In this section, we discuss our methodology, presenting details of our data sources, data collection procedure, and the reasons behind the choices we make. We also point out some limitations and how they might be alleviated in future studies.

One of our major goals is to account for the divergence observed in the case patterns among the heritage Kurmanji speakers. To that end, we compare the case patterns of the heritage speakers (Kurmanji_H) with the case patterns of their parents and immediate family members (Kurmanji_A speakers). Following Polinsky & Kagan (2007), we take the baseline language for a heritage speaker to be the language that they were exposed to as a child. We also discuss Kurmanji_B patterns as our heritage speakers converge on the case patterns of Kurmanji_B despite being exposed to Kurmanji_A. The data for Kurmanji_H and Kurmanji_A come from our own fieldwork mainly through translation and acceptability judgments. The data for Kurmanji_B is from Gündoğdu (2011), who reports data through fieldwork elicited from native speaker consultants.3

Another source of data that we make use of in this study is the monolingual acquisition data from Erzurum reported in Mahalingappa (2009), which helps us delineate the role of contact with Turkish in the shift towards accusativity in Kurmanji_H. We call the data from Erzurum Kurmanji_C to indicate its source and case patterns. We do not make any assumptions about its dialectal status with respect to dialects A and B.

In the following, we provide relevant information about our consultants and the linguistic settings in which they grew up. The heritage variety of Kurmanji we report developed in an environment where the baseline dialect is Kurmanji_A. Our heritage consultants are two Turkish-Kurmanji bilingual cousins from Adıyaman, Turkey. They grew up in slightly different language settings but they converge on the same case patterns shifting towards accusativity. As part of describing the linguistic profiles of our heritage Kurmanji consultants, we also provide details regarding the baseline speakers from whom they received primary input.

Kurmanji_A is a tense based split ergative language with an accusative alignment in non-past clauses and an ergative alignment in past tense clauses (Dorleijn 1996; Thackston 2006; Atlamaz & Baker 2018). Morphologically, Kurmanji has a two-case system traditionally called direct (dir) and oblique (obl). Subjects of intransitive clauses are always in direct (dir) form regardless of whether they are internal or external arguments indicating that ergative in Kurmanji is not inherent case associated with a thematic role. As for transitive clauses, in non-past clauses, subjects are in direct form (i.e. morphologically unmarked) while objects are marked with the oblique case. The sentences in (2) show the case patterns for finite present tense clauses. The sentences in (3) show that the embedded subjunctive clauses, which are devoid of tense information, also have a dir-obl alignment regardless of the tense of the matrix clause under which they are embedded.

In past tense clauses, subjects of intransitives and objects of transitives are in direct form whereas the subjects of transitive clauses are in oblique form. The sentences in (4) illustrate the case patterns in the past tense.

Table 4 provides a summary of the split ergativity patterns in Kurmanji_A.

Notice also that agreement tracks the arguments in the dir form in both tenses. This seems to be the case across all the varieties. Following Bobaljik (2008), we simply assume that agreement tracks case and ignore the agreement facts but nothing hinges on this assumption. See Atlamaz & Baker (2018); Atlamaz (2019) for further complications on agreement with oblique nouns in Kurmanji.

In terms of case patterns, Kurmanji_B differs from Kurmanji_A in the past tense (Gündoğdu 2011). Instead of an obl-dir pattern in past tense transitive clauses we observe a double oblique pattern (obl-obl), which we treat as a tripartite case system. The sentences in (5) illustrate dir-obl alignment in finite present tense clauses whereas the ones in (6) illustrate the dir-obl alignment in embedded subjunctive clauses that are devoid of tense information.

In past tense clauses, subjects of intransitives are in direct form whereas the subjects of transitive clauses and objects of transitives are in oblique form. (7) illustrates case patterns in past tense clauses.

Table 5 provides a summary of the split ergativity patterns in Kurmanji_B.

The double oblique pattern in various dialects of Kurmanji has been analyzed as an instance of diachronic shift towards an accusative system (Dorleijn 1996; Haig 2004; Gündoğdu 2017). This has been considered as the first step towards a fully accusative system as the objects in these dialects are treated consistently as accusative across all tenses, while the subjects still retain properties of a split ergative system.

One of the major questions regarding the language changes of this type has been about the source of the change. Is this shift due to some language contact with a neighboring language or is it a result of some language-internal change? Both have been proposed for Kurmanji and both make sense (see Gündoğdu (2017) for an overview.). The Kurmanji dialects we report are in close contact with Turkish, a purely accusative system. Plus, the heritage speakers are Turkish-Kurmanji speakers with stronger self-reported proficiencies in Turkish in a Turkish dominant setting. All of this could suggest that Turkish-Kurmanji bilinguals are somehow “copying” the accusative pattern from Turkish to Kurmanji. However, this line of reasoning fails to capture the fact that Kurmanji is a split ergative language and it already has the accusative alignment in all the non-past clauses.

Heritage languages provide us with a unique opportunity to observe and analyze these types of shift in a synchronic manner. The heritage dialect we report emerged in a setting where the primary linguistic data comes from Kurmanji_A but the speakers converge on the Kurmanji_B case patterns despite no contact with Kurmanji_B speakers.10 Clearly, language contact with Turkish has an impact on the emergence of the heritage patterns as the Kurmanji input they receive is consistently split ergative (the same as the one their parents received from their own parents) but they diverge from it. However, this impact cannot be merely “copying” the Turkish patterns as discussed above.

The facts described above require a theory of case that can account for all the variation observed across the Kurmanji dialects and a theory of learning that allows for the acquisition of various case patterns. In the following, we show that the Dependent Case Theory (Marantz 1991; Baker 2015) accounts for all the variation. We also propose a theory of staged learning that allows us to describe the acquisition of case patterns as outcomes of various learning paths made available by the Dependent Case Theory. We show that these acquisition paths have several well-defined checkpoints and each dialect is associated with a particular checkpoint on an acquisition path.

Dependent Case Theory is a configurational case theory first proposed by Marantz (1991) to account for accusative and ergative case assignment. It was later extended by Baker (2015) to account for tripartite case systems. The dependent case rules formulated by Baker (2015) are given below.

The dependent case rules in (10) cover all the case patterns observed across the Kurmanji dialects as well as Turkish. Let us start with the intransitive clauses, both unergative and unaccusative. DCT predicts no case on the arguments of intransitive verbs as there is no case competitor. This is realized as the absence of any overt case marking on the arguments, traditionally called dir case in Kurmanji and nom in Turkish. Following Kornfilt & Preminger (2015), we adopt the view that nominative, absolutive, and direct are simply overt realizations of the absence of case.11 Table 8 displays the predictions of the DCT for intransitive sentences across Turkish and all the relevant Kurmanji dialects. ‘Ø’ indicates the absence of any surface or deep form rather than implying a null morpheme.

For transitive clauses, the Dependent Case Theory predicts four distinct case patterns corresponding to the accusative, ergative, tripartite, and unmarked alignments shown in Table 9. While the downward dependent case rule assigns what is traditionally called the accusative case, the upward dependent case rule assigns the ergative case. In languages where both rules are actively employed, we end up with a tripartite system. Finally, the lack of dependent case leads to a possibly unmarked system with no dependent case on either argument. This covers the logical space of dependent case theory with intransitive and monotransitive clauses.

In the following, we present the learning problem and show how the predictions of the DCT combined with a learning framework where the dependent case rules are learnt on distinct stages accounts for all the variation observed across the languages and dialects under discussion. But before that, one terminological clarification is due.

Throughout the paper, we use traditional case names such as absolutive, direct, nominative, accusative, ergative, oblique, etc. only descriptively. We merely follow the traditional descriptions used in the literature. We use these conventions in our glossing as well. However, the crucial concepts for our analysis are the alignment patterns, accusative, ergative, and tripartite. We cover all these alignments with the downward dependent case and the upward dependent case. Languages vary in how these cases are realized and so do the traditional names associated with them. For example, downward dependent case is realized by the -(y)I suffix in Turkish, which is the dedicated accusative morpheme, while Kurmanji uses what has been traditionally called the oblique form. In Kurmanji, the upward dependent case is also realized through the oblique form.12 While the surface forms matter as they are the primary input to the acquisition mechanism, their names do not matter for our discussion.

The Dependent Case Theory is capable of accounting for all the variation observed in the case patterns of Kurmanji dialects and Turkish. In other words, the DCT has the descriptive adequacy to be a plausible theory of case.13 Descriptive adequacy is a necessary condition for any theory to be entertained seriously but it is not a sufficient condition. In addition to descriptive adequacy, a grammar posited by a theory must be learnable and allow room for predictable divergences from the input forms (e.g. overgeneralization errors, partial learning, etc.) (Pinker 1984). In this paper, we do not focus on the learnability problem, which deals with whether a particular grammar can be learnt from a particular set of input forms. Instead, we focus on the predictable divergences from the input forms within the tenets of the Dependent Case Theory. Even though we do not focus on the learnability problem, we need to make some assumptions regarding how the dependent case rules are learnt. These assumptions coupled with the DCT reveal that the observed dialectal variation is a natural outcome of the DCT (including the heritage patterns). In the following, we first position our proposal within generative approaches to language acquisition. Then, we provide the details of our account by defining the learning targets and then listing our assumptions regarding how these targets are learnt.

We propose a checkpoint based learning mechanism in the spirit of the Principles & Parameters framework (Chomsky 1981), where universal principles determine the parameter space and language acquisition is considered to be a process of setting the parameters of a grammar based on input data. In this sense, our proposal follows the line of acquisition research pursued in Hyams (1983; 1986), Wexler & Manzini (1987), Gibson & Wexler (1994), Yang (2000; 2004; 2011), Fodor & Sakas (2005), Fodor et al. (2007) among many others. We do, however, differ from some of them in the nature of parameters and where they come from. Much of the acquisition research within the Principles & Parameters framework assumes that the Universal Grammar determines the parameter space, where the parameters have binary values and they are set depending on the input. This parametric view has recently been challenged mainly because it predicts a massive amount of variation even with a relatively small number of parameters. The size of the parameter space is not only a problem for massive variation, most of which is never attested, but is also computationally intractable. See Newmeyer (2013) for a discussion.14

Following Arregi & Nevins (2012), we assume a highly modular architecture of grammar where each component of the grammar is dedicated to a particular task and has its own constraints. One such module is the Case Module which is constrained by the principles of the Dependent Case Theory. We assume that knowing the case system of a language involves inducing a set of rules from a finite amount of data.15 We further assume that the Dependent Case Theory limits the space of possible hypotheses by providing a set of possible rule types. To make the problem more explicit, we assume the following learning targets.

The precise characterization of the learning targets (i.e. dependent case rules) are slightly different for each language we discuss and they are provided in the relevant subsections below. Dependent Case Theory defines the logical space of target abstract rules to be learnt. We define the learning task as setting the direction parameters (upward, downward) of the dependent case rules for the right context.

Now that we have defined the learning targets and the learning task, we are in a position to articulate our assumptions regarding how these targets are learnt. We assume that the direction parameters of the dependent case rules are induced from the input data. Learning involves the following major operations:

Hypotheses are formed based on positive evidence from the data. Such hypotheses based on positive evidence are Plausible Hypotheses. The set of plausible hypotheses PH is a subset of the set of all the logically possible hypotheses H permitted by the Dependent Case Theory. Consider the hypothetical alignment pattern in (14).

Given the hypothetical input alignment in (14), the set of logically possible hypotheses are as in (15) and the set of plausible hypotheses are given in (16).

The plausibility restriction shrinks the hypothesis space to include the ones that are constructable from the input data and exclude the ones that are not.16 Once a hypothesis is formed, it needs to be checked for productivity. Productive hypotheses will be accepted while the unproductive ones will be rejected.17

We assume that generalization of the hypothesized rules simply follows from their underspecified nature. For example, h1 in (16) applies across all tenses, aspects, persons, etc. resulting in a consistently accusative system (as opposed to a tense/aspect based split). We further assume that learning is incremental and accepted hypotheses can be further revised in scenarios where they lead to errors (Widrow & Hoff 1960; Rescorla 1972; Fodor & Sakas 2005; Divjak et al. 2021).

Now that we have laid out our main assumptions regarding how abstract rules are learnt from input data, we turn to some specific assumptions regarding how the dependent case rules are learnt. In the previous section, we outlined the set of learning targets as the context specific dependent case rules. To achieve adult-like competence in case patterns, a learner must learn the dependent case direction parameters (upward/downward) as well as the correct context specifications (e.g. tense/aspect) associated with each dependent case rule. Our key assumption is that the direction parameters (i.e. downward and upward) and the context specifications can be learnt simultaneously or sequentially. When the learning is sequential, context specifications will have to follow the direction parameters as the reverse order would be neither a plausible hypothesis nor dependent case per se. We show that the heritage case patterns observed in Kurmanji emerge when the dependent case rules and their context specifications are learnt sequentially. Now that we have the main assumptions in place, we can start looking at the details more closely. For this, we start with Turkish, a relatively straightforward system.

Turkish has a consistently accusative alignment. The grammar described in (17) accounts for the facts in transitive clauses regardless of tense, aspect, or finiteness. We assume that a grammar is a state of an individual’s I-language at a given time. To achieve adult-like competence, a learner needs to reach the Final Attainment State in (17) within the tenets of the Dependent Case Theory.

Given that the logical space of rules to be learnt are restricted to one of the downward and upward case rules and the fact that Turkish is a consistently accusative language with no ergativity, a learner should be able to learn the downward dependent case rule given in (17), which is the only plausible hypothesis. Thanks to underspecification, a rule induced from input like in (18) readily applies to cases like (19).18

Kurmanji_A is a split ergative language with an ergative alignment in the past tense and an accusative alignment elsewhere (present tense and subjunctive). The final attainment state L_KAn given in (20) generates the correct case forms for Kurmanji_A. Following, Atlamaz & Baker (2018), we assume that the split ergativity in Kurmanji can be captured by the DCT and we assume the grammar in (20).

In order to produce the right case forms, a learner needs to attain the grammar in (20). This amounts to learning two dependent case rules (upward and downward) and the context specifications associated with each rule (past and non-past/elsewhere).

Primary linguistic data (PLD) is random but the speakers of a language usually converge on the same target patterns given “sufficient” data and time. Likewise, it has been shown that children usually follow similar acquisition paths even though the individual trajectories might show some differences.19 We assume that the acquisition paths for learning the case patterns are guided by the dependent case theory. For a split ergative system like Kurmanji_A, we assume that a single case grammar parameterized across tenses is learnt. What needs to be learnt is given in (21):

The key assumption is that the dependent case rules and context specifications can be learnt simultaneously or sequentially. The learning task in (21) combined with our assumption that the rules can be learnt sequentially leads to a variety of learning paths with various checkpoints for each learned rule or context specification. A checkpoint is defined as an accepted hypothesis state. When a hypothesis is accepted, a checkpoint is created and the hypothesis is maintained as a rule until it is further revised.20 In the following, we show how various distinct paths converge on the same target split ergative patterns, predicting individual differences among learners during the learning process. We show that some of the checkpoints on these acquisition paths correspond to the heritage case patterns (accusative in non-past clauses and double oblique in past tense clauses). We also present data to show that some of these paths are attested in monolingual acquisition data as well. This leads to the conclusion that the heritage case patterns in Kurmanji emerge naturally as checkpoints on the acquisition paths which weakens the hypothesis that the shift towards accusativity is due to contact with an accusative language. Instead, we speculate that the impact of language contact is more indirect and a result of reduced intake, specifically reduced production which reduces the amount of hypothesis testing.

The assumption that the items in the Kurmanji-specific learning task in (21) can be learnt simultaneously or sequentially leads to a set of paths that can be grouped as 2, 3, and 4-checkpoint paths. We start with the 2-checkpoint paths and quickly discard them as they do not render the observed heritage patterns and then move on to show that some of the 3 and 4-checkpoint paths render the observed patterns.

Let us first define the plausible hypotheses for Kurmanji before we show the logically possible 2-checkpoint paths that converge on the Kurmanji_A grammar. Table 10 lists the hypotheses. The arrows stand for the direction parameters for the dependent case rules and the subscript (_C) indicates that the hypothesis contains the correct context specification for that dependent case rule.

In a 2-checkpoint path, each of the dependent case rules (down or up) must be learnt simultaneously with their context specifications. Otherwise, the grammar does not converge on Kurmanji_A. Table 11 shows the logically possible 2-checkpoint paths that converge on the L_KAn described in (20) for Kurmanji_A. The checkpoint column indicates the checkpoint number in the learning path. The columns non-past and past indicate the predicted case alignments in transitive clauses at each checkpoint.

In Path 1, first the downward dependent case rule (accusative) is learnt for non-past clauses and then the upward dependent case rule (ergative) is learnt for past tense clauses. In Path 2, the order is reversed. Both of the paths converge on the Kurmanji_A grammar yielding the expected case patterns, DIR-OBL in non-past and OBL-DIR in past. While the final checkpoints are important as they indicate the convergence point, the previous checkpoints are equally important as they indicate the intermediate stages a learner could go through. Unless the transition from checkpoint 1 to checkpoint 2 is immediate, the paths predict a stage where the learners produce outputs that are not in the input. We call such case patterns that are not in the input but predicted at a particular checkpoint divergent. The divergent patterns in the 2-checkpoint paths are given in Table 12.

The 2-checkpoint paths predict unmarked alignment with DIR-DIR forms at the intermediate checkpoints. Path 1 Checkpoint 1 predicts DIR-DIR in past tense clauses only whereas Path 2 Checkpoint 1 predicts DIR-DIR in non-past tense clauses only. The monolingual acquisition data presented in Mahalingappa (2009) has 0 instances of DIR-DIR in the 104 past tense clauses. On the other hand, out of the 130 present tense clauses 20 have the DIR-DIR form.21 Our heritage consultants do not accept past tense clauses with DIR-DIR alignment either.

The fact that DIR-DIR is observed only in present tense clauses but not in past tense clauses does not follow from the Dependent Case Theory and requires an explanation either in terms of the lack of empirical coverage or some principled way of constraining the paths generating DIR-DIR alignments. In the following sections, we repeatedly observe that the DCT combined with our checkpoint-based learning hypothesis predicts certain paths and checkpoints whose predicted forms are not attested in any of the dialects discussed in this paper including the monolingual acquisition data and the heritage patterns. To address this we pursue a procedure that aims to capture the general tendencies across the dialects by constraining the learning paths to a proper subset of all the logical possibilities. Within this constrained space, our learning paths still predict certain alignments we do not observe in the dialects we discuss including the monolingual acquisition paths. In such cases, we survey some of the other dialectal variation data discussed by Dorleijn (1996) to show that some such patterns exist at least in some Kurmanji dialect. Even though we prefer to constrain the learning paths in a principled way to increase the predictive power of our theory, we do not rule out the possibility that some of the paths we exclude might be followed by some learners or some dialects for which we do not have data.

To close the discussion on the 2-checkpoint paths, the DIR-DIR pattern in past tense clauses predicted by Path 1 does not occur in the dialects we focus on. However, Dorleijn (1996: 121–122) reports presence of DIR-DIR alignment in past tense clauses in her survey of various dialects. Thus, we do not rule out Path 1 on the basis of the lack of empirical coverage in our own data. Some Kurmanji speakers might be following Paths 1 or 2.22

Although we do not rule out the Paths 1 and 2 for some Kurmanji speakers, we do rule them out for our heritage speakers and all of the monolingual children reported by Mahalingappa (2009). While these paths were available to them, they can’t have gone through these paths since the 2-checkpoint paths fail to predict the double oblique pattern (OBL-OBL) observed in Heritage Kurmanji, Kurmanji_B and the monolingual acquisition data from Kurmanji_C.23 Producing the divergent obl-obl forms necessitates learning paths where the direction parameter of the downward dependent case and its context restriction are learnt sequentially. In the next section, we enumerate the 3-checkpoint paths and show how some of the paths contain both the Kurmanji_A patterns as well as the Kurmanji_B, Kurmanji_H and some of the monolingual acquisition patterns from Kurmanji_C.

3-checkpoint paths (Table 13) differ from 2-checkpoint paths in that one of the dependent case rules is learned in two sequential steps. First, a context-free version of the rule is learnt. Then, its context specification is learnt at a subsequent checkpoint. Crucially, only one of the rules is learnt in two sequential steps while the other rule is learnt in a single step along with its context specification.

There are only four logically possible learning paths consisting of 3-checkpoints. While all of those paths converge on the Kurmanji_A patterns as their final checkpoints, they differ in their intermediate checkpoints. Path 3 and Path 6 contain checkpoints producing the double oblique pattern in past tense clauses and not in non-past clauses. Outputs generated at the second checkpoints of Paths 3 and 6 render the heritage patterns (double oblique in past tense clauses), a desired outcome. On the other hand, the second checkpoints of Paths 4 and 5 predict double oblique alignment in non-past tense clauses but not in past tense clauses. From the heritage grammar point of view (i.e. Kurmanji_H), this is an undesired outcome since these paths both undergenerate (no double-oblique in past tense clauses) and overgenerate (double-oblique in non-past clauses). In addition, out of the 130 transitive present tense clauses in monolingual acquisition data from Mahalingappa (2009) only one has OBL-OBL alignment. Dorleijn (1996) does not report any instances of OBL-OBL in present tense clauses and none of the dialects we report accept OBL-OBL in the non-past clauses either. Given that there is only one instance in the entire corpus and knowledge space available to us, we conclude that the single example might be a production error.24

Once again, the fact that non-past tense clauses are never (productively) OBL-OBL does not follow from the DCT and requires an explanation either in terms of the lack of empirical coverage or some principled way of constraining the paths to exclude Paths 4 and 5. In this case, we think that the culprit is not the lack of empirical coverage. Instead, we argue that Paths 4 and 5 are never pursued due to the nature of the context that accompanies the upward dependent case. In Paths 3 and 6, the downward dependent case is learnt in two steps whereas in Paths 4 and 5, it is the upward dependent case that is learnt in two steps.

The conclusion we arrive at is that the upward case is never learnt in two steps in Kurmanji. All the paths where the upward dependent case is learnt in two steps both overgenerate and undergenerate. We argue that this follows from the distribution of the upward dependent case, which in Kurmanji must make reference to a smaller set of environments (past tense only) as opposed to the downward dependent case which is the elsewhere case (present tense and subjunctive-imperative). Since the upward dependent case always co-occurs with past tense transitive clauses, the association of the upward dependent case with its context is more straightforward. This aligns well with the facts reported in the acquisition of ergative literature. Ergative has been shown to be learnt early with minimal overgeneralization errors. Pye (1990: 559–560) showed that children acquiring K’iche’ made very few errors with ergative subjects. Neither ergative nor absolutive was overgeneralized. Butt (2006: 176) mentions a study she carried out in 1991 on the acquisition of ergative in Hindi, a split ergative language like Kurmanji. She reports that there were “vanishingly few errors with ergative” (Butt 2006: 560). The ergative acquisition literature has shown that ergative does not lead to commission errors.25 We observe the same in Kurmanji and we believe that the tendency towards OBL-OBL in past tense clauses but not in non-past tense clauses across many varieties of Kurmanji is due to the narrower environment associated with ergative.

Now that we have ruled out Paths 4 and 5, we are left with Paths 3 and 6, which project trajectories where the upward dependent case is learnt in a single step whereas the downward dependent case is learnt in two steps. Learning an underspecified version of the downward dependent case rule at a checkpoint and setting its context specification at a later checkpoint allows room for the heritage case forms (double oblique in past tense clauses) to emerge naturally. We claim that the heritage speakers who use the double oblique alignment in past tense never set the context parameter for the downward dependent case rule and use checkpoint 2 in Paths 3 or Path 6 as their final attainment state diverging from the input forms. The final attainment state for heritage Kurmanji is given in (23).26

One of the key issues regarding the emergence of the heritage patterns in this model is the convergence problem. Why do the heritage speakers use Checkpoint 2 and diverge from the input patterns predicted by Checkpoint 3? We argue that this follows from the nature of the transition from Checkpoint 2 to Checkpoint 3, which requires a revision in a previously learnt underspecified rule. The downward dependent case rule which does not have a context restriction in Checkpoint 2 must be specified with the elsewhere condition.

Going from an underspecified rule to one that makes reference to the elsewhere condition requires production, hypothesis testing, and error-driven learning. The learner must realize that the grammar they acquired overgenerates (OBL objects in past tense clauses) and seek hypotheses to fix the overgeneralization problem.27 The relevant hypothesis in this case is the elsewhere condition, which has to make reference to the context for the upward dependent case (past tense) as it is the complement set of the more specific context (upward dependent case in past tense). In a nutshell, the transition from an underspecified rule to a context sensitive rule requires hypothesis testing and rule revision. We argue that this requires production and the heritage speakers have a reduced rate of production which prevents them from realizing the overgeneration problem caused by the underspecified downward dependent case rule. Thus, the impact of Turkish on the shift towards accusativity in Kurmanji dialects is indirect. Heritage Kurmanji speakers do not copy the accusative pattern from Turkish.28 Instead, we argue that the heritage speakers do not produce enough29 Kurmanji forms to do hypothesis testing and rule revising.30

Our proposal essentially follows the Reduced Intake Hypothesis. Putnam & Sánchez (2013: 480–481) draw a distinction between input and intake where input refers to the raw linguistic input whereas intake refers to the manipulation of this input which involves interpreting, extracting, and storing the formal features of language. In this model, in addition to the amount and quality of the input (Montrul 2002; Polinsky & Scontras 2019), a learner’s active interaction with the input (use for comprehension and production) is also of significant importance. Following Putnam & Sánchez (2013: 480–481), we argue that production is a fundamental piece of the acquisition process and reduced production can lead to divergences within a constrained logical space guided by the principles of grammar (the DCT in this case). The setting in which our consultants were raised points to a reduced intake, specifically reduced production, rather than a significantly reduced input. This is especially true for our Consultant 2, whose mother is a monolingual Kurmanji speaker and father is a late Kurmanji-Turkish bilingual,31 and the parents speak with Consultant 2 primarily in Kurmanji. We observe similar facts in the monolingual acquisition data from Mahalingappa (2009). The dialogue in (24) indicates that the child produces the divergent OBL-OBL pattern in the past tense even though the caregiver uses the OBL-DIR pattern, pointing to a divergence from the input.

In the previous section, we ruled out paths where the upward dependent case and its context specification are learnt in different checkpoints. This allowed us to account for the lack of OBL-OBL alignment in non-past clauses and it provided an explanation for why ergative is learnt relatively straightforwardly without overgeneralization errors (unlike the downward dependent case). Sticking to the same premise simply rules out all the 4-checkpoint paths as these are paths that project trajectories where each of the learning tasks listed in (21) is learnt at a distinct checkpoint, requiring a separation of the upward dependent case and its context.

For the sake of completeness of the discussion and in the interest of articulating all the predictions of the theory, we provide all the six logically possible 4-checkpoint paths in Appendix A. Five out of six paths (Paths 7–11) overgenerate by producing OBL-OBL in non-past clauses, three of them (Paths 8, 10, 12) overgenerate by predicting OBL-DIR alignment in non-past clauses and one of the paths (Path 11) undergenerates by not predicting the OBL-OBL in past tense clauses. Although all of the 4-checkpoint paths converge on Kurmanji_A in their final checkpoints, none of them provide a trajectory where all of the intermediate checkpoints are attested across the Kurmanji varieties and one of them undergenerates. In contrast, 3-checkpoint paths (Path 3 and 6) provide just the right amount of room for variation allowing the heritage patterns to emerge without significant overgeneration. The DIR-OBL pattern in the past tense predicted by Path 3 is not accepted by our heritage consultants. However, it does occur in the monolingual acquisition data reported by Mahalingappa (2009). Such patterns are observed in heritage speakers from the Batman area, indicating that there is variation in the potential paths followed by learners.

To conclude this section, if our assumption regarding the simultaneous learning of the upward dependent case along with its context specification is correct, then all of the 4-checkpoint paths are automatically ruled out, restricting the number of learning paths significantly while still allowing sufficient room for the observed variation across the Kurmanji dialects. It could well be the case that some of the learners follow 4-checkpoint paths which predicts a lot more variation than what we empirically cover here. It is also possible that some of the forms which are unattested to the best of our knowledge are in fact attested in some dialects or in some mono-lingual acquisition data. In that case, we would need to invoke the 4-checkpoint paths and work out the precise constraints that pick out the right set of paths that predict the attested forms without over or undergenerating. We are not aware of any data points that challenge our account that rules out the 4-checkpoint paths but we believe that the DCT based framework we developed here has the potential to provide the precise set of paths that can handle the systematic variation observed across Kurmanji dialects.