Anna Cardinaletti Universita di Venezia
Ian Roberts University of Wales
1. Agr1 as a Position for the Inflected Verb: Embedded V2
1.2 Old French
2. Agr1 as a Position for Clitics
2.1 The Wackernagel Position in German
2.2 "Tobler‑Mussafia" Effects
2.3 V3 Orders in Old English and Old High German
3. Subject Positions
4. Null Subjects
4.1. Expletive Null Subjects
4.2. Referential Null Subjects
The purpose of this paper is to propose a unified analysis of a range of "second‑position" phenomena that have been attested in various languages, and in so doing to motivate a more elaborated theory of Nominative Case assignment. The proposal is that many languages, including a number of Germanic and Romance languages, have a projection which intervenes between Comp and the highest Infl‑type projection (which, following Belletti (1990), we take to be AgrP). We refer to this projection as Agr1P and we refer to the lower, "traditional" AgrP as Agr2P. Thus our claim is that in the languages in question there are two Agr‑heads, and two projections of Agr. These two Agrs are both "subject" Agrs; in this respect, our proposal is distinct from but not exclusive with Chomsky's (1989) idea that in addition to the standard "subject" Agr, there is also an "object" Agr. We will show that this proposal is of considerable empirical value in that it offers a new perspective on a range of second‑position phenomena and allows us to connect "verb‑second" effects with various kinds of "clitic‑second" effects, known in traditional grammar as Wackernagel's Law and the Tobler/Mussafia Law. We will in fact suggest that the presence of Agr1P is fundamentally related to Nominative Case assignment, in that thebasic property of Agr1E seems to be that of assigning Nominative Case; the other properties that we ascribe to it (e.g. attracting clitics, or attracting the inflected verb) are intimately related to its Nominative‑assigning property. In this sense, it may be best to think of Agr1P as NomP. As a working hypothesis, then, we assume that in languages which have both Agr1P and Agr2P, Agr2E is not an assigner of Nominative Case. Because our focus is on the interaction of clause structure with structural Case assignment, we will concentrate almost exclusively on processes and properties of S‑structure. It is a classic tenet of generative grammar that inflectional affixes may be separate syntactic entities at pre‑phonological levels of representation; cf. the analysis of the English auxiliary system in Chomsky (1957). In the last two or three years, this idea has received new impetus, beginning with Pollock (1989). Pollock's proposal that the Infl‑node of Chomsky (1981) should be split into its morphological components has led to the working hypothesis in much recent research that any inflectional head which appears to be syntactically relevant heads its own maximal projection with the standard X‑bar theoretic structure (for a different view, see Williams (this volume)). Our proposal amounts to the claim that certain languages have a special X‑bar projection for the assignment of Nominative Case; in terms of the connection with inflectional morphology, it may be possible to relate the existence of Agr1P to the possession of morphologically‑realized Nominative Case. The structure that we are proposing is as follows:
Spec C' CE Agr1P Spec Agr1' Agr1E Agr2P Spec Agr2' Agr2E TP The order of TP and Agr2E varies in the languages under consideration: in West Germanic TP is on the left of Agr2E (cf. Giusti (1986) for the proposal that IP is head‑final in German); in North Germanic and Romance, it is on the right. On the other hand, Agr1E, like CE, precedes Agr2P in all the languages we discuss. The paper is organized as follows: in Section 1 we analyze the phenomenon of embedded verb‑second, basing ourselves largely on the best‑known case of this type: Icelandic (although we also analyse both Old French and Yiddish in these terms). In Section 2 we show how our system gives a natural analysis of the "clitic‑second" phenomena in Germanic and Romance; in our terms the traditionally‑recognized Wackernagel position of Germanic languages is Agr1E, as is the clitic position in those Romance languages which obey the Tobler‑Mussafia law. The last two sections deal with ways in which the properties of Agr1E vary parametrically: we discuss different modes of Nominative Case assignment in Section 3, where we present and elaborate the recent proposals of Koopman & Sportiche (1991), and different kinds of null subjects in Section 4.
1. Agr1 as a position for the inflected verb: embedded V2 1.1 Icelandic The postulation of Agr1P allows us to account in a straightforward way for certain differences within the class of verb‑second languages. While in many languages, e.g. German, Dutch and Mainland Scandinavian, verb‑second is essentially a root phenomenon, it appears to be generalized to all types of embedded clauses in Icelandic. The following data illustrate this, where in a variety of embedded clauses we have the order XP ‑ V ‑ subject:1 (1)a. Ég held aδ&egar hafi María lesiδ&essa bók.
I believe that already has Mary read this book.
'I believe that Mary has read this book already.'
b. Ég harma aδ&egar hafi María lesiδ&essa bók. (factive)
I regret that already has Mary read this book.
'I regret that Mary has already read this book.'
c. Ég spurδi hvort &egar hefδi María lesiδ&essa bók. (Wh)
I asked whether already had Mary read this book.
'I asked whether Mary had already read this book.'
d. sú staδreynd aδ&egar hefur María lesiδ&essa bók (NP)
the fact that already had Mary read this book
'the fact that Mary had already read this book'
e. bókin sem &egar hefur María lesiδ (relative)
book‑the that already had Mary read
'the book that Mary had already read'
(Rögnvaldsson & Thráinsson (1990); Thráinsson, p.c.) The Icelandic situation, as illustrated in (1), contrasts with what we find in German. In German embedded V2 is possible only in a limited class of embedded clauses, essentially the complements to "bridge verbs" of the type in (1a). On the other hand, embedded V2 is excluded in all the contexts parallel to (1b‑e):2 (2)a. Ich glaube, gestern hat Maria dieses Buch gelesen.
I believe yesterday has M. this book read.
'I believe Mary read this book yesterday.'
b.*Ich bedauere,(daß) gestern hat Maria dieses Buch gelesen.
I regret (that) yesterday has M. this book read.
c.*Ich frage mir, ob gestern hat Maria dieses Buch gelesen.
I ask me whether yesterday has M. this book read.
d.*die Tatsache, gestern hat Maria dieses Buch gelesen
the fact yesterday has M. this book read.
e.*das Buch, das gestern hat Maria gelesen
the book which yesterday has M. read For (2a), an analysis in terms of CP‑recursion seems to be in order. As noted in Rizzi & Roberts (1989), the same class of verbs allows an otherwise root phenomenon ‑‑ subject‑aux inversion triggered by a negative‑polarity item ‑‑ in its complement in English: (3)a. I believe that only in America could you do such a thing. b.*I wonder whether only in America could you do such a thing.
It seems then that, independently of verb‑second, the complements of bridge verbs are able to have root properties. We propose, still following Rizzi & Roberts (1989), that this is because bridge verbs allow CP‑recursion in their complements. More precisely, we propose that bridge verbs select a CE which selects another CE; to avoid unlimited recursion at the C‑level we clearly must propose that the two CEs have different properties (e.g. that the first allows a "propositional" complement while the second only allows a "predicational" complement in the terms of Rizzi (1990b)). In German the two CEs are different in form: the first is null, the second is filled by the inflected verb, like the CE of a matrix clause (in English, too, the CE which selects CP is different from other CEs: that in (3a) cannot be deleted, while other occurrences of that can be). Adopting this analysis for (2a), we propose the following partial structure: (4) .. glaube [CP O [CP gestern [C' hat [ Maria ...
Vikner (1990) proposes extending this analysis to embedded verb second in Icelandic. This entails that CP‑recursion is generalized in Icelandic, while it is limited to a specific class of complements in German, Mainland Scandinavian and English. In other words, the property of selecting CE is available for all classes of CE in Icelandic. However, if this were true, then there would be no way to avoid unlimited recursion of CE, clearly an undesirable consequence. Instead, our proposal provides a straightforward account of the data in (1). These examples have the following structure (although, to the extent that the class of verbs which allows CP‑recursion in German also allows it in Icelandic, (1a) may also have a structure like (4) with aδ in the higher CE and the inflected verb in the lower CE):3 (5) [CP CE [Agr1P TOP [Agr1' V+Agr1 [Agr2P NPNom [Agr2E t As we will see in more detail in Section 3, the special property of Icelandic is that SpecAgr1' is a topic position while the usual superficial subject position is SpecAgr2'. If we do not assume a double‑Agr structure, no other position would be available for the subject. Since there is no generalized CP‑recursion, SpecC' is not available. Although we assume that the subject is base‑generated in VP (cf. in particular Koopman & Sportiche (1991)), the base‑position of the subject is unavailable, at least for a definite NP, since it is not a position which can receive Nominative Case from AgrE (cf. Section 4 for some evidence that indefinite NPs can appear in this position). This idea is confirmed by the fact that definite subjects always precede VP‑adverbs, as in .. aδ Pegar hefur Maríaoft lesiδ Pessa bók, 'that already has Mary often read this book' (Thráinsson, p.c.) (cf. also Vikner (1990, Ch 2:45)). Furthermore, we follow Rizzi (1990a) in assuming that SpecT' is inherently an A'‑position and as such not a possible landing site for the subject. In fact, if we adopt Rizzi's (1991b) characterization of potential A‑positions as either Θ‑positions or specifiers of Agr, SpecT' must be an A'‑position. Hence SpecAgr2' is the position of the definite subject. Given this analysis of embedded clauses, we are not forced to treat matrix V2 in Icelandic as involving movement of the verb to CE. Movement to Agr1E would clearly suffice to derive the same orders (cf. Rögnvaldsson & Thráinsson (1990) and the references given there for recent discussion of similar proposals). At the same time, the data do not force us to reject a movement‑to‑C analysis.
One property of Icelandic, however, suggests that matrix V2 should in fact be handled in terms of verb‑movement to Agr1E rather than to CE. Icelandic makes much more frequent use of declarative V1 orders than do the other (Modern) Germanic languages (aside from Yiddish). Declarative V1 is illustrated in the following example: (6) Hitti hann pá einhverja útlendingar.
Met he then some foreigners.
'He then met some foreigners.'
(Sigurδsson (1985:1)) We propose that in (6) the inflected verb has undergone structure‑preserving topicalization (i.e. topicalization of an XE category to another head position YE). This operation probably takes place for reasons connected to information structure, since the examples in question seem to be presentational sentences. We propose that the landing‑site of this operation is CE; these are the only declarative clauses in which the verb is in CE (as in other Germanic and Romance languages, the inflected verb is typically in CE in matrix interrogatives, imperatives and hypotheticals). As we will see in 2.2, this kind of "verb‑topicalization" is not restricted to Icelandic, but is also found in Medieval Romance languages (cf. also Benincà (1989), Alberton (1990)). More generally, we expect this possibility to exist in all languages which realize V2 at the Agr1P‑level since in such cases CE is freely available as a landing site for structure‑preserving topicalization. Since the V2 requirement is satisfied elsewhere, SpecC' in such languages can remain empty. This analysis implies that declarative operators do not exist, since otherwise we would expect V1 declaratives to be generally possible on a par with V1 interrogatives, hypotheticals, etc.4 One reason to favour an analysis of (6) in which the verb moves to CE over one in which the verb moves only to Agr1E, while the subject stays in SpecAgr2', is that this kind of V1 is a root phenomenon. Since den Besten (1983), the simplest treatment of root phenomena has been to say that they involve movement to CE, a position available in principle in matrix clauses but unavailable in embedded clauses. If matrix clauses are Agr1P in Icelandic, this implies that verb second is not a unified phenomenon in the Germanic languages, at least in the sense that the landing site of the verb may vary cross‑linguistically. We will see further evidence in favour of this conclusion as we proceed. Cf. also Diesing (1988, 1990), Santorini (1988, 1989) for a similar conclusion based on Yiddish evidence (and cf. 1.3 for some discussion of Yiddish). To sum up, following on from our "double‑Agr" hypothesis about basic clause structure, the following conclusions emerge for Icelandic: (7)a. Agr1E can assign Nom under government (see Section 3);
b. SpecAgr2' is a subject position;
c. SpecAgr1' can be a topic position;
d. SpecC' is an operator position.
1.2 Old French The V2 nature of OF is illustrated clearly by the examples in (8) (non‑nominative clitics ‑‑ e.g. en in (8b) ‑‑ are effectively part of the inflected verb, and so do not "count" in the computation of the second position): (8) a. Einsint aama la demoisele Lancelot.
Thus loved the lady Lancelot
'Thus the lady loved L.'
(Adams (1987b:50)) b. Desuz un pin en est li reis alez
(Charroi de Nîmes, l. 20) Adams (1987a,b) shows that V2 is possible in the complements to bridge verbs. The class of bridge verbs in question is comparable to the class which in V2 Germanic languages typically allows complements with matrix properties (cf. 1.1). Here are some examples with null subjects: (9)a. Or voi ge bien, plains es de mautalant.
Now see I well full are (you) of bad‑intentions.
'And now I see clearly that you are full of bad intentions.'
(Le Charroi de Nîmes, l. 295)
b. Je cuit plus sot de ti n'i a.
I think more stupid than you not there has (it).
'I think that there is noone stupider than you.'
(Adams' (11a‑c), p. 17) In neither of these examples is que present, suggesting that these are cases of German‑style embedded V2 (this implies that "conjunctive discourse" is not specific to German, cf. Note 2). The embedded sentences in examples like (9) can be analysed as follows: (10) CP1 C1E CP2 AdvP C2' C2E AgrP V+Agr NP TP
It is clear that CP2 here is just like a matrix clause, and so V2 is possible, as expected. However, there are cases of V2 orders in Wh‑clauses: (11)a. Quant a eus est li rois venus, ..
When to them is the king come, ..
'When the king came to them, ..'
(Dupuis (1989), (40), p. 148) b. s'a la vostre bonté vousist mon pere prendre garde
if against the your good‑will wanted my father to‑take precaution
'if against your good will my father wanted to take precautions.' (Adams (1988b), (19c), p. 19)
(12) a. Por l'esperance qu'an lui ont, ...
For the hope which in him have (they) ..
'For the hope which they have in him, ...' b. Et si ne sait que faire puisse
And so not knows what to‑do can (he)
'And so he doesn't know what he can do' In terms of the standard assumption that the inflected verb cannot move to a [+wh] CE (cf. Rizzi & Roberts (1989), Rizzi (1991a) for an account of this), we are led to the conclusion that the verb is in Agr1E in these examples. So we assign the following structure to (11a), for example: (13) CP XP C' CE Agr1P
Spec Agr1' Agr1E Agr2P Spec Agr2' Agr2E TP Quant a eus est li rois t t venus Here the verb appears in Agr1E and assigns Case under government to the subject, li rois, in SpecAgr2'. What is the status of SpecAgr1' in OF? On the basis of examples like (11), it appears to be a topic position of the Icelandic kind. However, clear examples of the type in (11) are not very frequent. According to Dupuis (1989:151f.), this possibility is only attested with any real frequency in the Quatre LivresduRoi (a text from around 1170); in other texts (including some from the same period), there are very few cases of embedded V2 (in non‑bridge complements) with overt subjects. There are, however, cases of embedded V2 order with null subjects in a range of 12th‑century texts, as well as in some 13th‑century verse texts (cf. Hirschbuhler (1990) and 4.2). Such examples are not clear cases of topicalization and can instead be treated as Stylistic Fronting. Stylistic Fronting is an operation found in Icelandic, Faroese and Yiddish. The operation fronts some VP‑constituent, usually an adverbial, a participle or a complement (see Maling (1980) for details), to a position between CE and the inflected verb. The main condition on Stylistic Fronting is that the sentence must contain a subject gap, usually a trace of Wh‑movement, but also possibly the trace of cliticization (Platzack (1988)) or an NP‑trace (Sigurδsson (1989)). The following Icelandic example illustrates the application of Stylistic‑Fronting to a participle: (14) &arna er konan sem kosin var/var kosin forsetí.
There is woman that elected was/was elected president.
'There is a woman that was elected president.'
One test favours a Stylistic‑Fronting analysis of examples like (12) over a topicalization analysis. Platzack (1988) shows that the subject‑gap condition can be satisfied by a cliticized subject pronoun. This gives rise to the order Complementizer ‑ Subject ‑ XP ‑ V, which cannot be a case of embedded verb second. Such orders are found at the relevant period of OF: (15) quant il de ci departiront
when they from here will‑leave
'when they will leave here'
(Vance (1988), (11), p. 89)
The existence of this kind of order, and the absence of clear embedded topicalization, at the relevant period of OF suggests that apparently ambiguous examples of the kind in (12) are to be treated as involving Stylistic Fronting. For reasons of space, we make no proposal for the analysis of Stylistic Fronting in this paper, although we note that it does seem to correlate with the double‑Agr structure in VO languages: Icelandic and OF both have both properties, as arguably do Middle English and the Medieval Mainland Scandinavian languages (cf. Platzack (1988, 1990)). The conclusion is that, with the possible exception of some of the very early texts, OF does not allow generalized embedded topicalization, although it does have Stylistic Fronting. We interpret this to mean that SpecAgr1' was not a topic position in OF, but a subject position. Nevertheless, movement of the inflected verb to Agr1E was general, and so the movement of the inflected verb to Agr1E does not necessarily imply that SpecAgr1' is a topic position (cf. Section 3). In Section 4.2, we will see that the distribution of null subjects in OF supports this conclusion. To sum up, OF had the following properties: (16)a. Agr1E can assign Nominative under government; b. SpecAgr2' is a subject position; c. SpecAgr1' is also a subject position (see Section 3).
1.3 Yiddish The other Germanic language which has been claimed to allow generalized embedded topicalization is Yiddish (cf. Diesing (1988, 1990), Santorini (1988, 1989)). The following examples illustrate topicalization in [‑wh] and [+wh] complements:
(17)a. John bedoyert az zayn bukh hob ikh geleyent.
John regrets that his book have I read.
'John regrets that I have read his book.'
(Vikner (1990:Ch 2, 32)) b. Ikh veys nit far vos in tsimer iz di ku geshtanen.
I know not for what in room is the cow stood.
'I don't know why the cow has stood in the room.'
(Vikner (1990:Ch 2, 34)) c. Ikh veys nit tsi in tsimer iz di ku geshtanen.
I know not if in room is the cow stood.
'I don't know if the cow has stood in the room.'
(Santorini, p.c.) Since "regret" is a non‑bridge verb, (17a) is most likely not a case of CP‑recursion. (17b,c) are cases of topicalization in a [+wh] complement. Both Diesing (1988, 1990) and Santorini (1988, 1989) propose to analyse this kind of embedded V2 in Yiddish by treating SpecI' ‑‑ the canonical subject position in a language like English ‑‑ as a topic position, and taking the subject to be in its VP‑internal base position. However, this analysis will not carry over straightforwardly in our terms. If we split IP into TP and AgrP, and assume that Nominative Case is assigned by AgrE, then the VP‑internal subject position cannot receive Nominative Case.Instead, this position receives Partitive Case (cf. Belletti (1988)) and therefore can only be occupied by indefinite NPs. Since we are assuming that SpecT' is an A'‑position (see 1.1), no (non‑scrambled) NP can occupy this position. Hence the only position for a definite Nominative NP is SpecAgr'. For these reasons, the examples in (17) motivate an analysis in terms of the "double‑Agr" structure. So (17a) would have the following structure:
CE Agr1P NP Agr1' Agr1E Agr2P NP Agr2' Agr2E TP az zayn bukh hob ikh t t geleyent Further evidence in favour of embedded topicalization and against CP‑recursion comes from sentences like the following where an argument undergoes long Wh‑movement: (19) Vos hot er nit gevolt az in shul zoln di kinder
What has he not wanted that in school shall the children
'What didn't he want to the children to read in school?'
(Santorini (1989:59)) (As expected, the analogous long extraction of an adjunct is impossible (Vikner (1990, Ch. 2, p. 60), Vikner & Schwartz (1991, 3.1))). In general, extraction from complements with CP‑recursion is impossible (cf. Rizzi & Roberts (1989)). This, combined with the fact that CP‑recursion under verbs of volition is otherwise unattested, argues that there is no CP‑recursion here. Instead, then, this must be a case of embedded topicalization. However, there are a number of restrictions on embedded V2 in Wh‑complements in Yiddish. Only a limited class of Wh‑complements, e.g. those in farvos ('why') and tsi ('if') allow embedded topicalization. Compare (20) with (17b,c): (20)a. *Ikh veys nit vu nekhtn iz di ku geshtanen.
I know not where yesterday is the cow stood.
'I don't know where the cow stood yesterday.'
(Vikner (1990, Ch. 2, p. 34)) b. *Ikh veys nit ven zayn khaver hot Moyshe getrofn.
I know not when his friend has M. met.
'I don't know when M. met his friend.'
(Santorini, p.c.) c. *Ikh veys nit (far) vemen zayn khaver hot Moyshe forgeshtelt.
I know not (to) whom his friend has M. introduced.
'I don't know who M. has introduced his friend to.'
(Short extraction of an argument Wh‑element seems to be possible over a topicalized adjunct PP, however: Ikh veys nit vemen inrestoran hot Moyshe getrofn 'I don't know whom in the restaurant has Moyshe met'. This situation should be compared with that in Icelandic mentioned in Note 1).
On the other hand, topicalization in a Wh‑complement apparently becomes generally possible where (i) the subject is extracted, (ii) the subject is indefinite and apparently VP‑internal: (21)a. Zi iz gekumen zen ver frier vet kontshen.
She is come see who earlier will finish.
'She has come to see who will finish earlier.'
(Diesing (1988:132)) b. Ikh veys nit vu nekhtn iz geshtanen a ku.
I know not where yesterday is stood a cow.
(Vikner (1990, Note, Ch. 2, p. 34) Both contexts involve a gap in the canonical, preverbal subject position. As such, they appear to be cases of Stylistic Fronting (cf. 1.2). In relative clauses, verb‑second order arises where there is a subject gap, but not otherwise. Relatives on the object are possible, provided that a resumptive pronoun is used:
(22)a. nokh epes, vos oyfn hitl iz geven
still something, that on the hat‑DIM is been
'something else that was on the little hat' b. *Der yid vos in Boston hobn mir gezen iz a groyser lamdn.
The man that in Boston have we seen is a great scholar.
'The man whom we saw in Boston is a great scholar.'
(Lowenstamm (1977), cited in Santorini (1989:56))
(23) der yid vos in Boston hobn mir im gezen
the man that in Boston have we him seen
'the man that we saw (him) in Boston'
(Lowenstamm (1977), cited in Diesing (1990:63))
Here again, the possibility of embedded topicalization depends on the presence of a subject gap, confirming the idea that these are cases of Stylistic Fronting. The contrast between (22b) and (23) supports the idea that embedded topicalization is incompatible with any form of Wh‑movement (with the exception noted after (20)). It seems, then, that Yiddish allows embedded verb second under rather limited conditions: in non‑bridge complements and in Wh‑complements where the Wh‑element is not moved. Stylistic‑Fronting is possible in clauses where Wh‑movement affects the subject. In other words, embedded topicalization creates an island for extraction in Yiddish Wh‑complements (cf. Vikner (1990: Ch 2, 34f.)). This seems to be a property specific to Yiddish, since embedded topicalization does not have this effect in Icelandic (or perhaps the effect is limited to extraction of adjuncts over adjunct topics ‑‑ cf. Note 1). In spite of this difference, Yiddish shows the same basic properties as Icelandic: it has a double‑Agr clause structure, SpecAgr1' is a topic position, SpecAgr2' is a subject position and Nominative Case can be assigned under government.
2. Agr1 as the position for clitics
It has often been noticed, since the earliest work on Indo‑European syntax (cf. Wackernagel (1892)), that unstressed elements of various kinds tend to be found in the second position in the clause. In this section, we will propose a general analysis of this salient fact about the syntax of a variety of languages in terms of the following general idea:5 (24) Agr1E is a position for clitics. In the languages we discuss below, Agr1E is preceded by an element occupying its Spec. Since Agr1E is occupied by a clitic, the result is a "clitic‑second" structure. We will see that there is cross‑linguistic variation in these structures regarding the position of the inflected verb relative to the clitic; our system +makes it possible to analyse this variation in a straightforward way. 2.1 The Wackernagel position in German A striking property of German which differentiates it from other Germanic languages is the fact that pronouns can occur between CE and the subject. The following examples illustrate this phenomenon for both embedded and matrix clauses (in all our German examples the subject is unstressed; if the subject receives focal stress, the judgements can change, presumably owing to the fact that in this case the subject occupies a different position ‑‑ a matter which we leave aside): (25)a. .. daß es ihm der Johann gestern gegeben hat.
.. that it him‑DAT the J. yesterday given has.
'.. that John gave it to him yesterday.'
b. Gestern hat es ihm der Johann gegeben.
Yesterday has it him‑DAT the J. given.
'Yesterday John gave it to him.' In terms of our system, these examples have the following structure:
(26) C' CE Agr1P daß Agr1'
Agr1E Agr2P es ihm NP Agr2' der J. ... gegeben ... As (26) shows, the subject (der Johann) can remain in SpecAgr2'; it is not required to move to SpecAgr1'. However, it is also possible to have the order in which the pronouns follow the subject. In our system, it is possible to say that in such cases the clitics are in the same position as in (26), with the subject occupying SpecAgr1':
CE Agr1P daß NP Agr1'
der J. Agr1E Agr2P
es ihm NP Agr2' t .. gegeben .. b. .. daß der Johann es ihm gestern gegeben hat. c. Gestern hat der Johann es ihm gegeben.
The post‑subject position of the pronouns is not a VP‑internal position since the VP‑internal order of arguments is Dative‑Accusative in German. The fact that an Accusative pronoun must appear before a Dative NP indicates that it is outside VP: (28) .. daß der Johann es dem Hans gegeben hat.
.. that the J. it the‑DAT H. given has.
'.. that John gave it to Hans.'
Moreover, a post‑subject pronoun is typically unstressed and, like a pre‑subject pronoun, has clitic properties (cf. Boschetti (1986)). This suggests that post‑subject es occupies the same position as pre‑subject es, rather than being in a scrambled position of some kind (although other pronouns aside from es may be scrambled ‑‑ cf. Note 6). As indicated in (26) and (27), the clitic position is Agr1E.
A further argument in favour of our approach over a scrambling analysis comes from the absence of "clitic‑splitting" in German. That is, we do not find sentences where one pronoun precedes the subject and one follows:6 (29)a. * .. daß ihm der Hans es/sie wahrscheinlich gegeben hat.
.. that him‑DAT Hans it/her probably given has. b. * .. daß es/sie der Hans ihm wahrscheinlich gegeben hat.
.. that it/her the H. him‑DAT probably given has.
(29) is ungrammatical because there is only one position available for unstressed pronouns: Agr1E. Where there is more than one unstressed pronoun, they all appear in Agr1E. Furthermore, it is unlikely that (25) is a case of scrambling since many speakers reject scrambling to pre‑subject position, as shown by (30):
(30) ??.. daß den Roman ihm Johann gegeben hat.
.. that the‑ACC novel him‑DAT John given has. An important confirmation for this analysis of German comes from the fact that Dutch shows a pattern which differs minimally from German, something which we can treat very simply in our system.7 Zwart (1991) argues convincingly that Dutch has object clitics which occupy a VP‑external position (cf. also Jaspers (1989)). Zwart in fact proposes an analysis very similar to ours, involving a head‑initial functional projection whose head is to the right of the canonical subject position. This analysis provides a simple explanation for the order "subject ‑ pronouns" in Dutch: (31) .. dat Jan 't gisteren aan Marie gaf.
.. that J. it (cl) yesterday to M. gave.
'.. that John gave it to Mary yesterday.'
(31) is like the German examples in (27b,c), and we can analyse it in the same way: 't is in Agr1E and Jan is in SpecAgr1'.
Dutch differs from German, on the other hand, in that the order in (25) is ungrammatical: (32) *.. dat 't Jan aan Marie gaf. This means that the structure in (26) is impossible in Dutch. Since (31) shows, in our terms, that the clitic can occupy Agr1E, (32) must be interpreted as indicating that the subject cannot be in SpecAgr2'. The minimal difference between German and Dutch lies in the possible positions of the Nominative subject. We conclude that, although in both languages SpecAgr1' is a subject position, SpecAgr2' is a position which receives Nominative only in German (cf. Section 3).
2.2. "Tobler‑Mussafia" effects A related phenomenon to the ones just discussed is found in all the Medieval Romance languages (Old French, Old Italian, Old Spanish) and, at least prescriptively, in European Portuguese (Benincà (1989); see also Galves (this volume) for an analysis of the clitic systems of both European and Brazilian Portuguese in terms of the double‑Agr clause structure). The phenomenon in question amounts to a ban on clitic‑first orders; in constructions where a proclitic would appear in first position, enclisis is obligatory and proclisis is excluded (Mussafia (1983)). This phenomenon is known as the "Tobler/Mussafia law" in the traditional literature. The following examples illustrate the operation of the Tobler/Mussafia law in Old French (OF): (33)a. Toutes ces choses te presta Nostre Sires.
All these things you lent our Lord.
'All these things our Lord lent you.'
(de Kok (1985:74)) b. Voit le li rois.
Sees him the king.
'The king sees him.' (Le Charroi de Nîmes, l. 58) c. Fust i li reis, n'i oüssum damage.
Were here the king, not here had (we) damage.
'If the king were here, we wouldn't suffer any damage.'
In (33a), we have a regular V2 sentence (recall that OF was a V2 language; cf. 1.2). The direct object toutes ces choses is topicalized to SpecC'; and the finite verb appears in second position with the proclitic indirect‑object pronoun te. (33b) is an example of "narrative V1" (cf. Hirschbuhler (1990)). Here we see that the object clitic le follows the inflected verb; it is enclitic, not proclitic. (33c) is a verb‑initial conditional clause where the same type of enclisis affects the locative i. (34) illustrates the Tobler/Mussafia law in matrix interrogatives (de Kok (1985:82)): (34) a. Conois la tu?
Know her you?
'Do you know her?' b. Et quex chevaliers i avra il .. ?
And which knight there will‑be there .. ?
'And which knight will be there .. ?' In (34a), the clitic la is enclitic on the inflected verb (note that the subject pronoun follows it). In (34b), the clitic i is proclitic to the inflected verb (and subject pronoun still follows the verb). In (35) and (36), we illustrate the analogous effect in Old Italian (OIt): (35)a. Poi vi trovò tanto oro e tanta ariente
Then there (he) found much gold and much silver
'Then he found a lot of gold and silver there.'
(Novellino, LXXXIV; Alberton (1990))
b. Vogliolo sapere da mia madre.
(I) want‑it to‑know from my mother.
'I want to it from my mother.'
(Novellino, III; Alberton (1990))
(36)a. Hailo tu fatto per provarmi?
Have‑it you done to try me?
'Have you done it to try me?'
(Alberton (1990)) b. Chi si potrebbe tener di piangere e di lagri
mare in cotanto dolore?
Who self could (he) keep from crying and from weeping
in such pain?
'Who could keep himself from crying and weeping in such
pain?' (Alberton (1990)) A comparison of (35) and (36) with the French examples in (33) and (34) shows that the basic phenomenon is the same: wherever proclisis would place the clitic in first position, the clitic follows the inflected verb. There are, however, certain differences between OF and OIt; one important difference is that V1 declaratives are significantly rarer in OF than in OIt (Benincà (1989:5)). This may be due to the fact that the verb could remain in Agr1E in OIt matrix clauses (and so, to the extent that OIt was V2, V2 was realized at the Agr1‑level), while in OF movement to CE was required (i.e. V2 was realized at the C‑level). In this respect OIt would be like Icelandic (cf. 1.1), and OF differs from Icelandic (for other differences between OF and OIt, cf. Alberton (1990)).8 We gloss over these differences here since our main goal is to account for the obligatory enclisis.
Our account of the Tobler/Mussafia law runs as follows. Following Kayne (1991), we assume that clitics always occupy functional‑head positions. In particular, the clitic is in Agr1E prior to verb‑movement to CE. The difference between enclisis and proclisis lies in whether the clitic forms a complex with the verb, or whether the verb moves to CE independently of the clitic. Enclisis results from the latter situation: the verb moves to CE while the clitic remains in Agr1E. This is what we find in yes/no questions like (34a) and (36a) and conditionals like (33c), as well as in examples where the verb undergoes structure‑preserving topicalization to CE, e.g. (33b) and (35b). Proclisis, on the other hand, involves the formation of the complex [Cl + V] in Agr1E, which may then be moved to CE, as clearly seen in Wh questions like (34b) and (36b) (and V2 sentences where [Cl + V] moves to CE, as in OF examples like (33a)). A major advantage of this account is that it directly captures the fact that enclisis is a root phenomenon (Alberton (1990), Benincà (1989)); like other root phenomena, enclisis involves the presence of an inflected verb in CE. In this respect, our analysis is more straightforward than analyses of the type proposed by Alberton and Benincà, in which the clitic is in CE both with enclisis and with proclisis in matrix clauses and topicalizes to SpecC' to produce enclisis. The clause structure we wish to propose for these languages makes it possible to assume that the topicalization of V is movement to CE, rather than movement to SpecC'. As we said above, this integrates the account of enclisis more directly with accounts of other root phenomena (e.g. subject‑aux inversion in English or subject‑clitic inversion in contemporary French, and other phenomena discussed in den Besten (1983)). A further advantage of our account as compared to approaches of the type in Alberton and Benincà is that the topicalization rule is structure‑preserving, in that it moves the inflected verb, which is an XE, to CE. Our analysis leads to two questions: (i) why can the verb move to CE independently of the clitic in Agr1E? (ii) why must the verb move to CE independently of the clitic in cases of topicalization and yes/no questions? There are two possible answers to the first of these questions: either the verb "excorporates" from Agr1E to CE leaving the clitic behind in Agr1E (cf. Roberts (1991a) on excorporation), or the verb is able ‑‑ under the right conditions ‑‑ to move from Agr2E to CE "skipping" Agr1E. Of these two possibilities, we will assume the second in what follows (cf. Roberts (forthcoming)). For the second question, we see no alternative to the traditional idea that the languages which show this effect do not permit clitic‑first orders, presumably a phonological constraint. Hence the verb "skips" Agr1E exactly where the "regular" movement through Agr1E, picking up the clitic on the way, would give rise to a clitic‑first order with [Cl + V] in CE. The question which arises now is: why does V only "skip" Agr1E when moving through Agr1E would violate the ban on clitic‑first orders? As we will suggest in Section 3, the clitic in Agr1E attracts the inflected verb in the usual case (cf. Note 8). The clitic thus seems to impose two distinct requirements: (a) it cannot appear first, (b) it must combine with the inflected verb. Where some other element (e.g. a Wh‑constituent) appears in first position for independent reasons, requirement (a) is automatically satisfied and requirement (b) can be most economically satisfied by V‑movement to Agr1E and hence it must be satisfied in this way, following Chomsky (1989) (although other requirements, e.g. Rizzi's (1991a) Wh‑criterion, may lead to the [ Cl + V ] complex moving further). On the other hand, if no other element appears first, the two requirements imposed by the clitic are satisfied most economically if V moves to CE, skipping Agr1E, and the clitic left‑adjoins to V (the latter operation possibly taking place in PF). Thus we see that the Tobler‑Mussafia Law is a "last‑resort" operation in the sense of Chomsky (1989). Concerning comparative questions, our analysis makes the prediction that a language which shows the Tobler/Mussafia effect has Agr1E. We can relate the possibility of independent movement of the verb with respect to the clitic to the existence of this position in the following way: in a system with both Agr1E and Agr2E, Agr2E contains the inflectional affixes that are required to form the inflected verb. Hence the verb‑stem must move to Agr2E in order for an inflected verb to be formed. However, verb‑movement to Agr1E is not forced by such morphological factors (although the verb may be "attracted" to Agr1E independently of the need to pick up an inflectional affix; see below, Section 3, and Note 10). On the other hand, contemporary French and Italian have only one Agr position, and this is the position which contains the finite verbal inflection and to which clitics are adjoined. Thus the finite verb always forms a unit with object clitics in these languages, and so, when the finite verb appears initially in CE in yes/no questions ((37)) and certain types of conditionals ((38)), the clitic remains proclitic (we illustrate with a conditional in Italian, as no overt form of subject‑verb inversion is found in interrogatives; cf. Rizzi (1982) on this form of conditional inversion): (37) [CE [AgrE La connais ]] ‑ tu? (vs. (34a))
her know you?
'Do you know her?'
(38) [CE [AgrE L'avessi ]] io saputo in tempo, .... It‑had(subj.) I known in time, ..
'Had I known it in time, .. ' Note that our analysis does not claim that clitics form a unit with the non‑finite verb. As Kayne (1991) shows, it is probably desirable to maintain that enclitics on non‑finite verbs in some Romance languages occupy a syntactic position separate from the verb, e.g. in forms like Modern Italian farlo ('to do it'). Another point which is captured neatly by our approach is the fact that the Medieval Romance languages are like the contemporary Germanic languages in that both groups tolerate clitic elements which can be independent of the finite morphology, something which is completely impossible in most contemporary Romance languages (but not in Rumanian; cf. Motapanyane (1991) for an analysis of clitic placement in this language which makes use of the same "double‑Agr" system as that being proposed here). In our terms, what Medieval Romance and contemporary Germanic have in common is that their clitics occupy Agr1E, a functional‑head position which is independent in principle of the position of finite morphology ‑‑ Agr2E. Furthermore, our analysis captures the traditional idea that Tobler‑Mussafia effects are related to Wackernagel's Law; as we have seen, both sets of phenomena crucially involve the presence of an "autonomous" clitic in Agr1E.
2.3 V3 orders in Old English and Old High German In this section, we discuss one phenomenon of clitic‑placement found in the older Germanic languages Old English (OE) and Old High German (OHG). The data on OE are taken from van Kemenade (1987) and those on OHG from Tomaselli (1991). Both OE and OHG were V2 languages, although, as we shall see below, this does not necessarily imply that CP was activated in all matrix clauses in these languages. In this section, we focus on one case of V3 word‑order, found in both of these languages, in which the second element is a clitic (in OHG, this order is restricted to subject clitics; in OE complement clitics are found in this position too). This order is the same as that in the Medieval Romance languages in matrix declarative sentences: (39)a. God him worhte &a reaf of fellum.
God them wrought then garments of skin.
'God then made them garments of skin.'
(van Kemenade (1987:114) b. Forδon we sceolan mid ealle mod & mægene to Gode gecyrran.
Therefore we shall with all mind and power to God turn.
'Therefore we must turn to God with all our mind and
power.' (van Kemenade (1987:110))
c. Dhes martyrunga endi dodh uuir findemes mit urchundin
dhes heilegin chiscribes.
His martyrdom and death we demonstrate with evidence of
the holy writings.
'We demonstrate his martyrdom and his death with evidence
from the holy scriptures.' (Tomaselli (1991:3)) However, there is a major difference with the Romance Tobler/Mussafia effects discussed in the previous section. In contexts where the verb is in CE and an XP of a particular class is in SpecC', the clitic follows the verb. In OE, these contexts involve a Wh‑constituent, the negative element ne or the adverb &a ('then') in initial position, all of which can be plausibly treated as elements in SpecC' triggering movement of the inflected verb to CE (Tomaselli (1991:6) implies that the same is true in OHG; cf. Tomaselli (1989)): (40)a. Ne geseah hine nan man nates‑hwon yrre.
Not saw him no man so little angry.
'Noone ever saw him so little angry.'
(van Kemenade (1987:114)) b. Hwæt sægest &u, yr&lincg?
What saist thou, ploughman?
'What do you say, ploughman?'
(van Kemenade (1987:138‑9)) Parallel with our analysis of the Tobler‑Mussafia effect in 2.2, we propose that in examples like (39) the clitic occupies Agr1E. The verb can occupy this position, as the following OE case of verb‑subject order shows:
(41) Fela spella him sædon &a Beormas, æg&er ge of hiera
agnum lande, ...
Many stories him told the Permians, both of their own
'The Permians told him many stories, both about their
own country, ...' (van Kemenade (1987:114))
We cannot provide the parallel evidence in OHG, since the clitic in Agr1E is always a subject clitic. However, OHG examples like (39c) can be analysed as having the inflected verb in Agr1E together with the clitic. In Germanic matrix declaratives, then, the CP‑level is not activated.
The main verb can also appear in a position quite distant from that occupied by the clitic in embedded clauses and in the characteristically non‑V2 second conjuncts in main‑clause coordination (cf. Kiparsky (1990)). Notice that the clitic can appear either before or after the subject in OE, as in Modern German (cf. 2.1):
(42)a. .. &æt him his fiend wæren æfterfylgende.
.. that him his enemies were following.
'.. that his enemies were chasing him.'
(van Kemenade (1987:113)) b. & se cyng him eac wel feoh sealde.
and the king him also well property gave.
'and in addition the king gave him much property.'
(van Kemenade (1987:113))
c. .. daz íh nîeuuánne necúme in conventicula haereticorum
.. that I never not‑come in the circle of‑heretics
'.. that I never come into the circle of heretics.'
(Tomaselli (1991:29)) We take it that the inflected verb is in Agr2E in these examples. As in Medieval Romance languages, Agr2E is the position in which the inflected verb is formed by combining with the agreement affix; while Agr1E is the clitic position. The principal difference between these Germanic languages and the Medieval Romance languages discussed in 2.2 is that in the Germanic languages the verb only moves to Agr1E in matrix declaratives, while in the Romance languages the verb moves to Agr1E both in matrix and in subordinate clauses, and indeed in any clause where the disallowed clitic‑first order will not result from continued movement to CE. In Germanic interrogatives, etc., where the inflected verb is clearly in CE (cf. (40)), the verb "skips" Agr1E, and so the clitic appears in third position here. This is consistent with the fact that verb‑preposing is a root phenomenon in Germanic, while it is found also in embedded clauses in Romance. Our analysis captures the same range of facts as those proposed by van Kemenade (1987) for OE and Tomaselli (1991) for OHG. Compared to those analyses, ours has the advantage of greater generality, in that it relates the Germanic phenomena to the Tobler/Mussafia effects found in Romance. Also, the more elaborated clause structure that we propose makes it possible to distinguish the clitic position from the position of the inflected verb and to maintain that the inflected verb is always formed in the standard way in these languages, by head‑movement into a position containing the agreement affix (cf. Pollock (1989)). On the other hand, neither Tomaselli nor van Kemenade allow for a position reserved purely for verbal inflection; in such a framework, it is necessary to posit special rules for the formation of the inflected verb when it is "distant" from the clitic in examples like (42), an undesirable consequence that we are able to avoid.
3. Subject positions In this section, we will show how the subject positions are determined in the various languages which we have considered (we limit our attention throughout to non‑Θ subject positions, i.e. those outside VP). A central part of our discussion is based on the approach to Nominative‑Case assignment advocated in Koopman & Sportiche (1991). Koopman & Sportiche propose that UG makes available two mechanisms of Nominative‑Case assignment. Nominative Case can be assigned either under government, illustrated in (43a), or under agreement, illustrated in (43b): (43)a. X' XE YP NP Y'
b. XP YP X' XE Koopman & Sportiche consider that the choice among these two mechanisms is a "pure" parametric choice, in the sense that each language is in principle free to choose among these possibilities. We will suggest, however, that the choice of Nominative‑assignment under agreement depends on other factors, in particular on the nature of the Case‑assigning head XE, and that the only pure parametric choice is that of selecting (43a) or not. In the context of the general assumption that Agr1E is the Nominative‑Case assigner in all the languages we are concerned with (but see Note 9 for a refinement of this assumption), if a language chooses the government option in (43a) then SpecAgr2' will be the subject position; and if a language chooses the agreement option in (43b) then SpecAgr1' will be the subject position. We further assume that the two options can be combined, making both SpecAgr1' and SpecAgr2' possible subject positions. We understand "subject position" to mean both Case‑position, i.e. a position to which Nominative is assigned, and A‑position (although we are excluding Θ‑positions here, as we mentioned above). The generalization which emerges from a consideration of the languages we have looked at is that wherever Agr1E is a clitic position, its specifier is a subject position. On the other hand, where Agr1E is a "pure" verbal position (in the sense that the only element that ever appears there is the verb), its specifier is a topic position. Suppose, then, that it is the fact that Agr1E may host a clitic that makes it possible for it to assign Nominative Case under agreement. We will return to this idea below. First, however, let us consider the evidence from the various languages that points to this conclusion. Consider first the situation in Icelandic (and Yiddish, cf. 1.3). The evidence we presented in 1.1 shows that in this language Agr1E is a position which always hosts the inflected verb and that SpecAgr1' is a topic (non‑Case, A') position, while SpecAgr2' is a subject position. Schematically, then, we have the following:
(44) [Agr1P TOP [ [Agr1E V ] [Agr2P Subj. [Agr2E t ] This structure is common to main and embedded clauses. We can account for its basic properties in terms of Koopman & Sportiche's system by saying that Agr1E assigns Nominative under government only. Note that Agr1E always contains the inflected verb and never contains a clitic in Icelandic; we propose the lack of Case‑assignment under agreement is related to this point. Now compare German, on the basis of the analysis given in 2.1. In main clauses, we have the following: (45) [CP TOP [ [CE V ] [Agr1P Subj. [ [Agr1E Cl ] [Agr2P Subj. ]]
As in standard approaches, we take it that the verb moves to CE in main clauses and that SpecC' is a topic position. Agr1E is the Wackernagel position, and as such may contain one or more clitics. SpecAgr1' is a potential subject position in German, and not a topic position (cf. 2.1); this is the fundamental difference between German and Icelandic. SpecAgr2' is also a possible subject position in German, as it is in Icelandic. In terms of Koopman & Sportiche's system, we can say that in German Agr1E assigns Nominative both under government and under agreement. For this reason, SpecAgr1' is a subject position rather than a topic position. As we mentioned above, we believe that there is a non‑trivial connection between the possible presence of a pronoun in Agr1E and the fact that SpecAgr1' is a subject position. There is a further complication concerning the topic positions in both German and Icelandic, i.e. SpecC' in German and SpecAgr1' in Icelandic (this complication is probably common to all the V2 Germanic languages and to OF (cf. Roberts (forthcoming, 2.3)). These positions also display properties of subject positions. Apart from being available for topicalization of any XP, they accept elements which are clearly not topics: lexical expletives and non‑topic subjects (both pronominal and non‑pronominal); cf. Cardinaletti (1990b,c) for further arguments for this analysis. We can treat this situation in terms of Rizzi's (1991b) proposal that any specifier which agrees in Φ‑features with its head can be an A‑position (and therefore a subject position; other A‑positions are all structural complements, not specifiers). This agreement can naturally be thought of as a coindexation relation between the subject in SpecX' and XE. We refer to this possibility henceforth as