Apr 20, 1996 - do not apply to the VPs which are headed by getting clear or regretting it, .... (31) They managed to neutralize his brains, his most precious ...
A monostratal treatment of it extraposition without lexical rules Frank Van Eynde April 20, 1996 This paper deals with alternations of the type (1) That Sandy snores bothers Kim. (2) It bothers Kim that Sandy snores. In the monostratal framework of HPSG this alternation is described in terms of a lexical rule. The starting point is the lexical entry for the use of bothers in (1), cf. Pollard & Sag (1994, 150) h 2 3 i 66CAT j SUBCAT S comp : 1 , NP 2 7 377 2 66 7 66 66RELATION bother7777 64CONTENT 575 4BOTHERED 2 psoa
SOA-ARG
1
In words, bothers is subcategorized for an NP complement and a clausal subject which is introduced by a complementizer. Semantically, it expresses a two-place relation in which the complement corresponds to the bothered one and the subject to the state of a airs (SOA) which is bothering. Next to this entry there is a second one which accounts for the use of bothers in (2). It is derived from the rst entry by means of a lexical rule: \The Extraposition Lexical Rule removes an S[comp] from a SUBCAT list, replacing it by NPit , and appends the S[comp] to the end of the SUBCAT list, preserving role assignment." (Pollard & Sag (1994, 150)) As applied to bothers this yields the following entry 2 h i 3 66CAT j SUBCAT NPit , NP 2 , S comp : 1 77 2 37 66 7 RELATION bother 66 66 7777 64CONTENT 4BOTHERED 2 575 psoa
SOA-ARG
1
1
In this case the state of a airs corresponds to the most oblique complement, whereas the subject does not have a semantic role. This is made explicit by the type of its index: in contrast to the indices of NPs like Kim and you, which are of type referential, the one of the expletive pronoun is of a type sui generis, i.e. it. Since the lexical rule does not specify the position of the S[comp] on the SUBCAT list, it also covers the extraposition of object clauses, as in (3) We deeply regret that he has been red. (4) We regret it very much that he has been red. There are at least three problems with this treatment. One is that it does not account for the agreement between subject clause and nite verb in sentences of type (1); the second is that it erroneously predicts that the extraposed clauses in sentences like (2) and (4) behave like complements, and the third concerns the lexical rule, which is too restricted in some respects and too general in other respects. Each of these problems will be discussed in a separate section and an alternative will be developed which does not only provide a solution for these problems, but which also avoids the use of an extraposition lexical rule by making use of the computationally more attractive method of underspeci cation.
1 Subject clauses as nominal objects For modelling the agreement between subject and nite verb HPSG makes use of coindexation. The rst element on the SUBCAT list of a verb like bothers, for instance, is assigned the index [3rd, sing], and because of the Subcategorization Principle this index has to be token-identical to the one of the subject. This accounts for (5) This y bothers/*bother me. However, since clauses do not have an index it does not account for (6) That Sandy snores bothers/*bother Kim. The most straightforward way to bring these facts in line with the general treatment of agreement is to assume that also that clauses have a [3rd, sing] index. This would not only account for the subject-verb agreement, but also for the fact that that clauses can be coindexed with [3rd, sing] personal pronouns, as in (7) That Kim snores may bother Sandy, but it does not bother me. As for the origin of the index, there are basically two options. One is that it is projected from the clause's nite verb, but this option has to be ruled out since it erroneously predicts that the verb is invariably of the third person singular, cf. 2
(8) That you left without saying anything bothers her a lot. The other option is that it is projected from the complementizer. This is more plausible since the complementizer has no plural or non-3rd person counterparts, and since its omission causes ungrammaticality: (9)
* Kim snores bothers Sandy more and more.
This ungrammaticality is readily explained if one assumes that it is the complementizer which introduces the index, since subjects need an index in order to be coindexed with the nite verb. However, if it is the complementizer which introduces the [3rd, sing] index, then it cannot be treated as a marker, as proposed in Pollard & Sag (1994, 44-46), for if it is a marker, then the CONTENT value of the [COMP + S] combination is identical to the one of the clause without complementizer. As a consequence, since the latter is of type psoa (= parametrized state of a airs), the former will be of type psoa as well and hence lack an index. The only way to allow the complementizer to introduce the index is to treat it as the semantic head. Given the classi cation of headed structures in HPSG, this implies that it has to be either a head or an adjunct. If it is a head, it determines both the CATEGORY and the CONTENT values of the combination, giving rise to a complementizer phrase (CP), as in GB, cf. Chomsky (1986). A problem with this treatment, though, is that it does not account for the contrast in (10) I demand that he leave/*leaves at once. As pointed out in Pollard & Sag (1994, 44), the choice of the VFORM value of the subclause is determined by the matrix verb, and in order to model this the VFORM value of the that clause should be present in the latter's SYNSEM value, but if the complementizer is the syntactic head, it won't, since complementizers do not have VFORM values. If the complementizer is treated as an adjunct, this problem does not arise, for in that case the CATEGORY value of the combination is determined by the verbal projection. Adopting the adjunct analysis, the AVM of the complementizer can {provisionally{ be de ned as follows h i 2 3 1 psoa CAT j HEAD j MOD S finite : 66 2 h i3777 66 sing 77 64INDEX n3rd, 64CONTENT o 575 npro
RESTR
1
In words, that combines with a nite1 clause of type psoa and has a CONTENT value of type nominal-object, more speci cally of the subtype nonpronominal (= npro); the latter's index is [3rd, sing] and its set of restrictions includes the 1
I assume that ` nite' covers both the indicative and the subjunctive.
3
psoa of the clause. In order to distinguish this index from the ones of (pro)nouns I will slightly complicate the sort hierarchy of indices: index thematic referential
expletive
comp
The thematic indices are the ones which can take argument positions and comprise next to the usual referential ones, the indices of type comp, i.e. the ones which are introduced by complementizers; in contrast to the former they are invariably of the third person singular. An example of a non-thematic or expletive index is the one of existential there. Since the CONTENT value of the [COMP + S] combination is token-identical to the one of the adjunct, the addition of the complementizer has the e ect of turning a state of a airs into a nominal object2. The resulting object is of a hybrid nature: while verbal from a syntactic point of view, it is nominal from a semantic point of view. In that respect, there is a resemblance with the nonpredicative PPs which HPSG treats as prepositional from a syntactic point of view but as nominal from a semantic point of view. Because of this change in the treatment of that clauses we need a similar change in the AVMs of those verbs which take a that clause as their subject. Instead of de ning the value of their clausal argument as an object of type psoa, it has to be de ned as an index, as in the case of NP arguments. As applied to bothers, this gives the following result: h 2 3 i S finite , NP 1 [3rd;sing] 2 7 66CAT j SUBCAT 2 3 77 66 bother7 77 66 66RELATION 7 64CONTENT 4BOTHERING 1 comp75 75 psoa
BOTHERED
2
ref
In this analysis the presence of a complementizer in the subclause is not signalled by the syntactic feature [comp], but by the type of the index. At this point, though, we have a problem, for since the CONTENT value of the verb and its projection only contain the index of the subclause, its psoa value will get lost. In order to repair this I will push the analogy between that clauses and nonpronominal NPs one step further and assume that they behave like stored quanti ers. In order to model this I will enrich the AVM of the complementizer with a QSTORE feature, as in Because of the change of CONTENT value, the complementizer cannot be applied iteratively: that cannot be combined with a that clause, since the latter does not match its CATjHEADjMOD value. 2
4
2 2 33 h i 1 : psoa 66 66CAT j HEAD j MOD S finite 7777 3 h i 2 66 6 7777 INDEX 3rd, sing 66SYNSEM j LOCAL 66 757777 6 2 4 64CONTENT ncompo 66 577 RESTR 1 66 77 npro 66 77 8 " #9 < 66 77 DET the = 4QSTORE: : 5 RESTIND 2 ; quant
The RESTIND value of the element in the QSTORE is token-identical to the CONTENT value and hence contains the psoa value of the clause. The DET value is the same as for demonstrative and possessive determiners, i.e. the definite the. With the assignment of a QSTORE value to the complementizer's AVM, the latter shows some resemblance to the AVMs of determiners. This may seem puzzling to those who are accustomed to think of determiners and complementizers as totally di erent types of signs, but in this case the similarity is actually quite obvious. Notice, for instance, that the complementizer is homonymous to a demonstrative determiner, and that the nominal projections with which the latter combines, as in that small window, have to be [3rd, sing] as well. As a matter of fact, it has been argued in Jespersen (1927, 32) that the complementizer diachronically derives from the demonstrative. Since the nominalizing nature of the complementizer is now wired into its lexical entry, it is not only subject clauses which are treated as nominal objects but all that clauses. This might seem awkward in the case of verbs like say and think, whose clausal complement occurs both with and without complementizer (11) He said (that) he would like to see you. and whose SUBCAT list would hence contain a disjunctive statement if it made reference to the CONTENT value of the object. Notice, though, that the latter is by no means necessary: in order to prevent overgeneration it suces to require that the object is a saturated verbal projection, and this is a property which applies to both S and S[that]. 3
2 Extraposed that clauses The previous section focused on the analysis of sentences with a that clause in subject position. This one focuses on the sentences with an extraposed that clause. I will rst argue that extraposed clauses had better not be analyzed as complements and then use the dislocation based approach of Keller (1995) 3 In a previous version of the paper the treatment of that clauses as nominal objects was extended to for-to clauses and to VPs. For lack of space this extension could not be preserved in the nal version.
5
and Muller (n.d.) as the starting point for a treatment of extraposed clauses as dislocated appositions. The latter also incorporates some ideas from Kathol (1995).
2.1 Extraposition as a form of complementation In Pollard & Sag (1994) extraposed clauses are treated as complements. This correctly predicts that their presence is obligatory and that they have to follow the lexical head (12)
* It seems to me. * It that Sandy snores bothers Kim.
At the same time, though, this treatment does not square well with the rest of the grammar. One problem concerns the interaction with VP adjuncts. In English, the standard position of such adjuncts is before or after the VP, but not in between the verb and its complements (13) He often [takes a shower] in summer. * He takes often a shower in summer. * He often takes in summer a shower. This constraint is made explicit in Pollard & Sag (1994) by the requirement that the head sister of an adjunct be phrasal, which means that the head has to be saturated with respect to its complements before it takes any adjuncts: Head-Adjunct Schema: XP ! Y" [MOD [1]] , [1] XP ADJUNCT HEAD Head-Complement Schema: XP ! [2] , X0 [COMPS [2]] COMPS HEAD As a consequence, if extraposed clauses were complements, one would expect the VP adjuncts to occur either before the verb or after the subclause, as in (14) It very much [bothers Kim that Sandy snores]. We very much [regret it that he has been red]. (15) * It [bothers Kim that Sandy snores] very much. * We [regret it that he has been red] very much. but the post-VP position is ungrammatical and the pre-VP position is rather marked in comparison with the more commonly used (16) It bothers Kim very much that Sandy snores. We regret it very much that he has been red. A possible way out in this case is to allow for the interleaving of complements and adjuncts in the VP, but if that is done one will need special measures to rule out the ungrammatical 6
(17)
* He takes often a shower in summer. * It bothers very much Kim that Sandy snores.
A second problem for the complement analysis is exempli ed by (18) It started getting clear last year that he was not the right man for the job. She kept regretting it for years that she had not turned him down. What distinguishes these sentences from the ones above is that the adjuncts do not apply to the VPs which are headed by getting clear or regretting it, but rather to the higher VPs started getting clear and kept regretting it. As a consequence, if one takes word order seriously, the that clauses should be sisters of the higher VP as well, as in (19) It [[[started getting clear] last year] that he was not the right man for the job]. She [[[kept regretting it] for years] that she had not turned him down]. In order to make sure then that the extraposed clause is recognized as a complement of getting clear or regretting it one has to assume that the SUBCAT lists of kept and started inherit all elements which are still present in the SUBCAT list of their verbal complement, i.e. not only the one which corresponds to the subject, but also the ones which correspond to those complements for which the embedded verb is not yet saturated. This type of structure sharing, which is known as argument composition or generalized raising, has been described in detail in a number of articles by Hinrichs and Nakazawa, but the phenomena for which these authors propose the generalized raising device, such as verb clustering and auxiliary ip in German, do not occur in English. Furthermore, if English raising verbs were allowed to inherit the non-subject complements of their VP complement, special measures would have to be taken to exclude sentences like (20)
* He started wearing last year a hat.
A third problem with the complement analysis is that it makes the wrong predictions with respect to extractability. Whereas object clauses can be topicalized, as predicted by the complement extraction lexical rule in Pollard & Sag (1994, 378), the extraposed clauses cannot (21) That Kim would lose from Jack, nobody had expected. * That Sandy snores, it bothers Kim more and more. * That Sandy snores, Kim resents it more and more. A similar contrast holds for the extractability of the subject; whereas it is possible to extract the subject from a complement clause, as in (22) We think John has arrived. 7
Who do you think [ has arrived] ? such extraction causes ungrammaticality in the case of extraposed clauses: (23) It is a pity he doesn't know Russian. * Who is it a pity [ doesn't know Russian] ? A fourth piece of evidence against the complement analysis will be given at the end of section 3.
2.2 Extraposition as a form of dislocation The treatment of extraposed clauses as complements in Pollard & Sag (1994) is a consequence of the assumption that anticipatory it is an expletive. For, if the pronoun cannot be used as an argument, the only plausible supplier of the second argument of the matrix verb is the extraposed clause, and since the assignment of semantic roles is done lexically by structure-sharing between elements on the SUBCAT list and roles in the predicate-argument structure, the extraposed clause has to be on the SUBCAT list. It should be noticed, though, that the treatment of anticipatory it as an expletive is far from universally accepted. As a matter of fact, it is explicitly rejected in the GB analyses of Bennis (1986) and Koster (1987, 262) and in the HPSG analyses of Kathol (1995), Keller (1995) and Muller (n.d.). In the latter, for instance, the pronoun is treated as an argument and the extraposed clause as a dislocated constituent. The relation between them, which Pollard & Sag (1994) models in the lexical rule, is modeled in terms of a nonlocal dependency, which originates in the lexical entry of the pronoun4 2 2 33 h i CATEGORY NP nom 66 7 66 2 h i377777 66 66 66LOCAL 6CONTENT 64INDEX nref o3rd, sing 75777777 6 66 4 577 RESTR 1 66 77 npro h 64 75 i NONLOCAL j INHER j EXTRA S finite,that :
1
The pronoun has a referential index and a set of restrictions which includes the CONTENT value of the extraposed that clause. The latter's presence is enforced by the NONLOCALjINHERjEXTRA value. Following the Nonlocal Feature Principle, the EXTRA list is propagated up the tree and discharged at the point where the extraposed that clause is added. This analysis avoids the problems with the complement analysis, but it also raises some new ones. Notice, for instance, that the CONTENT value of the pronoun is of type nonpronominal: this is due to the fact that its restriction set This is a transposition into English of the AVM which Muller (n.d., 200) proposes for the German anticipatory es. 4
8
includes the CONTENT value of the extraposed clause, so that it is not empty as in the case of normal pronouns, but this is of course not in line with the fact that it behaves like a pronoun in all other respects. Another problem concerns the index of the pronoun: because of its argument status the index is referential, but this does not mesh well with Principle A of the Binding Theory. That principle states that a locally o-commanded anaphor must be locally o-bound, cf. Pollard & Sag (1994, 254), and hence seems to be violated in (24) Theyi made sure that it was clear to each otheri that this needed to be done. Theyi made sure that it was clear to themselvesi that this needed to be done. where the anaphors are not locally o-bound, since their SYNSEM values are not on the same SUBCAT list as the ones of their antecedents. In order to account for the fact that these sentences are nonetheless grammatical Pollard & Sag (1994, 258-262) argues that these anaphors do not need to be locally o-bound, since they are not locally o-commanded, and this they attribute to the fact that non-referential NPs like the anticipatory it cannot be (local) o-commanders. As a consequence, if it is treated as referential, one needs another way to account for the grammaticality of the above examples.
2.3 Extraposition as a form of apposition As in the analysis of Muller, I treat the pronoun as the origin of a nonlocal dependency, but instead of de ning the link with the extraposed clause in terms of CONTENT values, I de ne it in terms of contextual anchoring5 , as in 2 LOCAL 66 66 66 66 66NONLOC j INHER j EXTRA 66 64
3 77 82 39 77 > > CATEGORY XP > > 77 6 > 7 > 66CONT j INDEX 2 comp > 77> > > > > =777>77 > 6 > 75 7> 64ARG1 1 77 64CONTEXT > > 5 > > > > > > ; ;5 : ARG2 2 : NP:ppro
1 [3rd;sing]
In words, the pronoun anticipates the presence of an extraposed phrase with an index of type comp and provides an anchor for this index. Being part of the NONLOCALjEXTRA value the anchoring statement is propagated up the tree. The proposal to use the CONTEXT value for relating the pronoun to the subclause is also present in Kathol (1995, 290); in contrast to Kathol, though, I do not treat the subclause as a complement. 5
9
In contrast to the members of the leftward looking SLASH feature, which can be propagated across clause boundaries, the members of the rightward looking EXTRA feature cannot cross any clause boundaries. Compare (25) Which man [did you say [she claimed [she had seen ]]] ? * That [[it] didn't bother her anymore] is obvious [that her children snore]. This limitation, which was rst described in Ross (1967), is made explicit in (26) The Right Roof Constraint For any synsem object, if the LOCALjCATEGORY value is a saturated verbal projection, then the NONLOCALjINHERITEDjEXTRA value must be token-identical to the NONLOCALjTO-BINDjEXTRA value. >From this it follows that the distance which the EXTRA feature can span is rather small when the anticipatory pronoun appears in subject position, but if it appears in object position, the distance can be relatively large, as in (27) She [[kept [regretting it]] for years] that she had not turned him down. In Kathol (1995, 299) the existence of this constraint is treated as an argument against the nonlocal analysis of extrapostion, but I do not think that this is a valid objection, for while the distribution of the EXTRA feature is indeed more constrained than the one of the SLASH feature, it is at the same time less constrained than the one of the NONLOCALjREL feature, which is used for relating relative pronouns to their antecedents, as in (28) Here's the minister [[in [the middle [of [whose sermon]]]] the dog barked]. The propagation of this feature is restricted by the Clausal Rel Prohibition, i.e. the requirement that \the INHERjREL value of S must be empty" (Pollard & Sag (1994, 220)), and hence more limited than the one of the INHERjEXTRA feature. For the combination of the main clause with the extraposed subclause I will not make use of a separate Head-Extra Schema, as in Keller and Muller, but rather of a schema which is independently needed for the analysis of other phenomena, i.e. the parataxis schema. The need for such a schema is pointed out in the discussion of relative-correlative constructions in Pollard & Sag (1994, 228), but its properties are not spelled out and the only application concerns the formation of relative clauses in Hindi and Marathi. This paper will not propose a full- edged treatment of parataxis either, but it will discuss some salient properties of one type of parataxis, i.e. apposition, and demonstrate that those properties {when slightly generalized{ also apply to the phenomenon of it extraposition. Some prototypical examples of apposition are the [NP + NP] combinations in (29) He lives in Monrovia, the capital of Liberia. 10
We paid a visit to the castle, a splendid example of Bavarian rococo. That such structures are headed is clear from an example like the Italian (30) La Divina Commedia, il capolavoro di Dante, e stata scritta nel 1300. The Div. Comm., the masterpiece of Dante, is been written in-the 1300. The Div. Comm., Dante's masterpiece, was written in the 14th century. in which the past participles agree in number and gender with the rst NP, which is feminine, and not with the second one, which is masculine. What distinguishes appositions from adnominal adjuncts is that their CONTENT value is not of type psoa but rather of type nominal-object 6 . From this it follows that they have an index of their own, so that their PERSON, GENDER and NUMBER values are not necessarily identical to the ones of their antecedent, as in the Italian example above, or as in the English one below (31) They managed to neutralize his brains, his most precious possession. in which the antecedent NP is plural, and the apposed one singular. In spite of this lack of index agreement, though, the two NPs do refer to the same entity; in other words, they are not co-indexed, but they are co-referential. A straightforward way of making this explicit is by requiring that the index of the apposed NP be anchored to the one of its antecedent. Another characteristic of apposed NPs is that they combine with fully saturated NPs and that they invariably follow their head. From this it follows that they are right-peripheral in the NP and hence possible candidates for being extraposed7 (32) They spoke to Scalfaro yesterday, the President of the Republic. In such sentences the apposed NP is combined with an S. Interestingly, all of these properties also apply to the cases of it extraposition. First, in the [S + S'] combination it is the S which functions as the head; this is clear from the fact that the combination can be preceded by a complementizer as in (33) ... that [it bothers Kim that Sandy snores]. This would be impossible if the extraposed clause were the head, since complementizers cannot be combined with S'. Second, the extraposed clause is a nonpronominal object, whose index is anchored to the one of the anticipatory it. Third, the extraposed clauses combine with a fully saturated projection and are invariably preceded by that projection. The common properties are spelled out in the following AVM: More speci cally, they are of type nonpronominal, since pronouns cannot be used as appositions. From this it follows that appositions have to be o-free, cf. Principle C of the Binding Theory. This requirement is indeed ful lled, since appositions are not coindexed with their antecedents. 7 On the relation between right-peripheral occurrence and extraposability, see Stucky (1987) and Kathol (1995, 300-302). 6
11
2 3 HEAD-DTR j SYNSEM j LOCAL j CAT j SUBCAT h i 66 2 2 3377 66 7 CAT XP 66 66 777777 # " 66 66 66CONT 3 INDEX 2 777777 66 6 6 777777 RESTR fg 66 66 66 7 npro 66 82 397777777 66SYNSEM j LOC 6 6 7 > 66 > > 6 66 =777777 66APP-DTR 66 6 7 64CONX >4ARG1 1 5>75777777 66 66 : ARG2 2 ; 77 66 66 7777 9 8" 66 # 66 7777 = < DET 64 64 575 QSTORE : : RESTIND 3 ;
head-app-str
In a structure of type head-apposition, the head daughter is a fully saturated phrase and the apposed daughter is a phrase of type nonpronominal whose index is anchored by another (contextually determined) index and whose CONTENT value is put in storage. The latter brings out another di erence with adjuncts: while adjuncts incorporate the CONTENT value of their head sister, adding their own contribution and passing this on to the mother node, the semantic contribution of an apposition is stored and retrieved by means of the standard quanti er scoping devices. I do not want to suggest that this schema covers all types of parataxis, nor do I want to suggest that it covers all cases of extraposition, but it neatly subsumes the cases of it extraposition and has the advantage of being independently needed for the treatment of NP apposition. As compared to the other treatments of it extraposition, this analysis does not only avoid the problems with the complement analysis, but also the ones which originate from the decision to treat it as an argument. This is due to the fact that the sort hierarchy of indices makes a distinction between thematic and referential indices, so that the index of anticipatory it can be thematic without being referential: index thematic referential
it-anti
expletive comp
This guarantees proper interaction with Principle A of the Binding Theory. Apart from avoiding the problems with the other analyses the apposition treatment also clears the way for an elimination of the it extraposition lexical rule. This is the topic of the last section.
12
3 Underspeci cation Having discussed both the non-extraposed and the extraposed variants of sentences with clausal subjects, I now return to the device which is used to relate them, i.e. the lexical rule. In the formulation of Pollard & Sag (1994, 150) it \ removes an S[comp] from a SUBCAT list, replacing it by NPit , and appends the S[comp] to the end of the SUBCAT list, preserving role assignment." The problem with this rule is that it is both too general and too limited. It is too general, since it is not prevented from applying to its own output and hence from producing SUBCAT lists with more than one NPit , and it is too restricted, in the sense that it misses the generalization that predicates which are subcategorized for a that clause cannot only take it instead, but any kind of referential NP: (34) That Kim snores bothers Sandy. It bothers Sandy that Kim snores. This y bothers Sandy. Notice that this is not just a lexical property of bother, but a property of all predicates which take subject clauses. Both of these problems are avoided in the present treatment. The former does not rise since the object in the NONLOCALjINHERjEXTRA value of anticipatory it is required to have an index of another type than the one of the pronoun itself, and the latter can be solved by making judicious use of underspeci cation. In the case of bothers, for instance, it suces to relax the constraint on the speech part value of the rst element in the SUBCAT list and to require that its index be thematic, as in 2 3 XP , NP 1 [3rd;sing] 2 66CAT j SUBCAT 7 2 377 66 7 bother 66 66RELATION 7777 64CONTENT 4BOTHERING 1 thematic 575 psoa
BOTHERED
2
referential
De ned like this, the BOTHERING role can be supplied by either a referential NP or an S[comp] or the anticipatory it ; in the latter case, the presence of an extraposed S[comp] is not required by the verb, but by the interaction of the pronoun's AVM with the Right Roof Constraint. The underspeci cation technique can also be used for verbs like regret, which allow for the same range of possibilities in the object position: (35) We all regret that you did not come. He regrets it very much that you can't come. 13
I still regret my rudeness. As for verbs which require extraposition, such as seems and appears, it suces to put tighter constraints on the index of their rst argument: 3 2 h i XP , PP to CAT j SUBCAT 1 [3rd;sing] 2 7 66 2 377 66 7 66 66RELATION seem 7777 1 it-anti 64CONTENT 4ARG1 575 psoa
ARG2
2
referential
Because of this constraint the only possible subject is anticipatory it. This does not only exclude subject clauses but also referential NPs, and thus correctly prevents that the subject in John seems to be gaining weight is identi ed with the rst argument of the verb. Apart from being more accurate from an empirical point of view, the underspeci cation approach is also more attractive from a methodological point of view, since it does not necessitate the postulation of separate lexical entries for the di erent uses of the same verb. Instead of having a rather speci c lexical entry for one use and deriving another equally speci c entry for the other use, there is just one entry which generalizes over both. For computational purposes, this is a welcome result since it avoids the proliferation of lexical entries with the same phonological or orthographic form. 8 Another advantage of the underspeci cation approach is that it further minimizes the use of derivational devices in the grammar. In contrast to the lexical rule approach which literally derives lexical entries from other more basic lexical entries, the underspeci cation approach does not make any distinction between basic and derived. In this sense it can be seen as the next step in a chain which led from the elimination of transformations via the elimination of metarules to the elimination of this last vestige of derivationalism in HPSG, i.e. the lexical rules. That such elimination would not only be desirable from a methodological point of view but also from a formal one is clear from the remark in Pollard & Sag (1994, 395) that \we lack as yet any satisfactory declarative formalization. The fundamental diculty here is that lexical rules must be seen as implicative relationships between lexical entries; but lexical entries themselves are constraints on feature structures (not feature structures themselves), so evidently a higher-order formalism must be developed within which such relationships can be expressed." In other words, also from a formal point of view, the lexical rules are the odd man out in an otherwise declarative framework. In view of these considerations I see it as an extra argument in favour of the present treatment that it avoids the use of an it extraposition lexical rule. As Gosse Bouma pointed out to me, lexical rules do not lead to a proliferation of homographs if they are applied at run time and hence treated as nonbranching phrase structure rules. Notice, though, that the proliferation problem is then shifted to the grammar, where it resurfaces as the problem of avoiding a proliferation of nonbranching expansions. 8
14
Indirectly, this provides another argument against the complement treatment of extraposed clauses, for if the SYNSEM values of the extraposed clauses gure on the SUBCAT list of verbs, the only plausible way to exploit the possibilities of underspeci cation is to make the presence of it optional, as in , but this erroneously admits sentences like (36)
* Bothers Kim that Sandy snores.
In other words, the complement analysis does not square well with the underspeci cation approach. Hence, to the extent that the latter is preferable to the lexical rule approach, the complement analysis is less felicitous than the alternative apposition based analysis.
4 Conclusion This paper has provided a monostratal treatment of it extraposition which is based on underspeci cation. The starting point of the discussion was the treatment in Pollard & Sag (1994, 145-155). This treatment has been shown to meet a number of problems: rst it does not account for the agreement between nite verbs and that clauses in subject position, second its complement analysis of extraposed clauses does not square well with the rest of the grammar, and third the lexical rule which is used to derive entries for verbs with extraposed complements is at the same time too general and too limited. As an alternative it has been proposed to treat that clauses as verbal projections with a nominal CONTENT value, to analyse extraposed clauses as dislocated appositions and to capture the e ect of the it extraposition lexical rule in terms of the computationally more attractive method of underspeci cation.
References Bennis, H. (1986), Gaps and dummies, Foris, Dordrecht. Chomsky, N. (1986), Barriers, MIT Press, Cambridge (Mass.). Jespersen, O. (1927), A Modern English grammar on historical principles, Vol. III, Allen Unwin, London. Kathol, A. (1995), Linearization-based German Syntax, PhD thesis, Ohio State University. Keller, F. (1995), Towards an Account of Extraposition in HPSG, in `Proceedings of the 7th Conference of the European Chapter of the Association for Computational Linguistics', Dublin. Koster, J. (1987), Domains and Dynasties, Foris Publications, Dordrecht/Providence. 15
Muller, S. (n.d.), Head-driven Phrase Structure Grammar fur das Deutsche, Humboldt-Universitat zu Berlin. Pollard, C. & Sag, I. (1994), Head-driven Phrase Structure Grammar, CSLI and University of Chicago Press, Stanford/Chicago. Ross, J. (1967), Constraints on Variables in Syntax, PhD thesis, MIT. Stucky, S. (1987), Con gurational Variation in English: a Study of Extraposition and Related Matters, in `Discontinuous Constituency', number 20 in `Syntax and Semantics', Academic Press, New York, pp. 377{404.
16
