Natural Language & Linguistic Theory

, Volume 35, Issue 3, pp 581–634 | Cite as

Locality and variation in Finnish structural case

Article

Abstract

Finnish has both nominative and genitive objects. The two cases are normally in a complementary distribution based on the local syntactic context (Jahnsson’s Rule). The pattern breaks down in nonfinite clauses where the conditioning is non-local and the cases may occur in free variation. This puzzling pattern can be understood if we make the following assumptions: (i) structural case distinguishes the external argument from other arguments; (ii) structural case assignment is cyclic. In our optimality-theoretic analysis the choice of case is determined by the interaction of markedness constraints that apply cyclically and faithfulness constraints that protect case assigned on prior cycles. Non-locality arises because faithfulness is violable; free variation arises because constraint conflicts can be resolved in multiple ways. In addition to categorical well-formedness contrasts the analysis predicts degrees of well-formedness in cases of free variation.

Keywords

Case Locality Cyclicity Variation Intermediate well-formedness Optimality Theory 

1 Introduction

Structural case has been the subject of much cross-linguistic research and has provided evidence for very different theoretical positions. In early GB, the Case Filter required overt NPs to have abstract case (Chomsky 1981). Abstract case and morphological case were initially assumed to be directly related, but it was soon pointed out that they were relatively independent, perhaps the best-known example being that of dative subjects in Icelandic (Zaenen et al. 1985). What is characteristic of this research program is the focus on the licensing of NPs in which morphological case plays a somewhat secondary diagnostic role.

In more recent work, one can distinguish two main views on structural case. One view holds that case is assigned to NPs by functional heads via agreement (e.g., Chomsky 2000, 2001; Legate 2008; Vainikka and Brattico 2014); for an overview, see Bobaljik and Wurmbrand (2009). An alternative view holds that case is assigned to an NP depending on the presence of other NPs in the same local domain (e.g., Yip et al. 1987; Marantz 1991; Maling 1993; Harley 1995; McFadden 2004, 2009; Sigurðsson 2006; Bobaljik 2008; Baker and Vinokurova 2010; Baker 2014, 2015; Levin and Preminger 2015; Poole 2015, 2016).

Finally, researchers of a functionalist orientation (e.g., Mallinson and Blake 1981; Comrie 1989) have understood structural case in terms of the need to identify and distinguish arguments, in particular direct objects (P) in nominative-accusative languages and transitive subjects (A) in ergative-absolutive languages. Interpreted as violable constraints in the sense of Optimality Theory (Prince and Smolensky 1993/2004), such functional principles have yielded new insights into structural case assignment (e.g., Legendre et al. 1993; de Hoop and Malchukov 2008).

In this paper, we will focus on two empirical phenomena that pose problems for all these approaches. The first is the problem of non-locality. It is commonly assumed that structural case is assigned within some local domain. In GB and Minimalism, this has been understood as a local head-government relation or local agreement between a lexical or functional head and the nominal in question (e.g., Chomsky 1981; Pesetsky and Torrego 2004, 2011; Bobaljik and Wurmbrand 2005). Apparent instances of long-distance case assignment have been argued to be in fact local (McFadden 2009). In non-derivational frameworks such as HPSG or LFG, the locality of case assignment has been attributed to local constraints on lexical selection, local constructional properties (e.g., Pollard and Sag 1994; Przepiórkowski 1999; Malouf 2000; Kim and Choi 2004) or local constraints on the association of case and grammatical functions (Zaenen et al. 1985; Nordlinger 1998; Butt 2006). Similarly, in Optimality Theory (Prince and Smolensky 1993/2004) case assignment has been understood in terms of constraints distinguishing core arguments within a local domain (Legendre et al. 1993; Woolford 2000; Aissen 2003; de Hoop and Malchukov 2008; Anttila and Kim 2011). However, there are many known examples where case assignment seems less local than expected; see, e.g., Zaenen et al. (1985) and Sigurðsson (2006) for Icelandic; Dench and Evans (1988) and Blake (1994) for languages of Australia; Raposo and Uriagereka (1990) for European Portuguese; Polinsky and Potsdam (2001) for Tsez; Bhatt (2005) for Hindi-Urdu; Legate (2005) for English; and most relevantly for our purposes, Kiparsky (2001); Brattico (2012); and Vainikka and Brattico (2014) for Finnish. Much less often discussed is the problem of free variation. This is a situation where two distinct cases, e.g., nominative and accusative, are interchangeable in the same environment, with no difference in meaning.

Finnish objects occur in four cases: nominative (nom), genitive (gen), accusative (acc), and partitive (par).1 The distribution of the partitive is a long-standing puzzle that involves semantic conditions (see, e.g., Kiparsky 1998); the accusative is limited to human pronouns. Both will be set aside here. In this paper we will focus on the nominative and the genitive whose distribution is largely complementary and syntactically conditioned (see, e.g., Hakulinen and Karlsson 1975).2 The problems of non-locality and free variation arise in nonfinite clauses illustrated in (1) and (2). In both examples the matrix verb ‘think’ takes a non-finite complement ‘Matti to have shot a bear.’ The examples differ in matrix voice: (1) is active, (2) is passive. These examples show that matrix voice matters to case in the embedded clause: the embedded object is invariably gen under active, but varies nomgen under passive with no difference in meaning. The non-locality problem has been discussed in, e.g., Vainikka (1989, 1993, 2003), Maling (2004), Kiparsky (2001, 2010), Brattico (2012), and Vainikka and Brattico (2014). The free variation problem has to the best of our knowledge never been explored in depth, with the remarkable exception of Itkonen (1976, 1981) whose work will be reviewed below. For the sake of brevity, we shall call these structures Itkonen structures.
Other non-finite complements show similar patterns. One such structure is discussed by Ikola (1950, 1957, 1964, 1989) and illustrated in (3) and (4): Here ‘criticize’ takes an NP object which contains a VP complement. The embedded object varies nomgen under an active matrix clause with no difference in meaning, but is invariably nom under a passive matrix clause (see, e.g., Brattico 2012). For the sake of brevity, we shall call these structures Ikola structures.

Our solution to the non-locality problem builds on the assumption that case assignment is cyclic: case in a complex clause is a function of case in its constituents. The cycle has occupied a central place in linguistics since at least Chomsky et al. (1956) and continues to play a role under the notion of phase (Chomsky 2001). We interpret the cycle in terms of Stratal Optimality Theory (Kiparsky 2000): case is assigned by markedness constraints in a cyclic fashion; locality is imposed by faithfulness constraints that protect case assigned on prior cycles. Since faithfulness is violable the analysis predicts non-local effects under specific circumstances.

Our solution to the free variation problem starts from the insight that case is assigned by competition: the case assigned to an NP depends on the presence of other NPs in the same local domain. This “configurational” view of case assignment is discussed in, e.g., Zaenen et al. (1985), Yip et al. (1987), Marantz (1991), Maling (1993), Harley (1995), McFadden (2004, 2009), Sigurðsson (2006), Bobaljik (2008), Baker (2014, 2015), Levin and Preminger (2015), and Poole (2015, 2016). In certain environments, especially ones where markedness and faithfulness conflict, the competition results in multiple outcomes in the sense of Partial Order Optimality Theory (Anttila 1997; Anttila and Cho 1998/2003; Djalali 2014), hence variation. Crucially, the analysis does not stop at predicting variation, but also predicts degrees of well-formedness among the variants.

The paper is structured as follows. Section 2 lays out the basic empirical generalizations, gives an analysis of simplex clauses, and introduces our background assumptions about the place of variation in grammar. Section 3 derives the case patterns in Itkonen structures. In addition to well-formedness contrasts of the familiar categorical kind the analysis predicts intermediate degrees of well-formedness in cases of variation. Section 4 extends the analysis to Ikola structures. Section 5 addresses other nonfinite constructions. Section 6 discusses the alternative view that structural case is assigned through agreement, focusing on Vainikka and Brattico (2014), a recent agreement-based analysis of Finnish structural case, concluding that a cyclic analysis stated in terms of violable constraints goes further. Section 7 concludes the paper.

2 Preliminaries

2.1 Jahnsson’s rule

The basic generalization about object case in Finnish is known as Jahnsson’s Rule (Jahnsson 1871; Kiparsky 2001). The rule is stated in (5) and illustrated in (6) and (7).
  1. (5)

    Jahnsson’s Rule: If the verb has an overt subject, the object is genitive (gen); if the verb has no subject, the object is nominative (nom).3

     
The active transitive (6b) has an overt subject, therefore the object is gen. The imperative (7a) and passive (7b) have no subject, therefore the object is nom. While passivization demotes the subject, the object remains an object and is not promoted to subject; see, e.g., Manninen and Nelson (2004) and Kiparsky (2013).4
These case patterns have received a straightforward explanation in terms of the Case Tier Hypothesis (Yip et al. 1987; Maling 1993, 2004, 2009). The hypothesis is stated in (8):
  1. (8)
    Given the hierarchy of grammatical functions subj > obj > adv
    1. a.

      The most prominent NP gets nom; the remaining NPs get gen.

       
    2. b.

      NPs bearing lexical/inherent cases are skipped.

       
     
The Case Tier Hypothesis can be illustrated from examples (6)–(7). If the clause has a subject it receives nom (6ab); if in addition the clause has an object it receives gen (6b); if the clause lacks a subject, but has an object, the object receives nom (7ab).
The skipping of lexical/inherent cases can be illustrated from existential clauses (Ikola 1964: 32; Hakulinen and Karlsson 1979: Sect. 9.3.2; Vilkuna 2000: Sect. 4.4.2; Hakulinen et al. 2004: 893–894, 923) where we have a locative NP marked by a semantic case, an intransitive verb such as olla ‘be’, and a second NP. An example is given in (9). The first NP metsä-ssä ‘forest-ine’ is skipped because it has an inherent inessive case. This leaves karhu ‘bear’ as the most prominent available NP which therefore receives nom. Note that it does not matter whether karhu is a subject or an object, or neither, because the theory assigns case hierarchically, denying any direct connection between case and grammatical function.
Predicative clauses are an outstanding puzzle for the Case Tier Hypothesis because nom is assigned twice. As shown in (10), both subject and predicative are nom. The puzzle can be solved if we assume that case assignment operates on argument structure. Semantically, the predicative is not an argument, but a predicate that takes the subject as its argument showing agreement with it (Hakulinen and Karlsson 1979: Sect. 9.4; Vilkuna 2000: Sect. 3.2.2; Matushansky 2008). The predicative clause is thus like an intransitive clause in that it only contains one argument. The two NPs are thematically of equal prominence because they are linked to one and the same argument. This captures the two occurrences of nom.

2.2 Classical Optimality Theory

In this section, we will develop an analysis of Finnish structural case in Optimality Theory and work out its predictions in simplex clauses. For a concise survey of work on case in Optimality Theory, see de Hoop (2009). The intuition behind our analysis can be stated as follows: case marking serves to distinguish the external argument from all other arguments; case marking should be avoided, if possible, especially on the external argument; and case marking proceeds in a cyclic fashion. These ideas are mostly traditional. In particular, the distinguishing function of case, the costliness of case marking, and the principle of the cycle are all familiar from earlier literature. Where we depart from the simplest imaginable theory built on these ideas is in the repeated reference to the external argument: it is the external argument that needs to be distinct from all other arguments, and it is the external argument that tolerates case marking the least. This departure will be empirically motivated in the course of the subsequent discussion. Our optimality-theoretic analysis states these informal intuitions in terms of the following three violable constraints:
  1. (11)

    a.

    *

    MarkedCase

    Do not case-mark an argument.

    b.

    *

    MarkedCase/e

    Do not case-mark an external argument.

    c.

     

    Uniqueness

    The external argument on the current cycle must be distinct in case from all other arguments outside the external argument.

     
We assume that nominative is the unmarked case; any other case violates *MarkedCase (*mc). For precedents, see, e.g., Legendre et al. (1993), Aissen (1999, 2003), for similar views outside Optimality Theory, see, e.g., Vainikka (1989, 1993), McFadden (2009), McFadden and Sundaresan (2011). Marking an external argument is particularly costly and violates both *MarkedCase and *MarkedCase/e (*mc/e). At the same time, Uniqueness (*uniq) requires the external argument to be distinct in case from all other arguments. For precedents, see, e.g., Wiik (1972), Hakulinen and Karlsson (1975), Toivainen (1993), T. Mohanan (1994), Anttila and Fong (2000), Wunderlich and Lakämper (2001), de Hoop and Malchukov (2008). Finally, we assume that case assignment is cyclic; for the same conclusion in a different framework, see Baker (2014). The cyclicity assumption will become relevant in complex clauses.
We now illustrate how this analysis accounts for case patterns in simplex clauses. First, consider transitive clauses. Inputs are argument structures that specify the number of NP arguments, the identity of the lexical verb, and the lexical/inherent cases assigned by the verb, if any. NP/e is an external argument; NP is an internal argument. Violations are marked by integers. The desired winner is marked by the symbol ☞.5
  1. (13)

    Transitive clause

    Open image in new window

     
The competition takes place between the unmarked nom object and the marked gen object. We are assuming that other marked cases such as elative (ela) and inessive (ine) are ruled out independently unless licensed by faithfulness to lexical/inherent case, a point to which we will return shortly. In tableau (13) candidates (c) and (d) can never win, no matter how the constraints are ranked, because they are harmonically bounded: (c) has a superset of the violations of (b); (d) has a superset of the violations of (a); see, e.g., McCarthy (2008): 80–83. Harmonically bounded candidates are highlighted by graying out the entire row. To rule out candidate (a) we need the ranking uniq ≫ *mc. In other words, it is more important for the external argument to be distinct in case from other arguments than it is to avoid case. The required ranking is shown by an arrow above the tableau. The ordering of *mc/e with respect to the other two constraints is irrelevant.
Required rankings are easy to find with the help of a comparative tableau (Prince 2002a, 2002b; McCarthy 2008: Ch. 2; Brasoveanu and Prince 2011) where all loser rows have their constraints labeled W for ‘favors the winner,’ L for ‘favors the loser,’ or are else left unlabeled, as shown in (14).
  1. (14)

    Transitive clause (comparative tableau)

    Open image in new window

     
A ranking makes the desired winner optimal if it guarantees that all Ls are dominated by some W. It is easy to see that the ranking uniq ≫ *mc is necessary to rule out (a). No other rankings are required. Harmonically bounded candidates (c) and (d) only contain winner-favoring constraints and thus require no ranking. Comparative labels are useful for making ranking arguments and we will use them occasionally for that purpose.6
Second, consider existential clauses. Following Kiparsky (2001), we take existential clauses to be intransitive clauses with an internal argument.
  1. (16)

    Existential clause

    Open image in new window

     
We assume that the inessive case on the locative NP is protected by an undominated constraint Max-Lex that requires faithfulness to lexical/inherent case, making it invariant under structural case assignment. Only the second NP is available for structural case. Since the constraint uniq is idle, this NP receives the unmarked nom. No rankings are needed; the ungrammatical candidate (b) is harmonically bounded.
Third, we turn to predicative clauses. Since the two NPs represent the same external argument the constraint uniq is idle and both NPs receive the unmarked nom. No rankings are needed; all ungrammatical candidates are harmonically bounded.
  1. (18)

    Predicative clause

    Open image in new window

     
This analysis of simplex clauses will serve as the starting point for our discussion. In the following sections, we will extend the analysis to variation in complex clauses.7

2.3 Partial Order Optimality Theory

Before turning to the variation facts we will lay out our assumptions about the place of variation in grammar. In Classical Optimality Theory (Prince and Smolensky 1993/2004) a grammar is a language-specific total order of universal constraints: every constraint is ranked with respect to every other constraint. Given our three constraints we have the six possible total orders shown in (19), together with their predictions for the simplex clauses. The three grammars compatible with Finnish are enclosed in a box.
  1. (19)

    The six total orders

    Open image in new window

     
Which of the total orders 1–3 is the actual grammar of Finnish? Since all of them get the facts right it might seem that the choice does not matter. The problem is that making any choice at all forces one to say more than is warranted. For example, the evidence at hand does not determine the mutual ranking of *mc/e and *mc, yet the total order assumption will force one to choose either *mc/e ≫ *mc or *mc ≫ *mc/e. While analysts are often content to choose any total order that works, the choice is at best arbitrary and at worst incorrect given more data.

What the evidence does show is that the ranking uniq ≫ *mc must be part of the grammar of Finnish. This ranking is shared by all the empirically correct total orders, but none of the empirically incorrect ones. We can say this and nothing more if we assume that the grammar of Finnish is the partial order {uniq ≫ *mc}, or equivalently, the set of total orders {1, 2, 3}.8 This is the key assumption of Partial Order Optimality Theory (see, e.g., Anttila 1997; Anttila and Cho 1998/2003; Djalali 2014), the version of Optimality Theory adopted in this paper.

Just as (19) visualizes the six possible total orders hidden in a set of three constraints, (20) visualizes the nineteen possible partial orders. The partial orders compatible with the Finnish data are enclosed in boxes.
  1. (20)

    The nineteen partial orders

    Open image in new window

     
The total orders appear at the bottom, numbered as in (19), together with their predictions for the simplex clauses (transitive, existential, predicative). The partial orders are arranged vertically by the subset relation: each mother grammar is the intersection of its daughter grammars.
There are two points of linguistic interest here. First, every partial order can be translated into a set of total orders.9 This means that we can view an individual’s grammar in two different ways: as a ranking relation, i.e., a set of ordered pairs of constraints, or as a set of total orders, i.e., a set of classical grammars. The choice between the two is a matter of convenience. We illustrate this in (21) for the six partial orders consistent with the Finnish data.
  1. (21)

    Two views of grammar

    Ranking relation

    Total orders

    {uniq ≫ *mc}

    {1, 2, 3}

    {uniq ≫ *mc, uniq ≫ *mc/e}

    {1, 2}

    {uniq ≫ *mc, *mc/e ≫ *mc}

    {1, 3}

    {uniq ≫ *mc, uniq ≫ *mc/e, *mc/e ≫ *mc}

    {1}

    {uniq ≫ *mc, uniq ≫ *mc/e, *mc ≫ *mc/e}

    {2}

    {uniq ≫ *mc, *mc/euniq, *mc/e ≫ *mc}

    {3}

     
Second, grammars are ordered by simplicity, defined as the amount of ranking information. The simplest grammar compatible with the Finnish data has one single ranking: {uniq ≫ *mc}. There are five other grammars that are also compatible with the data, but all contain more ranking information. Note that the amount of ranking information is inversely correlated with the number of total orders: the simplest grammar {uniq ≫ *mc} is compatible with largest number of total orders {1, 2, 3}. Of all the grammars compatible with the Finnish data, {uniq ≫ *mc} is the simplest. Assuming that simpler theories are better, all else being equal, this grammar must be our choice.
  1. (22)

    The grammar of Finnish structural case:

    uniq ≫ *mc  (inferred from simplex clauses)

     
The key empirical advantage of Partial Order Optimality Theory is its ability to seamlessly combine invariant and variable patterns. We will assume that an individual randomly selects a total order compatible with the partial order at the moment of performance and evaluates it in the standard optimality-theoretic fashion. The grammar {uniq ≫ *mc} = {1, 2, 3} correctly predicts no variation in simplex clauses: the outcome is the same, no matter which of the three total orders is selected. In contrast, consider the hypothetical grammar {*mc/e ≫ *mc} = {1, 3, 5}. It is easy to see from diagram (20) that this grammar predicts nomgen variation in transitive clauses, a pattern not found in Finnish simplex clauses.10
The goal of the analysis is to find the simplest partial order that correctly predicts the Finnish case patterns. We will start by examining the empirical generalizations in complex clauses, showing that in some environments the case pattern is invariant (either nom or gen) while in other environments it is variable (nomgen). We will also observe that in the variable environments there are preferences: some environments prefer nom, other environments prefer gen. This raises two important general questions:
  1. (23)

    Why does variation (nomgen) occur in some environments, but not in others?

     
  1. (24)

    Why do different variable environments prefer different cases (nom vs. gen)?

     
In the sections to come, we will answer these questions. We first identify the simplest partial order compatible with the invariant patterns, and then show that it also predicts the variable patterns, including the quantitative preferences among the variants. This finding provides support for an interesting generalization succinctly formulated by Bane (2011): categorical data can explain much of the variable data.

3 Itkonen structures

3.1 Empirical generalizations

Itkonen structures, traditionally known as referative constructions (Hakulinen et al. 2004, Sect. 538–542), are nonfinite propositional complements of verbs like ‘say’, ‘think’, ‘want’, and ‘believe’. These complement clauses are headed by a participle inflected for voice, aspect, and case. An example is given in (25).11 The subject of the nonfinite clause is gen. This gen is independent of Jahnsson’s rule and does not participate in the case alternations described above. Neither is it assigned by the matrix verb unlike in English where similar sentences involve Exceptional Case Marking (ECM); witness the fact that it does not alternate with nom when the matrix verb is passivized, as shown in (1) and (2), repeated below. What is the source of the gen on the subjects of nonfinite clauses? Vainikka (1989: 306, 1993: 138, 2011) argues that it is a structural default case for [Spec,XP] where X is a lexical category N, A, P, V, in this case [Spec,VP]. Based on evidence from binding, extraction, and adverb placement, Kiparsky (2010) argues that the gen subject must in fact be located higher, in [Spec,CP], and proposes that the gen is assigned by an empty complementizer in C. Both proposals share the view that this gen is associated with a particular syntactic position. Here we will assume that this gen is a structural case associated with the specifier of any nonfinite XP where X is N, A, P, V, or C. Crucially, it lies outside the case alternation system related to Jahnsson’s Rule. We derive this from a high-ranking constraint [Spec,−fin] = gen that requires gen in these structural positions. The “default” nature of this gen is evident from the fact that it is overridden by lexical/inherent case if one is required by the embedded participle (Vainikka 1993: 132, fn. 6). In terms of our analysis, this means that [Spec,−fin] = gen is crucially dominated by the constraint Max-Lex which requires faithfulness to lexical/inherent case.
The most striking fact about Itkonen structures is the long-distance interaction between matrix clauses and embedded clauses (see, e.g., Vainikka 2003; Vainikka and Brattico 2014). Particularly interesting are the variable patterns where either nom or gen is possible, but not to an equal degree, resulting in intermediate well-formedness judgments which depend on the input. In Kiparsky’s (2010) words: “With respect to case assignment [Itkonen structures] are neither fully opaque domains, nor totally transparent ones, but complexly translucent.” The variable patterns are illustrated below.12
  1. (28)
    In embedded transitive clauses
    1. a.

      If the matrix verb is active, the object is gen.

       
    2. b.

      If the matrix verb is passive, the object varies nomgen.

       
     
In addition to case variation, (30) also illustrates a word order generalization: if there is no matrix subject, as in passive matrix clauses, some postverbal constituent is usually fronted to the left of the matrix verb, reflecting a stylistic preference by which canonical declarative sentences tend to start with a nominal constituent (Hakulinen and Karlsson 1979: 303–305; Vilkuna 1989: 40). The fronted constituent may be the embedded subject, here Matin ‘Matti-gen’. The process appears to have no effect on structural case.13
Next, consider embedded existential clauses. Here we get variation under both actives and passives, but nom and gen are not on an equal footing. The basic pattern is stated in (31) and illustrated in (32) and (33).
  1. (31)
    In embedded existential clauses
    1. a.

      If the matrix verb is active, the NP prefers gen.

       
    2. b.

      If the matrix verb is passive, the NP prefers nom.

       
     
Finally, we turn to embedded predicative clauses. Again, we get variation, but with different preferences. The basic pattern is stated in (34) and illustrated in (35) and (36).
  1. (34)
    In embedded predicative clauses
    1. a.

      If the matrix verb is active, the predicative prefers nom.

       
    2. b.

      If the matrix verb is passive, the predicative strongly prefers nom.

       
     

3.2 A cyclic analysis

Any analysis of the above data must somehow accommodate the systematic patterns of variation. In some environments the case pattern is invariant (either nom or gen); in other environments it is variable (nomgen). Our strategy will be to first identify the simplest partial order compatible with the invariant patterns. We will then show that this partial order also predicts the variable patterns, including the quantitative preferences documented by Itkonen (1976, 1981), with no adjustments or modifications.

3.2.1 Active matrix clauses

We start by considering case patterns under active matrix clauses. In (37) we have an active matrix clause combined with an embedded existential clause: What does our current analysis predict? uniq requires the external argument (Pekka) in the matrix clause to be distinct from the internal argument (karhu-n) in the embedded clause. The expected outcome is gen, as in the simplex transitive clause Matti ampui karhu-n ‘Matti.nom shot a/the bear-gen’ in (13). But why is nom (karhu) possible at all? In his article on case variation in complex clauses, Itkonen (1976: 77) makes an insightful suggestion: nom appears on the strength of analogy with simplex clauses. As shown in (38), in the simplex clause we have karhu ‘bear.nom’. According to Itkonen, it is the nominative in the simplex clause that is responsible for the optional appearance of karhu ‘bear.nom’ in the complex clause.14 Empirically Itkonen’s generalization seems to be exactly on the mark. Theoretically it provides evidence for the cycle: nom occurs in the simplex clause for transparent reasons and is optionally inherited into the complex clause because the derivation is cyclic. Following Stratal Optimality Theory (Kiparsky 2000), we implement the cycle in terms of Input-Output Faithfulness. The relevant constraints are stated in (39):
  1. (39)

    Faithfulness constraints:

    a.

    max/np

    ‘No case deletion in an NP’

    b.

    dep/np

    ‘No case insertion in an NP’

     
Faithfulness (dep/np) favors nom (karhu) inherited from the first cycle; markedness (uniq) favors gen (karhu-n) because it keeps the arguments distinct. As we will see shortly, in a system like Finnish where constraint ranking is partial the outcome is variation: gen respects markedness, nom respects faithfulness.
We now turn to the gen case on the participle.15 In Finnish, matrix verbs commonly mark their infinitival and participial complements with case. For example, verbs like ‘ask’ and ‘forbid’ mark the infinitive heading their VP-complement with the semantically appropriate local case (Fong 1997a, 1997b). Here we posit a general constraint that requires all nominal heads of a complement clause to be case-marked, i.e., to be non-nominative. The constraint is stated in (40); for precedents, see Aissen (2003: 447) and Anttila and Kim (2011). The formulation is intentionally general: the constraint applies to all nominal heads within the complement clause that are capable of case inflection, NPs as well as the participle.
  1. (40)

    *\(\varnothing_{\mathrm{C}}\)(cp) ‘All nominal heads in a complement clause must be case-marked.’

     
The tableau in (41) shows the constraint violation profiles for all six constraints. The cases inherited from the first cycle are shown in the input. Recall that the inessive case ine is protected by the undominated constraint Max-Lex which requires faithfulness to lexical/inherent case, ruling out all candidates where ine has been replaced by some other case, say, nom or gen. The grammatical candidates (a) and (b) are marked with the symbol ☞.
  1. (41)

    Embedded existential

    Open image in new window

     
Only candidates (a), (b), and (d) are viable; the rest are harmonically bounded. In particular, *\(\varnothing_{\mathrm{C}}\)(cp) renders harmonically bounded all the candidates (e)–(h) where the participle is nom (oleva). This is because there always exists a candidate with an identical violation profile except that it does better on *\(\varnothing_{\mathrm{C}}\)(cp). We can therefore omit candidates with a nom participle from all subsequent tableaux.
The analytical task is to find the simplest partial order that includes the grammatical (a) and (b), but excludes the ungrammatical (d), while maintaining the correct patterns in simplex clauses. With three constraints the simplest partial order was easy to find by visual inspection; with six constraints we need a computer. The simplest grammar compatible with all the data is shown in (42). This grammar was found with the help of OTOrder (Djalali and Jeffers 2015), a web application for working with Partial Order Optimality Theory. Only one additional ranking is needed for Itkonen structures: *mc/edep/np. The tableau updated with the new ranking is shown in (43).
  1. (42)

    The grammar of Finnish structural case (updated version)

     

    uniq ≫ *mc

    (inferred from simplex clauses)

    *

    mc/edep/np

    (inferred from Itkonen structures)

     
  1. (43)

    Embedded existential

    Open image in new window

     
The grammar correctly rules out the ungrammatical (d), but predicts variation (a) ∼ (b). How exactly does that happen? We first note that (d) loses because the new ranking guarantees that it will always lose against (a), as shown in (44):
  1. (44)

    Embedded existential: (a) vs. (d)

    Open image in new window

     
The variation (a) ∼ (b) arises in the following way. The grammar has six constraints, but only two rankings. This means that it contains all the total orders that respect the rankings uniq ≫ *mc and *mc/edep/np. There are 180 such total orders. Tableau (45) shows the competition between (a) and (b).
  1. (45)

    Embedded existential: (a) ∼ (b)

    Open image in new window

     
Let us assume that at the moment of performance the speaker randomly chooses a total order consistent with this partial order (Kiparsky 1993; Anttila 1997; Anttila and Cho 1998/2003; Riggle 2010). The winner will be (a) (= gen) or (b) (= nom), depending on the speaker’s choice. It is easy to see from (45) that only four constraints matter: dep/np and *mc are violated by (a) and uniq and *\(\varnothing_{\mathrm{C}}\)(cp) are violated by (b). It is the mutual ranking of these four constraints that decides the outcome. For example, (a) wins if uniq ranks highest and (b) wins if dep/np ranks highest. The outcome will thus sometimes be gen, sometimes nom, depending on the speaker’s choice of total order. In other words, the grammar predicts variation.
The grammar also correctly handles embedded transitive clauses. An example is given in (46), repeated from (29). We start from the full set of candidates. The input contains three NPs, each of which must choose between nom and gen. This yields the eight candidates in (47).
  1. (47)

    Embedded transitive: High-ranking constraints

    Open image in new window

     
Max-Lex is inactive since the input contains no lexical/inherent cases. [Spec,−fin]= gen rules out the candidates (e)–(h) where the subject of the embedded nonfinite clause is nom. This leaves the candidates (a)–(d). It now falls upon the remaining six constraints notated “…” in tableau (47) to select the correct output (a).
  1. (48)

    Embedded transitive: Lower-ranking constraints

    Open image in new window

     
This tableau shows that candidate (a) correctly wins. Candidate (b) is ruled out by the familiar ranking uniq ≫ *mc; candidates (c) and (d) are harmonically bounded. Note that candidate (a) does not violate uniq despite having two instances of gen. This is because uniq only applies to the external argument on the current cycle.
Finally, the grammar correctly predicts variation in embedded predicatives. All ungrammatical candidates are harmonically bounded.
  1. (50)

    Embedded predicative

    Open image in new window

     
We now summarize our analysis of case under active matrix clauses. uniq strives to distinguish the external argument from all other arguments. This favors gen in embedded clauses of all kinds. However, embedded clauses differ in terms of faithfulness. In transitive clauses the faithful choice is gen: markedness and faithfulness agree and there is no variation. In existential and predicative clauses the faithful choice is nom: markedness and faithfulness conflict, which yields variation reflecting the speaker’s free choice among the available total rankings at the time of performance.

3.2.2 Passive matrix clauses

Next, we turn to passive matrix clauses. The key difference is that passives have no external argument and hence there is no pressure to distinguish arguments. In embedded transitive clauses we have a conflict: markedness favors nom; faithfulness favors gen. This results in variation.
  1. (53)

    Embedded transitive

    Open image in new window

     
There are only two competitive candidates: [gen gen] and [gen nom]. The other possible candidates *[nom gen] and *[nom nom] with a nom-subject in the embedded nonfinite clause are ruled out by the high-ranking [Spec,−Fin] = gen as illustrated above. For this reason they have been omitted from the tableau. Note that the embedded subject Matin is optionally fronted to the left of the matrix verb. This does not make it part of the matrix cycle: we assume that optional stylistic fronting is invisible to cyclic evaluation. Since the matrix clause has no external argument uniq is not violated.
The situation is similar in embedded existential and predicative clauses. This time faithfulness favors nom in the embedded clause, but gen is also attested due to partial ranking as shown below in (56) and (59). These examples also provide independent evidence for *\(\varnothing_{\mathrm{C}}\)(cp): without it gen would be harmonically bounded and variation would not be possible.
  1. (56)

    Embedded existential

    Open image in new window

     
  1. (59)

    Embedded predicative

    Open image in new window

     
The analysis now covers both simplex clauses and Itkonen structures. It consists of six constraints and two rankings. The rankings were based on categorical judgments and were needed to rule out plainly ungrammatical candidates. We noted that in some environments the analysis predicts variation, but stopped short of trying to explain the intermediate well-formedness judgments among the variants. We will now take a closer look at the variation and the preferences among the variants.

3.3 Variation in complex clauses

The case variation in complex clauses was explored in a remarkable experimental study by Itkonen (1976, 1981). Itkonen’s strategy was to vary the syntactic context in order to see how that affects the choice of case. He constructed 28 stimulus sentences where the case of the second NP in the complement clause was left open and then presented the stimuli to 128 native speakers asking them to choose between nom and gen based on their intuitive judgment (kielikorva ‘language ear’). The subject population consisted of 38 first-year students majoring in Finnish and 90 students majoring in other subjects. The difference between the two groups turned out marginal (Itkonen 1981: 105). Itkonen’s 28 stimuli can be divided into three groups:
  1. (60)
     

    Matrix clause

    Embedded clause

    Sentences

    Stimuli

    1.

    Active/Passive

    Existential/Predicative

    4

    16

    2.

    Active/Passive

    Existential

    3

    6

    3.

    Active/Passive

    Other

    3

    6

     
Group 1 consists of four minimal quadruplets: the matrix clause can be active or passive; the embedded clause can be existential or predicative; and there are four distinct sentences. This yields 2 × 2 × 4 = 16 stimuli in all. Group 2 consists of three minimal pairs: the matrix clause can be active or passive; the embedded clauses are all existential; and there are three distinct sentences. This yields 2 × 1 × 3 = 6 stimuli in all. Group 3 is more heterogeneous and only reported on in the revised and expanded version of the original study (Itkonen 1981). It consists of three minimal pairs: the matrix clause can be active or passive and the three embedded clauses are all different in structure (transitive passive, transitive infinitival passive, permissive). This yields 2 × 3 = 6 stimuli in all. The design yields plenty of information about the active vs. passive contrast which is relevant in all 28 stimuli (14 actives, 14 passives). It gives somewhat less information about the existential vs. predicative contrast which is only relevant in 16 stimuli (8 existentials, 8 predicatives).
Let us consider some illustrative examples. In (61), we have a minimal quadruplet from Group 1 that combines matrix active vs. passive with embedded existential vs. predicative. For each example, the subjects were asked to report whether they preferred nom or gen in the NP highlighted in bold. Itkonen aggregated the results across subjects and reported them as percentages. In (62) we see the outcome from his four minimal quadruplets. Groups 2 and 3 were analogous.
  1. (62)

    Itkonen’s (1976, 1981) results for four minimal quadruplets (Group 1)

    Open image in new window

     
  1. (63)
    Generalizations:
    1. a.

      gen is more common under actives than under passives.

       
    2. b.

      gen is more common in existentials than in predicatives.

       
    3. c.

      The profiles of the four sentences are similar, but not identical.

       
     
These generalizations are consistent with the intermediate well-formedness judgments reported in Sect. 3.1. They can be summarized as follows: (i) gen is better under actives than under passives; (ii) gen is better in transitives (not included in (62)–(63)) than in existentials and better in existentials than in predicatives. The variation is clearly not free, but structured in a way that requires a grammatical explanation.
Itkonen also observed that individuals differed systematically in their case preferences: some favored gen, others favored nom. For example, if subjects favored nom in predicatives, they also favored nom in existentials under both actives and passives (Itkonen 1981: 110–111). Itkonen showed this by dividing his subjects into two groups: Group A who favored nom in predicatives (63 subjects) and Group B who did not (63 subjects).16 He then asked how these two groups behaved with respect to existential clauses. He found that the individual preferences carried over from predicatives to existentials: Group A had lower counts for gen than Group B for all seven stimulus sentences, for the actives in (64) as well as the passives in (65). Comparing the diagrams one can also see that the passive counts are systematically lower than the active counts.
  1. (64)

    gen % in embedded existentials under active matrix clauses. Group A favored nom in embedded predicatives (Itkonen 1981: 110).

    Open image in new window

     
  1. (65)

    gen % in embedded existentials under passive matrix clauses. Group A favored nom in embedded predicatives (Itkonen 1981: 110).

    Open image in new window

     
Itkonen provided similar evidence for actives and passives: if subjects favored gen under an active matrix verb, they also favored gen under a passive matrix verb (Itkonen 1981: 110–111). The effect was also found to hold in the reverse direction.

Itkonen interpreted these facts as evidence for syntactic analogy, proposing that the surface case patterns of particular constructions (e.g., existentials and predicatives, actives and passives) influence one another (Itkonen 1977: 86–87). An alternative interpretation is that individuals have a general preference for a particular case that is reflected across the board, with differences arising from the special properties of each construction. Under this view, one would expect a single change in the grammar of case to be reflected globally, affecting every construction in the same way modulo their special properties (see, e.g., Kroch 1989). This parametric view of change is common among practitioners of generative grammar and it is the view we will adopt here.

How did this situation arise historically? Itkonen noted that the variation is the result of an ongoing diachronic change: gen is receding and nom is gaining ground. Crucially, the spread of nom has followed a structured path, starting from predicatives and passive matrix clauses and moving on to existentials and active matrix clauses. According to Setälä’s (1901: 109) grammar, predicatives used to be gen under active matrix clauses. In the literary language of the late 1930’s, nom was still rare in this environment, but was becoming common under passive matrix verbs (Lehikoinen 1973, cited in Itkonen 1981: 92, 106, 112). At the time of Itkonen’s writing, nom had almost completely won out in predicatives, but existentials showed variation resembling the predicative situation in the 1930’s. The case pattern had thus been changing simultaneously in both constructions, with existentials lagging a generation behind the predicatives. The important question is why the change gen > nom occurred in this order. Why did predicatives change before existentials and passives before actives? Is there something necessary about this order or could it have been different? We will shortly see that this diachronic order falls out from our synchronic analysis.

3.4 Deriving variation from grammar

Grammatical analysis usually starts from the clear cases. Some expressions are clearly grammatical and should be predicted; other expressions are clearly ungrammatical and should be excluded. The question is what to do with the intermediate cases. One view is expressed in the following well-known passage:

[W]e may assume for this discussion that certain sequences of phonemes are definitely sentences, and that certain other sequences are definitely non-sentences. In many intermediate cases we shall be prepared to let the grammar itself decide, when the grammar is set up in the simplest way so that it includes the clear sentences and excludes the clear non-sentences. This is a familiar feature of explication. (Chomsky 1957: 14)

We will now see that our current grammar already derives the quantitative asymmetries in the intermediate cases studied by Itkonen. This is an instance of letting the grammar itself decide: given the simplest analysis of categorical cases the grammar automatically delivers the correct quantitative preferences in cases of variation, with no adjustments or modifications, providing strong support for the overall analysis.
The key observation is that some winners entail other winners (Prince 2002a, 2002b, 2007; Anttila and Andrus 2006). In order to see what that means, let us reconsider the inputs act[exist] and act[pred] discussed in (41) and (50). Both matrix clauses are active, but the embedded clauses are different: existential in the first, predicative in the second. The predicted outcome in both constructions is variation nomgen. All other candidates are ruled out as ungrammatical.
  1. (66)

    The predicted variants in act[exist] and act[pred]

     

    act[exist]

    act[pred]

    (a)

    nom [inegen]

    nom [gengen]

    (b)

    nom [inenom]

    nom [gennom]

     
The variation is localized in the last noun phrase which varies nomgen. Both variants are found for both inputs, but the variants turn out to have a very different standing in each construction. We can see this by asking what rankings must hold in order for nom (= (b)) to be optimal for each input. Let us look at the comparative tableaux in (67) and (68), arbitrarily choosing nom (= (b)) as the winner. No rankings are assumed.
  1. (67)

    Active matrix, embedded existential

    Open image in new window

     
  1. (68)

    Active matrix, embedded predicative

    Open image in new window

     
Recall that a desired winner is optimal if and only if all loser-favoring constraints (L) are dominated by some winner-favoring constraint (W). Looking at (67) we can see that (b) wins if both uniq and *\(\varnothing_{\mathrm{C}}\)(cp) are dominated by either dep/np or *mc. Looking at (68), we can see that this automatically makes (b) win here as well because the comparative rows are identical except for the addition of one W (*mc/e). The upshot is that nom in an existential clause entails nom in a predicative clause.17
To take another example, let us reconsider the inputs act[exist] and pass[exist] discussed in (41) and (56). This time, both embedded clauses are existential, but the matrix clauses are different: active in the first, passive in the second. Again, the predicted outcome in both constructions is variation nomgen. All other candidates are ruled out as ungrammatical.
  1. (69)

    The predicted variants in act[exist] and pass[exist]

     

    act[exist]

    pass[exist]

    (a)

    nom [inegen]

    [inegen]

    (b)

    nom [inenom]

    [inenom]

     
The variation is localized in the last noun phrase which varies nomgen. Both variants are found for both inputs, but again, the variants turn out to have a very different standing in each construction. We can see this by asking what rankings must hold in order for nom (= (b)) to be optimal for each input. Let us look at the comparative tableaux in (70) and (71), arbitrarily choosing nom (= (b)) as the winner. No rankings are assumed.
  1. (70)

    Active matrix, embedded existential

    Open image in new window

     
  1. (71)

    Passive matrix, embedded existential

    Open image in new window

     
Looking at (70), we can see that (b) wins if both uniq and *\(\varnothing_{\mathrm{C}}\)(cp) are dominated by either dep/np or *mc. Looking at (71), we can see that this automatically makes (b) win here as well because the comparative rows are identical except for the subtraction of one L (uniq). The upshot is that nom in an active clause entails nom in a passive clause.18
Entailments structure the space of variation in ways that result in quantitative asymmetries. This becomes evident by inspecting the predicted typology. We illustrate this in (72) for four input constructions: active vs. passive matrix clauses with existential vs. predicative complements. Our partial order predicts six possible invariant dialects. The last noun phrase (object or predicative) alternates between nom and gen depending on dialect and construction. We have shaded gen for perspicuity. The typology was computed using OTSoft (Hayes et al. 2003).
  1. (72)

    The six possible dialects

    Open image in new window

     
Each row represents a distinct dialect: 1 is a conservative dialect with gen in all four constructions; 6 is an innovative dialect with nom in all four constructions. 2–5 are intermediate dialects that have nom or gen depending on the construction. All six fall under the competence of a speaker of modern Finnish because all belong to the partial order that constitutes her grammar. Under this view, a single speaker controls six synchronic slices of the diachronic development that has been gradually replacing gen by nom over the past century. This view of grammar is sociolinguistically plausible. An individual acquiring Finnish under normal conditions is typically exposed to variation across generations as well as across registers, including the written standard that tends to be conservative as well as spoken colloquial registers that tend to be innovative. However, taken together, these dialects are not a random collection, but a grammatically natural class that constitute a single grammar characterized as a partial order.

This view of grammar lends itself to a straightforward quantitative interpretation. Recall our assumption that at the moment of performance the speaker randomly selects a total order from the partial order that constitutes her grammar. A number of quantitative predictions then follow. For example, gen is more likely to be selected with the input act[exist] than with the input act[pred]. This is because the total orders that predict gen for the former (dialects 1–5) are a superset of the total orders that predict gen for the latter (dialects 1–3). This correctly predicts that gen should be more frequent in act[exist] than in act [pred].

The entailments hidden in the typology (72) are easier to see if we visualize them as the directed graph in (73). Each entailment is depicted by an arrow. The graph was computed and drawn by T-order generator (Anttila and Andrus 2006).
  1. (73)

    Entailments among four variable Itkonen structures: the case of chaplain

    act[exist]

    ‘I have thought there to be a chaplain in every parish.’

    pass[exist]

    ‘There was thought to be a chaplain in every parish.’

    act[pred]

    ‘At least I thought that man to be a chaplain.’

    pass[pred]

    ‘In some villages that man was thought to be a chaplain.’

    1. a.

      Case patterns where the second noun phrase is gen

      Open image in new window

       
    2. b.

      Case patterns where the second noun phrase is nom

      Open image in new window

       
     
To understand these entailment graphs, consider the following examples. The first graph shows that if the alternating case is gen in pass[pred] it must also be gen in all the other three input structures: pass[exist], act[pred], and act[exist]. The graph further shows that if the alternating case is gen in act[pred] it must also be gen in act[exist], but nothing is entailed about the case under passives which may be either nom or gen. Finally, there can be no dialect with gen only under passives (i.e., in pass[pred] and pass[exist]) because this would entail gen under the corresponding actives as well. Similar entailments for nom can be read off the second graph. Since there are only two possible outcomes (nom, gen) the graphs are mirror images of each other.
The graphs in (73) are theoretical predictions derived from the grammar. The corresponding empirical data from Itkonen’s (1976, 1981) four minimal quadruplets are shown in (74).
  1. (74)

    Percentages of gen in Itkonen’s (1976, 1981) four minimal quadruplets

     

    act[exist]

    pass[exist]

    act[pred]

    pass[pred]

    1.

    91 %

    36 %

    33 %

    28 %

    2.

    66 %

    32 %

    29 %

    20 %

    3.

    98 %

    54 %

    30 %

    17 %

    4.

    75 %

    48 %

    36 %

    16 %

     
Comparing the predictions in (73) to the observations in (74) reveals three kinds of patterns. First, the percentage of gen is highest in act[exist] and lowest in pass[pred] in all four quadruplets, with pass[exist] and act[pred] falling in the middle. These quantitative asymmetries are correctly predicted by the entailments: the total orders that predict gen for act[exist] (dialects 1–5) are a superset of those that predict gen for any other input, and conversely, the total orders that predict gen for pass[pred] (dialect 1) are a subset of those that predict gen for any other input. Second, gen is systematically (slightly) more frequent in pass[exist] than in act[pred] across all four quadruplets. This is not predicted by the entailments as shown by the absence of an arrow between the two nodes. Is there any grammatical reason to expect this quantitative asymmetry or does it simply reflect external factors, such as social factors or construction frequency (see, e.g., Jarosz 2010)? We will return to this question in Sect. 4.3 after we have discussed Ikola structures. Third, the percentages vary widely across the quadruplets for reasons that remain a mystery. For example, the percentage of gen in act[exist] is 91 % in the first quadruplet, but only 66 % in the second. Is it possible to identify any grammatical reason for such sentence-specific differences? Again, we will put this question on hold and will return to it in Sect. 4.4 once we have discussed Ikola structures.
Finally, grammatical entailments hold in the domain of historical change. Our analysis predicts that the change gen > nom had to start from matrix passives and embedded predicatives before spreading to matrix actives and embedded existentials. To see this, consider the predicted case patterns for act[exist] and pass[exist]:
  1. (75)

    Three predicted dialects

     

    act[exist]

    pass[exist]

     

    1.

    nom [inegen]

    [inegen]

    gen in both

    2.

    nom [inegen]

    [inenom]

    alternation

    3.

    nom [inenom]

    [inenom]

    nom in both

    4.

    nom [inenom]

    [inegen]

    impossible, not predicted

     
Dialect 1 has gen in both constructions, i.e., the change gen > nom has not yet occurred. Dialect 2 has gen in act[exist], but nom in pass[exist], i.e., the change has taken place under passives resulting in a synchronic case alternation. Dialect 3 has nom in both constructions, i.e., the change has gone to completion. Modern Finnish is a mixture of these three types of dialects, with free variation among them. Crucially, Dialect 4 where act[exist] has nom, but pass[exist] has gen is excluded as grammatically ill-formed. It is precisely such a dialect that would have to arise if the change were to start from actives: it would provide the step where nom is possible under actives, but not under passives. This is a synchronically impossible state that violates one of the entailments. For this reason it is not a possible step along the path of change either. Under this view, all sorts of diachronic changes are possible, but the synchronic grammar filters out those that are not grammatically sustainable. This guarantees that language change remains grammatically benign (see, e.g., Kroch 1989; Kiparsky 2006).

We conclude with a methodological note. In variationist linguistics, it is common practice to start by excluding invariant data. Once the “envelope of variation” has been identified the analysis of variation proper begins, usually by fitting a statistical model to the variable data. Here we took the opposite approach. We started by finding the simplest grammar for the invariant patterns and discovered that variation and quantitative patterns emerged as a side-effect. Excluding invariant data would have been a bad move: it would have amounted to throwing out information that explains the structure of variation. It seems that there is much to be gained by studying variation against the backdrop of the invariant structure of the language instead of focusing on variation in isolation.

4 Ikola structures

4.1 Empirical generalizations

We now turn to Ikola structures that exhibit case variation similar to Itkonen structures, but with intriguing differences. Examples are given in (76) and (77): Ikola structures are NPs with a nonfinite VP complement. The embedded NP is usually nom, but nomgen variation is possible under limited circumstances. The two sentence types above were chosen for illustration because together they cover 97 % of all the gen variants in our corpus of 1,577 Ikola structures extracted from the Aamulehti 1999 corpus; see Appendix A for details. The examples differ in the matrix object: (76) has the genitive case; (77) has an oblique case. In both sentence types nom and gen are robustly attested, with nom being about twice as common as gen.
  1. (78)

    gen vs. nom in Ikola structures (raw counts)

    Open image in new window

     
Case variation is possible if the following three conditions are simultaneously satisfied (see also Hakulinen et al. 2004: Sect. 940):
  1. (79)
    The variation environment
    1. a.

      the matrix clause is active, i.e., has a nominative subject, and

       
    2. b.

      the embedded VP is transitive, i.e., the NP is an internal argument, and

       
    3. c.

      the matrix NP (= Ikola structure itself) is not an external argument.

       
     
First, the matrix clause must have a nominative subject in order for variation to occur. As shown in (80), variation is blocked under passives: Second, the embedded NP must be an internal argument in order for variation to occur. In (81) the embedded NP sankari ‘hero’ is a predicative that takes the subject Matti as its argument. No variation is possible. The tables in (82) contrast the case patterns in Itkonen and Ikola structures under matrix active vs. passive (vertical dimension), with embedded transitive vs. predicative (horizontal dimension).
  1. (82)
    1. a.

      Itkonen structures

       

      Transitive

      Predicative

      Active

      gen

      nomgen

      Passive

      nomgen

      nomgen

       
    2. b.

      Ikola structures

       

      Transitive

      Predicative

      Active

      nomgen

      nom

      Passive

      nom

      nom

       
     
The case patterns differ in all four contexts, but entirely systematically: in both structures active and transitive favor gen, passive and predicative favor nom. The difference is that Ikola structures show a stronger preference for nom across the board. Both the similarities and the differences require an explanation.
Finally, there is no variation if the Ikola structure itself is an external argument (NP/e): the outcome is an invariant nom. This option is not possible in Itkonen structures. The generalization holds no matter whether the structure is a subject as in (83) or a predicative as in (84).19 To summarize, variation in Ikola structures is limited to examples like (76) and (77).20 The variation appears to be a recent innovation: early 20th century grammarians (Setälä, Saarimaa) only recognized gen as a possibility, but a generation later Ikola (1964: 65–66) reported variation. This suggests that the change gen > nom observed in Itkonen structures had also been at work in Ikola structures over the same period of time.

4.2 Analysis

The key difference between Itkonen and Ikola structures is the category of the first cycle: in Itkonen structures it is a CP, in Ikola structures it is an NP. In the Itkonen structure (85) the predicate ampuneen ‘shot’ has two arguments: Matti and karhu ‘bear’. In the Ikola structure (86) the predicate ampua ‘shoot’ only has the internal argument karhu ‘bear’. This difference in number of arguments (two vs. one) plays a crucial role in explaining the differences in case patterns, as we will see shortly. In Ikola structures the embedded NP always gets nom on the first cycle, no matter whether it is an internal argument as in tilaisuus ampua karhu ‘opportunity to shoot a bear’ or an external argument as in tilaisuus olla sankari ‘opportunity to be a hero.’ Case distinctions are unnecessary because there is only one argument: uniq is idle and the unmarked nom wins. This is shown in (87).
  1. (87)

    1st cycle: ‘opportunity to shoot a bear (NP)’, ‘opportunity to be a hero (NP/e)’

    Open image in new window

     
Let us now embed the output of the first cycle in a larger context. If the matrix clause is active (i.e., has an external argument) and the embedded NP contains a transitive predicate (i.e., has an internal argument) the result is nom gen [nomgen] where the embedded internal argument varies in case. Variation arises from conflicting pressures: faithfulness favors nom (karhu); markedness strives to distinguish Matti from other arguments, favoring gen (karhu-n). Since the ranking is partial the outcome is variation. The tableau in (89) shows that only four candidates are viable; the rest are harmonically bounded. Note that the constraint *∅C(cp) is not violated because the complement is an NP, not a CP.
  1. (89)

    2nd cycle: ‘Matti (NP/e) got an opportunity (NP) to shoot a/the bear (NP).’

    Open image in new window

     
The analytical task is to find the simplest partial order that includes the grammatical (c) and (e), but excludes the ungrammatical (a) and (d), while maintaining the correct patterns in simplex clauses and Itkonen structures. This can be done with the help of OTOrder (Djalali and Jeffers 2015). Only one additional ranking is needed for Ikola structures: *mc/e ≫ *mc. The grammar contains 150 total orders.21
  1. (90)

    The grammar of Finnish structural case (final version)

     

    uniq ≫ *mc

    (inferred from simplex clauses)

    *

    mc/edep/np

    (inferred from Itkonen structures)

    *

    mc/e ≫ *mc

    (inferred from Ikola structures)

     
This ranking correctly rules out the ungrammatical candidates. Candidate (a) loses against candidate (c) by the familiar ranking uniq ≫ *mc:
  1. (91)
     
Candidate (d) loses against candidate (e) by another familiar ranking *mc/edep/np and the new ranking *mc/e ≫ *mc:
  1. (92)
     
The absence of variation in embedded predicatives is also correctly predicted:
  1. (94)

    2nd cycle: ‘Matti (NP/e) got an opportunity (NP) to be a hero (NP/e).’

    Open image in new window

     
As in simplex predicatives like Matti on sotilas ‘Matti is a soldier’, here the matrix subject ‘Matti’ and the embedded predicative ‘hero’ represent the same external argument which is part of the current cycle. Since case is assigned to arguments, not to NPs, we assume that uniq is violated if either NP/e representing this external argument is identical in case to the internal argument ‘opportunity’. The ungrammatical candidate (a) loses against (c) by the familiar ranking uniq ≫ *mc.
Under passives, Ikola structures show no variation: As shown in (96), this is correctly predicted: uniq is idle because there is no NP/e on the current cycle and both faithfulness and markedness prefer nom. All ungrammatical candidates are harmonically bounded. In this example the embedded NP is an internal argument; the prediction is similar for embedded predicatives.
  1. (96)

    2nd cycle: ‘An opportunity (NP) was obtained to shoot a/the bear (NP).’

    Open image in new window

     
Finally, consider examples where the Ikola structure itself is an external argument (subject, predicative): In both examples the internal argument karhu ‘bear’ is embedded inside an external argument. The result is an invariant nom. This suggests that within an external argument Uniq does not play its usual distinguishing role: arguments embedded inside an external argument are invisible to Uniq. This is reflected in our definition of Uniq given in (11): the external argument on the current cycle must be distinct in case from all other arguments outside the external argument. The details are shown in tableaux (99) and (100). All losers are harmonically bounded because the winner has no violations.
  1. (99)

    2nd cycle: ‘The opportunity (NP/e) to shoot a/the bear (NP) presented itself.’

    Open image in new window

     
  1. (100)

    2nd cycle: ‘It (NP/e) was an opportunity (NP/e) to shoot a/the bear (NP).’

    Open image in new window

     
Finally, note that in (100) the subject ‘it’ and the predicative ‘opportunity’ do not incur uniq violations because they represent the same external argument.

4.3 Interim summary

The predictions of our analysis are summarized in Appendix B. The six constructions where variation is predicted to be possible are shown in (101). Five of them are Itkonen structures, one is an Ikola structure.
  1. (101)

    Predicted cases of variation

    Input

    Case pattern

    Example

    act[pred]

    nom [gennomgen]

    Pekka uskoi Mati-n olevan sotilas∼sotilaa-n.

    ‘Pekka believed Matti to be a soldier.’

    act[exist]

    nom [inenomgen]

    Pekka uskoi metsä-ssä olevan karhu∼karhu-n.

    ‘Pekka believed there to be a bear in the forest.’

    pass[trans]

    [gennomgen]

    Mati-n uskottiin ampuneen karhu∼karhu-n.

    ‘Matti was believed to have shot a bear.’

    pass[pred]

    [gennomgen]

    Mati-n uskottiin olevan sotilas∼sotilaa-n.

    ‘Matti was believed to be a soldier.’

    pass[exist]

    [inenomgen]

    Metsä-ssä uskottiin olevan karhu∼karhu-n.

    ‘There was believed to be a bear in the forest.’

    act-NP[trans]

    nom gen [nomgen]

    Matti sai tilaisuude-n ampua karhu∼karhu-n.

    ‘Matti got an opportunity to shoot a bear.’

     
Our optimality-theoretic analysis does what any analysis must do: it declares some case patterns grammatical and others ungrammatical. It goes beyond this baseline in two respects. First, it predicts the environments where variation is possible, shown in (101). Second, it predicts particular quantitative preferences for nom and gen across the variable environments. These predictions arise because case patterns are connected by entailments and cannot vary or change independently of one another. Quantitative preferences in one construction are systematically related to quantitative preferences in another construction by the theory.
The entailments among the six variable patterns are summarized in (102). In addition to entailments, the graph also shows the ranking volume (rv) of each pattern: this is the number of total orders under which this particular pattern wins. To save space, we have left out case patterns with the ranking volume of zero, i.e., patterns that never win under any total ranking, and patterns with the ranking volume of 150, i.e., patterns that win under every total ranking and consequently show no variation. The graph was computed and visualized using OTOrder (Djalali and Jeffers 2015).
  1. (102)
    Entailments among variable patterns (both Itkonen and Ikola structures)
    1. a.

      Case patterns where the second noun phrase is gen

      Open image in new window

       
    2. b.

      Case patterns where the second noun phrase is nom

      Open image in new window

       
     
What we have here is a system of implicational laws in the sense of Greenberg (1963). The location of a pattern inside the graph indicates its markedness. Patterns high in the graph have small ranking volumes, i.e., are generated by few total orders. They are more marked and therefore less frequent. Patterns low in the graph have large ranking volumes, i.e., are generated by many total orders. They are less marked and therefore more frequent.

These grammatical entailments hold true in the quantitative data with no exceptions.22 A reviewer points out that the model also correctly predicts the gen variant to be more frequent in pass[exist] than in act[pred], a systematic pattern we observed in Itkonen’s data. This prediction is not an entailment, but based on the relative ranking volumes of gen in pass[exist] (rv = 84) and act[pred] (rv = 78). What is particularly interesting about this quantitative prediction is that it depends on the ranking *mc/e ≫ *mc inferred from an invariant pattern in Ikola structures. Omitting this ranking wrongly predicts gen to be more common in act[pred] (rv = 108) than in pass[exist] (rv = 96). In other words, a quantitative pattern in one construction (Itkonen structure) depends on an invariant pattern in another construction (Ikola structure). This is entirely expected in a theory where variation is not an isolated phenomenon, but embedded in a grammar where invariant and variable patterns interact seamlessly across constructions. This suggests that studying a variable construction in isolation may not be fruitful because the explanation for an observed quantitative pattern in that construction may not be found in the construction itself, but in other grammatically related constructions.

Finally, it would be tempting to go one step further and interpret the ranking volumes directly as observed frequencies (Anttila 1997). That would predict gen at a frequency of 84/150 = 56 % in pass[exist] and at a frequency of 78/150 = 52 % in act[pred]. This interpretation faces a problem in Itkonen’s data: the percentages vary widely across test sentences. It is therefore not clear which frequencies we should try to model. For the same issue in phonological variation, see Coetzee and Kawahara (2013). How can we explain such residual variation across sentences? Is it possible to say anything systematic about it? This is the question we will take up in the next section.

4.4 Are all NPs cyclic?

In a series of papers, Ikola (1950, 1957, 1964, 1989) studied variable Ikola structures where the matrix clause is active and the embedded VP transitive and suggested that the observed nomgen variation is not completely free. The essential content of his hypothesis is stated in (103):
  1. (103)

    Ikola’s hypothesis: The object of the infinitive prefers gen if the matrix verb and matrix object form a closely knit unit, else it prefers nom.

     
The following examples taken from Ikola (1964: 72–73) illustrate the hypothesis: The embedded object is gen in (104), but nom in (105). Ikola suggested that this is because myöntää oikeuden ‘grant the right’ is a closely knit unit whereas aiheutti vaikeuden ‘caused a difficulty’ is a loosely knit unit. In other words, the choice of case on the embedded object depends on how closely the matrix verb and the matrix object are connected. Penttilä (1963: 596–598) notes the same generalization. This is an intriguing hypothesis that to the best of our knowledge has never been empirically tested.
The key question is how to operationalize Ikola’s notion of “closely knit unit.” One possibility is corpus frequency: if a verb and an object form a closely knit unit they should co-occur in the corpus at a high frequency. In our corpus of 295 examples that consist of the Aamulehti 1999 examples of the variable contexts illustrated in (78), there are 167 distinct matrix verb + matrix object pairs. The most frequent pairs are antaa + mahdollisuus ‘give + opportunity’ (15 examples), saada + tehtävä ‘get + task’ (14 examples), antaa + lupa ‘give + permission’ and saada + lupa ‘get + permission’ (11 examples each). Taking high frequency as a proxy for Ikola’s “closely knit unit” the hypothesis makes a clear prediction: the higher the frequency of a pair, the more gen we should observe. The mosaic plot in (106) visualizes the relationship between the frequency of the matrix verb + matrix object pair and the frequency of the embedded object case in our corpus.
  1. (106)

    The distribution of nom vs. gen by matrix verb + matrix object pair frequency

    Open image in new window

     
On the x-axis we have matrix verb + matrix object pairs ordered in terms of ascending frequency: the leftmost column contains the pairs that occurred only once in the corpus (the lowest frequency); the rightmost column contains the pairs that occurred 15 times in the corpus (the highest frequency). On the y-axis we have case: black represents genitive, gray represents nominative. Bin size is proportional to the number of observations. The mosaic plot shows that Ikola’s hypothesis is on the right track: gen tends to be more common with high-frequency pairs. The plot and all subsequent calculations were done in the R statistical computing environment (R Core Team 2014).23

In order to understand the quantitative data better we modeled it using logistic regression; see, e.g., Baayen (2008), Dalgaard (2008). The question is whether the choice between nom vs. gen depends on the matrix verb + matrix object pair frequency. Note that pair frequency is simply our operationalization of Ikola’s notion of “closely knit unit.” It does not entail the claim that construction frequency plays an explanatory role in the choice of case (cf. Coetzee and Kawahara 2013). The advantage of logistic regression is that it allows us to consider several predictors at once. In particular, Ikola briefly remarks that the matrix object and the embedded object might exhibit attractio casus, a tendency to share the same case (Kholodilova 2013). We therefore included the case of matrix object (adv, gen, nom, par, where adv includes all semantic cases) as a predictor in the model. A third predictor we included is the number of the matrix object (pl, sg).

A summary of the regression model is shown in (107). A positive estimate means that the predictor favors gen. We find that Ikola’s hypothesis is supported: the log frequency of the matrix verb + matrix object pair is significant in the expected direction. In contrast, no support is found for the attractio casus hypothesis: neither matrix object case nor number come out significant.
  1. (107)

    A summary of the logistic regression model

     

    Estimate

    Std. Error

    z value

    Pr(>|z|)

    (Intercept)

    -3.25804

    0.53513

    -6.088

    1.14e-09 ***

    Case = GEN

    -0.04998

    0.34057

    -0.147

    0.883

    Case = NOM

    -15.15894

    819.30702

    -0.019

    0.985

    Case = PAR

    -0.68888

    0.46502

    -1.481

    0.139

    N num = SG

    0.79221

    0.48356

    1.638

    0.101

    VN logfreq

    1.32644

    0.21506

    6.168

    6.93e-10 ***

    - - -

    Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

     
How should Ikola’s hypothesis be interpreted in terms of our cyclic theory? The following interpretation suggests itself:
  1. (108)

    Ikola’s hypothesis interpreted: There are two kinds of matrix verb + matrix object pairs. In a loosely knit (typically low-frequency) pair the object NP is cyclic; in a closely knit (typically high-frequency) pair the object NP is noncyclic.

     
Ikola’s generalization can now be understood as follows. In a closely knit pair the object NP is not a cycle: the sentence is evaluated in one pass and we get invariant gen. As Ikola (1964: 70–71) insightfully notes, closely knit pairs are analogous to simple verbs: one can substitute haluta ‘desire’ for osoittaa + halua ‘show + desire’, suggesting that the pair is akin to a single predicate. In a loosely knit pair the object NP is a cycle: the sentence is evaluated in two passes, allowing faithfulness to nom to play a role, and we get nomgen variation. Since the structure is in diachronic flux the cyclic status of a particular verb + object pair may well vary across and even within individuals. It is possible that the sentence-specific differences in Itkonen’s data reflect analogous differences between cyclic vs. non-cyclic predicate + complement combinations.

We started from the assumption that all NPs are cyclic. We now revise this assumption by allowing both cyclic and noncyclic NPs. In terms of our model, noncyclic NPs ignore faithfulness (max/np, dep/np). The predictions for the two types of NPs are identical except in one case: if the matrix clause is active, i.e., has a nominative subject, and the embedded predicate is transitive, i.e., has an internal argument, only invariant gen is predicted if the NP is noncyclic. This is exactly the variation environment.

Finally, Ikola (1964: 71) suggests that there may even be phonological evidence for two kinds of NPs: “The object of a sentence [a cyclic NP] is in a stressed position whereas in fixed phrases [noncyclic NPs] it is often relatively unstressed.” This suggests that cyclic NPs receive phrasal stress whereas noncyclic NPs do not. The stress patterns of the two types of NPs in Finnish would thus parallel the stress patterns of phrases vs. compounds in English (blàck bóard vs. bláckbòard, Chomsky and Halle 1968). More evidence for a distinction between cyclic and noncyclic structures emerges in other complex constructions. This is the topic of the next section.

5 Other constructions

In the previous sections, we have seen that CPs and NPs may be cyclic categories in Finnish. In this section, we will briefly discuss a number of other nonfinite constructions. The key observation is that only some embedded structures are cyclic. Two additional types can be identified: precyclic structures that form a single cycle with the matrix clause and postcyclic structures that form a separate cycle, but do not feed subsequent cycles. We will focus on the constructions discussed in Vainikka and Brattico (2014) and use their nomenclature that originates in Vainikka (1989), but we will postpone the evaluation of Vainikka and Brattico’s solution until the following section.

Precyclic structures are part of the matrix clause for purposes of case: if the matrix clause has an external argument the embedded object is gen; if not, the embedded object is nom. In other words, the two form a monoclausal structure. An example of a precyclic structure is the A-infinitive, traditionally known as the first infinitive, which Vainikka and Brattico (2014: 93) call “the least clause-like of the non-finite forms in Finnish.” The nonfinite VP embedded in Ikola structures is an A-infinitive. Examples of A-infinitives as complements of verbs are shown in (109) where the matrix verb is active and in (110) where the matrix verb is passive. There is no variation. Another precyclic structure is the MA-infinitive, traditionally known as the third infinitive, illustrated with an active matrix clause in (111) and with a passive matrix clause in (112). Note that in the active both internal arguments ‘contest’ and ‘bear’ take the marked gen; in the passive both revert to the unmarked nom because there is no external argument present. The analysis is straightforward: A-infinitives and MA-infinitives are not cycles, but form a single case assignment domain with the matrix clause. Nothing else needs to be said.
Two cyclic structures have already been discussed in detail: Itkonen and Ikola structures. An additional example of a cyclic structure is the rationale adjunct, traditionally known as the translative form of the first infinitive, or the purpose clause. This structure shows nomgen variation parallel to Itkonen structures. The same analysis applies. If the matrix clause has an external argument the object in the embedded clause is gen, as illustrated in (113). The fact that Finnish allows optional pro-drop in first and second person is immaterial (Vainikka and Levy 1999; Holmberg 2005). If the matrix clause does not have an external argument, as in passive and necessive clauses, we get the familiar nomgen variation, as illustrated in (114). Finally, the temporal adjunct provides an example of a postcyclic structure. Here the embedded object is not sensitive to the matrix clause at all (Hakulinen et al. 2004: Sect. 543). Instead, it is strictly faithful to the first cycle gen, with no alternation or variation. This is illustrated in (115) for active matrix clauses and in (116) for passive matrix clauses. Under our analysis, the temporal construction constitutes a cycle of its own, but is not visible on the matrix cycle. This correctly predicts that the matrix clause has no chance to influence its case pattern, which thus remains faithful to the first cycle.

Is the three-way cyclic typology of clauses (precyclic, cyclic, postcyclic) only relevant for case or is it also reflected in other syntactic processes? The possible identity of case domains and extraction domains is discussed in Vainikka and Brattico (2014), Sect. 4.2, with negative results. Toivonen (1995) convincingly argues that extraction is possible out of nonfinite complements, but not out of nonfinite adjuncts, suggesting that the complement/adjunct distinction coincides with extraction domains; see Huhmarniemi (2012) for a detailed discussion as well as some cases of variation. It seems clear that the cyclic status of a clause is independent of its status as complement vs. adjunct and hence independent of extraction domains. Precyclic structures can be complements or adjuncts: the A-infinitive in (109)–(110) is a complement, but the MA-infinitive in (111)–(112) is an adjunct. Similarly, cyclic structures can be complements or adjuncts: Itkonen structures are complements, but rationale adjuncts are adjuncts. Finally, postcyclic structures can be complements or adjuncts: temporal adjuncts are adjuncts, but embedded finite clauses, which must be postcyclic as their case pattern is independent of the matrix clause, can be complements of verbs like ‘say’, ‘think’, ‘want’, and ‘believe’ and seem to allow extraction (Huhmarniemi 2012: 96–97; Vainikka and Brattico 2014: 104).

At this point, our cyclic typology is a descriptive one. We have no particular theoretical reasons to expect some categories to function as cyclic domains. However, we have seen evidence that the choice does not depend entirely on the category: some CPs are cyclic (Itkonen structures), others are postcyclic (embedded finite clauses); some NPs are cyclic (the “loosely knit” Ikola structures), others are precyclic (the “tightly knit” Ikola structures). It also remains to be seen whether the cyclic typology of clauses has reflexes in Finnish syntax beyond structural case.

6 Case and agreement

An interesting agreement-based analysis of Finnish structural case is offered in Vainikka and Brattico (2014). Their analysis provides a useful comparison to ours because it covers many of the same nonfinite structures from a sophisticated alternative perspective. Vainikka and Brattico put forward two main claims. First, the nom/gen alternation is argued to be genuinely long-distance: the case assigner may be situated arbitrarily far from the assignee. On this point we are in complete agreement. Second, the nom/gen alternation is argued to depend on subject-verb agreement higher in the clause. It is on this point that we disagree. We will now show that Vainikka and Brattico’s agreement analysis covers the core cases, but falls short in a number of respects, and that our cyclic analysis goes further.

Vainikka and Brattico assume a strict correlation between agreement and gen: agreement implies gen and gen implies agreement. This accounts for a number of core cases. Consider the simplex clauses in (117) and (118).24 According to Vainikka and Brattico, gen arises from the presence of agreement and nom arises from its absence. In active transitive clauses, agreement (3p.sg) assigns gen to the object. Since imperatives and passives have no agreement, so goes the argument, the object gets nom from the c-commanding C as a last resort. This rules out two kinds of phenomena: presence of agreement with nom and absence of agreement with gen. Let us call the first underapplication because gen is expected, but fails to appear, and the second overapplication because gen is not expected, but appears nevertheless. Both are ruled out, but both are found in Finnish.
In simplex clauses imperatives exhibit underapplication: they are overtly inflected in both number and person, but the object is always nom, as shown in (119). In complex clauses we find both underapplication and overapplication. According to Vainikka and Brattico the object of a nonfinite clause can receive gen from two sources: the primary source is the agreement on the nonfinite verb realized as the possessive suffix (px); the secondary source is agreement on the matrix verb. If neither clause has agreement, the embedded object receives nom from the c-commanding C as a last resort. Summarizing, the system has four ways to assign case to the object of a nonfinite clause:
  1. (120)

    Embedded object case by agreement (Vainikka and Brattico 2014)

    Matrix

    Embedded

    Case

    +agr

    [+agr]

    gen

    +agr

    [−agr]

    gen

    agr

    [+agr]

    gen

    agr

    [−agr]

    nom

     
Vainikka and Brattico’s analysis correctly predicts the case of the embedded object in control sentences. In (121) the possessive suffix (px) in the embedded clause assigns gen to the object: If the embedded clause has an overt subject, as in Itkonen structures, the px is absent. Under active matrix clauses the embedded object is gen. This is also correctly predicted under Vainikka and Brattico’s analysis: here the source of gen is the matrix clause agreement. However, problems arise under passive matrix clauses. The embedded object should be nom because passive has no agreement, but as Vainikka and Brattico note, we get variation nomgen. The gen variant is unexpected, an instance of overapplication. Where does the gen variant come from? Under our analysis it is optionally inherited from the first cycle through faithfulness. Vainikka and Brattico propose that it comes from agreement on the participle. On the surface this is plainly not the case: there is no px when the embedded clause has an overt subject, as shown in (122). Following their hypothesis to its logical conclusion, Vainikka and Brattico conclude that the px is optionally present, but inaudible. The postulation of an inaudible suffix is not in itself objectionable, of course. The problem is that there is no independent evidence for it.
The same problem arises in temporal constructions where px is not possible, but the embedded object is invariably gen under passive matrix clauses: In these examples, gen has no plausible agreement source. However, it is entirely expected under the cyclic analysis: gen is assigned to objects of transitive clauses on the first cycle and preserved intact because the structure is postcyclic.
However, for the sake of the argument, let us follow Vainikka and Brattico in assuming that the case variation in fact does reflect optional inaudible agreement on the participle. This results in the following predictions:
  1. (125)
    1. a.

      Active matrix clause (agreement):

      +agr [+agr] → gen

      +agr [−agr] → gen

       
    2. b.

      Passive matrix clause (no agreement):

      agr [+agr] → gen

      agr [−agr] → nom

       
     
This predicts invariant gen under active matrix clauses and variation nomgen under passive matrix clauses. This is correct for embedded transitive clauses. The problem is that the analysis does not generalize to embedded existentials or predicatives which show nomgen variation even under active matrix clauses: This variation cannot be derived from optional inaudible agreement in the embedded clause because matrix agreement will always be available to assign gen to the embedded NP. This time the nom variant is unexpected, an instance of underapplication. The upshot is that optional agreement does not predict enough variation.

As for Ikola structures, Vainikka and Brattico (2014: 93–94) give examples of nom objects, suggesting that gen objects are ruled out because an infinitive inside the complement of a noun is independent of the matrix verb and does not support gen. However, gen is robustly attested in such examples in our corpus. This is unexplained under the agreement analysis which has no plausible source for the gen variant. Under our analysis the variation is correctly predicted.25

More problems emerge if we try to generalize the agreement analysis to predicatives. Consider Ikola structures with matrix transitives (nomgen) and matrix predicatives (nom): This case difference is not explained by agreement: the agreement patterns are identical, but the case patterns are different. The correct empirical generalization refers to argument structure: the alternating NP is embedded inside an internal argument (NP) in (128), but inside an external argument (NP/e) in (129), hence the difference in case patterns.

In sum, Vainikka and Brattico’s analysis posits a strict correlation between agreement and gen: agreement implies gen and gen implies agreement. This works well in a number of cases, but not all: sometimes gen does not appear in the presence of agreement (underapplication) and sometimes gen appears in the absence of agreement (overapplication). This suggests that agreement is not responsible for structural case assignment in Finnish. In contrast, a theory where structural case distinguishes the external argument from other arguments in a cyclic fashion gets the facts right.

7 Conclusion

We have proposed a new solution to two outstanding problems in Finnish structural case: non-locality and free variation. The solution builds on two assumptions: (i) structural case distinguishes the external argument from other arguments and (ii) structural case assignment is cyclic. Formulated in terms of Stratal Optimality Theory (Kiparsky 2000) and Partial Order Optimality Theory (Anttila 1997; Anttila and Cho 1998/2003; Djalali 2014) the analysis correctly predicts a wide range of case patterns in both simplex and complex clauses, including intermediate well-formedness judgments in complex clauses (Itkonen 1976, 1981). More generally, the analysis illustrates the intimate relationship between categorical and quantitative patterns: as Itkonen (1981, 1976) demonstrates, quantitative patterns are just as systematic as categorical patterns and both are conditioned by the same grammatical factors. We have shown that it is possible to derive both types of patterns from one and the same grammar in terms of a conservative generalization of classical Optimality Theory, without any numerical parameters.

Footnotes

  1. 1.

    Abbreviations: acc ‘accusative’, act ‘active’, ela ‘elative’, gen ‘genitive’, imp ‘imperative’, ine ‘inessive’, nom ‘nominative’, p ‘person’, par ‘partitive’, pass ‘passive’, past ‘past tense’, perf ’perfective aspect’, pl ‘plural’, pres ‘present/imperfective aspect’, px ‘possessive suffix’, sg ‘singular’.

  2. 2.

    The terminology surrounding Finnish structural case is somewhat confusing. One common view posits an abstract accusative (acc) which is morphologically realized as zero (acc/∅) homophonous with the nominative and as -n (acc/n) homophonous with the genitive; see, e.g., Vainikka and Brattico (2014). In this paper, we take the view that the zero accusative is the nominative (see, e.g., Jahnsson 1871, and subsequently Timberlake 1974; Milsark 1985; Taraldsen 1985; Mitchell 1991; Maling 1993; Toivainen 1993; Nelson 1998; Kiparsky 2001) and the -n accusative is the genitive (see, e.g., Penttilä 1963; Vainikka 1989; Kiparsky 2001).

  3. 3.

    The statement in (5) cannot be found as such in Jahnsson’s book. What we do find is a statement that the presence of a subject matters to making the object accusative (Jahnsson 1871: 10) and the absence of a “personal” subject matters to making it nominative (Jahnsson 1871: 14). Jahnsson’s Rule is problematic if taken as an inviolable constraint (see Vainikka and Brattico 2014), but what is correct about it can be captured in terms of violable constraints, as we will see shortly.

  4. 4.

    Any statement of Jahnsson’s Rule must mention two special cases. First, human pronouns have a dedicated accusative form that is insensitive to the presence vs. absence of subject, e.g., Ota minu-t take.imp 1p.sg-acc ‘Take me!’ Second, plural objects are always realized in the unmarked nominative, e.g., Matti ampui karhu-t Matti.shot bear-pl.nom ‘Matti shot the bears’. Both will be set aside here.

  5. 5.

    Certain adverbs of duration, measure, and frequency also receive structural case and behave more or less like objects, e.g., Pekka nukkui tunni-n ‘Pekka slept an hour-gen’, Kala painoi kilo-n ‘The fish weighed a kilo-gen’, Pekka luki kirja-n kerra-n ‘Pekka read the book-gen once-gen’; see, e.g., Maling (1993). However, adverbs allow variation not found in argument case. For discussion of adverb case and an optimality-theoretic analysis, see Anttila and Kim (2011).

  6. 6.

    A reviewer notes that the harmonically bounded candidate *gen nom with marked case on the subject and unmarked case on the object resembles the pattern in ergative languages, raising the question of how they would be analyzed. The reviewer suggests two possible solutions: (i) Ergative is inherent; see, e.g., Woolford (2006), Anand and Nevins (2006), Legate (2008), and McFadden (2009). Under this view, the ergative case would be protected by faithfulness to inherent case; (ii) Ergative is structural, see, e.g., Marantz (1991) and Baker (2014). Under this view, one could propose an additional constraint *mc/i ‘Do not case-mark an internal argument’ and ergative and accusative languages would be distinguished based on how this constraint is ranked relative to the others. We leave the choice between these two alternatives open.

  7. 7.

    A reviewer inquires about the analysis of human pronouns. The basic generalization is that human pronouns are acc as internal arguments: Minä näin sinu-t ‘I saw you-acc’ (active transitive); Sinu-t nähtiin ‘You-acc were seen’ (passive transitive); Minu-lla on sinu-t ‘I have you-acc’ (lit., ‘On me is you’, possessive); cf. Minä olen hän ‘I am he.nom’ (predicative). This generalization could be stated as an undominated constraint. For the potential exception of unaccusatives, see Nelson (1998: 83–84). Not all human pronouns are alike. For example, joku ‘someone’ behaves like a lexical noun in terms of case, e.g., Minä näin jo-n-ku-n ‘I saw someone-gen’ (active transitive); Joku nähtiin ‘Someone.nom was seen’ (passive transitive); Minu-lla on joku ‘I-ade have someone.nom’ (possessive). We will return to the case marking of joku briefly below. The differential case marking of pronouns vs. lexical nouns is a complex problem that we cannot satisfactorily solve here; for an OT approach, see Aissen (2003).

  8. 8.

    An optimality-theoretic grammar can be defined as a set of ordered pairs R in the constraint set C, i.e., as a binary relation in C. In Classical Optimality Theory R is irreflexive, asymmetric, transitive, and connected. Partial Order Optimality Theory omits the connectedness assumption, see, e.g., Anttila and Cho (1998/2003). An introduction to ordering can be found in Partee et al. (1993: 39–53); for a formalization of Partial Order Optimality Theory, see Djalali (2014).

  9. 9.

    The reverse does not hold: there are many sets of total orders that are not partial orders. For example, the set {1, 6} is not a partial order.

  10. 10.

    A reviewer inquires about the empty grammar at the top of the diagram. The empty grammar allows any possible ranking of the three constraints, but that does not mean it allows any possible output pattern. In the case of (20), the empty grammar predicts variation in transitive clauses (nomgennomnom), but no variation in existential (inenom) or predicative (nomnom) clauses. This is like actual Finnish except that it allows nomgen variation on direct objects. To the best of our knowledge, there is no such native dialect, but it might well be the dialect of a second language speaker whose first language lacks case inflection. Assuming that constraints are universal and rankings language-particular, the empty grammar is what a child brings into the world before exposure to language data. Such a grammar permits extensive variation, but in a way tightly constrained by Universal Grammar. In particular, the predicted dialect conforms to all the implicational universals that follow from these constraints, as discussed in Sect. 3.4.

  11. 11.

    The following underlying forms may be assumed: the voice morphemes are /-∅/ ‘active’ and /-ttA/ ‘passive’; the aspect morphemes are /-vA/ ‘present’, /-nee/ ‘perfect’, /ttA-vA/ ‘passive, present’, and /-tU/ ‘passive, perfect’. The case morpheme is invariably /-n/ ‘genitive’.

  12. 12.

    A reviewer asks about possible interpretational or structural differences between nom and gen. We have the explicit affirmation of Hakulinen and Karlsson (1975: 339): “No meaning difference is associated with the morphological difference between [genitive] and [nominative] […].” Itkonen (1981: 105) discusses one potential difference, the influence of prescriptive rules, but dismisses it as irrelevant. We will see that the choice between nom and gen is largely predictable. If there were some unidentified structural difference involved, the evidence would simply show that the choice between the two structures is largely predictable.

  13. 13.

    Interrogative pronouns and topicalized NPs appear to be different. Itkonen (1981: 115–116) notes that gen is preferred in cases like Kene-n [hän sanoi vieraa-n olevan t]? who-gen [he.nom said guest-gen be-act.pres-gent] ‘Who did he say the guest to be?’ Yhtä hullu-n [minä luulen hänen olevan t kuin ennenkin] as crazy-gen [I believe he-gen be-act.pres-gent as before] ‘I believe him to be as crazy as ever.’ It is not clear to us what to make of this pattern.

  14. 14.

    This proposal is anticipated in Hakulinen and Karlsson (1975: 345). For Itkonen, the data provided evidence against the transformational grammar of the day: he pointed out that his generalization would presuppose peculiar “sideways derivations” where output structures communicate with one another.

  15. 15.

    The theoretical literature on Finnish case is remarkably silent about the source of this gen. Setälä (1901: 109) calls it an accusative singular. We are not aware of any competing analyses. Parallel examples exist in Sakha (Yakut) (Baker and Vinokurova 2010: 615) and Uzbek (Gribanova 2016).

  16. 16.

    Itkonen’s criterion for “favoring nom” was defined as choosing nom in at least 7 sentences out of 8; the rest were grouped as “not favoring nom.” He appears to have chosen the cutoff point with an eye towards making the two groups approximately the same size (Itkonen 1981: 110); here both groups have 63 subjects.

  17. 17.

    This entailment corresponds to the rule of W-extension of Prince (2002a, 2002b): a row entails any other row that can be derived from it by replacing an empty cell with a W (see McCarthy 2008: 124–132).

  18. 18.

    This entailment corresponds to the rule of L-retraction of Prince (2002a, 2002b): a row entails any other row that can be derived from it by replacing an L with an empty cell (see McCarthy 2008: 124–132).

  19. 19.

    Human pronouns take acc even here: Se oli tilaisuus tavata häne-t ‘It was an opportunity to meet him/her-acc.’ We noted above that joku ‘someone’ behaves like a common noun in terms of case, but internet searches turn up examples like Sanotaan, että häät ovat paras tilaisuus tavata jo-n-ku-n ‘It is said that a wedding is the best opportunity to meet someone-gen’ where joku takes gen inside a predicative, although nom is also found. We find no such examples in our corpus. While we have no satisfactory explanation to offer it is probably not a coincidence that these gen examples involve a human pronoun.

  20. 20.

    This observation is also made by Brattico (2012: 277), although his generalization is different from ours. We will discuss Brattico and Vainikka’s analysis in Sect. 6.

  21. 21.

    It turns out that tableau (89) contains all the information needed to infer the full ranking for Finnish. This means that datum (88) allows the learner to infer all the case patterns discussed above, including the quantitative patterns in Itkonen’s data.

  22. 22.

    The graph includes one prediction that cannot be tested given the available data: gen should be less frequent in active-NP[transitive] (an Ikola structure) than in active[existential] (an Itkonen structure). Our data are not appropriate for checking this prediction: the Itkonen data come from an experiment conducted in the mid-1970’s; the Ikola data come from a newspaper published a quarter century later.

  23. 23.

    We also tested Ikola’s hypothesis by defining pair frequency as the frequency of the 167 adjacent verb + noun pairs in the full Aamulehti 1999 corpus. The results were similar. The advantage of this alternative measure is the larger spread of frequencies; the disadvantage is that the full Aamulehti 1999 corpus was not manually checked and the data contain plenty of noise due to homonymy and ambiguity.

  24. 24.

    Vainikka and Brattico assume that Finnish has an abstract accusative with three morphological variants: the /-t/-accusative which only occurs on human pronouns (acc/t), the zero accusative which is homophonous with the nominative (acc/∅), and the /-n/ accusative which is homophonous with the genitive (acc/n). In keeping with our usage in the present paper, we will call the latter two nom and gen, respectively.

  25. 25.

    The variation in Ikola structures is noted by Brattico (2012: 277): “if the matrix verb shows full phi-features, both the n-accusative [= gen] and ∅-accusative [= nom] are possible inside the object NP.” Here Brattico invokes a Telescopic Object Principle: “Long distance case assignment between a probe and a goal takes place only if the goal is a direct object.”

Notes

Acknowledgements

This paper has benefited from presentations at the 12th Korea-Japan Workshop on Linguistics and Language Processing (Seoul, March 24, 2012), the Stanford Syntax and Morphology Circle (March 14, 2013), the Berkeley Syntax and Semantics Circle (April 10, 2013), the Structure of Finnish seminar at Stanford University (Winter Quarter 2014), the Linguistic Society of America Annual Meeting, Portland, Oregon (January 8, 2015), and the Workshop on the Formal Structure of OT Typologies, Rutgers University (May 29, 2015). We thank Mark Baker, Pauli Brattico, Joan Bresnan, Alex Djalali, Vivienne Fong, Vera Gribanova, Jane Grimshaw, Boris Harizanov, Heidi Harley, Tarja Heinonen, Cameron Jeffers, Peter Jenks, Elsi Kaiser, Lauri Karttunen, Paul Kiparsky, Victor Kuperman, Joan Maling, Line Mikkelsen, Ethan Poole, Alan Prince, Shigeo Tonoike, Anne Vainikka, Maria Vilkuna, the editors of NLLT, and three anonymous reviewers whose extremely helpful comments improved the paper significantly in both content and presentation. We are responsible for any errors. The first author gratefully acknowledges the financial support of Kyung Hee University during his term as an International Scholar (April 2011–March 2012). If not otherwise stated, the well-formedness judgments are those of the first author. We dedicate this paper to the memory of Osmo Ikola (1918–2016), a pioneer in Finnish syntax, whose work was an inspiration for us.

Supplementary material

References

  1. Aissen, Judith. 1999. Markedness and subject choice in Optimality Theory. Natural Language and Linguistic Theory 17: 673–711. CrossRefGoogle Scholar
  2. Aissen, Judith. 2003. Differential object marking: Iconicity vs. economy. Natural Language and Linguistic Theory 21: 435–483. CrossRefGoogle Scholar
  3. Anand, Pranav, and Andrew Nevins. 2006. The locus of ergative case assignment: Evidence from scope. In Ergativity: Emerging issues, eds. Alana Johns, Diane Massam, and Juvenal Ndayiragije, 3–27. Dordrecht: Kluwer. CrossRefGoogle Scholar
  4. Anttila, Arto. 1997. Deriving variation from grammar. In Variation, change and phonological theory, eds. Frans Hinskens, Roeland van Hout, and Leo Wetzels, 35–68. Amsterdam: Benjamins. CrossRefGoogle Scholar
  5. Anttila, Arto, and Young-mee Yu Cho. 1998/2003. Variation and change in Optimality Theory. Lingua 104: 31–56. Reprinted 2003 in Optimality theory in phonology, ed. John J. McCarthy, 569–580. Malden: Blackwell Sci. Google Scholar
  6. Anttila, Arto, and Curtis Andrus. 2006. T-order generator. Software package. Stanford University. Available at stanford.edu/~anttila/research/software.html. Accessed 3 October 2016.
  7. Anttila, Arto, and Vivienne Fong. 2000. The partitive constraint in optimality theory. Journal of Semantics 17: 281–314. CrossRefGoogle Scholar
  8. Anttila, Arto, and Jong-Bok Kim. 2011. On structural case in Finnish and Korean. Lingua 121(1): 100–127. CrossRefGoogle Scholar
  9. Baayen, R. H. 2008. Analyzing linguistic data: A practical introduction to statistics using R. Cambridge: Cambridge University Press. CrossRefGoogle Scholar
  10. Baker, Mark. 2014. On dependent ergative case (in Shipibo) and its derivation by phase. Linguistic Inquiry 45(3): 341–379. CrossRefGoogle Scholar
  11. Baker, Mark. 2015. Case: Its principles and its parameters. Vol. 146 of Cambridge studies in linguistics. Cambridge: Cambridge University Press. CrossRefGoogle Scholar
  12. Baker, Mark, and Nadya Vinokurova. 2010. Two modalities of case assignment: Case in Sakha. Natural Language and Linguistic Theory 28: 593–642. CrossRefGoogle Scholar
  13. Bane, Max. 2011. Deriving the structure of variation from the structure of non-variation in the English dative alternation. In 28th West Coast Conference on Formal Linguistics (WCCFL), eds. Mary Byram Washburn et al., 42–50. Somerville: Cascadilla Proceedings Project. Google Scholar
  14. Bhatt, Rajesh. 2005. Long-distance agreement in Hindi-Urdu. Natural Language and Linguistic Theory 23: 757–807. CrossRefGoogle Scholar
  15. Blake, Barry. 1994. Case. Cambridge: Cambridge University Press. Google Scholar
  16. Bobaljik, Jonathan David. 2008. Where’s phi? Agreement as a post-syntactic operation. In Phi-theory: Phi features across interfaces and modules, eds. Daniel Harbour, David Adger, and Susana Béjar, 295–328. Oxford: Oxford University Press. Google Scholar
  17. Bobaljik, Jonathan David, and Susi Wurmbrand. 2005. The domain of agreement. Natural Language and Linguistic Theory 23(4): 809–865. CrossRefGoogle Scholar
  18. Bobaljik, Jonathan David, and Susi Wurmbrand. 2009. Case in GB/minimalism. In The Oxford handbook of case, eds. Andrej Malchukov and Andrew Spencer, 44–58. Oxford: Oxford University Press. Google Scholar
  19. Brasoveanu, Adrian, and Alan Prince. 2011. Ranking and necessity: The fusional reduction algorithm. Natural Language and Linguistic Theory 29: 3–70. CrossRefGoogle Scholar
  20. Brattico, Pauli. 2012. Long-distance case assignment in Finnish. Studia Linguistica 66(3): 245–285. CrossRefGoogle Scholar
  21. Butt, Miriam. 2006. Theories of case. Cambridge: Cambridge University Press. CrossRefGoogle Scholar
  22. Chomsky, Noam. 1957. Syntactic structures. The Hague: Mouton and Co. Google Scholar
  23. Chomsky, Noam. 1981. Lectures on government and binding. Dordrecht: Foris. Google Scholar
  24. Chomsky, Noam. 2000. Minimalist inquiries: The framework. In Step by step: Essays on minimalist syntax in honor of Howard Lasnik, eds. Roger Martin, David Michaels, and Juan Uriagereka, 89–155. Cambridge: MIT Press. Google Scholar
  25. Chomsky, Noam. 2001. Derivation by phase. In Ken Hale: A life in language, ed. Michael Kenstowicz, 1–52. Cambridge: MIT Press. Google Scholar
  26. Chomsky, Noam, and Morris Halle. 1968. The sound pattern of English. New York: Harper and Row. Google Scholar
  27. Chomsky, Noam, Morris Halle, and Fred Lukoff. 1956. On accent and juncture in English. In For Roman Jakobson: Essays on the occasion of his sixtieth birthday, eds. Morris Halle et al., 65–80. The Hague: Mouton and Co. Google Scholar
  28. Coetzee, Andries, and Shigeto Kawahara. 2013. Frequency biases in phonological variation. Natural Language and Linguistic Theory 31: 47–89. CrossRefGoogle Scholar
  29. Comrie, Bernard. 1989. Language universals and linguistic typology. Chicago: University of Chicago Press. Google Scholar
  30. Dalgaard, Peter. 2008. Introductory statistics with R. Berlin: Springer. CrossRefGoogle Scholar
  31. Dench, Alan, and Nicholas Evans. 1988. Multiple case marking in Australian languages. Australian Journal of Linguistics 8: 1–47. CrossRefGoogle Scholar
  32. Djalali, Alex J. 2014. A constructive solution to the ranking problem in partial order optimality theory. Ms., Stanford University. Google Scholar
  33. Djalali, Alex J., and Cameron Jeffers. 2015. OTOrder. Software package. Stanford University. Available at http://rc-linguistics.stanford.edu/. Accessed 3 October 2016.
  34. Fong, Vivienne. 1997a. A temporal interpretation for locative case. In Fifteenth West Coast Conference on Formal Linguistics (WCCFL), eds. Brian Agbayani and Sze-Wing Tang, 145–159. Stanford: CSLI Publications. Google Scholar
  35. Fong, Vivienne. 1997b. The order of things: What directional locatives denote. PhD diss., Stanford University. Google Scholar
  36. Greenberg, Joseph H. 1963. Some universals of grammar with particular reference to the order of meaningful elements. In Universals of language, ed. Joseph Greenberg, 58–90. Cambridge: MIT Press. Google Scholar
  37. Gribanova, Vera. 2016. Case, agreement, and differential subject marking in Uzbek nominalized clauses. Ms., Stanford University. Google Scholar
  38. Hakulinen, Auli, and Fred Karlsson. 1975. Suomen akkusatiivi: funktionaalinen näkökulma [The Finnish accusative: A functional perspective]. Virittäjä 79: 339–363. Google Scholar
  39. Hakulinen, Auli, and Fred Karlsson. 1979. Nykysuomen lauseoppia [On the syntax of modern Finnish]. Helsinki: Suomalaisen Kirjallisuuden Seura. Google Scholar
  40. Hakulinen, Auli, Maria Vilkuna, Riitta Korhonen, Vesa Koivisto, Tarja Riitta Heinonen, and Irja Alho. 2004. Iso suomen kielioppi [A comprehensive grammar of Finnish]. Helsinki: Suomalaisen Kirjallisuuden Seura. Google Scholar
  41. Harley, Heidi. 1995. Subjects, events, and licensing. PhD diss., Massachusetts Institute of Technology. Google Scholar
  42. Hayes, Bruce, Bruce Tesar, and Kie Zuraw. 2003. OTSoft 2.1, software package. OTSoft 2.4 available at linguistics.ucla.edu/people/hayes/otsoft/. Accessed 3 October 2016.
  43. Holmberg, Anders. 2005. Is there a little pro? Evidence from Finnish. Linguistic Inquiry 36(4): 533–564. CrossRefGoogle Scholar
  44. Huhmarniemi, Saara. 2012. Finnish A′-movement: Edges and islands. PhD diss., University of Helsinki. Google Scholar
  45. de Hoop, Helen. 2009. Case in optimality theory. In The Oxford handbook of case, eds. Andrej Malchukov and Andrew Spencer, 88–101. Oxford: Oxford University Press. Google Scholar
  46. de Hoop, Helen, and Andrej Malchukov. 2008. Case-marking strategies. Linguistic Inquiry 39: 565–587. CrossRefGoogle Scholar
  47. Ikola, Osmo. 1950. Infinitiivin objektista [On the object of an infinitive]. Virittäjä 54: 468–474. Google Scholar
  48. Ikola, Osmo. 1957. Infinitiivin objektia koskevien sääntöjen soveltamisesta [On the application of rules governing the object of an infinitive]. Virittäjä 61: 342–345. Google Scholar
  49. Ikola, Osmo. 1964. Infinitiivin objektista [On the object of an infinitive]. In Lauseopin kysymyksiä: Tutkielmia nykysuomen syntaksin alalta [Problems of sentence structure: Studies on the syntax of modern Finnish]. Vol. 26 of Tietolipas, 64–80. Helsinki: Suomalaisen kirjallisuuden seura. Google Scholar
  50. Ikola, Osmo. 1989. Mistä infinitiivin objektin horjuvuus johtuu? [Why the variation in the object of an infinitive?]. Virittäjä 93: 452–456. Google Scholar
  51. Itkonen, Terho. 1976. Syntaktisten vaikutusyhteyksien luonteesta [On the nature of syntactic interactions]. Virittäjä 80: 52–81. Google Scholar
  52. Itkonen, Terho. 1977. Tutkimuksen vaihtoehdoista ja vähän paradigmoistakin [On alternatives in research and something about paradigms]. Virittäjä 81: 78–89. Google Scholar
  53. Itkonen, Terho. 1981. Katson minulla olevan oikeus. In Näillä näkymin: Kirjoituksia nykysuomesta ja sen huollosta, 86–116. Helsinki: Suomalaisen kirjallisuuden seura. Google Scholar
  54. Jahnsson, Adolf W. 1871. Finska språkets satslära: För läroverkens behof. Helsingfors: Finska Litteratur-sällskapets tryckeri. Google Scholar
  55. Jarosz, Gaja. 2010. Implicational markedness and frequency in constraint-based computational models of phonological learning. Journal of Child Language 37: 565–606. CrossRefGoogle Scholar
  56. Kholodilova, Maria. 2013. Inverse attraction in Ingrian Finnish relative clauses. Linguistica Uralica 2: 96–116. CrossRefGoogle Scholar
  57. Kim, Jong-Bok, and Incheol Choi. 2004. The Korean case system: A unified constraint-based approach. Language Research 40(4): 885–921. Google Scholar
  58. Kiparsky, Paul. 1993. Variable rules (an OT perspective on phonological variation). Handout from Rutgers optimality workshop 1993, also presented at NWAV 1994, Stanford University. Available at stanford.edu/~kiparsky/Papers/nwave94.pdf. Accessed 3 October 2016.
  59. Kiparsky, Paul. 1998. Partitive case and aspect. In The projection of arguments: Lexical and compositional factors, eds. Miriam Butt and Wilhelm Geuder, 265–307. Stanford: CSLI Publications. Google Scholar
  60. Kiparsky, Paul. 2000. Opacity and cyclicity. The Linguistic Review 17: 351–367. CrossRefGoogle Scholar
  61. Kiparsky, Paul. 2001. Structural case in Finnish. Lingua 111: 315–376. CrossRefGoogle Scholar
  62. Kiparsky, Paul. 2006. The amphichronic program vs. evolutionary phonology. Theoretical Linguistics 32(2): 217–236. CrossRefGoogle Scholar
  63. Kiparsky, Paul. 2010. Nonfinite clauses in Finnish. Ms., Stanford University. Google Scholar
  64. Kiparsky, Paul. 2013. Towards a null theory of the passive. Lingua 125: 7–33. CrossRefGoogle Scholar
  65. Kroch, Anthony S. 1989. Reflexes of grammar in patterns of language change. Language Variation and Change 1: 199–244. CrossRefGoogle Scholar
  66. Legate, Julie Anne. 2005. Phases and cyclic agreement. MIT Working Papers in Linguistics 49: 147–156. Google Scholar
  67. Legate, Julie Anne. 2008. Morphological and abstract case. Linguistic Inquiry 39(1): 55–101. CrossRefGoogle Scholar
  68. Legendre, Géraldine, William Raymond, and Paul Smolensky. 1993. An optimality-theoretic typology of case and grammatical voice systems. In 19th meeting of the Berkeley Linguistic Society. Also available on Rutgers Optimality Archive (ROA-3) at roa.rutgers.edu. Accessed 3 October 2016. Google Scholar
  69. Lehikoinen, Maija. 1973. Partisiippirakenteen totaalisen predikatiivin muodosta [On the form of total predicatives in participial constructions]. Virittäjä 77: 190–194. Google Scholar
  70. Levin, Theodore, and Omer Preminger. 2015. Case in Sakha: Are two modalities really necessary? Natural Language and Linguistic Theory 33: 231–250. CrossRefGoogle Scholar
  71. Maling, Joan. 1993. Of nominative and accusative: The hierarchical assignment of grammatical case in Finnish. In Case and other functional categories in Finnish syntax, eds. Anders Holmberg and Urpo Nikanne, 49–74. Berlin: de Gruyter. Google Scholar
  72. Maling, Joan. 2004. Case and grammatical functions: Structural vs. hierarchical approaches to case assignment. In 2004 Linguistic Society of Korea International Conference (LSK), Seoul: Linguistic Society of Korea, Yonsei Institute of Language and Information Studies. Google Scholar
  73. Maling, Joan. 2009. The case tier: A hierarchical approach to morphological case. In The Oxford handbook of case, eds. Andrej Malchukov and Andrew Spencer, 72–87. Oxford: Oxford University Press. Google Scholar
  74. Mallinson, Graham, and Barry J. Blake. 1981. Language typology: Cross-linguistic studies in syntax. Amsterdam: North-Holland. Google Scholar
  75. Malouf, Robert P. 2000. Mixed categories in the hierarchical lexicon. Stanford: CSLI Publications. Google Scholar
  76. Manninen, Satu, and Diane Nelson. 2004. What is a passive? The case of Finnish. Studia Linguistica 58(3): 212–251. CrossRefGoogle Scholar
  77. Marantz, Alec. 1991. Case and licensing. In Arguments and case: Explaining Burzio’s generalization, ed. Eric Reuland, 11–30. Amsterdam: John Benjamins. Google Scholar
  78. Matushansky, Ora. 2008. A case study of predication. In Studies in formal Slavic linguistics: Contributions from formal description of Slavic languages 6.5, eds. Franc Marušič and Rok Žaucer, 213–239. Frankfurt am Main: Peter Lang. Google Scholar
  79. McCarthy, John J. 2008. Doing optimality theory. Malden: Blackwell Sci. CrossRefGoogle Scholar
  80. McFadden, Thomas. 2004. The position of morphological case in the derivation: A study on the syntax-morphology interface, PhD diss., University of Pennsylvania. Google Scholar
  81. McFadden, Thomas. 2009. Structural case, locality, and cyclicity. In Explorations of phase theory: Features and arguments, ed. Kleanthes Grohmann, 107–130. Berlin: de Gruyter. CrossRefGoogle Scholar
  82. McFadden, Thomas, and Sandhya Sundaresan. 2011. Nominative case is independent of finiteness and agreement. Ms., Universitetet i Tromsø and CASTL (Tromsø)/Universität Stuttgart. Google Scholar
  83. Milsark, Gary. 1985. Case theory and the grammar of Finnish. In New England Linguistics Society (NELS), eds. Steve Berman, Jae-Woong Choe, and Joyce McDonough. Vol. 15, 319–331. Amherst: University of Massachusetts. Google Scholar
  84. Mitchell, Erika. 1991. Case and the Finnish object. Cornell Working Papers in Linguistics 9: 193–228. Google Scholar
  85. Mohanan, Tara W. 1994. Case OCP: A constraint on word order in Hindi. In Theoretical perspectives on word order in South Asian languages, eds. Miriam Butt, Tracy Holloway King, and Gillian Ramchand, 185–216. Stanford: CSLI Publications. Google Scholar
  86. Nelson, Diane. 1998. Grammatical case assignment in Finnish. New York: Garland. Google Scholar
  87. Nordlinger, Rachel. 1998. Constructive case: Evidence from Australian languages. Stanford: CSLI Publications. Google Scholar
  88. Partee, Barbara H., Alice ter Meulen, and Robert E. Wall. 1993. Mathematical methods in linguistics. Dordrecht: Kluwer. CrossRefGoogle Scholar
  89. Penttilä, Aarni. 1963. Suomen kielioppi [Finnish grammar]. Helsinki: WSOY. Google Scholar
  90. Pesetsky, David, and Esther Torrego. 2004. Tense, case, and the nature of syntactic categories. In The syntax of time, eds. Jacqueline Guéron and Jacqueline Lecarme, 495–538. Cambridge: MIT Press. Google Scholar
  91. Pesetsky, David, and Esther Torrego. 2011. Case. In The Oxford handbook of linguistic minimalism, ed. Cedric Boeckx, 52–72. Oxford: Oxford University Press. Google Scholar
  92. Polinsky, Maria, and Eric Potsdam. 2001. Long-distance agreement and topic in Tsez. Natural Language and Linguistic Theory 19: 583–646. CrossRefGoogle Scholar
  93. Pollard, Carl, and Ivan A. Sag. 1994. Head-driven phrase structure grammar. Stanford: CSLI Publications. Google Scholar
  94. Poole, Ethan. 2015. A configurational account of Finnish case, University of Pennsylvania Working Papers in Linguistics 21(1): Article 26. Available at: repository.upenn.edu/pwpl/vol21/iss1/26. Accessed 3 October 2016.
  95. Poole, Ethan. 2016. The locality of dependent case. Paper presented at the 34th West Coast Conference on Formal Linguistics (WCCFL), University of Utah. Google Scholar
  96. Prince, Alan. 2002a. Entailed ranking arguments. Available on Rutgers Optimality Archive (ROA-500) at roa.rutgers.edu. Accessed 3 October 2016.
  97. Prince, Alan. 2002b. Arguing optimality. Available on Rutgers Optimality Archive (ROA-562) at roa.rutgers.edu. Accessed 3 October 2016.
  98. Prince, Alan. 2007. The pursuit of theory. In The Cambridge handbook of phonology, ed. Paul de Lacy, 33–60. Cambridge: Cambridge University Press. CrossRefGoogle Scholar
  99. Prince, Alan, and Paul Smolensky. 1993/2004. Optimality theory: Constraint interaction in generative grammar. Malden: Blackwell Sci. Google Scholar
  100. Przepiórkowski, Adam. 1999. Case assignment and the Complement-Adjunct Dichotomy: A non-configurational constraint-based approach. PhD diss., Universität Tübingen. Google Scholar
  101. R Core Team. 2014. R: A language and environment for statistical computing. Vienna: R Foundation for Statistical Computing. Available at http://www.R-project.org/. Accessed 3 October 2016.
  102. Raposo, Eduardo, and Juan Uriagereka. 1990. Long distance case assignment. Linguistic Inquiry 21(4): 505–537. Google Scholar
  103. Riggle, Jason. 2010. Sampling rankings. Ms., University of Chicago. Available on Rutgers Optimality Archive (ROA-1075) at roa.rutgers.edu. Accessed 3 October 2016.
  104. Setälä, Eemil N. 1901. Suomen kielen lauseoppi: Oppikouluja varten [Finnish syntax for secondary schools], 5th edn. Helsinki: K.E. Holm. Google Scholar
  105. Sigurðsson, Halldór Ármann. 2006. The nominative puzzle and the low nominative hypothesis. Linguistic Inquiry 37: 289–308. CrossRefGoogle Scholar
  106. Taraldsen, Tarald. 1985. On the distribution of nominative objects in Finnish. In Features and projections, eds. Pieter Muysken and Henk van Riemsdijk, 139–161. Dordrecht: Foris. Google Scholar
  107. Timberlake, Alan. 1974. The nominative object in Slavic, Baltic, and West Finnic. Vol. 82 of Slavistische Beiträge. Munich: Otto Sagner. Google Scholar
  108. Toivainen, Jorma. 1993. The nature of the accusative in Finnish. In Case and other functional categories in Finnish syntax, eds. Anders Holmberg and Urpo Nikanne, 111–128. Berlin: de Gruyter. Google Scholar
  109. Toivonen, Ida. 1995. A study of Finnish infinitives. Senior thesis, Brandeis University. Google Scholar
  110. Vainikka, Anne. 1989. Deriving syntactic representations in Finnish. PhD diss., University of Massachusetts. Google Scholar
  111. Vainikka, Anne. 1993. The three structural cases in Finnish. In Case and other functional categories in Finnish syntax, eds. Anders Holmberg and Urpo Nikanne, 129–159. Berlin: de Gruyter. Google Scholar
  112. Vainikka, Anne. 2003. Postverbal case realization in Finnish. In Generative approaches to Finnic and Saami linguistics, eds. Diane Nelson and Satu Manninen, 235–266. Stanford: CSLI Publications. Google Scholar
  113. Vainikka, Anne. 2011. Genitive case in Finnish reconsidered. Biolinguistica Fennica Working Papers 2: 1–32. Google Scholar
  114. Vainikka, Anne, and Pauli Brattico. 2014. The Finnish accusative: Long distance case assignment under agreement. Linguistics 52(1): 73–124. CrossRefGoogle Scholar
  115. Vainikka, Anne, and Yonata Levy. 1999. Empty subjects in Finnish and Hebrew. Natural Language and Linguistic Theory 17(3): 613–671. CrossRefGoogle Scholar
  116. Vilkuna, Maria. 1989. Free word order in Finnish: Its syntax and discourse functions. Helsinki: Suomalaisen Kirjallisuuden Seura. Google Scholar
  117. Vilkuna, Maria. 2000. Suomen lauseopin perusteet [Foundations of Finnish syntax]. Vol. 90 of Kotimaisten kielten tutkimuskeskuksen julkaisuja. Helsinki: Oy Edita Ab. Google Scholar
  118. Wiik, Kalevi. 1972. Suomen akkusatiiviobjektin muoto [The form of the Finnish accusative object]. In Turun yliopiston fonetiikan laitoksen julkaisuja 12 [Publications of the Department of Phonetics, University of Turku, Finland]. Google Scholar
  119. Woolford, Ellen. 2000. Case patterns. In Optimality-theoretic syntax, eds. Géraldine Legendre, Jane Grimshaw, and Sten Vikner, 509–543. Cambridge: MIT Press. Google Scholar
  120. Woolford, Ellen. 2006. Lexical case, inherent case, and argument structure. Linguistic Inquiry 37(1): 111–130. CrossRefGoogle Scholar
  121. Wunderlich, Dieter, and Renate Lakämper. 2001. On the interaction of structural and semantic case. Lingua 111: 277–418. CrossRefGoogle Scholar
  122. Yip, Moira, Joan Maling, and Ray Jackendoff. 1987. Case in tiers. Language 63: 217–250. CrossRefGoogle Scholar
  123. Zaenen, Annie, Joan Maling, and Höskuldur Thráinsson. 1985. Case and grammatical functions: The Icelandic passive. Natural Language and Linguistic Theory 3: 441–483. CrossRefGoogle Scholar

Copyright information

© Springer Science+Business Media Dordrecht 2016

Authors and Affiliations

  1. 1.Department of LinguisticsStanford UniversityStanfordUSA
  2. 2.School of EnglishKyung Hee UniversitySeoulRepublic of Korea

Personalised recommendations