The syntax and semantics of split NPs in LFG ... - Semantic Scholar

1 downloads 0 Views 139KB Size Report
indefinite NP in the sentence-initial position. Thus, the analysis can ..... will simply use the indefinite pronoun welche 'some' as in (15a) and assume that it is ...
The syntax and semantics of split NPs in LFG*

Jonas Kuhn

In the split NP construction (SNP) in German, the descriptive material contributing information about one participant in a predicate’s argument structure is realized discontinuously as two nominal phrases in different positions—in the sentence-inital topic position and within the Mittelfeld. This is a challenge to syntactic accounts of argument selection and unbounded dependency constructions. I propose an account in LexicalFunctional Grammar making use of unification at the level of functional structure, which allows argument doubling. This mechanism is constrained by the semantic types of the elements involved. The Mittelfeld NP in SNP construction acts like an isolated elliptical NP in that it requires an anaphorical link to the topic of the sentence, which has to be of the semantic type of a property. In SNP, the topic is introduced sentence-internally as an indefinite NP in the sentence-initial position. Thus, the analysis can explain the data without stipulating any construction-specific mechanisms.

1 Introduction The split NP (or split topicalization) construction—henceforth SNP—of German has received much attention in the literature, as a challenge for syntactic accounts of argument selection and of unbounded dependency constructions (Bayer, 1987; van Riemsdijk, 1989; Fanselow, 1988; 1993; Haider, 1990; Hinrichs/Nakazawa 1994) and in the discussion of the semantic status of indefinites (Diesing, 1992; van Geenhoven, 1996). In this construction, the descriptive material about one underlying argument is not confined to a single NP as normal, but is split over two nominal constituents: (1)

Bücher sieht Anna drei Books sees Anna three ‘As for books, Anna can see three.’

The SNP construction constitutes a problem in view of the widely assumed grammatical principle that the syntactic complement phrases of a *

For discussion and comments on various versions of this paper, I’d like to thank Judith Berman, Norbert Bröker, Miriam Butt, Mary Dalrymple, Stefanie Dipper, Martin Emele, Christian Fortmann, Anette Frank, John Fry, Christian Rohrer and two anonymous reviewers. Parts of sec. 1, 2 and 3.1 of this paper appear also in (Kuhn, 1998).

Jonas Kuhn

head stand in a one-to-one correspondence to argument slots in an underlying conceptual representation—formulated e.g. as the Theta Criterion of Government-and-Binding Theory and as Biuniqueness within Lexical-Functional Grammar (LFG). In SNP apparently two NPs can realize a single underlying argument. I’d like to argue that previous accounts of the SNP construction make unnecessary assumptions of some extra devices and often fail to explain the striking parallelism of the Mittelfeld NP and elliptical NPs. The architecture of LFG with its system of correspondence between parallel structures—constituent (c-)structure, functional (f-)structure, argument (a-)structure and semantic structure—turns out to provide the means for an explanatory analysis of the SNP construction that doesn’t require the stipulation of any construction-specific mechanisms. The starting point for my analysis is the observation that on a “low” level, both parts of SNP behave like autonomous NPs. The topic part (ein Schwimmbad ‘a pool’ in (2a)) has to satisfy standard requirements for determiner selection: singular mass nouns can stand without a determiner, while count nouns can’t; the Mittelfeld NP surfaces exactly like an isolated elliptical NP (cf. keins ‘none’ in (2a) and (2c)); both parts must independently meet constraints on declension class (cf. the contrasting acceptibility of keins/kein in the SNP example (2a) vs. example (2b) containing only a single object NP). (2) a.

b. c.

(van Riemsdijk, 1989: p. 109, ex. (11)) Ein Schwimmbad hat er sich noch keins/*kein gebaut. A pool has he REFL yet none/no built ‘As for a swimming pool, he hasn’t yet built one.’ Er hat sich noch *keins/kein Schwimmbad gebaut. ‘He hasn’t yet built a swimming pool.’ Er hat sich noch keins gebaut. ‘He hasn’t yet built one.’

This behaviour is precisely predicted in LFG under two assumptions: (A) In a case-marking language like German, the functional annotations in cstructure rules are underspecified as to the exact grammatical function (e.g., SUBJ or OBJ) an NP bears; not the configurational position, but case and agreement principles1 are decisive for function specification (cf. e.g., 1

For example, (↓ CASE) = NOM ⇒ (↑ SUBJ) = ↓.

The syntax and semantics of split NPs in LFG

Bresnan, 1995: ch. 5; 1996: p. 17). Thus, on the basis of c-structure rules and case principles, an NP bearing a particular function (say, OBJ ) can occur in different configurational positions—e.g., in the Mittelfeld or in the preverbal position. Since the annotations are confined to local trees, nothing prevents the simultaneous occurrence of an OBJ NP in both these positions. (B) In the case of SNP, the f-structure specification for the two participating NPs (and their semantic interpretation) is compatible, i.e., their contribution can be unified in a single f-structure, 2 as in the unification analysis of clitic doubling by Andrews (1990). The general idea of this SNP analysis is illustrated by the simplified c-structure and f-structure analysis of (3). 3 The functional annotations given in the tree, below the category symbols, should be seen as the result of applying the mentioned case principles to underspecified configurational annotations. (3)

Einen Frosch sah er einen kleinen. a.ACC frog.ACC saw he.NOM a.ACC small.ACC ‘As for frogs, he could see a small one.’ .

CP 0

NP

C

("topic)=# "comp* obj)=#

(

 HH

einen Frosch

"=#

C

VP

"=#

"=#

sah

NP

NP

"subj)=#

"obj)=#

 HH

(

2 6 6 6 6 6 6 4





er

subj pred topic

pred `er' `sehenhsubj obji'

obj

4

2

pred `Frosch' spec ein  pred adjuncts

(

einen kleinen

3

`klein'



7 7 7 3 7 7 7 5 5

The NP nodes for einen Frosch and einen kleinen are both mapped to the same f-structure, as the OBJ of the verb sehen. C-structurally, both NPs 2 3

Since the merging of two constituents as a single underlying argument is captured by unification at the level of f-structure already, we can ignore the LFG level of argument (a-)structure in this study. The arrows (↑ and ↓) in the functional annotations (which interact to constrain the construction of the f-structure) are metavariables; ↑ refers to the f-structure projected from the mother node, ↓ refers to the f-structure projected from the node itself. For more background information on the mechanisms of standard LFG, the reader is referred to the articles in (Dalrymple et al., 1995b).

Jonas Kuhn

are treated as autonomous canonical NPs, under application of the general rules. Consequently, constraints on determiner selection and morphological marking (declension class) apply to both NP parts. As the further discussion will reveal, a full and sufficiently constrained LFG account of the construction requires a more fine-grained view on the semantic contribution of the individual NPs, in particular their semantic type. On this basis, the observed parallelism of the Mittelfeld NP and elliptical NPs is explained as their anaphoricity: both require an antecedent of the semantic type of a property, which in the case of SNP is provided sentence-internally by the topic NP.4 Before addressing the analysis in more detail in the main section 3, I will in sec. 2 briefly list and comment on further empirical properties of SNP discussed in the literature. A discussion of previous accounts will be presented in sec. 4, and is followed by a conclusion in sec. 5. 2 Properties of the split NP construction In addition to the mentioned morphological autonomy of the nominal projections involved in SNP, the following properties have been observed: 5 Agreement. The two NPs agree in number (and case and gender): Other than in the ‘as for...’ paraphrase in English, the SNP construction is ungrammatical when a mismatch in number occurs (4). (4)

Ein Buch/*Bücher ist erst eins von ihm erschienen A book/Books has only one by him appeared ‘As for books, only one by him has appeared yet.’

This property follows trivially under a unification analysis. Unbounded dependency. The standard diagnostics for an unbounded dependency construction apply (for instance, extraction out of a wh-clause is impossible: (5b)). (5)

a. b.

4 5

Unbeschädigte Exemplare glaube ich, daß ich nur noch 2 auf Vorrat habe. Undamaged copies believe I, that I only still 2 in stock have *Unbeschädigte Exemplare wollte er wissen, wer noch 2 auf Vorrat hat. Undamaged copies wanted he to-know, who still 2 in stock has

Compare (van Geenhoven, 1996), who argues that such a property-denoting indefinite is licensed under particular circumstances of semantic incorporation. For an extensive discussion, see (van Riemsdijk, 1989); this section is entirely based on his data.

The syntax and semantics of split NPs in LFG

In the proposed LFG analysis, the topic part of the NP is (like any NP occuring in this position) related to the f-structure canonically projected from arguments in the Mittelfeld by the LFG standard mechanism for unbounded dependency construction, a functional uncertainty equation (cf. the annotation (↑COMP* OBJ)=↓ in (3) saying that the topic NP einen Frosch projects to an f-structure embedded under OBJ either directly under the mother’s f-structure, or with an intervening chain of an arbitrary number of f-structures, recursively embedded under the function COMP, i.e., embedded in clauses like the one in (5a)). VP topicalization. SNP can be part of (partial) VP topicalization: In (6), the (usually sentence-final) participle gekauft ‘bought’ appears in topicalized position, together with an indefinite argument NP, and another NP realizing the same argument slot is “left” in the Mittelfeld. (6)

[Einen Wagen gekauft] hat er sich noch keinen. A car bought has he REFL yet none ‘As for cars, he has not yet bought one.’

While this property is fatal for a movement analysis as the one by van Riemsdijk (1989) (cf. the discussion in sec. 4.1 below), it fits in naturally with the unification analysis. The remaining three empirical properties 6 are not immediately predicted by the basic unification idea laid out in the introduction. They will be taken care of in the discussion of constraints on the analysis in sec. 3. Indefiniteness. The topic NP must be non-quantificational (cf. the ungrammatical manche ‘some’ in (7)). Some dialects allow only a bare plural (like in (1)), others alternatively allow a singular indefinite (cf. (2a)). 6

A further property van Riemsdijk (1989) discusses is the fact that restrictions on the serialization of multiple adjectives bear over to the split NP situation suggesting that the topic NP has to be reinserted into the “remnant” NP for checking this condition (i-ii). The condition may however just as well (and more sensibly) apply on a more semantic level of representation, not requiring that there be some configurational level with a single “unsplit” NP. (i) a. b. (ii) a. b.

ein neues amerikanisches Auto ‘a new American car’ *ein amerikanisches neues Auto Ein amerikanisches Auto kann ich mir kein neues leisten. ‘As for an American car, I cannot afford a new one.’ *Ein neues Auto kann ich mir kein amerikanisches leisten.

Jonas Kuhn (7)

(*Manche) Häuser habe ich auch noch kleinere gesehen. (Some) houses have I also even smaller seen.

Topic position. Other than in the closely related floating quantifier construction (8a),7 the higher NP in SNP is structurally confined to the (preverbal) topic position: (8b) is infelicitous.8 (8)

a. b.

Er hat die Zigaretten gestern alle vier geraucht. He has the cigarettes yesterday all four smoked. *Er hat einen Wagen sich noch nie einen leisten können. He has a car REFL yet never one afford could

Headless “remnant”. The NP in the Mittelfeld may not contain an overt nominal head (9).9 (9)

*Bücher hat er noch keine Romane geschrieben. Books has he yet no novels written.

This property does follow under the “naive” analysis of (3)—the PRED values of Bücher and Romane would clash. However, under the more fine-grained view on PRED values taken below in sec. 3.1, this property will need reconsideration.

7

8

My assumption is that the floating quantifier construction is also an instance of f-structure unification, with the difference lying in the restrictions on semantic types; however, further research is required to work out the exact relationship between the two phenomena. Van Geenhoven (1996: p. 24, fn. 5) points out that example (i) (attributed to Fanselow, p.c.) shows that verb-final (embedded) clauses may also contain SNP (note however that in terms of intonation and interpretation, Bücher has the characteristics of a topic, suggesting that it is located in a Mittelfeld-internal topic position). (i) ....weil man Bücher damals in den Osten keine mitbringen durfte ....because one books then to the East none bring was-allowed-to ‘because in the past it was not allowed to take books to East Germany.’

9

Fanselow (1993: p. 63) cites (i) (attributed to Santorini, p.c.) as a counterexample to this. To me, examples of this type are marginal, cf. also fn. 22 below. (i) Raubvögel glaube ich kennt Gereon nur Bussarde. birds-of-prey believe I knows Gereon only buzzards ‘As for birds of prey, I believe that Gereon knows only buzzards.’

The syntax and semantics of split NPs in LFG

3 Constraints on the unification analysis The general idea of the unification analysis outlined in the introduction is fairly straightforward. However, it remains to be shown that there are independent principles constraining the mechanism in a way that explains the data.10 Given the mechanism of f-structure unification—why cannot simply any canoncial NP dissociate and spread its contribution to the sentence over various c-structure NPs? The constraints on morphological marking following from autonomy of the NPs at the level of c-structure do not explain the specificity and topichood restrictions and headlessness; moreover the question arises whether assumption (B) of sec. 1—that the f-structures of the two NPs in SNP can unify—is tenable without an undermining of the key concepts regulating subcategorization in LFG. The latter question will be addressed first, thereby providing the notational prerequisites to come to the empirical properties. 3.1 On the role of PRED values Classically, a central role of f-structure is to constrain the great freedom at c-structure, excluding that two canonical NPs are assigned the same grammatical function. This is achieved by functional uniqueness (disallowing two distinct appearences of the same grammatical function within a single f-structure) combined with the fact that the PRED values of the NPs will fail to unify. Even if the PRED values are of the same form, standard LFG interprets them as instantiated symbols, i.e., as semantically distinct from any other symbol (Kaplan/Bresnan, 1995: p. 77ff). “Doubling” of a function—as in the ungrammatical example (10)—is thus ruled out.

10 It would be certainly possible to explicitly stipulate restrictions on the rules that ensure the empirical generalization—similarly as in the lexical rule of (Hinrichs/Nakazawa, 1994)—but as will be shown, these restrictions follow from independent properties of the elements involved.

Jonas Kuhn (10)

*den Peter sieht der Karl den Peter. the.ACC Peter sees the.NOM Karl the.ACC Peter . . .

. . .

NP

"

( ...

DET

obj)=# N

"pred)=`Peter'27

...

"obj)=#

(

DET

(

den

Peter

2

NP

4

pred subj obj

hsubj obji'

`sehen

 

pred `Karl'  pred ?

3 5

N

"pred)=`Peter'58

(

den

Peter

It has been observed however (Andrews, 1990) that to a limited degree, some languages allow for exceptions to this rule. While in French, the presence of an object clitic on the verb makes the appearence of a full NP with the same function impossible, Spanish does allow clitic doubling (Andrews, 1990: sec. 3.1). 11 Andrews models this option making the introduction of the clitic’s PRED value optional. The initial SNP example for which I sketched the “naive” analysis in (3) seems to suggest that for this German construction no provisions about PRED are necessary at all: since the Mittelfeld NP doesn’t contain an overt head noun, it is just the attributive adjective that contributes a PRED value – as a member of the ADJUNCT set. Thus, no clash occurs when the topic NP introduces the PRED value to the common f-structure under the verb’s OBJ. Looking at a larger set of data and taking into account the parallelism of the Mittelfeld NP and elliptical NPs however, the type of semantic contribution of the various components needs a revision: Firstly, an elliptical NP (being identical to the Mittelfeld part of SNP) is perfectly capable of filling the slot of an argument function on its own (cf. (2c))— suggesting that it introduces its own NP level PRED value, not just the contribution under ADJUNCT . Secondly, the “naive” analysis wrongly predicts that in SNP there can be an iteration of Mittelfeld NPs, which would all unify (11). (11)

*Einen Frosch sah er einen kleinen heute einen grünen. [a frog].ACC saw he.NOM [a small].ACC today [a green].ACC

1 1 A simple example is (i):

(i) Lo vimos a él. him we-saw at him ‘We saw him.’

The syntax and semantics of split NPs in LFG

As a matter of fact, the Mittelfeld part of NP is existentially quantified (or introduces a variable/discourse referent in terms of DRT 12), i.e., it should have the character of an instantiated symbol, not tolerating a contribution from elsewhere to the same slot—so the topic NP cannot also introduce an instantiated P R E D value. This result converges with observations by Fanselow (1988: p. 105) and van Geenhoven (1996): the topic NP has a semantic contribution of type property, rather than being an entity or quantifier. As further evidence for this assumption, it is possible to have more than one topic NP in a particular contrastive context, using the appropriate intonational marking for topic on both these NPs. The first one introduces a topic that gets narrowed down by the second one: 13 (12)

Bücher durfte man politische damals in den Osten keine mitbringen. books was-allowed one political then to the East none bring ‘Talking about books—as for political ones, one wasn’t allowed to bring any to East Germany.’

How can we capture this behaviour in the LFG formalism? As Dalrymple et al. (1995a: p. 14) note, the PRED value notation in classical LFG bundles together a number of types of information (including (i) the introduction of an instantiated symbol to mark predicate uniqueness and (ii) the specification of the semantic relation); they observe that apart from instantiation all this information is meanwhile encoded in different components of the theory: the level of a-structure with linking theory or linear-logic based LFG semantics with its resource accounting (cf. Dalrymple et al., 1997). As it turns out, even instantiation can be taken over by resource sensitivity in linear-logic based semantics (cf. Kuhn, 1998: sec. 4.1). Nevertheless, the classical notation is often handy as an abbreviatory device. Now in the SNP construction—contrary to the canonical situation— the introduction of an instantiated symbol (a variable/discourse referent) on the one hand and the specification of (at least some of) the semantic relations and properties on the other hand have their origin in distinct cstructure constituents: the Mittelfeld NP and the topic NP respectively. The classical PRED notation is not fine-grained enough to encode this 1 2 Discourse Representation Theory, cf. (Kamp/Reyle, 1993). 1 3 The example is based on (i) in fn. 8

Jonas Kuhn

distinction. In (Kuhn, 1998: sec. 4.4), I therefore develop the SNP analysis with a linear logic-based semantics. To save space for a more general discussion in the present paper, I keep up the general format of classical LFG PRED values, simulating the effects just discussed by using more internal structure. 14 Only elements that do introduce instantiated symbols will come with an ordinary PRED value (and if they introduce just an instantiated symbol, without specifying a semantic relation, the value of PRED will be ‘pro’). Elements that contribute just a semantic relation or property (like the topic part of SNP) are essentially modificational and will introduce their semantic value as the element of an embedded ADJUNCT set.15 Note that this means that these elements can unify with an instantiated symbol from elsewhere. Applied to the SNP example (13a), we now get the f-structure analysis in (13b) (including a sketch of the respective contribution of the two NPs). (13)

a.

Frösche sieht Karl kleine. Frogs sees Karl small ‘As for frogs, Karl can see small ones.’

b.

2

f

6 6 6 6 6 6 4





subj pred topic

pred `Karl' `sehenhsubj obji'

obj

4

2

Fr osche:

(f

kleine:

(f (f

pred adjuncts

3

`pro'

 

pred pred

  ;

`Frosch' `klein'

7 7 7 3 7 7 7 5 5

obj adjuncts 3 pred) = `Frosch' obj pred) = `pro' obj adjuncts 3 pred) = `klein'

3.2 Anaphoricity of the Mittelfeld NP Having introduced the basic notational apparatus, we are now in a position to address the question what explains the open empirical properties from sec. 2—specificity restrictions and the topichood requirement for the higher NP, and headlessness of the lower NP. From the discussion it will become clear that despite the simulation of semantic types, the classical 1 4 It is impossible however to simulate the full resource sensitivity that way. I will come back to this

limitation in sec. 3.2. 15 This way of capturing the property status within classical f-structures was suggested to me by Mary Dalrymple (p.c.).

The syntax and semantics of split NPs in LFG PRED -based formulation has its limitations. The resource sensitivity

involved in accounting for the various semantic contributions would require a non-unificational mechanism at the level of semantics. The crucial point of the explanation that I propose is the initially observed parallelism between the Mittelfeld NP of the SNP construction and elliptical NPs. I assume that underlying both NP usages is a common anaphoric behaviour requiring an antecedent of the semantic type of a property. For elliptical NPs, it should be uncontroversial to assume anaphoric behaviour, in the sense that they take up a property that is salient in the context.16 In English, this anaphoricity can even be lexically tied to the item one (as in a new one).17 In the corresponding case in German (14), no overt element is required. (14)

Ottos Auto ist kaputt gegangen. Er will sich ein neues kaufen. Otto’s car has broken gone. He wants REFL a new buy ‘Otto’s has broken down. He wants to buy a new one.’

I won’t go into the question whether syntactically, the effect should be captured by an empty nominal head like the pro that Fanselow (1988) assumes, or by category change, or at the level of the construction, since this is orthogonal to the main point under discussion. For illustration I will simply use the indefinite pronoun welche ‘some’ as in (15a) and assume that it is lexically anaphoric to a property (and, being existentially quantified, it furthermore introduces its own PRED value: ‘pro’). The choice of the antecedent property is controlled by discourse function. It has to be the current topic 18 . The topic for the second sentence is the property of being a rabbit, recovered from the context. 16 As to discourse semantics, the resolution of property or kind anaphora is not constrained by the same discourse accessibility restrictions as the paradigm case of anaphora, which require individuals as antecedents. The discourse (i) is coherent, while (ii) is ill-formed. (i) Jones didn’t own a Porsche. Ann bought a red one for him. (ii) Jones didn’t own a Porsche. #Ann sold it for a lot of money. In DRT, this could be captured by assuming that property referents are not introduced in the local universe, but at some higher level, like proper names (cf. Rooth, 1992: p. 87, fn. 8). 1 7 However, where no adjective appears in the elliptical NP, one is not required: I saw some/two/many/none. 18 One should be careful with the term topic, since it is being used in very different senses, and often there is no clear formal interpretation. Unfortunately, I presently can’t do more than make use the intuitive concept of the topic being what the sentence is uttered about, i.e., typically an entity (but possibly also a property) that is at the centre of discussion. The topic need not be expressed in the sentence itself.

Jonas Kuhn

Let us assume this is encoded under the discourse function TOPIC , as sketched in (15b). (15)

a.

b.

Wir wollten Kaninchen beobachten. Aber nur Anna hat welche gesehen. We wanted rabbits watch. But only Anna has some seen. ‘We wanted to watch rabbits. But only Anna could see any.’ 2 6 6 4

subj pred topic obj





pred `Anna' `sehen hsubj obji'   pred g  adjunct  h pred `pro'

3

`Kaninchen'

7   7 5

g

= h [ ]

The anaphoricity of the elliptical object NP welche establishes a link to the property in the topic: the two f-structures project to the same semantic structure (via the semantic projection σ). 19 This is expressed by the anaphoric condition20 in the annotation of the lexicon entry for welche: (16)

welche

NP

(↑PRED) = ‘pro’ ((GF* GF↑) TOPIC)σ = ↑σ ¬(→TOPIC)

Bear in mind that the unificational account of semantics just sketched is to be taken with a grain of salt. Intuitively what the anaphoric welche does is consume a copy of the property of being a rabbit and use it for its own purpose. Now, what happens in the SNP construction (17a) is that the topic that was not expressed sentence-internally in the second sentence of (15a) is overtly introduced by the topic NP Kaninchen.

1 9 The actual meanings are left out here, since their content doesn’t affect the composition.

20 Following Dalrymple (1993), anaphoricity is modelled in LFG using so-called inside-out functional uncertainty equations. The respective equation in (16) asserts identity between the semantic structures projected from (i) welche’s own f-structure, say h, and (ii) the value of the feature TOPIC in an embedding f-structure. The inside-out functional uncertainty constrains the path under which h may be embedded in the larger f-structure, here a path of 1 to n grammatical functions ( GF) of which only the outermost may co-occur in an f-structure with a feature TOPIC. In other words, the closest embedding TOPIC feature is picked out as the antecedent.

The syntax and semantics of split NPs in LFG (17)

a. b.

Kaninchen hat Anna welche gesehen. rabbits has Anna some seen 2 6 6 4

subj pred topic obj





pred `Anna' `sehen hsubj obji'   pred g  adjunct  h pred `pro'

3

`Kaninchen'

7   7 5

g

= h [ ]

As a consequence, the anaphoric relation of (15b) is in (17b) overlaid by the f-structure unification due to object doubling as discussed in sec. 1. The effect at semantic structure is identical. However, the token identity of the semantic structures projected from the NP Kaninchen and from the NP welche is now established twice: firstly, through the anaphoric annotation lexically introduced by welche, and secondly, due to the syntactic doubling configuration that assigns both NPs the same grammatical function, thus establishing identity already at f-structure (g=h). Is this apparent redundancy desirable? And does this analysis bring us closer to an explanation of the observed empirical restrictions on the SNP construction? The two links have a different status. While f-structure unification due to doubling merely directs the contribution of the two NPs to the same argument slot without looking at their semantic type, the anaphoric link is really under control of semantic composition. Even though the notation chosen here for illustrative simplicity doesn’t enforce it (but cf. Kuhn, 1998: sec. 4.4), the semantics of the complete sentence could not be successfully construed if it wasn’t for the anaphoric link. Recall that intuitively, welche requires a property as its antecedent that it can consume to build up its own semantic contribution. And recall furthermore that this antecdent is confined to the discourse function of the topic. So, it is essential that if the partner of the “elliptical NP” is present in the same sentence as in SNP, it must be located in a position that allows it (i) to introduce (just) a property and (ii) to introduce this property as the TOPIC. One may still be concerned about redundancy and ask why the doubling at f-structure should occur if it is not directly required to put together the semantic contribution. Wouldn’t it be enough to have just the anaphoric link?

Jonas Kuhn

As a matter of fact, the doubling is not there in constructions where the property-type antecedent for the elliptical NP is introduced by a phrase that is syntactically an adjunct, like the English ‘as for ...’ construction (a German equivalent would be was ... betrifft): (18)

Was Kinder betrifft, hat Anna welchen geholfen. what children.ACC concerns has Anna some.DAT helped

Note that here, we do not observe case agreement. The German SNP construction is special in that for introducing the property-type antecedent (whose semantic nature is modificational), it takes advantage of the existing mechanism of function specification on NPs, originally devised to introduce full canonical arguments. Since in certain specific positions, including the topic position, property-denoting indefinites are licensed (cf. van Geenhoven’s (1996) concept of semantic incorporation), lacking their own variable/referent (i.e., character of an instantiated symbol), there is no need to make use of syntactic modification as in (18). So, finally we have an explanation for the open observations from sec. 2. Since SNP essentially involves an elliptical NP, the topicrestriction on the antecedent for anaphoric elliptical NPs bears over to SNP. The sematic type restriction on the antecedent furthermore explains the observed confinement to indefinite NPs—properties can only be introduced by these. Finally, the headlessness of the Mittelfeld NP is also a direct consequence of the central role of the elliptical NP: if the Mittelfeld NP is non-elliptical, no one will consume the extra NP’s contribution, so the construction is ruled out by resource accounting. 4 Discussion of previous accounts In this section, I will briefly address the literature on SNP, pointing out problems and relating the analyses to my own proposal. 4.1 Movement In the late 1980s, there was a debate within the GB framework about the status of SNP. Van Riemsdijk (1989) argues for a movement account: the topic NP is a preposed part of the “remnant” NP in the Mittelfeld. This

The syntax and semantics of split NPs in LFG

analysis saves the Theta Criterion and gives a straightforward explanation for empirical properties like unbounded dependency behaviour, headlessness of the remnant NP (the gap is the trace of the moved N') and serialization effects of adjectives. The unexpected morphological autonomy of the two NPs (the moved N' ends up as a maximal phrase, potentially with its own determiner—like in (2)) is explained by positing that under certain circumstances, non-maximal projections that have been moved can “regenerate” some amount of structure to meet X-bar requirements applying at S-structure. Only unmarked lexical material may be introduced in regeneration; this explains indirectly the indefiniteness property: the indefinite article is the unmarked form of the determiner. Apart from the potentially controversial status of the regeneration mechanism (which lacks independent motivation), the account has the empirical problem that it predicts ungrammaticality for the VP topicalization examples containing SNP (6). Van Riemsdijk (1989: 134, note 29) rejects the data, but I found wide agreement that they are perfectly grammatical (cf. also Haider, 1990: p. 108; Fanselow, 1993: p. 59; Hinrichs/Nakazawa, 1994: sec. 11; van Geenhoven, 1996: p. 147). There is no obvious way to adapt the analysis such that it allows movement of a VP that contains an N', without containing the maximal projection of this N'. 4.2 Base generation As alternatives to a movement analysis, various base-generation accounts have been proposed (a.o. Fanselow, 1988; Hinrichs/Nakazawa, 1994; van Geenhoven, 1996). Base-generation of two NPs gives a straightforward explanation for the morphological autonomy of the parts of SNP and is compatible with the VP topicalization data. However, the relation between the two constituents and their argument status and contribution to interpretation require some more explanation. Due to space limitations, I can only address some key points. Hinrichs/Nakazawa (1994) propose an HPSG account stipulating a special lexical rule for SNP. The rule applies to the verb and changes an NP on the subcategorization list into a headless variant, simultaneously introducing an indefinite NP with the same content specification to the slash set, thus ensuring that the two parts restrict the same underlying

Jonas Kuhn

argument. The nonlocal feature mechanism causes the slash element to be realized higher up to the left in the topic position. While this account allows to stipulate the empirically observable restrictions explicitly, no explanatory reduction to more general properties of the items is made; in particular the specificity restrictions, and in a sense even case agreement have the appearance of being coincidental properties of the construction rather than the consequence of more general principles, like I tried to argue for in the present paper. Fanselow (1988) assumes that two NPs are generated in the Mittelfeld, one of which contains the empty pronoun pro. The other NP is non-referential, denoting a property, and is coindexed with the pro. For the pro NP to satisfy binding principle C (which excludes that referential expression are A-bound), the non-referential NP moves to the sentenceinitial non-argument position. Fanselow argues that the Theta Criterion is not violated since it applies only to term-denoting argument NPs—nonargument NPs can go through without a theta role. The reasons for case agreement remain vague in Fanselow’s analysis. He claims that SNP is confined to NPs bearing structural case, i.e., nominative and accusative, and shows how accusative can be assigned to both NPs under the assumption that the non-referential NP is adjoined to V˚. Apart from the empirical problem that SNP in dative is possible (19), it is unclear what enforces that the two NPs get the same case—coindexation between an NP and a pronoun is certainly allowed under differing case. (19)

Kindern hat er vielen geholfen. children.DAT has he many.DAT helped ‘As for children, he has helped many.’

The anaphoricity aspect of the analysis I proposed is very similar to Fanselow’s approach, making the exact nature of “coindexation” between the two NPs explicit. In the LFG analysis, case agreement follows from the additional assumption of f-structure unification of the two NPs. Such a configuration becomes possible because the version of the Theta Criterion—or its LFG equivalent Biuniqueness—I assume, 21 does not strictly exclude the contribution of descriptive information from several categorial sources to a single argument slot, as long as their semantic 21 This is not even a modification of standard Biuniqueness but follows from allowing a divergence of the classical functions of PRED values discussed in sec. 3.1.

The syntax and semantics of split NPs in LFG

types are compatible to form an appropriate variable and restrictions on it (which is controlled by a separate, non-unificational component). 4.3 Base generation and semantic incorporation Van Geenhoven (1996) regards SNP as a semantic counterpart for morphological noun incorporation (into verbs) in West Greenlandic. She claims (in her sec. 4.3) like Fanselow that in SNP, two autonomous indefinite NPs are generated in the Mittelfeld, one of which is moved to the sentence initial topic position (because only one of the NPs can get “weak case” from the verb). The denotation of both these indefinites is of the semantic type of a predicate, i.e., they do not themselves introduce a variable/discourse referent. The variable for them is rather introduced and existentially bound by the verb—in this sense, van Geenhoven speaks of semantic incorporation of the indefinite into the verb. As empirical support for this analysis, she cites examples like (20), in which the SNP indefinite cannot outscope negation—as expected since quantification of the variable for the indefinite takes place within lexical semantics of the verb. (20) a.

b.

(van Geenhoven, 1996: p. 41) *Schwarze Spinnen hat Lisa im Keller einige nicht gesehen. black spiders has Lisa in-the cellar some not seen [intended interpretation:] ‘As for black spiders, there are some that Lisa didn’t see in the cellar.’ Schwarze Spinnen hat Lisa im Keller keine gesehen. black spiders has Lisa in-the cellar NEG-some seen ‘As for black spiders, it is not the case that Lisa saw some in the cellar.’

Van Geenhoven (1996:147) remarks that in her account no coindexation of the two NPs is required, as they “can be unambiguously interpreted as predicates of the internal argument’s variable introduced by the verb”. In a footnote (p. 40, fn. 13), van Geenhoven mentions an empirical problem for her account in examples with “wide scope split topics” like (21), claiming however that they are restricted to “contrastive contexts”.

Jonas Kuhn (21)

(van Geenhoven 1996: p. 40, fn. 13, attributed to Kamp, p.c.) Orthographische Fehler waren ihm sogar drei nicht aufgefallen. Orthographical mistakes were him even three not struck ‘As for spelling mistakes there were even three that he didn’t recognize.’

How this contrastive usage could be explained is unanswered. Note that this usage is by no means exceptional—according to the recent semantic theory of topic and focus in (Büring, 1995), which builds on top of Rooth’s (1992) alternative semantics, all usages of an s-topic (sentence topic) can be reduced to the contrastive case: the s-topic picks out one element of a contextually constrained set of alternative questions. Apparently, van Geenhoven’s assumption that both NPs involved in SNP are necessarily of type property is too strong; the one that stays in the Mittelfeld can (at least alternatively) also introduce a variable itself, just like an ordinary full NP or an elliptical NP of identical form. This is essentially the analysis that I argued for; the coindexation that was not required in van Geenhoven’s analysis has to come back in, since the Mittelfeld NP is no longer semantically incorporated into the verb. However, the parallelism with elliptical NPs case gives this coindexation for free (as anaphoric binding). 22 Further evidence for the asymmetry between the two parts of SNP comes from the empirical fact that the two cannot always be swapped (contrary to the facts from West Greenlandic, cf. van Geenhoven, 1996: p. 20), as van Geenhoven herself remarks (p. 27): (22)

*Zwei hat Julius Briefe gekriegt. two has Julius letters got ?#‘As for two things, Julius has got letters.’

(van Geenhoven, 1996: p. 27)

22 The “buzzard example” from fn. 9 above is not compatible with an elliptical NP analysis, as van Geenhoven (p.c.) notes. However since the acceptability of such examples is significantly lower than of other SNP examples, I assume that they should be explained as a performance effect (like other duplications), which allows to merge the two predicates when the lower one can be interpreted as a specialization of the higher one. This explains why (i) is out, while the adjectival “elliptical” SNP case (ii) with the same intended meaning is perfect. (i) *Volleyballspieler hat sie nur Freunde eingeladen. volleyball-players has she only friends invited (ii) Volleyballspieler hat sie nur befreundete eingeladen. volleyball-players has she only friendly invited ‘As for volleyball players, she only invited friends of her.’ As far as I can see, van Geenhoven’s (1996) “double incorporation” account could not exclude (i).

The syntax and semantics of split NPs in LFG

5 Conclusion In this article, I argued for an LFG analysis of the Split NP construction, exploiting the idea of argument doubling as f-structure unification, and anaphoricity of elliptical NPs for a topic of the semantic type property. Through an interplay of independently motivated mechanisms, the observed empirical behaviour could thus be predicted correctly, without having to make any construction-specific stipulations. The analysis relies on the assumption that the functional-syntactic specification—regulating which argument slot a certain category contributes to—does not by itself exclude doubling (i.e., the situation of more than one category contributing to the same slot), as long as the relevant syntactic features like case unify. This somewhat generous level of functional specification is complemented by a semantic component performing resource accounting based on the semantic types of the contributing categories. This component cannot be adequately modelled on the basis of purely monotonic unification, as the limitations addressed in sec. 3.2 showed. Kuhn (1998) presents a linear logic based analysis of the SNP construction that provides the natural basis to express the resource sensitivity of the elements involved, and that completely abandons the bookkeeping of P R E D values at the level of f-structure, thereby establishing a clear-cut distinction between the resource-sensitive and the non-sensitive parts of grammar.

References ANDREWS, A. D. 1990. Unification and Morphological Blocking. Natural Language and Linguistic Theory, 4, 507–557. BAYER, J. 1987. The syntax of scalar particles and so-called “floating quantifiers”. Ms., Max-Planck-Institute, Nijmegen. B RESNAN , J. 1995. Lexical-Functional Syntax (Barcelona version). Manuscript, Stanford University, reproduced at ESSLLI 95, Barcelona. BRESNAN, J. 1996. Morphology competes with syntax: explaining typological variation in weak crossover effects. Manuscript, Stanford University. BÜRING, D. 1995. The 59th Street Bridge Accent. On the Meaning of Topic and Focus. PhD thesis, Universität Tübingen. D ALRYMPLE , M., HINRICHS , A., LAMPING , J., and SARASWAT , V. 1995a. The Resource Logic of Complex Predicate Interpretation. Technical report, Xerox Technical Report. D ALRYMPLE, M., KAPLAN , R. M., MAXWELL , J. T., and ZAENEN , A., eds. 1995b. Formal Issues in Lexical-Functional Grammar. CSLI Publications, Stanford, CA.

Jonas Kuhn D ALRYMPLE, M., LAMPING, J., PEREIRA, F., and SARASWAT, V. 1997. Quantifiers, Anaphora, and Intensionality. Journal of Logic, Language, and Information, 6, 219–273. DIESING, M. 1992. Indefinites. No. 20 in Linguistic Inquiry Monographs. MIT Press, Cambridge. FANSELOW, G. 1988. Aufspaltung von NPn und das Problem der ‘freien’ Wortstellung. Linguistische Berichte, 114, 91–113. F ANSELOW , G. 1993. Die Rückkehr der Basisgenerierer. Groninger Arbeiten zur germanistischen Linguistik, 36, 1–74. H A I D E R , H. 1990. Topicalization and other Puzzles of German Syntax. In G. Grewendorf and W. Sternefeld, eds., Scrambling and Barriers, vol. 5 of Linguistik Aktuell. John Benjamins. HINRICHS, E. W. and NAKAZAWA, T. 1994. Partial-VP and Split-NP Topicalization in German – An HPSG Analysis. Arbeitspapiere des SFB 340 Nr. 58, Universität Tübingen. KAMP, H. and REYLE, U. 1993. From Discourse to Logic. Kluwer, Dordrecht. K APLAN , R. M. and BRESNAN , J. 1995. Lexical-Functional Grammar: A Formal System of Grammatical Representations. In Dalrymple et al. 1995b. Originally appeared in The Mental Representation of Grammatical Relations, ed. Joan Bresnan (Cambridge, MA: The MIT Press, 1982) 173-281. KUHN, J. 1998. Resource Sensitivity in the Syntax-Semantics Interface and the German Split NP Construction. In T. Kiss and D. Meurers, eds., Proceedings of the ESSLLI X workshop “Current topics in constraint-based theories of Germanic syntax”, Saarbrücken. Also available via http://www.ims.uni-stuttgart.de/~jonas/ ROOTH, M. 1992. A Theory of Focus Interpretation. Natural Language Semantics, 1, 75–117. VAN GEENHOVEN, V. 1996. Semantic Incorporation and Indefinite Descriptions. PhD thesis, Universität Tübingen. Appeared as SfS-Report-03-96. VAN R IEMSDIJK, H. 1989. Movement and Regeneration. In P. Benicà, ed., Dialect Variation and the Theory of Grammar. Foris, Dordrecht.