1 Introduction
In clausal ellipsis constructions, the sentential part of an utterance (i.e., IP, S, or TP depending on one’s preferred terminology) is not overtly pronounced, but some sub-part of the sentence may be overt. In (1), we have a simple case of sluicing in Icelandic, where the TP of a wh-question is not pronounced, leaving just the wh-phrase hvern ‘who.acc’ in SpecCP overt. Adopting Merchant’s (Reference Merchant2001) terminology, the overt wh-phrase is called the ‘remnant’.
Following Merchant (Reference Merchant2004), Griffiths & Lipták (Reference Griffiths and Lipták2012), and Weir (Reference Weir2014), among others, fragments receive the same analysis, with the pronounced material undergoing movement to the left periphery prior to TP deletion:Footnote [2]
In clausal ellipsis, the remnant typically corresponds, in some intuitive sense, to a particular phrase in the antecedent, called the ‘correlate’. In (1), the correlate for hvern ‘who.acc’ is einhvern ‘someone.acc’ and in (2), the correlate for mig ‘me.acc’ is hvern ‘who.acc’. We will refer to the clause containing the correlate as the ‘antecedent clause’ and the clause that undergoes ellipsis as the ‘ellipsis clause’.
In this paper, we investigate Icelandic clausal ellipsis, a relatively understudied phenomenon, with a special focus on fragment responses. For the most part, Icelandic clausal ellipsis is unexceptional in having the properties we expect from such constructions as found in many other languages. Independent properties of the Icelandic case system, however, shed potentially important light on the analysis of clausal ellipsis. Like most languages with robust case morphology, Icelandic generally requires the case of a remnant to match the case of the correlate. However, we show that under constrained circumstances, certain kinds of case mismatches are possible. We argue that the correct characterization of when such mismatches are possible and when they are impossible supports the existence of silent syntactic structure at the ellipsis site, as well as a syntactic identity condition on ellipsis (perhaps in addition to a semantic condition). Moreover, we will show how the Icelandic facts support a post-syntactic approach to at least some aspects of morphological case, and will develop an analysis within such an approach.Footnote [4]
2 Background
2.1 Case Matching
Ross (Reference Ross, Binnick, Davison, Green and Morgan1969) was the first to note that in clausal ellipsis, specifically sluicing, the remnant and correlate must match in case. We will refer to this as the ‘Case-Matching Generalization’ (CMG). The CMG is detectible in languages that overtly mark case on nominals, illustrated below with a German sluice. German schmeicheln ‘flatter’ assigns dative case to the correlate, whereas loben ‘praise’ assigns accusative; in (3a)–(3b) we see that the remnant must bear whichever case its correlate does.
Merchant (Reference Merchant2004) shows the same facts hold for fragments.Footnote [5] There has been very little research on Icelandic clausal ellipsis, but in the research that does exist, the same has generally been shown or assumed to hold. In one of the first papers taking a generative approach to Icelandic, Thráinsson (Reference Thráinsson, Grossman, San and Vance1975) focused on the role of case matching in Icelandic gapping constructions. E.F. Sigurðsson & Stefánsdóttir (Reference Sigurðsson and Stefánsdóttir2014) and E.F. Sigurðsson, Sigurjónsdóttir & Örnólfsdóttir (Reference Sigurðsson, Sigurjónsdóttir and Örnólfsdóttir2018) use case matching in fragment responses to probe the structure of the ‘New Impersonal Passive’ construction. Ott (Reference Ott2014) and Ott & de Vries (Reference Ott and de Vries2016) argue that contrastive left dislocation and right dislocation in Icelandic and related languages should be analyzed as clausal ellipsis, and the fact that these constructions show case matching (or case-connectivity) is part of what is argued to be explained by this analysis.
Case-matching effects are robustly attested in languages with overt case marking. Some counterexamples have been uncovered thus far in a few languages.Footnote [6] Ince (Reference Ince2012) notes that Turkish genitive correlates correspond to nominative remnants obligatorily in sluicing. Kim (Reference Kim2015) discusses case mismatches in Korean, where an expected case suffix can be optionally dropped only if there is an overt correlate (but not when there is not, as in what is referred to as ‘sprouting’ in the ellipsis literature). Vicente’s (Reference Vicente2015) short survey cites counterexamples attested in Mongolian, Korean, Uzbek, Japanese, German and Chamorro, though it has been argued for some of these, namely Japanese and Uzbek, that what appears to be sluicing is actually a reduced copular clause, so that the relevance of these languages to the status of the CMG is questionable.Footnote [7]
The CMG is often assumed to follow from two assumptions: first, that there is silent syntactic structure in the ellipsis clause, as described above, and second, that this structure is identical to the structure of the antecedent clause. Under the assumption that the remnant is extracted from fully present, though unpronounced, syntactic structure, we expect its case to match that of the correlate, since they both share identical base positions at the relevant level of representation, as illustrated in (4a)–(4b).
However, these two assumptions are at odds with interpretive approaches, which reject the assumption that there is silent structure (Ginzburg & Sag Reference Ginzburg and Sag2000; Culicover & Jackendoff Reference Culicover and Jackendoff2005, Reference Culicover and Jackendoff2012; Sag & Nykiel Reference Sag, Nykiel and Müller2011; Barker Reference Barker2013; Jacobson Reference Jacobson2016). Such theories have various ways of resolving the interpretation of the remnants, but they generally must stipulate case matching directly, rather than derive it.Footnote [8] For example, Sag & Nykiel (Reference Sag, Nykiel and Müller2011: 196), in an HPSG framework, claim that ‘a grammatical constraint must dictate directly that there be identity of (category and) case between the remnant and its correlate’ (as in Ginzburg & Sag (Reference Ginzburg and Sag2000)). Barker (Reference Barker2013), in a Categorial Grammar framework, makes case values part of the syntactic category, and requires that the remnant matches the correlate in case value as part of how it combines syntactically with the antecedent clause; see Jacobson (Reference Jacobson2016) for a similar mechanism.Footnote [9]
As pointed out by Sag & Nykiel (Reference Sag, Nykiel and Müller2011) and Jacobson (Reference Jacobson2016), silent structure theories that derive case-matching and interpretive theories that encode it directly make different predictions for verbs which can assign more than one case. For example, Jacobson (Reference Jacobson2016) points out that the Hungarian verb hasonlít ‘resemble’ may assign sublative or allative case to its object, with no difference in meaning. Silent structure theories, in principle, predict that case mismatches should be possible, whereas interpretive theories that encode case matching directly predict that case mismatches should not be possible. According to Jacobson (Reference Jacobson2016: 356), mismatches are in fact not possible, confirming the prediction made by interpretive theories.
However, we will show below that Icelandic does tolerate case mismatches in fragments in some cases. This undermines the idea that case matching – which is certainly the norm in Icelandic – is directly encoded in the way that it is in interpretive approaches. We will show that Icelandic also has instances that disallow case mismatches, much as in the Hungarian facts above.Footnote [10] We argue that the nature of when such mismatches are allowed and when they are not allowed supports silent structure approaches to ellipsis, along with a hybrid syntactic and semantic identity condition along the lines of Chung (Reference Chung2006, Reference Chung2013).Footnote [11]
The remainder of the paper is organized as follows. Section 2.2 provides a basic overview of the Icelandic case system, especially the aspects of it that are relevant to the present paper. Section 3 discusses case matching in Icelandic fragments, and introduces the identity condition that we will adopt. Section 4 discusses case mismatches, and works out when such mismatches are possible and when they are impossible. Section 5 details the analysis of the case alternations and shows how the availability and unavailability of mismatching follows from the identity condition first presented in Section 3. Section 6 concludes, highlighting the broader implications of the study.
2.2 Icelandic Case
In this section we provide a brief overview of the Icelandic case system. Icelandic distinguishes four morphologically distinct cases (nominative, accusative, dative and genitive). The majority of the time, subjects are nominative, direct objects are accusative, and indirect objects are dative (Barðdal Reference Barðdal2001), as illustrated in the ditransitive sentence in (7).
However, the relationship between case and grammatical role is by no means one-to-one. In addition to the canonical cases given above, subjects can be accusative, dative or genitive; indirect objects can be accusative; and direct objects can be nominative, dative or genitive.Footnote [12] Not all combinations of these are possible. Ditransitives, for example, always have a nominative subject, but may otherwise have dat-acc, dat-dat, dat-gen, acc-dat, acc-gen, or (very rarely) acc-acc objects (Zaenen, Maling & Thráinsson Reference Zaenen, Maling and Thráinsson1985, Yip, Maling & Jackendoff Reference Yip, Maling and Jackendoff1987, Jónsson Reference Jónsson2000). Verbs with dative subjects are generally either intransitive or take a nominative object. Verbs with accusative subjects are generally either intransitive or take an accusative object.
Descriptively, it seems as though different verbs select different case frames. For example, the verb leiðast ‘be bored’ takes a dative subject and, optionally, a nominative object. The verb aka ‘drive’ takes a dative object, while the verb keyra, also ‘drive’, generally takes an accusative object (although some speakers allow dative, as discussed below). Studies of the verbs involved in case-marking patterns beyond the ordinary nom-(dat-)acc pattern reveal various lexical-semantic subregularities, but also a lot of idiosyncrasy.
There is also a fair amount of variation in certain case-marking patterns, and recent studies have shown that there may be even more variation than previously thought (Árnadóttir & E.F. Sigurðsson Reference Árnadóttir and Sigurðsson2013, Thráinsson et al. Reference Thráinsson, Eythórsson, Svavarsdóttir, Blöndal, Thráinsson, Angantýsson and Sigurðsson2015, Eythórsson & Thráinsson Reference Eythórsson and Thráinsson2017, Jónsson Reference Jónsson2017, Nowenstein Reference Nowenstein2017). Connected to the present study, we will introduce just a few of the well-known cases of case variation. First, Dative Substitution involves a verb that historically took an accusative experiencer subject taking a dative subject instead.Footnote [13]
The choice between cases makes no semantic difference (Jónsson & Eythórsson Reference Jónsson and Eythórsson2005: 235–236; H.Á. Sigurðsson Reference Sigurðsson2012b: 197). There is a lot of both inter- and intra-speaker variation, such that some speakers may find dative ungrammatical with these verbs, whereas others may only find dative grammatical. Many speakers, however, show intra-speaker variation, allowing both cases. (See discussion in, e.g., Halldórsson (Reference Halldórsson1982), Svavarsdóttir (Reference Svavarsdóttir1982), Eythórsson (Reference Eythórsson and Lightfoot2002), Jónsson (Reference Jónsson, Brandner and Zinsmeister2003), Barðdal (Reference Barðdal2001, Reference Barðdal2011), Jónsson & Eythórsson (Reference Jónsson and Eythórsson2003, Reference Jónsson and Eythórsson2005), Eythórsson & Jónsson (Reference Eythórsson, Jónsson, Dufter, Fleischer and Seiler2009), Viðarsson (Reference Viðarsson2009), Ingason (Reference Ingason2010), Nowenstein (Reference Nowenstein2012, Reference Nowenstein2014a, Reference Nowensteinb, Reference Nowenstein2017), and further discussion below.) The most common Dative Substitution verbs are langa ‘want’ and vanta ‘need’. Other verbs are attested as well, but there is more inter-speaker variation for them.
An interesting kind of Dative Substitution is found with the verbs hlakka til ‘look forward to’ and kvíða fyrir ‘be anxious about’.
In these cases, the verb traditionally took a nominative subject, and people have started to use accusative and/or dative instead. From a diachronic perspective, these cases are interesting in that they involve a rare instance of an oblique subject case replacing a structural nominative (Jónsson & Eythórsson Reference Jónsson and Eythórsson2005, Reference Jónsson, Eythórsson, Simon and Wiese2011). For our purposes, they will turn out to be interesting because speakers may allow three different cases on the subjects of such verbs, allowing us to investigate the nature of case matching and mismatching in more detail.
Second, Nominative Substitution involves a verb that historically took an accusative theme subject taking a nominative subject instead.
Once again, the choice between cases makes no semantic difference, but there is a lot of inter- and intra-speaker variation. Although both Dative Substitution and Nominative Substitution have some similarities – both involve the loss of traditional accusative subjects, for example, and both involve intra-speaker variation – researchers generally distinguish them as they show somewhat different properties. For example, while Dative Substitution seems to be on the increase, it has been claimed that Nominative Substitution is historically stable variation. Moreover, Dative Substitution seems to involve ‘thematic case’ (case marking connected with theta-role assignment) replacing ‘idiosyncratic case’ (case marking connected idiosyncratically with specific verbs), while Nominative Substitution seems to involve ‘structural case’ replacing idiosyncratic case (Jónsson Reference Jónsson, Brandner and Zinsmeister2003).
Finally, we note briefly that there is variation in object cases as well. There are many subcases, and they are not all relevant to the study here (see E.F. Sigurðsson (Reference Sigurðsson2017) for extensive discussion). What is most relevant is that many verbs vary between assigning accusative or dative to their objects (Maling Reference Maling2002a, Reference Malingb; Jónsson Reference Jónsson, Barðdal and Chelliah2009, Reference Jónsson2013a). We will discuss such cases in more detail below.
3 Icelandic Fragment Responses: Case Matching
There are two main issues in the literature on clausal ellipsis constructions. First, is there unpronounced syntactic structure at (what looks like) the ellipsis site? Second, what is the nature of the identity condition relating the silent (or understood) material to the antecedent clause? Existing approaches to the latter question posit either a syntactic identity condition, a semantic identity condition, or a hybrid condition. A syntactic identity condition would say that the syntax of the elided material has to be identical to some aspect of the antecedent clause. A purely semantic identity condition would say that the elided material can be anything, as long as it connects in some semantically defined way to the antecedent clause. A hybrid condition incorporates both syntactic and semantic factors.
It is well known that both solely syntactic and solely semantic identity conditions face significant empirical challenges. Merchant (Reference Merchant2001) shows that purely syntactic approaches undergenerate, whereas Chung (Reference Chung2006) shows that purely semantic approaches overgenerate. Various hybrid approaches have therefore been proposed, often adopting an overarching semantic identity condition alongside one or more syntactic codicils to reign in overgeneration (Merchant Reference Merchant2005; Chung Reference Chung2006, Reference Chung2013; AnderBois Reference AnderBois2011; Barros Reference Barros2014; Weir Reference Weir2014). In this section, we discuss case matching in Icelandic fragment responses, with an eye toward identifying the relevant identity conditions on ellipsis.
Like in other case-rich languages that have been studied (see references above), fragment responses in Icelandic generally require case matching. We show this for dative indirect objects in (11), and subjects in (12) (for a nominative subject) and (13) (for a dative subject).Footnote [14]
The same holds for direct objects, as illustrated in (14) and (15). The (A) examples show that aðstoða ‘assist’ takes an accusative object while hjálpa ‘help’ takes a dative object. The (B) examples show that the fragment answers must match the case of the correlate.
Consider these facts from the perspective of a purely semantic identity condition, assuming silent syntactic structure, as in Merchant (Reference Merchant2001, Reference Merchant2004). In Merchant’s (Reference Merchant2004) proposal for fragments, TP can be elided under focus-assisted mutual entailment, that is, when the antecedent entails the existential focus closure (in the sense of Schwarzschild (Reference Schwarzschild1999)) of the ellipsis clause, and vice versa. For example, in (15), (A) means that John helped speaker A, whereas (B) means that John helped speaker B. These of course are not mutually entailing. If we construe the objects as focused material, however, we factor them out, and both sentences have the meaning in (16) as their existential focus closure.
(A) now entails the existential focus closure of (B), and vice versa. The sentences are identical at this level, so they are mutually entailing, and ellipsis is possible.Footnote [15]
However, by itself, this does not necessarily explain case matching. Suppose, for example, that speaker B in (14) responded with dative, with the structure in (17).
That is, speaker A uses the verb aðstoða ‘assist’, which assigns accusative case, but speaker B responds with the verb hjálpa ‘help’, which assigns dative case, and elides the TP. Should this be possible? Certainly the non-elliptical change of verb is possible.
To the extent that ‘assist’ and ‘help’ mean the same thing (see Svenonius (Reference Svenonius2002) for discussion), if John assisted speaker A, it is also true that John helped speaker A; speaker B can respond fully naturally that John helped him/her as well. This kind of response is a fully natural and coherent discourse. However, all of that is not enough to license ellipsis of everything but the focus, as in (14). Part of the question hinges on whether (19) and (20) are mutually entailing. Are the sets of people who John helped and the people John assisted distinct sets?
Jacobson (Reference Jacobson2016) emphasizes that any condition of this sort would have to be based on truth conditions, and not implicatures or other layers of meaning. She illustrates the point with the following pair.
B’s response here clearly lacks A’s implicature about their attitude toward Bozo, which does not prevent ellipsis from being licensed. In most cases, the distinction between hjálpa ‘help’ and aðstoða ‘assist’ is arguably even more subtle. In her discussion of the Hungarian verb segít ‘help’, which gets slightly different meanings depending on the case of the object (related to telicity), Jacobson (Reference Jacobson2016: 19) suggests that it may be enough, to make the point, to construct truth-conditionally equivalent minimal pairs.Footnote [16]
In fact, it has long been noted that Icelandic has many pairs of verbs that mean more or less the same thing but assign different cases to their subjects or objects (Andrews Reference Andrews and Bresnan1982, Zaenen & Maling Reference Zaenen, Maling, Cobler, MacKaye and Wescoat1984). They all pattern like (14) and (15) in fragments. The verbs aka and keyra, both ‘drive’, take dative and accusative objects, respectively.Footnote [17]
The two verbs mean the same thing in that they are mutually entailing, so that examples like (23) are contradictions. (Note that Jacobson (Reference Jacobson2016: 358) makes her argument about case marking with hasonlít ‘resemble’ using contradiction in this way.)
As above, speakers can switch between keyra and aka and stay fully coherent, relevant, etc.
However, aka ‘drive’ does not license case mismatches, despite the existence of a synonymous verb that assigns a different case.
Thus, if we assume silent syntactic structure, it is not enough to propose a semantic identity condition like focus-assisted mutual entailment. At the very least, it would seem that we need to refer to the lexical content of the verb and make sure it is identical to the antecedent (or directly enforce case matching, as in Barros (Reference Barros2014) or the interpretive approaches cited above). If we make sure that the verbs are the same, the above facts follow from an ellipsis analysis of fragment responses along the lines of (26) and (27): the remnant is moved to a high position (perhaps [Spec,CP]), and the rest of the sentence is deleted under syntactic identity with the antecedent clause in the question (Merchant Reference Merchant2001, Reference Merchant2004).
Assuming this much, (27) cannot be used to derive dative case in (14B), and (26) cannot be used to derive accusative case in (15B).Footnote [18]
What kind of identity condition will make sure that the verbs are the same? Here we adapt the limited syntactic identity condition proposed by Chung (Reference Chung2013).Footnote [19]
In unpacking these, we would first like to point out that perhaps contrary to what one might expect, the Case condition will not derive Icelandic case-matching effects. Chung (Reference Chung2013) makes it clear that the Case condition is intended to apply to abstract Case. It derives, for example, the contrast between (29b) and (30b).
In (29b), the remnant in the ellipsis clause is licensed by finite T, whereas in the antecedent clause, the correlate DP (PRO) is licensed by non-finite control T. In (30b), however, the remnant in the ellipsis clause is licensed by the verb (or little v/Voice), just the same as in the antecedent clause.
In Icelandic, abstract Case is crucially distinct from morphological case, as has been much discussed (Jónsson Reference Jónsson1996, H.Á. Sigurðsson Reference Sigurðsson2012b, Árnadóttir & E.F. Sigurðsson Reference Árnadóttir and Sigurðsson2013).Footnote [20] For example, PRO in Icelandic is licensed in similar environments as English, but can be shown by various agreement phenomena to be case-marked (H.Á. Sigurðsson Reference Sigurðsson1991, Reference Sigurðsson2008; Landau Reference Landau2003, Reference Landau2006).
In (31), PRO is marked dative because it is the subject of leiðast ‘be bored’, as can be seen from the fact that the floating modifier einum ‘alone’ agrees with it in case. But although its case marking is determined by the verb (or perhaps an Appl head selected by the verb, as discussed below), it is Case-licensed by non-finite control T, leading to it being null PRO.
Now consider examples of the sort discussed by Chung (Reference Chung2013).
Despite the fact that PRO and the remnant in (32b) get their morphological case (dative) determined by the same head (the verb leiðast ‘be bored’ or the Appl head it projects), ellipsis is not possible. A straightforward reason why is that this is for the same reason as the English examples above: the sentence does not satisfy the Case condition of (28b). Although PRO and the remnant get their morphological case from the same place, PRO is Case-licensed by non-finite control T, and the remnant is Case-licensed by finite T. This shows that case-matching effects, and the mismatches discussed below, will not derive from the Case condition, and we will not discuss the Case condition further in this paper.
Instead, we will pursue a solution in terms of the Argument Structure Condition in (28a). As worded, it automatically requires that the verbs match in the ellipsis clause and the antecedent clause: there must be a ‘corresponding predicate’, which we can assume must be identical. (28a) also derives the impossibility of several other well-known mismatches (examples taken from Chung, Ladusaw & McCloskey (Reference Chung, Ladusaw, McCloskey, Gutiérrez, Mikkelsen and Potsdam2011: 37), Merchant (Reference Merchant2013: 82), Weir (Reference Weir2014: 142–144)).
(33a) is ungrammatical because the antecedent clause is active and the ellipsis clause is passive (see Merchant (Reference Merchant2013) for extensive discussion). (33b) is ungrammatical because the antecedent clause has a different argument structure from the ellipsis clause. The ungrammaticality of mismatches such as these have been taken to motivate the claim that some kind of syntactic identity condition must be involved, if one assumes silent syntactic structure at the ellipsis site.
The wording of the ‘Argument Structure Condition’ in (28) above leads to the question of what the representation of argument structure is, and what it means for it to be the same or different. In what follows, we will adopt a syntactic approach to argument structure (Pylkkänen Reference Pylkkänen2002; Cuervo Reference Cuervo2003; Borer Reference Borer2005a, Reference Borerb, Reference Borer2013; Ramchand Reference Ramchand2008; Schäfer Reference Schäfer2008; Wood Reference Wood2015). According to this view, different verbs (or verb roots) may appear with different combinations of verbal heads in the VoiceP domain. Having ‘identical argument structure’ thus means having the same such heads, with the same features. This allows us to ground the Argument Structure Condition in more concrete terms, in a way that is inspired by what has come to be known as the ‘No New Words’ constraint of Chung (Reference Chung2006: her (29)):
Assuming that the domain of argument structure is VoiceP, and that we want that structure to be identical (as in (28a) above), we can adopt a slightly different version, which makes no claims about material outside of VoiceP:Footnote [21]
This says that argument structure must be identical by stating that an elided VoiceP must reuse the same lexical items – including roots and functional heads – used in the VoiceP of the antecedent clause.
This accounts for the impossibility of the mismatches in (33) above, which use different Voice heads in (33a) and different prepositions and verbal heads in (33b).Footnote [23] We will show below that, when combined with independently motivated analyses of Icelandic case phenomena, an identity condition along the lines of (35) can account for the existence of case matching and case mismatching, correctly predicting when each may occur.Footnote [24]
4 Case Mismatches and Impossible Mismatching
In contrast to the approaches discussed in the previous section, some approaches to fragments do not assume silent syntactic structure. For those approaches, the question is not how to constrain the content of silent structure, but how to specify the properties of the fragment to connect it to the antecedent clause. One way this is done is by specifying case matching (or something close to it) directly, as in Ginzburg & Sag (Reference Ginzburg and Sag2000), Culicover & Jackendoff (Reference Culicover and Jackendoff2005, Reference Culicover and Jackendoff2012), Sag & Nykiel (Reference Sag, Nykiel and Müller2011). Another is to connect the fragment syntactically to the antecedent clause, as in Barker (Reference Barker2013) and Jacobson (Reference Jacobson2016), where case matching is again directly encoded in the way that such structures are built. The impossibility of mismatching with verbs that assign multiple cases has been taken as an argument in favor of this approach. As we will see in this section, however, certain mismatches of exactly this sort are possible in Icelandic. We will argue that this supports silent structure with at least a hybrid identity condition referring to that structure, specifically the Argument Structure Condition as formulated in (35) above.Footnote [25]
4.1 Case Mismatches
As discussed above, case mismatches are ordinarily not possible in Icelandic fragment responses. We have shown that this holds for subjects, direct objects and indirect objects, regardless of whether the case is structural (as with nominative subjects and accusative objects) or oblique (as with dative subjects and objects). However, recall that some verbs in Icelandic assign more than one possible case to their subjects or objects. This is so for langa ‘want’, which takes either an accusative subject (standardly) or a dative subject (under Dative Substitution).
Unlike the Hungarian facts discussed above, the availability of accusative or dative with langa ‘want’ does make a case mismatch available.
While we will illustrate most matching and mismatching effects with exchanges along the lines of the above examples, we should point out that such mismatches are by no means limited to them. (39) shows that such mismatches are possible in question–answer pairs and (40a) shows that this is even possible within one sentence.Footnote [26] (40b) shows that such mismatching is not possible with a verb like vilja ‘want’, which, as illustrated in (12) above, only takes a nominative subject.
The same paradigm is found for other Dative Substitution verbs, such as vanta ‘need’ illustrated in (41)–(42) below.
We will see another example of a dative/accusative subject alternation below with the verb hlakka til ‘look forward to’, which we discuss separately because there, the possibility of nominative also comes into play. That is, this is not a fact about just one or two verbs, it is a general fact about what happens with Dative Substitution verbs and other verbs that vary between accusative and dative subjects.Footnote [27] Footnote [28]
A similar dative/accusative mismatch can be found in object position. As mentioned above, some speakers allow keyra ‘drive’ to take either an accusative or dative object (Jónsson Reference Jónsson, Barðdal and Chelliah2009: 209).
Such speakers allow mismatches of the sort described above.
These facts show clearly that case matching is not a surface true generalization, and therefore cannot be stipulated as such. How do we reconcile this with the observation above that Icelandic generally does require case matching? As a first pass, the data so far seems to suggest a generalization as in (45).
This would look distinct from the Hungarian situation above. However, we will see several ways in which (45) overgenerates for Icelandic as well, leading us to refine the Case-Mismatching Generalization and the analysis of it.
4.2 Impossible Mismatching
It turns out that it is not enough that a verb can assign two different cases (to the same argument), however. In this section we discuss several ways in which case mismatching is impossible, even when the verb assigns more than one case in principle. The first case we discuss is when the object case makes a semantic difference. We then turn to alternations between nominative and oblique subjects, where there is some speaker variation, despite the absence of a semantic distinction, and we discuss how this supports the claims in this paper.
4.2.1 Object Case Makes a Semantic Difference
It has long been known that some kinds of case alternations do have semantic consequences. For example, there are distinct classes of verbs which can take either an accusative or a dative object (H.Á. Sigurðsson Reference Sigurðsson1989, Barðdal Reference Barðdal1993, Maling Reference Maling2002a, Svenonius Reference Svenonius2002). For one such class, if the dative is chosen, the object is understood to benefit from the event. Consider the example in (46):
If accusative is chosen, it means she affected me physically, and probably hurt me or damaged my skin. If dative is chosen, it means I benefitted from the event, as if she had scratched me kindly or scratched an itch. With case alternations like this, a case mismatch in fragment answers is not possible.Footnote [29]
Another, more subtle example comes from cases discussed by Jónsson (Reference Jónsson2013a), drawing in part on the references above. Jónsson (Reference Jónsson2013a) noticed that verbs of contact, like skalla ‘(hit with one’s) head’, can take either accusative or dative objects.Footnote [30]
According to Jónsson (Reference Jónsson2013a: 154), ‘While both the accusative and the dative variant assert contact with the object, only the latter variant asserts motion of the object.’ Thus, a sentence with the dative entails the corresponding sentence with an accusative, but not vice versa.
Similarly to klóra ‘scratch’ above, case mismatches with skalla ‘head’ are not possible (although the contrast is perhaps sharper with klóra ‘scratch’ than with skalla ‘head’, as pointed out to us by Jóhannes Gísli Jónsson).
The contrast between Dative Substitution and objects of keyra ‘drive’, which seem to allow case mismatches, and the examples in this subsection, which do not, seems to relate to the fact that when mismatch is impossible, a difference in case assignment correlates with a difference in interpretation, whereas when mismatch is possible, it does not. While this may seem to suggest that the identity condition on ellipsis is semantic, the independently proposed syntactic identity condition makes this unnecessary. Specifically, this follows from the Argument Structure Condition, as in (28a), or our revised version in (35). The different meanings, and case patterns connected to them, in fact correspond to different argument structures. These argument structures may stem from a phrase structural difference, so that there is a different tree geometry for, say, dative vs. accusative, as proposed in Jónsson (Reference Jónsson2013b) and E.F. Sigurðsson (Reference Sigurðsson2015, Reference Sigurðsson2017). Alternatively, it could be that the dative and accusative variants involve featurally distinct VoiceP-internal heads as in Schäfer (Reference Schäfer2008) and Wood (Reference Wood2015).
4.2.2 Alternations With Nominative
In the previous sections, we have seen possible and impossible mismatches connected with dative and accusative subjects and objects. In this section, we discuss some alternations with nominative subjects. The facts here will undermine a purely semantic identity condition, but are compatible with a syntactic identity condition, given independently motivated assumptions about how nominative/oblique alternations work. Importantly, in all of the alternations discussed in this subsection, the difference in case assignment has always been said to make no semantic difference (Eythórsson Reference Eythórsson2000: 33; H.Á. Sigurðsson Reference Sigurðsson, Alexiadou, Hankamer, McFadden, Nuger and Schäfer2009: 266; Reference Sigurðsson2012b: 197, 204).
We begin with the verb hlakka ‘look forward to’, which is interesting because it is a rare case where a traditionally nominative subject verb began to be used with accusative and/or dative for some speakers.Footnote [31]
There is a lot of inter-speaker variation with hlakka, even before we get to the question of case (mis)matching in ellipsis. However, one clear fact is that for speakers who allow dative or accusative, the mismatch is possible as indicated earlier.
At this point, hlakka simply adds another example of a verb that takes either accusative or dative, with no difference in meaning, and allows mismatching.
When nominative is involved, there is more variation. In the absence of ellipsis, speakers can respond freely with the case of their choice.
When the response is a fragment, however, some speakers reject nominative in the response when dative or accusative are used initially.
We will refer to two groups of speakers: Group A speakers, who reject the mismatch, and Group B speakers, who allow it.Footnote [32] As we will discuss further below, the existence of both groups of speakers raises interesting issues regarding the identity condition. In short, the existence of Group A speakers seems to support a syntactic identity condition over a semantic one. The existence of Group B speakers will be argued to be compatible with this conclusion. When nominative is used initially, we see a similar picture.
Most speakers accept (59A), so case matching in the response is generally an option.Footnote [33] As for dative and accusative, we have a similar division between Group A speakers, who disallow the mismatch, and Group B speakers, who allow it. We wish to emphasize here that the Group A speakers under consideration are speakers who generally allow accusative or dative with hlakka, and thus do allow case mismatches, as described in (54)–(55) above; they simply do not allow such mismatches with nominative.
We find a similar picture with another nominative alternation, known as Nominative Substitution. Nominative Substitution describes verbs that historically took an oblique case – here, accusative – but began for many speakers to be possible in the nominative (Eythórsson Reference Eythórsson2000, Jónsson Reference Jónsson, Brandner and Zinsmeister2003, Jónsson & Eythórsson Reference Jónsson and Eythórsson2005, Eythórsson & Thráinsson Reference Eythórsson and Thráinsson2017).
In these cases, we again find that in non-elliptical responses, speakers can use whichever subject case they like. We illustrate this with reka ‘run ashore’.
In elliptical responses, there is variation in whether mismatch is allowed, essentially along the lines discussed above for hlakka ‘look forward to’.
All speakers allow case matching, while only some allow nominative/accusative mismatches with Nominative Substitution verbs. Importantly, the case variation has no semantic effect in these cases, whether mismatching is allowed or not. As we discuss below, this casts doubt on a purely semantic identity condition, but is compatible with a syntactic identity condition. We will suggest below that for some speakers, variation in accusative and nominative is grounded in a syntactic difference (despite there being no semantic effect of this syntactic difference), while in other speakers it is not. The availability of case mismatching under ellipsis is connected with the way that speakers treat nominative/accusative mismatches more generally.
4.2.3 Case Mismatching Revised
We have now presented two ways in which our original Case-Mismatching Generalization, presented in (45), overgenerates. First, when a difference in case assignment corresponds to a semantic difference, case mismatching may not be possible. Second, for some speakers, mismatches involving nominative and oblique may not be possible, despite there being no semantic effect of the case alternation. As we alluded to, we will argue below that for such speakers, the distinction between nominative and oblique subjects is grounded in a syntactic difference, despite there being no semantic effect of this difference. We therefore revise our generalization as follows:
We have already shown how the semantic effects follow from a syntactic identity condition (specifically the Argument Structure Condition). In the next section, we turn to a detailed analysis of the other case alternations, and show how those facts too follow from such a condition.
5 Case (Mis)Matching and the Syntactic Identity Condition
In this section, we propose that the constellation of facts presented above supports a syntactic identity condition on ellipsis. There is silent syntactic structure at the ellipsis site, and this structure must be identical, in the relevant respects, to the antecedent clause. First we provide an overview of how a syntactic identity condition derives the facts presented above. We then describe in more detail Dative Substitution, followed by Nominative Substitution.
First, let us repeat the relevant syntactic identity condition from Chung (Reference Chung2013). For reasons discussed above, we do not discuss the Case condition, but focus on the Argument Structure Condition.
Adopting a syntactic (‘neoconstructivist’) approach to argument structure, we reformulated the Argument Structure Condition along the lines of Chung’s (Reference Chung2006) ‘No New Words’ constraint.
A syntactic identity condition such as (28) and (70) derives case matching as the general case, much as Ross (Reference Ross, Binnick, Davison, Green and Morgan1969) anticipated. Morphological case assignment generally reflects either argument structure or Case licensing. If the argument structure heads and Case-licensing heads must be the same in the ellipsis clause as in the antecedent clause, then the case of the remnant will have to be the same as the corresponding DP in the antecedent clause. We see this as the basic fact of the system that needs to underlie the account of when mismatching is and is not allowed.
The Dative Substitution facts then suggest that the argument structure is the same, whether the subject is dative or accusative, and that the Case licensing is identical. The second point is not controversial, and as for the first, we will propose that Dative Substitution alternations are not encoded anywhere in the syntax, but are instead post-syntactic. They thus meet the requirements for ellipsis according to the identity condition. However, applying such post-syntactic case-manipulation rules crucially requires the presence of silent syntactic structure.
Finally, the impossibility of mismatch with nominatives for Group A speakers suggests that for such speakers, sentences with nominative subjects have a distinct syntax from those with oblique subjects. In fact, there is independent support for this. Jónsson (Reference Jónsson, Brandner and Zinsmeister2003) has argued that nominative subject experiencers are systematically distinct from oblique subject experiencers. For example, oblique subject experiencers can never passivize, while nominative subject experiencers can passivize at least sometimes. Various authors have argued that accusative-subject constructions actually have a silent external argument present in the syntax (Haider Reference Haider2001; Platzack Reference Platzack2006; Schäfer Reference Schäfer2008, Reference Schäfer2012; Wood Reference Wood2017), which is not (necessarily) present in nominative subject unaccusative constructions. Therefore, we propose that Group A speakers have a genuine syntactic difference correlating with the nominative/oblique distinction, making case mismatches impossible: the accusative-subject variant takes a silent external argument, and the nominative subject variant does not. As for Group B speakers, the simplest claim is that this is syntactic microvariation: Group B speakers do not make the syntactic distinction that Group A speakers do. One can imagine various implementations of this idea, but we will present one specific proposal based on Wood (Reference Wood2015, Reference Wood2017).Footnote [34]
5.1 Dative Substitution
In this section, we flesh out and support the idea that Dative Substitution alternations are not encoded in the syntax. There is a longstanding intuition in work on Icelandic case variation that Dative Substitution involves ‘thematic lexical Dat case’ replacing ‘idiosyncratic lexical Acc’ (Eythórsson & Thráinsson Reference Eythórsson and Thráinsson2017: 61) (emphasis added). The idea is that people have to memorize, word by word, which experiencer verbs want an accusative subject. With dative experiencers, there is a more regular pattern of assigning dative to experiencers of a certain sort. Dative Substitution reflects that pattern. We will first discuss how we derive ‘thematic lexical’ dative, and then turn to ‘idiosyncratic lexical’ accusative.
5.1.1 Thematic Lexical Dative
In a syntactic theory of argument structure, experiencers are dative because they are merged in a syntactic position that then corresponds to dative, such as ApplP (Cuervo Reference Cuervo2003, McFadden Reference McFadden2004, Wood Reference Wood2015).
Dative could be assigned by the Appl head to its specifier in the syntax (E.F. Sigurðsson Reference Sigurðsson2017). Alternatively, in the spirit of McFadden (Reference McFadden2004, Reference McFadden, Hole, Meinunger and Abraham2006) and H.Á. Sigurðsson (Reference Sigurðsson2012a, Reference Sigurðssonb), there could be a general post-syntactic rule to the effect that dative case is added to a DP base-generated in SpecApplP.
The pattern that many experiencers are dative stems from the ability of Appl to assign an experiencer role, along with the general post-syntactic rule in (72).Footnote [35] This is a syntactic implementation of the ‘thematic lexical’ dative intuition, the idea that the dative case is somehow connected to the thematic role. Here, it is Appl that is responsible for dative case and the experiencer role.
5.1.2 Idiosyncratic Lexical Accusative
Accusative experiencer subjects are subject to a far less ‘regular’ rule (though see Ingason (Reference Ingason2010) for some attenuation of this claim), and must be memorized word by word. Following McFadden (Reference McFadden2004), suppose we assume that they have the same general syntactic structure as dative experiencers, namely (71). The accusative arises by a lexically specific impoverishment rule that applies after (72) in the morphology.
First, the features [+inferior, +oblique] (=dative) are assigned to a DP in SpecApplP. Then, for some verbs, the impoverishment rule in (74) applies, resulting in the loss of the [+oblique] feature, leading to accusative case. The lexically idiosyncratic nature of accusative experiencer subjects is captured by assuming that this impoverishment rule only applies to a specific, memorized list of verbs.
5.1.3 Dative Substitution
Under this view, Dative Substitution makes sense in the following way: at a deeper morphological level, accusative experiencers want to be – or even are – dative. An extra morphological rule has to apply to make them accusative. But impoverishment rules can be the source of considerable variation (see e.g., Nevins & Parrott (Reference Nevins and Parrott2007)). If or when the impoverishment rule does not apply, it is dative that will surface. Thus, Dative Substitution is not syntactic – it is morphological. This explains why dative/accusative alternations are no problem for a syntactic identity condition. The syntax is exactly the same, with the same Appl head introducing the experiencer.Footnote [36]
As a morphological phenomenon, Dative Substitution is sensitive to very specific, somewhat idiosyncratic pressures. As documented by Nowenstein (Reference Nowenstein2012, Reference Nowenstein2014a, Reference Nowensteinb, Reference Nowenstein2017), whether DS applies is sensitive to person: 3rd person plural subjects are most likely to be dative, while 1st and 2nd person singular subjects are most likely to show up as accusative. It is also sensitive to morphological form: since masculine third person singular pronouns are syncretic for nominative and accusative, they are more likely to show up as dative (overtly ‘marking’ their oblique status). The role of these factors makes sense, given historical pressures and the ‘morphologically shallow’ nature of DS alternations.
Another kind of data discussed by Nowenstein (Reference Nowenstein2012, Reference Nowenstein2014a, Reference Nowensteinb, Reference Nowenstein2017) and Jónsson (Reference Jónsson2013b) also make sense in terms of an impoverishment analysis. Consider the data in (75). Floating modifiers like sjálf(ur) ‘self’ typically agree with their subject in case, and this is what we see in (75).
However, we find certain mismatches – but generally only in one direction. An accusative subject can show up with a dative floating modifier, but a dative subject cannot show up with an accusative floating modifier.
An impoverishment analysis can explain why. If case marking is post-syntactic, agreement should be too (cf. Bobaljik (Reference Bobaljik, Harbour, Adger and Béjar2008)). But then we can ask: should floating modifier agreement take place before, or after impoverishment? In fact there is no intrinsic need to order them – they could apply in either order. The idea that agreement could apply in either order has some general independent support. Arregi & Nevins (Reference Arregi and Nevins2012: 344) propose that PF agreement (which they call ‘Agree-Copy’) normally takes place before impoverishment rules, but may be deferred until after such rules (where it is sensitive to linear order, accounting for some kinds of closest-conjunct agreement). Nevins (Reference Nevins2014) proposes that varieties of Indo-Aryan discussed by Deo & Sharma (Reference Deo and Sharma2006) vary in the timing of Agree-Copy and ergative-to-nominative impoverishment. In Standard Puṇe Marathi, agreement precedes case-impoverishment, whereas in Gowari, agreement follows case-impoverishment.
The Icelandic variation is accounted for if we assume exactly this – floating modifier agreement may apply in either order.
Notice that there is no natural way to derive dative on the DP, but accusative on the modifier.
This kind of phenomenon is not restricted to sjálf(ur) ‘self’ – it applies to various kinds of floating modifiers. Two more examples are floating quantifiers, as in (78), and sem ‘as’-phrases, as in (79).
The existence of, and nature of the mixing supports the view that DS is a morphological phenomenon. The impoverishment analysis accounts for the lexically idiosyncratic nature of accusative subjects in the first place, the existence/directionality of Dative Substitution, and the kind of mixing we find in floating modifier agreement. Returning to the main issue, a syntactic identity condition provides an understanding of why case mismatches are allowed with Dative Substitution: Dative Substitution does not reflect a syntactic distinction, so it does not interfere with a syntactic identity condition.
5.2 Nominative/Accusative Alternations
We now turn to alternations between nominative and accusative, and show how the speaker variation we find there is compatible with a syntactic identity condition. This discussion is based on the theory in Wood (Reference Wood2015, Reference Wood2017). To begin, the analysis of accusative subjects in Wood (Reference Wood2017) takes them to involve a silent external argument clitic, which bears nominative case.Footnote [37] The accusative then moves around the nominative clitic into the subject position.
Wood (Reference Wood2017) argues in favor of a ‘Dependent Case’ analysis of accusative (Marantz Reference Marantz and Reuland1991/2000, McFadden Reference McFadden2004, Wood Reference Wood2011). The idea is that if a DP does not get some lexically specified case, such as dative in SpecApplP, then it is generally either accusative or nominative. Given two DPs within a given case domain, the lower DP is accusative and the higher one, if it is not c-commanded by another, non-lexically case-marked DP, will be nominative. (In ECM constructions, however, the higher DP could also be accusative.)Footnote [38]
The question, on this analysis, is what happens when nominative substitution occurs? One way to get nominative would be for the external argument clitic to lose its nom case property. If the clitic is not nominative, then the internal argument will not be accusative.
It would, in this way, become more like the anticausative -st clitic as analyzed in Wood (Reference Wood2014, Reference Wood2015). In (82b), the -st clitic marks the anticausative use of ‘open’. Building on the analysis of marked anticausatives cross-linguistically in Schäfer (Reference Schäfer2008) and Alexiadou, Anagnostopoulou & Schäfer (Reference Alexiadou, Anagnostopoulou and Schäfer2015) (among others), Wood (Reference Wood2014, Reference Wood2015) proposes -st is merged in the external argument position, and the internal argument moves to the subject position around it (see Eythórsson (Reference Eythórsson1995), H.Á. Sigurðsson (Reference Sigurðsson2012b), and E.F. Sigurðsson (Reference Sigurðsson2017) for related analyses of the -st clitic).
Importantly, the -st clitic neither bears case nor conditions accusative case (Wood Reference Wood2015: 68–69).
If (81) is the analysis of Nominative Substitution sentences, then the syntactic structure is the same whether the subject is nominative or accusative. Specifically, the argument structure is the same, per (69), since both take an external argument. Or, drawing on (70), both VoicePs are built from the same lexical items. Assuming that case features are assigned post-syntactically, case features themselves cannot cause a mismatch for the syntactic identity condition. The difference is morphological: in one case, CL gets a morphological case feature, while in the other, it does not. Therefore, we expect case mismatches in fragment responses with Nominative Substitution just as with Dative Substitution.
However, another option to get nominative is that the external argument clitic is genuinely not present. Instead, VoiceP can either be specifierless or absent altogether. For present purposes, we will assume (following Wood (Reference Wood2015: 152–155)) that a specifierless Voice would be present.
In this case, the syntactic structure would be different, depending on whether nominative or accusative surfaces. The VoicePs are built from distinct Voice heads, resulting in the presence or absence of an external argument. Thus, we would expect mismatches in fragment answers to be impossible.Footnote [39]
The fact that speakers vary as to whether mismatches with nominative are possible suggests that speakers genuinely vary as to whether they internalize the first or the second option for accusative/nominative variation. Moreover, it is possible that speakers internalize different options for different verbs. Speaker variation of this kind is found overtly in other cases, where some speakers may use caseless -st while others use a case-marked reflexive pronoun (see Wood (Reference Wood2015: 192)). It also allows for the possibility that even with a given verb, speakers vary in terms of whether they use/allow (80), (81) or (84). The fact that we find such variation, across and within speakers, thus supports the present analysis. (Or, more generally, the claim that nominative/accusative variation is sometimes, but not always, grounded in a syntactic distinction; the point here is to give an explicit idea of what this means.)
Finally, it is important to note that the choice between (80), (81) and (84) does not need to correlate with any semantic distinction. If the present suggestion is on the right track, they must not, because Nominative Substitution has always been claimed to have no semantic effects. The question here amounts in part to whether the clitic in (80) and (81) is more like an expletive or more like a referential pronoun. Wood (Reference Wood2017) (drawing on Schäfer (Reference Schäfer2008, Reference Schäfer2012)) suggests that the clitic affects the semantics in two ways: it has its own semantics, referring to ‘the referentially underspecified agent responsible for forces that are not in human power’ (264), and it conditions idiomatic interpretations on the vPs it combines with. The latter is not a problem here, because idiomatic interpretations can be learned in the absence of a clitic. As for the former, it is only a problem here if we assume that this ‘referentially underspecified agent’ cannot be part of the implied meaning without a clitic or pronoun pointing to it. If it cannot, then either (84) is not a genuine option for Nominative Substitution, or, contrary to previous reports, there in fact is some subtle semantic distinction between nominative and accusative subjects with Nominative Substitution. We find neither of these to be particularly likely, so we will assume that either the clitic is more like an expletive, or else the ‘referentially underspecified agent’ can become encoded in the meaning of the VoiceP even in the absence of the clitic.Footnote [40]
We now return to the Hungarian facts discussed above that were taken to argue against silent syntactic structure, and in favor of stipulating case matching directly.
Recall that according to Jacobson (Reference Jacobson2016: 356), case mismatching is not possible, and there is no semantic difference in the choice. We have now seen that accusative/nominative alternations, for a subset of Icelandic speakers, matches this description. But we have also seen that the very fact that case mismatching is disallowed does not mean that case matching should be directly encoded. The Hungarian facts would follow if there is a syntactic difference between the allative and sublative case-marking options, even if that difference does not make a difference semantically. If, for example, the case markers correspond to distinct postpositions, then the syntax is not identical, and the Hungarian facts follow in the way that the following English data follow:
Another possibility is that the distinct case markers correspond to distinct argument structures (i.e., distinct VoiceP-internal heads), even if there is no detectable semantic difference. Either way, it is clear that the Hungarian data do not argue against silent syntactic structure; there are various factors that can enforce case matching. However, the fact that case mismatching is possible in some cases does strongly suggest that case matching cannot be directly encoded as part of how clausal ellipsis works.
6 Conclusion
In this paper, we have taken a detailed look at Icelandic fragment responses, focusing on what case matching and mismatching tell us about the analysis of clausal ellipsis in general. In particular, we have argued that the existence of case mismatches within a system that generally enforces case matching argues against interpretive approaches to ellipsis that deny the existence of silent syntactic structure and encode case matching directly.
We examined when mismatches are possible and when they are impossible, and showed that the facts follow from silent syntactic structure plus a hybrid identity condition along the lines of Chung (Reference Chung2013), in particular her Argument Structure Condition. It is not enough to say that case mismatching is possible when the choice of case makes no difference in meaning, because we find some instances where case mismatching is not possible, despite the fact that choice of case makes no difference in meaning. We proposed that in such cases, the choice of case nevertheless corresponds to a syntactic distinction (in argument structure), but one that does not affect the semantics.
Finally, we developed our analysis within a framework that crucially assumes that at least some (but not necessarily all) case marking is determined post-syntactically, and the overall picture seems to support this perspective (at least in the places where it is crucial to the analysis). If so, then the analysis here not only supports silent structure in the sense generally intended in research on ellipsis, but also ‘silent structure’ in the grammatical architecture, in the sense that there is an articulated mapping from syntax to PF, manipulating abstract features in a step-by-step fashion to derive surface forms from underlying structured representations.
