (7), and thus of “equal rank in animacy” (Himmelmann & Primus 2015ː 52), and also when A ranks higher on the animacy hierarchy, as in (5).

To sum up briefly and concisely, the AVO construction in Tima is considered basic when taking into account frequency of word order in elicitation contexts, including having a default and exclusively agent-first interpretation with unmarked A and O.

By contrast, if we look at longer chunks of discourse, the picture that emerges regarding the frequency of word order is less clear. In ten narratives consisting of 851 clauses, we find 344 transitive clauses; see Table 2 for an overview of intransitive and transitive clauses, including information about the overtness of S/A and P arguments. Note that the ergative argument needs to be overtly expressed, i.e., it cannot be reduced to zero as is possible with other arguments in Tima (non-ergative-marked S/A, O, BEN or INS). As we can see in Table 2, still more than half (54%) of the A arguments are not overt. By comparison, 38% of the P arguments are covert. Conversely, we find only 46% overt A arguments in transitive clauses, but 62% overt P arguments.

The transitive clauses are those that are relevant if we want to compare how many clauses exhibit AVO word order as opposed to OVAerg word order. Since, depending on the context, Tima allows for zero S/A and O, only 157 out of 344 clauses have an overt A argument realized as a full noun phrase or a pronoun. Of these 157 overt A arguments, 91 are preverbal A arguments as opposed to 66 postverbal Aerg arguments. That is, roughly 42% of the transitive clauses with overt A have the ergative construction. Considering all transitive clauses (i.e., including those with zero A), the ergative construction occurs in 19% of the clauses. While an evaluation of the conditioning factors of this distribution is still pending, our findings from stimulus-based work reveal several factors playing a role in determining which word order is preferred. One of them is illustrated in (8), in which the same event as in (7) is described but by using the OVAerg construction instead of the AVO construction. The contextual difference lies in the identifiability of the agent (compare the stimulus pictures). While the human stimulus agent is uniquely identifiable in (7), in (8) it is backgrounded, to the effect that one cannot identify the agent.

The same construction occurred when the speaker was asked to describe photos taken in the area in which the agent was not entirely visible or recognizable; compare, for instance, example (9), with OVAergObl word order, to example (5) above, displaying AVOObl word order.

Thus, while AVO is chosen if the agent is clearly identifiable, word order (together with ergative case marking) changes as a function of this property. Once the agent is no longer identifiable (or if it is backgrounded), the OVAerg construction is chosen to reflect the increased prominence of O as opposed to A. In other words, and more generally speaking, AVO (in line with an agent-first preference) seems to be used as the default, while OVAerg is used to accommodate the fact that O becomes more prominent than A.

In addition to the simple AVO and OVAerg constructions, the stimulus-based work brought forth a third option for coding a transitive event, i.e., the decomposition of the event into two (or more) sub-events, as illustrated with (10) and (11). This option is chosen when the agent is inanimate/non-human while the patient is human, i.e., when the agent is lower on the animacy scale but when both participants are identifiable. The human patient is mentioned first in an intransitive clause and is thereby made the attentional centre, and then the agent is introduced in the following clause as an ergative-marked participant. Thus, if A is inherently less and O inherently more agentive, the speaker tends to, but does not necessarily have to, choose O as the more prominent argument.

Note that in almost all responses to stimuli with human patients and non-human agents, the patient is not overtly expressed in the later transitive construction. Thus, (11) is a more typical example of non-prominent agents. We reflect the fact that P is not overt by bracketing the pronoun ‘him’ in the English translation, and regard the expression of P with zero as indicating the attentional prominence of P. The overtness of P in (10) may thus be due to the fact that two agents as well as two patients are contrasted. Further research needs to be carried out before a final verdict can be made.

Note also that in Tima, any causer, be they animate or inanimate, can easily take the agent role. Thus, language-specifically, causation is one of the main features mentioned by Dowty (1991) that is used to build up agency. Sentences like ‘the knife cut me’ or ‘the rope/stone downed me’ are common in Tima.

It is worth mentioning that information structure also plays a role with regard to sentence structure in Tima. Specifically, as shown by Schneider-Blum (2023: 105), “[t]he information structural notion of givenness interacts with the two parameters of agent-identifiability and animacy”. In an elicitation context involving two subsequently presented pictures with either A or P shown in the first picture, we find that given A always leads to the AVO construction in the following clause, while the connection between given P and OVAerg word order is less tight (see Schneider-Blum & Hellwig 2018; Schneider-Blum 2023). This latter, unclear correlation of given P with the ergative construction can be explained in terms of the assumptions speakers make. When being confronted with the first picture, they assign prominence to the only participant, i.e., they make it the attentional centre. Prominence is associated with agency. When, in the second picture, this same participant turns out to be not A but P, the speakers are caught in conflict: Should they keep their focus of attention on the previously established participant (now P), or should they make the newly presented A more prominent? Thus, the speaker has to choose between given prominence and inherent A prominence and the choice results in the two different constructions. Preliminary results (see Schneider-Blum 2023: 99) show that given P leads more often than not to an ergative construction (out of 35 second clauses, 25 have a postverbal ergative-marked agent, while 10 have AVO word order).

To put it concisely, in Tima, the sentence-initial position is reserved for the most prominent / prominent-made participant (labelled as ‘topical constituent’ in Casaretto et al. 2020, and as ‘attentional centre’ in Schneider-Blum 2023). The prominence of O correlates, as described, with the alternation of the constituent order from AVO to OVAerg.

Taking focus marking into account, further alternations (which we show in turn) are possible, namely AfocVO, AOfocV, OfocVAerg and OAerg.focV. In a nutshell, focus marking always occurs after wh-questions, be they explicit or assumed by the speaker, as well as in correction contexts (see Becker & Schneider-Blum 2020). Like ergative marking, focus marking also interacts with word order in that a focus-marked constituent is invariably in the preverbal position, although it is not necessarily sentence-initial. With regard to AfocVO, then, A is maximally prominent, since it is not only the attentional centre but additionally receives focus marking, thereby highlighting A (see example (7) above). With AOfocV, as displayed in (12), A also occupies the a-centre, but O is relatively more prominent due to focus marking than it would be without focus marking in AVO order.

Considering OfocVAerg worder, as exemplified with (13), we find O maximally prominent, being in the a-centre and focus-marked. Answering the question ‘What do you see?’, the speaker assigned the ‘roof’ more prominence than the agent, correlating with the former being visually foregrounded, i.e., more prominent, than the agent (although the agent is identifiable). The visual foregrounding of O thus goes together with the linguistic foregrounding of O. This is a good example to show that the speaker does indeed have a choice regarding who/what to assign (more) prominence to and that the form of the utterance is based on the speaker’s assessment of the situation.

The most complex situation is found with OAerg.focV, as in (14), since ergative-marked A additionally receives focus marking. In that case, the rule that focus-marked constituents have to be preverbal overrides the default rule that ergative-marked constituents occur postverbally.

To sum up: The ergative construction makes a shift in prominence from A to O possible. Focus marking of a constituent contributes to its prominence status and allows for a more fine-grained scaling of prominence. The Tima data definitely show that prominence cannot be equated to focus-marking or contrast, because both inherently prominent (considering the animacy hierarchy, for example) and less-prominent entities can receive focus marking, suggesting that these parameters can be dissociated. The close connection between word order and ergative marking does not allow us to assign the triggering factors to one or the other, i.e., we must necessarily relate the conditions to both marking strategies. Information structural factors do have an influence on whether or not to choose the ergative construction. Thus, the findings for Tima are in line with the generalization made by Dunn & Meakins (2023: 224) that “all optional ergative languages have in common […] the interaction of information structure and the (non)appearance of ergative marking”.

The examples showed that the (visual) backgrounding of the agent as well as an agent’s lower animacy status and the givenness of the patient run counter to the expectation of the agent being mentioned first. Nevertheless, even for Tima we argue that these are explainable deviations from the default. Consequently, Tima appears to adhere to the cross-linguistic principle in that agents tend to be placed first (under the condition that they exhibit certain Proto-Agent features, of which animacy is but one, and that discourse factors are minimalized), while deviations from this pattern arise under special contexts (e.g., Dryer 2013; Napoli & Sutton-Spence 2014). When comparing the frequency of occurrences of AVO word order and OVAerg constructions in elicitation contexts as opposed to occurrences in narratives, this points in the same direction, i.e., the need for special conditions that lead to the use of the ergative construction.

3.1 Symmetrical voice systems and agent prominence

The second challenge to the cross-linguistic validity of agent prominence is illustrated by means of Totoli (Western Austronesian, Indonesia), a language with a symmetrical voice system. Symmetrical voice (also known in the literature as Austronesian alignment or the Philippine focus system) is specific to a number of western Austronesian languages, including languages from Indonesia, the Philippines, Taiwan and Madagascar (Himmelmann 2005; Riesberg 2014; Chen & McDonnell 2019). These languages exhibit at least two different voices, called here actor voice and undergoer voice, which are equally transitive and equally morphologically marked on the verb.⁸ As illustrated in example (15a) from Totoli, in the actor voice, the subject has the semantic role of actor, whereas the other core argument (the non-subject) has the semantic role of undergoer. The undergoer voice, exemplified in (15b), shows the opposite alignment: the subject function is assigned to the undergoer argument, whereas the non-subject argument has the role of actor. Each voice predicate is obligatorily marked by a different set of voice-mood affixes: in the examples at hand, the actor voice is marked by the actor-voice prefix noN- and the undergoer voice construction is marked by the prefix ni-. These affixes mark voice and mood at the same time. That is, each voice can come in two different moods (realis and non-realis). This can be represented in the paradigm in Table 3. The same table shows that Totoli exhibits two different undergoer voices (called here UV1 and UV2). Lexical roots are determined for one or the other and there does not seem to be a semantic difference between the two.

The symmetrical voice system differs significantly from the (asymmetrical) voice system found in European languages and many other languages of the world in three main regards.⁹ First, the two systems differ with regard to transitivity. In an active-passive alternation like the one in (16), the active construction is transitive (it has the two core arguments the coconut and Jui), while the passive construction is intransitive (the agent is oblique and, moreover, can be omitted). By contrast, as shown in (15), both the actor and the undergoer voice are transitive; i.e., both Jui and the coconut are core arguments in (15a), as well as in (15b). Second, asymmetrical and symmetrical systems differ with regard to morphological marking. The English active-passive alternation in (16) is asymmetrical in that the active verb is morphologically unmarked, whereas the passive verb is marked with the verb be and the past participle form. In contrast, both constructions in (15a) and (15b) are marked for voice: the actor voice predicate is marked with the actor voice prefix non- and the undergoer voice predicate takes the undergoer voice prefix ni-.

Third, and most importantly for the current argumentation, asymmetrical and symmetrical voice alternations differ with respect to usage. In languages like English, the use of a passive construction is not a neutral choice: passive voices are typically less frequent than active voices and are used only under certain conditions (for instance, when the undergoer argument has a lower information value or a higher level of activation than the actor; see, e.g., Biber et al. 1999: 938–943). Otherwise, when actor and undergoer are equal with regard to the relevant parameters at a specific point in discourse (e.g., animacy, topicality, givenness, etc.), there is a default preference for the active construction. This is a case in point for the higher prominence status of the agent with regard to the patient (or any other role), as the least morphologically marked and most frequent construction is the one which assigns the subject function to the agent.

In symmetrical voice languages, by contrast, there is no such distribution between actor and undergoer voices. Neither of the two voices is more basic or less marked than the other: that is, both are equally morphologically marked on the verb and identical in terms of transitivity (no argument is demoted). In terms of frequency, symmetrical alternations also contrast with asymmetrical alternations because constructions in which the agent is subject are not significantly more frequent than constructions with undergoer subjects. In fact, in Totoli, the corpus frequency of undergoer voice constructions (70%) is higher than for the actor voice (30%), and this holds true across different genres and text types.¹⁰ Similar distributions are reported also for a corpus of Tondano (North Sulawesi, Indonesia; Brickell 2014). In corpora of other symmetrical voice languages, the percentages of actor and undergoer voices are often relatively close to a 50/50 distribution (see Pastika 1999 for Balinese, Quick 2005 for Pendau and McDonnell 2016 for Besemah).

Languages with symmetrical voice systems deviate from the notion of agent prominence in that there is no default syntactic alignment of agents with the subject function. Transitive constructions in which the undergoer argument is linked to the subject function (i.e., an undergoer voice) are as morphologically marked and as frequent or even more frequent than those with an actor voice.

This also has consequences with respect to the agent-first principle. As mentioned above, if agents had a prominent syntactic status, one would not only expect them to be linked to the most prominent syntactic function (i.e., the subject) but also for them to occur in the most prominent position within the sentence (i.e., clause-initially). These two predictions align in languages which tend to link agents to the subject function and exhibit a basic word order pattern in which the subject precedes all other arguments. However, in symmetrical voice languages, a subject-initial word order does not always imply an agent-initial word order and vice versa. As can be seen in the undergoer voice in (15b) above, PVA orders are allowed in Totoli. In fact, Riesberg et al. (2019: 528–539) show that it is also common for other symmetrical voice languages to allow for both basic word order patterns, those in which agents precede patients and those in which patients precede agents. In this regard, then, symmetrical voice languages also seem to violate the agent-first principle.

3.2 Agent prominence persists, even in symmetrical voice systems

Totoli and other symmetrical voice languages thus seem to be in confrontation with the universality of agent prominence. They do not link the subject function to agent arguments by default and they exhibit basic word order patterns that do not follow the agent-first principle. Nevertheless, this does not necessarily entail that agent prominence does not play a role in the organization of the grammar of such languages. The hypothesis put forward by Himmelmann & Primus (2015) and Riesberg & Primus (2015) that agents will receive more prominence than other semantic roles in linguistic structures does not dictate how this prominence is linguistically realized. That is, there can be variation in the way in which languages conventionalize agent prominence in grammar. First, the hypothesis does not imply that the linking of semantic roles to syntactic functions is the only area of grammar where the prominence of the agent is made visible. And second, it does not imply that the linking of semantic roles to syntactic functions always responds to the semantic hierarchical ordering of the arguments. This is made particularly explicit in Riesberg & Primus (2015: 552), who note that “the linking of the privileged syntactic argument (PSA) to patient does not predict that the grammar of the respective language is patient-oriented” and that “systems of grammatical relations reflect a number of different factors, of which position on the semantic role hierarchy – or, as we would rather have it, agent prominence – is but one”.

In fact, symmetrical voice languages provide evidence for agent prominence beyond the alignment of semantic roles and syntactic functions. In the following, we support this argumentation with our own data which illustrate two distinct morphosyntactic phenomena in Totoli: word order variation (3.2.1) and causative and limited control morphology (3.2.2).

4.1 Psycholinguistic and neurophysiological studies on split ergative languages

As outlined in the Tima section, we find evidence for agent prominence and an agent-first bias. Formerly, it has been argued for ergative languages that the agent role seems to be outranked by the patient role in terms of its prominence status because these languages follow an S/O-alignment (e.g., Dowty 1991: 582). Furthermore, because ergative case marking is restricted to agents, speakers of ergative languages may be encouraged to interpret unmarked NPs as patients, rather than as agents.

However, there is neurophysiological evidence indicating an agent-first preference in split ergative languages like Hindi, which uses ergative marking in the perfective aspect (e.g., Choudhary 2011; Bickel et al. 2015). Specifically, it has been proposed for Hindi that “when readers or listeners encounter a base-form noun phrase (NP) in a sentence […], the processing system first assumes that this NP refers to the S argument […] or to the A argument of a transitive verb […]”. (Bickel et al. 2015: 2). To test this proposal, Bickel and colleagues investigated how speakers of Hindi comprehended unmarked (nominative) sentence-initial NPs in comparison to accusative-marked sentence-initial NPs in an ERP study. Sentence-initial NPs were inanimate, since in Hindi accusative case marking of the syntactic object is optional for inanimate NPs. Hence, a sentence-initial inanimate NP that is left unmarked is then ambiguous between subject (S/A argument) and object (P argument) interpretation. By contrast, a sentence-initial accusative unambiguously identifies the P argument. If Hindi follows an agent-first preference for incremental argument interpretation, then sentence-initial unmarked (and hence ambiguous) NPs are assumed to receive the S/A interpretation and a subsequent disambiguation towards a P reading will lead to enhanced processing costs. Contrary to this, if Hindi follows an undergoer-first strategy in the case of an unmarked sentence-initial NP, no disambiguation costs are predicted, compared with sentence-initial accusatives, since a sentence-initial argument would receive an undergoer interpretation by default. Indeed, Bickel and colleagues found evidence in favour of the first hypothesis. Specifically, for conditions with sentence-initial base-form NPs, in contrast to sentence-initial accusatives, Hindi speakers showed an increased neurophysiological response (N400 and late positive shift) at the position of the subsequent transitive verb that identifies the previous NP as the P-argument. Thus, even though in Hindi initial base-form NPs can turn out to be patients (OVS structures), speakers of Hindi nevertheless seem to interpret an initial base-form NP as the S or A argument, similar to speakers of other languages that do not use ergative marking (see, e.g., Haupt et al. 2008 for German). These findings are in line with the assumption of a general agent-first bias across languages (for converging evidence see also Isasi-Isasmendi et al. 2024 for Basque, another language with ergative case marking).

However, some caveats deserve to be mentioned: First, ergative case marking in Hindi is restricted to the perfective aspect. Given the restricted use of ergative alignment, speakers of Hindi may interpret ambiguous noun phrases in line with the nominative-accusative pattern they frequently encounter in non-perfective contexts. This scenario is even more likely since all of the Hindi-speaking participants were also fluent in English, a language devoid of ergative case marking. Thus, an agent-first preference in Hindi might result from a default interpretation strategy in which nominative-accusative alignment is assumed. Second, pertinent ERP literature on processing of split ergative languages makes a subtle but important distinction between the default preference for nominative-accusative alignment (S/A-alignment) and an agent-first interpretation strategy. In particular, Bornkessel-Schlesewsky and colleagues (2008), as well as Bickel et al. (2015), interpret their ERP findings as evidence for a general preference for nominative-accusative alignment. However, Choudhary (2011) reports ERP data that demonstrate processing costs for sentence-initial ergatives compared with sentence-initial nominatives. Although ergative morphology on the sentence-initial NP would unambiguously call for agent interpretation and therefore satisfy the agent-first preference, initial ergatives were found to elicit a late positivity relative to sentence-initial nominatives in an auditory ERP experiment. Experiment 2 replicates this finding in the visual modality. Thus, these data suggest that morphological markedness (i.e., ergative case marking) in sentence-initial position leads to additional processing costs compared to unmarked forms.

Interestingly, this is in line with the argumentation in Himmelmann & Primus (2015), who assume that, in a transitive argument relation, the most prominent argument is morphologically unmarked, precedes other arguments and receives the agent role. However, they do not point out how these factors (i.e., markedness, constituent position, role assignment) might interact. The ERP literature, as well as our observations from Tima, favour the assumption that argument prominence can be captured as an integrating function of multiple information types. The data on Hindi seem to support the view that linear precedence and unmarked morphology are both crucial for computing argument prominence, whereas a sentence-initial ergative “lacks” unmarked morphology.

While the agent-first bias has mainly been discussed as a processing mechanism operative during language comprehension (e.g., Bornkessel-Schlesewsky & Schlesewsky 2009; Bickel et al. 2015), some additional insights about agent prominence can be obtained when focusing on sentence planning during the production of ergative languages (e.g., Sauppe et al. 2021; Egurtzegi et al. 2022). Since these studies often use eye-tracking, they provide information about whether agents can be considered attentional centres (i.e., whether they attract initial visual attention, as argued by Himmelmann & Primus 2015; for experimental evidence, see, e.g., Cohn & Paczynski 2013). Indeed, Sauppe and colleagues (2021) found that when speakers of Hindi were asked to describe visual scenes depicting transitive events, speakers initially looked towards the agent more often, thus likely reflecting an agent-first bias. The same early attention towards the agent has been confirmed for Basque, another language with ergative case marking (see, e.g., Egurtzegi et al. 2022; see also Isasi-Isasmendi et al. 2023 for Basque speakers’ early attention towards the agent character during event apprehension).

Taken together, there is converging evidence to suggest that even if languages do not align S and A, as is the case for (split) ergative languages, speakers of these languages still tend to interpret an ambiguous (i.e., unmarked) first noun phrase as the agent argument. A similar bias has been observed during production experiments with ergative languages, supporting the notion that an agent-first preference may span both language comprehension and language production (i.e., sentence planning) and occurs regardless of a language’s means of morphosyntactic encoding (in line with the proposals by Himmelmann & Primus 2015; see also Bornkessel-Schlesewsky & Schlesewsky 2009; Bickel et al. 2015). Though similar psycholinguistic studies cannot currently be conducted for Tima, we can assume with reasonable certainty that they would produce comparable results.

4.2 Psycholinguistic and neurophysiological studies on symmetrical voice languages

While psycholinguistic evidence on language processing in languages with symmetrical voice systems is relatively sparse, recently there has been an increased interest in studying sentence processing in typologically diverse languages. Most central to the notion of an agent-first preference is a recent neurophysiological study conducted by Sauppe and colleagues (2023) examining language comprehension in speakers of Äiwoo, a language belonging to the Oceanic branch of the Austronesian language family (Ross & Næss 2007). Äiwoo is argued to have a symmetrical voice system (Næss 2015; 2021) and, most critically, the default word order is SVO, including in the undergoer voice, in which SVO means PVA. Given that the undergoer voice is the more frequent voice in this language (Sauppe et al. 2023: 7), Äiwoo shows a default patient-first order (even for human agents), which arguably poses a challenge to the notion of an agent-first preference. In reflection of this patient-first order, Sauppe and colleagues found that speakers of Äiwoo preferentially interpreted role-ambiguous non-human nouns (such as stone) as patients, which led to an increased N400 when the verb disambiguated the noun toward an agent role in a clause. Critically, however, for human referents, Sauppe and colleagues (2023) observed the opposite effect. That is, Äiwoo speakers preferentially interpreted role-ambiguous human nouns as agents, hence resulting in an increased N400 effect when the verb disambiguated the noun toward a patient role. Together, these results suggest that an agent-first bias can be overridden by speakers’ long-term experience with word order, as is the case for non-human referents. However, when the referent exhibits properties of a prototypical agent (i.e., when the referent is human and comprises features such as volition, sentience, and control; see, e.g., Dowty 1991), the agent-first preference persists even in speakers of a language like Äiwoo that preferentially places patients before agents. These findings provide evidence that prototypical agents do indeed trigger an agent-first preference, even if typological features of a language work against this. Specifically, an agent-first preference seems to persist even if a language exhibits a basic transitive construction in which the agent is not obligatorily linked to the syntactic function of subject (symmetrical voice system) and even if a language has a basic word order pattern in which the patient precedes the agent.

As mentioned above, although the agent-first principle may be more central to language comprehension, there appears to be complimentary evidence from sentence planning in symmetrical voice languages (e.g., Sauppe et al. 2013). That is, regardless of the word order that was eventually produced, speakers of Tagalog tended to first fixate on the agent character in a depicted event scene, suggesting that the agent remains the attentional centre even for speakers of symmetrical voice languages.

4.3 Evaluating the notion of agent prominence more closely

While the challenges posed by split ergative languages and by symmetrical voice languages are very different in nature – the former concerning syntactic alignment and the markedness of arguments, the latter, word order and the mapping of syntactic functions to semantic roles – there are a number of commonalities in these findings.

First, we see that in (split) ergative systems the most common word order pattern is represented by the preference for putting an unmarked argument at the beginning of the sentence. In contrast to this, the use of ergative constructions is much more restricted. Thus, ergative systems in general and ergative constructions in particular are less common than the nominative-accusative pattern. Within most ergative languages, ergative-marked constructions are used only under specific conditions. Furthermore, we see that although Tima and symmetrical voice languages allow both agent-first and non-agent-first word orders, the orders in which the agent precedes the patient are preferred. In elicitation contexts – i.e., when the influence of discourse factors is reduced to a minimum – speakers of Tima almost exclusively produce AVP (≙ AVO) orders. As for symmetrical voice languages, corpora from languages like Totoli and Tagalog show that AVP/VAP orders are used to a much greater proportion than PVA/VPA orders. Thus, although available, constructions in which the agent receives more morphological marking and does not precede the other roles are dispreferred.

A critical case in point might be Äiwoo, which has been observed to use the PVA pattern (undergoer voice) more frequently than the AVP pattern (see 4.2 above). However, behavioural and neurophysiological evidence indicates that an (unmarked) agent-initial word order induces lower processing costs despite the higher frequency of usage of PVA (cf. Sauppe et al. 2023) and points to the agent-first preference as an underlying processing strategy. As the data from Äiwoo show a three-level difference in ERPs according to the animacy scale (humans > animals > inanimate objects), we argue that a multi-dimensional approach to semantic roles in terms of basic role properties (see Dowty 1991) can capture these differences: Human referents can take over the full set of agent features while animate non-humans (animals) are excluded from some role features (volition/control) while keeping others (causation, sentience, independent movement). Inanimate objects, in contrast, are excluded from having most of the agent features (volition/control, sentience, independent movement). The ERP patterns from Sauppe et al. (2023) seem to reflect this three-level distinction, but they also exhibit a qualitative distinction between humans and animals/inanimates. At the level of morphosyntactic encoding, the limited control and causative constructions of western Austronesian symmetrical voice languages indicate that agent prominence plays a role beyond grammatical relations, since agent features (in this case, control and volition) are morphologically marked on the predicate even when the agent is not the subject of the construction. With regard to ergative languages, we did not tackle the topic of optional ergativity without word order change and its relation to agentivity. As summarized in Dunn & Meakins (2023: 224), in some Australian languages, the absence versus presence of the ergative marking might correlate with characteristics of the agent such as intention, animacy, the expectedness of a particular referent being an agent, etc. It remains to be investigated, however, whether the patterns in languages with such types of split ergativity conform to the universal prominence of agents that we observed in the languages reported in this paper. Finally, concerning word order, the Tima data show that the loss of Proto-Agent features and discourse- or context-related features (animacy, givenness and/or identifiability) that make an agent prominent correlate with the choice of a non-agent-first order.

This latter point goes hand in hand with a further issue. Grammatical structures are uttered in a context – discourse – which is dynamic by definition. As discourse unfolds, the prominence status of arguments shifts (Himmelmann & Primus 2015; von Heusinger & Schumacher 2019). Thus, while we argue for a universal principle of agent prominence that is based on fundamentals of human cognition, we observe that the way in which this prominence is manifested in language use will be influenced by a number of discourse factors. The prominence status of an agent in a particular clause is relative to the prominence status of the other arguments at a particular point in discourse. The role of factors such as information structure and the competition between arguments constitute a complexity which we have not elaborated on in the context of this paper. However, while we acknowledge that discourse-related factors certainly modulate the linguistic prominence given to an agent in a specific context, this does not necessarily provide counter-evidence to the assumption of a more general (or universal) concept of agent prominence. As shown for instance in Tima, some discourse contexts involve a situation in which the patient argument outranks the agent in terms of prominent features (such as animacy and identifiability). It is to be expected that such contexts facilitate grammatical structures that deviate from the agent-centred ones (e.g., a patient-initial word order). By contrast, when arguments are of equal rank, the claim is that the agent is given privileged treatment. Even when this privileged status of agents is not apparent in the grammatical organization of a clause, psycholinguistic and neurophysiological studies add evidence in favour of agent prominence. The fact that both speakers of ergative languages and speakers of symmetrical voice languages show a preference (in terms of ease of processing) for agent-first constructions supports the transversality and the cross-linguistic character of the agent-first bias.

With respect to the work on Äiwoo (Sauppe et al. 2023), the crucial question arises as to whether individual language experience can indeed affect or shape fundamental language principles like agent prominence. By broadening the scope of psycholinguistic studies to typologically diverse languages – especially those that seem to de-emphasize the importance of the agent – we can gain important insights into which aspects of agent prominence may indeed be universal and which ones are shaped by the particular features of a language.