1.1 The normal argument order of German experiencer-object verbs

Both subject-object (SO) and object-subject (OS) serialisations are possible in German. The choice, however, is not arbitrary: there is a normal order, and deviations from this order may affect interpretation (Höhle 2019/1982). Hence, both sentences in (1) are grammatical, but (1a) is contextually relatively unrestricted, while (1b) is restricted to contexts in which Emma is focussed.¹

Closer scrutiny into verb classes other than transitive action verbs reveals that normal order appears to be predicate-(class-)dependent. In addition, verb-extrinsic properties such as the animacy of the arguments (Müller 1999; Fanselow 2003) are often considered relevant. Hence, argument serialisation appears to be driven by predicate(-class)-specific properties, predicate-extrinsic properties of the individual elements ordered (such as animacy, or weight), and predicate-extrinsic properties affecting the meaning of the whole clause (such as focus, or scope).

A particularly problematic class in this respect are experiencer-object (EO) verbs. The normal order of their arguments has been discussed controversially in theoretical and experimental linguistics. EO verbs are a subclass of psych verbs, verbs that can be characterised by an entailment about the mental state of an experiencer,² namely those that realise their experiencer argument as their object. The subject of EO verbs³ generally refers to the semantic stimulus of the psychological state. Following the analysis of Italian psych verbs by Belletti & Rizzi (1988), the class of EO verbs is subdivided into accusative EO verbs with an accusative object as in (2) and dative EO verbs (3). Experiencer-subject (ES) verbs form the third class in their subclassification (e.g. English fear). It is widely held that ES verbs are semantically stative and syntactically transitive, and dative EO verbs stative and unaccusative (their object is taken to c-command their subject at some point), while the syntactic and aspectual nature of accusative EO verbs is debated (Rozwadowska & Nowak & Bondaruk 2020; cf. Belletti & Rizzi 1988; Pesetsky 1995; Arad 1998; Reinhart 2002; Landau 2010 among many others).

There is a near-consensus that the normal order with dative EO verbs in German is OS (see Lenerz 1977; Fanselow 1992; Wegener 1999; Haider & Rosengren 2003; Hirsch 2018 among others).⁴ For accusative EO verbs, however, linearisation properties are debated. Lenerz (1977), Haider & Rosengren (2003), and Primus (2004) exclude agentive readings of some accusative EO verbs from consideration (as these are analysed to prefer the same linearisation as prototypical action verbs, SO) and do not propose differences between accusative and dative EO verbs on the non-agentive reading. From Fanselow’s (1992) analysis it seems to follow that accusative EO verbs have an SO normal order, while he later (2003: 204 seq.) proposes that both orders are normal for accusative EO verbs (although he argues for OS with interessieren ‘to interest’ (p. 203)). Hirsch (2018) adopts this view for a subclass of accusative EO verbs, namely those he considers stative, and assumes an SO normal order for the others. Primus (2004) assumes a free order for non-causal EO verbs (both accusative and dative).

Previous experimental studies point towards OS for dative verbs and SO for accusative verbs. Most studies agree that accusative EO verbs tend towards SO (although to a lesser extent than action verbs, or verbs with experiencer-subjects) (Scheepers & Hemforth & Konieczny 2000; Temme & Verhoeven 2016; Verhoeven & Temme 2017; Ellsiepen & Bader 2018). While Temme & Verhoeven (2016) conclude that dative EO verbs tend towards OS, Fanselow & Häussler & Weskott (2016) find that OS is preferred with dative EO verbs whose perfect tense auxiliary is sein ‘be’, but both orders are acceptable with those whose perfect tense auxiliary is haben ‘have’. The animacy of the subject is regarded as a decisive factor with animate subjects pushing towards SO (Scheepers et al. 2000). In Verhoeven’s (2015) corpus study the OS rate was above 50% in the midfield for all kinds of EO verbs, and around 80% for dative EO verbs, if the subject was inanimate, but it dropped to less than 10% if the subject was animate with accusative EO verbs (dative around 50%). In a corpus study presented by Ellsiepen & Bader (2018: 30), accusative EO verbs occur primarily with an SO order.

Three remarkable aspects must be noted about all previous experimental studies: First, there is no experimental study in which accusative and dative EO verbs are compared directly in the same experiment. Given that experimental data are not independent from one another both in terms of the participants and the items presented in a study, this is a major methodological drawback. Secondly, an experimental separation of accusative and dative EO verbs hard-codes the case-based classification instead of experimentally testing its viability. This is exacerbated by the recent observation that verbs assumed to belong to the same class show ample heterogeneity, which raises doubts about taking the case-based classification for granted (Hirsch 2018; Poppek et al. 2021).⁵ The classes, of course, must be represented in the statistical analysis of the data if hypotheses are formulated with the classes in mind. An awareness of potential differences between class members, however, should lead to a choice of verbs that facilitates subsequent explorative analyses searching for reflexes of these differences. Thirdly, participants in previous experimental studies (except for Scheepers et al. 2000 and Ellsiepen & Bader 2018) were presented verb-second clauses, where one argument was placed in the prefield. Constituents in the prefield may introduce confounding effects, as they may be used to signal contrastive focus (Frey 2006). Hence, using verb-second clauses is methodologically problematic.

We may thus conclude that there is no consensus regarding the normal order of EO verbs in theoretical linguistics, and that previous studies in experimental linguistics suffer from not comparing the case-based classes directly, from taking them for granted in the first place, and from using structures in the studies that may yield confounding effects.

To address these issues, we have conducted two Two-Alternative-Forced-Choice experiments that target the linearisation preferences of German EO verbs with inanimate (study A) and animate (study B) subjects. Both studies show that a small number of verbs deviates from the general pattern of their case-based class. Study B shows that certain EO verbs with a dative object do not prefer a linearisation at all if both subject and object are animate. This is hard to reconcile with accounts of German constituent order that assume that the normal order is a direct reflection of a single base order (for the distinction between normal order and base order, see section 1.2). Building on the experimental results, we will present an account that uses free base generation in combination with violable linearisation constraints to explain the observed patterns. The account does not rely on the case of the verbs’ object, but on a semantic distinction between different kinds of stimulus arguments introduced to the linguistic literature by Pesetsky (1995), namely causers and objects of emotion (see section 4.2).

1.2 Normal order vs. base order

Research on word order variability in German is abundant and cannot be reviewed with due appreciation here (see e.g. Abels 2015; Frey 2015; Salzmann 2023 for overviews). Most importantly, however, the concept normal order must be distinguished from the concept base order. While the normal order is the order that is contextually least restricted (Höhle 2019/1982), a base order emerges in theoretical works if arguments must be combined in a specific order with the verbal projection.

Most researchers assume a specific order of combination in general, for verb classes, for predicates, or possibly for predicates and constructions, as can be witnessed in Frey (1993), Müller (1999), Haider & Rosengren (2003), among many others. In these approaches (except for Müller’s), normal order is taken to reflect a base order, which is the simplest configurational structure derived from the fixed order of combinations. Deviations from this order are taken to be the result of scrambling, i.e. movement to the left. It is important to note that with such an approach, there can only be one base order (in general, per predicate, class, or whatever subdivision is taken to hold), and hence only one normal order. The base order for transitive verbs may differ from verb/predicate/class/… to verb/predicate/class/…, but either the object or the subject is bound to merge first; and if they are found in a different order, at least one element must have moved. If normal orders reflect the base order (in a setting where there are no relevant differences between subject and object), there should thus only be one normal order (for a set of sentences differing only regarding serialisation).⁶

The approach by Fanselow (2001; 2003) reflects a minority view, according to which a head can realise its arguments in (almost) any order. Hence, there is no fixed base order. Fanselow presents both conceptual and empirical arguments against the stipulation of a base order. He raises doubts on the idea that free word order requires movement, while fixed word order comes for free and provides an analysis where fixed word order follows from movement.

Another term closely linked to normal order is unmarked order. In one sense, a sentence is marked if it is contextually more restricted than the other sentence(s) from its comparison set (Höhle 2019/1982, see also ibd. for the intricate difference between normal and unmarked). In another sense, markedness is equated to suboptimality with regard to a constraint hierarchy (Müller 1999). A long research tradition (e.g. Lötscher 1981; Uszkoreit 1987; Hoberg 1997; Keller 2000; Ellsiepen & Bader 2018) tries to model serialisation in the midfield using constraints encoding some of the factors taken to influence constituent order in German from a descriptive perspective, such as animacy, definiteness or case. Ellsiepen & Bader (2018) provide an overview of this literature and they report on a series of experiments conducted to establish a constraint ranking, or a set of weights. They arrive at the hierarchy in (4):⁷

They particularly assume the dominance of a constraint that places nominatives before accusatives. The decisive experiment involves accusative EO verbs and is taken to show that case is more important than thematic roles. In view of our experiments and the semantic distinction between causers and objects of emotion introduced by Pesetsky (1995), we argue that the postulation of case-based constraints is not necessary to capture the linearisation preferences of EO verbs.

For our experiments, it is essential that all factors not exclusively depending on the predicate be controlled for as far as possible, one of the most important ones being focus. According to Höhle (2019), a sentence that allows a focus projection from one constituent to maximal focus has normal order. If we can ensure a maximal focus interpretation, the variant with normal order should thus be preferred.

1.3 The roadmap

The analysis presented in section 4 will follow Fanselow (2001) in assuming that arguments of a verb can be discharged in any order, thus yielding free word order in principle. We also follow most research on German syntax in assuming a strictly binary-branching midfield. Furthermore, we concur with Haider (2010) that there are no functional projections in German clause structure between V and C. However, we assume that the structures are subject to violable and potentially conflicting, weighted linear precedence (LP) rules, and we assume competition between alternative structures.

We will see that the distinction between accusative and dative EO verbs hides a semantic distinction, namely the one between causers and objects of emotion (as introduced by Pesetsky 1995). We propose three LP rules, as illustrated in (5).

The analysis rests on the results of two studies on the preferred argument order of EO verbs considering both accusative and dative EO verbs in single studies. An important result of study B is that a subset of EO verbs allows more than one normal order. This is in line with the idea of regulative LP rules but cannot easily be analysed in terms of scrambling as movement based on a fixed base order if this fixed base order is equated with the normal order.

The experimental studies will be presented in section 2, section 3 discusses their theoretical implications, and section 4 presents the analysis.

2.1 Test environment

The target sentences were constructed as verb-final sentences embedded in matrix clauses to avoid prefield effects. According to Höhle (2019/1982), a sentence that allows a focus projection from one constituent to maximal focus has normal order. To ensure a maximal focus interpretation, each pair of sentences was displayed with an introductory question, as illustrated in example (6). Participants were asked to choose the answer that they perceive as more natural.

Each participant saw and judged all items in a pseudo-randomised order subject to some constraints, e.g. no subsequent test items without at least one filler item in between. Both variants of the sentence were presented simultaneously, the horizontal alignment of choices (SO or OS right or left) was pseudo-randomised. A screenshot of an item from study B can be found in the OSF directory.

Since factors like animacy, constituent weight and definiteness are known to influence linear order in German, they had to be controlled for. We only used indefinite subjects and objects, singular NPs containing only the noun and the indefinite article and aimed for differences in length of maximally one syllable (for the whole NP).

In study A, subjects are inanimate, and objects animate, while in study B both are animate. Experiencer objects, of course, must be animate – at least in a wider sense.⁹ All object NPs (in study B also the subjects) in the test items were masculine because feminine and neuter NPs would have led to ambiguities or processing difficulties due to case syncretism.

Test items were constructed based on corpus examples following the procedure of Modified Stimulus Composition (Börner & Pieper & Kiss 2019), both from the GerEO database (Masloch et al. 2021; Poppek et al. 2022) for sentences containing an EO verb and the DWDS corpora (Berlin-Brandenburgische Akademie der Wissenschaften n.d.) for some of the filler items and the sentences containing an action verb in study B.

The verbs used in both studies were carefully vetted for two reasons: to avoid interference effects and to represent potential subclasses. Corpus findings show that accusative EO verbs differ with respect to the syntactic patterns they partake in (Poppek et al. 2021) and it has been proposed that EO verbs belong to several different verb classes (Hirsch 2018). Verbs from various potential subclasses were included. We selected the EO verbs for both studies based on their behaviour in annotated corpus data. For this, we used GerEO, which contains syntactic and semantic annotations for ca. 10,000 sentences (up to 200 per verb) containing one of 64 German verbs – 16 selecting the dative, 48 the accusative – that occur as an EO verb (Masloch et al. 2021; Poppek et al. 2022). Verbs that possess a frequent reading in the database in which the verb does not refer to a mental state were avoided. In both studies, the verbs co-occur with an experiencer object and a stimulus subject, but some verbs are frequently used in other syntactic patterns, e.g., a reflexive one, where the experiencer is the subject and the stimulus is expressed in a PP (if expressed at all). We computed the share of occurrences in the transitive and an object-drop pattern (where the experiencer is not expressed overtly and receives an arbitrary interpretation) and used only verbs that regularly displayed these patterns. Other exclusion criteria include frequent usage within a collocation or idiomatic expression and the overall frequency of the verb, see the documents describing the selection process in the article’s data directory (see section Data availability/Supplementary Files). Regarding the stimulus argument, both studies required different criteria: Since the agentive readings, which many accusative EO verbs possess in combination with animate subjects, are taken to have SO normal order (even by authors who argue for OS with accusative EO verbs otherwise), we included only verbs in study B that had a decent share of animate subjects in GerEO and scored low on the agentivity test rating studies by Verhoeven (2014) and Hirsch (2018). Agentivity is not an issue in study A, because stimuli are inanimate there. All other criteria being fulfilled, we aimed for the inclusion of verbs with different morphological structures as well as verbs that belong to different classes according to Hirsch (2018) or that displayed different distributional properties in corpus data (Poppek et al. 2021). Full lists of the verbs used in the experiments are provided in Table 1 and Table 2.

2.2 Implementation

Participants were recruited via the survey tool Prolific (www.prolific.com). We conducted Monte Carlo power simulations that estimated power at > 90% (α = 0.05) for relevant effects and the intended statistical models and expected parameter values based on 25 participants. Since some of the parameters of the models are hard to estimate (e.g. by-participants random effect correlations), we opted for a larger number of participants.

To identify non-cooperative or distracted participants as well as participants who provided their answers implausibly fast, we implemented attention and control items and used the response times for a latency-based ReMFOD analysis (see Pieper & Börner & Kiss to appear).

We also asked participants to guess the topic of the study and excluded (the few) participants who guessed correctly or displayed substantial linguistic background knowledge.

2.3 Study A: inanimate stimuli

Study A was designed to determine the normal constituent order with accusative and dative EO verbs, an inanimate stimulus and an animate experiencer, using a total of 16 EO verbs (see Table 1).

2.3.3 Results

40 native speakers of German participated in the study, but 11 surveys were excluded following the exclusion criteria. The remaining 29 participants judged 16 test items each, resulting in a total of 464 observations. Figure 1 shows the empirical distribution of choices for both conditions. We see that OS is preferred with dative EO verbs (177 OS, 55 SO), and SO is preferred with accusative EO verbs (165 SO, 67 OS).

Figure 1

Empirical distribution of choices in study A.

We analyse the data using a binomial (logit) generalised linear mixed model (GLMM) using the R environment (R Core Team 2023) and the package lme4 (Bates et al. 2015). This means that (as in all models we use) the influence of the factors manipulated on the independent variable order is modelled via a linear predictor incorporating the factors manipulated treatment-coded (one level is coded as 0, the other one as 1) to which the logistic function logit^–1 is applied to yield the probability of a specific order. Here, we only have a fixed effect of case (and an intercept), and all variables are treatment-coded, the reference levels being SO and dative. Hence:

$P\left(order=OS\right)=logi{t}^{-1}\left({\beta }_0+case\times {\beta }_{case}\right)$ (eq. 1)

However, equation 1 only covers the fixed effects of the model. In addition, we include random intercepts for participants and items as well as a random slope for case for participants (a random slope for case for items would not be meaningful since case is manipulated between items). This means that we assume individual effects for items and participants stemming from normal distributions with the fixed effect as their mean. The width of the distribution is a parameter estimated in the model, as is the correlation between the participants’ random effects. The intercept β₀ captures the situation in the dative condition (where case = 0) here: If it were 0, the distribution of choices predicted by the model for this condition would be 50% SO, 50% OS. The fixed effect of case (β_case), on the other hand, represents the change to the accusative condition. If it were 0, the distribution expected by the model in this condition would be the same as the one in the dative condition. An accessible introduction to GLMMs in linguistics can be found in (Winter 2020).

We observe significant effects for the intercept ( $\widehat{\beta }$ = 1.34, 95% confidence interval (CI): [0.77, 1.9], p < 0.001¹¹) as well as for case ( $\widehat{\beta }$ = −2.41, 95% CI: [−3.18, −1.63], p < 0.001). The positive effect of the intercept indicates that for dative EO verbs the preferred order is OS (logit^–1 (1.34) = 79.25% probability of OS). The negative effect of case implies that dative and accusative EO verbs differ, the latter having a stronger tendency towards SO (logit^–1 (1.34 – 2.41) = 25.54% probability of OS). Both effects exceed the expectations.

Regarding the random effects, we find relatively high standard deviations for the items’ ( $\widehat{SD}$ = 0.52) and the participants’ ( $\widehat{SD}$ = 0.69) random intercepts as well as for the participants’ random slope for case ( $\widehat{SD}$ = 0.9). The negative correlation of −0.56 for the participants’ random effects indicates that higher intercepts come with lower slopes, i.e., the stronger a participants’ comparative tendency towards OS with dative EO verbs, the stronger their comparative tendency towards SO with accusative EO verbs.

Since we used treatment coding in the model, we also build a model with accusative as the reference level. While all other parameters of this model are similar to the model with dative as the reference level, it displays a significant negative intercept ( $\widehat{\beta }$ = −1.07, 95% CI: [−1.63, −0.51], p < 0.001), indicating that the accusative EO verbs’ tendency towards SO is also significant.

2.3.4 Discussion

We take these results to mean that accusative and dative EO verbs differ in their linearisation preferences and that the normal order (with inanimate subjects and all other factors equalled out) is OS with dative EO verbs and SO with accusative EO verbs. Since the items had to contain an unresolvable asymmetry in animacy, the observed SO tendency for accusative EO verbs is worth noting. However, we observed relatively high standard deviations for the random effects. A short explorative investigation of the items’ random effects in Figure 2 illustrates the frequencies of SO choices for the individual items grouped by case, purple for items containing an accusative EO verb (t01–t08) and yellow for dative EO verbs (t09–t16). While 29 judgments per verb are a small sample and individual properties of the items besides the verb might have added to the results, we note that some verbs appear not to comply with the general pattern of their classes.

Figure 2

Number of SO choices for individual items in study A.

While most verbs display a rather clear preference in this setting, interessieren ‘to interest’ shows a very mild tendency towards OS, which fits previous accounts that argue for OS normal order for this verb (e.g., Haider & Rosengren 2003). Yet, its behaviour is generally untypical for an accusative EO verb (the same holds for nerven ‘to annoy’). Among the dative EO verbs, behagen ‘to please’ stands out, which closely resembles interessieren in its distribution. However, the tendencies of both verbs towards any order are so weak that they might point either towards no actual preferences on the participants’ side or to idiolectal variation.

2.4 Study B: animate stimuli

Study B was conducted to shed light on animacy effects in study A. Since animacy differences were inherent to the design of study A, study B was designed to contain only animate subjects and objects. This has the disadvantage that some verbs may display an agentive reading with animate subjects – which is standardly assumed to behave like an action verb –, but it has the additional advantage that it allows us to directly compare EO verbs to action verbs as a contrasting verb class with generally non-controversial linearisation behaviour.

2.4.3 Results

33 native speakers of German participated in the study. However, 8 surveys had to be excluded, so that data from 25 participants remained. Each participant provided a judgment for 32 test items, resulting in 800 data points. Figure 3 depicts the empirical distribution of choices for the four conditions.

Figure 3

Empirical distribution of choices in study B.

We fitted a binomial (logit) generalised linear mixed model to the data, including case, agentivity and their interaction as fixed effects, random intercepts for items and participants and random slopes for case, agentivity and their interaction for participants. The model formula for the fixed effects part is:

$P\left(order=OS\right)=logi{t}^{-1}\left({\beta }_0+case\times {\beta }_{case}+ag\times {\beta }_{ag}+case\times ag\times {\beta }_{case\times ag}\right)$ (eq. 2)

All factor variables were treatment-coded with SO, dative, and EO as reference levels. The intercept of the model therefore reflects the dative EO condition, while case and agentivity are not the main effects of case and agentivity but the effects of changing from the reference level to the other one along that dimension. So, we can obtain the model’s expected proportions of choices for the accusative EO condition by adding the effect of case to the intercept (and applying the inverse logit to the result), and we will have to add the intercept, both effects and the interaction (β_case×ag) to get the expectations for the accusative agentive condition. The random effects capture the individual variability around the fixed effects.

The observed effects generally match the expected ones: There is a significant strong positive effect of case ( $\widehat{\beta }$ = 2.28, 95% CI: [1.17, 3.39], p < 0.001¹³) and a significant strong positive effect of agentivity ( $\widehat{\beta }$ = 4.61, 95% CI: [2.09, 7.13], p < 0.001), which – given our setting – shows that dative EO verbs differ strongly both from accusative EO verbs as well as from dative action verbs. The interaction effect ( $\widehat{\beta }$ = −2.58, 95% CI: [−5.68, 0.53], p = 0.1) is negative (as expected) and not significant. The estimate could be taken to counter the effect of case, in the sense that the difference between the dative EO condition and the accusative action condition roughly equals the difference between the dative EO condition and the dative action condition, but the effect is not significant and at this end of the scale, a change of −2.58 in the linear predictor translates only into a very small change in predicted probabilities (ca. 1%).¹⁴ Surprisingly, the intercept is smaller than expected and not significant ( $\widehat{\beta }$ = 0.12, p = 0.74). One could take this to indicate that there is no preferred order with dative EO verbs (note that the null hypothesis here would correspond to a 50/50 distribution of choices between the two orders), but we will discuss differences between the verbs below.

To see whether the difference between accusative EO and action verbs is also significant, we fitted an additional model that differed from the original one only in having accusative as the reference level of case. Indeed, the effect of agentivity is significant then ( $\widehat{\beta }$ = −2.03, 95% CI: [−3.82, −0.24], p < 0.05). The intercept of this model is also rather large, negative and significant ( $\widehat{\beta }$ = −2.4, 95% CI: [−3.31, −1.50], p < 0.001), which means that accusative EO verbs tend towards SO.¹⁵ We will switch back to the first model in the following.

Concerning the random effects, participants do not differ that much in their judgments ( $\widehat{SD}$ = 0.46). Also, there is not much variation in how participants change their behaviour from dative to accusative EO verbs ( $\widehat{SD}$ = 0.28). However, the variance of the items’ random intercepts is more notable ( $\widehat{SD}$ = 0.89): Items differ in which variant is preferred even if all other factors are considered (recall that the SD of ca. 0.9 is on the same scale as the fixed effects). We will explore these differences below and see that they are likely linked to the verbs. The variances of the participants’ random slopes for agentivity ( $\widehat{SD}$ = 1.75) and the interaction ( $\widehat{SD}$ = 1.65) are huge, and they have a considerable negative correlation (−0.78), mainly caused by five participants with high values for the interaction and huge negative slopes for agentivity, who also chose OS for some dative action verbs.

2.4.4 Discussion

As the effects of case and agentivity illustrate, accusative EO verbs and dative action verbs differ significantly from dative EO verbs. This cannot be attributed to agentive readings of the accusative EO verbs since those used in this study scored low on agentivity tests with animate subjects. Given the observed proportion of choices, we can assume an SO normal order for action verbs. The same holds for accusative EO verbs, although their tendency towards SO is weaker (but see below).

The results for dative EO verbs, however, are unexpected. The model even predicts a slight preference for SO (53%) for them. This is surprising since dative EO verbs are almost universally taken to have OS normal order (see section 1.1). Another interesting aspect is the high variation of the items’ random intercepts. Items must differ within the conditions. With action verbs SO is virtually always chosen, so the observed high variance must thus be rooted in the judgments for EO verbs.

Figure 4

Number of SO choices per item for experiencer-object verbs in study B.

An explorative view on the individual verbs sheds light on the matter. Figure 4 shows the number of SO choices per item for EO verbs with dative EO verbs on top (IDs t17–t24) and accusative EO verbs below (t01–t08). Clearly, the verb appears to be the decisive factor since differences between the individual items for each verb are small. We observe differences within the case-based classes: Among the dative EO verbs, there is no clear linearisation preference for leidtun ‘to feel sorry’ and gefallen ‘to appeal to’, while imponieren ‘to impress’ leans towards SO and auffallen ‘to strike’ towards OS. Auffallen is the only verb in this study that uses sein ‘to be’ instead of haben ‘to have’ as its perfect auxiliary. Imponieren seems to have an agentive reading: We are not aware of a discussion of agentive readings with dative EO verbs in the literature, but we find examples like (9) in GerEO. (9) contains an impersonal lassen-middle. Since agentivity of the implicit argument is a core feature of middles (see Pitteroff 2014: 43 sqq. and the literature cited there), imponieren must have an agentive reading (and it is clearly interpreted this way here). Arguably, (10) shows that imponieren must also have a change-of-state reading independent of animacy since it contains imponieren in the adjectival/stative passive and it has been argued that an adjectival passive is only possible in German with verbs that have such a reading (Gehrke 2015). However, the acceptability of examples like (10) is subject to idiolectal variation and they are rejected by many (probably most) speakers.¹⁶ DeReKo (Kupietz et al. 2010), also contains occasional examples of imponieren in the verbal passive on its agentive reading (11).

1. (9)

1. NZZ_1996_02_23_a187_seg4_s4 in GerEO
1. […]
2.  
1. Mit
2. with
1. ihnen
2. them
1. lässt
2. lets
1. sich
2. refl
1. trefflich
2. splendidly
1. imponieren.
2. impress
1. ‘With them, it’s easy to impress’

1. (10)

1. NZZ_1994_05_25_a98_seg7_s13 in GerEO
1. […]
2.  
1. dass
2. that
1. Vertreter
2. representatives
1. der
2. the.gen
1. Europäischen
2. European.gen
1. Union […]
2. Union.gen
1. von
2. by
1. der
2. the
1. Wahlfreiheit
2. freedom.of.choice
1. imponiert
2. impressed
1. gewesen
2. been
1. seien.
2. were.sbjv
1. ‘that representatives of the European Union were impressed by the freedom of choice’

1. (11)

1. (U07/JUN.04350 Süddeutsche Zeitung, 26.06.2007)
1. Wem
2. who
1. soll
2. shall
1. wohl
2. mp
1. damit
2. with.this
1. imponiert
2. impressed
1. werden?
2. become
1. ‘Who is supposed to be impressed by this?’

Among the accusative EO verbs, interessieren ‘to interest’, displays a much weaker tendency towards SO than the other accusative EO verbs (this was already the case in study A).

The question arises whether mixed judgments for certain verbs are caused by variation between participants or by the same participants not preferring either option. Figure 5 suggests that the latter is the case: Very often, participants chose the SO variant only once. For gefallen ‘to appeal to’ and leidtun ‘to feel sorry’, the choices are almost exactly what is expected if participants are forced to choose between two alternatives they like equally well. However, we will also assume some degree of individual variation, particularly for imponieren ‘to impress’ and interessieren ‘to interest’ (see section 4.3). Naturally, these considerations are speculative to some degree since each participant only saw two sentences for each verb. However, the conclusion remains that gefallen ‘to appeal to’ and leidtun ‘to feel sorry’ appear not to enforce a specific linearisation in this setting.

Figure 5

Number of SO choices of each participant for noticeable verbs in study B (with all other verbs, SO is almost always chosen).

4.1 Syntactic framework

We follow Haider (2010) in assuming that there are no functional projections between C and the maximal verbal projection in German, that this verbal projection is binary branching with the verb on the right and that the subject is contained in the maximal verbal projection. Furthermore, we assume that arguments and adjuncts may combine with the verb in any order, but that their ordering is subject to violable and interacting constraints. Therefore, we need a model of constraint interaction. For concreteness, we will make use of Maximum Entropy Grammar (Goldwater & Johnson 2003; Hayes 2022), a probabilistic variant of Optimality Theory (Prince & Smolensky 2004) assuming weighted constraints (however, any model providing the means for constraint cumulation is possible in principle). Using a probabilistic model has the advantage that we can account for the observation that participants do not choose one variant or both in equal proportions, as expected if there were a set of winners (containing usually only one element). As the grammatical backbone providing the structures to be compared, any framework within the limits mentioned above is applicable given some level on which surface order constraints can be checked, e.g. a list containing all non-head daughters within a projection (as in Kiss & Pieper & Börner in preparation).¹⁹ Having access to all features of the relevant elements, linearisation constraints can apply to this list.²⁰ (13) contains a partial analysis of a fragment of one of the test sentences (repeated in (12)). We only indicate animacy here, but all elements will have many other (possibly irrelevant) features. NPs have case, and φ-features, but we will also assume that the verbal projections carry features for argument structure and semantics. They may hence impose certain features on their arguments (via feature sharing or some other mechanism).

Ordering constraints will take the form animate ≺ inanimate, meaning that animate elements should precede inanimate ones. We assume that syntactic and semantic structures are built up in parallel, so that we do not have to replicate semantic information in the syntax.

In Maximum Entropy Grammar, the probability of a candidate y given a context/input x is as in equation 3 (Goldwater & Johnson 2003: 114):

$\begin{array}{l}P\left(y|x\right)=\frac{1}{Z\left(x\right)}\text{exp}\left({\displaystyle \sum }_{i=1}^m{w}_i{f}_i\left(y,x\right)\right), \text{where}\\ Z(x)={\displaystyle \sum }_{y \in Y(x)}\exp \left({\displaystyle \sum }_{i=1}^m{w}_i{f}_i\left(y,x\right)\right)\end{array}$ (eq. 3)

This is to be understood as follows (see Hayes 2022 for a conceptual motivation): There is a set of weighted constraints C = {c₁,…,c_m}. We can think of Y as the generator function, which represents the inviolable part of the grammar. It takes some input x, the nature of which won’t concern us here, and returns all candidates (in our case, these will be the two alternatives compared in the experiments). f_i(y,x) is candidate y’s number of violations of c_i in context x, either 0 or 1 in all cases we will consider here. w_i is the weight of c_i. We sum over the weighted constraint violations and rescale the sum by exponentiating it. Dividing this value by the sum of the corresponding values for all candidates mainly serves two purposes: 1. Comparing candidates. 2. Pressing the real-valued weighted-constraint-violation-profile sums into the interval [0,1], as required for a probability.

4.2 Causation

Before we introduce the pertinent constraints, we take a step back to consider the nature of emotions. Although it is subject to ongoing debate in philosophy and psychology (see e.g. Deigh 2009; Scarantino & de Sousa 2021 for overviews), an important tradition views “emotions as intentional states of mind, that is, states of mind that are directed at or toward some object” (Deigh 2009: 17). That being said, the verbs called “psych verbs” do not necessarily denote these states: E.g. Marín & McNally (2011) argue that a certain class of Spanish reflexive psych verbs denotes only the beginning of a state, not the state itself. Pesetsky (1995) points to a distinction among the arguments of psych verbs one may naively describe as stimuli or themes. Some of them are the objects of emotion (which can further be subdivided into targets of emotion and subject matters), while others are mere causers. He shows that they can be distinguished by truth conditional differences:

The PPs in (14a) and (15a) indicate objects of emotion while the subjects in (14b) and (15b) are causers of the emotion. Pesetsky points out that example (14b) is compatible with a situation in which Bill considers the article itself great but is angry at e.g. the government, whose corruption the article exposes. This is not possible in (14a): The object of the anger must be the article itself. In (15b), the television set need not be the subject matter of John’s worries but only needs to cause them: If John is a detective, a television set owned by a purportedly blind man can make John worry about completely different things. In (15a), however, the television set must be the subject matter of John’s worries.

Since we will propose below that causers push towards the beginning of the clause, the causer/object-of-emotion distinction is crucial. As shown above, causer subjects with EO verbs can be identified by finding a context in which a pure causer reading is possible, i.e., the subject is not the object of emotion. E.g., in (16a) my fear need not be directed towards the news broadcast, it is fully compatible with a situation in which I watched the news and am afraid of nuclear war afterwards. This is in stark contrast to (16b): For (16b) to be true, I must have assessed the qualities of the news broadcast itself and found it appealing (see also Fanselow 1992: 292).²¹

In the appendix, we list a sentence and a context for each verb used in our studies in which a pure causer reading is salient, if possible, and we discuss some crucial cases where it is not possible. In general, there is a strong correlation between object case and nature of the subject such that the accusative EO verbs used in our experiments have causer subjects and the dative EO verbs object of emotion objects. This comes as no surprise since causation plays a major role in most major accounts of psych verbs (see the overview in Rozwadowska 2017) and it is often assumed that accusative EO verbs are causative. E.g., their causativity has been taken to explain their linking behaviour in comparison to ES verbs (see i.a. Grimshaw 1990; Reinhart 2002). The assumption that German accusative, but not dative EO verbs are causative is widespread in the literature (see e.g. Rothmayr 2009; Marelj 2013; Hirsch 2018). However, we were unable to find or to come up with an example in which the subject of (accusative EO) interessieren ‘to interest’ is not an object of emotion,²² and the subject of (dative EO) imponieren ‘to impress’ seems to license a pure causer interpretation in (17). In our judgment, (17) allows a line of argumentation parallel to the one for (15b).

However, there may be some degree of (idiolectal) variation for these judgments: As discussed in section 2.4.4, imponieren ‘to impress’ may enter the adjectival/stative passive, indicating a change-of-state reading (which would fit well with a causer subject, although causation is possible without a change of state, see footnote 26), but many native speakers reject the adjectival passive with imponieren ‘to impress’.

4.3 Constraints

We assume the following constraints to be among the linear precedence rules of German. None of them is new or should be too controversial (see e.g. Lötscher 1981; Frey 2015).

From the results presented in section 2, it follows that Act must be weighted strongly (accounting for the behaviour of action verbs in study B), while Caus and Anim must be ranked lower and in this order of importance. Using Maximum Entropy Grammar, however, we do not have to resort to stipulation. Instead, we can learn the weights almost directly from the data.

The formula for the Maximum Entropy Grammar in equation 3 corresponds to a logistic regression model. In order to fit it using common software, one can subtract the constraint violation profile of the OS variant from the one of the SO variant (see appendix). If there are only two candidates x and y, the probability of x is (we drop the input to the generator function here since all alternatives are listed):

$\begin{array}{l}P\left(x\right)=\frac{exp\left({{\displaystyle \sum }}_{i=1}^m{w}_i{f}_i\left(x\right)\right)}{exp\left({{\displaystyle \sum }}_{i=1}^m{w}_i{f}_i\left(y\right)\right)+exp\left({{\displaystyle \sum }}_{i=1}^m{w}_i{f}_i\left(x\right)\right)}\\ =logi{t}^{-1}\left({\displaystyle \sum }_{i=1}^m{w}_i\left(f_i\left(x\right)-f_i\left(y\right)\right)\right)\end{array}$ (eq. 4)

We can then encode our experimental data in an appropriate way and feed it to logistic regression models to obtain the constraint weights. However, we must discuss which constraints apply to which items first. We take the accusative-object action verbs used in study B (anrufen ‘to call sb.’, informieren ‘to inform’, verhaften ‘to arrest’) except for betrügen ‘to deceive’, but not the dative-object action verbs (applaudieren ‘to applaud’, antworten ‘to reply’, danken ‘to thank’, zujubeln ‘to cheer’) to be causative (see appendix).²⁵ Of course, the subject is an actor for all of them. With EO verbs, all accusative EO verbs except for interessieren ‘to interest’ are taken to be causative,²⁶ all dative EO verbs except for imponieren ‘to impress’ not (but see below) following the discussion in section 4.2. The only EO verb used in our studies that we take to have a reading on which the subject is an actor is imponieren ‘to impress’ (see the discussion in section 2.4.4.). To explain the OS preference with auffallen ‘to strike’, we will need an additional constraint. Since auffallen is the only verb in study B selecting sein ‘to be’ as its perfect tense auxiliary, we will assume (19) here (which also affects nahegehen ‘to affect deeply’ from study A).²⁷

A summary of the constraint violation profiles assumed is contained in the appendix. As discussed above, we re-encode the constraint violation profile by subtracting the vector for the OS variant from the one for the SO variant. In a simple model, the linear predictor (i.e. the part within logit^–1 () in eq. 4) will then look as follows:

${\beta }_{act}\times act+{\beta }_{caus}\times caus+{\beta }_{anim}\times anim+{\beta }_{\mathrm{sein}}\times \mathrm{seina}$ (eq. 5)

However, we will complicate this slightly by adding varying intercepts for items and varying slopes for participants. The former means that each item will be allowed to have its own tendency towards SO. The inclusion is helpful because there may be properties of individual items that we failed to control for properly. The latter means that each participant will receive their own constraint weights. Once linearisation constraints are assumed to be weighted, it seems plausible that speakers of a language may differ in the weights they assign to them. By adding the random slopes, it becomes possible to assess the degree of variability of each constraint.²⁸ However, since each participant saw at most two items for which Sein is relevant, we do not include a random slope for Sein. We assume that the random slopes for the other constraints are uncorrelated.²⁹ We call this model M_base. We fitted M_base to our experimental data (code available via the OSF directory). However, we did not include the data for all items containing imponieren ‘to impress’ or interessieren ‘to interest’ in the model. We discuss several models where they are included in the constraint-weight-computation script available on OSF, but at this point we exclude them for the following reason: As discussed in section 2, we have to assume individual variation for these verbs. Specifically, we suspect that speakers differ in whether they consider their subjects causers or not. Additionally, imponieren ‘to impress’ has an agentive reading (as discussed in section 2.4.4), but also a non-agentive one. There could also be variation between speakers in whether they assume the agentive or the non-agentive reading (which will differ in their constraint violation profiles) more often. Thus, whatever constraint violation profile we assume, it won’t be appropriate for all speakers. There is no way to feed this information to the model. The random intercept for items can in principle absorb this bias to a certain degree, but it will still have an impact on the constraint weights we are primarily interested in here, so we exclude the data for these two verbs. M_base offers outstanding discrimination on its training data (Somers’ D = 0.82, concordance index = 0.91). Its constraint weights are presented in (20).

The higher the absolute value, the more dominant the constraint. It will be helpful to see how the values in (20) capture the data (see also Table 2 in the appendix): In study A, there was an asymmetry in animacy such that objects were animate and subjects inanimate. The only constraint applicable to the items with dative EO verbs (minus the sein-selecting ones and imponieren) is Anim, which explains why they prefer OS: Anim is violated with the SO variant, but not the OS variant, thus f_anim (SO) – f_anim (OS) = 1. It is the only constraint relevant, so the predicted P(order = SO) = logit^–1 (0 × –3.6 + 0×–2.4 + 1 ×–1.1 + 0 × –0.9) ≈ 0.25. Accusative EO verbs, however, have a causer subject. SO is preferred for these verbs even though it violates Anim since only OS violates Caus, which is dominant (logit^–1 (–1×–2.4 + 1 × –1.1) ≈ 0.79). In study B, where animacy does not interfere, no constraint pushes an argument in any direction with gefallen ‘to appeal to’ and leidtun ‘to feel sorry’, so participants should not have a preference (logit^–1 (0) = 0.5). With auffallen ‘to strike’, Sein leads to a preference for OS (logit^–1 (–0.9) ≈ 0.29). For accusative EO verbs, only Caus (violated by OS) is relevant (logit^–1 (2.4) ≈ 0.92). With action verbs, Act (and with the accusative-object ones also Caus) leads to the strong preference for SO (logit^–1 (3.6) ≈ 0.97, logit^–1 (3.6 + 2.4) ≈ 0.998).

The standard deviation of the random intercepts for items, which is on the same scale as the constraint weights in (20), is estimated at 0.22. So, our constraints seem to capture the factors at play in the test items well (except for imponieren ‘to impress’ and interessieren ‘to interest’ of course). The estimates for the standard deviations of the random slopes for participants are 0.65 (Anim), 0.24 (Caus), 1 (Act), so it seems that participants differ quite substantially in their constraint weightings, especially for Act (although one has to keep in mind that differences at the extremes of the linear predictor translate into lower differences on the probability scale than comparable differences around 0).

Note that this analysis does not need to postulate case-based or grammatical-role related constraints (except for Sein; see below on ES verbs). Indeed, while it would be generally compatible with a (low weighted) constraint nominative ≺ accusative, it is incompatible with constraints that demand nominatives to precede datives because such a constraint would destroy the equilibrium with dative EO verbs and animate subjects (unless counterbalanced by another constraint). A model M_int that differs from M_base only in having an intercept (whose inverse corresponds to a constraint subject ≺ object since it would be violated in all OS but no SO examples, so f_SO (SO) – f_SO (OS) would always be −1) has a worse AIC (Akaike information criterion) and BIC (Bayesian Information Criterion) than M_base (so the latter is the preferable model). Its estimate for the intercept is close to 0 and not significant. Thus, we assume that there is no constraint like subject ≺ object.

We will now look at the two misbehaving verbs: In study B with the animate subjects, the SO variant of the two examples with imponieren ‘to impress’ was chosen by 20 and 22 of the 25 participants respectively (80/88%), in study A with the inanimate subjects, it was 9 out of 29 (31%). In section 4.2 we argued that imponieren has a causer subject, in section 2.4.4 that it has an agentive reading. The data from study A is only explainable if we assume that for most participants imponieren does not have a causer subject (predicted probability of SO for verbs with a causer subject in this study: logit^–1 (2.4 – 1.1) ≈ 0.79). In study B (with different participants), however, imponieren clearly differed from the other dative EO verbs. This could also be due to the agentive reading. However, we cannot extrapolate from the experiments which participants assumed an agentive reading in which example, and since an agentive interpretation should lead to a strong preference for SO in the given setting, the source of the SO preference cannot be identified with certainty. With interessieren ‘to interest’ an agentive reading is not possible. In section 4.2, we suggested that it does not have a causer subject. In study B, the SO variant was chosen 19 and 16 times (out of 25, 76/64%), in study A it was chosen 13 times (out of 29, 45%). If it did not have a causer subject, we would expect an SO rate of 50% in study B and of ca. 25% in study A, so we suspect that it may have a causer subject for some participants (but not for all since there is still a striking difference to the values predicted for causative EO verbs mentioned above).

Having dealt with the linearisation preferences of EO verbs, we will now briefly comment on the behaviour of experiencer-subject verbs like mögen ‘to like’: Their preferred argument linearisation is universally taken to be SO. Yet, there is also a (near-)consensus that they are not causative. How can this preference come about if we do not want to assume case- or grammatical-role-based LP rules? While we cannot dive further into this topic here, we note that ES verbs are often taken to be individual-level predicates (see e.g. Pylkkänen 2000 on Finnish; Fábregas & Marín 2015 on Spanish). If such an analysis is feasible for German, too, as recently argued for by Sommer (2023), one could argue that the subject of an ES verb is thus a very likely aboutness topic (see Krifka 2008 on this notion), and topics are usually assumed to precede comments in linear order (see e.g. Siewierska 1993 among many others).³⁰

An obvious question is why almost all accusative EO verbs fall into one class and almost all dative EO verbs into the other. Since it would require a discussion of many grammatical properties of EO verbs whose empirical basis is far from clear, we cannot provide linking constraints here. Nevertheless, we may speculate that the constraints governing linking are mostly similar or equal to some of the ones governing linearisation, one difference being that linking constraints apply to lexical items while linearisation constraints apply to structures that are built on the fly. This would explain why idiosyncrasies are possible with the former, but not the latter. Accusative case on the object of interessieren ‘to interest’ could be such an idiosyncrasy, at least in the grammar of some speakers. The reason for the close resemblance of these constraints might be that they reflect general cognitive preferences. If Haider (2020) is correct in assuming that scrambling comes about as the utilisation of a syntactic potential by other grammatical subsystems, we may assume that among the factors determining the ordering are also general cognitive preferences, some of which are also at work in linking/argument selection (Dowty 1991). Since (by assumption) the demands of the other subsystems, which make use of the syntactic potential, may stand in conflict to each other, one will need some system of managing conflicting constraints anyway. The difference between a fixed-base-order-plus-movement and a free-base-generation account is then at which point broadly speaking thematic features become relevant: “before” syntax, when a base order is determined, within syntax via functional heads that syntactify thematic notions, or later as an important criterion when choosing between available structures? While our data cannot decide the issue, it seems to be more naturally explained by a free-base-generation approach (see section 3). A further point in case comes from the behaviour of comitative adverbials, which are not listed on argument structure representations. As shown by Kiss et al. (in preparation), they adhere to Act if (and only if) they link their internal argument to an actor (an affirmative comitative takes over the role of a c-commanding antecedent and is subject to the same constraints then), so that postulating linearisation constraints on a higher level is necessary on independent grounds.