+ All Categories
Home > Documents > Arnold&Sadler 1992

Arnold&Sadler 1992

Date post: 04-Jun-2018
Category:
Upload: albano-dalla-pria
View: 217 times
Download: 0 times
Share this document with a friend
32
8/13/2019 Arnold&Sadler 1992 http://slidepdf.com/reader/full/arnoldsadler-1992 1/32 NOUN-MODIFYING ADJECTIVES IN HPSG Doug Arnold Louisa Sadler Department of Language and Linguistics, University of Essex, Wivenhoe Park, Colchester, CO4 3SQ, UK louisa @ essex.ac.uk doug @ essex.ac.uk ABSTRACT This paper proposes an analysis of adjectival constructions in HPSG, focussing on the behaviour of adjectives in NP, and develops an account of the familiar observation that adjectives without complements typically precede the noun, while those with comple- ments or post-modiers typically follow the noun. 1. INTRODUCTION Adjectives in English appear in a number of different contexts, including the following: (1) a. a tall person pre-nominal b. a person fond of children post-nominal c. a hard to pronounce name ‘complex As’ d. a hard problem to solve ‘hard nuts’ e. somebody interesting indenite pronouns f. too good a chance to miss ‘big messes’ (2) a. Sam is guilty. subject complement b. They consider Sam guilty. object complement/small-clause c. They hammered the metal at. resultative d. Sam drinks beer cold. depictive e. Drunk, Sam came in, and fell over. free adjunct f. With Sam drunk, the problem was compounded. absolutive (1) are sometimes (misleadingly) called ‘attributive’ uses of adjectives, those in (2) ‘predicative’. The difference is that while those in (1) are noun-modifying uses within NP, those in (2) have adjectives in positions governed (loosely speaking) by verbs. (2a) is usually taken as canonical for predicative adjec- tives, because it is not subject to the restrictions characteristic of the other contexts. Similarly, (1a) is the ‘typical’ position for adjectives in NP. ‘Central’ or typical adjectives characteristically appear in both of these contexts. The names for the constructions in (1c,f) are from Berman (1974) April 25, 1992
Transcript
Page 1: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 1/32

NOUN-MODIFYING ADJECTIVES IN HPSG

Doug Arnold Louisa Sadler

Department of Language and Linguistics,University of Essex,

Wivenhoe Park,Colchester, CO4 3SQ, UK

louisa @ essex.ac.uk doug @ essex.ac.uk

ABSTRACT

This paper proposes an analysis of adjectival constructions in HPSG, focussing on the

behaviour of adjectives in NP, and develops an account of the familiar observation thatadjectives without complements typically precede the noun, while those with comple-ments or post-modiers typically follow the noun.

1. INTRODUCTION

Adjectives in English appear in a number of different contexts, including the following:

(1)a. a tall person pre-nominalb. a person fond of children post-nominal

c. a hard to pronounce name ‘complex As’d. a hard problem to solve ‘hard nuts’e. somebody interesting indenite pronounsf. too good a chance to miss ‘big messes’

(2)a. Sam is guilty. subject complementb. They consider Sam guilty. object complement/small-clausec. They hammered the metal at. resultatived. Sam drinks beer cold. depictivee. Drunk, Sam came in, and fell over. free adjunct

f. With Sam drunk, the problem was compounded. absolutive

(1) are sometimes (misleadingly) called ‘attributive’ uses of adjectives, those in (2) ‘predicative’. Thedifference is that while those in (1) are noun-modifying uses within NP, those in (2) have adjectives inpositions governed (loosely speaking) by verbs. (2a) is usually taken as canonical for predicative adjec-tives, because it is not subject to the restrictions characteristic of the other contexts. Similarly, (1a) is the‘typical’ position for adjectives in NP. ‘Central’ or typical adjectives characteristically appear in both of these contexts. The names for the constructions in (1c,f) are from Berman (1974)

April 25, 1992

Page 2: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 2/32

2 HPSG/ADJ

In the rst part of this paper, we will describe the syntax and semantics of these ‘central’ adjectives,using the theoretical framework of Head-Driven Phrase Structure Grammar (HPSG: Pollard and Sag,1987, 1991; henceforth ISS-1, and ISS-2, respectively), and with the analysis outlined in ISS-2, Ch8 as astarting point. This account will be extended to deal with the other contexts in (1b-d), in particular wewill focus on the contrast between (1a,b), and try to account for the familiar observation that adjectives

without complements typically precede the head noun in NP, those with complements follow the noun:

(3)a. A tall mana’. *A man tallb. A man fond of childrenb.’ *A fond of children man

This phenomenon turns out to be rather difcult to account for. No satisfactory account exists, and someproposals have involved rather extreme theoretical modications. For example, Hendrick (1978) andNanni (1978,1980) proposed that complements of A are not part of AP at all underlyingly (in violationof all accepted approaches to subcategorisation), and Arnold (1989) suggested that the normal X

rela-

tionship of speciers, heads, and complements should be inverted in the case of As (the idea being thatcomplements of A would be excluded from A

, and prenominal position would be a submaximal posi-

tion).

Among the less radical proposals are Hudson’s (1984) suggestion that prenominal adjectives form com-pounds with their head Ns, Emonds’s (1976:19) restriction on surface recursion (which prohibits pre-head modiers from branching into heads and complements), and Williams’ (1982) Head Final Filter.These are all subject to empirical difculties (Arnold, 1989:204ff).

One of the most interesting accounts to date is Flynn’s (1983) Categorial Grammar account, whichderives these facts from a general word order convention, which predicts whether a functor category willprecede or follow its arguments on the basis of whether it contains an expression such as an NP, S, or V

(i.e. an expression whose category ‘results in S’). This entails that lexical heads and speciers precedetheir complements, as do adjectives without complements, but that APs that contain NPs, such as fond of children follow the nouns they modify (which they take as arguments: attributive or noun-modifying Asand APs are N/Ns in Categorial Grammar). Again, there are empirical difculties (e.g. with the treat-ment of complementisers, and coordination), as well as the normal objection to Categorial Grammaraccounts of adjectives, which is that they fail to recognize a unied category of adjectives at all (makingattributive and predicative adjectives appear totally unrelated).

The account we will explore here tries to overcome some of the drawbacks of these alternatives. Itinvolves a reformulation of the ID rule for combining heads and adjuncts in ISS-2 and uses a new LPstatement (independently required for compounds). We discuss in some detail the lexical properties of different sorts of adjectives and formulate a lexical redundancy rule for relating predicative and noun-

modifying adjectives. This is very much a rst draft, and at a number of points we note or sketch outalternatives.

2. THE HPSG FRAMEWORK

April 25, 1992

Page 3: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 3/32

HPSG/ADJ 3

2.1. Introduction

In this section, we summarize the theoretical framework, and describe the treatment of adjectives withinHPSG outlined in ISS-2, Ch8.

In HPSG, the basic units of linguistic description are typed feature structures interpreted as providingsimultaneous descriptions of different levels of linguistic organisation (i.e. they are signs ). The typesigns has two subtypes - lexical sign which has the attributes PHON, SYNSEM and QSTORE and

phrasal sign with the additional attribute DTRS. DTRS encodes the ‘parse tree’ of a phrasal sign,depending on its type it has attributes ADJUNCT, HEAD, COMPLEMENT, MARKER, or FILLERDAUGHTERS (COORD-DTR will presumably also be needed). Different Immediate Dominance (ID)schemata are used to admit local trees of different types. Principles governing the sharing of structurebetween mothers and daughters make crucial reference to the different types of daughter: e.g. the HEADFEATURE principle (HFP) requires that the values of the HEAD attribute be the same on mother andhead daughter (ie the sign which is the value of the HEAD-DTR attribute), the SUBCAT principlegoverns the relation between SUBCAT information on the HEAD-DTR and the mother, the SemanticsPrinciple species the sharing of either NUCLEUS or CONTENT between the semantic head and themother (the semantic head is the adjunct daughter in a head-adjunct structure, and the head daughterelsewhere). The value of NUCLEUS is shared if the CONTENT of the semantic head is of type soa ,and otherwise the CONTENT itself is shared (the difference is simply whether or not the QUANTSvalues are identied).

In this paper, the most important principles and ID-schemata will be the Semantics Principle, and SUB-CAT principles, and the ID-Schemata described below.

2.2. ID Schemata

(4) ID1 : A well-formed phrase (wfp) may consist of a saturated ([SUBCAT <> ]) phrase with a DTRSvalue of sort head-comp-struc in which the HEAD-DTR is a phrasal sign and the COMP-DTR value is a

list of length one.(5) S <>

→ NP, VP < NP >

(6) ID2 : A wfp may consist of an almost saturated (SUBCAT list of length one) phrase with DTRSvalue of sort head-comp-struc in which the HEAD-DTR value is a lexical sign.(7) VP < NP >

→ V< NP,NP,PP > , NP, PP

(8) ID3 : A wfp may consist of a saturated ([SUBCAT <> ]) phrase with DTRS value of sort head-comp-struc in which the HEAD-DTR value is a lexical sign. (In English, this is used only for SUBJ-AUX inversion structures, i.e. where the lexical head is also [INV +], and will play no further part inour discussion).(9) S <> → AUX, NP, VP < NP >

(10) ID5 : A wfp may consist of a phrase with DTRS value of sort head-adj-struc , such that the MODvalue of the adjunct daughter is token identical with the SYNSEM value of the head daughter.(11) e.g. X → X, X −ADJUNCT

(There are two other ID-Schemata: ID4, dealing with head-marker-structures , such as S

→ Comp S andID6, dealing with head-ller-structures , such as S

→ XP S).

April 25, 1992

Page 4: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 4/32

4 HPSG/ADJ

Once the relevant principles are taken into account, Schemata ID1, ID2, and ID5 license structures likethe following:(12) ID1 : Saturated HEAD-COMP-STRUCTURES:(13)

phrasal −sign

DTRS head −comp −struc COMP −DTRS < phrasal −sign [1] >

HEAD −DTR phrasal −sign SYNSEM LOC SUBCAT < [1] > SYNSEM LOC SUBCAT <>

(14)

HEAD [1]

SUBCAT <>

----------------------

| |

|H |C

| |

| |HEAD [1] [2]

SUBCAT <[2]>

(15) ID2 : Almost Saturated HEAD-COMP-STRUCTURES:

phrasal −sign

DTRS head −comp −struc COMP −DTRS < phrasal −sign [2], phrasal −sign [3], ..., phrasal −sign [n] >

HEAD −DTR lexial −sign SYNSEM LOC SUBCAT < [2], [3], ..., [n] >

SYNSEM LOC SUBCAT < [2]>

HEAD [1]

(16)

HEAD [1]

SUBCAT <[2]>

-----------------------------------

| | |

|H |C |C

| | |

| | |

HEAD [1] [3] .... [n]

SUBCAT <[2],[3],...[n]>

(17) ID5 : HEAD-ADJUNCT STRUCTURES: 1

__________________

1 We simplify: if the CONTENT of the ADJUNCT, or more generally the semantic head, is of sort psoa , then theNUCLEUSes of the mother, and the ADJUNCT daughter are token identical. Otherwise, their entire CONTENTS areidentical, as shown here.

April 25, 1992

Page 5: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 5/32

HPSG/ADJ 5

phrasal −sign

DTRS head −ad j −struc

ADJ −DTRS < SYNSEM LOC CONTENT [1]CAT HEAD MOD [3]

>

HEAD −DTR SYNSEM LOC SUBCAT [4]HEAD [2]

SYNSEM [3] LOC CONTENT [1]

CAT SUBCAT [4]HEAD [2]

(18)

HEAD [1]

CONTENT [2]

SUBCAT [4]

-------------------------------------

| |

|H |ADJ

| |

| |

SYNSEM [3] MOD [3]

SYNSEM|LOC|CAT|HEAD [1] CONTENT [2]

SUBCAT [4]

It should be noted that the treatment of adjuncts assumed in ISS-2 (and here) is different from that inISS-1. In ISS-1 heads were assumed to select adjuncts in something like the way they select comple-

ments (a set-valued feature ADJUNCTS was assumed, for an adjunct to occur in a construction, it had tounify with one of the members of this set). However, as discussed in ISS-1:163ff, this approach raised aserious formal problem, since it involved a relational (non-functional) dependency between twoattribute-values). Here it is the adjunct that selects its head, via the value of the MOD feature. Theadjunct also supplies the CONTENT for the whole phrase, having incorporated the CONTENT of thehead.

2.3. Linear Precedence

Ordering relations between constituents are stated in terms of a collection of Linear Precedence (LP)statements. These are intended to be carried over from ISS-1 (by assuming a further LOCAL featureLEX: see ISS-2:Ch1,p15, note 16). The important ones are the following:(19) LP1 : HEAD-DTR [ LEX: +] < []

That is, lexical heads precede everything else.(20) LP2 : COMPLEMENT << COMPLEMENT [LEX - ] (2nd formulation) 2

That is, less oblique complements precede more oblique (phrasal) ones (the [LEX-], i.e. phrasal, stipula-tion is to allow for alternative orderings of particles).__________________

2 The nal formulation generalises the obliqueness hierarchy (to include adjuncts and heads) and accounts forordering between sister adjuncts and complements: COMPLEMENT[MAJ -V] < < [LEX -]

April 25, 1992

Page 6: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 6/32

6 HPSG/ADJ

Other LP statements allowing for other orders of, e.g. FILLERS, and FOCUSsed items, are also forseen.

Finally, we should note some abbreviations, and conventions:

X

stands for a non-lexical phrase with a SUBCAT list of length one. For example, N

denotes a nominalphrase that has found all its complements except the determiner (determiners are assumed to be sub-categorized), as in (21), V

denotes a verb that has found all its complements except the subject.

(21)

N

=def SUBCAT < DETP >

HEAD N

X:[i] is an abbreviation for an X with the CONTENT [i]. The INDEX can be mentioned as a subscript,for example:(22)

NP [1] =def LOC CONTENT INDEX [1]

CAT SUBCAT <>

HEAD

N

To make the reading of CONTENT parts of feature structures easier, we will sometimes abbreviate tosomething like normal FOPC. For example, (23) abbreviates the following structure:(23) [1] {athlete([1]),albanian([1]) }(24)

RESTR

NUCLEUSINST [1]

RELN athlete

QUANTS <>

NUCLEUSINST [1]

RELN albanian

QUANTS <>

INDEX [1]

The denition of the sorts phrasal-sign , and lexical-sign will also be important. The initial assumption isthat a phrasal sign is a sign for which a DTRS value is dened, and which is [LEX −]. A lexical sign isone for which no DTRS value can be dened, and which is [LEX +]. However, the suggestion wasmade in ISS-1 (p72-3) that the distinction between ±LEX does not coincide exactly with the distinctionbetween phrasal and lexical signs. In particular, it was suggested that there might be [LEX +] signswhich had internal constituent structure, and rules which introduced [LEX +] on the mother, to accountfor cases like (25), which can be part of a (noun-noun) compound, but has internal constituent structure.

(25)a. toxic wasteb. [[ toxic waste ] dump ]

Notice that these are head-nal constructions, even though the head is lexical. Their LP is not so farpredicted. Our proposal exploits this divergence between the LEX feature and the type (phrasal or lexi-cal) of the sign.

April 25, 1992

Page 7: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 7/32

HPSG/ADJ 7

3. ADJECTIVES IN HPSG

3.1. Adjectives in NP

ISS-2 briey discusses both prenominal attributive adjectives and adjectives in predicative position. Werst sketch out the analysis and then ll in some further details, pointing out some inadequacies andattempting to remedy them. As is reasonable, ISS-2 takes adjectives within NP as ADJUNCTS. Thusexamples like Albanian athlete fall under ID5.

Noun-modifying (‘attributive’) adjectives like Albanian are treated as adding restrictions to the anchor of a parameter (i.e. Albanian adds a further restriction to the parameter of the N it modies). TheSYNSEM LOCAL value for an ‘intersective’ adjective, such as Albanian used inside NP will be some-thing like the following (cf ISS-2:Ch8,p14):(26)

CONT re f −ob j

RESTR

NUCLEUS INST [1]RELN Albanian

QUANTS <>

[3]

INDEX [1]

CATSUBCAT

<>

HEAD ad jective MOD N : RESTR [3]

INDEX [1]

Abbreviating:(27)

CONT [1] Albanian([1]) [3]

CAT SUBCAT <>

HEAD ad jective MOD N : [1] [3]

That is, Albanian is an expression whose HEAD value is of sort adjective , and which modies an N .

This N

has an INDEX [1], with RESTRictions [3]. The interpretation of the Albanian N involves union-ing the restrictions on the INDEX of Albanian with those on the INDEX of the N.

In the case of an N

like athlete , [3] would be something like (28a), more fully (28b):(28) a. {athlete([1]) }

(28)b.

April 25, 1992

Page 8: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 8/32

8 HPSG/ADJ

NUCLEUS INST [1]RELN athlete

QUANTS <>

Thus, for Albanian athlete one would get an N

with CONTENT formed by unioning the restrictions thatcome from Albanian and athlete , giving a CONTENT like the following. Unioning the restrictions fromthe N and those from the A captures the idea that these adjectives are intersective — that somethingsatises Adj-Noun just in case it satises the restrictions of Adj, and those of Noun.(29)

CONT re f −ob j [1] Albanian([1]), athlete([1])

(30)

CONT re f −ob j

RESTR

NUCLEUS INST [1]RELN Albanian

QUANTS <>

NUCLEUS INST [1]RELN athlete

QUANTS <>

INDEX [1] (29)

This treatment extends easily to the other main classes of adjective. Measure adjectives ( red , good , tall ,etc.) can be analysed as making reference to a futher hidden parameter, whose value xes a scale andstandard against which measurement is made. This involves introducing a further argument for the

corresponding relations (called STANDARD), whose value is determined contextually.For ‘modal’ adjectives, such as alleged , instead of unioning the restrictions that come from the modiedN , one treats these restrictions as arguments of the adjective. An adjective like alleged corresponds to a

two place relation between an ‘alleger’, and an allegation — the latter being a state of affairs type objectsupplied by the N

restrictions. The difference between the entry for alleged , and that for Albanian is in

the NUCLEUS part of the content:(31)

NUCLEUSSOA −ARG [3]ALLEGER [4]RELN alleged

Thus, instead of the restrictions from the N

being unioned with those of the adjective, they are embed-

ded as an argument to the adjective. The effect is that alleged athlete will get an interpretation like:(32) [1] {alleged([4], {athlete([1]) })}.

Roughly, an alleged athlete gets interpreted as an entity x ([1]) such that someone ([4]) alleges that x isan athlete (alleges that [1] must be anchored to something which is an athlete).

April 25, 1992

Page 9: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 9/32

HPSG/ADJ 9

3.2. Adjectives in VP

Following ISS-2, we assume adjectives in ‘predicative’ (e.g. post-copula) position are COMPLE-MENTS of the verb ( be , seem , etc.), which typically subcategorizes for a [PRD +] complement.

In head complement structures, HEAD features are shared between head daughter and mother. TheNUCLEUS of the mother is shared with the HEAD daughter if it is of type psoa ; otherwise, the wholeCONTENT value is shared. The former occurs in constituents headed by a verb or predicative P, N, orA. The SUBCAT value of the mother is derived from that of the HEAD by removing the occurringcomplement daughter(s) from the SUBCAT list. The ID rules for such structures permit a lexicalHEAD to combine with all except one of the complements (ID2), and an almost saturated HEAD tocombine with its remaining complement (ID1). Predicative APs occur as complements to HEADSwhich subcategorise a [PRD +] complement ( be , appear , seem ). The AP subcategorises an NP, theframe for raising be (or the other raising verbs) equates this NP on the SUBCAT list of its [PRD +]complement with its own subject — this is full SYNSEM identity. The LOCAL part of the entry for beis :(33)

CONT [1]

CAT SUBCAT < [2]NP, XP PRD +, SUBCAT < [2]> :[1]

HEAD V+AUX

(32)

where [1] is a CONT value associated with the XP. The idea is that the copula subcategorizes for anNP subject, which is identical as regards SYNSEM values with the subject of its XP complement. TheCONTENT of the phrase headed by the copula is identical to the content of the XP. The copula is thustransparent, semantically. Since the CONTENT of declarative sentences is of type psoa , this means theCONTENT of the XP must be of type psoa , also. That is, in general, the CONTENT of predicativeadjectives must be of type psoa . Note that this is different from the CONTENT of noun-modifying

adjectives.The LOCAL value for typical post-copula adjectives is then along the following lines:(34)

CONT psoa

NUCLEUS INST [1]RELN Albanian

QUANTS <>

CAT SUBCAT < NP[1], ... >

HEAD ad jective PRD +

Abbreviating:(35)

CONT psoa Albanian([1])

CAT SUBCAT < NP[1], ... >

HEAD ad jective PRD +

April 25, 1992

Page 10: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 10/32

10 HPSG/ADJ

3.3. Discussion, and Lexical Redundancy Rules

In this section we discuss a problem with the account of adjectives just described, to do with theassumptions about the subcategorization of noun-modifying adjectives, and state a lexical redundancyrule for relating predicative and noun-modifying entries for adjectives.

3.3.1. Subcat for Noun-Modifying Adjectives

Most of the discussion of adjectival constructions in Pollard and Sag (ISS-2) concerns the attributive ornoun-modifying use of adjectives like red . These are taken to be saturated (i.e. [SUBCAT <> ])adjuncts. This does not give problems with interpretation because the MOD feature ensures that theCONTENT of the N

that is modied is correctly combined with the CONTENT of the adjective. In their

predicative usage (e.g. as complements of be ), such adjectives are monovalent, subcategorising a SUB-JECT. A consequence of this view, therefore, is that predicative adjectives and their attributive counter-parts differ in terms of SUBCAT. While this is an attractive view of noun-modifying adjectives (for areason we will come to later) we do not think it tenable (for the version of HPSG described in ISS-2) asa general treatment of all noun-modifying adjectives.

To see this, consider an adjective like fond , or more strikingly, grateful , or dependent , which takerespectively two and three complements, when used predicatively:

(36)a. Sam is fond of children.a’. a person fond [ of children ]b. Sam is grateful to everyone for anything they do.b’. a person grateful [ to everyone ] [ for anything they do ]c. Sam is dependent on the state for healthcare.c’. a person dependent [ on the state ] [ for healthcare ]

The structure associated with a predicative use of grateful to NP for NP , will be almost saturated. If

noun-modifying uses of adjectives are saturated, the entry associated with this use of grateful must differand the AP will be saturated (37b).

(37)a. grateful (predicative) SUBCAT < NP, PP to, PP for>

Sam [i] is [ A,SUBCAT < NP:[i] > grateful to everyone for everything ].b. grateful (noun-modifying) SUBCAT < PP to, PP for>

a person [ A,SUBCAT <> grateful to everyone for everything ]

There are two difculties here: (a) in the SUBCAT list, the PP to is indistinguishable from a SUBJECT(it is the rst element on the list) and (b) technically, noun-modifying grateful would not be able tocombine with its complements to produce grateful to everyone for everything . This is because the ID

rule (ID2) which combines lexical heads with their complements produces an almost saturated phrase;but here we require a lexical head to combine with complements to produce a saturated phrase. Noticethat ID1 (which does produce a saturated phrase) cannot be used to produce the desired effect: (a) itrequires a non-lexical head; and (b) it only allows one complement (here we have two). Recall also thatID3, which allows a lexical head to combine with all its complements, is only used in English for struc-tures involving subject-auxiliary inversion.

There are two alternatives which should be investigated. The rst is to recast this analysis in terms of the proposals for COMP and SUBJECT selection made by Borsley (which immediately eliminates therst problem and may also address the second). This looks a promising direction, but we leave it to one

April 25, 1992

Page 11: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 11/32

HPSG/ADJ 11

side for further investigation. The second alternative seems to be to assume that noun-modifying APsare almost-saturated, that is, noun-modifying adjectives have the same SUBCAT lists as their predicativecounterparts. This would give results like (38).(38) a person [ A,SUBCAT < NP:[1] > grateful to everyone for everything ]

Here the noun-modier grateful to everyone for everything is not saturated. This is not problematic (i.e.it does not violate any principles — unsaturated (actually almost saturated) signs also appear, in controland raising constructions. In particular, notice that the semantics of the adjective is complete, because of the way the semantics of the head noun is shared within the MOD and CONTENT slots of the adjective.We will pursue this second alternative.

This would give an entry for the noun modifying use of Albanian along the lines of (39), differing fromthat given previously only in the SUBCAT value. Note the reentrance between the INDEX of the subjectNP (NP:[1]), and the INDEX of the N

that the adjective will modify.

(39)

CONT re f −ob j

RESTR

NUCLEUS INST [1]RELN Albanian

QUANTS <>

[3]

INDEX [1]

CAT SUBCAT < NP:[1] >

HEAD ad jective MOD N

RESTR [3]INDEX [1]

Abbreviating:(40)

CONT [1] Albanian([1]) [3]

CAT SUBCAT < NP:[1] >

HEAD ad jective MOD N :

[1] [3]

Most adjectives appear both predicatively (as complements to be , etc.), and as noun modiers, withidentical subcategorisation, and closely related senses. These regularities can be captured by means of alexical redundancy rule, which we discuss and formulate in the next section.

This concludes our presentation of the ISS-2 approach to adjectives. We have noted a technicaldifculty involving the SUBCAT lists of noun-modifying adjectives, and proposed an alternative. In thefollowing section we will attempt to formulate a lexical redundancy rule for relating the different lexicalentries for adjectives. In subsequent sections we return to the data noted at the start of the paper and inparticular to the issue of NP-internal word order.

April 25, 1992

Page 12: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 12/32

12 HPSG/ADJ

3.3.2. Lexical Redundancy Rule

Predicative and attributive adjectives have the same phonology, and nonlocal values. The differences arein their SYNSEM LOC values. They have the same CONTEXTUAL restrictions. The CONT valuesdiffer in sort. As regards the CAT values, the SUBCAT is the same, but the HEAD values differ in tworespects:

• the attributive adjective contains a MOD value

• they have different values for PRED 3

The differences are summarized in the following.(41)

CONTEXT [8]

CONT psoa pcont

CAT SUBCAT [2]

HEAD ad jective PRD +

−−−−>

CONTEXT [8]

CONT re f −ob j acont

CAT SUBCAT [2]

HEAD ad jective MOD modPRED −

What remains is to describe the relation between the respective CONTents, designated acont and pcont and between the MOD value and the subject of the predicative adjective (which is also the subject of thenoun-modifying one). These matters are closely linked.

The precise difference between the subject and the MOD depends on just what properties are sub-categorized for. But if we assume it is the normal ones then there will be only two obvious differences:

• the MOD will always be nominal, while the subject may be e.g. verbal (VP, or S). We get thislatter effect by restricting the rule to applying to items whose HEAD value is of sort noun . Sinceall predicative adjectives allow nominal subjects (whatever else they may allow), this does notactually cut down the possibilities.

• the subject will typically have to be saturated (NP), while the MOD will always be almostsaturated (N

)

To get these effects, we have to look at the internal structure of the SUBCAT list. So, let [2] be under-stood as a structure: < [3],....,[n] > and consider the relation between [3], and the value mod above: Therelation between their LOC CAT values will be as follows: (the sortal restriction only needs to bestated on one side)

__________________

3 This assumption about the PRED value for noun-modifying adjectives is not very well motivated yet. The MODfeature will already prevent the attributive adjective from appearing in post copula position (even on our assumptionsabout its SUBCAT, so the specication [PRED -] is not needed for this. The feature seems just to correlate with thedifference in semantics as we use it here.

April 25, 1992

Page 13: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 13/32

HPSG/ADJ 13

(42)

SUBCAT <>

HEAD [4]−−−−>

SUBCAT < []>

HEAD noun [4]

So we get:(43)

CONTEXT [8]

CONT psoa pcont

CAT SUBCAT [2]

HEAD ad jective PRD +

−−−−>

CONTEXT [8]

CONT re f −ob j acont

CAT

SUBCAT [2]<SUBCAT <>

HEAD [4], ...>

HEAD ad jective

MOD LOC CAT SUBCAT < []>

HEAD noun [4]

PRED −

This leaves the CONT of the noun-modifying adjective to be considered, (and this includes the CONTof the MOD). The CONT of the adjective will be something like:(44)

RESTR NUC [6] [7]

INDEX [5]

Where [7] denotes the set of restrictions that come from the N

it modies, [6] is related to the content of the predicative adjective (in particular, [6] will be the NUCLEUS from the predicative adjective —which is all there is to the CONT in that case), and [5] has to be shared with the INDEX of the N

. So

the problem is to get at the restrictions that come from the N . We do this making [7] equal to

MOD LOC CONT RESTR. To tie up any contextual restrictions on the MOD and the subject, weshould also associate the INDEX on the MOD with that of the subject. So we treat [3] as actually[3]:[5] (that is, [5] is to be the INDEX of the subject). All this gives:

(45)

CONTEXT [8]

CONT psoa [6]

CAT SUBCAT [2]

HEAD ad jective PRD +

−−−−>

April 25, 1992

Page 14: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 14/32

14 HPSG/ADJ

CONTEXT [8]

CONT re f −ob j

RESTR [6] [7]

INDEX [5]

CAT

SUBCAT [2]<SUBCAT <>

HEAD [4]

[5], ... >

HEAD ad jective

MOD LOCCONT re f −ob j REST [7]

INDEX [5]

CAT SUBCAT < []>

HEAD noun [4] PRED −

This has some other nice effects.

For example, the fact that the INDEX of the MOD and the INDEX of the SUBJ coincide will mean thatany agreement restrictions from the predicative A will carry over to the attributive A. (INDEX is a bun-dle of agreement features; of course, this will only work for semantically conditioned agreeement, whichis argued to be the case in English).

Similarly, suppose the adjective is something like easy , which has a SUBCAT list like the following(ISS-2:Ch4,12):(46) < NP [1],VP inf ⁄ NP acc,[1] >

That is, it subcategorizes for an NP subject, and a non-nite VP which has an accusative gap, where the

NP and the gap are coindexed (that is have the same value for INDEX). In the attributive use of easy ,INDEX of the N

that is modied will share with that of the subject NP, and with the ‘gap’. Thus we

automatically get that in (47) the person in question is understood as the object of please .(47) person i easy to please −i

It is not so obvious how to get this in e.g. trace based formalisms, where one would expect the ller of the gap to be an NP — there is no NP available here, just an N

.

4. WORD ORDER IN NP

4.1. The Problem

In this section we return to the problem of ensuring that APs which contain complements follow thenoun they modify while those without complements typically precede (there are exceptions, like present in editors present , asleep and other "a" adjectives and transient property cases like rivers navigable ).The account we have so far gets the word order in APs like grateful to John for his help and VPs right(by LP1, lexical heads precede complements), but otherwise for head adjunct combinations predicts onlythat lexical heads will precede phrasal adjuncts.

April 25, 1992

Page 15: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 15/32

HPSG/ADJ 15

We begin with a few observations.

The account in ISS-2 does not give an adequate treatment of the ID/LP properties of noun-modifyingadjectives. The discussion of zero-valent (for us, mono-valent) prenominal adjectives in ISS-2 statessimply that adjectives like attributive red combine with nominal N

constituents (recall that this means

almost saturated - still subcategorising a Det) to form N

constituents (chap 8, p14). The problem, how-

ever, is that both the structure and the LP are underdetermined. Consider rst the ID. By ID1, the N

which combines with a Det to form an NP must be an almost saturated phrasal head daughter, but thisnode can itself of course be formed in various ways from lexical head nouns and adjectival adjuncts. Inparticular, notice that ID5 places no restrictions on the lexicality of the adjuncts and heads it combines.Thus an attributive adjective like red can be combined with a noun like book in the following ways:

(48)

(48a)

phrasal sign

N

[SUBCAT <[]>]

----------------------

| |

ID5: |H |A

| |

| |

phrasal sign lexical sign

SUBCAT <[]> red

|

ID2: |

lexical sign

book

(48b)

phrasal sign

N

[SUBCAT <[]>]

----------------------

| |

ID5: |H |A

| |

| |

lexical sign lexical sign

SUBCAT <[]> red

book

As we pointed out above, there is no structure admitted at all for a combination of a noun and acomplement-taking adjunct, such as man fond of children . ID5 would allow a phrasal adjunct to com-bine with a lexical or phrasal head, and ID2 could apply to the almost saturated lexical sign man to givea phrasal sign, but there is no way to actually build the phrasal adjunct unless the adjective subcategor-ises a SUBJECT.

As for the LP of cases like (48a,b), the only LP statement which is relevant is LP1, by which lexicalheads [LEX +] precede their sisters, which then gives the following predicted strings from these struc-tures:

April 25, 1992

Page 16: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 16/32

16 HPSG/ADJ

(48)a’. *book red (from (48a))b’. red book, *book red (from (48b))

But matters are worse still for these adjective-noun combinations when we take these adjectives to bemonovalent, introducing the almost saturated SUBCAT we argued for above. As well as (48a,b), wenow have the possibility of ID2 applying within the adjunct, producing in addition:

(48c)

phrasal sign

N

[SUBCAT <[]>]

----------------------

| |

ID5: |H |A

| |

| |

phrasal sign phrasal sign

SUBCAT <[]> || | ID2

ID2: | lexical sign

lexical sign red

book

(48d)

phrasal sign

N

[SUBCAT <[]>]

----------------------

| |

ID5: |H |A| |

| |

lexical sign phrasal sign

SUBCAT <[]> |

book | ID2

lexical sign

red

and the strings:

(48)

c’. red book (from (48c))d’ *book red, *book red (from (48d))

Now it is possible to derive the following trees for combinations of almost saturated phrasal AP adjunctssuch as fond of children and nominals like man :

April 25, 1992

Page 17: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 17/32

HPSG/ADJ 17

(48e)

phrasal sign

N

[SUBCAT <[]>]

----------------------

| |

ID5: |H |A

| |

| |

phrasal sign phrasal sign

SUBCAT <[]> fond of children

|

ID2: |

lexical sign

man

(48f)

phrasal sign

N

[SUBCAT <[]>]

----------------------

| |

ID5: |H |A

| |

| |

lexical sign phrasal sign

SUBCAT <[]> fond of children

man

and the strings:

(48)e’. *fond of children man, man fond of childrenf’. man fond of children (from (48f))

There are two points to note here - the rst is that ID5 for head-adjunct structures might be too permis-sive and the second is that there are no LP statements to constrain the ordering between adjuncts(phrasal or lexical) and phrasal heads. Finally, whatever the solution to this problem is, we want it toexpress in some direct way the intuition that APS which contain complements or adjuncts occur posthead and others do not.

In the following section, we briey discuss a number of directions which we do not pursue here. Wethen pursue an alternative which starts out from the idea that A+N combinations are lexical structures.Finally, we suggest a variant of this analysis which requires a revision to our earlier assumptions aboutSUBCAT in prenominal adjectives but which seems more adequate.

4.2. Some Non-Solutions

An attractive idea is that post-nominal position is a predicative position. (All adjectives that appearpost-nominally can be used after the copula, and for adjectives like present , which have distinct attribu-tive and predicative readings (roughly ‘current’, and ‘here’), it is the predicative reading that occurs

April 25, 1992

Page 18: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 18/32

18 HPSG/ADJ

post-nominally: the editors present means the same as the editors who were present , which is not thesame as the present editors ). The same reading also occurs in relative clauses. It is not easy to see howthis can be exploited to predict the observed ordering, however. Perhaps, contrary to what we haveassumed, the boolean syntactic feature PRED could be used (post-nominal adjectives being [PRED +]).This would involve two separate lexical redundancy rules for noun-modifying adjectives, and the

account would not extend to explain the post-head ordering of relative clauses. On current assumptions,the relevant ordering principle cannot be stated in terms of the semantics (CONTENT) of the ADJUNCT- the semantics (CONTENT) of predicative adjectives is not compatible with noun-modication. Therole of postnominal adjectives is the same as that of pre-nominal adjectives, relative clauses, and PPswithin in NP. Semantically, all place constraints on the INDEX of the NP. Intuitively, there is noimportant difference between the semantic contributions made by the emphasised parts of (49a-d):

(49)a. a happy personb. a person happy about the weather c. a person who is happy about the weather d. a person in a good mood (about the weather)

This means, of course, that the CONTENT of post-nominal APs must be ref-obj s, like pre-nominal ones(rather than psoa , like predicative adjectives).

Another possibility we can exclude is to exploit FOCUS: which permits focussed constituents to appearphrase nally (e.g. heavy constituents) — this is appropriate for some cases, such as a person tall, thinand balding , but not for the cases we are considering (for one thing, the order we describe is the onlyone possible).

We also want to avoid non-concatenative operations in the phonology (e.g. wrapping). This is just forideological reasons.

Flynn’s solution is also not straightforward - it depends on a feature (whether its category ‘results in S’)

of an expression that can be buried arbitrarily deeply within a construction. The only way to capture thiswould seem to be to have a feature which is percolated (and can come from head, complement, oradjunct), and which records categorial makeup — presumably one would just collect the relevant valuesinto a set, and check this for membership when deciding LP: adjuncts precede heads if their ynn-setdoes not contain categories X,Y, etc.

4.3. Towards a Solution

4.3.1. The Basic Idea

A common assumption, originating with Hudson (1984) is that A+N combinations are actually com-

pounds , that is, lexical structures. In the current framework, this can mean two things:

(a) they are [LEX +]

(b) they are of sort lexical-sign .

As things are set up in ISS-1, (b) and (a) are mutually entailing. However it is suggested that this maybe wrong, and that at least three kinds of structure may be attested. The innovation is (iii), and oneshould probably dene a new sort to allow it. This might be the intention behind the discussion of toxicwaste dump in ISS-1 but it is not elaborated. 4

April 25, 1992

Page 19: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 19/32

HPSG/ADJ 19

(i) [LEX -]; sort phrasal-sign (i.e. with DTRS values)

(ii) [LEX +]; sort lexical-sign (i.e. without DTRS values)

(iii) [LEX +]; sort phrasal-sign (i.e. with DTRS values)

Essentially, our proposal is that pre-nominal adjectives are possibly phrasal signs which are marked

[LEX +], and which are restricted to modifying [LEX +] signs.Something like was hinted at by Fodor (1983), and proposed by Hukari and Levine (1987) for degree-word+Adj structures ( too tall ), and there are many parallels. Most signicant from our perspective is thefact that pre-head adjuncts of A are themselves subject to the same restriction excluding complementsand post-head adjuncts:

(50)a. Sam is too stupid to take with us.a’. *Sam is too to take with us stupid.

Suppose this is correct. We will have an ID rule to admit structures like (51):

(51)+LEX

-----------------------

| |

|A |H

| |

+LEX +LEX

(52) ID5A : a wfp marked [LEX +] can consist of [LEX +] HEAD, and a single [LEX +] ADJUNCT.

For [LEX +] structures, regardless of whether they are lexical or phrasal , the same LP principle willapply (equivalent to, e.g. William’s Right Hand Head Rule): 5

(53) [] < HEAD [ LEX +]

We assume the same treatment will work for Degree words, and adjectival intensiers, giving structureslike the following:

__________________

4 Once you admit this new type, ID2 must be reformulated for a [LEX +] not a lexical sign head daughter (i.e.,reverting to the ISS-1 proposal). Such reformulation would seem to be needed anyway for compounds like ( overcook ?and coordinations).

5 Note the clash between this and the current LP1 which orders lexical heads before sisters. Our new rule must berestricted to operate in [LEX +] phrases. Other LP rules do not mention restrictions on the mother node.

April 25, 1992

Page 20: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 20/32

20 HPSG/ADJ

(54)

+LEX

N

----------------

| |

|A |H

| |

+LEX +LEX

A N

---------------- man

| |

|A |H

| |

+LEX +LEX

Deg A

very tall

4.3.2. Some Problems

There are some empirical issues which we should deal with at once:

• Examples like (55a) raise an objection to Hudson’s original claim: words are generally anaphoricislands, so it should not be possible for one to have part of an A+N structure as its antecdent. Infact, this is the case with A+N combinations that clearly do form single words (55c). Moreover,one is generally taken to have the distribution of an N

, which would exclude it from compounds

(cf (55d)).

(55)

a. brown bears and white onesb. ?polar bears and brown onesc. *blackboards and white onesd. *polar bears and teddy ones

None of this is fatal to our proposal — which should perhaps be taken as a renement of Hudson’s.

We have to assume that it is lexical signs , rather than [LEX +] structures which form anaphoricislands — hence anaphora into an A+N structure will be possible if it is a phrasal sign (which weassume examples like (55a) are). To say that one is an N

is say that it has the distribution of an

almost saturated N. There is no reason why such structures should not occur inside [LEX +] struc-

tures (for example, most lexical nouns are exactly of this kind).• A second potential problem for Hudson is the fact that pre-nominal adjectives can be coordinated

(of course, prexes can coordinate:(56) tall, thin and balding people

This is not problematic here, since it is clear that [LEX +] signs must be able to coordinate pro-ducing [LEX +] signs — this is the natural analysis of examples like (57) — hugged and kissed should form a [LEX +] sign so that it can combine with its object by ID2 (other analyses are pos-

April 25, 1992

Page 21: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 21/32

HPSG/ADJ 21

sible, of course, involving RNR, or extraposition).(57) [ hugged and kissed ] Leslie

• We get an interesting prediction in the case of Ns with complements or adjuncts, such as:

(58)a. Sam is an alleged burglar fond of country houses.b. This is a fake antique in nest mahogany.c. He is a former policeman honest in his dealings with the public.

It does not seem possible to get a reading of (58a) where alleged has scope over both burglar and fond of country houses , (58a) asserts that Sam is fond of country houses, and is an alleged burglar,at least on its most salient reading. Similarly, (58b) does not suggest that the material has beenfaked, only the artifact, and (58c) does not seem to be synonymous with former honest policeman .In general, it seems that postnominal modifers are outside the scope of prenominal modiers, atleast on the most salient reading. This could be explained on this account, since the pre-nominaladjective has to combine with the N before the N combines with its arguments. Thus, on theaccount suggested earlier, only the CONTENT that comes from the N will be an argument to theallege relation. Gazdar (pc) has suggested that in cases like fake photograph of Martin Borman ,

fake does have scope (loosely speaking) over the N + PP combination. This is not a counterexam-ple on the assumption that the CONTENT of the noun (same for all picture nouns) is somethinglike:(59)

RESTR NUC

ARG2

Martin Borman

ARG1 [1]RELATION ‘photogragh ′

INDEX [1]

4.3.3. Post-Nominal Adjuncts

Now, it is natural to make the complementary assumption about non-lexical adjuncts, namely that theyonly combine with non-lexical heads. Thus, we have a restriction on ID5, so that head, adjunct, andmother must all be [-LEX]. Of course, this can be collapsed with ID5A above: the restriction on ID5 is

just that head, adjunct, and mother must agree on LEX.(60) ID5 : a wfp marked [LEX α ] can consist of [LEX α ] HEAD, and a single [LEX α ] ADJUNCT.

April 25, 1992

Page 22: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 22/32

22 HPSG/ADJ

(61)

HEAD [1]

CONTENT [2]

SUBCAT [4]

LEX [4]

-------------------------------------

| |

|H |ADJ

| |

| |

SYNSEM [3] MOD [3]

SYNSEM|LOC|CAT|HEAD [1] CONTENT [2]

SUBCAT [4] LEX [4]

LEX [4]

One possibility is that we introduce a further LP statement to the effect that phrasal ADJUNCTS follow

the head. However, it is also possible to deal with this in terms of obliqueness: suppose that (as sug-gested in ISS-1) heads and adjuncts gure in the obliqueness hierarchy (which is no longer denedsolely in terms of the SUBCAT list). Then if heads are less oblique than adjuncts (though more obliquethan complements), the desired ordering is a special case of (62).(62) [LEX -] << [LEX -]

In any case, Sag (1987:334) proposed (63) for independent reasons.(63) HEAD [LEX -] < ADJUNCT [LEX -]

Of course, this requires adjectives with complements or non-degree modiers to follow the N — theymust be [LEX -] because they have combined with a complement (by ID2) or with a phrasal adjunct.

4.3.4. Keeping Lexical Adjectives before the Noun

So far, we can explain why (64a,c) are possible, and (64d) is excluded

(64)a. tall personb. *person tallc. person fond of childrend. *fond of children person

We are not yet out of the woods. One problem is that there is nothing to stop tall forming a [LEX -]

sign, by combining with its empty list of complements (in the same way that e.g. intransitive verbsbecome VPs) under ID2. But once it is phrasal, it will have to appear post-nominally, as in (64b).Again, a number of alternatives suggest themselves. We will discuss one approach and then sketch outwhat appears to be a more promising alternative, but which requires an alteration to our previousassumptions about the SUBCAT of prenominal adjectives.

One possibility is that adjuncts describe the kind of head they can modify. Thus, it is quite possible foran adjunct to be restricted to modifying [LEX +] heads. Suppose that prenominal adjectives are like this— it follows they will not be able to operate as modiers at all once they have become phrasal — sincethe new head-adjunct schema requires that head, adjunct and mother all have the same values for LEX.

April 25, 1992

Page 23: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 23/32

HPSG/ADJ 23

Thus, they will never appear postnominally.(65) happy: MOD LOC CAT HEAD LEX +

(66)a. *a man happyb. a happy man

We can account for the behaviour illustrated in (67) by assuming that asleep is restricted to modifying a[LEX -] head.(67) asleep:MOD LOC CAT HEAD LEX -

(68)a. *an asleep manb. a man asleep

This means it cannot appear in head-adjunct structures until it becomes phrasal ([LEX -]) by combiningwith its zero complements. (Perhaps it is lexically specied as itself intrinsically [LEX -], but we do notwant to have to recognise a further, fourth type: a [LEX -] lexical sign ). Note that since MOD is a head

feature, this restriction will be passed on to phrases like completely asleep .

The prediction is that MOD LEX- adjectives will only be able to modify non-lexical projections of N.This means that MOD LEX - adjunts will never be able to appear before complements. This seems tobe borne out: 6

(69)a. teacher of English asleepb. *teacher asleep of Englishc. teacher of English fond of childrend. *teacher fond of children of English

Note that we need both this idea and the modied ID schema: if we have just the ID schema, we getman happy , as already noted. If we just have the idea of MOD LEX+/-, then we restrict what the A cancombine with, but not what the result can combine with — in particular, we do not predict that LEX+and LEX+ gives LEX+, (recall that LEX is not a head feature). But we have to have this, because e.g.very happy has the same distribution as happy .

What is the distribution of the values of the MOD LEX feature? Suppose fond is were marked asMOD LEX+. It would be able to appear pre-nominally (as in fond man ), but not post-nominally, andnot if it had its complements (when it would be LEX-, and so only able to undergo the LEX- case of ID5) which would mean combining with a LEX- head, which it could not do, because of its MOD LEXvalue.

It seems clear that the vast majority of monovalent noun-modifying adjectives are specied MOD LEX+. This means that IF they (vacuously) undergo ID2 and become phrasal, they can only combine with aphrasal HEAD and their MOD LEX requirement will not satised. This is therefore ruled out. So, wewill have an LRR which says that subcat < []> entails MOD LEX +. This can be evaded by lexicalspecication. Now, many As do not need to be specied for MOD LEX. Consider fond : if it isunspecied for MOD LEX it can undergo either case of ID5, but if it undergoes the LEX+ case, it will__________________

6 Actually, examples like: fact, obvious to us all, that Sam has gone are alright but need the right intonation. Theyinvolve some sort of focussing/extraposition.

April 25, 1992

Page 24: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 24/32

24 HPSG/ADJ

not be able to combine with its complements. This is not possible for fond , but it is for other adjectives.

It is evident that the MOD LEX feature is not very nice. Indeed, it is little more than a device to ensurethe right LP predictions by ensuring that certain adjectives do not become phrasal by application of ID2.This device does not seem very insightful and seems to miss the point, which is that it is the occurrenceof complements or adjuncts to A containing verbal or nominal material which forces the post-nominalposition. Worse still, there ARE cases in which " prenominal " adjectives occur post-nominally:

(70)a. a cry too loud to ignoreb. a wall too high to jump

So we now sketch out (very roughly) another alternative which seems to overcome this descriptivedeciency.

This alternative differs in two respects:

• There are two lexical redundancy rules, not one. Rule A applies to monovalent adjectives (exceptlexical exceptions like asleep ) and derives zerovalent adjectives. Rule B (which is essentially the

redundancy rule formulated in section 2. above) applies to bi- and tri-valent adjectives and doesnot alter SUBCAT. The effect of this is simply that zero-valent adjectives, which are saturated,cannot undergo ID2 vacuously and become [LEX -].

• There is no selection for the LEX feature of the noun (ie, there is no MOD LEX requirement).Thus in principle zero-valent adjectives (like red ) CAN combine with phrasal heads. In order todo so, they must themselves become phrasal by combining with some adjunct and its complement.We maintain the assumption that too high is a complex [LEX +] sign which can combine with thecomplement VP (probably by ID2).

Of course a complete treatment of cases like (70) requires us to formulate some sort of complement-sharing mechanism to pass the SUBCAT requirements of too to the [LEX +] phrasal node too high . Weoffer some preliminary thoughts on this matter in the following section.

There is a nice feature of this approach. Since we distinguish in principle between two classes of noun-modifying adjectives we can say something about why post-nominal adjectives " feel " predicative. Thisis because, in the general case, they DO subcategorise for subjects (like relative clauses and unlike pre-nominal adjectives).

5. OTHER CONSTRUCTIONS

In this section we turn our attention to some of the other problematic constructions noted in (1), andrepeated here, and make some very brief, general remarks.

(71)a. a hard to pronounce name ‘complex As’b. a hard problem to solve ‘hard nuts’c. too tall (for us) to ignore ‘degree complements’d. so tall that we could not ignore them

April 25, 1992

Page 25: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 25/32

HPSG/ADJ 25

5.1. Complex Adjectives

The standard assumption about Complex As, such as hard to pronounce is that they are in some sense(complex) lexical items (e.g. Nanni 1978, 1979; Chomsky 1981, Keenan and Timberlake, Montalbetti etal). 7

Our account will be that these are phrasal-signs which are [LEX +], as in (72). Thus, they will haveinternal constituent structure, but have the distribution of ordinary ‘simple’ adjectives.

(72)a. an [ LEX + easy to understand ] proposal

This not only offers an account of many of the standard observations about complex-As, but overcomessome of the most obvious problems with these accounts.

(73)a. an [ LEX + easy to understand sounding ] proposalb. *easy to understand+ly (=‘in an easy to understand manner’)c. *a hard-to-pronounce enough name

d. a hard enough to pronounce name

For example, an account of the observation (from Nanni 1978) that Complex-As take part in compound-ing processes, as in (73a), is available if the relevant compounding processes are open to [+LEX] items,rather than just lexical signs (not all compounding processes are open in this way, we assume 8). Of course, this does not recommend our proposal over any of the other standard views. However, ouraccount may also provide an explanation of why (73b) is ungrammatical. On a standard account, if easyto understand is a lexical adjective, one would expect it to take part in (inectional) morphological pro-cess such as adverb formation. However, as (73b) indicates, this is not the case. On our account, thedomain of processes like adverb formation will be lexical-signs, not [+LEX] items, accounting for this(notice that the ill-formedness of (73b) is probably not semantic in nature, there is nothing semantically

odd about something being done in an easy to understand manner).Similarly with (73c,d), the generalization about enough is that it normally follows the adjective itmodies. If easy to understand is an adjective, one would expect (73c) to be acceptable. We can accountfor this if we assume that enough follows the rst lexical-sign in whatever it modies. 9

A formulation of the rule for forming Complex-As might be as in (74):(74) ID complex − As: A LEX + → A tough , VP ⁄ NP acc

__________________7 Chomsky 1981, and much following work has assumed that hard to please is a complex lexical item in both

attributive and predicative positions. Some of the evidence for this is the impossibility of non-peripheral extractions(which on this account follows from the impossibility of extracting material from inside a lexical item):

a. Which violins i are these sonatas j hard to play e j on ei.b. *Which sonatas j are these violins j hard to play ei on e j.

However, we will assume that only the pre-nominal use of such expressions are complex-As, and follow Kaplan andBresnan (1982) in relying on some restriction on crossing dependencies (perhaps based on some kind of processingaccount) to account for this phenomenon.

8 It is important that this is the case, since otherwise one would expect adjectives to appear inside (e.g.) noun-nouncompounds, giving * shoe big shops , instead of big shoe shops (this was pointed out to us by Dan Flickinger). Oneaccount of this would be that noun-noun compounding is restricted to lexical signs — since big shops is not a lexicalsign on our account, it is not possible to form the compound * shoe big shops .

9 Of course, this is not an explanation of the distribution of enough , which poses non-trivial problems, for thisaccount or any other.

April 25, 1992

Page 26: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 26/32

26 HPSG/ADJ

This is given as an a further special, and English-specic ID rule, which generates +LEX structures outof tough adjectives, and VPs with accusative gaps (e.g. easy + to understand ). This means that it is a‘syntactic’ rule. Notice that the existence of syntactic rules that introduce some sort of ‘lexicality’feature (making lexical items out of syntactic constructions) is independently argued for in the analysisof a number of languages (cf. Spencer, 1992: and references there). This accounts for the apparent inter-

nal constituent structure, and the presence of to . One problem with this formulation is that it does notaccount for the restrictions on the VP, which as the following indicate, cannot contain more than onelevel of embedding, or material after the verb (or after the accusative ‘gap’). This might suggest a for-mulation as below, which is in fact very similar to Nanni’s proposal for a lexical rule (here V nft abbrevi-ates "a non-nite verb with SUBCAT list that contains a subject, and an accusative object ").

(75)a. *an [ easy to try to understand ] proposalb. *an [ easy to understand quickly ] proposal

c. A LEX + → A tough , to Vnft

5.2. Hard-Nuts and Degree Complements

Here we will sketch out two possible analyses. The rst is a fairly standard sort of extrapositionanalysis, and our aim will be to show that we can accommodate such an analysis, and actually explainone feature of the behaviour of degree words such as too , and so . The second analysis is more specula-tive, and is based on the idea that what is involved in these constructions is some sort of merger or com-bination of the SUBCAT lists of the adjunct and the head items.

The standard account of these cases probably involves some sort of extraposition. We do not have amechanism for handling extraposition here — in particular, it should be noted that the treatment of extraposition given in ISS-1 relies on heads, complements, and adjuncts being given ‘at’ structures (theidea was that in such a structure a daughter could be taken either as a complement or adjunct of thehead, or as an adjunct of one of the complements). However, it is easy enough to imagine a fairly stan-dard extraposition kind of account. 10

First, consider what the lexical entry of a degree word like so , as in (76a), would look like. Notice thatthe nite S that we could not ignore them is a complement of so , but cannot occur next to it.

(76)a. so tall that we could not ignore themb. *[ so that we could not ignore them ] tallc. *tall [ so that we could not ignore them ]

If we assume that so is an adjunct 11 , we can expect an entry like (77).

__________________10 This has a family resemblance to GPSG style analyses of these constructions such as those in Chae (1990), and

Hukari and Levine (1987), though there are important differences.11 This is not a standard assumption. The more normal assumption is that degree words are Speciers. It is not clear

that very much hangs on our assumption, but it is useful because it makes the parallel between the degree wordconstructions and the ‘hard nuts’ clearer.

April 25, 1992

Page 27: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 27/32

HPSG/ADJ 27

(77)

SYNSEMNONLOCAL INHERITED SLASH [1]

LOCAL CAT SUBCAT < [1]:S n>

MOD ad jective LEX +

PHON ⁄ so ⁄

This states that so must modify a [+LEX] item, but subcategorizes for a nite S (it may also subcategor-ize for a SUBJECT; a matter we will return to), and that this nite S also appears as a member of itsSLASH set. This will start a SLASH value percolating up the tree, although there is no trace, until suchtime as it is realized. Such dependencies can be bound off by the introduction of a ller (the displacedelement). This involves the head-ller ID schema, schematically:(78) Head-Filler Schema :

X → HEAD

TO −BIND SLASH [1]

INHER SLASH [1], . . .CAT s n yes

FILLER [1]

The intuitive force of this is just the following (so long as we are concerned with singleton SLASHsets):(79) X → X ⁄ YP, YP

Notice that this ID schema is restricted to heads of category S and must be altered if it is to apply to the

case in question. Assuming an appropriate modication, it would give rise to trees like the following:

(80)

AP

-----------------------

| |

| |

A/{[1]:S} [1]

LEX+ S

--------------- that we could not ignore them

| |

A/{[1]:S} A

LEX+ LEX+

so tall

This is rather standard, but one interesting feature of this particular analysis is that the requirement thatso is both MOD LEX+, and subcategorizes for a nite S is, in some sense a contradiction. To beMOD LEX+ means that it must combine with a lexical head, while it is itself still lexical; but of course,the fact that it subcategorizes for something means that it should combine with that, and so becomenon-lexical. The only way that this tension can be resolved is if so rst combines with the item itmodies, and then combines with its complement — extraposition is, in this sense, obligatory. This

April 25, 1992

Page 28: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 28/32

Page 29: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 29/32

HPSG/ADJ 29

different analysis is adopted, which does not involve the use of the SLASH apparatus at all, and whichis reminiscent of a lexically restricted version of function composition, as found in Categorial Grammaranalyses.

Such an analysis could be developed along the following lines. Suppose that adjuncts (of certain classes,but including degree words, and the tough adjectives) can combine their SUBCAT lists with those of their heads, and pass this to their mother. In essence, and with one caveat regarding the status of thesubject of the adjunct, this would mean that hard teacher subcategorises for the complements of hard and for those of teacher (i.e. the determiner, and the PP of ). The ordering effects noted above follow, if we assume that the complements of the adjunct are appended after those of the head. This would givethe following effects:

(84)a. hard SUBCAT < NP, [1]:VP inf ⁄ NP acc >

b. teacher SUBCAT < DETP, PP of >

c. hard teacher SUBCAT < DETP, PP of , [1]:VP inf ⁄ NP acc >

a. so SUBCAT < S>

b. hard SUBCAT < NP, VP inf ⁄ NP acc >

c. so hard SUBCAT < NP, VP inf ⁄ NP acc , S>

Since the complements of the adjunct are later on the SUBCAT list of the derived Adj+N, or Deg+Astructure, they are more oblique, and the normal ordering rule for complements will order them laterthan the complements of the head, thus giving the appearence of nesting.

(85)

AP

------------------------------------

| | |

| | |A,LEX+ | |

<NP,VP,S> VP S

--------------- for us to solve that we had

| | to give up

A,LEX+ A,LEX+

<S> <NP,VP>

so hard

April 25, 1992

Page 30: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 30/32

30 HPSG/ADJ

(86)

N <DETP>

---------------------------------

| | |

| | |

N,<DETP,[1]:PP,[2]:VP> [1] [2]

LEX+ PP VP/NP

--------------- of English to please

| |

A,<NP,[2]:VP> N,<DETP,[1]:PP>

LEX+ LEX+

hard teacher

There are several problems with this analysis. One is that the disappearence of the the subject from theadjunct is very surprising, given the standard view of the SUBCAT list. However, some ideas of Borsley’s may be useful here (essentially the idea is to remove the subject from the SUBCAT list). We

also do not have any good account as to why it is such a restricted class of adjectives (essentially justthe tough adjectives) which occur in this construction.

6. OUTSTANDING PROBLEMS

There are two remaining contexts in which noun-modifying adjectives appear. We have not attempted toprovide an account of them here.

(87)a too good a chance to miss ‘big messes’b somebody interesting indenite pronouns

The triggering condition for the ‘big mess’ construction, where the degree word and adjective appearbefore the determiner of the noun they modify, is the degree word (e.g. so and too require it, enoughallows it marginally, and items like very , sufciently , excessively do not allow it at all). As things stand,indenite pronouns are straightforward counter-examples to our analysis (if the pronoun is [LEX +], onewould expect the adjective to precede, if the pronoun in [LEX -], one would not expect ‘ordinary’ (i.e.[MOD LEX+]) adjectives to appear at all (any more than they do in * man interesting ). Of course, anatural thing to say about such pronouns is that they are in some sense ‘fused’ from determiners ( some )and nouns ( body ), and this fusion in some way makes it impossible for the adjective to take its normalprenominal position. In this sense, they would be evidence for a kind of markedness analyis, involvingboth pre- and post- nominal positions being possible for central adjectives, but post-nominal positiononly being acceptable if pre-nominal position is ruled out for some reason. This is an attractive view,but the default logic it implies is problematic.

April 25, 1992

Page 31: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 31/32

HPSG/ADJ 31

7. REFERENCES

D.J. Arnold Theoretical and Descriptive Issues in Machine Translation , PhD. Thesis, University of Essex, Colchester, 1989.

A. Berman Adjectives and Adjective Complement Constructions in English , PhD Thesis Dept of Linguis-

tics, Harvard University, 1974, (also published as Report NSF-29).R. Borsley ‘‘Subjects, Speciers and Complements in HPSG’’, Readings in Information Based Syntax

and Semantics , C. Pollard and I. Sag, eds., CSLI, Stanford, Ca., forthcoming .

H-R. Chae ‘‘Gap Licensing in tough and similar constructions’’, WCCFL 9 , 1990.

N. Chomsky Lectures on Government and Binding , Foris Pub., Dordrecht, 1981.

J. Emonds A Transformational Approach to English Syntax , Academic Press, New York, 1976.

M. Flynn ‘‘A Categorial Theory of Structure Building’’, Order Concord and Constituency , G. Gazdar,E. Klein and G.K. Pullum, eds., Foris Pub., Dordrecht, 1983.

Janet D. Fodor ‘‘Phrase Structure Parsing and the island constraints’’, Linguistics and Philosophy , Vol.6, 1983 , pp 163-223.

R. Hendrick ‘‘The Phase Structure of Adjectives and Comparatives’’, Linguistic Analysis , Vol. 4, 1978,pp 255-299.

R. Hudson Word Grammar , OUP, Oxford, 1984.

T.E. Hukari and R.D Levine ‘‘Rethinking Connectivity in Unbounded Dependency Constructions’’,WCCFL 6 , 1987.

R.M. Kaplan and J. Bresnan ‘‘Lexical Functional Grammar: a Formal System for GrammaticalRepresentation’’, The Mental Representation of Grammatical Relations , J. Bresnan, ed., MITPress, Cambridge, Mass, 1982, pp 173-282.

E.L. Keenan and A. Timberlake ‘‘Natural Language Motivations for Extending Categorial Grammar’’,

Categorial Grammars and Natural Language Structures , R.T. Oehrle, E. Bach and D. Wheeler,eds., D. Reidel Publishing Co, Dordrecht, 1988, pp 265-295.

M. Montalbetti , M. Saito and L. Travis ‘‘Three Ways to Get Tough’’, Papers from the Regional Meet-ing of the Chicago Linguistic Society , Vol. 18, 1982, pp 348-366.

D. Nanni The easy Class of Adjectives in English, PhD Thesis, Dept of Linguistics, UMass, Amherst, Mass, 1978.

D. Nanni ‘‘On the Surface Syntax of Constructions with Easy -type Adjectives’’, Language , Vol. 56,1980, pp 568-581.

C. Pollard and I. Sag ‘‘An Information Based Approach to Syntax and Semantics: vol 2: AgreementBinding and Control’’, forthcoming, (draft June 1991).

Carl Pollard and Ivan Sag Information Based Syntax and Semantics , Chicag University Press, Chicago,CSLI Lecture Notes, 13, Vol. 1: Fundamentals, 1987.

Ivan Sag ‘‘Grammatical Hierarchy and Linear Precedence’’, Discontinuous Constituency , Geoffrey J.Huck and Almerindo E. Ojeda, eds., Academic Press, Syntax and Semantics, Vol. 20, 1987, pp304-342.

M. Siegel ‘‘Measure Adjectives in Montague Grammar’’, Linguistics, Philosophy and Montague Gram-mar , S. Davis and Mithun M., eds., University of Texas Press, Austin, 1979, pp 223-262.

April 25, 1992

Page 32: Arnold&Sadler 1992

8/13/2019 Arnold&Sadler 1992

http://slidepdf.com/reader/full/arnoldsadler-1992 32/32

32 HPSG/ADJ

A. Spencer Morphological Theory , Blackwell, Oxford, 1992.

E.S. Williams ‘‘Another argument that Passive is Transformational’’, Linguistic Inquiry , Vol. 13, 1982,pp 160-63.


Recommended