1.1.1 Co-reference

Based on comprehensive corpus and empirical data, Bosch et al. (2003; see also Bosch et al. 2007) suggested that DPros avoid referents of DPs as antecedents that function as the grammatical subject of the immediately preceding sentence (see also Kaiser 2011a). In contrast, PPros do not show such an anti-subject bias and, on the contrary, even display a general preference for subject referents as antecedents. Similar contrasts have been observed for other languages with personal pronouns and demonstrative pronouns, such as Finnish, Dutch, and Romance languages, which have null pronouns in addition to overt pronouns (see Kaiser & Trueswell 2008; Mayol & Clark 2010; Kaiser 2010; 2011b; 2013).

However, judging from data such as the ones presented in (4), Bosch and Umbach (2006) argued that DPros do not necessarily avoid referents of grammatical subjects, but rather discourse topics. That is, the DPro in (4a) can only refer to the referent of the grammatical subject of the preceding sentence, Peter, and not the referent of the indirect object, Karl. According to Bosch and Umbach (2006), this is due to the fact that Karl has been established as the discourse topic: The first sentence poses a question about Karl and the two subsequent sentences address this question. With respect to PPros, Bosch and Umbach (2006) propose that these pronouns have a (weak) preference for discourse topics, which is attested by the observation that the PPro in (4b) is typically interpreted as referring to Karl, although it could also, at least in principle, refer to Peter.

Now, Bosch and Umbach (2006) can account for contrasts that have been taken as evidence for subject avoidance in previous research by assuming that grammatical subjects are, by default, discourse topics. In other words, what seems to be (strong) subject avoidance in the case of DPros and (weak) subject preference in the case of PPros might in fact be an artifact of the observation that grammatical subjects are often also the discourse topic (but see Schumacher et al. 2017).

Extending Bosch et al.’s work, Schumacher et al. (2016) conducted three experiments comparing the German PPro er with the German DPro der. In Experiment 1, a forced choice antecedent selection task, sentences either contained an accusative verb (to rescue) or a dative experiencer verb (to attract attention). Furthermore, the two potential antecedents of the pronoun appeared in canonical (SVO) or non-canonical word order (OVS). For accusative verbs, Schumacher and colleagues found a bias for PPros to refer to the nominative marked agent antecedent and a bias for DPros to refer to the accusative marked patient antecedent, and this data pattern was independent of word order. For dative experiencer verbs and for canonical word order, PPros were mostly interpreted to refer to the dative marked agent/experiencer antecedent and DPros to refer to the nominative marked patient antecedent. For non-canonical word order, no interpretation bias was found for PPros or DPros. The authors concluded that multiple prominence cues negotiate during pronoun interpretation, with thematic role presumably being a stronger predictor than grammatical role (see also Schumacher et al. 2015 for evidence from event-related potentials).

1.1.2 Binding: Wiltschko (1998)

While the co-referential use of DPros has been targeted in theoretical and empirical investigations, there has been relatively little systematic research on the binding behavior of DPros, in particular in the empirical domain. Wiltschko (1998) proposed that DPros are referential expressions that correspond to definite DPs consisting of an overtly realized determiner and an empty NP introducing a free predicate variable. Consequently, these pronouns cannot be bound by c-commanding DPs without violating Principle C of Binding Theory (Chomsky 1986) and not be interpreted as bound variables either. The idea is that DPros can only pick up contextually given individuals or be interpreted as donkey pronouns (see Wiltschko 1998 for details), where donkey pronouns are pronouns that behave as if they were bound by indefinite DPs that do not have scope over them due to their containment in a syntactic island or different sentence (see Geach 1962; Heim 1990; Elbourne 2005 and the references therein). For PPros, in contrast, Wiltschko (1998) assumes that they lack a DP-layer entirely and are just projections of agreement features. As such they introduce variables which can either receive a value from the context or be bound by a c-commanding DP. Evidence for Wiltschko’s claim comes from sentences such as those in (5a–d) (reproduced from (1) above) and (5e–h), for which the use of DPros is infelicitous (at least in the absence of a contextually salient male individual, which the DPro could refer to).

Note that while Wiltschko does not discuss the contrasts between DPros and PPros in their co-referential behavior, her analysis could easily be extended in a way that it can account for these contrasts as well. One could, for example, assume that being assigned as value to a free individual variable is a process that is preferably applied to contextually maximally prominent individuals. The more indirect process of making a DPro co-referential with an individual via assigning a suitable value to the predicate variable, in contrast, could be assumed to preferably apply to less prominent individuals.

1.1.3 Binding: Hinterwimmer (2015)

Hinterwimmer (2015) questions the core arguments presented by Wiltschko (1998) and argues, on the basis of examples like those in (6a–b), that DPros can in principle be bound by DPs c-commanding them either on the surface or at LF, i.e. after Quantifier Raising has applied.

Hinterwimmer (2015) proposes that both DPros and PPros are DPs that consist of a(n) (phonetically differently realized) overt determiner and an empty NP introducing a free predicate variable, which follows Elbourne’s (2005) analysis of PPros. Setting aside the case of donkey pronouns, which are irrelevant for the discussions in the present paper, the free predicate variable can be resolved in two ways: First, to the property of being identical to a contextually provided individual, which results in a co-referential interpretation. Second, to the property of being identical to (the value of) a variable that is bound by a c-commanding DP. Note that when the binding DP is a referential DP such as a proper name or definite description, the resulting interpretation is indistinguishable from a co-referential interpretation. Nevertheless, following Reinhart (1983; 2006) and Heim (1998), Hinterwimmer (2015) assumes that whenever binding is possible, it is preferred over co-valuation when the two resulting interpretations are equivalent, where co-valuation is a process in which two terms are assigned the same value accidentally, i.e. without being co-indexed.

The critical argument for the assumption that there is a real difference between bound and co-referential interpretations even in cases where the putative binder is a referential DP comes from the following observation (see Reinhart 1983; Heim 1998 and Reinhart 2006 for detailed discussion and additional references): A sentence such as (7), in both the version with VP-ellipsis and the version without VP-ellipsis, allows for a strict as well as a sloppy reading. On the sloppy reading, Paul is said to adore his own (i.e. Paul’s) cat, while on the strict reading Paul is said to adore Mike’s cat. The strict reading comes about via co-valuation: The pronoun receives the contextually salient individual Mike as value. The sloppy reading, in contrast, comes about via binding: The variable introduced by the pronoun is co-indexed with and thus bound by a lambda-operator, turning the VP adore(s) his cat into a predicate. That predicate is then applied to the individual denoted by the subject DP, Mike. In cases such as (7), where co-valuation makes available a reading that could not be achieved via binding, co-valuation is licit. In cases where binding is an option and where there is no difference between a bound and a co-referential interpretation, binding is assumed to be preferred over co-valuation.

When the DP on whose interpretation the pronoun depends is a quantifier (e.g., (6a)), binding is the only option to begin with. As in cases in which the binder is a referential DP, it comes about via a lambda-operator binding the variable introduced by the pronoun. Because of the higher semantic type, however, the resulting predicate is not applied to the binding DP. Rather, it is the other way round: The quantifier is applied to the predicate, resulting in a bound variable interpretation of the pronoun.

On Hinterwimmer’s (2015) account, neither PPros nor DPros are expected to give rise to Principle C effects, unlike DPs with an overt NP (cf. Schlenker 2005). The crucial difference between DPros and PPros is rather that DPros, by virtue of being the marked pronoun variant in German, signal to the reader/hearer that the default process of identifying a potential antecedent or binder does not apply.

Technically, this is implemented in the following way: DPros, in contrast to PPros, come with a lexical presupposition which prevents the predicate variable introduced by the empty NP from being resolved in a way that makes it dependent on the (currently) most prominent DP. The idea is that the default process of identifying a potential antecedent or binder would pick out the most prominent one. Importantly, the lexical presupposition that prevents this process is the “frozen” effect of a general principle: Formal markedness corresponds to a non-default interpretation. Crucially, what counts as the (currently) most prominent DP differs in binding and non-binding configurations. Binding configurations are defined in structural terms: A (potential) binder has to be contained in the same sentence as its bindee, and the former needs to c-command the latter either on the surface or, at the latest, at LF (see discussions in Barker 2012 for evidence that binding does not require surface c-command, but only linear precedence in combination with scope, contra Reinhart 1983; Heim 1998; Reinhart 2006). We would therefore expect that prominence is defined in structural terms as well. Non-binding configurations, in contrast, are cases in which the pronoun is neither c-commanded on the surface nor at LF by the DP on which its interpretation depends. That is because that DP is either contained in a separate sentence or in a syntactic island from which it cannot be moved out at LF.

Now, grammatical subjects are the syntactically most prominent DPs within their sentences (for example, in terms of operations such as controlling implicit arguments or reflexive binding being readily available to them, but not to other DPs). Consequently, the lexical presupposition of DPros precludes the predicate variable from being resolved to the following predicate: the property of being identical to (the value of) a variable bound by the respective grammatical subject. In that vein, the contrast between the DPro and PPro variants in (5) is simply due to DPros avoiding grammatical subjects as binders. In (6a) and (6b), in contrast, an interpretation on which the DPro is bound by the respective prepositional or indirect object is unproblematic. That is, the free predicate variable can be resolved to the property of being identical to the value of a variable bound by the respective DP.

Turning to non-binding configurations, we first note that prominence in these contexts is defined in terms of discourse properties, with discourse topics (Bosch & Umbach 2006; or agents, see Schumacher 2016) being maximally prominent (but see, e.g.: Grosz, Joshi & Weinstein 1995 for the claim that structural factors such as subjecthood play an important role in the interpretation of cross-sentential pronouns). In the absence of a potential binder, the lexical presupposition of DPros thus precludes them from referring to antecedents that function as discourse topics. In other words, the free predicate variable may not be resolved to the property of being identical with the individual functioning as the current discourse topic.

Concerning the co-referential behavior of DPros, the analysis of Hinterwimmer (2015) thus makes exactly the same predictions as the analysis provided by Bosch and Umbach (2006). Although Hinterwimmer (2015) does not provide an explicit discussion of languages other than German, his reasoning generalizes to any language that has marked and unmarked pronouns, as long as prominence in terms of topicality and subjecthood makes clear whether to use the unmarked pronoun. Consequently, the null hypothesis would be that marked pronouns in other languages come with the same presupposition as the one assumed for DPros in German, as this presupposition is assumed to be the “frozen” effect of the general principle mentioned above: Formal markedness corresponds to a non-default interpretation.

1.1.4. Binding: Patel-Grosz and Grosz (2017)

Patel-Grosz and Grosz (to appear) agree with Hinterwimmer (2015) that DPros can in principle be bound. Their analysis, however, differs from Hinterwimmer (2015) in that the authors do not assume DPros to come with a lexical presupposition which prevents them from being bound by/co-referring to the (currently) most prominent DP. Instead, Patel-Grosz and Grosz assume that the contrast between PPros and DPros emerges as the artifact that DPros are structurally more complex than PPros. That is, both types of pronouns are analyzed as DPs with an overt determiner and a covert NP (as in Hinterwimmer 2015). However, DPros, unlike PPros, contain an additional functional layer above the D-layer and it is this additional layer that is ultimately responsible for the different phonetic realizations.¹ Patel-Grosz and Grosz (2017) follow Schlenker (2005) in assuming that the pragmatic economy principle Minimize restrictors! is in effect. This principle basically precludes items with more (complex) structure to be used (a) whenever there is a structurally less complex alternative available and (b) when using the less complex alternative does not lead to a difference in interpretation. Whenever there is a pragmatic benefit, though, using the more complex variant is felicitous. Consequently, DPros should only be used when there is some benefit that could not be achieved by using a PPro. Since Minimize restrictors! is a general pragmatic principle, it should hold for the marked pronoun variants in other languages as well, as Patel-Grosz and Grosz also state in their paper.

2.1.2 Results

Overall accuracy for comprehension questions was 94%. Mean reading times and standard errors for each word region (words 1–5) are provided in Table 1. Reading times and confidence intervals are also plotted in Figure 1. Effect sizes of comparisons were calculated using Cohen’s d and are provided in Table 1. Prior to statistical analysis, reading times faster than 200 ms and slower than 2000 ms were excluded from further analysis (0.1% of the data). All other reading times were log-transformed individually for each word region (words 1–5) using Box-Cox power transformations.

Figure 1

Mean reading times in milliseconds and 95% confidence intervals of word regions in Experiment 1. Demonstrative = demonstrative pronoun; Personal = personal pronoun; Male object = female subject referent, male object referent (male object condition); Male subject = male subject referent, female object referent (male subject condition).

Linear mixed effects regression models were used for each word region individually to test for differences between male object and male subject conditions for the two pronoun types. While reading times were included as dependent measure, pronoun type (DPro or PPro) and grammatical role (male object or male subject), and their interaction, were included as independent variables. We also included random intercepts and random slopes for pronoun type and grammatical role, both for participants and items. We started with the most complex random effects structure and reduced the complexity of the slopes by successively reducing the random effects structure until the model converged. For words 1, 4, and 5, models converged with the most complex random effects structure. For word 2, the final model only included main effects and no interaction, while, for word 3, only the interaction in the by-items slopes was removed.

Grammatical role and pronoun type were centered prior to analysis and p-values calculated on the assumption that our models’ intercepts are normally distributed. For each word region, we fitted an individual model.

For first, fourth, and fifth words of our regions of interest, we failed to find any reliable reading time differences between conditions, ts < 0.9, ps > 3. However, for the second word region, the male object condition was read faster than the male subject condition, but only when the sentence included a DPro, leading to a Grammatical role × Pronoun type interaction β = 12.75e-04, SE = 5.57e-04, t = 2.29, p = .022. The same was true for the third word region, β = 5.63e-04, SE = 2.85e-04, t = 1.97, p = .049. To ensure that the observed reading time differences generalized to both prepositional and non-prepositional object versions of test sentences, we reran the models for the second and third word, while including object type (prepositional object or non-prepositional object) as a co-variant. Log-likelihood ratio testing revealed no improvement of model fit for the second, χ²(1) = 0.96, p = .327, or the third word, χ²(1) = 0.03, p = .862.

2.1.3 Discussion

The reading time data of Experiment 1 show that, first, the male object condition with a DPro was read faster than the male subject condition with a DPro and, second, that the male object condition with a DPro was read as fast as the male object condition with a PPro. These data speak against the claim that DPros can generally not be bound. If this were the case, the male object variants of sentences should have led to longer reading times in pronoun regions when the sentence included a DPro than when it included a PPro. This is not what we observed. A more straightforward explanation of the observed data pattern is provided by Hinterwimmer (2015): DPros avoid grammatical subjects as binders (see also Bosch & Umbach 2006; Kaiser 2010).

We should point out here that one might argue that participants adopted a relative clause interpretation for the male object DPro condition, as dessen matches in form with the genitive form of the German relative pronoun. If this were correct, the male subject DPro sentences were read more slowly because readers were surprised with a male relative pronoun used for a female marked head noun. We think that this alternative interpretation of the data is highly unlikely, though. First, in 12 out of 20 experimental sentences the pronoun was embedded in a prepositional phrase, thereby blocking a relative clause reading. Second, relative clauses require the use of a comma, which was never present in our items. Third, if readers had adopted a relative clause reading, we should have observed a strong reading slow-down for the male object DPro condition at the point the relative clause reading turned out to be false. We should then have observed a reading slow-down for male object DPros in the later spillover region, which is not what we found (cf. Figure 1).

2.2.2 Results

Accuracy of comprehension questions was 97%. Reading times were elicited and analyzed individually for five regions of interest, comparing the male object with the male subject condition. Mean reading times and standard errors for each word region (word 1–5) are presented in Table 2 and plotted in Figure 2. Effect sizes of comparisons, which are provided in Table 2, were again calculated using Cohen’s d. Prior to statistical analyses, reading times faster than 200 ms and slower than 2000 ms were identified as outliers and therefore excluded (1% of the data). The remaining reading times were log-transformed for each word of interest individually using Box-Cox power transformations.

Figure 2

Mean reading times in milliseconds and 95% confidence intervals of word regions in Experiment 2. Male object = female subject referent, male object referent (male object condition); Male subject = male subject referent, female object referent (male subject condition).

We again fitted mixed effects models for each word of interest individually, using the same fitting procedure as described for Experiment 1. The independent variable was grammatical role (male object or male subject), and was centered. Random intercepts and random slopes were included for participants and items. Final models for all words only included the main effect of grammatical role in the by-participants and by-items random slopes.

Results for Experiment 2 replicated those of Experiment 1. We failed to find statistically reliable reading time differences between male object and male subject condition for the first, second, and fifth word, ts < 1.5, ps > .2. However, we found a reliable difference for the third, β = 2.71e-03, SE = 1.28e-03, t = 2.11, p = .035, and a marginal difference for the fourth word of interest, β = 7.69e-05, SE = 4.05e-05, t = 1.90, p = .058. To investigate whether the observed data pattern generalized to both prepositional and non-prepositional object versions, we again compared the final regression models with models that additionally included object type (prepositional vs. non-prepositional) as predictor using log-likelihood ratio tests. No improvement of model fit obtained for word three, χ²(1) = 0.63, p = .478, or word four, χ²(1) = 1.27, p = .260.

2.2.3 Discussion

The close resemblance of the reading time differences between the DPro conditions in Experiment 1 and Experiment 2 are difficult to entertain within a framework that generally disallows DPros from receiving a bound variable interpretation (Wiltschko 1998). However, our data are fully in line with the proposal that DPros can be interpreted as bound variables as long as the (potential) binder is not the grammatical subject of the respective sentence (Hinterwimmer 2015).

Before we continue, there is an important point to address: Reading times were overall longer in Experiment 2 than Experiment 1. Considering that only the test materials of Experiment 2 unambiguously forced bound readings, one might argue that, relative to PPros, DPros are still marked variants, at least when they trigger bound readings. In other words, although the data from Experiment 2 provide evidence that, in binding configurations, DPros can more easily be bound by a preceding object than a preceding subject quantifier, they might still be poorer candidates, in both binding contexts, than their PPro counterparts. To test for this possibility, we conducted an offline rating study including DPros and PPros in quantification contexts.

If the longer reading times observed in Experiment 2 were due to the fact that bound readings of DPros are generally less acceptable than their (potentially) co-referential interpretations, the test sentences of Experiment 2, which included DPros, should be less acceptable than when these sentences include a PPro, which is arguably not marked. If, on the other hand, the longer reading times in Experiment 2 were simply due to the fact that we recruited overall slower readers in the second than the first experiment, the experimental sentences of Experiment 2 should be rated as similarly acceptable when they involve object binding irrespective of whether they include a DPro or a PPro.

2.3.2 Results

Sentences with male objects were rated as overall more acceptable than sentences with male subjects, in both the DPro and the PPro conditions. More interestingly, though, sentences with a DPro and a male subject were rated as much less acceptable than their PPro counterparts. Sentences with a DPro and a male subject received a mean acceptability score of 4.23 (SE = 0.10). Corresponding sentences with a PPro received a mean acceptability score of 4.88 (SE = 0.09). Sentences with a masculine marked object received a mean acceptability score of 5.51 (SE = 0.08) when the sentence contained a DPro, and a mean acceptability score of 5.38 (SE = 0.09) when it contained a PPro.

To test for statistical reliability, we fitted a linear mixed effects model with grammatical role (male object or male subject) and pronoun type (DPro or PPro) as well as their interaction as predictors and acceptability scores as dependent variable. The model included random intercepts and random slopes for participants and items. Random slopes converged with the Grammatical role × Pronoun type interaction. The output of the model reveals a significant Grammatical role × Pronoun type interaction, β = –0.78, SE = 0.20, t = –3.84, p < .001. This statistically confirms the observation that there was a much larger difference in acceptability between DPro and PPro versions when the pronoun was bound by the subject quantifier than when it was bound by the object quantifier.

Results of Experiment 3, then, are a nice extension of the reading time data of Experiment 1 and suggest that reading times in Experiment 2 were not overall longer because bound versions of DPros were generally less acceptable than potentially co-referential versions, which were tested in Experiment 1. Indeed, we elicited very similar acceptability scores for DPros and PPros when the pronoun could be bound by the object quantifier, which is not compatible with the view that bound readings of DPros are generally dispreferred. Instead, DPros only seem to be dispreferred when the only available binder is a subject DP, which is fully compatible with our reading time data from Experiments 1 and 2.

Before turning to the General Discussion, we would like to point out a potential shortcoming of the data tested in Experiments 2 and 3.² As shown by Roberts (1989) and on the basis of examples like the one in (16), quantificational DPs headed by every and each allow telescoping. That is, under specific conditions, quantificational DPs seem to bind pronouns that they clearly do not c-command, neither at the surface nor at LF. In contrast to this, downward entailing quantifiers do not allow telescoping.

Because we only used sentences with quantificational DPs headed by jeder ‘every’ in Experiments 2 and 3, we cannot rule out the possibility that the object quantifiers in our test materials do not enter into standard binding configurations with the DPros, and that telescoping might therefore be involved. In addition, the variant of the DPro version of our test sentence in (15a) given in (17a), in which the universally quantified DP is replaced by a negative DP, is judged as slightly degraded by some informants and does not sound entirely natural to us either. Note that the variants in (17b) and (17c), where the respective quantificational DP does not have an overt NP as its complement, seem completely unacceptable. In contrast, the variants with the PPros in (18) all seem felicitous.

While we do not have an ultimate explanation for the unacceptability of (17b) and (17c) and the intuition that (17a) is somewhat degraded, albeit not unacceptable, we consider it unlikely that telescoping can explain these contrasts. First, the contrast in acceptability between (17a), on the one hand, and (17b) and (17c), on the other, is not expected if telescoping were involved: Replacing the universally quantified DP in (16) by a negative DP, as shown in (19), does not only make telescoping more difficult; it makes telescoping completely unavailable.

Second, whatever the correct account of telescoping is, it seems to be a rather complicated process that is only available under very specific and still poorly understood discourse conditions involving factors such as narrative continuity (see the brief discussion in Roberts 1989). It does not seem to be an all-purpose repair mechanism that applies whenever standard variable binding fails. Finally, we are not aware that telescoping has ever been proposed for cases in which standard binding is possible, i.e. cases in which the quantifier c-commands the pronoun on the surface already or precedes it on the surface and can easily take scope over it at LF (which, recall, are the conditions which we, following Barker 2012, consider to be decisive for binding configurations).

Of course, the discussed contrasts between (17), on the one hand, and the sentences in (18), on the other, need to be accounted for. For now, we do not have a conclusive explanation. Intuitively, what seems to be relevant is the availability of a specific set of individuals as the domain of quantification for a quantifier that is to bind a DPro. This is further suggested by the observation that (20), which also contains a negative quantifier, sounds entirely natural:

Needless to say, that the above-discussed contrasts do not follow from anything said in Hinterwimmer (2015) (or any of the other discussions on DPros). However, with the data at hand, we do not see the putative existence of the discussed restriction as a reason to reject the idea that DPros can be bound by quantificational DPs c-commanding them on the surface or at LF at the latest.