In our experiments, we assess six surface constraints that can be assigned to three different categories, namely lexical-semantic constraints, syntactic constraints and discourse constraints:

ANI (animate > inanimate) Constraint (4a) refers to the animacy status of arguments and states that animate NPs should precede inanimate ones. This pattern is in line with the general tendency to put dative NPs, which are prevalently animate, before accusative NPs in the German middle field. Some authors regard this constraint as one factor within a set of competing surface constraints (Lenerz 1977; Hoberg 1981; Müller 1999). Other authors dismiss this constraint and consider it a mere confound of case or thematic roles (Uszkoreit 1986; Haider & Rosengren 2003).

AG (agent > non-agent) Constraint (4b) refers to the thematic roles of arguments and states that an agentive NP should precede all other argument NPs. While this constraint could be partly responsible for the canonical order of subject before object in German, it might also be confounded with either the syntactic constraint equivalent NOM in (5a), or with ANI, as agents are mostly animate entities. This constraint is part of Uszkoreit’s (1986) set of constraints3 and is entailed by Haider & Rosengren’s (2003) account of base order in terms of thematic roles.

REC (recipient/goal/benefactive > theme) Constraint (4c) is an alternative to ANI in accounting for the predominant order of indirect before direct object in German, making reference to thematic roles. It is advocated in slightly different forms in Haider & Rosengren (2003) and Uszkoreit (1986).

NOM (nominative > non-nominative) Constraint (5a) refers to the syntactic status of arguments and states that nominative subjects precede accusative or dative objects. While this constraint has been taken to be the source of the canonical SO word order in German and is often considered more important than other constraints (Müller 1999), it has also been noted that it may be violated in certain contexts without a cost (Lenerz 1977).

DAT (dative > accusative) Constraint (5b) is the syntactic counterpart of REC and states that dative NPs precede accusative NPs. This constraint has been proposed by different authors (Uszkoreit 1987; Müller 1999), sometimes in addition to seemingly competing constraints like ANI. While it appears to be at odds with traditional case hierarchies, e.g., the accessibility hierarchy of Keenan & Comrie (1977), it could be triggered by an underlying constraint requiring structural case (accusative) to appear adjacent to the verb, as argued in Heck (2000).

DEF (definite > indefinite) Constraint (6a) states that definite NPs should precede indefinite NPs. It was proposed in this simple surface oriented form in Lenerz (1977) and Müller (1999). Alternatively, it is conceptualized as a result of banning indefinite NPs from scrambling (Büring 2001; Lenerz 2001).

Some of the surface constraints described above form part of the account of Müller (1999), which extends standard OT to accommodate markedness. In accordance with much research in syntactic theory, standard OT classifies sentences as either grammatical or ungrammatical and offers no direct way to differentiate between degrees of grammaticality or markedness. In Müller’s variant of OT, alternative word orders can be grammatical and nevertheless differ in terms of markedness, in the sense of a dispreference of a marked structure compared to other less marked structures.

Müller’s account is also notable for combining minimalist syntax and OT. Like Haider & Rosengren (2003), Müller assumes that argument order variation is derived by scrambling. The underlying base order is not determined by semantic properties of the verb, however, but is assumed to be the uniform configuration “subject > direct object > indirect object” at D-structure. Deviations from this configuration can be achieved by scrambling the indirect or direct object when this is licensed by one of the constraints in (7). The last constraint, PER, counteracts existing lower ranked faithfulness constraints (STAY, PAR-MOVE) by favoring linearizations where the order of arguments is reversed relative to the base-generated order.

Any word order resulting from a triggered scrambling operation is grammatical, in addition to the base-generated order, which is always grammatical. This is implemented by means of a subhierarchy, subsumed under the constraint SCR-CRIT, which may be replaced by any of its subconstraints at evaluation. Formally then, every structure is grammatical that is optimal for at least one replacement of SCR-CRIT.

Müller (1999) accounts for relative grammaticality, or markedness, by imposing an internal ranking (>) on the constraints in the subhierarchy. To find the unmarked candidate for a given content, SCR-CRIT is replaced with the whole subhierarchy and evaluation takes place as usual. As one or more constraints in the subhierarchy might favor a scrambled word order, the base order is not in general the unmarked one, contrary to Haider & Rosengren (2003). For example, the direct object is base generated before the indirect object, but the dative constraint favors the scrambled order and renders it the unmarked candidate, if the candidates do not differ on higher ranked dimensions. The candidate set is defined by Müller (1999) as consisting of <D-structure, S-structure> pairs, which share the same numeration, i.e., the same lexical content. He also seems to assume the same meaning, which results in the set of <D-structure, S-structure> pairs with the same D-structure.

All unmarked structures that are identified by the extended evaluation including the whole subhierarchy, are expected to be of approximately equal, high acceptability, as long as they do not differ substantially with respect to complexity. For the remaining candidates, an order of markedness can be established using the whole subhierarchy and the concept of suboptimality introduced in Keller (1996): Candidate C_i is more marked than Candidate C_j, iff C_i is suboptimal to C_j. This definition imposes a ranking in terms of markedness, but does not quantify the differences in acceptability. In addition, it does not allow to compare sentences from different candidate sets, i.e., sentences that do not directly compete because they differ in content or information structure.

Linear Optimality Theory (LOT) proposed in Keller (2000) provides a fully quantified model of constraint application. Similar to Müller (1999), Keller uses surface constraints to account for word order preferences in German, in particular testing the relative importance of the equivalents of Müller’s NOM, DAT and FOC constraints. In his theoretical framework, each constraint is associated with a numeric weight and the harmony of a sentence is defined as the weighted sum of the constraint violations, as in the formula in (2).

Keller assumes that differences in harmony are directly reflected in acceptability differences that can be observed empirically. When two candidate structures S₁ and S₂ differ only in that candidate S₁ violates constraint C_i whereas S₂ does not, then S₁ will be less acceptable than S₂ to an extent that is proportional to the numeric weight of C_i. The decrease in acceptability that is caused by a violation of C_i should be approximately the same in all candidate sets containing two candidates differing only with regard to C_i.

What is not defined directly in this model is the effect that the violation of a constraint has on the optimal candidate, that is, the candidate with the highest harmony. When a constraint C_i is violated in the optimal candidate, this should also lead to a decrease in acceptability, but this decrease cannot be measured because there is no better candidate that differs from the optimal one only in not violating C_i (otherwise, this candidate would have an even higher harmony and thus would be the optimal candidate). An obvious way to quantify the effect that violating C_i has on an optimal candidate would be to compare two optimal candidates that differ only with respect to C_i. In the formal statement of LOT given in Keller (2000), this is not allowed because the relationship between harmony and acceptability is limited to sentences within the same candidate set, that is, different realizations of the same content. Below, we discuss an informal proposal by Keller that allows a comparison across candidate sets under certain narrow conditions.

Keller’s account makes interesting and testable predictions. First, constraint violations are cumulative. Therefore, in contrast to OT’s notion of strict domination, multiple violations of low-weighted constraints might outweigh a single violation of a higher-weighted constraint.4 Secondly, the violation of a constraint leads to the same numeric acceptability decrease in every candidate set. In contrast to Müller’s proposal, we can thus quantify the degree of markedness directly and are not limited to a categorical ranking of candidates. These predictions are summarized in the Uniform Penalty Hypothesis given in (8). The weaker version only includes the comparison of candidates from a single candidate set, as stated in (8a), and corresponds to Keller’s formalization. The strong version, which includes both (8a) and (8b), claims that comparisons across candidate sets are possible if all constraints and their weights are known.

As noted above, the Uniform Penalty Hypothesis contrasts with standard OT, where constraint violations by the optimal candidate do not affect acceptability. These include violations common to all candidates in the same candidate set and violations of constraints ranked below the fatally violated constraint.

In Experiment 1, we test the applicability and ranking of ANI, REC, DAT, and NOM with regard to the order of arguments in the German middle field. We concentrate on ditransitive verbs and manipulate the order of the two objects, the animacy of the direct object, and the voice of the verb. Let us first consider cases where ANI and REC make the same predictions, that is, sentences containing an animate recipient and an inanimate theme. For active sentences, ANI, REC, and DAT all predict a preference for orders where the dative object precedes the accusative:

In order to tease apart the influence of ANI and REC on the one hand and DAT and NOM on the other hand, Experiment 1 includes the passive counterparts of the sentences in (9). The examples below only include the dative 1st order, the full example can be found in Table 2.

In contrast to the active sentence, DAT is not applicable here, as there is no accusative object. Furthermore, NOM now competes with ANI and REC: If the dative 1st order is still preferred, we can conclude an influence of ANI or REC and their ranking above NOM. A preference for dative 2nd, on the other hand, would be in line with a high-ranked NOM constraint as advocated by Müller (1999) and others.

In a last step, we tease apart REC and ANI by including sentences with two animate NPs:

For active sentences, where we expect a strong preference for dative 1st orders with inanimate themes, a modulation of this preference caused by the animacy manipulation would be strong evidence in favor of ANI. If the preference is completely neutralized, REC and DAT can be dismissed. The passive versions instantiate a competition of REC and NOM and will thus enable us to establish a ranking between them.

In this experiment, we aim to replicate the effects of Order and Voice on acceptability that we found in Experiment 1. Furthermore, we want to integrate DEF in the preliminary hierarchy. In the hierarchy of Müller (1999) given in (7), DEF is higher-ranked than ANI. This ranking derives from the claim that sentence (14a) is unmarked and sentence (14b) marked. That is, when DEF and ANI are in conflict, the order that respects DEF but violates ANI wins the competition.

In order to keep the complexity moderate, we only used inanimate theme sentences here and manipulated three factors. Order and Voice were parallel to Experiment 1. In contrast to Experiment 1, one of the two full NPs was indefinite. We manipulated whether definiteness aligned with animacy or not. When definiteness and animacy are aligned (i.e., the animate NP is definite and the inanimate NP indefinite), ANI and DEF make the same predictions in favor of the dative 1st order. When animacy and definiteness are not aligned (i.e., the inanimate NP is definite and the animate NP is indefinite), ANI and DEF pull in different directions: ANI predicts dative 1st to be preferred, while DEF makes the opposite prediction.

The preference in the conflicting condition will therefore allow us to establish a ranking between ANI and DEF and evaluate Müller’s proposal that DEF is higher ranked than ANI.

In Experiment 3, we turn our attention to the order between subject and direct object in transitive structures. As has often been noted, German has a strong preference for subject-object order, but this preference does not hold across the board (Lenerz 1977, and much subsequent work). For example, as confirmed by Experiments 1 and 2, passive sentences with an animate dative object and an inanimate theme subject show a preference for object-subject order. The otherwise strong subject-object preference can therefore not be attributed to a purely syntactic constraint requiring the subject to precede all other arguments. A prototypical sentence with preferred OS order has a subject that is inanimate and not an agent. In order to test whether the thematic role constraint AG or the animacy constraint ANI is the driving factor for the canonical SO order observed with standard transitive verbs, Experiment 3 contrasts AG with ANI by investigating agentive verbs that take animate as well as inanimate subjects.6 In addition, the influence of DEF is taken into account by varying the definiteness of the agentive subject. The object in all sentences is a definite animate NP. The sentences in (17) illustrate the contrasts that are tested in Experiment 3.

The experiment is organized in three subexperiments. Subexperiment 3a contrasts AG with ANI to test whether ANI is in general more influential than thematic roles or whether the highest thematic role takes precedence. This is achieved by varying the animacy of the agentive subject, (17a) vs (17b), while definiteness is kept constant. If we find a preference for SO in general, we can conclude that AG is ranked higher than ANI, while the opposite holds if OS is preferred for inanimate subjects. Subexperiment 3b contrasts AG with DEF by varying the definiteness of an animate agentive subject, (17a) vs (17c). In Subexperiment 3c, we combine both violations (ANI and DEF) to see whether violations accumulate and gang up against a presumably higher ranked AG by comparing a definite inanimate subject to an indefinite inanimate subject, i.e., (17b) vs (17d).

The preference for dative-before-nominative orders in the passive conditions of Experiments 1 and 2 motivated a relatively low ranking of NOM in the constraint hierarchy in (18). Following Vogel & Steinbach (1995), Heck (2000) argues that subject and accusative object are generated in a fixed order by OT’s generator GEN, whereas dative arguments can be rather freely attached inside the clause. In this view, we would have to split NOM into the following two constraints, where (19a) is the highest ranked constraint, whereas (19b) is relatively low ranked.

In Experiment 3, we argued that the preference for SO is rooted in the lexical-semantic constraint AG, rather than NOM, as NOM was found to be low-ranked for orders including dative objects. To tease apart the influence of AG and NOM_ACC, which were confounded in Experiment 3, we conducted an experiment with accusative object experiencer psych verbs. In addition to word order, we manipulated the definiteness of the accusative object:

Thematic roles and animacy favor OS order here, while NOM_ACC supports SO order. If SO were preferred, we would have to discard the constraint ordering ANI > NOM in favor of NOM_ACC > ANI > NOM_DAT.

We begin our review with Experiment 3, which is at odds with the strong version of the Uniform Penalty Hypothesis. Let us assume the following partial ranking of the weights: w_NOMACC > w_AG > w_ANI > w_DEF. While these weights would clearly lead to a general SO preference, as observed, weighted constraint models make an additional prediction: For inanimate or indefinite subjects, SO should score lower than for animate definite subjects, as ANI or DEF are violated. The OS orders, on the other hand, are not expected to differ in acceptability, as they violate the same constraints: NOM_ACC and AG, but not ANI or DEF. What we observed is the opposite. There is no difference in acceptability between SO sentences, regardless of whether they violate ANI, DEF, or both. Surprisingly, however, we see a difference between OS sentences: The more constraints their SO counterpart violates, the better the OS sentence is rated. This pattern is consistent with the weak version of the Uniform Penalty Hypothesis, as the difference between the optimal candidate and its competitor is smaller if the optimal candidate violates a constraint its competitor does not violate.

There is, however, an alternative explanation which also captures the observed pattern. So far, we have only considered positive weights that lead to penalties. Instead, we could argue to include negative weights that lead to a reward for a structure that satisfies the constraint non-vacuously, i.e., a sentence with arguments of different animacy or definiteness levels (see Pater 2009: 1008, for a discussion of positive and negative weights). In Experiment 3, this would apply to sentences in which OS order accommodates ANI or DEF. We reformulate the two constraints ANI and DEF in (22) such that they only apply in relevant contexts. In our previous definition, ANI was violated by structures with inanimate > animate, but not by structures with either animate > inanimate or animate > animate. The reward version should apply only to animate > inanimate orders, but not to animate > animate orders

This account is challenged by the results of Experiment 1. Here, we see a similar pattern in the conditions with an inanimate theme: the two optimal candidates score equally good, but there is a significant difference between the two dative 2nd sentences with passive sentences scoring better than active sentences. When only considering this part of the experiment, we might come to the conclusion that NOM_DAT is not an active constraint and therefore remains without penalty, while the difference between dative 2nd orders could be attributed to DAT, which is violated only in the active voice. This argumentation breaks down, however, when we consider the results from the animate theme conditions, where we found clear evidence in favor of NOM_DAT. In this case, we cannot argue for a satisfaction reward for NOM_DAT, as this would lead to a difference between the dative 1st orders for inanimate themes, because it is satisfied only by the active sentence.

The observed patterns seem to be consistent with the weak Uniform Penalty Hypothesis, but could be captured to a fair degree by a model including negative and positive weights, while the strong Uniform Penalty Hypothesis seems to fare worst. On the other hand, the strong Uniform Penalty Hypothesis is the most simple model that provides us with the most explicit predictions, that is, a predicted acceptability value for each sentence, which can be compared against the empirical acceptability measure. The model derived from the weak Uniform Penalty Hypothesis gives us numerical values for the expected acceptability differences only. In such a model, all optimal candidates could, e.g., score equally good or the differences between optimal candidates could reflect their complexity, plausibility or other factors which are often associated with performance. In comparison with Müller (1999), we still get more fine-grained predictions, as constraint violations should numerically have the same effect in different contexts, although only affecting the difference between candidates from the same candidate set. In this view, there is still a penalty associated with each constraint violation even for optimal candidates, but this penalty does not necessarily affect the absolute acceptability measure, but rather the magnitude of the lead of the optimal candidate over its competitors.

With the mixed results reported above, it seems appropriate to assess the explanatory benefit of weighted constraint models compared to the standard view. While the latter usually focuses on predicting preferences, weighted constraint models aim at predicting quantitative differences, in particular acceptability differences between suboptimal candidates. In our informal review above, we have already noted that a simple additive model with only positive constraint weights is not able to capture the data. Major adjustments as the ones outlined above are therefore necessary. In order to determine whether these changes, and therefore a numerical model, is motivated by the data, we will build regression models incorporating the different model specifications. There are three aspects of explanatory adequacy that we will assess. The first one is to check whether the qualitative pattern, i.e., the observed preferences, are captured equally well as in the OT model of Müller (1999) with the modified constraint hierarchy in (21). The second one is to check whether constraint weights remain stable across experiments – a particular well-suited test case for this is to compare the weights of ANI and DEF as derived from Experiments 1 and 2 to their weights as derived from Experiment 3. In the first case, they affect order preferences between theme and recipient in double object constructions, while in the latter case, they affect the order between subject and direct object. Lastly, we can also compare the model fit between the different models directly.

The statistical results presented in this section show that fully quantified models of acceptability face problems with regard to our data. In particular, the strong penalty model and the rewards model cannot account for all of the data and show relatively poor stability of constraint weights across experiments. The weak penalty model, which only accounts for differences within candidate sets, on the other hand, shows stable constraint weights and good model fits for all data subsets. In addition to the predictions we can derive from the OT-based markedness model with the hierarchy in (21), we can now make precise numerical predictions with regard to the difference between optimal candidate and suboptimal candidate for unseen cases. If, e.g., we use object experiencer psych verbs as in Experiment 4, but with animate subjects, we expect the benefit of SO over OS to increase by roughly .1. Also, this model accommodates the possibility of ganging-up effects which might become necessary when we extend the model to include more phenomena.

While the majority of existing accounts of argument order in German relies on individual judgments or corpus examples, our work tested and compared a set of constraints by gathering judgment data from native speakers in a principled fashion by using factorial designs and keeping the participants unaware of the manipulated factors. As expected, our results mostly match informal data from earlier accounts with one exception detailed below. Importantly, our method allowed us to disentangle constraints that are often confounded in examples and to assess their individual contribution. As a result, we can confirm the influence of the lexical-semantic constraint ANI, the syntactic constraints NOM_ACC, NOM_DAT, and DAT and the discourse constraint DEF on word order acceptability. For REC, we do not find direct evidence, whereas an influence of AG and/or EXP can only be derived by the numerical patterns, not by categorical preferences alone.

The minimal constraint hierarchy necessary to account for the results of Experiments 1–4 is given in (23), where constraints in parentheses cannot be placed conclusively with the evidence provided by our experiments.

If we take into account the numerical information, we arrive at the following constraint ranking along with their estimated weights:

Both hierarchies are incompatible with the subhierarchy of SCR-CRIT in (25) as proposed by Müller (1999).

Firstly, his hierarchy predicts dative 2nd to score better than dative 1st in passive sentences, because of the high-ranked constraint NOM. The same holds for all other hierarchies in Table 1 in which NOM is the highest-ranked constraint. Secondly, he ranks DEF above ANI, while we find evidence for the opposite ordering when looking at those sentence pairs where ANI and DEF compete directly (conditions 5 and 2 and conditions 7 and 4 in Figure 1). Contrary to the intuition expressed in example (32) of Müller (1999), given above as (14), sentences with an indefinite animate dative object preceding the definite inanimate accusative object were judged better than their counterparts in our experiment. While our results therefore do not confirm Müller’s specific constraint hierarchy, they are compatible with his model architecture, provided the splitting of NOM and the different ordering between ANI and DEF given in (23).

If we commit ourselves to weighted constraints, a striking feature of the constraint set we established is that it incorporates constraints related to thematic roles and case simultaneously, namely AG and NOM. This contrasts with the majority of the literature on word order variation in German, which shows a divide between accounts relying on thematic roles (Uszkoreit 1986; Haider & Rosengren 2003) or case (Uszkoreit 1987; Müller 1999) to identify unmarked or base-generated orders, with exceptions being Hoberg (1997) and to some extent Heck (2000).

In summary, our results show that a suitable set of surface-based constraints can capture observed word order preferences. The constraints in this set make reference to different types of linguistic information, including thematic roles, animacy, case and definiteness. Our results therefore argue against the view that the unmarked base order is determined solely by verb semantics, a view entertained by the scrambling account of Haider & Rosengren (2003). In particular, we were able to show that the effect of animacy cannot be reduced to thematic roles, as we observed significant differences even when using the same verbs with animate/inanimate direct objects.

In the first major corpus study on argument order in the German middle field, Hoberg (1981) derived the word order template in (26) for NP arguments that are neither pronouns nor part of an idiomatic expression (see also Hoberg 1997).

The template in (26) amounts to a constraint hierarchy in which ANI is ranked highest and case decides when there is a tie with respect to ANI. Bader & Häussler (2010b) applied the word order template in (26) to their corpus sample and found that for about 90% of all sentences, order was correctly predicted. The same result was obtained when order was predicted from verb semantics, which roughly corresponds to constraints in terms of thematic roles. Bader & Häussler (2010b) could therefore not decide whether word order in the German middle field is determined by animacy or by thematic roles. A possibility not considered by either Hoberg (1997) or Bader & Häussler (2010b) is the intermingling of case constraints and the animacy constraint. This, however, is exactly what our experimental results argue for. While NOM_ACC is higher ranked than ANI, ANI is in turn higher ranked than NOM_DAT.

The evidence for ranking NOM_ACC higher than ANI comes from the finding that in sentences with an accusative object, SO order is preferred even if this violates the animacy constraint ANI, that is, even if the subject is inanimate and the object animate. In Experiment 3, this was shown for sentences containing an agentive verb. The use of agentive verbs with an inanimate subject and an animate object is rare in comparison to other associations between animacy and syntactic function, but it still occurs with some regularity. A typical authentic example is provided in (27) (from the deWac corpus; Baroni et al. 2009). Note that in this example, both NPs are definite and the first NP is longer than the second, so there are no confounding factors that could explain the order “inanimate subject before animate object.”

In an ongoing corpus study that complements a series of production experiments, about 3500 sentences containing one of 24 action verbs were retrieved from the deWac corpus. 193 sentences, or about 5%, contained an inanimate subject and an animate object. In all sentences, the subject preceded the object, thus confirming that NOM_ACC is ranked above ANI. Given the rareness of examples of this kind, they probably occurred not at all or only with very low frequency in corpora of the sizes investigated by Hoberg (1997) or Bader & Häussler (2010b), resulting in a premature acception of the schema in (26). The ranking of ANI above NOM_DAT is already contained within the word order template in (26), for which – as pointed out above – a large body of corpus evidence is available.

Evidence for ranking NOM_ACC higher than ANI was also found in Experiment 4, which investigated sentences with an accusative object experiencer verb appearing together with an inanimate subject and an animate object. In this experiment, SO sentences like (28a) received higher ratings than OS sentences like (28b).

In the syntactic literature, there is no consensus concerning object-experiencer psych verbs with an accusative object (for comprehensive discussion, see Verhoeven 2015). Whereas some authors treat such verbs on a par with other verbs selecting a direct object (e.g., Vogel & Steinbach 1995; Fanselow 2000), other authors have claimed that accusative object-experiencer verbs are like dative object-experiencer verbs in that OS order is unmarked for them (e.g., Lenerz 1977; Haider & Rosengren 2003). In a sentence selection experiment, Haupt et al. (2008) found an SO preference for object-experiencer verbs with an accusative object when the subject was definite and the object indefinite whereas OS order was preferred when the object was definite and the subject indefinite. This was true independently of animacy. In a recent corpus study, Verhoeven (2015) found a moderate preference for middle field-internal OS order for accusative object-experiencer verbs when the object was animate and the subject inanimate. When both were animate, in contrast, a strong preference for SO was observed.

In the ongoing corpus study mentioned above, we are also analyzing a set of 24 object-experiencer verbs with an accusative object. As in Verhoeven (2015), all sentences containing a subject or an object pronoun were removed. For the remaining 171 examples with an animate object and an inanimate subject, both occurring within the middle field, 23% of the sentences occurred with OS order and 77% with SO order. For a set of 82 sentences in which both subject and object were animate, a strong SO preference emerged, as in the study of Verhoeven (2015). For sentences with inanimate subject and animate object, the results for the object-experiencer psych verbs clearly differ from the results for the agentive verbs, for which not a single OS sentence was found.

For agentive verbs and for object-experiencer verbs with two animate arguments, the corpus findings of Verhoeven and our corpus findings are almost identical. For object-experiencer psych verbs with an animate object and an inanimate subject, in contrast, they differ. Verhoeven found a moderate preference for OS order whereas we found a moderate preference for SO order. We do not know what accounts for this difference. One reason could be that psych verbs in general are more variable with regard to order than other verbs. For psych verbs, an OS preference turned up in Verhoeven’s study only when both arguments were contained in the middle field. When one of the arguments was contained within the prefield, SO order was preferred. For agentive verbs, in contrast, a strong preference for SO order was observed independently of where the arguments were located. In addition, in a production experiment involving accusative object-experiencer psych verbs, Temme & Verhoeven (2016) found a strong context effect on order. Given that the number of sentences with inanimate subject and animate object was neither large in Verhoeven (2015) nor in our study, the preceding discourse context may have had particularly strong effects. More research is needed in order to resolve this issue.