Parasitic gap patterns and hierarchy preservation in German

1 Introduction

This paper focuses on parasitic gap (PG) data from German that involve scrambling of two non-pronominal internal argument DPs. The core configurations are schematized in (1c, d), but by way of introduction, consider first (1a, b), with scrambling of just one DP. In (1a, b), a single internal argument – an accusative direct object (DO) in (1a) (from Heck and Himmelreich 2017: 53) or a dative indirect object (IO) in (1b) (from Assmann 2010: 114) – is the antecedent of a gap (the PG) contained within a vP-level adjunct that it has scrambled past. Such adjuncts are typically islands for extraction, but nevertheless the scrambled DP can serve as the antecedent for the gap inside the adjunct. I follow the common analysis that such a gap inside an island crucially is licensed by (i.e. is parasitic on) A′-movement (here, scrambling) of its antecedent from a position below the adjunct to a position above it (cf., for example, Chomsky 1982; Engdahl 1983). Indeed, this line of analysis has been taken in numerous works that discuss comparable German examples (see, e.g., Felix 1985; Mahajan 1990; but for alternative views of scrambling and PGs in German, see, e.g., Haider and Rosengren 2003; Kathol 2001).^[1] The core empirical concern of this paper, now, is which DP can function as the PG’s antecedent when both an accusative DO and dative IO scramble past the adjunct. Given the different word orders possible from scrambling, a simple representation of this can be as in (1c) or (1d).

As we have seen in (1a, b), in the case of one DP scrambling, either an accusative or a dative can be the antecedent of the PG, and there does not appear to be variation in this regard for the two varieties of German that I will presently introduce.^[2] Nevertheless, when it comes to which of the two scrambling DPs can license a PG in (1c, d), a puzzling asymmetry emerges when we carefully examine what has been reported in the literature (actual German examples given in Section 2). In recent work, Heck and Himmelreich (2017) provide data showing how only the accusative DP in (1c, d) can be the PG antecedent. Strikingly, German data reported in earlier literature (e.g. Müller 1995) reveal a different pattern. In this other dataset, it is the DP (either dative or accusative) that is linearly closer to the PG in (1c, d) that can serve as the PG’s antecedent.

The systematic nature of these contrasting data patterns, when there are multiple scrambled DPs, has not been noticed before.^[3] Such a systematic contrast can be taken as evidence that there is grammatical variation across German speakers. For descriptive purposes I will use the terms Accusative-antecedent Grammar (AG) to refer the grammar of the first pattern (i.e. that of Heck and Himmelreich) with only an accusative antecedent allowed for (1c, d), and Linear Order Grammar (LOG) for the second pattern (i.e. that of Müller), where only a linearly closer antecedent is allowed for (1c, d).^[4]

This now raises the question of whether a unified analysis of this variation can be given. The fact that accusative case and not word order is critical for the AG data suggests that a structure-based analysis is relevant. This is indeed the approach of Heck and Himmelreich, who give an analysis of their dataset that crucially relies on the two DPs scrambling to the vP edge in a way that preserves their hierarchical relation with each other prior to movement. For Heck and Himmelreich (2017: 62), such hierarchy-preserving movement is implemented via a movement stack, which is a buffer/temporary storage area in the workspace where constituents that are attracted by a particular feature and copied from the tree are placed before they are re-merged, and which – when more than one constituent is attracted – provides an order according to which these constituents are re-merged in order to satisfy that feature (see details in Section 3, and cf. Stroik 2009 for more on buffers). However, little analysis has been proposed for the older LOG data from the literature, and a unified analysis seemingly faces a challenge with what appears to be a sensitivity to linear order, a sensitivity that appears to be at odds with Heck and Himmelreich’s analysis that crucially relies on a movement stack and the hierarchy related to it.^[5] Indeed, the seeming relevance of linear order runs counter to the prevailing view that it is considerations of structural hierarchy that determine PG licensing.

In response to this challenge, I propose that Heck and Himmelreich’s proposal can indeed form the basis of a unified analysis of the data with multiple scrambling dependencies, but with two key modifications (see (a) and (b) below).^[6] Thus, hierarchy and a movement stack (and not linear order) underlie resolving PG dependencies in both grammars. And although they do not suggest it at first, the LOG facts also provide further evidence (see e.g. McGinnis 1998; and Richards 2001) that multiple movement dependencies attracted by the same type of feature on a head (here: v) can be hierarchy preserving. Instead, the loci of grammatical variation concern (a) the availability of scrambling positions other than the vP edge, and (b) which phrase markers constitute domains for the linearization of syntactic structure. In short, the LOG contains a key position that can trigger scrambling within vP (by hypothesis in ApplP), but which cannot trigger scrambling within the AG. Further, the LOG contains a linearization domain within vP that effectively fixes the relative hierarchy of the scrambling DPs (as a result of fixing their relative word order), as per Fox and Pesetsky’s (2005a) proposal for cyclic linearization. In contrast, the AG just has a larger linearization domain that properly contains vP. These differences between the grammars are outlined schematically in (2), where numbered positions indicate positions available for scrambling, and arcs indicate linearization domains.

(2)

a.	Accusative-antecedent Grammar (AG)
b.	Linear Order Grammar (LOG)

A consequence of this proposal is to highlight two different ways that hierarchy preservation can emerge in grammar for constituents involved in multiple movement dependencies of the same type, even when (as we will see) this hierarchy can also sometimes be disrupted in these same dependencies. The first arises locally when multiple constituents are attracted by a head via the same movement stack, and involves maintaining a pre-movement hierarchical relation (cf. Richards 2001). This is what we see in both the AG and LOG when multiple DPs scramble to the vP edge. The second is the result of linearizing a chunk of structure (i.e. a linearization domain), which has the effect here of globally fixing the hierarchical relation of two constituents with respect to each other for all remaining iterations of linearization in the derivation (cf. Fox and Pesetsky 2005a).^[7] To the extent that the discussion here motivates both local and global mechanisms, it can be taken as introducing an argument against the approaches in Müller (2001) or Müller (2007) that advocate a single (type of) mechanism for preservation effects more generally, and is in line with Fox and Pesetsky’s (2005b: 256–258) suggestion that two such mechanisms might be necessary. Accordingly, given the differences proposed in (2), we will see different consequences stemming from how the two scrambling DPs have their relative structural height fixed. In the LOG, their relation to each other observed at the ApplP edge is maintained throughout higher phrase markers, whereas in the AG, their relation at the vP edge can subsequently be permuted higher in the structure. This rigid hierarchy we see with the scrambled DPs relative to each other in the LOG thus constitutes novel evidence (in addition to Heck and Himmelreich’s data) against the view (e.g. Biskup 2017; Haider and Rosengren 2003) that scrambling is insensitive to hierarchy preservation and can freely permute the relative hierarchy of DPs within the German Mittelfeld (i.e., descriptively, the material that appears between a complementizer/fronted finite verb on the left and a clause-peripheral verbal complex on the right).

In what follows, I first describe the core German PG data in Section 2. I then briefly review in Section 3 Heck and Himmelreich’s proposal for the AG and how it cannot capture the LOG facts, before discussing the modified proposal in Section 4. Section 5 concludes with further discussion of some theoretical implications of the analysis and some questions about language acquisition and further variation for future research, as well as how the analysis here could potentially shed light on these questions. An Appendix follows that presents further data (beyond PGs) that provide support for the current proposal.

2 Core data

In this section, I present the core data that illustrate the contrasting patterns of PG-licensing in German. The data here focus on ditransitive examples, in which the dative and accusative-marked non-pronominal internal arguments scramble past a vP-level adjunct containing a single PG.

Heck and Himmelreich (2017) present the data in (3), which show that only the accusative DP can be the antecedent when a dative DP also moves past the adjunct. Regardless of the whether the scrambled dative DP precedes the scrambled accusative DP, the dative is not a licit antecedent (3a–b). However, under either order, the accusative DP is a licit antecedent (3c–d).

The conclusion is that scrambled accusative DPs block datives from acting as antecedent to the PG here, and this is supported by the observation that dative DPs are licit antecedents when the accusative stays low as we have already seen in (1b), or when there is no accusative DP, illustrated in (4) from Heck and Himmelreich.^[8]

Supporting data partially illustrating the pattern above with multiple scrambled DPs can also be found in Lee and Santorini (1994:267).

Müller (1995) presents a contrasting paradigm. The examples in (5) show that when both dative and accusative DPs scramble past the adjunct, either one can be a licit antecedent in principle, but word order matters. Of the dative and accusative DPs, it is only the one that is linearly closer to the adjunct that can be the antecedent.

Supporting data showing that it is the linearly closer dative or accusative DP that is the licit antecedent when both those DPs scramble past the adjunct containing the PG are given by Müller and Sternefeld (1994:375) and Fanselow (1993: 34), who gives examples parallel to (5a) and (5c–d) (though they involve more across-example lexical differences than in (5)), and who reports (p.c.) that (5b) is also grammatical.

To conclude, I take the systematic differences across the paradigms in (3) and (5) as evidence for variation across the grammars of German speakers (and for some discussion of the possibility of further variation in German see Section 5).^[9] I now turn to an analysis of the data.

3 Heck and Himmelreich’s (2017) proposal

In this section I first review how Heck and Himmelreich’s (2017) proposal captures the AG pattern, and then show the difficulty this proposal faces with the LOG data.

Heck and Himmelreich assume that v in German can have an Edge Feature (cf. Chomsky 2007) that can trigger movement (such as scrambling) of one or more constituents.^[10] When, for example, the Edge Feature attracts both dative and accusative DPs from within its c-command domain, the two XPs are placed in an ordered buffer in the form of a (movement) stack. Recall that such a buffer functions as a storage area for attracted constituents and places them in an order for re-merge. Dative indirect objects c-command accusative direct objects within v’s sister for the data under consideration (see also Fanselow 2000; Müller 1995 on this point), and for concreteness, I follow Georgala (2011) in assuming an Applicative Phrase (ApplP) hosts the IO (6a). Given this c-command relation, because datives are encountered first by the Edge Feature in its search down the tree for its two goals within its c-command domain, datives are placed in the stack first (at the bottom), and accusatives placed in the stack last (at the top). This is illustrated schematically in (6b): from their base positions, the attracted DPs have been placed in a movement stack.^[11] The DPs will subsequently be merged with vP iteratively, from the top of the stack to the bottom of the stack, in a first-in last-out manner. This also holds in cases where the derivation involves attaching an adjunct containing a PG to the vP; when the accusative DP from the top of the stack merges with this vP, the dative DP remains in the stack (6c).^[12] Upon re-merging and occupying its position c-commanding the adjunct, the accusative DP immediately establishes an Agree relation (Chomsky 2000, 2001) with the PG, thereby licensing the PG.^[13] Only at this point in the derivation does the dative DP merge with vP, emptying the stack (6d).^[14] But this step of merge is too late for the dative to license the PG, as this has already been done by the accusative DP. This accounts for the pattern we saw above in (3a) and (3c). Then, after the Edge Feature is no longer active, the external argument merges with vP (6e).

Further, if the accusative DP subsequently scrambles past the dative DP as in (6f), which also involves movement of the subject, the Agree relation between the accusative and the PG, and the attendant licensing by the accusative of the PG is maintained. This is what we see given the word order in (3b) and (3d), where even though the ultimate DO_ACC < IO_DAT hierarchy (a < indicating c-command when discussing hierarchy) does not parallel the hierarchical IO_DAT < DO_ACC relation below v in (6a), that hierarchy is maintained among the multiple specifiers at the vP level in (6d), resulting in only the accusative DP licensing the PG.^[15] And if there is no scrambling accusative, as in (1b) or (4), the stack will contain only the dative DP, and upon merging with vP, the dative is free to agree with, and license, the PG.

In sum, to capture the AG, Heck and Himmelreich’s proposal of using a stack for multiple movement dependencies ensures hierarchy-preserving movement at the vP level triggered by the Edge Feature, and has the result of the accusative DP blocking the dative from licensing an adjunct containing a single PG. I will follow their proposal, then, for the analysis of PG licensing in the AG data here (but see notes 9, 12, and 13 for qualifications).

However, it is clear that as it stands, Heck and Himmelreich’s proposal makes the wrong predictions for the LOG. Given the structure in (6c), where the accusative DP agrees with the PG, Heck and Himmelreich predict that in (5) the accusative antecedent should always be good, and the dative antecedent should never be good. In contrast to Heck and Himmelreich’s proposal, in (5) the hierarchical relation between the internal arguments prior to moving to the vP edge – and thus the role of any movement stack (or hierarchy-preserving movement) – is seemingly irrelevant for PG-licensing. In this way the LOG grammar poses a challenge to any unified account of PG-licensing in German. To see how these factors can be relevant, though, in the next section I will consider an important assumption embedded in Heck and Himmelreich’s proposal, which involves the number of positions a DP can scramble to.

4 Modifying the proposal

In this section I show that positing grammatical variability in the positions that scrambling targets and variability in which phrase markers are linearization domains allows for a unified analysis of both grammars under Heck and Himmelreich’s (2017) basic framework.

For Heck and Himmelreich, scrambling in the German Mittelfeld crucially targets only the vP edge within the verbal domain, with further scrambling possible at the TP level. Consider now a minimally modified proposal for scrambling within the Mittelfeld that will capture the LOG.

Instead of scrambling targeting just TP and vP, as per the AG, I propose that in the LOG, scrambling could additionally target some other phrasal marker dominated by vP. For the sake of concreteness, let us assume this vP-internal position is at the edge of ApplP. Similar to how T can optionally attract a DP via a scrambling feature, Appl would be able to optionally attract a DP to its edge via a comparable feature. A schematic structure for ApplP prior to this movement is given in (7a), where we see the canonical IO_DAT < DO_ACC hierarchy. (7b) then shows movement of the accusative DO to ApplP, resulting in the completed ApplP not being hierarchy preserving. Importantly, in the LOG, the mechanics of multiple attraction of DPs by the Edge Feature in v work just as in the AG. After merging in a v that will attract multiple DPs, those DPs again will be placed in a stack. However given (7b), it is now the accusative DP that will be placed in the stack first (at the bottom), because it is closer to the c-commanding v (7c). The stack will again be emptied by re-merging the DPs with vP in a first-in last-out manner. But given (7c), it is now the dative DP (on the top of the stack) that will first be re-merged with vP, allowing it to agree with, and license, the PG (7d). This agreement relation also blocks the accusative from subsequently agreeing with the PG when it merges with vP in (7e).

Introducing a new position for scrambling provides the first step in capturing the two points of contrast between the LOG and the AG, namely (5b) and (5d): when the dative DP is linearly closer to the PG than the accusative, as is the case when the structure containing (7e) is linearized and no further IO/DO scrambling occurs (see below for more on linearization), the dative must be the licensor. Note that this analysis does involve a movement stack and thus hierarchy-preserving movement to the vP edge triggered by the Edge Feature. This hierarchy is established at ApplP and can differ from that found in the AG at the vP edge because of the possibility of scrambling to ApplP in the LOG. Accordingly, because dative PG antecedents are impossible in (3), I assume that scrambling cannot target ApplP in the AG.

It is important to point out here that proposing that ApplP is a landing site for scrambling is in line with the null hypothesis that any phrase marker could in principle provide such a landing site. The null hypothesis is especially pertinent given the well-documented cross-linguistic variation in scrambling across languages. Now, this variation can be viewed as parametric variation regarding which heads can trigger scrambling. For example, the fact that scrambling cannot cross finite CP boundaries in German, but can in Japanese (Müller 1995), can simply reflect whether finite C is parametrically chosen as a scrambling head in the two languages. Thus introducing Appl as another scrambling head helps round out the typological picture, allowing for a more parsimonious view of scrambling, as per the null hypothesis – C, T, v, and Appl all having been identified as scrambling heads now – and the variation between the AG/LOG with respect to ApplP scrambling would be a parametric difference (cf. Section 5 for further discussion of parameter setting).

A question that now arises is why the structure in (7e) cannot feed the structure in (8). (8) is derived by scrambling the PG-antecedent dative IO in (7e) to TP and past the accusative DO. (8) should not be possible in the LOG, as it would give us the ungrammatical order in (5a) of a dative antecedent DP followed by the accusative (In the remaining discussion, I abstract away from the possibility of string vacuous scrambling of IO/DO to TP, though nothing substantively changes if such scenarios are considered).

(8)

Illicit dative-scrambling to TP in the LOG (cf. (7e))

The second step of the analysis is to assume that the LOG is distinguished from the AG by having a linearization domain within vP (in the sense of Fox and Pesetsky’s 2005a “spell-out domain”; see Fox and Pesetsky 2005a for the possibility that different grammars can have different linearization domains).^[16] For concreteness, I assume that the sister of v constitutes this linearization domain, as indicated by the arc in (8). For (7c), this gives us the linearization domain in (9a). According to Fox and Pesetsky, building a linearization domain triggers generating a set of ordering statements that define the linear order of the elements within the linearization domain (for details on this relation, see Fox and Pesetsky 2005a: 40–42).^[17] Being specifiers of the same phrase, with the DO c-commanding the IO, then translates into the schematic ordering statement in (9b) for (9a), which establishes the order of the DO as preceding the IO (a < indicating precedence when discussing word order).^[18]

Crucially, once a set of ordering statements has been generated for a linearization domain, it cannot be contradicted by any further sets of ordering statements resulting from subsequent construction of higher linearization domains. This still allows the IO and DO to scramble to a higher linearization domain, as in (7d) and (7e). As such movement is hierarchy-preserving, creating specifiers of vP, it effectively recreates what we saw with the specifiers of ApplP in (9). The result is that in this higher linearization domain, there will be a new ordering statement that is identical to (9b) (that is, if there is no further IO/DO scrambling in the domain containing (7e)), thereby preserving the relative order of the DPs that was established in the lower domain in (9a). However, linearizing the structure containing (8) so as to give us (5a) would involve contradicting the statement in (9b). As the c-command relation between the DO and IO specifiers has now been reversed (with IO now c-commanding DO) given (8), for the higher linearization domain we generate ordering statements such that the IO now precedes the DO, in contradiction of (9b). It is this ordering contradiction that I propose is the source of the ungrammaticality of (5a), given the derivation begun in (7).

The lower linearization domain in the LOG thus has the effect here of freezing the hierarchical relation of the IO with respect to the DO. This is the second type of hierarchy preservation mentioned in Section 1. Regardless of what subsequent movement dependencies are established later in the derivation by creating new specifiers, the relative order within lower linearization domains must be preserved upon the completion of all subsequent linearization domains.

Importantly, this proposal does not always block scrambling to TP in the LOG. The analysis here predicts that scrambling to TP is possible so long as it preserves the order established in the lower linearization domain. This prediction is borne out in (10a) (which parallels (4) of the AG), from Fanselow (1993: 34), who also provides data representative of the LOG (see also note 19 for further supporting data). In (10a), the dative DP first undergoes Edge Feature-driven scrambling to vP, where it establishes an Agree relation with the PG, before scrambling again because of T’s scrambling feature, as shown in (10b).

Unlike what we saw in (8), this second instance of scrambling in (10) is now possible because there is no constituent (such as the other internal argument in (8)) from within the lower linearization domain that will trigger an ordering contradiction along with the dative DP. Further, examples such as (10) show that structural considerations are critical in PG-licensing, and not linear order by itself, as the linearly closer subject DP does not disrupt the dative DP’s ability to participate in Agree and license the PG.

Let us now consider the other possibility available under the LOG, namely one where the accusative DP is the antecedent, and not the dative. I propose that this results in cases where the accusative DO does not scramble to ApplP, as in the AG. If the step of movement in (7b) does not occur, then movement to the vP edge will proceed as per the steps in (6a–d) above. As we saw there, in this type of derivation the accusative DO can license the PG, giving us (5c), but the dative IO cannot, resulting in (5a). As no scrambling has occurred within ApplP in (6), DP-hierarchy from this domain yields a IO_DAT < DO_ACC word order, and this is not contradicted when linearizing a structure containing (6e) and no further IO/DO scrambling. And similar to (8), the additional scrambling in (6f) is not possible in the LOG, as shown in (11).

(11)

Illicit accusative-scrambling to TP in the LOG (cf. (6f))

Again, as there is no scrambling within the lower linearization domain in (11), the IO_DAT < DO_ACC word order from the lower domain must be preserved with subsequent scrambling in (11). But as the higher scrambling in (11) disrupts the IO_DAT < DO_ACC hierarchy from the vP, linearizing the structure in (11) so as to give us (5d) would thus involve a contradictory DO_ACC < IO_DAT word order. This then correctly rules out the accusative antecedent in (5d), given the derivation begun in (6).^[19]

Having discussed the PG data, I point out here that further support for the analysis can be found with data involving floating quantifier alles ‘all’, which are discussed in the Appendix. Independent of PGs then, we see familiar AG and LOG patterns (in the AG, an accusative DO associate of alles instead of a dative IO; and in the LOG, a linear order effect involving the ditransitive’s internal arguments), and this can be analyzed in a parallel way as per the preceding sections. I refer the reader to the Appendix, and in particular the contrast in (14).

5 Conclusion, implications, and future research

The analysis in Section 4 thus captures the data in (5) and, along with the analysis in Section 3, allows us to maintain the core of Heck and Himmelreich’s (2017) proposal in accounting for both the AG and the LOG. For PG antecedents in both grammars, then, there is movement to the vP edge that is crucially dictated by an Edge Feature movement stack that preserves hierarchical relations below v. However, the relations of constituents that feed this stack can vary. In the data here, the LOG allows two different relations to precede movement to the vP edge, depending on whether Appl triggers scrambling. Further, the grammars vary in whether these relations can subsequently be disrupted after moving to the vP edge via scrambling to TP, with only the AG allowing such permutations. This is attributed to only the LOG having a linearization domain contained within vP, which acts to fix the final ordering, and as a consequence here, the hierarchy of the internal arguments. As the AG lacks such a low linearization domain, either ordering (and thus hierarchy) of the internal arguments is possible in the Mittelfeld (after scrambling to vP/TP) when linearization occurs with a higher domain (say, CP).

Thus, although any hierarchy of DPs involved in multiple scrambling dependencies within the Mittelfeld is in principle possible in both grammars, the analysis here advances novel evidence for hierarchy preservation as a constraining factor for scrambling. This view stands in contrast to accounts where scrambling in German is taken to be movement that more or less freely reconfigures constituents’ relations with each other (e.g. Biskup 2017; Haider and Rosengren 2003), and instead helps support Heck and Himmelreich’s proposal that restricts scrambling (and PG-licensing), in part, through feature-triggered movement and hierarchy-preserving movement stacks.

The analysis also highlights how hierarchy preservation phenomena involving multiple movement dependencies of the same type can stem from independent mechanisms in the grammar – namely the effects of movement stacks and linearization domains – although sometimes these mechanisms’ effects overlap. Thus on the one hand, when the Edge Feature attracts two DPs from a lower linearization domain, as in (7d) and (7e) in the LOG for example, the hierarchy preservation resulting from the movement stack matches the hierarchy preservation requirements that can be tied to the low linearization domain. Yet on the other hand, these mechanisms still have independently discernible effects. The first concerns how the hierarchy-preserving movement to vP of a movement stack feeds subsequent agreement outcomes, even when this phrase marker contains no linearization domain, as in (6) for the AG. And further, in the cases where scrambling is blocked in the LOG (cf. (8) and (11)), this is because of a hierarchy and consequent word order determined in a lower linearization domain, and is not related to multiple constituents being buffered in a movement stack. In sum, the data here have motivated an approach where such preservation effects are ultimately due to both local and global mechanisms: locally at particular derivational points via stacks, which no longer play a role in the derivation once they have been emptied; and globally via earlier linearization domains, which continuously play a role upon completion (as their ordering statements cannot be contradicted in the output of all subsequent domains). This is problematic for the approaches in Müller (2001 or 2007, which adopt only a single global or local mechanism respectively.

Next, an important question to consider is how child learners come to acquire the different German varieties. Properly addressing this goes far beyond what can be covered here, but I would like to point to an available direction for future research that is consistent with the proposal here that these varieties differ in terms of scrambling positions and linearization domains. The question of acquisition is especially pertinent given what is likely to be the total (or near total) absence of evidence for the learner that directly bears on distinguishing these varieties. Consider that the evidence here for the two varieties (and for the analytical differences proposed for them) relies on PGs, as well as floating quantifier alles ‘all’ in the Appendix, co-occurring with ditransitive verbs and very specific multiple movement dependencies. Such highly particular data points are likely to be vanishingly rare in the learner’s linguistic input. Indeed, a preliminary search through eight corpora in CHILDES (MacWhinney 2000) of child-directed speech in German (the Caroline, Leo, Manuela, Miller, Rigol, Szagun, Wagner, and Weissenborn databases) yielded no examples of PGs in the data (my thanks to Zhuqing Wang for informing me of these results). This is in line with what Pearl and Sprouse (2013: 54) report as an absence of PGs in their preliminary search through corpora of child-directed speech. Similarly, in a more detailed search for floating quantifier alles throughout the Caroline corpus, I identified 98 instances where alles was not adjacent to a potentially relevant wh-word associate. Fourty five tokens involved copular constructions with sein ‘to be’, where alles is not clearly a floating quantifier. In the remaining 53 tokens, alles is clearly a floating quantifier, but none of these examples involve ditransitive verbs, let alone the necessary multiple movement dependencies.

The conclusion, then, is that there appears to be a poverty of the stimulus acquisition puzzle (cf. Chomsky 1980): regardless of the theory adopted for PGs or alles, how do learners end up acquiring such similar, yet distinct, varieties of German given the paucity of evidence? If learners are not exposed to the kinds of data discussed in this paper, then to sharpen the issue in terms of the theory here, we can observe that the (ditransitive) input that the learners do end up receiving is in fact ambiguous as to whether there is scrambling within ApplP or whether there is a low linearization domain.

However, this ambiguity regarding scrambling positions and linearization domains might be key to understanding the variation we see. Gould (2017) presents a model for the acquisition of syntax that can crucially learn in systematic ways from ambiguous evidence. Gould shows that when given an input corpus of entirely ambiguous data for certain syntactic parameters, the model can arrive at different parameter settings for different learners (provided these parameters do not interact to a sufficient degree with other parameters), thereby providing an account of systematic variation that can be observed across speakers. An advantage of the proposal here is that the way it distinguishes the German varieties can be understood in terms of motivated sets of parameters that can be fed into this kind of learning model.

The full range of cross-linguistic variation (and thus the full details of these parameters for the model) remains to be explored. Still, as discussed in Section 4, (a) we can assume that one set of parameters will concern which phrase markers can allow for scrambling (including ApplP), which is motivated by the cross-linguistic variation we see regarding scrambling; and (b) given apparent cross-linguistic variation in what constitutes a linearization domain (cf. Fox and Pesetsky 2005a), we can also hypothesize that another set of parameters concerns which phrase markers delimit linearization domains. Different parameter settings for scrambling to ApplP and having a low linearization domain would then yield the AG and LOG along the lines of the analysis here.

Now, given these ingredients – (i) independently observed variation across languages as regards scrambling and linearization domains (represented via parameters), (ii) the ambiguous evidence facing German learners, and (iii) the independently motivated learning model in Gould (2017) – it could very well be the case that the variation in German is actually expected. And if such variation is indeed expected and can be modelled accordingly, then this would provide a different direction of support for the proposal here.

The details and results of actually applying the theory here to the learning model by running modeling simulations are a topic for future research (including more fine-grained details of the input corpus that the model learns from), but initial considerations indicate that this approach has the potential for success. That is, the ambiguous German input could result in the model learning the relevant parameter settings for the AG/LOG, thereby supporting the analysis here and addressing the acquisition puzzle regarding these varieties.

A final point here concerns the possibility of further variation in German. While I am not aware of additional variation regarding PGs and alles in German, this could be a fruitful area to investigate further, especially in light of the modeling considerations above. Indeed, given the ambiguity of the learner’s input and the learning scenarios sketched above, one reasonable expectation is that some learners would acquire a grammar where there is scrambling within ApplP, but no low linearization domain (a kind of cross between the AG and LOG). This would be reflected by judgments for (3)/(5), for example, where either the displaced accusative or dative DP could be the PG antecedent, regardless of linear order. It remains to be seen, though, whether such a sub-population of German speakers exists.