head-driven phrase structure grammar - github pages · 2020-02-05 · subj 9 pred % agreement pers...

Head-driven Phrase Structure GrammarArchitecture

Weiwei Sun

Institute of Computer Science and TechnologyPeking University

May 29, 2019

Formal Foundation: Informal Introduction A System of Sign

Developing a feature structure based syntactic theory (1)

Question

Manipulate feature structures in a systematic way

I How to specify which feature structures are well-formed?

I How to combine small feature structures to larger ones?

Mathematically sound model is employed to guarantee that richfeatures can be widely defined for modeling various linguisticphenomena.

HPSG describes linguistic objects with typed feature structures

B. Carpenter. The Logic of Typed Feature Structures

Weiwei Sun Head-driven Phrase Structure Grammar 2/77



Proposal

Languages are systems of sorts of linguistic objects at avarietyof levels of abstraction.

I Grammars assign structural descriptions and semanticinterpretations to linguistic objects.

I Grammars are best represented as process-neutral systemsof declarative constraints.

Question

Define feature structures in a linguistic way

I What kinds of features go together?

I What values are appropriate for each particular feature?




Question

Define feature structures in a linguistic way

I What kinds of features go together?

I What values are appropriate for each particular feature?

References

I Sag, Wasow and Bender. Syntactic Theory: A FormalIntroduction

I Pollard and Sag. Head-Driven Phrase Structure Grammar

* Sign-based Construction Grammar



Head-driven Phrase Structure Grammar

Pioneers

I C. Pollard:I Ph.D., Linguistics, Stanford, 1984I M.A., East Asian Languages and CulturesI M.S./B.S, Mathematics,

I I. A. Sag (Nov. 9, 1949 - Sep. 10, 2013):I Ph.D., Linguistics, MIT, 1976I M.A./B.A., Linguistics,



Outline

Formal Foundation: Informal Introduction

A System of Sign



Signature (1)

Signature

A signature defining the appropriate space of feature struc-tures, including a multiple-inheritance type hierarchy, feature andvalue-type declarations.

feature-structure

pos

conjadjprepagr-pos[agr]

detverb[aux]

noun

val-cat[spr,comps]

expression[head pos]

phraseword



Signature (2)

The signature defines what exists:

I which kind of objects are distinguished, and

I which properties of which objects are modeled.

The signature consists of

I the type (or sort) hierarchy andI the appropriateness conditions, defining which type has

which appropriate attributes (or features) with whichappropriate values.

* Some atomic types have no feature appropriate for them

Closed World Assumption

There are no linguistic species beyond what is specified in thetype hierarchy.



Typed feature structure: Informal idea

TFS’s can be represented as rooted directed labeled graphs:

I Nodes are labeled with types

I Edges are labeled with features.

I A feature path is a list of features through a featurestructure leading to a particular value.

I Reentrancy/structure-sharing: two attributes denoteidentical information.

Example

noun

AGR&&

3rd

predicate

SUBJ

99

PRED%%

agreement

PERS

77

NUM''

verb

AGR

88

singular



TFS: Graphs or AVMs

Graph

noun

AGR&&

3rd

predicate

SUBJ

99

PRED%%

agreement

PERS

77

NUM''

verb

AGR

88

singular

AVM

predicate

subj

noun

agr 1

agreement

[person 3rd

num singular

]pred

verb

[agr 1

]

Path: 〈SUBJ, AGR, NUM 〉 = singular



Feature structures in HPSG: Typed

predicate

subj

noun

agr 1

agreement

[person 3rd

num singular

]pred

verb

[agr 1

]

Well-typed

I Feature structures (FS) employed in HPSG are sorted.

I Each node is labelled with a type symbol

I HPSG requires feature structures to be well-typed.

I What attribute/feature labels can appear in a FS aredetermined by its sort (i.e. ontological category).

I The value of a feature must also be of a appropriate type.




predicate

subj

noun

agr 1

agreement

[person 3rd

num singular

]pred

verb

[agr 1

]

Well-typed

I Feature structures (FS) employed in HPSG are sorted.I Each node is labelled with a type symbol







feature-structure

pos


detverb[aux]

noun

val-cat[spr,comps]


phraseword

Well-typed








feature-structure

pos


detverb[aux]

noun

val-cat[spr,comps]


phraseword

Well-typed


I HPSG requires feature structures to be well-typed.I What attribute/feature labels can appear in a FS are

determined by its sort (i.e. ontological category).





feature-structure

pos


detverb[aux]

noun

val-cat[spr,comps]


phraseword

Well-typed


I HPSG requires feature structures to be well-typed.I What attribute/feature labels can appear in a FS are

determined by its sort (i.e. ontological category).I The value of a feature must also be of a appropriate type.



Feature structures in HPSG: Totally well-typed

feature-structure

pos


detverb[aux]

noun

val-cat[spr,comps]

expression[head]

phraseword

FS that serves as models of linguistic entities is required to satisfyfurther criteria of completeness. More precisely, they are requiredto be both totally well-typed and sort-resolved.

Totally well-typed

For each node, every feature that is appropriate for the sort as-signed is actually present.



Feature structures in HPSG: Sort-resolved

feature-structure

pos


detverb[aux]

noun

val-cat[spr,comps]

expression[head]

phraseword

‘agr-pos’ and ‘ noun∨verb∨det’ are equivalent. ⇒ A type (orsort) is really a disjunction of its maximally specific subtypes.

Sort-resolved

An FS is sort-resolved provided every node is assigned a sort labelthat is maximal in the sort ordering.A type without subtypes is called a maximal type.



Linguistic description

Theory

A theory (in the formal sense) is a set of description languagestatements, often referred to as the constraints.

I The theory singles out a subset of the objects declared inthe signature, namely those which are grammatical.

I A linguistic object is admissible with respect to a theory iffit satisfies each of the descriptions in the theory and sodoes each of its substructures.

Descriptions of feature structures need not be complete.



An example

Example

(1) a. Ja spal. � Imasc.sg sleptmasc.sg

b. On spal. � Hemasc.sg sleptmasc.sg

word

synsem | loccat |head noun

cont | index[num sing

gen masc

]

I This AVM specifies the “partial” constraints on thecomplete (totally well-typed) feature structure of thesubject.

I This AVM doesn’t specify the entire (totally well-typed)feature structure, just what needs to be true in the featurestructure.



Subsumption

Example

word


cont | index[num sing

gen masc

]

subsume

word


cont | index

per 1st

num sing

gen masc

word


cont | index

per 3rd

num sing

gen masc



Structure sharing

Examplephrase

head 1 verb

val

[comps itr

spr −

]→

word

head 1

val

[comps itr

spr −

]

Structure sharing is a key descriptive mechanism in HPSG.



Auxiliary data structures

>

...list

non-empty-list

first >rest list

empty-list

boolean

false(−)true(+)



Outline

Formal Foundation: Informal Introduction

A System of Sign



An ontology of linguistic objects

Signs: form+meaning (Saussure)

I The type sign has the features phon and synsem.

I sign is the basic sort/type in HPSG used to describe lexicalitems (of type word) and phrases (of type phrase).

sign

phon list(phonstring)

synsem synsem

[

phrase

dtrs constituent-structure

][word

]

synsem

local local

nonlocal non-local

local

category category

content content

context context

category

head head

subcat list



An ontology of linguistic objects (1)

synsem

local local

nonlocal non-local

local

category category

content content

context context

category

head head

subcat list

I The feature synsem has a value of type synsemI This type itself has relevant features ( local and nonlocal)

I local introduces cat(cateogry), cont(content) andconx(context)

I These correspond to syntactic, semantic, and pragmaticnotions, all of which are locally determined.

I non-local is connected with unbounded dependencies

I Most linguistic constructions can be handled locally.Non-local constructions require different mechanisms



An ontology of linguistic objects (2)

category includes,

The sign’s syntactic category

I Given via the feature [ head head], where head is thesupertype for noun, verb, adjective, preposition,determiner, marker

each of these types selects a particular set of features

The sign’s subcategorzation frame/valence

I List-valued feature[synsem | loc |cat | subcat list(synsem)

]I If the list is non-empty, the sign has the potential to

combine with another sign



Part-of-speech

head

[substantive

]

[preposition

pform pform

][noun

case case

]verb

vform vform

aux boolean

inv boolean

adjective

[functional

]

deteminermarker

vform

passive-partpast-partpresent-partgerundbaseinfinitefinite



subcat feature

synsem, not sign

The subcat list takes as values the list of synsems instead ofsigns

Locality

I Word does not have access to the dtrs list of items on itsvalence lists

I A word cannot access properties of the daughters of itsdaughters.

⇒ Constructions are thus restricted to local relations



Content features

synsem

loc

cat

[head ...

subcat ...

]cont content

conx ...

nonloc non-loc

I The cont feature specifies different semantic information.

I The sort that is the value of the cont feature variesaccording to wether the sign’s head value is of type noun,quantifier, or something else.

I For most parts of speech, the cont value is aparameterized-state-of-affairs (psoa).



Semantic representations

(2) Every bird flies.

nominal-object: an individual/entity (or a set of them), associ-ated with a referring index, bearing agreement featuresparameterized-state-of-affairs: an event relation along withrole names for identifying the participants of the eventquantifiers

content

nom-obj

index index

restriction set(psoa)

psoa

think′

thinker ref

thought ref

drink′

drinker ref

drunken ref

give′

giver ref

given ref

gift ref

[

laugh′

laugher ref

]



Index

nom-obj

index index

restriction set(psoa)

index

person person

number number

gender gender

ittherereferential

person

thirdsecondfirst

number

pluralsingular

gender

neuterfemininemasculine

I index: A feature appropriate for nominal-object objectsI Agreement features is stated through the index featureI case was put somewhere else (within head), so case

agreement is treated differently than person, number, andgender agreement.



Semantic representation: An example

nominal-object: an individual/entity (or a set of them),associated with a referring index, bearing agreement features

book

npro

index 1

per 3rd

num sing

gend neut

restriction

[

book’

instance 1

]



Abbreviations of common AVMs

NP 1

synsem

locallocal

cat cat

[head noun

subcat 〈〉

]cont | index 1

S: 1

synsem

locallocal

cat cat

[head verb

subcat 〈〉

]cont 1

VP: 1

synsem

locallocal

cat cat

head verb

subcat⟨

synsem⟩

cont 1



Syntax and semantics

Using structure sharing

word

phon⟨

laughs⟩

ss | loc

cat | subcat

⟨NP[nom] 1 [3rd,sing]

⟩cont

[laugh′

laugher 1

]

NP 1

synsem

[local

local

[cont | index 1

]]

The index of the NP on the subcat list is said to unify withthe value of laugher.



More examples

sees

word

phon⟨

sees⟩

ss | loc

cat | subcat

⟨NP[nom] 1 [3rd,sing],NP[acc] 2

⟩cont

see′

seer 1

seen 2

NP 1

synsem

[local

local

[cont | index 1

]]



More examples

gives

word

phon⟨

gives⟩

ss | l

c | sc⟨NP[nom] 1 [3rd,sing], NP[acc] 2 ,NP[acc] 3

⟩

cont

give′

giver 1

given 2

gift 3

NP 1

synsem

[local

local

[cont | index 1

]]



A lexicalist framework (1)

I HPSG relies crucially on complex lexical information todetermine the grammatical properties of phrasal expressions.

I The lexicon defines the ontologically possible words that aregrammatical:

word → lexical entry1 ∨ lexical entry2 ∨ ...I Properties of lexical entries and relationships among them are

expressed in a concise and principled fashion by employingTFS.

Lexical entry: laughs

word

phon⟨

laughs⟩

synsem | local

cat

head verb

[vform fin

]subcat


⟩

contlaugh′

[laugher 1

]



A lexicalist framework (2)

HPSG doesn’t rely on complex lexical stipulations

Properties of lexical entries and relationships among them areexpressed

I in a concise and principled way

I in terms of classification by a multiple inheritancehierarchy and

I lexical rules.

The multiple-inheritance architecture

I Distinct lexical entries with shared properties leads tomassive redundancy within the lexicon.

I To cross-classify words according to properties sharedacross word-classes which are represented by “generic”lexical types.



The multiple-inheritance architecture

How to specify lexical information?

I Each generic lexical type specifies certain constraints i.e.values of certain features or relationships among values ofdifferent features (structure sharing).

I Such constraints must hold of all actual lexical entries thatinstantiate the type.

Type hierarchy

The hierarchical organization of lexical entries amalgamates

I the information associated with any one actual entry with

I the information associated with all of the generic entriesthat it instantiates.



A real-world ontology of linguistic objects



The lexicon: Example entries

Lexical entry: gives

phon⟨

gives⟩

ss | loc

cat

head verb

[vform fin

]subcat


,NP[acc] 2 ,PP[to] 3

⟩

cont

give′

giver 1

gift 2

given 3

A textbook solution (reference grammar)

I gives might specify that it inherits all constraints imposedby the two types ditransitive and third-singular-finite.

I The only specific information about gives is the phonologyof the base form, and semantic information.



The lexicon: Example entries

Lexical descriptions from English Resource Grammar

g i v e v 1 := v np−n p l e &[ ORTH < ” g i v e ” >,

SYNSEM [ LKEYS . KEYREL . PRED ” g i v e v 1 r e l ” ,PHON.ONSET con ] ] .

g i v e v 2 := v n p n o a r g 3 l e &[ ORTH < ” g i v e ” >,

SYNSEM [ LKEYS . KEYREL . PRED ” g i v e v 1 r e l ” ,PHON.ONSET con ] ] .

g i v e v 3 := v − l e &. . .

A language technology solution (resource grammar)

I gives might specify that it inherits all constraints imposedby type v np-np le.

I The only specific information about gives is the phonologyof the base form, and semantic information.



Using lexical rules

Once we have,

phon⟨

drink⟩

ss | loc

cat

head verb

[vform fin

]subcat

⟨NP[nom] 1 [plur]

,NP[acc] 2

⟩

cont

drink′

[drinker 1

drunken 2

]

using lexical rules, we get,

phon⟨

drinks⟩

ss | loc

cat

head verb

[vform fin

]subcat


,NP[acc] 2

⟩

cont

drink′

[drinker 1

drunken 2

]



The building blocks of HPSG grammars

In HPSG, words, phrases and sentences are all represented as signs

= complexes of phonological, syntactic/semantic, and discourseinformation.

An HPSG grammar consists of

I a lexicon licenses basic words

I lexical rules license derived words

I immediate dominance (id) schemata license constituentstructure

I linear precedence (LP) statements constrain word order

I a set of grammatical principles express generalizationsabout linguistic objects



Types of phrases

I HPSG does not treat phrase structure trees as formalobjects.

I Phrase structure is represented by the various daughtersattributes of phrasal signs.

I Trees are used as a convenient graphic representation.

I Each phrase has a dtrs attribute which has aconstituent-structure value

I This dtrs value corresponds to what we view in a tree asdaughters

I By distinguishing different kinds of constituent-structures, wecan define different kinds of constructions in a language



Types of phrases

constituent-struc

coord-struc

[conj-dtrs set(sign)

conjunction sign

]head-struc

[head-dtr sign

]

...[

head-comp-struc

comp-dtrs list(sign)

]



Example: Kim walks

phrase

phon⟨

kim, walks⟩

synsem S[fin]

dtrs

head-comp-struc

head-dtr

phrase

phon ⟨walks

⟩synsem VP[fin]

comp-dtrs

⟨phrase

phon ⟨kim

⟩synsem NP[nom]

⟩

Abbreviation:

S[fin]

VP[fin]

walks

NP[nom]

Kim

C H



Example: she drinks winephon

⟨she, drinks, wine

⟩ss | loc |cat

[head 3

subcat 〈〉

]dtrs head-comp-struc

phon

⟨drinks, wine

⟩ss | loc |cat

head 3

subcat⟨

1

⟩dtrs head-comp-struc

phon ⟨wine

⟩ss 2

phon

⟨drinks

⟩ss | loc |cat

head 3

verb

[vform fin

]subcat

⟨1 , 2

⟩

H C

phon ⟨she

⟩ss 1

C H



Universal principles

How do we specify the well-formed phrases?

A candidate phrase will be well-formed provided it satisfies allthe principles of grammar, including both universal principles andlanguage-specific principles.

Example

How exactly did the last example work?

I drink has head information specifying that it is a finiteverb and subcategories for a subject and an object.

I The head information gets percolated up (the HeadFeature Principle)

I The valence information gets “checked off” as one movesup in the tree (the Subcategorization Principle)



Head Feature Principle

Head Feature Principle (HFP)

The head value of any headed phrase is structure-shared withthe head value of the head daughter

phrase

synsem[loc |cat |head 1

]dtrs

[head-struc

head-dtr | synsem | loc |cat |head 1

]

The effect is to guarantee that headed phrases really are projec-tions of their head daughters.



Subcategorization Principle

Subcategorization Principle

In a headed phrase the subcat value of the head daughter isthe concatenation of the phrase’s subcat list with the list ofsynsem values of the complement daughters.

synsem

[loc |cat | subcat 1

]

dtrs

head-struc

head-dtr[synsem | loc |cat | subcat 1 ⊕ 2

]comp-dtrs synsem2sign( 2 )

The effect is to check off the subcategorization requirements ofthe lexical head as they become satisfied.



Again: she drinks winephon

⟨she, drinks, wine

⟩ss | loc |cat

[head 3

subcat 〈〉

]dtrs head-comp-struc

phon

⟨drinks, wine

⟩ss | loc |cat

head 3

subcat⟨

1

⟩dtrs head-comp-struc

phon ⟨wine

⟩ss 2

phon

⟨drinks

⟩ss | loc |cat

head 3

verb

[vform fin

]subcat

⟨1 , 2

⟩

H C

phon ⟨she

⟩ss 1

C H



A lexicalist framework (3)loc |cat [head 4

subcat 〈〉

]

loc |cathead 4

subcat⟨

1

⟩

3

Fido

2

Sandy

loc |cathead 4 verb[fin]

subcat⟨

1NP[nom][3rd,sing], 2NP[acc], 3NP[acc]⟩

gives

H C1 C2

1

Kim

C H

HPSG relies crucially on complex lexical information, which de-termines, in accordance with general principles such as the HFPand the Subcategorization Principle, the essential grammaticalproperties of phrasal expressions.



Immediate Dominance Schemata (1)

Rule or rule schema?

CFG X-bar CCG

VP → V NP V’ → V DP (s\np)/np np ⇒ s\npS → NP VP IP → DP I’ np s\np ⇒ s

/ construction-specific rules, highly schematic immediate dominance templates

Immediate Dominance Schemata in HPSG

I ID schemata serve as templates for permissible local phrasestructure trees or configurations of immediate constituency.

I ID schemata in HPSG occupy a position analogous to thatof X-bar schemata in GB as well combinatory rules in CCG.



Immediate Dominance Schemata (2)

Rule or rule schema?

CFG X-bar CCG

VP → V NP V’ → V DP (s\np)/np np ⇒ s\npS → NP VP IP → DP I’ np s\np ⇒ s

Universality

I ID schemata in HPSG are a small, universally available setof disjunctive constraints on the immediate constituency ofphrases, from among which each language makes aselection.

I The set of ID schemata should be regarded as adisjunctively specified principle of universal grammar.



Immediate Dominance Principle

The disjunction of the ID schemata itself constitutes a universalprinciple.

Immediate Dominance Principle (IDP)

The universally available options for a well-formed phrase are:

I (SCHEMA 1) a saturated ([subcat 〈〉]) phrase with dtrsvalue of sort head-comp-struc in which the head-dtrvalue is a phrasal sign and the comp-dtrs value is a listof length one; or

I (SCHEMA 2) an almost-saturated (subcat list of lengthone) phrase with dtrs value of sort head-comp-struc inwhich the head-dtr value is a lexical sign; or

I (SCHEMA 3) a saturated ([subcat 〈〉]) phrase with dtrsvalue of sort head-comp-struc in which the head-dtrvalue is a lexical sign.

I ...Weiwei Sun Head-driven Phrase Structure Grammar 51/77


ID schemata (1)

SCHEMA 1

A saturated ([subcat 〈〉]) phrase with dtrs value of sort head-comp-struc in which the head-dtr value is a phrasal sign andthe comp-dtrs value is a list of length one.

synsem | loc |cat

head[

verb

inv −

]∨ ¬ verb

subcat 〈〉

dtrs

head-comps-struc

head-dtr phrase

comp-dtrs⟨

sign⟩



ID schemata (1)synsem | loc |cat

[subcat 〈〉

]

dtrs

head-comps-struc

head-dtr phrase

comp-dtrs⟨

sign⟩

With Head Feature Principle and SubcategorizationPrinciple [

head 1

subcat 〈〉

]

head 1

subcat⟨

2

⟩2

C H



ID schemata (2)

SCHEMA 2

An almost-saturated (subcat list of length one) phrase withdtrs value of sort head-comp-struc in which the head-dtrvalue is a lexical sign.

synsem | loc |cat

head[

verb

inv −

]∨ ¬ verb

subcat⟨

synsem⟩

dtrs

[head-comps-struc

head-dtr word

]



ID schemata (2)

synsem | loc |cat

head verb

[inv −

]∨ ¬ verb

subcat⟨synsem

⟩

dtrs

[head-comps-struc

head-dtr word

]

With Head Feature Principle and SubcategorizationPrinciple head 1

subcat⟨

2

⟩

n...3head 1

subcat⟨

2 ,..., n⟩

H C1 Cn−2



ID schemata (3)

SCHEMA 3

A saturated ([subcat 〈〉]) phrase with dtrs value of sort head-comp-struc in which the head-dtr value is a lexical sign.

synsem | loc |cat

head[

verb

inv +

]subcat 〈〉

dtrs

[head-comps-struc

head-dtr word

]



ID schemata (3)synsem | loc |cat

head verb

[inv +

]subcat 〈〉

dtrs

[head-comps-struc

head-dtr word

]

With Head Feature Principle and SubcategorizationPrinciple [

head 1

subcat 〈〉

]

n...2head 1

subcat⟨

2 ,..., n⟩

H C1 Cn−1



Practice

I Nothing in HPSG forces you to use constituency trees. Thereare only features which encode objects akin to trees.

I Technically, types license particular schemata (e.g.,head-comps-struc), and a dtrs list keeps track of theconstituent daughters

head−complement−r u l e −0 := p h r a s e &[ HEAD #h ,

SPR #s ,COMPS #c ,ARGS < word & [ HEAD #h , SPR #s , COMPS #c& < > ] > ] .



she drinks wine

phrase

s | l |c[head 3

subcat 〈〉

]

dtrs

head-comps-struc

comp-dtrs

⟨phon ⟨she

⟩synsem 1

⟩

head-dtr

phrase

s | l |c

head 3

subcat⟨

1

⟩

dtrs

head-comps-struc

head-dtr

word

phon⟨

drinks⟩

s | l |c

head3

[verb

vform fin

]subcat

⟨1 , 2

⟩

comp-dtrs

⟨phon ⟨wine

⟩synsem 2

⟩



Word order (1)

(3) a. Kim can go

b. Can Kim go

[head 1

subcat 〈〉

]

head 1

subcat⟨

2

⟩

3

go

head 1 verb[fin,+AUX,−INV]

subcat⟨

2NP[nom], 3VP[base]⟩

can

H C

2

Kim

C H

How to describe an auxiliary word?



Word order (2)

Schemata 1 and 2 are employed in uninverted structures, whileSchema 3 is employed in inverted structures.

I In availing itself of Schemata 1 and 2, English imposes onthem the further parochial condition that the head valuemust bear the specification [-INV], while

I in availing itself of Schema 3, English imposes on it theparochial condition that the head value must bear thespecification [+INV].



Word order (3) [head 1

subcat 〈〉

]

3

go

2

Kim

head 1 verb[fin,+AUX,+INV]

subcat⟨

2NP[nom], 3VP[base]⟩

can

H C1C2

No movement.

Question

How can we build the predicate-argument relation between Kimand go?



Control

Question

How can we build the predicate-argument relation in a controlconstruction?

word

phon try

synsem | loc

cat

subcat⟨NP 1 ,

loccat

head verb

subcat⟨NP 1

⟩cont 2

⟩

cont

try’

tryer 1

tried 2



Complementizers (1)

Problem of GB’s analysis

(4) a. Kim said that Sandy left.

b. I demand that he leave/*leaves immediately.

/ The complementizer heads the complementized clause ⇒How demand selects the complement’s head verbinflection?

, The head of that-S, namely demand, subcategorizes forS[bse] rather than S[fin].

Proposal: marker

that is a marker, a new type of part-of-speech.



Complementizers (2)

Proposal: Add marking to catss | loc |cathead ...

subcat ...

marking marking

marking

unmarkedmarked

comp

...forthat



Complementizers (3)

A marker selects its head

I Markers are not heads, but they resemble heads inasmuchas they select the phrases that they mark.

I that selects S[fin∨bse]I for selects S[inf]

I Markers bear the head feature specified ([spec synsem]).

I A marker’s spec value is structure-shared with thesynsem value of the (head) sign

SCHEMA 4

A phrase with drts value of sort head-marker-structure whosemarker daughter is a marker whose spec value is structure-sharedwith the synsem value of the head daughter, and whose mark-ing value is structure-shared with that of the mother.



Complementizers (4)

Schema 4 licenses head 3

subcat 4

marking 1

2

[head 3

subcat 4

]head

marker

[spec 2

]subcat 〈〉marking 1marked

M H



Complementizers (5)

cat value of thathead

marker

[spec S[fin∨bse,unmarked]

]subcat 〈〉marking that

Example

S[fin,that]

S[fin]

John left

that

M HS[bse,that]

S[bse]

John leave

that

M H



Adjunct

Proposal: An adjunct selects its head

cat

head

adj

mod N’:

[index 1

restr 2

]prd −

subcat 〈〉

cont

index 1

restr

[reln red

arg 1

]∪ 2

SCHEMA 5

A phrase with dtrs value of sort head-adjunct-structure suchthat the mod value of the adjunct daughter is structure-sharedwith the synsem value of the head daughter.



Key properties of HPSG and their consequences

I HPSG is monostratal, declarative, non-derivational.I No transformations, no rule ordering.I Analyses are surface oriented, with a desire to avoid abstract

structure such as traces and functional categories.I HPSG is pure constraint-based

I A structure is well-formed iff it satisfies all relevant constraints.I Constraints are not violable

I Lexical entries, ID rules, principlesI HPSG is a lexicalist theory

I Strong lexicalism; Word-internal structures and phrasestructure are handled separately.

I ‘Parallel’ architecture organized by signs; Grammar is infusedby semantics; Modular, yet interactive

I Structure is ‘minimal’ (not full of empty elements and extrastructure)

I HPSG is a unification-based linguistic frameworkI All linguistic objects are represented as typed feature

structures.



HPSG vs. Classical Phrase Structure Grammar

Similarities

I Both are monostratal: every analysis is represented by asingle structure

I Grammar rules have local scope: parent phrase and itsimmediate children

Differences

I HPSG uses complex categories while classical PSG usessimple/atomic ones

I The relation between entities in HPSG is formulated aswell-formedness constraints on typed feature structures.

I ID schemata specify the parent and children in a local tree.I Universal principles are specified to constrain the set of local

trees admitted by the ID schemata.

I HPSG analyses include semantic representations in addition tosyntactic representations



HPSG vs. Transformational Grammar

Similarities

I Both try to account for a similar range of data (e.g. valence)

I Both are theories of generative grammar

Differences

I HPSG is non-derivational, TG is derivational

I HPSG uses more complex categories than TG

I HPSG is more committed to precise formalization than TG

I HPSG is better suited to computational implementation thanTG



Representational or Derivational (1)

Two categories of grammarsI Derivationally oriented grammarsI Representationally oriented grammar

Derivationally oriented grammar

A grammar generally include a set of structural atoms (the basis)of the derivation.The derivational procedure constructs syntactic structures usingoperations of two types.

1. Structural composition: Either previously constructedsyntactic representations or elements of the basis arecombined to form larger representations.

⇒ Fundamental: Such operations provide a way to generatethe requisite infinity of possible structures.

2. Transformations: Modify an individual syntacticrepresentation in some specified fashion.




Two categories of grammarsI Derivationally oriented grammarsI Representationally oriented grammar

Representationally oriented grammar

A grammar determines the set of linguistic expressions using asystem of well-formedness constraints.

I Each constraint provides an evaluation of some part of thelinguistic expression. The well-formedness of the entirelinguistic expression is determined by combining togetherthe evaluations of the individual constraints.

I Representationally oriented grammarsI don’t specify how to find well-formed linguistic

expressions,I but only what properties well-formed expressions must

have.




A key concern

[...] there are typically powerful generalizations sur-rounding particular surface forms that are more broadthan those captured by derivations or transformations.

Surface Generalization Hypothesis

There are typically broader syntactic and seman-tic generalizations associated with a surface argumentstructure form than exist between the same surfaceform and a distinct form that it is hypothesized to besyntactically or semantically derived from.

from A. Goldberg’s Constructions at Work: The nature ofgeneralizations in language



Universal Grammar

I Linguistic ontology: the inventory of universally availablesorts of linguistic entities, together with a specification oftheir appropriate attributes and their value sorts.

I Schemata: a small, fixed inventory of universally availablephrase types (schematic immediate dominance rules), forexample, head-complement, head-adjunct, head-marker, etc.

I Universal constraints on well-formed phrases: HeadFeature Principle, Subcategorization Principle, etc.



Particular Grammar

I Lexicon: a system of lexical entries (possibly interrelated bylexical rules).

I Linguistic ontology: selection from and further articulationof the universal linguistic ontology.

I Schemata: selection from and further specification (e.g. forthe constituent order) of the universally available schemata.



Reading

I Chapter 1, Head-driven Phrase Structure Grammar


head-driven phrase structure grammar - github pages · 2020-02-05 · subj 9 pred % agreement pers...

Documents